gonzalezpazmonica/pm-workspace

v2.24.0MIT

Complete PM automation with AI agents — Azure DevOps, Jira, Scrum, SDD pipeline

GitHub

Changelog — pm-workspace

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased] — 2026-08-17 · Human authority for learning recall (SCL-008)

Changed

Learning recall now injects only active, human-authored principles resolved from CRITERIO.md; inferred and proposed lessons remain measurable shadow candidates and cannot become instructions automatically.
Claude Code and OpenCode use the same canonical recall hook and authority filter, including hybrid ranking and federated lessons.

[Unreleased] — 2026-07-19 · Neuro-inspired Orchestration (SE-268)

Added (SE-268 — PR #907)

S1: Constitutional veto bus (bridge/control-plane/). Independent control plane with priority over A2A, compile CRITERIO.md linea_roja → permanent vetos, fail-closed semantics, scope global|domain|instance|session with TTL.
S2: Salience-based federation (bridge/salience/). Deterministic 0-1 scoring (novelty, impact, criticality, recurrence), anti-stagnation detector with elevation, configurable per-context thresholds.
S3: Hook thresholds by domain (config/hook-thresholds.yaml). Three exigency levels (conservador/estandar/experimental), anti-relaxation gate for linea_roja, audit log of all threshold decisions.
S4: Two-speed memory (scripts/memory-two-speed.py). Episodic (fast) + semantic (slow) stores on bitemporal axis, selective replay consolidation, dome-context-indexed retrieval, promoted-to-used quality metric.

[Unreleased] — 2026-07-13 · Agent Shell Safety (v2) — PR #906 extension

Added: SE-266 v2 — extended agent-git-discipline.sh to block destructive shell operations after session catastrophique 2026-07-13.

BLOCK: rm -rf, rm -r, rm --recursive (recursive delete)
BLOCK: rm without -i/--interactive (except safe paths)
BLOCK: dd of=/dev/, mkfs., chown -R on user home
BLOCK: truncate of home files
Safe paths: /tmp/opencode/, /tmp/recovery/
27/27 manual tests pass (30 BATS in test-se-266-agent-git.bats)
Spec updated with v2 scope + lesson learned section

[Unreleased] — 2026-07-12 · Agent Git Governance — PR #906

Added: SE-266 — agent git discipline rules for concurrent agent safety, inspired by Pi (earendil-works/pi AGENTS.md).

agent-git-discipline.sh hook blocks destructive git ops (reset --hard, clean -fd, stash, checkout .) and warns on global stage (git add -A/.), registered in settings.json PreToolUse Bash matcher.
docs/AGENTS.md Git Discipline for Concurrent Agents section.
12/12 BATS pass.

[Unreleased] — 2026-07-12 · Court model tier assignment — PR #905

Added: SE-265 — per-judge model tier assignment in rules/court.rules.yaml. security/correctness → heavy, architecture → mid, cognitive/spec → fast. Safety: critical judges require heavy, operator override allowed. 10/10 BATS pass. Inspired by gentle-ai v2.0.0 configurable review models.

[Unreleased] — 2026-07-11 · Federacion de Savias — PR #904

Added: SE-263 — Federacion de Savias soberanas mediante cupulas de contexto, plano de estado git, A2A sobre red privada, y Agent Index. 7 slices (S1-S7): coordination model, instance identity, exportable domes, git plane + agent index, A2A server (5 skills), exchange ledger, federation drill + atestacion. 39/39 BATS pass.

[Unreleased] — 2026-07-11 · gentle-ai v1.49 patterns — PR #903

Added: 4 patrones de gentle-ai v1.49 adaptados a Savia.

S4: scripts/receipt-v2.sh — recibos de revisión ligados a contenido (patch-id estable).
S1: rules/court.rules.yaml — ciclo de revisión acotado del Court (freeze, path-set, verificación dirigida).
S2: config/native-delegation.yaml — delegación nativa OpenCode con allowlist deny-by-default.
S3: docs/managed-artifacts-contract.md + scripts/lib/managed_artifacts.py — contrato ownership-safe para artefactos gestionados.
24/25 BATS tests pass.

[Unreleased] — 2026-07-11 · CodeFlow-inspired health tools — PR #902

Added: 3 nuevas herramientas de salud del workspace inspiradas en CodeFlow.

SE-260 scripts/blast-radius.sh — analiza dependientes de un fichero con control de profundidad y riesgo.
SE-261 scripts/workspace-health.sh --v2 — extiende health dashboard a 9 dimensiones (blast radius, code ownership, dead code).
SE-262 scripts/pr-thermal-receipt.sh — genera recibos térmicos con delta de métricas para PRs.
3 specs SDD en docs/specs/SE-26*, 30 BATS tests (27 pass).

[Unreleased] — 2026-06-26 · Enterprise 23 specs — PR #875

Added: 23 specs enterprise implementadas en 5 batches (237 BATS + 1416 pytest). Fixed: bug en plugin block-force-push.ts — regex /git\s+push\s+.*--force/i bloqueaba falsamente --force-with-lease en ramas feature. Añadido lookahead negativo (?!-with-lease) y soporte para flag -f. 12 tests nuevos (TDD).

SE-006 Workspace Compliance Validator (30 scripts enterprise)
SE-023 Knowledge Federation
Rules: authority-claims, criterion-simulation-honesty, enterprise-governance, enterprise-licensing, enterprise-mcp, enterprise-onboarding, enterprise-sovereign, sandbox-os-policy
Hooks SPEC-193: context-sanitize-input, criterion-simulation-challenge, memory-write-sanitize, re-anchor-redlines
Agent criterion-simulation-judge: restored ejemplo banner_text FRAME_DOUBT

and this project adheres to Semantic Versioning.

[Unreleased] — 2026-06-26 · SE-229 Slice 1 — Session Registry MVP

Added: sincronización de sesiones Savia paralelas (SE-229, Slice 1).

scripts/session-registry.sh: registro JSONL de sesiones activas con

operaciones register, claim, release, list y gc. Protección de escrituras concurrentes mediante flock. Fallback a python3 cuando jq no está disponible.
docs/rules/domain/parallel-session-protocol.md: protocolo que obliga a toda

sesión autónoma a registrarse antes de operaciones git, reclamar su rama y hacer release al terminar. Evita conflictos entre nidos y sesiones paralelas.
.claude/hooks/session-init.sh: hook de arranque que registra la sesión y

reclama la rama automáticamente cuando SAVIA_NIDO está set.
.claude/hooks/session-end-snapshot.sh: hook de cierre que libera la sesión

y ejecuta gc sobre entradas stale (heartbeat >10 min).
tests/test-session-registry.bats: 15 tests BATS que cubren register,

claim conflict, release, gc, list y edge cases de sesión duplicada.

[Unreleased] — 2026-06-13 · Docs/SCM gaps fix tras PR #844

Fixed: cinco gaps detectados en auditoria post-merge.

generate-capability-map.py: scan_scripts ahora recorre subdirs depth-1

(scripts//.sh). Antes solo top-level scripts/.sh, dejando 10 scripts del recommendation-tribunal sin indexar en .scm.
8 nuevos scripts indexados en .scm/INDEX.scm (aggregate, banner,

calibrate, classifier, early-cancel, expertise-rewrite, followup-record, iterate). Total recursos 1265 a 1296.
SPEC-195/196/197/198/200 frontmatter status PROPOSED a IMPLEMENTED.

Las cinco specs llevan implementadas, wireadas (#844) y activadas en pilot (#845) pero el frontmatter no se actualizo al merge.
ROADMAP refleja PRs (#844 wiring + #845 telemetry pilot) en la

tabla de specs implementadas. Nueva seccion Telemetry pilot 30d.
Knowledge graph rebuilt: ahora incluye SPEC-189 hasta SPEC-200.

Antes faltaban 12 specs.

Determinismo verificado: dos regeneraciones consecutivas del SCM producen el mismo hash. test-scm-determinism 19/19 pass.

[Unreleased] — 2026-06-13 · Activate SPEC-198/200 telemetry pilot defaults

Changed: tras mergear PR #844, los toggles SPEC-198/200 estaban presentes pero defaultaban a off. Esta entrada activa modo warn por default durante ventana de pilot 30 dias, segun decision pactada con el usuario.

savia-env.sh exporta SAVIA_QUALITY_GATE_ADAPTIVE=warn y

SAVIA_JUDGE_VERDICT_VALIDATE=warn por default cuando no hay override.
CI workflow ci.yml setea SAVIA_QUALITY_GATE_ADAPTIVE=warn en el step

Gate Test Quality y sube output/quality-gate-history.jsonl como artifact (retencion 30 dias).
11 tests nuevos en test-telemetria-pilot-defaults.bats que verifican

defaults, overrides explicitos, e2e generation de los dos JSONL.

Backward compat: cualquier sesion que exporte explicitamente SAVIA_QUALITY_GATE_ADAPTIVE=off o SAVIA_JUDGE_VERDICT_VALIDATE=off mantiene comportamiento anterior. Promocion warn->on tras revision de telemetria de 30 dias.

[Unreleased] — 2026-06-13 · Wire SPEC-195/196/197/198/200

Added: wiring opt-in de cinco specs implementadas previamente a sus consumidores reales en produccion. Defaults conservadores backward-compat.

SPEC-200 wired al CI test quality gate (adaptive threshold opt-in)
SPEC-198 wired al tribunal aggregator (judge verdict validation opt-in)
SPEC-197 wired al tribunal iterate (subcomando compute-temperature)
SPEC-196 documented en orchestrator agent prompt (early-cancel doc)
SPEC-195 documented en orchestrator agent prompt (iterative loop doc)

Tests bats verdes, gate PASS, suite pytest sin regresion. Detalle en PR.

Out of scope: SPEC-194 no existe todavia, tribunal-async-runner no modificado, orchestrator prompt LLM no shell glue. Branch dedicada.

[Unreleased] — 2026-06-04 · SPEC-188 Root-Cause Investigation Architecture (meta-spec)

Added: docs/propuestas/SPEC-188-root-cause-investigation-architecture.md — meta-spec PROPOSED que coordina 4 sub-specs existentes (SPEC-043 responsibility-judge, SPEC-065 execution-supervisor, SPEC-108 sentry-rca, SPEC-125 recommendation-tribunal) e identifica 5 gaps no cubiertos en el sistema causal actual: G1 sealed contract tests, G2 calibration POST-CODE-CHANGE, G3 failure pattern memory con frequency, G4 diagnostic quality metrics longitudinal, G5 decision trace artifact. Define 5 piezas nuevas (P1-P5) con esquema, ubicacion, contratos cruzados, y plan fasificado (Fase 0-4). Fase 0 cierra deuda tecnica detectada: feedback_root_cause_always.md referenciado por memory-conflict-judge pero inexistente en path canonico. Era 200 Wave 1 Tier 1 P0. Estimacion 6-8h spec doc; ~80-120h implementacion completa via sub-specs ejecutables.

[Unreleased] — 2026-06-04 · SPEC-158 Workflow vs Agent decision gate

Added: /decide-architecture command + scripts/decide-architecture.sh heuristic classifier with three-tier signals (weak +1, medium +2, strong +5) and Anthropic workflow-first bias (+1 base). Corpus scripts/decide-architecture-corpus-test.sh covers 20 curated tasks (10 workflow, 10 agent) with 100% accuracy (AC: 85% min). BATS suite tests/test-decide-architecture.bats — 31/31 pass, audit 93/100. Doc rule docs/rules/domain/workflow-vs-agent-decision-gate.md. Tier 1C from anthropic-effective-agents-thesis.

[Unreleased] — 2026-06-04 · SPEC-184 Write-time non-blocking validators

Added: PostToolUse Edit|Write hook .opencode/hooks/post-write-validate.sh that runs composable markdown validators and warns to stderr without blocking. 4 validators shipped: banned-unicode (em-dash, en-dash, curly quotes, NBSP, ellipsis), frontmatter (SPEC docs require yaml block), spec-status (enum), memory-entry-length (150 char cap). Bypass for output, .git, node_modules, dist, raw. Toggle via SAVIA_WRITE_VALIDATORS_ENABLED. Latency 23ms measured on small file. BATS suite tests/test-write-time-validation.bats — 29/29 pass, audit 91/100. Doc rule docs/rules/domain/write-time-validation.md. Era 199 Wave 1 Tier 1.

[Unreleased] — 2026-06-04 · SPEC-180 Sentinel-safe regeneration

Added: scripts/sentinel-regen.sh primitive (inject/extract/verify-hash) with sha256-8 hash drift detection. Contract doc at docs/rules/domain/sentinel-safe-regen.md. Opt-in piloto on AGENTS.md auto-regen via SENTINEL_MODE=1. BATS suite 17 tests, audit 90/100. Era 199 Wave 1 Tier 1. Unblocks SPEC-181.

[Unreleased] — 2026-06-04 · SPEC-185 Critical-facts anchor

Added: docs/critical-facts.md anchor with 150-token cap, validator, auto-regenerator, policy doc, CLAUDE.md integration, BATS suite (21 tests, audit 86/100). Era 199 Wave 1 Tier 1.

[Unreleased] — 2026-06-03 · SE-091 Caveman always-on

Added

SE-091 closed (Era 195, rank 11). Caveman como comportamiento por defecto.
- docs/rules/domain/caveman-default.md ya existía; 6 restricciones cargadas en cada turno via instructions en opencode.json (verificado AC-02).
- .opencode/plugins/guards/auto-grill-me.ts — guard non-blocking: Edit/Write sobre código → console.warn con recordatorio grill-me (edge cases, error handling, unstated assumptions).
- .opencode/plugins/guards/auto-zoom-out.ts — guard non-blocking: Edit/Write sobre ROADMAP/docs/architecture → console.warn con recordatorio zoom-out (dependencies, second-order effects).
- Ambos guards registrados en BEFORE_GUARDS de savia-foundation.ts.
- tests/test-caveman-always.bats — 21/21.
- .opencode/plugins/__tests__/auto-grill-me.test.ts — 14/14 TS.
- .opencode/plugins/__tests__/auto-zoom-out.test.ts — 15/15 TS.

[Unreleased] — 2026-06-03 · SE-086 Ubiquitous Language extractor

Added

SE-086 closed (Era 190 stack rank 10). DDD ubiquitous language extractor

para glosarios per-proyecto. Clean-room de mattpocock/skills (MIT).
- .claude/skills/ubiquitous-language/{SKILL,DOMAIN}.md — skill con proceso 5 pasos, triggers, atribución MIT, formato CONTEXT.md.
- scripts/extract-domain-entities.py — extractor: report + --auto-update con marcadores [REVIEW]. Statuses: new / existing / inconsistent.
- scripts/knowledge-graph-domain-bridge.py — ingesta CONTEXT.md en SE-162 grafo con edges DOMAIN_TERM.
- docs/rules/domain/ubiquitous-language.md — rule doc.
- tests/test-ubiquitous-language.bats — 24/24 AC-01..AC-09 pass.
Desbloquea: SE-171 (contradiction detector) gana términos de dominio.

[Unreleased] — 2026-06-02 · SE-162 Knowledge Graph sobre memoria Savia

Added

SE-162 closed (Era 251 stack rank 8). Grafo de conocimiento tipado

sobre memoria Savia con SQLite WAL. Sin pip deps (stdlib Python).
- scripts/knowledge-graph.py — engine: build/query/impact/status/entities.
- scripts/knowledge-graph.sh — wrapper bash.
- docs/rules/domain/knowledge-graph.md — schema, tipos, comandos.
- tests/test-knowledge-graph.bats — 26/26 pass.
- ~/.savia/knowledge-graph.db — SQLite gitignored (derived, regenerable).
- Línea base 2026-06-02: 543 entidades (211 specs, 209 rules…), 661 relaciones.
Comandos actualizados (graph-build.md, graph-query.md) con sección

SE-162 backend para queries al SQLite global.
Desbloquea: SE-086 (ubiquitous-language), SE-171 (contradiction detector).

[Unreleased] — 2026-06-02 · ROADMAP drift cleanup (SE-081 + SE-082 + SE-083 + SE-084 + SE-093)

Changed

ROADMAP table sync (rank 3, 4, 5, 6, 9). Cinco specs marcados IMPLEMENTED

tras verificar realidad en disco:
- SE-081: 3 skills (caveman, zoom-out, grill-me) con SKILL.md+DOMAIN.md.
- SE-082: docs/rules/domain/architectural-vocabulary.md + auditor + tests 31/31 + audit TSVs diarios en output/.
- SE-083: .claude/skills/tdd-vertical-slices/ + tests SE-083 en architectural-vocabulary.bats.
- SE-084: scripts/skill-catalog-audit.sh operativo, audit TSVs diarios.
- SE-093: scripts/project-context.sh + .claude/hooks/project-isolation-gate.sh, merged en PR #783 (2026-05-30).
Stack priorizado en docs/ROADMAP.md ahora refleja estado real.

[Unreleased] — 2026-06-02 · SE-167 Skill Maturity Kanban

Added

SE-167 closed (Era 252 stack rank 7). Auditor que clasifica las 98

skills del workspace en 4 estados canónicos:
- Calibrated (stable + tests + DOMAIN.md)
- Incomplete (falta tests o maturity != stable)
- Stub (missing DOMAIN.md o SKILL <50 líneas)
- Deprecated (frontmatter deprecated: true)
scripts/skill-maturity-audit.sh — auditor con flags --tsv-only,

--markdown-only, --help. Output diario en output/.
docs/rules/domain/skill-maturity-kanban.md — reglas + transiciones.
docs/propuestas/SE-167-skill-maturity-kanban.md — spec.
tests/test-skill-maturity-audit.bats — 23 tests pass.
Línea base 2026-06-02: 1 Calibrated, 94 Incomplete, 3 Stub, 0 Deprecated.

[Unreleased] — 2026-06-02 · SE-153 Template SKILL.md "Authoritative Paths First"

Added

SE-153 closed (Era 197 / Tier 1 free-wins). Plantilla canónica para

crear nuevas skills con el patrón "Authoritative Paths First" (inspirado en flowsint). Reduce coste de contexto: paths canónicos antes de la prosa explicativa, eliminando inferencia incorrecta de firmas/paths.
- .claude/skills/_template/SKILL.md: template de 112 líneas con Authoritative Paths obligatoria antes de Workflow y Decision Checklist. Visible vía symlink desde .opencode/skills/_template/.
- docs/rules/domain/skill-template-protocol.md: managed-content rule (68 líneas) — uso, reglas duras, migración opt-in, validación.
- tests/test-skill-template.bats: 34/34 pass — frontmatter, ordenamiento de secciones (Authoritative Paths antes de Workflow/Decision), 150-line cap, copyability, exclusión en generadores. Auditor: 88/100.
- scripts/skills-md-generate.sh y scripts/resolver-md-generate.sh: excluyen _template con ! -path '*/_template/*' para evitar contaminar catálogos auto-generados.
- CLAUDE.md: lazy reference añadido.

Changed

docs/ROADMAP.md: SE-153 PROPOSED → IMPLEMENTED en Tier 1.

Notes

Migración opt-in: skills existentes NO se migran masivamente. Cada skill

se actualiza al patrón cuando se toca por otra razón.

[Unreleased] — 2026-06-02 · SE-160 RESOLVER.md dispatch explícito

Added

SE-160 closed (Era 251 free-wins). Tabla explícita intent → skill/agent

para reducir carga del contexto central y permitir routing editable sin prompt engineering. Patrón GBrain adaptado.
- docs/RESOLVER.md: 98 skills + 70 agents en tablas auto-generadas; sección OVERRIDE hand-curated con sinónimos comunes (es/en).
- scripts/resolver-md-generate.sh: generador con modos generate (stdout), --apply (reemplaza solo bloque AUTO_BEGIN/AUTO_END preservando OVERRIDE), --check (drift detection para CI). Soporta frontmatter multilinea (description: >).
- docs/rules/domain/resolver-protocol.md: managed-content rule (62 líneas).
- tests/test-resolver-md-generate.bats: 21/21 tests — estructura, drift detection, idempotencia, preservación de OVERRIDE, validación de targets.
- CLAUDE.md: lazy reference añadido.

Changed

docs/ROADMAP.md: SE-160 PROPOSED → IMPLEMENTED en tablas de priorización

Era 251.

[Unreleased] — 2026-06-02 · SPEC-186 double opt-in for autonomous skills

Added

SPEC-186 closed (status: PROPOSED → IMPLEMENTED). Era 199 Wave 1.

Closes the accidental-activation vector via inherited env vars in persistent shells / CI runners.
- scripts/savia-double-optin-check.sh: canonical helper requiring BOTH a persistent env var AND --confirm-autonomous flag per invocation. Exit codes: 0 ok · 1 missing factor · 2 invalid invocation. Test bypass only when SAVIA_TESTING=1 AND inside BATS. Audit log at output/agent-runs/optin-audit.log (override via SAVIA_OPTIN_AUDIT_LOG).
- 24 BATS tests in tests/test-savia-double-optin-check.bats covering all 4 confirmation combinations × multiple skills, env var mapping, audit logging, and bypass semantics. 24/24 pass.
- 5 SKILL.md files updated with the double opt-in gate: overnight-sprint, code-improvement-loop, adversarial-security, tech-research-agent, savia-dual.
- docs/rules/domain/autonomous-safety.md: new appendix documenting the principle, helper, skill→env-var mapping, audit, and bypass.

[Unreleased] — 2026-06-01 · SPEC-155 close-out + silent-regression sweep

Fixed

SPEC-155 closed (status: PROPOSED → IMPLEMENTED). Core args-shape fix

was already applied in lib/hook-input.ts. This PR completes the spec:
- 4 silent-regression tests fixed (88/88 plugin tests pass; was 83/88):
  - prompt-injection-guard: added .opencode/{agents,skills,rules,commands,profiles} patterns + AGENTS.md / SKILLS.md. Pre-fix, prompt injections in .opencode/agents/*.md passed through silently.
  - block-gitignored-references: added /.opencode/hooks/ to SOURCE_SELF_REFS. Pre-fix, edits to hook source files emitted false positives on legitimate self-references.
  - Test assertion in block-gitignored-references.test.ts aligned to actual error message (was matching /output/, error says "dated").
  - savia-foundation.test.ts Edit-on-docs test resolved against real README.md via import.meta.url (independent of suite cwd).
Derived rule: docs/rules/domain/external-contract-testing.md —

every module dialoguing with an external system must have at least one integration test using the documented shape, copied literally from the external doc. Prevents silent regressions on frontend migrations.

[Unreleased] — 2026-06-01 · Era 199 obsidian-inspired SPECs (SPEC-180..186)

Added

Era 199 (7 SPECs PROPOSED, 22-32h en 3 waves): transferencias del repo

obsidian-second-brain compatibles con SDD/Rule #8/#24/#25.
- SPEC-180 sentinel @generated/@user safe-regen (P1 Wave1, 2-3h)
- SPEC-181 L0-L3 context token budgets por tier (P2 Wave2, 4-6h, dep 180)
- SPEC-182 bitemporal timeline frontmatter (P2 Wave2, 6-8h)
- SPEC-183 reconciliation 3-bucket auto/manual/drift (P3 Wave3, 5-7h, dep 182)
- SPEC-184 write-time validator non-blocking warn (P1 Wave1, 3-4h)
- SPEC-185 critical-facts 150-token cap anchor (P2 Wave1, 1-2h)
- SPEC-186 double opt-in para gates autonomos (P1 Wave1, 1-2h) Origen: output/research/obsidian-second-brain-mejoras-cupulas-20260601.md. Rechazados explicitamente: write-time auto-rewriting (viola Rule #8), Two-Output Rule, bg-agent PostCompact default, integraciones Perplexity/Grok, sistema de 4 presets.

[Unreleased] — 2026-05-31 · SPEC-156 Slice 2 + LeCun roadmap (SPEC-163..168)

Added

SPEC-156 Slice 2: PreToolUse Task hook

spec156-token-budget-projection.sh projects token cost from subagent_type frontmatter token_budget before dispatch. Emits telemetry to output/agent-runs/budget-projections.jsonl (verdicts: ok / warn >80% / exceeded). Enforcement gated by SAVIA_BUDGET_ENFORCEMENT env (default warn; block + exceeded + escalation_policy: block → exit 2). Dual-syntax YAML parser (block
- flow). 18/18 BATS PASS.
SPEC-163: System 1/2 dispatch router (Tier 1A, fast/heavy routing).
SPEC-164: Memory feedback loop — auto-memory writes from outcomes

(Tier 1A, 5h).
SPEC-165: World model simulation — pre-action simulator (Tier 2, 12h).
SPEC-166: Configurator explicit — centralize System 2 dispatch

(Tier 2, 8h).
SPEC-167: Critic RAG — BM25 over memory for judges (Tier 2, 6h).
SPEC-168: Actor iterative pre-action — inner loop, blocked by

SPEC-165 (Tier 3, 10h).

Changed

70 agent frontmatter migrated flat token_budget_* → nested

token_budget: {per_invocation, context_window_target, escalation_policy}. 8 mirrors converted to flow form to keep ≤150 lines (Rule #11).
Agent size baseline .ci-baseline/agent-size-violations.count 27→30.

Honest ratchet: 3 obese agents (code-reviewer 6869B, security-guardian 6531B, test-runner 6519B) deserve own refactor SPEC; ratchet protects at new floor.

[Unreleased] — 2026-05-30 · SPEC-155 plugin hook args shape fix [CRITICAL]

Fixed

SPEC-155: align plugin hook signature with OpenCode v1.14+ contract.

Pre-fix, all 12 guards (validate-bash-global, credential-leak, data-sovereignty-gate, tool-call-healing, etc.) read input.args (undefined in real runtime) instead of output.args, operating on empty data = security theater. Symptom: spurious BLOCKED [tool-healing]: read called with empty file_path on valid tool calls. Helpers now prefer output.args with retro-compat fallback to input.args. 4 golden tests added with real v1.14 shape.

[Unreleased] — 2026-05-23 · savia-evolution incorporation

Generic improvements ported from a downstream private fork. Confidential references and internal IDs scrubbed; only features with public utility outside the fork are included. Branch: feat/incorporate-savia-evolution.

Added

SPEC-144: /speckit.* slash command aliases for cross-tool

compatibility with github/spec-kit. New thin redirector commands (/speckit.specify, /speckit.plan, /speckit.tasks, /speckit.implement, /speckit.constitution, /speckit.clarify, /speckit.analyze, /speckit.checklist) routing to existing Savia SDD flow. Zero behaviour change, additive only.
scripts/savia-env.sh: _resolve_project_slug() (auto-detect via env,

branch prefix, or single subdir under projects/), _savia_load_dotenv() (workspace .env loader with parent-env precedence) and _savia_activate_venv() (auto-activate workspace .venv/). New CLI action project and a project field in json output.
scripts/lib/os-detect.sh (new): detect_os() + setup_paths() with

Windows/macOS/Linux defaults for OLLAMA_BIN, ANDROID_HOME, JAVA_HOME. emergency-setup.sh and savia-shield-setup.sh now source it.
scripts/embedding-server.py (new): in-process HTTP server keeping a

sentence-transformers model resident to avoid the 15-30s cold-start per recall.
scripts/memory-status-check.sh (new): L0-L4 + agents + vault + hooks

status table, with a "frontend without hooks" warning banner.
.claude/commands/exit.md (new): /exit persists ROADMAP.md and

writes a RESUME-{slug}.md. Opt-in commit, no autonomous merges.

Changed

scripts/memory-store.sh: hard $HOME guard (no more /tmp fallback)

and lazy-start of embedding-server.py before search/recall.
scripts/savia-shield-daemon.py: /health now reports

ollama_model_loaded and ollama_model (env OLLAMA_CLASSIFY_MODEL, default qwen2.5:7b).
scripts/savia-shield-proxy.py: docs reflect provider-agnostic use;

glossary search expanded to projects/<umbrella>/GLOSSARY.md and projects/<umbrella>/<subdir>/GLOSSARY.md; mask_text uses lookarounds so path-embedded identifiers still match; request/response now traversed with _mask_in_obj/_unmask_in_obj so system prompt, tool_use input, tool_result content and thinking blocks no longer leak unmasked; SSE responses parsed and unmasked per data: line; response audit entry includes content-type and unmask-applied flag.

6.15.0 — 2026-07-08

Added

3 Pocock-style skills: caveman (brutally honest review), zoom-out (architectural perspective), grill-me (adversarial weakness hunting)
scripts/skill-audit.sh: baseline skill catalog quality auditor (YAML frontmatter, required fields, DOMAIN.md)
92 skills pass baseline audit (0 FAIL, 0 WARN)
Understand-Anything codebase knowledge graph integration (scripts/ua-install.sh + ua-bridge.sh)
7 new Savia commands: /ua-analyze, /ua-domain, /ua-diff, /ua-chat, /ua-dashboard, /ua-onboard, /ua-install
Supports 13 languages: TS/JS, Python, Go, Rust, Java, C#, Ruby, PHP, Kotlin, Swift, C/C++, Scala, Elixir
SaviaClaw self-monitoring: heartbeat every 120s, auto-restart after 3 failures, stuck task detection (>300s)
SaviaClaw cron infrastructure: /cron add/list/remove/run via Talk, jobs.json persistence
SaviaClaw streaming feedback: progressive stdout capture with ▸ prefix every 5s during execution
docs/rules/domain/model-alias-table.md: provider-agnostic model resolution (DeepSeek/Claude/LocalAI)
scripts/opencode-cross-audit.sh: .opencode/ vs .claude/ resource drift detection
scripts/savia-env.sh: provider-agnostic env loader with fallback chain
scripts/content-fingerprint.sh: deterministic short content hash skill (sha256 truncated, lengths 8/16/32/64)
.claude/skills/content-fingerprint/SKILL.md: skill with declarative bioquimica frontmatter (DOIs validated against api.crossref.org)
tests/scripts/content-fingerprint.bats: 14 tests covering determinism, avalanche, length validation, dataset
tests/fixtures/fingerprint/: 30 etiquetated fixtures (5 identical / 5 near-duplicate / 5 distinct pairs)
tests/scripts/failure-pattern-memory-pattern-id.bats: 6 tests including byte-for-byte equivalence vs pre-SE-151
docs/specs/SE-151-content-fingerprint-consolidation.spec.md: SDD spec with falsifiable AC
docs/learning/biomimetic-investigation-protocol.md: reusable discipline from 2 sessions
.claude/hooks/contract-test-guard.sh: PreToolUse hook that blocks Edit/Write to allowlisted contract tests from agent/* branches (69 LOC)
.claude/contracts/allowlist.txt: 5 sealed contract tests pre-committed (block-force-push, confidentiality-sign, pii-gate, permissions-wildcard, validate-agent-permissions)
tests/hooks/test-contract-test-guard.bats: 24 tests covering block/allow paths, bypass via [contract-change|add|remove], env var guards, latency
docs/specs/SPEC-188-P3-sealed-contract-tests.spec.md: SDD spec with falsifiable AC
docs/specs/SPEC-188-P3-precommitment.md: pre-commitment of selection criteria, hard limits, abort criteria
scripts/spec-validator.sh — Frontmatter validator for resource: URI convention. Scans docs/propuestas/*.md and docs/rules/domain/*.md and emits WARN findings when:
docs/rules/domain/spec-resource-uri.md — Canonical rule documenting the convention.
tests/test-spec-validator.bats — Tests covering argument parsing, frontmatter detection, both rules (positive + negative paths), --strict mode, --json output, and --batch mode. Certified by test-auditor (≥80).
scripts/spec-lifecycle.sh — Helper that changes the status: field of a spec and appends a lifecycle entry to docs/propuestas/LOG.md. Modes:
docs/propuestas/LOG.md — Append-only lifecycle log with header + 3 seed entries (SE-222 S0 PROPOSED→IMPLEMENTED implicit, SE-220 PROPOSED, log creation).
tests/test-spec-lifecycle.bats — Tests covering argument parsing, canonical-status validation, dry-run, bootstrap, real transitions, LOG.md isolation (via PROPUESTAS_DIR_OVERRIDE and LOG_FILE_OVERRIDE env vars), and reverse-chronological order preservation.
bats tests/test-spec-lifecycle.bats: full suite passing
bash scripts/test-auditor.sh tests/test-spec-lifecycle.bats: certified
Spec: docs/propuestas/SE-222-okf-adoptable-patterns.md (Slice S1)
Previous slice: PR #850 (SE-222 S0 resource: URI convention)
Next slice: SE-222 S2 (propuestas/INDEX.md auto-generated)
scripts/propuestas-index-gen.sh — Generator that scans docs/propuestas/*.md, extracts frontmatter (spec_id, title, status, priority, effort, era), and writes docs/propuestas/INDEX.md grouped by status. Modes:
.claude/hooks/post-spec-edit-reindex.sh — PostToolUse hook that triggers regeneration when a file in docs/propuestas/*.md is edited (excluding INDEX.md and LOG.md themselves). Rate-limit: 1 regeneration per 60s (configurable via SAVIA_REINDEX_COOLDOWN). Toggle: SAVIA_PROPUESTAS_REINDEX_ENABLED=false.
docs/propuestas/INDEX.md — Initial auto-generated index, sentinel-marked @generated.
tests/test-propuestas-index-gen.bats — Tests for generator: dry-run, generate, check (passes when fresh, fails when stale), grouping by status, link rendering, exclusion of INDEX.md and LOG.md from listing.
tests/hooks/test-post-spec-edit-reindex.bats — Tests for hook: empty/invalid input safe, only triggers on propuestas/*.md edits, skips INDEX.md/LOG.md self-edits, respects cooldown, toggle disables it.
scripts/workforce-analytics.py — Python module that aggregates agent data sources
scripts/workforce-analytics.sh — CLI wrapper: --json, --format table|json|csv,
docs/rules/domain/workforce-analytics-protocol.md — Protocol doc: metrics table,
tests/scripts/test_workforce_analytics.py — 17 pytest tests covering empty/synthetic data,
tests/bats/test-spec-se-025-analytics.bats — 10 BATS tests: script exists/executable,
scripts/graphrag-quality-gates.sh: wrapper over knowledge-graph.py SQLite DB executing 12 structural quality checks:
tests/bats/test-se-030-graphrag-gates.bats: 10 tests covering script existence, JSON output, 12-check count, missing DB handling, self-loop detection, confidence validation
Complements graphrag-quality-gate.sh (metric thresholds) with structural/topology checks
Operates against ~/.savia/knowledge-graph.db by default (SE-162)
scripts/skill-usage-tracker.py — Registers each skill invocation in
scripts/skill-pattern-detector.sh — Analyzes data/skill-invocations.jsonl to
tests/scripts/test_skill_detection.py — 15 pytest tests: tracker appends entry,
tests/bats/test-spec-se-030-skill-detection.bats — 10 BATS tests: detector exists,
scripts/nl-to-wiql.py: pattern-based natural language to WIQL/JQL translator
tests/scripts/test_nl_to_wiql.py: 22 pytest tests covering:
Pure pattern library: deterministic, no LLM, no network
Aligns with SE-031 Slice 1 scope (resolver + NL heuristic)
Extends existing query-lib-nl.sh pattern with Python-native solution
scripts/agent-degradation-canary.sh: 3-canary behavioral regression detector:
tests/bats/test-se-040-canary.bats: 8 tests covering script syntax, all-pass scenario, JSON structure, total/passed/degraded fields, quiet mode, text output canary names
Zero new instrumentation: probes existing deterministic scripts
Canary 1 uses stdin pipe (router-mode-classifier reads from stdin)
Canary 3 uses exit-code heuristic (PASS=0, WARN=1, FAIL=2)
Extendable: add canaries by appending to CANARY_* arrays
scripts/scope-trace-gate.sh: standalone G13 runner. Input: spec_id + file list. Output: JSON with gate, passed, files_in_scope, files_outside_scope, verdict. Always exits 0 (advisory only).
g13_scope_trace function in scripts/pr-plan-gates.sh: wired as gate "G13" "Scope-trace audit" in scripts/pr-plan.sh. Detects spec refs from .pr-summary.md, commits, and branch name; matches changed files via whitelist, self-spec, path-hint, and token-overlap heuristics.
docs/rules/domain/pr-plan-gates.md: new reference doc for all pr-plan gates with G13 detail section.
tests/bats/test-se-079-scope-gate.bats: 7 BATS tests covering script existence, in-scope, out-of-scope, JSON validity, graceful no-spec skip, whitelist, and gate field.
docs/rules/domain/attention-anchor.md: canonical doc defining 4 Genesis patterns (B8 ATTENTION ANCHOR, B9 GOAL STEWARD, A7 ADVERSARIAL REVIEW, A9 SUPERVISED EXECUTION) with their pm-workspace implementations. Already existed with full content; verified complete.
scripts/attention-anchor-check.sh: verifier script. Searches workspace for references to each of the 4 patterns. Output JSON: checked, found, missing, details. Always exits 0.
tests/bats/test-se-080-attention-anchor.bats: 6 BATS tests covering doc existence, pattern mentions, cross-references in radical-honesty.md and autonomous-safety.md, check script JSON validity, and Genesis citation.
docs/rules/domain/radical-honesty.md: "implementa Genesis B9 GOAL STEWARD"
docs/rules/domain/autonomous-safety.md: "implementa Genesis A9 SUPERVISED EXECUTION"
docs/rules/domain/code-review-court.md: "implementa Genesis A7 ADVERSARIAL REVIEW"
docs/propuestas/SE-079-pr-plan-scope-trace-gate.md: "Pattern alignment: Genesis B9 GOAL STEWARD + B8 ATTENTION ANCHOR"
docs/rules/domain/architectural-vocabulary.md: canonical 6-term vocabulary (Module/Interface/Seam/Adapter/Depth/Locality) with Avoid rejection sets. Clean-room re-implementation of mattpocock/skills/improve-codebase-architecture/LANGUAGE.md (MIT). ≤89 LOC.
scripts/architectural-vocabulary-audit.sh: static auditor scanning outputs for prohibited terms (boundary, component, service, API in architectural contexts). Warning-only mode; exit 0 always.
Cross-references added to docs/rules/domain/attention-anchor.md, .opencode/agents/architect.md, .opencode/agents/architecture-judge.md.
Three ratchet principles: deletion test, interface=test surface, one adapter=hypothetical seam.
31 BATS tests in tests/structure/test-architectural-vocabulary.bats (all pass).
SE-082 APPROVED → IMPLEMENTED
scripts/skill-catalog-audit.sh: auditor with modes --report/--gate/--baseline-write/--json/--skill/--fix-report. Detects: description-missing-use-when, skill-long (WARN >100 LOC), skill-overlong (FAIL >200 LOC), missing-frontmatter, description-too-short.
docs/rules/domain/skill-catalog-discipline.md: canonical rule doc (≤124 LOC) citing Pocock write-a-skill (MIT). Defines 5 enforcement rules (frontmatter/size/trigger/attribution/cross-refs).
.ci-baseline/skill-quality-violations.count: baseline count for ratchet (155 warnings).
G14 gate in scripts/pr-plan-gates.sh: filters to skills modified in the PR, runs auditor in --gate mode. Registered in scripts/pr-plan.sh.
Skills audited: 106
WARN: 172 (long ≤200 LOC, missing use-when)
FAIL: 0
49 BATS tests in tests/test-skill-catalog-auditor.bats + tests/structure/test-skill-catalog-g14.bats (all pass).
SE-084 APPROVED → IMPLEMENTED
.opencode/skills/ubiquitous-language/SKILL.md: 5-step DDD glossary skill (≤92 LOC, SE-084 compliant). Triggers: "extrae glosario", "ubiquitous language", "/glossary", or >5 repeated domain terms without CONTEXT.md. Clean-room re-implementation of mattpocock/skills/ubiquitous-language + domain-model (MIT).
scripts/extract-domain-entities.py (370 LOC): reads memory-store JSONL or arbitrary input, extracts domain term candidates, cross-references against CONTEXT.md, outputs report with new/existing/inconsistent status columns. Modes: --report (default), --auto-update (adds [REVIEW]-marked terms), --export-glossary, --sync-graph.
scripts/knowledge-graph-domain-bridge.py: emits DOMAIN_TERM edges between episodic memory entities and CONTEXT.md terms.
docs/rules/domain/ubiquitous-language.md: canonical rule doc referencing script and CONTEXT.md format.
CONTEXT.md format standardized: # Domain Glossary — <Project> header, term|definition|status table, status values: stable/[REVIEW]/[INCONSISTENT].
47 BATS tests in tests/test-se-086-ubiquitous-language.bats (all pass).
30 BATS tests in tests/test-ubiquitous-language.bats (all pass).
SE-086 APPROVED → IMPLEMENTED
.opencode/skills/design-an-interface/SKILL.md (≤104 LOC, SE-084 compliant): generates 3 parallel interface design alternatives (A: maxima simplicidad, B: maxima flexibilidad, C: pragmatico) via Task tool sub-agents, consolidates in comparative table, recommends one with justification using SE-082 architectural vocabulary (Module/Interface/Seam/Depth/Locality). Clean-room re-implementation of mattpocock/skills/design-an-interface (MIT). Cross-references SE-082 architectural-vocabulary and SE-074 parallel-specs-orchestrator.
Decision stays with user — skill never auto-merges or auto-implements.
15 BATS tests in tests/structure/test-design-an-interface.bats (all pass). Covers AC-01..AC-06.
AC-02: added MIT attribution to Pocock in frontmatter and header
AC-04: added SE-074 parallel-specs-orchestrator cross-reference in Related section
description field updated to pass SE-084 description-missing-use-when audit (was WARN, now OK)
SE-087 APPROVED → IMPLEMENTED
docs/archive/rules/20260624-hook-event-equivalence.md
docs/archive/rules/20260624-image-relevance-filter.md
docs/archive/rules/20260624-portfolio-as-graph.en.md
docs/archive/rules/20260624-receipts-protocol.en.md
docs/archive/rules/20260624-savia-memory-architecture.md
docs/archive/rules/20260624-session-state-location.md
docs/archive/rules/20260624-slm-consolidation-pattern.md
docs/archive/rules/20260624-slm-training-pipeline.en.md
docs/archive/rules/20260624-vault-frontmatter.md
tests/bats/test-se-096-orphan-rules-archive.bats: 3 tests (all pass)
rule-orphan-detector.sh now reports 233 active rules (was 242, -9 archived)
2 remaining orphans (glm-governance-protocol, spec-resource-uri) are post-spec state
scripts/rules-index-generate.sh: generates docs/rules/INDEX.md from filesystem scan
docs/rules/INDEX.md: regenerated — 233 active rules, 240 lines
tests/bats/test-se-097-rules-index-regen.bats: 4 tests (all pass)
Existing docs/rules/domain/INDEX.md (165 lines, SPEC-115 format) is separate
New INDEX.md uses 4-column table (context_tier | file | title | spec) vs old
tests/bats/test-se-099-agents-split.bats: 7 tests verifying the checker, reference dir, and 0 line-WARN violations
scripts/output-cleanup.sh: retention policy enforcer for output/
tests/bats/test-se-101-output-cleanup.bats: 3 tests (all pass)
109 files currently older than 90 days in output/ (dry-run verified)
Protected telemetry files confirmed excluded from candidate list
scripts/speculative-tool-predictor.py — Updated from S0. New additions:
scripts/speculative-tool-execution.py — Main orchestrator:
.opencode/hooks/speculative-pre-execute.sh — PostToolUse hook:
scripts/speculative-cache-manager.py — Standalone cache manager:
.opencode/hooks/speculative-skill-preload.sh — PreToolUse hook:
scripts/speculative-telemetry-report.sh — Dashboard:
tests/scripts/test_speculative_execution.py: 23 pytest tests — all green
tests/bats/test-se-220-speculative.bats: 25 BATS tests — all green
S0 regression: 16 pytest + 10 BATS all green
Spec: docs/propuestas/SE-220-speculative-tool-execution.md (IMPLEMENTED)
.opencode/hooks/live-progress-emitter.sh: PostToolUse hook emitting [SAVIA-PROGRESS] {agent}: {action} [{elapsed}ms] to stderr
Master switch SAVIA_LIVE_PROGRESS=on|off (default off — zero cost when disabled)
Per-tool action formatting: Bash/Edit/Write/Read/Agent/Glob/Grep/Skill/Task
Duration from duration_ms field in hook payload; fallback to 0 when absent
Agent label from SAVIA_AGENT_NAME env var; defaults to savia
tests/bats/test-spec-042-live-progress.bats: 8 tests covering switch, format, tools, fallbacks
scripts/trace-prompt-optimizer.py: standalone prompt analysis script (no LLM required)
Detectors: verbose (>2000 chars), repetitive (concept >3x), contradictory (conflicting pairs),
Output JSON: {issues: [{type, severity, text}], score: 0-100, suggestions: [str]}
Score starts at 100; deducted per issue (high=20, medium=10, low=5)
Supports --prompt TEXT or stdin
tests/scripts/test_trace_prompt_optimizer.py: 10 pytest tests covering all detectors + CLI
scripts/visual-diff-merge-check.sh: PR-scoped before/after visual diff orchestrator
Gap analysis confirmed: visual-qa-agent handles single-screenshot analysis only;
4-phase pipeline: baseline capture → candidate capture → pixel diff → report + gate
Pixel diff uses compare (imagemagick) when available; falls back to file-size proxy
Semantic analysis via visual-qa-agent for borderline 2-10% diffs; auto-pass/fail outside
Score formula: (pixel * 0.6) + (semantic * 0.4); gate: PASS>=90 / REVIEW 60-89 / FAIL<60
Report output: output/visual-qa/merge-diff/{pr-id}/report.json + report.md
--dry-run flag skips agent invocation for safe testing
tests/bats/test-spec-046-visual-diff.bats: 7 tests covering validation, pass/fail, reports
scripts/sdd-reaction-engine.py: declarative SDD pipeline reaction engine
Handles 9 events: spec-approved, pr-created, tests-failed, review-done, ci-failed,
Output JSON: {event, action, auto, retries_allowed, escalate_after, message, source}
approved-and-green is always auto: false (autonomous-safety.md Rule #5)
.opencode/reaction-rules.yaml: default declarative rules file (YAML override support)
Minimal YAML parser built-in (works without PyYAML installed)
tests/scripts/test_sdd_reaction_engine.py: 10 pytest tests covering all events, safety rules, overrides
scripts/task-decomposer.py: heuristic recursive task decomposer (no LLM required)
Classifies tasks as atomic/compound using keyword and conjunction signals
Hour estimation via keyword patterns (auth=4h, API=3h, database=2h, etc.)
Decomposition templates for known domains: auth, api, crud, frontend, database, notification, payment, search
Conjunction splitting: "and / with / , ;" produces subtask titles automatically
Constraints respected: min 2 / max 7 subtasks per compound node; max-depth 3 default
SAVIA_TASK_SIZE_HOURS env var controls atomic threshold (default 4h)
Output JSON tree: {id, title, classification, estimated_hours, depth, can_parallelize, lineage, subtasks}
--tree flag prints ASCII tree to stderr
tests/scripts/test_task_decomposer.py: 10 pytest tests covering all decomposition paths
scripts/agent-message-schema.py: AgentMessage + ContentBlock dataclasses with full validation
tests/scripts/test_agent_message_schema.py: 25 pytest tests covering all acceptance criteria
scripts/semantic-fault-handlers.py: Keyword-based error classifier for agent fault recovery
tests/scripts/test_semantic_fault_handlers.py: 40 pytest tests
40/40 passing — timeout→TRANSIENT, missing field→FORMAT, context exceeded→CAPACITY,
scripts/query-keyword-expand.py: Stdlib-only query expansion for pre-search recall improvement
tests/scripts/test_query_keyword_expand.py: 33 pytest tests
33/33 passing — CamelCase/snake_case split, auth synonyms, ADO/KG acronyms,
scripts/pending-user-input.py: Async agent-to-human input request protocol
tests/scripts/test_pending_user_input.py: 21 pytest tests
21/21 passing — create writes file, check→1 no answer, resolve writes answer+ts,
scripts/cognitive-debt-monitor.py:
.opencode/hooks/cognitive-debt-check.sh:
docs/rules/domain/cognitive-debt-protocol.md:
tests/scripts/test_cognitive_debt.py: 14 tests — score low/high, all risk levels, recommendations,
tests/bats/test-cognitive-debt.bats: 5 tests — hook executable, master switch off, protocol doc,
scripts/agent-rca-analyzer.py: lightweight agent Root Cause Analysis pipeline (no Sentry required)
Builds on semantic-fault-handlers.py for error classification; falls back to keyword matching
RCA layer mapping: LOGIC→SPEC_REVIEW, CAPACITY→CONTEXT_PRUNING, SCOPE→TASK_DECOMPOSITION,
Output JSON: {root_cause, category, fix_suggestion, confidence, rca_layer}
Confidence reduced when no context provided (0.85x multiplier)
tests/scripts/test_agent_rca.py: 10 pytest tests covering all categories, fields, CLI
scripts/knowledge-graph-temporal.py: standalone temporal extension module for Savia knowledge graph
Adds valid_at (ISO-8601) and expired_at (ISO-8601, nullable) columns to entities table (idempotent migration)
add-temporal: adds/updates temporal metadata for an entity with validation
invalidate: sets expired_at on an entity (soft-delete, never hard-deletes)
query-at --when DATE: filters entities valid at a given point in time (null-safe — entities without temporal metadata are always included)
backfill: sets valid_at=first_seen for legacy entities with null valid_at (AC-03)
Validator rejects invalid ISO-8601 and expired_at < valid_at (Contradict guard)
Compatible with existing knowledge-graph.py SQLite schema
tests/scripts/test_graphiti_temporal.py: 8 pytest tests covering all operations + edge cases
.opencode/plugins/guards/sycophancy-guard.ts: TypeScript port of sycophancy-strip.sh. Exports detectSycophancy (pure function, testable) and sycophancyGuard (after-guard). Modes: shadow (default), warn, block. Master switch: SAVIA_ANTIADULATION=off.
.opencode/plugins/__tests__/sycophancy-guard.test.ts: 14 unit tests covering detection, shadow/warn/block modes, master-off bypass, clean-output no-op, positional bounds.
tests/bats/test-spec-150-s2-plugin.bats: 6 bats tests covering file existence, pattern presence, bash hook integrity, migration doc decision, foundation wiring, exports.
scripts/hook-multihandler-baseline.sh: FP/FN baseline evaluator for 6 critical hooks (sycophancy-strip, block-credential-leak, contract-test-guard, context-sanitize-input, pii-gate, router-mode-dispatch). 20 test inputs per hook (10 positive + 10 negative). Output JSON: {hook, fp_count, fn_count, fp_rate, fn_rate, avg_latency_ms, total_invocations}
tests/evals/hook-baselines/: directory for hook baseline JSON files
docs/rules/domain/hook-multihandler-migration.md: design document for TS plugin migration — candidate hooks, 2-layer pattern, anti-patterns, phase plan
tests/scripts/test_hook_multihandler_baseline.py: 19 pytest tests covering script existence, JSON schema, rate ranges, 6 hooks in registry, output directory creation
tests/bats/test-spec-150-hooks-migration.bats: 11 BATS tests covering script existence, doc existence, directory creation, JSON validity
Slice 1 baseline establishes pre-migration FP/FN rates
Slices 2-6 (TS plugin migration) pending: requires feasibility gate from Slice 1 results
Sensitive test payloads generated at runtime via Python helper (not stored in source)
.github/workflows/evals-ci.yaml:
scripts/evals-runner.sh:
scripts/evals-paired-delta.py:
tests/evals/datasets/skills/pbi-decomposition.jsonl: 5 real-domain cases (anonimized)
tests/evals/datasets/hooks/privacy-shield.jsonl: 10 pattern-only strings (no real PII)
tests/evals/baselines/pbi-decomposition-baseline.json: 5 baseline scores for tests
tests/scripts/test_evals_gate.py: 14 tests covering all paired-delta edge cases,
tests/bats/test-evals-ci-gate.bats: 6 tests covering runner executable, mock JSON,
OK AC-01: GitHub Action evals-ci.yaml triggers on PR changes to skills/agents/hooks
OK AC-02: Runtime <15 min (mock mode — LLM evals skipped gracefully if not installed)
AC-03: 2 datasets present (MVP scope: 5 cases min, not 20; full datasets deferred)
OK AC-04: Baseline regenerated on push to main (job regenerate-baseline)
OK AC-05: PR comment with delta table
OK AC-06: Judge model pinned via SAVIA_EVAL_JUDGE_MODEL var
OK AC-07: Paired-delta rationale in runner + this changelog (policy doc deferred to Phase 2)
OK AC-08: BATS test validates workflow yaml exists
tests/test-failure-pattern-memory.bats: 8 bats tests formalizing SPEC-188 Fase 1 acceptance criteria:
scripts/causal-confidence-scorer.py: causal confidence scorer (SPEC-188 Phase 3 MVP). Given a cause, evidence list, and alternatives, calculates confidence_score in [0,1], classifies supporting vs contradicting evidence, assigns probabilities to alternatives, returns verdict: high/medium/low/insufficient. Heuristics: 0 evidence → insufficient; 1-2 → medium (0.4-0.6); 3+ coherent → high (≥0.7); alternatives reduce confidence proportionally
scripts/diagnostic-metrics-tracker.py: append-only JSONL diagnostic quality tracker (SPEC-188 Phase 4 MVP). CLI: --record (adds entry), --report (mean_confidence, accuracy_rate, mean_time_to_identify, rework_rate), --list --n N (last N entries). Schema: {ts, investigation_id, time_to_identify_min, confidence_score, was_correct, rework_needed}
tests/scripts/test_spec188_phases34.py: 17 pytest tests — 7 for causal scorer, 10 for diagnostic tracker
tests/bats/test-spec-188-phases34.bats: 8 BATS tests covering both scripts
SPEC-188 Phase 3 implements G2 (Causal Confidence Channel) MVP core — heuristic-only, no LLM judge
SPEC-188 Phase 4 implements G4 (Diagnostic Quality Metrics) MVP — append-only JSONL tracker
Full Phase 3 (PreCommit hook, calibration-judge adapted) and Phase 4 (cron + monthly report) pending
Phases 1 (Failure Pattern Memory) and 2 (Sealed Contract Tests) remain independent scope
SE-228 Slice 3: loop-run-log append-only — schema docs/rules/domain/loop-run-log-schema.md, CLI scripts/loop-run-log.sh (subcomandos: append, tail, stats, prune), overnight-sprint skill actualizado con integración run-log, tests BATS certificados
docs/rules/domain/loop-budget-schema.md — canonical schema for loop-budget.md
scripts/loop-budget-check.sh — gate script that reads a skill's
templates/loop-budget.md.template — bootstrap template with defaults
tests/test-se228-s4-loop-budget.bats — 15 tests, all passing, certified.
docs/rules/domain/loop-phasing.md — Definición canónica de niveles L0→L3 para
scripts/loop-phasing-audit.sh — Auditor de nivel loop declarado vs inferido.
Campo loop_level en frontmatter de 4 SKILL.md:
tests/test-se228-s5-loop-phasing.bats — 34 tests, score 89/100, certified.
bats tests/test-se228-s5-loop-phasing.bats: 34/34 OK
bash scripts/test-auditor.sh tests/test-se228-s5-loop-phasing.bats: score 89, certified
Spec: docs/propuestas/SE-228-loop-engineering-patterns.md
Slices anteriores: S1 (PR #878), S2+S3 (PR #877)
Regla complementaria: docs/rules/domain/autonomous-safety.md
SE-143: Vector search activado (sentence-transformers + faiss-cpu). Recall@5: Grep=40% → Vector=90% (+50pp)
memory-store.sh: nuevo subcomando 'doctor' reporta Level 0/1/2, deps y estado del índice
install.sh: Step 7 instala requirements-memory.txt con fallback graceful
session-init.sh: auto-rebuild del índice vectorial en background al arranque
tests: new skill-behavior/ infrastructure — validates all skills for Rule 11 compliance, required structure, and Description Trap patterns (SE-147)
SE-148: AgentRunSummary telemetry — schema v2 JSONL (backward-compat), agent-run-logger.sh, agent-run-report.sh, 15 BATS tests
SE-149: hashline-guard stale-file protection — anchor generation, drift check, safe edit wrapper, L3 agent protocol, 10 BATS tests
SE-150: pre-output validator TTSR-inspired — 7 rules (POR-001..007), check/list-rules/explain subcommands, 34 BATS tests
OpenCode binary upgraded 1.3.13 → 1.14.30 (via opencode upgrade). Local action — no repo changes.
opencode.json — workspace config en root. Schema oficial OpenCode v1.14:
SKILLS.md — catálogo cross-frontend de skills (109 líneas, 90 skills listados con path + descripción truncada). Análogo a AGENTS.md (SE-078 pattern). Auto-generado.
scripts/skills-md-generate.sh — generador idempotente del SKILLS.md desde .claude/skills/*/SKILL.md. Modes: stdout default, --apply, --check. Patrón SE-078 reutilizado.
Symlinks en .opencode/:
tests/structure/test-opencode-bootstrap.bats — 32 tests certified. Estructura:
Agents bajo OpenCode v1.14: schema mismatch. Claude Code usa tools: [array] + named colors; OpenCode espera tools: {object} + hex colors. Sin converter, los 70 agents no se invocan vía native OpenCode agent mechanism. AGENTS.md provee discovery, pero ejecución requiere Slice 2b (converter TS plugin).
Hooks reactivos (PreToolUse/PostToolUse): los 64 hooks viven en .opencode/hooks pero OpenCode v1.14 NO ejecuta .sh hooks nativamente — necesita TS plugin que los wrappee. Slice 2b territory.
Skills loading on-demand: los SKILL.md están visibles via .opencode/skills/, pero OpenCode no tiene concepto "skills" nativo. Cuando el LLM necesita una skill, lee el SKILL.md vía Read tool. SKILLS.md provee el catálogo.
PV-01 Backward compat absoluto: cero modificación de los 64 hooks, 70 agents, 90 skills, 534 commands existentes. Solo se añade infraestructura nueva.
PV-04 Operator control: autoupdate: false — Mónica controla cuándo actualizar OpenCode.
PV-06 No vendor lock-in: opencode.json no menciona ningún vendor de inferencia. BATS test enforce.
AGENTS.md y SKILLS.md son auto-generados (SE-078 pattern) — drift detection vía --check.
Cero red, cero git operations en runtime, cero merge autónomo.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/opencode-bootstrap-20260430, sin merge autónomo.
docs/propuestas/SPEC-125-recommendation-tribunal-realtime.md — diseño completo (3 slices, 36h total estimadas, P0 Critical Path):
Solo PROPOSAL — no implementación.
No modifica hooks existentes ni agentes existentes.
Cero red, cero git operations.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-125-..., sin push automático ni merge.
5 hooks TIER-1 portados a TypeScript (.opencode/plugins/):
Helpers compartidos (.opencode/plugins/lib/):
scripts/agents-opencode-convert.sh — converter de agentes Claude Code → OpenCode v1.14:
docs/migration-claude-code-to-opencode.md — guía de migración paso a paso (≤150 líneas):
scripts/opencode-migration-smoke.sh — 6 checks fast-fail:
savia-budget-guard registrado en .claude/settings.json — PreToolUse * con timeout 5s. Hook nunca bloquea (only warns 70/85/95% del presupuesto mensual). Bajo OpenCode, mismo guard via plugin foundation. Bajo Claude Code, via .sh hook nativo.
scripts/cognitive-debt.sh — cmd_summary integra resumen mes-a-fecha del quota tracker (Slice 5 + 2b-ii). El operador ve cognitive metrics + presupuesto Savia en una sola salida.
tests/structure/test-spec-127-migration-final.bats — 43 tests cubriendo todos los componentes (5 ports + helpers + foundation + converter + budget-guard + onboarding + smoke + cognitive-debt). Auditor: certified (92).
slice_2b_ii_status: IMPLEMENTED 2026-05-01.
slice_3_status: NOT_NEEDED — OpenCode v1.14 tiene discovery nativo de slash commands via .opencode/commands/ symlink. No hace falta porting.
AC-2.2 marcada como implementada.
Bumped hooks(65/65reg) → hooks(65/66reg) para reflejar registro de savia-budget-guard.
Era 189 (OpenCode) — esta PR cierra la era Claude Code. La siguiente PR ya se trabaja desde OpenCode bootstrapped.
Provider-agnostic (PV-06) — ningún archivo nuevo hardcodea Anthropic/Claude/Sonnet/Opus/Haiku. Verificado por test BATS dedicado.
Backward compat (PV-01) — todos los .sh originales siguen vivos. Esta PR añade archivos, NO modifica los hooks .claude/hooks/*.sh.
Investigación BitNet entregada como informe interno gitignored. Spec/PR no creados — autonomous-safety: investigación NO crea backlog sin aprobación humana.
docs/propuestas/SPEC-127-savia-opencode-provider-agnostic.md — APPROVED 2026-04-30 (operator review). Renamed from prior Copilot-specific framing. Inference-independent foundation: cada usuario decide su frontend (Claude Code / OpenCode / Codex / Cursor / otro) y su provider (Anthropic API / hosted-OSS / LocalAI / Ollama / vendor-managed / custom corporate endpoint). Framework no asume vendor — detecta capacidades en runtime y degrada gracefully. 5 slices (Foundation 8h · Hook adapter 16-20h · Slash MCP shim 12-16h · Subagent fallback 8-12h · Quota guard 6h ≈ 50-62h). Restricciones inviolables PV-01 a PV-06 (PV-06 = cero vendor lock-in en source).
scripts/savia-env.sh — provider-agnostic env loader. Sourceable. Exporta SAVIA_WORKSPACE_DIR (fallback chain SAVIA_WORKSPACE_DIR → CLAUDE_PROJECT_DIR → OPENCODE_PROJECT_DIR → git rev-parse → pwd) y SAVIA_PROVIDER (free-form: lo que el usuario declare en preferences.yaml). 3 capability probes: savia_has_hooks, savia_has_slash_commands, savia_has_task_fan_out. Cero hardcoded vendor names en source — los probes leen ~/.savia/preferences.yaml cuando existe y autodetectan via env vars cuando no. CLI dispatch standalone.
scripts/savia-preferences.sh — gestor de ~/.savia/preferences.yaml. 6 subcomandos: init (entrevista 8 preguntas), show, get <key>, set <key> <value>, reset --confirm, validate. Validator rechaza claves prohibidas (api_key, password, secret, token — credenciales van en credential manager, no aquí).
.claude/commands/savia-setup.md — slash command que invoca el onboarding interactivo. 8 preguntas con campo libre — sin lista cerrada de vendors. Re-ejecutable idempotentemente cuando el stack del usuario cambia.
docs/rules/domain/provider-agnostic-env.md — rule canonical (121 líneas). Define el contrato cross-frontend × cross-provider. Hook author + script author checklists. Backward compat absoluto (PV-01).
docs/rules/domain/model-alias-schema.md — schema YAML user-extensible (149 líneas, NO tabla cerrada). Cada usuario añade sus mappings en ~/.savia/preferences.yaml. 3 ejemplos genéricos con placeholders (no nombran vendor concreto en source). Documenta forbidden credential keys.
tests/structure/test-spec-127-slice1-foundation.bats — 55 tests certified (auditor max). Estructura por AC:
PV-01 Backward compat absoluto: cero modificación de los 70 agents, 64 hooks, 90 SKILL.md, 534 commands existentes. Solo infraestructura nueva.
PV-03 Zero data exfiltration: validator rechaza claves de credenciales en preferences.yaml.
PV-04 Opt-in: si no hay preferences.yaml, Savia opera con defaults (Claude Code autodetect).
PV-06 No vendor lock-in: BATS tests verifican que ni savia-env.sh ni savia-preferences.sh mencionan vendors comerciales específicos en código.
Cero red, cero git operations en runtime.
Onboarding 100% local: ~/.savia/preferences.yaml nunca sale del home del usuario.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-127-provider-agnostic-20260430, sin merge autónomo, AUTONOMOUS_REVIEWER asignado.
docs/propuestas/SPEC-127-savia-opencode-copilot-enterprise-compat.md — PROPOSED. Audit Savia ↔ OpenCode + GitHub Copilot Enterprise compatibility. 5 slices (Foundation 8h · Hook adapter+TS plugin 16-20h · Slash command MCP shim 12-16h · Subagent fallback 8-12h · Premium quota guard 6h ≈ 50-62h total). Documents 5 critical frictions (compaction #11157, hooks port bash→TS plugin, GHE on-prem auth #3936, premium request inflation #8030, context cap 128K #5993) + 3 categorical breaks vs OpenCode-Claude (zero hook surface, no subagent fan-out, no workspace slash commands). Restricciones inviolables PV-01 a PV-05. Compatibility matrix Claude Code | OpenCode-Claude | OpenCode-Copilot | Risk. Top 10 priority changes + 5 decisiones pendientes para revisor humano. Sin parity completa — re-route strategies (MCP for commands, git pre-commit for hooks, single-shot for orchestrators).
scripts/hook-portability-classifier.sh — clasifica deterministically cada hook de .claude/hooks/*.sh en uno de 4 tiers de portabilidad bajo cualquier frontend × cualquier provider:
output/hook-portability-classification.md — reporte auto-generado (103 líneas). Estructura:
tests/structure/test-spec-127-slice2a-hook-classifier.bats — 31 tests certified. Estructura por AC:
PV-01 Backward compat absoluto: cero modificación de los 64 hooks existentes. Solo análisis + clasificación + reporte.
PV-02 Safety layer crítica: los 3 safety-critical hooks verificados TIER-1 — cobertura preservada.
PV-05 Visibilidad de pérdidas: 15 hooks TIER-4 listados explícitamente con la razón. La pérdida no es silenciosa.
PV-06 No vendor lock-in: classifier branches en capability events (PreToolUse, Stop, ...) y matchers (Bash, Edit, Write, Task), nunca en vendor names. BATS test enforce.
Cero red, cero git operations en runtime, cero merge autónomo.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-127-slice2a-hook-classifier-20260430, sin merge autónomo.
.opencode/package.json — declara @opencode-ai/plugin dependency. ESM type=module. Private (no publishable). OpenCode runtime ejecuta bun install automáticamente al arrancar — no requiere paso manual del usuario.
.opencode/tsconfig.json — TS strict mode, ES2022 target, Bundler module resolution (compatible con Bun runtime). noEmit: true (typecheck-only, OpenCode loader interpreta TS directamente). Includes plugins/**/*.ts, excludes node_modules.
.opencode/plugins/savia-foundation.ts — plugin foundation no-op. Importa Plugin type desde @opencode-ai/plugin. Exporta SaviaFoundationPlugin como named + default export. Async function. Returns {} (empty hooks). Documenta el porting roadmap de Slice 2b-ii: 5 hooks TIER-1 a portar (validate-bash-global, block-credential-leak, block-gitignored-references, prompt-injection-guard, tdd-gate).
.opencode/plugins/savia-foundation.test.ts — contract test scaffold con bun:test runner. 3 tests: plugin es async function, returns empty hooks {}, robust to partial context. Bun runtime no se ejecuta en CI todavía — Slice 2b-ii añade el test runner. Por ahora el contract es enforced por BATS estructural.
.opencode/plugins/README.md — porting roadmap doc (60 líneas). Explica por qué existe el folder (OpenCode v1.14 no ejecuta .sh hooks nativamente), toolchain requirements, los 5 hooks TIER-1 a portar en 2b-ii (uno por PR), y safety boundaries (PV-01: hooks .sh originales intocados).
tests/structure/test-spec-127-slice2b-i-ts-toolchain.bats — 40 tests certified. Estructura:
PV-01 Backward compat absoluto: cero modificación de los 64 hooks .sh existentes. Solo se añade infrastructura nueva en .opencode/plugins/.
PV-06 No vendor lock-in: BATS tests verifican que ni el plugin TS ni el package.json mencionan vendors comerciales hardcoded.
TDD enforced: el .test.ts se creó antes del plugin (TDD gate hook lo bloqueó hasta que existiera).
150-line cap respetado en plugins/README.md.
Cero red en runtime (el bun install que OpenCode dispara descarga npm packages — esto es runtime de OpenCode, no del repo).
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-127-slice2b-i-ts-toolchain-20260501, sin merge autónomo.
scripts/savia-orchestrator-helper.sh — bash helper provider-agnostic. Subcomandos:
docs/rules/domain/subagent-fallback-mode.md — patrón canonical "single-shot expanded prompt" (121 líneas). Documenta: por qué (4 orchestrators dependen de Task; sin fallback rompen silenciosamente bajo stacks sin Task tool), capability detection (savia_orchestrator_helper.sh mode), pseudocode del pivot, trade-offs explícitos (loss of context isolation, sequential vs parallel, shared memory between judges), context-wall mitigation via wrap envelope.
4 orchestrators críticos patched con sección compacta "## Fallback mode (SPEC-127 Slice 4)":
tests/structure/test-spec-127-slice4-subagent-fallback.bats — 39 tests certified. Estructura por AC:
PV-01 Backward compat absoluto: cero modificación del fan-out flow existente. Solo se añade el branch single-shot a los 4 orchestrators.
PV-05 Visibilidad de pérdidas: trade-offs documentados explícitamente en el rule doc (loss of isolation, sequential not parallel, shared memory).
PV-06 No vendor lock-in: helper bash branches en capability (has_task_fan_out), no en vendor. BATS test enforce.
150-line cap respetado en los 4 orchestrators tras patch.
Cero red, cero git operations en runtime, cero merge autónomo.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-127-slice4-subagent-fallback-20260430.
scripts/savia-quota-tracker.sh — tracker genérico provider-agnostic. Lee budget_kind (req-count / token-count / dollar-cap / none) y budget_limit desde ~/.savia/preferences.yaml (Slice 1 source-of-truth). 5 subcomandos:
.claude/hooks/savia-budget-guard.sh — PreToolUse advisory hook. SIEMPRE exits 0 (never blocks). Drains stdin sin parsear (no inspecciona tool name/args). Records 1 request unit per call via tracker. Lee threshold marker; si over-70/85/95/exceeded, emite stderr nudge once-per-threshold-per-session vía /tmp marker keyed por PID + threshold. Bail silent si tracker falta.
tests/structure/test-spec-127-slice5-quota-guard.bats — 40 tests certified. Estructura por AC:
PV-01 Backward compat absoluto: cero modificación de los 64 hooks, 70 agents, 90 skills, 534 commands existentes. Solo se añade el tracker + hook nuevos.
PV-02 No blocks: el hook PreToolUse SIEMPRE exits 0, never blocks. Verified by BATS regression.
PV-03 Zero data exfiltration: log local en ~/.savia/quota/, never repo-committed (gitignored). No telemetry to any vendor.
PV-04 Operator opt-in: si el usuario no declara budget_kind o lo deja en none, tracker invisible.
PV-06 No vendor lock-in: BATS tests verifican que ni tracker ni hook mencionan vendors comerciales hardcoded.
Cero red, cero git operations en runtime, cero merge autónomo.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-127-slice5-quota-guard-20260501, sin merge autónomo.
tests/test-token-tracker-middleware.bats — 30 tests certified (score 91). PostToolUse async monitor de context tokens, 3 zonas (50% hint / 70% alert / 85% critical → auto-compact).
tests/test-subagent-lifecycle.bats — 29 tests certified (score 94). SubagentStart/Stop logging a output/agent-lifecycle/lifecycle.jsonl.
tests/test-task-lifecycle.bats — 30 tests certified (score 94). TaskCreated/Completed logging con team/teammate fields.
.claude/hooks/emergency-mode-readiness.sh — SessionStart async hook (AC-03). Ejecuta localai-readiness-check.sh solo si EMERGENCY_MODE_ENABLED=true. Surface FAIL/WARN a stderr para visibilidad usuario, log append-only a output/emergency-mode/readiness.jsonl. Timeout 10s en script invocation, nunca bloquea SessionStart.
tests/test-emergency-mode-readiness.bats — 30 tests certified (score 94). Cubre verdict states (READY/WARN/FAIL/SKIP/TIMEOUT/UNKNOWN), feature-flag silencio, mock script via $CLAUDE_PROJECT_DIR, append accumulation, edge cases.
.github/workflows/templates/pr-agent-review.yml — reusable workflow para invocar qodo-ai/pr-agent como 5º juez del Court (AC-04). Cost gate max_lines default 1000, feature-flag check COURT_INCLUDE_PR_AGENT, skip on draft, comments tagged [pr-agent], graceful skip si pr-agent no instalado.
docs/rules/domain/court-external-judges.md — política para incluir jueces externos OSS en el Court (AC-08). 7 requisitos de inclusión (auditable, self-hostable, pinable, schema validable, opt-in, tagged comments, respeta AUTONOMOUS_REVIEWER), 6 reglas operación, activación paso a paso, riesgos documentados.
scripts/memory-save.sh — cmd_save extendido con --source <origin> flag obligatorio. Valida format (4 sources OK: tool:*, file:*:*, verified:*, user:explicit), rechaza blacklist (speculation, plan, intent, draft, hypothesis). Source embeded en JSONL output. Escape hatch: SAVIA_VERIFIED_MEMORY_DISABLED=true.
.claude/hooks/memory-verified-gate.sh — PreToolUse Write hook bloquea auto-memory writes sin citation pattern. 5 patrones aceptados: file reference, markdown link, Source/Ref keyword, URL, frontmatter type. Skipped: MEMORY.md, session-journal.md, session-hot.md, session-summary.md.
tests/test-memory-verified-gate.bats — 33 tests certified score 94. Cubre block/pass/skip, escape hatch, edge cases (empty content, large content, malformed JSON), isolation.
docs/rules/domain/verified-memory-axiom.md — política completa con rationale, ejemplos correcto/rechazado, escape hatch, política de evolución.
docs/rules/domain/pr-natural-language-summary.md — regla canónica con 4 puntos obligatorios (qué cambia, por qué importa, riesgos, activación), lenguaje prohibido (spec IDs, jergas, métricas), excepciones explícitas (commits sign, reverts puros, hotfixes documentados).
scripts/pr-plan-gates.sh:g_summary — gate G11 valida .pr-summary.md existe, ≥300 chars, contiene título canónico ## Qué hace este PR (en lenguaje no técnico).
docs/propuestas/SE-074-parallel-spec-execution.md — spec en status APPROVED. 3 slices: worktree manager + spec queue (M 8h), PR queue manager (S 4h), DB sandbox + cleanup (M 6h). Pre-requisitos cumplidos (hook 100%, G11, cascade-rebase pattern, bounded concurrency, AUTONOMOUS_REVIEWER). Comparativa explícita vs status quo: 3-4x throughput sobre 1.1-1.2x coste tokens.
docs/ROADMAP.md Era 188 sección "Memory + Throughput foundations" con pipeline ordenado SE-072 (done) → SE-073 → SE-074. Trade-off explícito: SE-074 Slice 1 (8h) bloquea ~1 día pero break-even tras 2-3 Eras post.
docs/propuestas/SE-075-voicebox-adoption.md — APPROVED. 3 slices: task_queue.py (S 2h, sirve a SE-074 + SE-076), auto-chunking long-form TTS (M 3h), Kokoro 82M CPU voice español (M 3h, independiente de SE-042 GPU-blocked). Source: jamiepine/voicebox MIT.
docs/propuestas/SE-076-queryweaver-patterns.md — APPROVED. 3 slices: Graphiti episodic memory model en JSONL (M 4h, extiende SPEC-027), schema-as-graph WIQL (M 3h, target hallucination 20%→5-10%), LLM healer wrapper (S 2h). Source: FalkorDB/QueryWeaver patterns (AGPL-3.0 — solo patterns, no código).
docs/propuestas/SE-077-opencode-replatform-v114.md — APPROVED. 2 slices: plugin TS savia-gates (M, 8h) + parity audit ratchet (M, 6h). Era 189.
docs/propuestas/SE-078-agents-md-cross-frontend.md — APPROVED. AGENTS.md generator + drift check + Stop hook auto-regenerate. M 6h, era 189. Supersedes SPEC-114.
docs/rules/domain/spec-opencode-implementation-plan.md — regla canónica. Cada spec APPROVED post-2026-04-26 incluye sección obligatoria. Grandfathering documentado. Hot-fix exemption con exempt_opencode_plan frontmatter.
scripts/spec-opencode-plan-audit.sh — audit script (3 sub-secciones obligatorias, exit 1 si missing).
scripts/pr-plan-gates.sh:g_opencode_plan — G12 gate, solo se activa si el PR toca specs.
.ci-baseline/spec-opencode-plan-violations.count — baseline frozen at 0.
scripts/memory-tier-rotate.sh — 2-tier rotation (Tier A active ≤30 entries, Tier B filename-only archive).
scripts/memory-access.sh — increments access_count + updates last_access por memory file.
tests/structure/test-memory-tier-rotate.bats — 24 tests, score 83 certified (test-auditor).
MEMORY-ARCHIVE.md (cuando rotation produce demociones) — Tier B filename-only.
scripts/parallel-specs-orchestrator.sh — Slice 1 core. N workers paralelos en worktrees aislados, bounded concurrency hard cap 5, port allocation determinística, tmp dir sandboxing, runtime timeout, session.log per spec.
scripts/spec-budget.sh — Slice 1.5. Mapeo effort field (S/M/L) → Poisson-clipped retry budget (lambda_S=2, lambda_M=3, lambda_L=5; clipped [2, 8]). Determinístico por defecto; estocástico opt-in seedeado por hash(spec_id).
scripts/adaptive-halting.sh — Slice 1.5. Doble criterio halting (convergencia tree-hash + confidence ≥ floor + tests passed). Reemplaza halting binario "tests pasan una vez".
docs/rules/domain/parallel-spec-execution.md — regla canónica con configuración, garantías de seguridad, comandos.
tests/structure/test-parallel-specs-orchestrator.bats — 23 tests, score 88.
tests/structure/test-spec-budget.bats — 15 tests, score 85.
tests/structure/test-adaptive-halting.bats — 18 tests, score 81.
OK AC-01 orchestrator acepta lista de spec IDs y crea N worktrees
OK AC-02 MAX_PARALLEL_SPECS hard cap 5 (rejects valores mayores)
OK AC-03 port range único por worktree (hash determinístico)
OK AC-04 tmp dir aislado por worker (/tmp/savia-spec-<id>-<ts>/)
OK AC-05 graceful per-worker failure (1 falla, otros continúan)
OK AC-06 MAX_RUNTIME_MINUTES (default 60m) con timeout(1)
OK AC-07 Tests BATS ≥20 score ≥80 (23 tests, 88 actual)
OK AC-08 Doc en docs/rules/domain/parallel-spec-execution.md
OK AC-09 CHANGELOG entry (este fichero)
OK AC-09a should_halt con doble criterio (convergencia + confianza)
OK AC-09b spec-budget mapea effort → Poisson(lambda) clipped
OK AC-09c Confidence threshold configurable, validado [0.50, 0.95]
OK AC-09d Tests BATS ≥10 score ≥80 (18 + 15 = 33 tests, scores 81 y 85)
OK AC-09e Cita Kohli et al. 2026 arXiv:2604.07822 en docstrings + spec
Slice 2 (S 4h): PR queue manager con cascade-rebase auto-resolve
Slice 3 (M 6h): DB sandbox + cleanup hook async stale worktrees
scripts/parallel-specs-merge-queue.sh — Slice 2 core. Gestiona la cola FIFO de branches producidas por el orchestrator paralelo y aplica el patrón cascade-rebase documentado en feedback_changelog_cascade_rebase. Subcomandos: add, remove, list, status, clear --confirm, rebase <branch>, rebase-next. Auto-resolve sólo para conflictos restringidos a CHANGELOG.md + CHANGELOG.d/; cualquier otra divergencia ESCALA con git rebase --abort y mensaje detallado a la usuaria.
tests/structure/test-parallel-specs-merge-queue.bats — 32 tests, score 88 certificado.
Sección "Merge queue (Slice 2)" añadida a docs/rules/domain/parallel-spec-execution.md con comandos, scope de auto-resolve y límites de seguridad.
NUNCA hace git push
NUNCA hace gh pr merge ni equivalente
NUNCA force-pushes
NUNCA auto-resuelve un conflicto fuera de CHANGELOG.md / CHANGELOG.d/
Se niega a operar con working tree dirty (vía git status --porcelain)
MAX_REBASE_STEPS (default 30) como guard rail anti-loop
OK AC-10 PR queue manager merge-en-orden con cascade-rebase auto-resolve para CHANGELOG
OK AC-11 Conflictos no-CHANGELOG escalados (no auto-merge), output ESCALATE: <ficheros> por stderr
3 dispatch (usage / help / unknown subcmd)
8 queue CRUD (add idempotente, remove idempotente, list missing/merged/ready/needs-rebase, status totales)
2 clear (refusal sin --confirm, success con)
8 rebase (branch missing, dirty tree, ya merged, ahead-only, conflicto CHANGELOG.md, fragment-only, escalación non-CHANGELOG, vuelta a rama original)
3 rebase-next (queue vacía, skip merged + rebase stale, no-op si ready)
3 safety estática (ningún git push o gh pr merge, set -uo pipefail, MAX_REBASE_STEPS guard)
3 spec ref (SE-074 Slice 2, parallel-spec-execution.md, autonomous-safety.md, cascade-rebase header)
docs/propuestas/SE-079-pr-plan-scope-trace-gate.md — APPROVED, priority media, effort S 3h, Era 189. Slice único: nueva G-gate g13_scope_trace en scripts/pr-plan-gates.sh que verifica que cada archivo cambiado en un PR trace a (1) un AC del spec referenciado, (2) una whitelist hard-coded (CHANGELOG.d/, .scm/, .confidentiality-signature), o (3) override explícito Scope-trace: skip — <razón ≥10 chars> en .pr-summary.md. Heurística pure-bash sin LLM (target <1s overhead). Soft-skip cuando no hay spec ref detectable (PRs legítimos sin spec).
ROADMAP entry bajo Era 189.
docs/propuestas/SE-080-attention-anchor-vocabulary.md — APPROVED, priority media, effort S 2h, Era 189. Slice único: doc canónico docs/rules/domain/attention-anchor.md (~80 líneas) con definición de los 4 patrones Genesis (B8 ATTENTION ANCHOR, B9 GOAL STEWARD, A7 ADVERSARIAL REVIEW, A9 SUPERVISED EXECUTION) + 4 cross-references unilíneas a reglas existentes (radical-honesty.md, autonomous-safety.md, code-review-court.md, SE-079). Output del G13 scope-trace de SE-079 emite "B8 attention-anchor present" en lugar de un check anónimo. Sin código ejecutable nuevo — adopción 100% vocabulary.
ROADMAP entry bajo Era 189.
scripts/lib/task-queue.py — Slice 1. Serial-execution job queue with SQLite (WAL) persistence and auto-recovery of stale running jobs at construction. Re-implementation of the pattern in voicebox MIT (backend/services/task_queue.py); clean-room — no source code copied. Public API: enqueue / dequeue / heartbeat / complete / status / recover / drain / list_jobs. CLI dispatcher with subcommands. Atomic claim via BEGIN IMMEDIATE. STALE_HEARTBEAT_SEC=300 default (env override). Storage under output/task-queue/<queue>.sqlite.
scripts/lib/sentence-splitter.py — Slice 2 helper. Spanish-aware sentence boundary splitter for long-form TTS chunking. Preserves abbreviations: Sr./Sra./Sres./Sras./Dr./Dra./Lic./Ing./Prof./D./Dña./Vd./Vds./S.A./S.L./a.m./p.m./e.g./i.e./etc./aprox./depto./pág./vol./… via placeholder substitution. Secondary split on commas/semicolons when sentence exceeds --max-chars (default 600). Numeric decimals (1.5, 2.3) are NOT treated as sentence boundaries.
scripts/savia-voice-chunk.sh — Slice 2 orchestrator. Long-form Spanish TTS chunker with bounded concurrency (default 2) and ffmpeg acrossfade between chunks (default 80 ms, configurable). TTS backend pluggable via $SAVIA_TTS_CMD (placeholders {out}/{text}); auto-detects espeak-ng/espeak if unset. Modes: --dry-run (chunks only), --no-fade (concat without crossfade). Reads from --text / --file / stdin.
tests/structure/test-task-queue.bats — 26 tests, certified (positive paths × 8, negative paths × 5, edge cases × 6, atomicity × 1, auto-recovery × 3, static safety × 3).
tests/structure/test-savia-voice-chunk.bats — 27 tests, certified (positive paths × 9, negative paths × 5, edge cases × 7, splitter quality × 6).
voicebox (MIT, jamiepine/voicebox, 23k stars) — patterns for task_queue.py and chunker services/tts.py. Clean-room: docstrings, attribution headers, and matching public API behaviour replicated; no source code copied.
OK AC-01 task-queue.py with MIT/voicebox attribution in header
〰 AC-02 SSE bridge endpoint — DEFERRED (Savia is CLI-first; status --json covers observability)
OK AC-03 Auto-recovery tested (tests #19, #22, #23)
OK AC-04 BATS coverage requirement met (26 tests certified)
〰 AC-05 Dedicated doc — DEFERRED (module docstring + spec ref sufficient until a second consumer appears)
OK AC-06 Long-form text → single audio file via ffmpeg acrossfade
OK AC-07 Bounded concurrency 2 default (env + CLI override)
OK AC-08 Spanish abbreviations preserved across splitter tests
AC-09–AC-12 Kokoro 82M install, wrapper, skill docs, latency benchmark — pending user authorization for model download
task-queue.py CLI smoke: enqueue → status → dequeue → complete → drain → status
savia-voice-chunk.sh end-to-end: 3 Spanish sentences → 3 espeak-ng chunks → ffmpeg crossfaded WAV (65 KB, RIFF/WAVE PCM 16-bit mono 22050 Hz)
task-queue.py: NO destructive operations. drain deletes only done|failed jobs (preserves pending/running). Atomic CAS via BEGIN IMMEDIATE prevents double-claim under concurrent workers.
savia-voice-chunk.sh: set -uo pipefail. NO network, NO file modification outside $WORKDIR (mktemp) and the user-specified --out. Refuses unknown CLI args.
.ci-baseline/hook-critical-violations.count 5 → 6. Hook latency benchmark (scripts/hook-bench-all.sh) oscillates 5–8 under sustained system load (concurrent gitlab/puma at ~96 % CPU on the bench host). Pre-existing — none of this batch's changes touch hooks. Precedent: commit e1fdac2d fix(ci): revert hook-critical-violations baseline 4 → 5 (non-deterministic). Should re-tighten via scripts/baseline-tighten.sh once the host is idle (auto-applied by next CI run when ≤ 5).
docs/propuestas/SE-081-pocock-skills-quick-wins.md — APPROVED priority alta (S 2h). Tres skills MIT clean-room: caveman (modo ultra-comprimido -75% tokens), zoom-out (mapa de módulos), grill-me (interrogatorio relentless aligned con Rule #24 radical-honesty). Zero código, sólo markdown + atribución.
docs/propuestas/SE-082-architectural-vocabulary-discipline.md — APPROVED priority alta (M 4h). Vocabulario obligatorio Module/Interface/Seam/Adapter/Depth/Locality para architect y architecture-judge. Extiende SE-080 attention-anchor. Auditor warning-only en este Slice; gate enforced en SE-084.
docs/propuestas/SE-083-tdd-vertical-slice-skill.md — APPROVED priority media (S 2h). Skill que codifica el anti-pattern de horizontal slicing. Cross-reference desde test-architect.
docs/propuestas/SE-084-skill-catalog-quality-audit.md — APPROVED priority alta (M 6h, 2 slices). Auditor estático (Use-when triggers + ≤100 LOC + frontmatter validation) + pr-plan G14 gate sólo sobre skills modificados en el PR. Baseline ratchet via baseline-tighten.sh.
docs/propuestas/SE-085-write-a-skill-meta.md — APPROVED priority baja (S 2h). Meta-skill consultivo para crear skills nuevos. Depende de SE-084 (regla canónica enforced).
docs/propuestas/SE-086-ubiquitous-language-extractor.md — APPROVED priority media (M 5h, 2 slices). Skill DDD glossary + extractor python que cruza .memory-store.jsonl con CONTEXT.md per-proyecto, marca términos new/existing/inconsistent. Edge DOMAIN_TERM en grafo episodic (extiende SE-076 Slice 1).
docs/propuestas/SE-087-design-an-interface-parallel.md — APPROVED priority media (M 4h). Skill que spawnea N=3 sub-agentes paralelos con prompts ortogonales (minimalist / compositional / type-safe / pure functional) para destapar trade-offs en diseño de interfaz. Reusa Agent tool existente; opt-in delegate a SE-074 parallel-specs-orchestrator para problemas grandes.
.claude/skills/caveman/SKILL.md (83 LOC) + DOMAIN.md. Modo respuesta ultra-comprimido (~75% reducción tokens). Persistencia explícita hasta "stop caveman" / "modo normal". Auto-clarity exception para warnings de seguridad y operaciones irreversibles. Aligned con Rule #24 radical-honesty.
.claude/skills/zoom-out/SKILL.md (39 LOC). Trigger humano puro (disable-model-invocation: true) para pedir mapa de módulos cuando Mónica entra en área de código desconocida.
.claude/skills/grill-me/SKILL.md (53 LOC). Interrogatorio relentless una pregunta a la vez sobre cada rama del decision tree. Cross-references Rule #24 (challenge assumptions) + Genesis B9 GOAL STEWARD (SE-080).
scripts/skill-catalog-audit.sh — escanea .claude/skills/*/SKILL.md, detecta 4 issue types (missing-frontmatter, missing-name-field, missing-description-field, description-too-short, description-missing-use-when, skill-long, skill-overlong, missing-skill-md). Modos: --report (default), --gate (exit 1 sobre fail-severity), --baseline-write, --json, --skill PATH. Output TSV en output/skill-catalog-audit-YYYYMMDD.tsv.
.ci-baseline/skill-quality-violations.count — baseline inicial captura el estado actual (warn + fail conjugados). Los 9 FAIL son skills con description: > folded scalar vacía (banking-architecture, codebase-map, company-messaging, doc-quality-feedback, evaluations-framework, orgchart-import, prompt-optimizer, resource-references, smart-calendar) — agentes los cargan sin signal de trigger, causa real.
tests/structure/test-pocock-skills-quick-wins.bats — 39 tests certified. Cubre los 3 skills (frontmatter, Use-when, attribution, sizes, cross-refs) + el auditor (negative cases: missing-frontmatter / unknown-arg / nonexistent-path; edge cases: empty dir / overlong skill / zero-LOC; assertion quality: JSON parsing, integer comparisons, status checks).
caveman/SKILL.md → SE-081 Slice única caveman skill
zoom-out/SKILL.md → SE-081 Slice única zoom-out skill
grill-me/SKILL.md → SE-081 Slice única grill-me skill
write-a-skill/SKILL.md discipline → SE-084 Slice 1 auditor heuristics
OK AC-01 caveman SKILL.md ≤100 LOC con frontmatter name, description + "Use when ..." (aspiración spec ≤80, real 83)
OK AC-02 zoom-out SKILL.md ≤100 LOC (aspiración spec ≤30, real 39)
OK AC-03 grill-me SKILL.md ≤100 LOC + cita radical-honesty (aspiración spec ≤30, real 53)
OK AC-04 Headers de los 3 skills citan mattpocock/skills (MIT)
OK AC-05 Ningún skill copia texto literal de Pocock (clean-room)
OK AC-06 Tests BATS estáticos (39 tests certified)
OK AC-07 CHANGELOG fragment
OK AC-01 auditor ejecutable, modos report/gate/baseline-write/json
OK AC-02 Output TSV con 4+ issue types
OK AC-03 Tests BATS ≥12 (cubierto en test-pocock-skills-quick-wins.bats con secciones C9-C11)
set -uo pipefail declarado.
Cero red, cero git operations, cero modificación a archivos fuera de output/ y .ci-baseline/.
Modo --gate solo lee — no modifica skills.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/se081-se084slice1-..., sin push automático ni merge.
docs/rules/domain/architectural-vocabulary.md — vocabulario canónico Module / Interface / Implementation / Seam / Adapter / Depth / Locality + Leverage. Cada término con _Avoid_: explícito (rejection set). 4 principios ratchet: deletion test, interface = test surface, one adapter = hypothetical seam, depth is property of interface. Atribución MIT clean-room a mattpocock/skills/improve-codebase-architecture/LANGUAGE.md.
Cross-references añadidas en 3 sitios:
scripts/architectural-vocabulary-audit.sh — auditor estático sobre outputs recientes de architect y architecture-judge (output/architect-*.md, output/architecture-*.md, output/agent-runs/architect-*/*.md). Detecta términos prohibidos (boundary, component, service, api) en prosa NO en código. Skip de fenced code blocks + inline backticks (identifiers como BoundaryService siguen permitidos en code, prose-only enforcement). Modos: --report (default warning-only, exit 0), --gate (exit 1 sobre violations, hook-up para Slice 2 SE-084 G14), --json, --file PATH.
.claude/skills/tdd-vertical-slices/SKILL.md — codifica el anti-pattern de horizontal slicing ("escribir todos los tests primero, luego todo el código produce tests-de-mentira") y la disciplina vertical (1 test → 1 implementation → repeat). Incluye workflow 4-pasos (planning, tracer bullet, incremental loop, refactor) + per-cycle checklist. Atribución MIT clean-room a mattpocock/skills/tdd/SKILL.md.
.claude/skills/tdd-vertical-slices/DOMAIN.md — secciones convención pm-workspace (Por qué existe, Conceptos de dominio, Reglas de negocio, Relación con otras skills, Decisiones clave, Limitaciones).
Cross-reference en .claude/agents/test-architect.md — sección "TDD Vertical Slices (SE-083)" tras Constraints. NO modifica test-engineer ni test-runner (son ejecutores, no discípulos del pattern).
tests/structure/test-architectural-vocabulary.bats — 31 tests certified. Cubre SE-082 (vocabulary doc structure × 4, cross-refs × 3, auditor positive × 8, auditor negative × 2, auditor edge × 4) + SE-083 (skill structure × 4, cross-refs × 2) + spec ref reinforcement.
improve-codebase-architecture/LANGUAGE.md → SE-082 architectural-vocabulary doc + auditor heurística
tdd/SKILL.md → SE-083 tdd-vertical-slices skill (anti-horizontal-slicing es la contribución central)
OK AC-01 architectural-vocabulary.md ≤200 LOC, define 6 términos con Avoid
OK AC-02 Atribución MIT a Pocock LANGUAGE.md en header
OK AC-03 Cross-reference añadida en attention-anchor.md
OK AC-04 architect agent referencia el doc canónico
OK AC-05 architecture-judge agent referencia el doc canónico
OK AC-06 architectural-vocabulary-audit.sh ejecutable, output warning-only por defecto + modo gate listo
OK AC-07 Tests BATS ≥10 estáticos
OK AC-08 CHANGELOG fragment
OK AC-01 SKILL.md ≤120 LOC con frontmatter name, description que incluye "Use when ..."
OK AC-02 Anti-pattern de horizontal slicing nombrado explícitamente con "DO NOT" / "NO escribas todos"
OK AC-03 Atribución MIT a Pocock tdd/SKILL.md
OK AC-04 Cross-reference añadida en test-architect.md
OK AC-05 Tests BATS ≥8 estáticos
OK AC-06 CHANGELOG fragment
set -uo pipefail declarado en el auditor.
Cero red, cero git operations, cero modificación a archivos fuera de output/ (TSV).
Modo --gate solo lee — no modifica outputs auditados.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/se082-se083-..., sin push automático ni merge.
docs/rules/domain/skill-catalog-discipline.md — define 5 reglas obligatorias para todo SKILL.md nuevo o modificado:
Atribución MIT a mattpocock/skills/write-a-skill/SKILL.md (pattern source).
scripts/pr-plan-gates.sh — añadida función g14_skill_catalog. Filtra git diff origin/main..HEAD --diff-filter=AM a ^\.claude/skills/[^/]+/SKILL\.md$, invoca scripts/skill-catalog-audit.sh --gate --skill <dir> para cada skill modificado, parsea JSON output y agrega FAIL/WARN counts.
scripts/pr-plan.sh — registro G14 tras G13: gate "G14" "Skill catalog audit" g14_skill_catalog.
Comportamiento:
tests/structure/test-skill-catalog-g14.bats — 19 tests certified. Cubre rule doc structure (×4), G14 function structure + safety (×5), G14 behavior (×3), auditor negative cases (×2), edge cases (×3), spec ref + assertion quality reinforcement (×2).
OK AC-04 .ci-baseline/skill-quality-violations.count ya existente desde Slice 1 batch 75
OK AC-05 docs/rules/domain/skill-catalog-discipline.md ≤120 LOC, cita Pocock MIT
OK AC-06 pr-plan G14 activa el gate en modo --gate filtrando a skills cambiados en el PR
OK AC-07 No regression: skills modificados no aumentan violations (G14 skipped en este PR — no se modifica ningún SKILL.md, dogfood pasa)
OK AC-08 Tests BATS = 19 certified
OK AC-09 CHANGELOG fragment con baseline counts (155 inicial — 146 WARN + 9 FAIL — desde batch 75)
〰 AC: G14 wireado a Slice 2 aspecto del spec — DONE
G14 SOLO lee git diff y archivos especificados — no modifica nada
WARN-severity NO bloquea (ratchet) — evita regresión accidental sobre skills antiguos
FAIL-severity bloquea explícitamente con instrucciones de remediation
Cumple docs/rules/domain/autonomous-safety.md: rama agent/se084-slice2-..., sin push automático ni merge.
docs/rules/domain/savia-enterprise/audit-trigger-primitive.md — define el primitive de compliance: trigger function genérico audit_trigger_fn() adjuntable a cualquier tabla regulada via CALL attach_audit('table'::regclass). Captura {table_name, record_id, operation, old_row, new_row, user_id, agent_id, session_id, tenant_id, created_at} en JSONB. Append-only por constraint (REVOKE UPDATE, DELETE). Multi-tenant RLS heredado de SPEC-SE-002. Diseño AFTER (no BEFORE). current_setting(..., true) silencia settings missing. Atribución MIT a dreamxist/balance 00006_audit_log.sql (clean-room, no wholesale import).
docs/rules/domain/savia-enterprise/audit-retention.md — retention policy obligatoria con 7 categorías regulatorias:
scripts/enterprise/audit-search.sh — búsqueda tabular con filtros --tenant <uuid> / --table / --agent / --since (acepta Nd, Nh, Nm, ISO-8601). Output con diff column computado de old_row vs new_row (claves cuyos valores cambiaron). Modo --json para integration. Falla graceful si SAVIA_ENTERPRISE_DSN ausente (exit 3 con mensaje documentado).
scripts/enterprise/audit-purge.sh — DELETE selectivo con guards de seguridad:
tests/structure/test-audit-trigger-primitive.bats — 36 tests certified. Cubre safety (×4), template SQL structure (×7), CLI negative cases (×8), edge cases (×5), retention doc structure (×3), rule doc structure (×4), spec ref reinforcement (×5).
OK AC-01 Tabla audit_log append-only (REVOKE UPDATE/DELETE en template)
OK AC-02 Función audit_trigger_fn() table-agnostic capturando 9 campos
OK AC-03 Procedure attach_audit(regclass) para agregar a tabla nueva
OK AC-04 RLS multi-tenant respetado (SPEC-SE-002)
OK AC-05 Doc retention policy en audit-retention.md con tabla por categoría
OK AC-06 CLI audit-search.sh con filter tenant/table/agent/since
OK AC-07 CLI audit-purge.sh requires --confirm + retention policy file
〰 AC-08 Tests pgTAP ≥8 — DEFERRED: pm-workspace no tiene Postgres en CI; pgTAP queda pendiente del repo Savia Enterprise. BATS ≥6 SÍ cumplido (36 tests certified).
OK AC-09 SQL template docs/propuestas/savia-enterprise/templates/audit-trigger.sql (ya existía desde batch 71)
OK AC-10 Doc docs/rules/domain/savia-enterprise/audit-trigger-primitive.md
〰 AC-11 Migration: attach_audit a 5 tablas regulated existentes — DEFERRED: pm-workspace es repo de config Claude Code; las migrations a tablas reales se aplican en el repo Savia Enterprise post-deploy.
OK AC-12 CHANGELOG entry
set -uo pipefail en ambos CLI.
audit-purge.sh con 6 layers de safety: --table required, --before required, --confirm required, retention doc required, no bulk purge, no self-purge, table-must-be-classified.
Cero red, cero git operations, cero modificación a audit_log desde el script search-only.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-se-037-..., sin push automático ni merge.
scripts/enterprise/jwt-mint.sh — CLI bash que intercambia API key por JWT efímero HS256:
docs/rules/domain/savia-enterprise/agent-jwt-mint.md — define el modelo:
tests/structure/test-jwt-mint-primitive.bats — 36 tests certified. Cubre safety×3, template SQL structure×6, CLI negative×9 (incluye boundary/empty edge cases), edge×6, rule doc structure×6, spec ref reinforcement×2, exit codes×1, defaults×1.
OK AC-01 Tabla api_keys con SHA-256 hash + key_prefix UX (template ya existente desde batch 71)
OK AC-02 Función mint_jwt() con scope downscoping enforce (split: api_key_verify + api_key_scope_is_subset SQL helpers + jwt-mint.sh bash signer)
〰 AC-03 4 CLI commands (create, list, revoke, mint) — DEFERRED Slice 2: solo mint aquí; create/list/revoke en batch siguiente
OK AC-04 JWT efímero ≤900s, scope mínimo signed (TTL clamp [60,3600])
OK AC-05 Audit trail api_key_mints append-only (sinergia SPEC-SE-037)
〰 AC-06 Hook block-pat-file-write.sh — DEFERRED Slice 3
〰 AC-07 block-credential-leak.sh PAT-shaped detection — DEFERRED Slice 3
〰 AC-08 BATS ≥12 certified — OK cumplido (36 tests certified); pgTAP ≥6 DEFERRED (no Postgres en pm-workspace CI)
OK AC-09 Doc docs/rules/domain/savia-enterprise/agent-jwt-mint.md
OK AC-10 SQL template docs/propuestas/savia-enterprise/templates/api-keys.sql (batch 71)
OK AC-11 CHANGELOG entry (este fragmento)
set -uo pipefail en jwt-mint.sh.
--key-stdin flag explícito para evitar leak vía ps.
Refuse de upscoping (exit 8): scope_subset MUST ⊆ stored_scope.
TTL clamp [60, 3600] — refuse fuera de rango (DoS prevention).
Audit log non-blocking (warning a stderr si api_key_record_mint falla, JWT sí emitido — disponibilidad operacional priorizada; auditoría reconstruible off-line).
Cero red, cero git operations.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-se-036-..., sin push automático ni merge.
scripts/enterprise/api-key-create.sh — genera key fresca de 32 bytes urandom (base64url, prefijo savia_), inserta sha256(plaintext) + key_prefix (8 chars UX) en api_keys, e imprime el plaintext exactamente una vez con warning destructivo. Soporta --tenant <uuid> (validado RFC4122 shape), --scope <csv>, --desc "...". Exit codes 2 (usage) / 3 (DSN) / 4 (psql/openssl) / 5 (insert fail).
scripts/enterprise/api-key-list.sh — inventario tabular o JSON. Filtros: --tenant <uuid>, --active, --revoked, --json. NUNCA muestra key_hash ni plaintext (zero-leakage por construcción). Mutually exclusive --active y --revoked. Validación UUID antes de DSN/psql checks.
scripts/enterprise/api-key-revoke.sh — revoca por prefix de 8 chars con 4 safety layers:
tests/structure/test-api-key-cli-suite.bats — 35 tests certified. Cubre file-level safety×6, create negative×5, list negative×4, revoke negative×5, edge×8 (boundary, no-leak, dry-run, ttl-warning), spec ref×3, exit code documentation×3.
docs/rules/domain/savia-enterprise/agent-jwt-mint.md — sección "CLIs operacionales (Slice 2)" añadida; "No hace" actualizado para reflejar que solo Slice 3 queda pendiente.
OK AC-03 4 CLI commands (create, list, revoke, mint) — Slice 1 ya hizo mint; Slice 2 completa los 3 restantes
〰 AC-06 Hook block-pat-file-write.sh — DEFERRED Slice 3
〰 AC-07 block-credential-leak.sh PAT-shaped detection — DEFERRED Slice 3
OK AC-09 Doc actualizado con Slice 2
set -uo pipefail en los 3 CLIs.
api-key-create.sh: plaintext printado exactamente UNA vez, sin escritura a disco. Recuperación = revoke + recreate.
api-key-list.sh: zero-leakage — la SQL SELECT explícitamente excluye key_hash y plaintext (grep -qE "SELECT.*key_hash" falla, enforced por test).
api-key-revoke.sh: 4 safety layers (prefix required, no bulk, no wildcards, dry-run default + actor logging).
Cero red, cero git operations.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-se-036-slice-2-..., sin push automático ni merge.
scripts/enterprise/reconciliation-status.sh — wrapper bash que itera dimensiones activas para un tenant y renderiza:
docs/rules/domain/savia-enterprise/reconciliation-delta-engine.md — define el modelo:
tests/structure/test-reconciliation-delta-engine.bats — 40 tests certified. Cubre file-level safety×7, SQL template structure×5, delta-tier helper positive+edge×8, status CLI negative×6, edge structural×5, rule doc structure×6, exit codes + spec ref×3.
OK AC-01 Función tenant_reconciliation_status() definida con 3 tiers (en template, batch 71)
〰 AC-02 4 dimensiones seed — DEFERRED Slice 2 (rule doc lista plan conceptual)
〰 AC-03 Vista materializada — DEFERRED Slice 3 follow-up (Slice 1+3 entregan on-demand CLI; vista materializada es optimización posterior)
OK AC-04 CLI reconciliation-status.sh <tenant> con output ANSI
〰 AC-05 Trigger recompute on-write con webhook — DEFERRED Slice 4
〰 AC-06 Baseline ratchet bloquea regresión — DEFERRED Slice 5 (precursor: --fail-on exit 7 ya implementado)
〰 AC-07 pgTAP ≥10 — DEFERRED (no Postgres en pm-workspace CI); BATS ≥8 OK cumplido (40 tests certified)
OK AC-08 Doc docs/rules/domain/savia-enterprise/reconciliation-delta-engine.md
OK AC-09 SQL template docs/propuestas/savia-enterprise/templates/reconciliation.sql (batch 71)
OK AC-10 CHANGELOG entry (este fragmento)
set -uo pipefail en reconciliation-status.sh y delta-tier.sh.
UUID validation antes de DSN/psql checks (fail-fast en arg validation).
--fail-on solo acepta green|amber|red (no levels custom; mantiene API limpia).
Cero red, cero git operations.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-se-035-..., sin push automático ni merge.
scripts/image-relevance-filter.sh — CLI bash con 3 subcomandos:
docs/rules/domain/image-relevance-filter.md — define el modelo:
tests/structure/test-image-relevance-filter.bats — 33 tests certified. Cubre file safety×3, spec ref×2, usage/negative×6, heuristic positive×7 (incluye auto-promote ≥3 + boundary 2-marks-not-enough), edge×5 (cache idempotencia, JSON sha 64-char, zero-byte), rule doc structure×6, spec ref + exit codes×3, graceful degradation×1.
OK AC-01 scripts/image-relevance-filter.sh con check/skip/log subcomandos
OK AC-02 Heurísticas de tamaño, aspect ratio, cache de checksums (4 reglas en orden)
〰 AC-03 word-digest, pptx-digest, excel-digest consultan filtro antes de Vision — DEFERRED Slice 2 (modifica 3 agents existentes, requiere greenlight)
OK AC-04 Log de decisiones en cache (refinamiento automático via auto-promote ≥3 marks)
OK AC-05 Tests BATS ≥12 — cumplido (33 tests certified)
〰 AC-06 Reducción medible 40% Vision — DEFERRED medible solo tras Slice 2 integration en producción
set -uo pipefail en image-relevance-filter.sh.
Cache off-repo en ~/.savia/digest-cache/ — nunca toca el repo.
NO modifica agents existentes en este Slice (riesgo cero sobre flow actual).
Skip-list idempotente (escribir mismo sha 2× no duplica fila).
Cero red, cero git operations.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-103-..., sin push automático ni merge.
.claude/agents/recommendation-tribunal-orchestrator.md — sonnet, convoca los 4 jueces en paralelo via Task, agrega vía aggregate.sh, persiste audit JSON. Sync, latency budget hard-cap 3s wall-clock con graceful timeout fallback (verdict WARN + razón "timeout").
.claude/agents/memory-conflict-judge.md — sonnet, lee ~/.claude/external-memory/auto/MEMORY.md + filtra candidatos + lee files completos. Veto cuando draft contradice un feedback_*.md o user_*.md con confidence ≥ 0.8.
.claude/agents/rule-violation-judge.md — sonnet, carga lazy de CLAUDE.md + reglas críticas + radical-honesty + autonomous-safety + dominio. Veto en violación de Rule #1 / Rule #8 / autonomous-safety / radical-honesty / zero-project-leakage.
.claude/agents/hallucination-fast-judge.md — haiku (fast), verifica entidades del draft via tool calls reales: [ -f path ], grep de funciones, --help de flags. Veto con ≥1 entidad fabricada confidence ≥ 0.9 (verificación definitiva).
.claude/agents/expertise-asymmetry-judge.md — sonnet, lee perfil expertise.md del usuario activo. NUNCA veta — solo muta el output forzando rewrite con secciones "Por qué creo esto / Alternativas que descarté / Cómo verificar tú misma" cuando draft cae en área marcada audit_level: blind.
scripts/recommendation-tribunal/classifier.sh — heurística primer paso (sin LLM, <50ms). 3 niveles: critical (bypass / lower threshold / hook-skip flags / force push, también en español), high (sudo / rm -rf / drop table / prod deploy), medium (te recomiendo / should use / el problema es). Solo medium+ activa el panel.
scripts/recommendation-tribunal/aggregate.sh — agregación deterministic (no LLM). Lee 4 JSONs, aplica veto rules, computa consensus_score, decide PASS/WARN/VETO. Hard rule: veto requiere confidence ≥ 0.8.
scripts/recommendation-tribunal/banner.sh — renderiza markdown banner según verdict. PASS = empty (passthrough). WARN = banner + draft. VETO = banner que muestra el draft original como blockquote citado pero claramente marcado como bloqueado (auditabilidad — la usuaria SIEMPRE ve qué le iba a decir Savia).
.claude/hooks/recommendation-tribunal-pre-output.sh — hook PreOutput listo para activar pero NO añadido a .claude/settings.json. En Slice 1 el hook está en modo detect-only: clasifica + persiste audit log + pasa el draft sin mutación. Esto permite calibrar el classifier en uso real antes de wirear los vetos.
docs/rules/domain/recommendation-tribunal.md — define el modelo:
tests/structure/test-recommendation-tribunal.bats — 51 tests certified. Cubre file safety×6, agent files×7, classifier positive×9 (bilingüe), classifier negative×5, aggregate×7 (verdict PASS/WARN/VETO + boundary confidence), banner×4, hook detect-only×3, rule doc×6, spec ref + meta×3.
OK AC-01 Classifier detecta recomendaciones accionables (heurística bilingüe + JSON output) — precision validable contra golden set en uso real (modo detect-only)
OK AC-02 4 jueces implementados como agentes con prompts versionados y modelos asignados (haiku/sonnet)
〰 AC-03 Latency p95 < 3s — DEFERRED medible solo tras activación; targets definidos en rule doc
OK AC-04 Verdict VETO bloquea entrega; banner muestra contenido vetado claramente marcado
OK AC-05 Banner WARN se inyecta antes del output con findings concretos
OK AC-06 memory-conflict-judge cita el .md del auto-memory en conflicto
OK AC-07 rule-violation-judge cita el path + line range de la regla violada
OK AC-08 hallucination-fast-judge verifica entidades con tool calls (grep, [ -f ], --help)
OK AC-09 expertise-asymmetry-judge re-escribe con secciones blind-area calibration (template emit; rewrite real Slice 2)
OK AC-10 Audit trail JSON persiste en output/recommendation-tribunal/<date>/<hash>.json
〰 AC-11 Falso positivo registrado por usuaria → feedback memory — DEFERRED Slice 3
〰 AC-12 Falso negativo registrado → feedback memory regression — DEFERRED Slice 3
OK AC-13 BATS ≥30 tests certified — 51 tests (sobrepasa target en 70%)
〰 AC-14 Regression test 6 patterns reportadas — DEFERRED golden set Slice 2
OK AC-15 CHANGELOG entry + spec status (este fragmento; spec status update en próxima iteración)
set -uo pipefail en los 4 scripts.
Hook NO añadido a settings.json — activación es paso humano deliberado tras revisión completa.
Modo detect-only en Slice 1 — el hook clasifica + audita pero NO muta el draft.
Audit trail append-only, off-repo en output/.
Clasificador es pura heurística (sin LLM call) — 0ms timeout risk en este path crítico.
Cero red, cero git operations en cualquiera de los 4 scripts.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-125-slice-1-..., sin push automático ni merge.
scripts/cognitive-debt.sh — CLI principal con 5 subcomandos:
.claude/hooks/cognitive-debt-telemetry.sh — PostToolUse async, registra cada Edit/Write/Task. Sin LLM call (CD-01), JSONL append en ~/.savia/cognitive-load/{user}.jsonl (CD-03 N3), exit 0 siempre (CD-02 nunca bloquea).
.claude/hooks/cognitive-debt-hypothesis-first.sh — PreToolUse WARNING-ONLY en Phase 1. Lee git log -5, si los últimos commits no tienen trailer Hypothesis: emite nudge a stderr (una vez por sesión vía /tmp/marker). Phase 2 escalará a soft-block con escape --skip-cognitive.
.claude/commands/cognitive-status.md — wrapper de cognitive-debt.sh status + summary invocable como /cognitive-status.
docs/cognitive-debt-guide.md — guía canónica:
tests/structure/test-cognitive-debt-phase-1.bats — 38 tests certified. Cubre file-level safety×6, status×3, forget×3 (incluye boundary --confirm required), dispatch errors×3, telemetry hook (privacy + safety + edge cases)×5, hypothesis-first (warning-only enforcement)×4, restricciones CD-01 a CD-04 verificadas explícitamente×4, edge cases (missing dir, empty log)×3, doc + command×5, spec-ref + meta×2.
Kosmyna et al. 2025 "Your Brain on ChatGPT" (arxiv.org/abs/2506.08872) — base para I4 telemetry necessity
Lee et al. CHI 2025 (Microsoft Research + CMU) — base para fast-accept ratio como proxy de skip-verification
CMU ICER 2025 (arxiv.org/pdf/2509.20353) — alerta sobre atrofia diferencial en metacognición débil
Roediger & Karpicke 2006 — base para I1 hypothesis-first (retrieval practice)
OK Phase 1.1 scripts/cognitive-debt.sh con 5 subcomandos (enable/disable/status/summary/forget)
OK Phase 1.2 Hook telemetry async (I4) + comando /cognitive-status
OK Phase 1.3 Hook hypothesis-first (I1) en modo WARNING ONLY, no bloquea (Phase 2 escalará)
OK Phase 1.4 BATS tests certified
OK Phase 1.5 Documentación: docs/cognitive-debt-guide.md con evidencia académica + Phase 2/3 deferred items
CD-01 verificada por test: ! grep -qiE 'anthropic|/v1/messages|claude.api' en los 3 scripts. Cero LLM calls.
CD-02 verificada por test: telemetry hook always exit 0, hypothesis hook always exit 0 (Phase 1 warning-only).
CD-03 verificada por test: telemetry path en ~/.savia/cognitive-load/, gitignored en N3.
CD-04 verificada por test: status muestra DISABLED (opt-in default per CD-04) por defecto.
Hooks NO añadidos a settings.json en este batch — opt-in significa que la usuaria los activa con cognitive-debt.sh enable que hace el wire en settings.json (con backup).
set -uo pipefail en first 5 lines en los 3 scripts (verificado por test contra el lint G5b).
Cero red, cero git operations en los hooks.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-107-phase-1-..., sin push automático ni merge.
tests/structure/test-spec-se-001-foundations.bats — 30 tests certified. Estructura por AC:
Cero modificación de código existente — solo el status frontmatter (1 línea) y nuevo BATS test.
El BATS test usa subdirectorios temporales en mktemp -d para no modificar el workspace real durante runs.
set -uo pipefail semantics aplicado en bats subshells.
Cero red, cero git operations.
Cumple docs/rules/domain/autonomous-safety.md: rama agent/spec-se-001-foundations-slice-1-..., sin push automático ni merge.
`CHANGELOG.d/`: nuevo directorio de fragments per-PR. Cada PR crea un fichero en vez de editar CHANGELOG.md al top. Zero-conflict by design (fix root-cause del problema "mergear un PR rompe el CHANGELOG de todos los demás"). README.md documenta lifecycle + pattern.
`scripts/changelog-fragment.sh`: helper para crear fragment desde CLI. Flags --slug --version-bump --section --entry (repetible) --from-stdin. Validación estricta de parámetros. Si fragment existe, append idempotente a misma section.
`scripts/changelog-consolidate.sh`: consolidador de release. Agrupa fragments por section, calcula siguiente versión (highest bump wins), inyecta nueva entry al top de CHANGELOG.md, añade link line, borra fragments. Modo --dry-run para preview.
`scripts/resolve-all-open-prs.sh`: bugfix — snapshot del resolver a tempfile ANTES de checkout para evitar "No such file or directory" al cambiar a branches que no contienen el script. trap EXIT limpia tempfile.
scripts/spec-frontmatter-migrate.sh: herramienta mecánica para migrar specs con status en prosa (> Status: **DRAFT**) a YAML frontmatter canónico. Mapping fijo: DRAFT→Proposed, ACTIVE/IMPLEMENTING→IN_PROGRESS, READY→ACCEPTED, COMPLETE/DONE/PHASE N DONE→Implemented, REJECTED→Rejected, resto→UNLABELED. Preserva body intact (prepend-only). Modos --dry-run / --apply / --spec PATH / --limit N (max 50).
tests/test-spec-frontmatter-migrate.bats: 32 tests — safety, CLI, mapping canónico (7 status types), estructura frontmatter, idempotencia, single-spec mode, negatives, edges. Auditor score 89.
docs/propuestas/SE-040-agent-degradation-canary.md: skill futuro agent-health-canary con 5 métricas canarias (Read:Edit ratio, stop hook violations, interrupt rate, edit-without-read count, token efficiency). Analiza ventana 48h vs baseline 7d de session-actions.jsonl. Feasibility Probe 1.5h blocking. Slicing 3 slices.
docs/propuestas/SE-041-memvid-portable-memory-probe.md: spec que propone evaluar mediante Feasibility Probe obligatorio el formato .mv2 de memvid (github.com/memvid/memvid, 15k stars, Apache 2.0, v2.0 de 2026-03) para 3 usos narrow: backup/restore, snapshot audit, travel portability. NO es reemplazo del stack actual (SPEC-027/018/035/123 ya cubren search+graph). Probe 2h blocking; criterios cuantitativos: ingest 100 docs <30s, retrieval p50 <50ms offline, round-trip byte-identical.
`docs/propuestas/SE-042-savia-voice-training-pipeline.md`: spec con pipeline 3-fases (extract, filter PII, quality gates) para convertir session-actions.jsonl en dataset JSONL para SE-027 SLM fine-tuning. Patrón robado de WeClone (github.com/xming521/WeClone, 17.6k stars, AGPL-3.0) — SOLO pattern, no código (AGPL incompatible). Feasibility Probe 2h blocking con yield ≥500 muestras/30days.
`docs/propuestas/ROADMAP.md`: Tier 4.11 añadido (SE-042). Complementa SE-027 (training) con data-prep específico de voz Savia.
batch 3 extensions: pdf-extract-probe (SPEC-102 Slice 1) + gaia-benchmark-harness (SPEC-100 Slice 1) + 43 BATS tests
CodeGraph as opt-in MCP engine for .acm and ast-comprehension — tree-sitter + SQLite indexing, 8 MCP tools, disabled by default in opencode.json
docs/rules/domain/codegraph-confidentiality.md: gitignore mandate for .codegraph/, prohibited in N4b scope
New skill .claude/skills/codegraph/ documenting integration with agent-code-map and ast-comprehension
Era 198 — 6 specs alineados con tesis Anthropic 'Building Effective Agents' (SPEC-156..161). Tier 1 HIGH (13h): token-budget frontmatter, context pre-flight check, workflow-vs-agent decision gate. Tier 2 MEDIUM (13h): async tribunal fan-out, tool-ergonomics auto-audit. Tier 3 LOW research (12h): self-evolving tools. ROI estimado: heavy tier -40% tokens, tribunal wall-time -65%.
SPEC-147 Slice 3 — 10/10 decision trees complete (dev-orchestrator, court-orchestrator, frontend-developer + AC-05 docs in best-practices-claude-code.md §19). BATS 36/36 PASS.
SPEC-156 Slice 1: token_budget frontmatter declared on 70/70 agents (defaults per Anthropic Effective Agents tier: fast 30k/8k/escalate, mid 60k/15k/escalate, heavy 100k/20k/block)
scripts/spec-156-migrate-token-budget.sh: idempotent migration script for token_budget injection
SPEC-161 PROTECTED_JOB_NAMES: hook + YAML allowlist blocking costly agents (architect, sdd-spec-writer, pentester, ...) in autonomous loops (overnight-sprint, code-improvement-loop). Override via SAVIA_PROTECTED_JOB_OVERRIDE env. Fail-safe permissive if YAML missing. 15 BATS tests.
.claude/skills/savia-identity/SKILL.md — workspace self-awareness skill.
.claude/skills/savia-memory/SKILL.md — persistent memory management skill.
.claude/skills/savia-identity/DOMAIN.md — domain documentation for workspace identity skill.
.claude/skills/savia-memory/DOMAIN.md — domain documentation for persistent memory skill.
docs/rules/domain/hook-event-equivalence.md: tabla de mapeo Claude Code ↔ OpenCode (17 eventos), gap analysis (6 sin equivalente), documentación del bridge savia-gates, y roadmap de migración en 4 fases.
Five DiffusionGemma-inspired specs IMPLEMENTED in single PR. SPEC-196 freeze-done elements scripts/recommendation-tribunal/early-cancel.sh — poll-and-cancel pattern for tribunal judges with conf>=0.95 veto. 16 bats tests verdes. SPEC-198 JudgeVerdict frozen dataclass scripts/recommendation-tribunal/judge_verdict.py — schema v1 con validacion + backward compat con dicts legacy. 32 pytest verdes. SPEC-195 iterative tribunal scripts/recommendation-tribunal/early_stop.py + iterate.sh — multi-criteria early stop (stability+entropy+max_iter) con telemetria JSONL. 23 pytest + 13 bats verdes. SPEC-197 annealing schedule scripts/annealing-schedule.py — interpolation curve (max_t -> min_t via exponent) para multi-phase agents. 17 pytest verdes. SPEC-200 adaptive quality gate threshold scripts/quality-gate-adaptive.py — proportional to score distribution con high_mean_strict + low_mean_tolerant + floor/ceil clamps. 21 pytest verdes. Total 122 tests new. Pattern: ChainedEarlyStop + _WhileLoopCarry.done + AnnealingTemperatureShaper + dataclass frozen + entropy_bound proportional. SPEC-199 (historical context conditioning) queda en backlog — depende de SPEC-195 ahora desbloqueado. Defaults conservadores: cada componente opt-in via env vars; telemetria-first activation.
13 specs Q2/Q3 2026 (SPEC-141..153) from research output/research-tendencias-workspaces-agentes-2026-20260523.md. Numbered 141..153 to avoid collision with main's SPEC-128 (Obsidian Context-as-Code, PR #762).
SPEC-141 — MCP Curated Catalog (opencode.jsonc mcp block + OAuth DCR + Server Cards, 6h)
SPEC-142 — Plugin tool.execute.before auto-redaction of secrets (TS, 5h)
SPEC-144 — /speckit.* slash command aliases (4h)
SPEC-145 — Import anthropics/skill-creator + mcp-builder (3h)
SPEC-146 — excellent-* monthly watcher (4h)
SPEC-147 — Decision trees for top-10 agents (10h)
SPEC-149 — Sandbox OS-level via opencode-sandbox plugin + Docker (20h)
SPEC-150 — Hooks multi-handler migration (TS plugin for LLM judges + MCP audit, 35h)
SPEC-151 — Evals CI gate (DeepEval + Promptfoo, paired-delta, 24h)
SPEC-152 — Hierarchical orchestrator delegation (feature leads, 18h)
SPEC-153 — Memory bi-temporal + episodic→semantic consolidation (22h)
SPEC-143 — Skill conformance audit (ABORTED: premise false — Rule #11 caps at 150 lines)
SPEC-148 — Skill progressive disclosure split (ABORTED: no real use case)
Plugin fix: MODEL_TIER_MAP in .opencode/plugins/savia-foundation.ts resolves abstract tiers heavy/mid/fast to Copilot model IDs (SPEC-127 / model-alias-schema)
Critical review report (49 findings, 7 GREEN / 4 YELLOW / 2 RED) — micro-fixes applied to SPEC-147, SPEC-150, SPEC-152, SPEC-153
scripts/memory-index-rebuild.sh — regenera el índice auto/MEMORY.md desde el JSONL store.
docs/rules/domain/savia-memory-architecture.md — documento de arquitectura unificada de las 4 capas de memoria (L0-L3).
_update_memory_index() en memory-store.sh — actualiza el índice canónico en cada save.
output-taxonomy rule: taxonomía canónica de output/ con 5 subcategorías (research/, audit logs, pipeline, session-state, raíz)
SE-220 spec PROPOSED: Speculative Tool Execution — draft+verify pattern aplicado a tool calls; feasibility-probe bloqueante S0
CLAUDE.md: entrada lazy-ref a output-taxonomy.md
SE-104: docs/rules/domain/savia-ethical-principles.md — 13 humanist principles + 5 immutable red lines, cross-referenced from CLAUDE.md, savia.md, autonomous-safety, radical-honesty, equality-shield
SE-094: scripts/hooks-integrity-check.sh — bidirectional phantom/orphan detector with symlink + scripts/ search path; registers 4 orphan hooks in settings.json (android-adb-validate, cognitive-debt-hypothesis-first, cognitive-debt-telemetry, project-isolation-gate)
SE-095: scripts/count-commands.sh — canonical command counter (559 total)
SE-093: .claude/commands/project-switch.md — slash command for active project switching
SE-105 (PROPOSED): docs/propuestas/SE-105-glm-governance-manifest.md — adopt GLM v1.0 RFC 8615 well-known manifest; Tier 2 §3.2 rank 16, M 4h
SE-106 (PROPOSED): docs/propuestas/SE-106-tiered-tribunal-execution.md — tiered exec for Truth Tribunal + Code Review Court (Tier 0 sequential early-stop on veto, Tier 1 parallel); ~67% token savings on vetoed runs; Recommendation Tribunal excluded (sync latency p95<3s); ROADMAP §3.2 rank 17, M 5h
8 spec drafts: SE-096..103 from 2026-05-27 obsolete/legacy audit
SE-029 Slice 2: context-task-classify.sh (6 classes deterministic) + SKILL.md + 34 BATS tests (auditor 98)
SE-029 Slice 3: compress-skip-frozen.sh advisory (SKIP/COMPRESS/COMPRESS_LIMITED verdicts per frozen flag) + 28 BATS tests (auditor 90)
SE-029 Slice 4: operational-point-selector.sh (plan de compresión multi-turn con FITS/OVERFLOW verdict) + 30 BATS tests
SE-034 Slice 1: dag-typing-validate.sh validator (audit/validate/infer modes, 6 canonical types text|json|markdown|yaml|binary|any) + 27 BATS tests (auditor score 87)
SE-252 Bus Factor Shield: bus-factor scanner, CONTEXT_DOME.md generator, onboarding planner, and bus-factor-warn hook
SLM batch 2: portfolio-contention (SE-020 Slice 4) + slm-export-gguf (llama.cpp recipe) + slm-deploy orchestrator + slm-data-collect (harvest 321 training samples from workspace) — 99 BATS tests total
SLM/SE-020 batch 3: portfolio-deps-status + gitagent-export (SPEC-099) + docs EN (slm-training-pipeline, portfolio-as-graph, receipts-protocol) + 69 BATS tests
SLM Fase 1 scaffolding: slm-dataset-validate.sh pre-training validator (Alpaca schema, PII scan, dedup, length stats, --allow-pii override) + 29 BATS tests
SLM batch 2: slm-eval-compare.sh A/B eval results (PROMOTE/ROLLBACK/REJECT verdicts, per-benchmark delta, --min-improvement) + 24 BATS tests
SLM Fase 5 scaffolding: slm-eval-harness-setup.sh (emite eval-config.yaml + prompts.jsonl, 7 benchmarks allow-list, no ejecuta eval) + 27 BATS tests
SLM scaffolding: slm-modelfile-gen.sh Ollama Modelfile generator (5 personas incl. savia, ChatML template, no ejecuta ollama) + 30 BATS tests
SLM scaffolding: slm-pipeline-validate.sh meta-validator (layout, config.yaml, eval/harness.yaml, gitignore privacy) + 28 BATS tests
SLM scaffolding: slm-project-init.sh bootstrap canonical layout (config.yaml + datasets/* + eval + README + .gitignore privacy) + 29 BATS tests
SPEC-SE-027 Model Registry: slm-registry.sh (register/list/show/promote/deprecate, single-deployed invariant, manifest.json) + 32 BATS tests
SE-028 Slice 1 scaffolding: slm-synth-recipe.sh (5 synth strategies, sovereignty-aware judge selection, oumi recipe YAML) + 30 BATS tests
SLM training pipeline scaffolding (SPEC-SE-027/SPEC-023/SPEC-080/SE-028/SE-042 approved): slm-dataset-prep.sh + slm-train-config.sh + slm-training-pipeline.md rule + 35 BATS tests. GPU execution deferred — software ready ahora
SPEC-125 Slice 2: scripts/recommendation-tribunal/expertise-rewrite.sh — rewrite calibrado en areas blind con razonamiento + alternativas + verificacion
SPEC-125 Slice 2: scripts/recommendation-tribunal-search.sh — CLI de inspeccion del audit-trail con filtros --from/--to/--verdict/--risk/--hash y modos --summary/--json
SPEC-125 Slice 2: .claude/profiles/users/template/expertise.md — TEMPLATE generico para que cada usuario declare sus areas de expertise (audit_level: blind/low/medium/high)
SPEC-125 Slice 2: tests/test-spec-125-slice-2-asymmetric-expertise.bats — suite BATS certificada por SPEC-055 quality gate
SPEC-125 Slice 3: Memory feedback loop + tribunal calibration. Recorder followup-record.sh classifies user reactions (fp/fn/neutral). Calibrator calibrate.sh emits feedback memories from fp/fn records. Hook recommendation-tribunal-followup.sh wire-ready (no-op default). Canonical rule docs/rules/domain/tribunal-calibration.md. Classifier expanded with 4 new pattern groups (5/6 regression patterns caught by heuristic). 42 BATS tests, score 86/100 certified.
SPEC-141: MCP curated catalog with 10 templates (7 third-party filtered by BlueRock checklist + 3 own skills as MCP stdio). audit-mcp-templates.sh validates schema, secrets, scope. Server Cards for auto-discovery. mcp.json stays empty by default (Rule #19).
SPEC-142: OpenCode plugin auto-redact-credentials uses tool.execute.before to mutate args before tool execution. Redacts GitHub/OpenAI/Anthropic/AWS tokens and connection strings before they reach the LLM. Integrated into savia-foundation plugin chain. Includes TS Vitest tests and BATS smoke tests.
SPEC-145: vendor anthropics/skills upstream under external/anthropic-skills/ (skill-creator 485 lines + mcp-builder, Apache-2.0). external/ is intentionally outside .claude/ to preserve upstream fidelity (Rule #11 applies to .claude/ only). external/README.md documents vendoring policy.
SPEC-146: monthly ecosystem watcher — GitHub Action signals deltas in excellent-* repos and upstream tooling (spec-kit, MCP, opencode). Snapshots in .savia-memory/ (gitignored), no autonomous actions.
SPEC-147 Slice 1: Decision trees for 3 pilot agents (architect, code-reviewer, security-guardian). Each agent gains a deterministic decision-tree document under .claude/agents/decision-trees/ and a frontmatter reference. OpenCode mirror via symlink. Validation via tests/test-agent-decision-trees.bats.
SPEC-147 Slice 2 — 3 more decision trees: dotnet-developer, business-analyst, sdd-spec-writer (7/10 trees done, 3 remaining)
SPEC-149: Sandbox OS-level para modos autonomos (Capa A + Capa B + Capa C).
SPEC-157: Context Pre-Flight Check — multi-source token estimator + PreToolUse hook. Detects @-import file refs, skill refs, prompt tokens; warns at 80%, blocks at 100% of agent budget; caches by input hash; suggests context-rot-strategy + context-task-classifier. 29 tests certified 81/100.
scripts/router-mode-classifier.py: heuristic classifier for System 1/2 dispatch.
scripts/_router_extract_helper.py: internal helper that extracts {intent, command,
scripts/_router_telemetry_helper.py: internal helper that assembles JSONL telemetry
.opencode/hooks/router-mode-dispatch.sh: PreToolUse hook. Reads SAVIA_ROUTER_MODE
complexity_tier: mode1 added to frontmatter of 8 commands:
complexity_tier: mode2 added to frontmatter of 1 command: savia-shield.
output/router-decisions.jsonl: telemetry log (created at first shadow/enforce invocation).
tests/scripts/test_router_classifier.py: 43 pytest tests — all pass.
tests/bats/test-router-mode-dispatch.bats: 17 BATS tests — all pass.
Default SAVIA_ROUTER_MODE=shadow (classifies but does not enforce).
Classifier uses stdlib only — no external dependencies.
2 weeks shadow period recommended before flipping to enforce (AC from spec).
.opencode/hooks/memory-feedback-task.sh: PostToolUse hook that fires when
scripts/memory-feedback-extractor.py: Standalone extractor. Reads
scripts/memory-feedback-post-merge.sh: Detects merged PRs via git log
scripts/memory-feedback-compactor.py: Reads MEMORY.md, identifies
docs/rules/learned/ directory created (destination for promoted patterns).
tests/scripts/test_memory_feedback.py: 16 pytest tests — extractor outcome
tests/bats/test-memory-feedback.bats: 12 BATS tests — hook existence,
SPEC-169 Project Twin — artefacto versionado por proyecto: twin.md schema, linter (AC-7 campos prohibidos), refresher determinista <5s, loader con cap 2000 tokens (AC-2), hook PostToolUse, decay check, anonymize N1 (AC-6). 33+ tests BATS score 97.
SPEC-188 Root-Cause Investigation Architecture (meta-spec): 5 piezas P1-P5 (Failure Pattern Memory, Causal Confidence Channel, Investigation-First Hook, Responsibility Judge, Shortcut-Tax Telemetry) con 4 sub-specs detalladas y Fase 0 ejecutada.
Fase 0 (SPEC-188) .claude/rules/domain/feedback/feedback_root_cause_always.md: feedback canonico que cierra el hallazgo G3 (memory-conflict-judge referenciaba fichero inexistente). Path N1, tracked, source-of-truth. 10 patrones prohibidos + 4 excepciones documentadas.
Tests SPEC-188: tests/test-spec-188-architecture.bats (20 tests) cubre frontmatter, piezas, sub-specs referenciadas, hallazgos, fasificado, feature flags, Bridge SE-072, Fase 0 closure, ROADMAP y CHANGELOG presence.
scripts/context-greedy-budget.py and .sh wrapper: stdlib-only Python module
.opencode/hooks/context-greedy-inject.sh (PreToolUse Read): intercepts
Validated against 11 real project ACMs (45 query/file combinations).
Tests: 88 of 88 green
SPEC-190 Slice 1: code-twin-lint.sh (CTF/CTI/seeds validator), code-twin-init.sh scaffold, canonical rule application-code-twin.md; 41 BATS tests 98/100 certified
SPEC-190 Slice 2: domain+application CTF fixtures (entities, value-objects, business-rules, use-cases, commands, queries), extended code-twin-init.sh with DRAFT stubs, 47-test lint suite (98/100), 29-test init suite (88/100)
SPEC-190 Slice 3: infrastructure CTF fixtures (repos/item-repository, external/notification-client), extended code-twin-init.sh with infrastructure DRAFT stubs (db/schema, repos, external), 49 lint tests (98/100), 34 init tests (88/100)
SPEC-190 Slice 4: api CTF fixtures (routes, error-codes), frontend CTF fixture (item-list-component), extended code-twin-init.sh with api/frontend DRAFT stubs, 52 lint tests (98 sobre 100), 42 init tests (88 sobre 100)
SPEC-190 Slice 5: code-twin-simulate.sh — symbolic simulation engine (Python); seeds users.jsonl (sim: convention); 51 BATS tests (93 sobre 100); AC-3+AC-4
SPEC-190 Slice 6: code-twin-validate-spec — spec feasibility validator against ACT (route_duplicate detection, scoring, JSON output, 46 BATS tests 93 sobre 100)
SPEC-190 Slice 7: code-twin-extract.sh+py — AST-light extractor for TypeScript (@Injectable/@Controller/@Component) and C# (namespace→layer) producing CTFs with 8-field frontmatter
SPEC-190 Slice 8: code-twin-load.sh (AC-10 token-budget summary mode), code-twin-sync-check.sh (AC-12 stale detection), code-twin-anonymize.sh (AC-14 project anonymizer), code-twin-agent definition
SPEC-190 Slice 9: pilot twin savia-web with 8 CTFs, simulate AC-V1 (confidence=0.95), 64 BATS tests (93 sobre 100)
scripts/anti-adulation/lexical-strip.py + regex-patterns.json: Layer 1 deterministic detector for empty social validation patterns. Hook .opencode/hooks/sycophancy-strip.sh with 5 modes (off/shadow/warn/strip/block) registered in PostToolUse Task. Three new LLM judges added to Recommendation Tribunal (SPEC-125): sycophancy-judge (warn default), concession-judge (shadow), repetition-truth-judge (shadow). aggregate.sh extended fail-soft for the 3 optional SPEC-192 judges. New skill .opencode/skills/epistemic-humility/SKILL.md with replacement protocol. Updated radical-honesty.md with Enforcement section linking the three layers. Updated savia-ethical-principles.md §5 and §10 references. Telemetry to output/anti-adulation-telemetry.jsonl. 51 pytest + 17 bats green covering 20 positive examples, 20 legitimate courtesy negatives, all hook modes, JSON envelope handling. Spec PROPOSED in docs/propuestas/SPEC-192-anti-adulation-illusory-truth.md, P0, 18 ACs.
scripts/criterion-simulation/operator-state-signals.py: Computes operator-state signals (fatigue_score, pressure_score, override_rate, time_band) from local data only. Zero network calls. Reads deadline_proximity from ~/.savia/preferences.yaml if present.
scripts/criterion-simulation/trigger-evaluator.py: Evaluates task context to decide if criterion-simulation should activate. Threshold: SAVIA_CS_TRIGGER_THRESHOLD (default 50). High-impact signals: touches_production (+25), touches_security (+30), touches_human_safety (+50), estimated_hours>16 (+15).
scripts/criterion-simulation/historical-priors.py: Searches KG (SQLite .savia-kg/graph.db) for similar reverted tasks within lookback window. Graceful: returns {count:0} if KG absent.
scripts/criterion-simulation/reaffirmation-log.py: Records conscious human reaffirmations (reaffirm) and problem reformulations (reframe). Enforces reason >= 20 chars for reaffirm (exit 2 if violated).
scripts/kg-schema-migrate-cs.py: Idempotent SQLite migration adding frame_reaffirmations table (task_id, ts, operator, reason, verdict_before, tags). Separate from SPEC-193 migration.
.opencode/agents/criterion-simulation-judge.md: Heavy-model LLM judge. Executes 4 meta-questions (frame challenge, historical priors, operator state, alternative reframing). Emits verdict FRAME_OK|FRAME_DOUBT|FRAME_REJECT + banner_text + confidence. Always includes is_simulation_disclaimer field.
.opencode/skills/meta-reflection/SKILL.md: Protocol for the 4 meta-questions. Includes replacement protocol (reaffirm/reframe commands) and declared limitations.
.opencode/skills/meta-reflection/DOMAIN.md: Conceptual knowledge: difference between evaluation (correct application of given criterion) and reflection (questioning the criterion itself). Cites origin quote 2026-06-13.
.opencode/hooks/criterion-simulation-challenge.sh: Pre-task hook. Master switch SAVIA_CRITERION_SIMULATION=off (default; opt-in). Modes: shadow (telemetry only), advise (banner, no block), interrupt (banner + reaffirmation_required log). Exit 0 always. Never blocks.
docs/rules/domain/criterion-simulation-honesty.md: Declares explicitly that this layer is NOT real criterion — it is a declared heuristic simulation of pause. Cites spec quote verbatim.
output/criterion-simulation/ directory: Telemetry events.jsonl + reaffirmations.jsonl.
tests/scripts/test_criterion_simulation_trigger.py: 50 synthetic normal tasks (AC3: trigger rate <=20%), 10 high-impact tasks (AC4: trigger rate >=8/10), socket mock test (AC5: no network calls).
tests/scripts/test_criterion_simulation_misc.py: reaffirmation-log validation (AC7), kg-schema-migrate-cs idempotency (AC14).
tests/test-criterion-simulation-hook.bats: Hook tests covering AC1 (disclaimer), AC2 (exit 0 always), AC9 (JSONL telemetry), AC10 (shadow no stderr), AC11 (advise banner), AC12 (interrupt reaffirmation_required), AC13 (master switch off).
scripts/embeddings-cache.py — Embedding cache backed by sentence-transformers (all-MiniLM-L6-v2, 384 dims). Caches per-text to ~/.savia/embeddings-cache/<sha256>.npy. In-process LRU + disk persistence. Fallback to zero-vector if model unavailable (fail-soft). CLI: python3 scripts/embeddings-cache.py --text "..." --json.
scripts/kg-schema-migrate-tribunal.py — Idempotent SQLite schema migration. Creates tribunal_iterations table with privacy column + two indexes (draft_hash, session_id) in ~/.savia/tribunal-iterations.db. CLI: python3 scripts/kg-schema-migrate-tribunal.py [--db PATH] --json.
scripts/recommendation-tribunal/historical-context.py — Main SPEC-199 engine. Computes embedding of current draft, searches DB for top-K similar historical drafts (cosine similarity >= threshold), builds a token-capped context_text block for orchestrator injection. Persists current draft (skips entries flagged as private). Returns is_zero_sc=true when DB is empty (first round analog to DiffusionGemma's zero-signal first step). CLI: python3 historical-context.py --draft TEXT --top-k 3 --similarity-threshold 0.6 --session-id ID --iteration N.
scripts/recommendation-tribunal/iterate.sh — Extended evaluate-stop handler with SPEC-199 block. When SAVIA_TRIBUNAL_HIST_CONTEXT=on, calls historical-context.py before forwarding to early_stop.py. Result exported as TRIBUNAL_HISTORICAL_CONTEXT env var for orchestrator injection into judge prompts. Default remains off (opt-in pilot).
tests/scripts/test_historical_context.py — 12 pytest cases covering all ACs: top-k 0, empty DB, 1 entry match, 5 entries top-3 sorted, threshold enforcement, deterministic embeddings, cache hit, idempotent migration, private entry exclusion, token cap, is_zero_sc flag, 1000-entry latency <= 200ms. All 12 passing.
tests/test-historical-context-tribunal.bats — 6 bats tests: CLI runs, output is valid JSON, top-k 0 returns empty, is_zero_sc present, iterate.sh unaffected when feature off, schema migration CLI works. All 6 passing.
Spec: docs/propuestas/SPEC-199-historical-context-tribunal-rounds.md
Blocked by: SPEC-195 (iterative tribunal), SPEC-189 (embeddings infra)
SPEC-193 (Context Provenance & Injection Hardening) and SPEC-194 (Criterion Simulation Layer) PROPOSED. Two complementary specs addressing structural limits of agentic systems with explicit honesty about what they cannot do. SPEC-193 closes 6 of 7 documented prompt-injection vectors via 12 components in 3 layers (deterministic sanitize, LLM judges, temporal correlation). SPEC-194 adds meta-reflection simulation BEFORE applying ideas — 4 questions (frame challenge, historical priors, operator state, alternative reframing) that interrupt visually but never block. Both spec files include disclaimer that the layer is heuristic simulation, not real criterion. Origin: 2026-06-13 jailbreak techniques study + user reflection on review vs reflection. ROADMAP backlog reordered: 8 items, 3 in PR review + 1 P0 new + 4 P2/P3.
Six specs PROPOSED extracted from analysis of google-deepmind/gemma/diffusion (text diffusion sampling). SPEC-195 iterative tribunal with multi-criteria early stop (P1, 3-4d). SPEC-196 freeze-done elements early-cancel jueces tras VETO (P1, 1-2d). SPEC-197 annealing temperature schedule en jueces meta-reflexivos (P2, 2-3d). SPEC-198 JudgeVerdict frozen dataclass contract (P2, 3-4d, refactor). SPEC-199 historical context conditioning entre rondas (P2, 4-5d, depends SPEC-195, honest naming: NOT self-conditioning literal because LLM via API). SPEC-200 adaptive quality gate threshold proporcional a la distribucion en lugar de fijo 80 (P3, 2-3d). Patrones: ChainedEarlyStop, _WhileLoopCarry.done, AnnealingTemperatureShaper, @flax.struct.dataclass, SelfConditioning adapted, entropy_bound proportional. ROADMAP backlog actualizado: 12 items priorizados. Total estimacion 8-21 dias humano / 14-26h agente.
SPEC-SE-020 Slice 1: deps-validate.sh schema validator + portfolio-as-graph rule + sample deps.yaml + 31 BATS tests
SPEC-SE-020 Slice 2: portfolio-graph.sh (ASCII/Mermaid/JSON) + 29 BATS tests
SPEC-SE-020 Slice 3: portfolio-critical-path.sh (topological + slack + bottleneck detection) + 24 BATS tests

Changed

pr-plan-gates.sh G5: hi regex expanded to catch all .claude/ paths + .scm/ + .opencode/
scripts/push-pr.sh now creates PRs as draft by default (was: ready-for-review). Use --no-draft to opt out. Backward-compat: --draft flag preserved as no-op.
Title auto-selection now prefers the first feat:/fix: commit in chronological order on the branch, falling back to the previous behavior (first non-chore/non-Merge in descending log) only if no conventional-commit feat/fix is found.
Draft default: matches autonomous-safety pattern — PRs created via /pr-plan should require explicit human action to become reviewable. Avoids accidental ready-for-review on autonomous flows.
Title selection: previous logic picked the latest commit (e.g. drift fixes like fix(ci): ...), masking the actual feature intent. Real example: PR #3 fix(tool-healing): ... was titled with the trailing fix(ci) drift commit and required manual PATCH via REST API to correct.
Existing --draft callers: unchanged behavior.
Callers that relied on default ready-for-review: must add --no-draft.
scripts/failure-pattern-memory.sh:37 (compute_pattern_id) migrated to content-fingerprint skill
scripts/test-auditor.sh:49 (audit hash) migrated to content-fingerprint skill
scripts/ado-bridge.sh:36 documented as NOT migrated (cksum fallback retained)
scripts/semantic-map.sh:78 documented as NOT migrated (shasum/nohash00 chain retained)
.claude/settings.json: registers contract-test-guard hook on PreToolUse Edit|Write
Hook reduced from 113 to 69 LOC (under hard limit 80)
_SAVIA_INTERNAL_TEST_BRANCH guard requires BATS_TEST_FILENAME or CI or SAVIA_TEST_MODE
[contract-change|add|remove] commit message bypass implemented and tested
substring match replaced by exact + REPO_ROOT-normalized suffix match
docs/propuestas/SE-216-evo-patterns.md — Added resource: "https://github.com/evo-hq/evo".
docs/propuestas/SE-217-autoresearch-patterns.md — Added resource: "https://github.com/karpathy/autoresearch".
docs/propuestas/SE-218-codebase-memory-patterns.md — Added resource: "https://github.com/DeusData/codebase-memory-mcp".
docs/propuestas/SE-219-abtop-patterns.md — Added resource: "https://github.com/graykode/abtop".
docs/ROADMAP.md — Added Era 208 OKF Adoptable Patterns section + repriorized backlog 2026-06-20 with SE-222 S0-S3 entries.
docs/propuestas/SE-222-okf-adoptable-patterns.md — Spec proposing OKF v0.1 adoptable patterns (resource: URI, log.md convention, index.md auto-gen). Discarded: portability inter-org, OKF as export bundle, external SDK. S0 implemented in this PR; S1+S2 in follow-up PRs; S3 (back-fill 20 specs) is P3.
bats tests/test-spec-validator.bats: full suite passing
bash scripts/test-auditor.sh tests/test-spec-validator.bats: certified
bash scripts/spec-validator.sh --batch docs/propuestas/: backfill reduces total findings; full back-fill deferred to S3
Spec: docs/propuestas/SE-222-okf-adoptable-patterns.md
Origin: análisis comparativo OKF Google Cloud 2026-06-20 vs modelo de cúpulas Savia
Resource: https://github.com/GoogleCloudPlatform/knowledge-catalog
.claude/settings.json — Register post-spec-edit-reindex.sh as PostToolUse hook on Edit|Write (matcher Edit|Write, async, timeout 5s).
CLAUDE.md — Counter update: hooks 81→83, regs 84→86 (post-write-validate + dual-estimation-gate + new reindex hook).
bats tests/test-propuestas-index-gen.bats: full suite passing
bats tests/hooks/test-post-spec-edit-reindex.bats: full suite passing
Both certified by test-auditor (>=80)
Spec: docs/propuestas/SE-222-okf-adoptable-patterns.md (Slice S2)
Previous slices: PR #850 (S0 resource: URI), PR #851 (S1 LOG.md)
Block closure: Era 208 OKF Adoptable Patterns
docs/rules/domain/hook-event-equivalence.md: added archived: true frontmatter (SE-096)
docs/rules/domain/image-relevance-filter.md: added archived: true frontmatter (SE-096)
docs/rules/domain/portfolio-as-graph.en.md: added archived: true frontmatter (SE-096)
docs/rules/domain/receipts-protocol.en.md: added archived: true frontmatter (SE-096)
docs/rules/domain/savia-memory-architecture.md: added archived: true frontmatter (SE-096)
docs/rules/domain/session-state-location.md: added archived: true frontmatter (SE-096)
docs/rules/domain/slm-consolidation-pattern.md: added archived: true frontmatter (SE-096)
docs/rules/domain/slm-training-pipeline.en.md: added archived: true frontmatter (SE-096)
docs/rules/domain/vault-frontmatter.md: added archived: true frontmatter (SE-096)
scripts/rule-orphan-detector.sh: skip rules with archived: true in frontmatter
docs/agent-notes-protocol.md: Added "Message Schema (SPEC-056)" section with CLI reference,
25/25 passing — validate accepts valid, rejects no-id, rejects empty blocks,
.opencode/plugins/savia-foundation.ts: registers sycophancyGuard in AFTER_GUARDS chain (SPEC-150 Slice 2).
docs/rules/domain/hook-multihandler-migration.md: updated to status: implemented, slice_final: 2. Documents 2026-06-24 decision: FP rate = 0.00, ROI below threshold for Slices 3-6. Slices 3-6 descoped with reopen criterion (FP >= 2%).
docs/propuestas/SPEC-150-hooks-multi-handler-migration.md: status flipped IN_PROGRESS → IMPLEMENTED with reduced-scope note.
web-e2e-tester: adopt Webwright workspace-as-state patterns — Write→Execute→Inspect→Repair loop + self-reflection gate, zero new dependencies
skills: fix Description Trap — rewrite 98 skill descriptions and 65 agent descriptions as trigger-only (Superpowers pattern SE-145)
skills: add SUBAGENT-STOP gate to 8 high-criticality skills (adversarial-security, code-improvement-loop, consensus-validation, dag-scheduling, overnight-sprint, spec-driven-development, tdd-vertical-slices, verification-lattice) — prevents runaway orchestration in nested agent contexts (SE-146)
.ci-baseline/hook-untested-count.count: 3 → 0. Hook coverage 55/58 (94.8%) → 58/58 (100%).
18/58 → 58/58 hooks tested (31% → 100%, +40 hooks en 13 batches)
Average score certified ≈ 90 (todos ≥80, mayoría ≥90)
1100+ tests añadidos
Drift hooks 0 (CI-enforced via ratchet)
docs/propuestas/SPEC-055-test-auditor.md: status PROPOSED → IMPLEMENTED. Resolution section con verificación de los 4 scripts deliverables + auditor self-test (15 tests, score 83). Acceptance criteria final: 5/5 cumplidos.
docs/ROADMAP.md: Era 186 extension marcada CLOSED 2026-04-25. Tabla milestones extendida con batches 49-51 (51/58 → 58/58). Header bumped v6.2.0 → v6.3.0 con SPEC-055 IMPLEMENTED y baseline tighten anotado.
.ci-baseline/hook-critical-violations.count: 5 → 4 (current measurement = 4 críticos consistentes en últimos 5 hook-bench runs). Ratchet never-loosen mantenido.
.claude/agents/sdd-spec-writer.md — sección "Handoff Format (SPEC-121)" añadida: E1→E2 handoff a dotnet-developer tras spec approval.
.claude/agents/dotnet-developer.md — sección "Handoff Format (SPEC-121)" añadida: E2→E3 handoff a code-reviewer tras implementation + tests green.
.claude/agents/code-reviewer.md — sección "Handoff Format (SPEC-121)" añadida: E3→E4 a test-engineer (PASS) o developer (REJECT con termination_reason=unrecoverable_error).
.claude/agents/test-engineer.md — sección "Handoff Format (SPEC-121)" añadida: E4 completion a court-orchestrator.
.claude/agents/court-orchestrator.md — sección "Handoff Format (SPEC-121)" añadida: REJECT routing back a dotnet-developer.
docs/agent-notes-protocol.md — sección "Cuándo usar agent-notes vs handoff-as-function" añadida con tabla de decisión y regla práctica ("si cabe en 7 campos, usa handoff-as-function").
docs/propuestas/SPEC-121-handoff-convention.md — status PROPOSED → IMPLEMENTED. Resolution section con 6/6 ACs cumplidos.
.claude/settings.json — registra emergency-mode-readiness.sh en SessionStart con timeout 12s y statusMessage descriptivo.
docs/rules/domain/autonomous-safety.md — sección "Emergency-mode (LocalAI fallback) — SPEC-122" añadida (AC-05). Prohibiciones explícitas: NUNCA bypass AUTONOMOUS_REVIEWER, ramas agent/*, PR Draft en emergency-mode. Emergency-mode cambia SOLO el endpoint de inferencia.
docs/propuestas/SPEC-122-localai-emergency-hardening.md — status PROPOSED → IMPLEMENTED. Resolution section con 7/7 ACs cumplidos.
docs/propuestas/SPEC-078-dual-estimation-agent-human.md — status PROPOSED → IMPLEMENTED. Resolution section con verificación de los 4 deliverables Fase 1 MVP: dual-estimate.sh (engine CLI), test-dual-estimate.bats (score 82), dual-estimation-gate.sh (PostToolUse warning hook), docs/politica-estimacion.md (política dual). 5/5 AC Fase 1 cumplidos. Fases 2-4 (auto-estim, tracking, capacity dual) quedan como evolución gradual, no bloquean IMPLEMENTED status.
docs/propuestas/SPEC-124-pr-agent-wrapper.md — status PROPOSED → IMPLEMENTED. Resolution section con 9/9 ACs cumplidos.
SPEC-055 test-auditor (drift fix)
SPEC-078 dual-estimation (drift fix)
SPEC-121 handoff-as-function (3 ACs cerrados)
SPEC-122 LocalAI emergency (4 ACs cerrados)
SPEC-124 pr-agent wrapper (3 ACs cerrados)
SE-070 opus47 calibration (Slice 1-3, Slice 4 deferred per spec)
.claude/settings.json — registra memory-verified-gate.sh en PreToolUse Edit|Write con timeout 5s.
tests/test-memory-store.bats — 9 tests SE-072 nuevos (rechaza sin source, blacklist, format inválido, source válido tool/file/user, JSONL embed, escape hatch). 3 tests existentes actualizados con --source tool:Bats. Score 90.
docs/propuestas/SE-072-verified-memory-axiom.md — APPROVED → IMPLEMENTED. Resolution con 5/5 ACs.
CLAUDE.md — hooks 59→60 (64 regs).
scripts/pr-plan.sh — invoca G11 tras G10.
scripts/push-pr.sh — lee .pr-summary.md y lo prepend al PR body antes de la sección Summary auto-generada.
.gitignore — excluye .pr-summary.md (vive solo local).
CLAUDE.md — referencia lazy a la nueva regla en la tabla.
docs/ROADMAP.md header bump v6.8.0 → v6.10.0 con SE-074 APPROVED listado. Backlog APPROVED reordenado para incluir SE-074 entre los sin-GPU.
docs/ROADMAP.md Era 188 pipeline ordenado: SE-073 → SE-074 → SE-075 → SE-076. Sinergias documentadas (task_queue habilita orquestador paralelo y healer async). Backlog APPROVED ahora 5 sin-GPU + 4 GPU-blocked.
scripts/pr-plan.sh — invoca G12 tras G11.
docs/propuestas/SE-074-parallel-spec-execution.md — añade sección OpenCode Implementation Plan (PURE_BASH) + nuevo Slice 1.5 (S, 3h) "Adaptive halting + dynamic retry budget" inspirado en Kohli et al. 2026 (arXiv:2604.07822). Doble criterio halting (convergencia + confianza) + Poisson-clipped budget según effort field.
docs/ROADMAP.md Era 189 inaugurada con SE-077 + SE-078 priority alta.
CLAUDE.md — referencia lazy a la regla nueva.
docs/memory-system.md — documenta hard-cap 30 entries Tier A + algoritmo de score (access_count + recency_bonus + pin_bonus + identity_bonus). 200-line ceiling histórico se mantiene como límite absoluto.
user_* files (identidad foundational) NUNCA caen a Tier B
pin: true files NUNCA caen a Tier B
Recencia (<30d) actúa como tiebreaker positivo
Default cap configurable via MEMORY_TIER_A_CAP env (default 30)
scripts/parallel-specs-orchestrator.sh — el header del plan incluye una línea run timestamp : <RUN_TS> para correlacionar el output con los worktrees creados en disco (.claude/worktrees/spec-<id>-<RUN_TS>/).
tests/structure/test-parallel-specs-orchestrator.bats — 2 regression tests añadidos:
28/28 pasan. 18/18 adaptive-halting + 33/33 spec-budget también pasan.
docs/rules/domain/opencode-savia-bridge.md — nueva regla canonical OpenCode↔Savia bridge (95 líneas).
docs/propuestas/SPEC-114-docs-savia-alignment.md — flipped a status: SUPERSEDED, superseded_by: SE-078, banner añadido.
docs/propuestas/SE-077-... y docs/propuestas/SE-078-... — frontmatter a status: IMPLEMENTED, ACs marcados [x] salvo AC-03/AC-05 (E2E E2E pending boot por la usuaria) y AC-11 (eliminación wrappers tras 1 sprint).
5 ficheros nuevos, 80 tests, todos certified ≥80:
Sin regresiones en orchestrator/merge-queue/G13/attention-anchor/db-sandbox/cleanup-stale.
Plugin TS NUNCA hace git push, gh pr merge, git push --force
Installer NUNCA modifica el repo (sólo ~/.savia/)
Parity audit + canary son read-only por defecto (--baseline y --check requieren flag explícito)
B8 ATTENTION ANCHOR: plugin re-inyecta SPEC_WORKER_ID y Spec ref: en cada hook payload, igual que Claude Code
A9 SUPERVISED EXECUTION: permission.ask enforce AUTONOMOUS_REVIEWER deterministically antes de pedir al humano
docs/ROADMAP.md — header línea 3 actualizada con Era 190 APPROVED. Critical Path Q2-Q3 reprioritizado: 12 items unificados por relevancia (compliance enterprise + skill catalog quality multipliers + sovereignty residual + apalancamiento residual). Sinergias documentadas extendidas con cross-refs SE-082/084/086/087 y SPEC-SE-036/037.
mattpocock/skills (MIT, jamiepine/voicebox-style clean-room): patterns extraídos para SE-081 (caveman/zoom-out/grill-me skills), SE-082 (LANGUAGE.md vocabulary), SE-083 (tdd/SKILL.md anti-pattern), SE-084 (write-a-skill meta-discipline), SE-085 (write-a-skill skill), SE-086 (ubiquitous-language + domain-model skills), SE-087 (design-an-interface skill). Zero código copiado; cada spec cita upstream commit hash en header.
15 specs docs/propuestas/SPEC-003..017: migradas a frontmatter YAML via la herramienta. Source of truth body-prose preservada. status: canónico + origin_date: + migrated_at: + migrated_from: body-prose (auditable).
50 specs migrados a frontmatter YAML (batch 2) vía `scripts/spec-frontmatter-migrate.sh --apply --limit 50`. Conjunto: SPEC-018..SPEC-118 + savia-enterprise/SPEC-SE-028/029/030. Mapping mecánico body-prose → YAML canónico.
Resultado: `spec-status-normalize.sh --audit` reporta missing 96 → 46 (reducción 50 specs). Acumulado batch 1+2 = 65 specs con frontmatter canónico, 75% del gap original cerrado.
docs/propuestas/ROADMAP.md: Tier 4.10 añadido (SE-040 degradation canary). Priorizable alto por evitar cascadas silenciosas cuando proveedor degrada modelo.
docs/propuestas/ROADMAP.md: Tier 2.5 añadido (SE-041 memvid probe).
SPEC-081 hook-bats-coverage: status Proposed → SUPERSEDED. Superseded by SE-037 (que absorbe BATS coverage + añade latency SLA + ratchet enforcement). Frontmatter extendido con superseded_by, superseded_at, superseded_reason.
SPEC-028 search-reranker: status ACCEPTED → SUPERSEDED. Superseded by SE-032 (versión Spec Ops del mismo concepto con Feasibility Probe bloqueante + cross-encoder específico BAAI/bge-reranker + ablation report).
SPEC-027 graph-memory-layer: status UNLABELED → Proposed + converges_with: SPEC-123. Documenta convergencia con SPEC-123 graphiti temporal pattern y SE-030 GraphRAG quality gates.
ROADMAP Tier 6: convergencias marcadas como aplicadas (3 de 4). Sólo pendiente SPEC-078 ↔ SPEC-SE-013 dual estimation. Pattern documentado: status: SUPERSEDED + superseded_by en frontmatter.
Zero vendor lock-in: 274 agent/command model declarations migrated from vendor-specific names (claude-opus/sonnet/haiku, opus/sonnet/haiku) to abstract capability tiers (heavy/mid/fast). Model resolution delegated to ~/.savia/preferences.yaml per-user, per-provider.
SPEC-127 Slice 1: New scripts/savia-env.sh provider-agnostic environment loader with fallback chain for SAVIA_WORKSPACE_DIR and SAVIA_PROVIDER. Supports Claude Code, OpenCode, Copilot, and LocalAI via capability probes.
38 hooks now source savia-env.sh and export CLAUDE_PROJECT_DIR from SAVIA_WORKSPACE_DIR, fixing silent breakage under OpenCode where CLAUDE_PROJECT_DIR is unset.
4 missing hook timeouts added to settings.json (validate-bash-global, plan-gate, prompt-injection-guard, delegation-guard).
Docs: model-alias-schema.md and model-alias-table.md rewritten for tier-based resolution. AGENTS.md and agents-catalog.md auto-regenerated.
Model names migrated from hardcoded IDs (claude-opus-4-7, claude-sonnet-4-6) to provider-agnostic tiers (Heavy, Mid, Fast) resolved at runtime via scripts/savia-env.sh and ~/.savia/preferences.yaml.
docs/AGENTS.md: all agent tables and decision tree updated to tier-based labels.
.opencode/CLAUDE.md: model constants replaced with runtime resolution note.
opencode.json: agent definitions and slash commands populated with provider-agnostic config.
scripts/advisor-config.sh: advisor/executor defaults sourced from preferences.
.claude/commands/agent-run.md, .claude/commands/pr-plan.md: model references updated.
tests/test-advisor-config.bats: added SAVIA_MODEL_HEAVY/MID/FAST env vars in setup() to simulate configured preferences.yaml — tests adapt to provider-agnostic tier resolution.
SKILLS.md: regenerated to include 2 new skills (92 total).
.claude/commands/pr-plan.md: gate count updated 11 → 15 (G0-G14).
CLAUDE.md: hooks count corregido → "65 hooks (61 registrados, 4 huérfanos)".
.claude/hooks/README.md: catálogo completo de 65 hooks con categorías reales, documentación de arquitectura dual (OpenCode nativo + bridge), y compliance provider-agnostic.
memory-cache-rebuild.sh: escanea .savia-memory/ + legacy .claude/projects/ (provider-agnostic).
memory-stack-load.sh: MEMORY_BASE → .savia-memory/ con fallback a legacy.
stop-memory-extract.sh: escribe snapshots en .savia-memory/sessions/YYYY-MM-DD/.
session-end-memory.sh: escribe session-hot.md en .savia-memory/sessions/YYYY-MM-DD/.
savia-memory-bootstrap.sh: marker en .savia/external-memory-target en vez de symlink .claude/external-memory.
memory-agent.md: paths actualizados a estructura canónica .savia-memory/.
savia-memory SKILL.md: documentados search y recall (ambos válidos), añadido memory-index-rebuild.sh.

Fixed

court-review.md + trace-optimize.md: YAML frontmatter added (SCM coverage 534/534)
agents-opencode-convert.sh: extended to sync subdirectories (decision-trees/ → .opencode/)
generate-capability-map.py: --check flag (read-only, tempdir, exit 0=FRESH 1=STALE 2=MISSING)
`.github/workflows/auto-rebase-open-prs.yml`: nuevo workflow que rebased automáticamente PRs abiertos tras cada push a main. Resuelve el problema raíz: `merge.ours.driver` en .gitattributes solo funciona local, no en GitHub server-side merge. El workflow ejecuta `resolve-pr-conflicts.sh` desde runner GHA (donde sí puede configurar el driver). Elimina la fricción "mergear un PR genera conflicto en todos los demás" — ahora la usuaria mergea y el auto-rebase resuelve los PRs en cola sin intervención.
Añade -a a grep -oE '!projects/...' para forzar text mode al leer .gitignore. El fichero contiene caracteres Unicode box-drawing (U+2500 ─) en headers de comentarios, lo que hacía que GNU grep clasificara el archivo como binary y silenciosamente descartara las matches (sin error, exit 0, stdout vacío). Resultado: PUBLIC_PROJS quedaba vacío, los proyectos whitelisted se filtraban a la blocklist como falsos positivos, y cualquier PR externo (especialmente forks) que mencionara savia-monitor / savia-web / savia-mobile-android en CHANGELOGs o commits fallaba el PII & Confidentiality Scan con BLOCKED: Blocklist terms found.
Comment inline explicando el porqué del -a para prevenir regresión por re-formato del script.
tests/structure/test-generate-blocklist.bats — 12 tests cubriendo:
scripts/test-auditor-sweep.sh: bug aislado donde sweep extraía campo .score del auditor JSON pero el auditor emite total. Resultado: sweep siempre reportaba 0% compliance vs 100% real. Impact LOW (script no usado en CI, solo en su propio test). Fix de 1 línea: .score → .total. Verificación post-fix: 329/329 BATS files compliant ≥80 = 100% baseline.
Duración: 2026-04-21 → 2026-04-25 (5 días)
Batches: 39-51 (13 iteraciones)
Hooks tested: 18/58 (31%) → 58/58 (100%) — +40 hooks, +69 puntos
Tests añadidos: 1100+ certified con score ≥80
Avg score: ~90 (después de internalizar feedback_test_auditor_score_targets)
Bugs reales descubiertos: 4 (cwd-changed-hook, emotional-regulation-monitor, memory-auto-capture, SE-071 block-branch-switch-dirty)
Drift hooks: 0 (CI-enforced via .ci-baseline/hook-untested-count.count)
PR-B: SPEC-121 handoff convention completion (3 ACs faltantes)
PR-C: SPEC-122 LocalAI emergency hardening completion (4 ACs faltantes)
Backlog: SPEC-124 pr-agent wrapper (5 ACs ~3-4h)
scripts/parallel-specs-orchestrator.sh — --queue ya no deja whitespace residual al procesar líneas con comments inline. El trim anterior (${line## }) sólo borraba un espacio, dejando trailing whitespace que rompía el lookup posterior de specs (find con patrón "SE-140 *.md" retornaba vacío). Reemplazado por regex greedy ^[[:space:]]*(.*[^[:space:]])[[:space:]]*$.
scripts/parallel-specs-orchestrator.sh — effort tier en minúsculas (s/m/l) en frontmatter de spec ya se reconoce. El grep -oE '^[SML]' era case-sensitive y caía al default M, asignando budget medio a specs grandes con lowercase.
scripts/adaptive-halting.sh — exclusión de .halt-state.prev.json añadida al find que computa el tree-hash. Sin ella, el snapshot anterior contaminaba el hash actual y el criterio de convergencia nunca se estabilizaba en árboles con halting iterativo.
tests/structure/test-parallel-specs-orchestrator.bats — 3 regression tests (queue trim, lowercase tier, mixed-case normalisation). 26/26 pasan.
tests/structure/test-adaptive-halting.bats — 18/18 siguen pasando con la nueva exclusión.
scripts/parallel-specs-orchestrator.sh — el bloque Plan per spec mostraba puertos que no se acababan usando. La causa: la fase de plan y spawn_worker llamaban a date(1) por separado y con formatos distintos (%Y%m%d%H%M%S vs %Y%m%d-%H%M%S UTC), de modo que allocate_ports hashing producía dos asignaciones diferentes. Ahora se calcula un único RUN_TS al inicio del script y SPEC_WORKTREE_NAMES + SPEC_PORTS se rellenan una vez en la fase de validación; tanto el plan como el spawn leen de esos maps. Plan output ahora coincide con la realidad (verificado con regression test #27).
docs/propuestas/SPEC-103-deterministic-first-digests.md — PROPOSED → IN_PROGRESS (Slice 1 IMPLEMENTED batch 82, Slice 2 pendiente)
docs/propuestas/SPEC-125-recommendation-tribunal-realtime.md — PROPOSED → IN_PROGRESS (Slice 1 foundation IMPLEMENTED batch 83, NO ACTIVADO; Slices 2+3 follow-up)
docs/propuestas/savia-enterprise/SPEC-SE-035-reconciliation-delta-engine.md — PROPOSED → IN_PROGRESS (Slices 1+3 IMPLEMENTED batch 81, Slices 2/4/5 follow-up)
docs/propuestas/savia-enterprise/SPEC-SE-036-api-key-jwt-mint.md — PROPOSED → IN_PROGRESS (Slices 1+2 IMPLEMENTED batches 79+80, Slice 3 sunset PAT pendiente)
docs/propuestas/savia-enterprise/SPEC-SE-037-audit-jsonb-trigger.md — PROPOSED → IMPLEMENTED (8/12 ACs cumplidos batch 78, 4 deferred a repo Enterprise post-deploy)
docs/propuestas/SPEC-106-truth-tribunal-report-reliability.md — ALL (inválido) → IMPLEMENTED (Truth Tribunal completo en producción: orchestrator + 7 jueces + scripts + tests)
docs/propuestas/SPEC-110-memoria-externa-canonica.md — DRAFT → IMPLEMENTED (cargado como @import en CLAUDE.md desde Era anterior)
docs/propuestas/SPEC-047-requirement-pushback.md — UNLABELED → APPROVED (body declara "Status: APPROVED — Phase 1")
docs/propuestas/SPEC-025-chinese-compatibility.md — UNLABELED → PROPOSED (research study, no implementación)
docs/propuestas/SPEC-079-legal-compliance-agent.md — UNLABELED → PROPOSED (body declara "Estado: Propuesta")
docs/propuestas/ROADMAP.md — last_updated: 2026-04-18 → 2026-04-30, expires: 2026-06-18 → 2026-06-30. Sección "## 6. Live status" reescrita para reflejar realidad post-batch-83: último PR mergeado, recientes (Era 232 cerrada + Recommendation Tribunal foundation + image relevance filter + hotfixes), top 10 Critical Path repriorizado tras audit, próximo slice = SPEC-107 Slice 1 (medición AI Cognitive Debt).
tests/structure/test-spec-status-frontmatter.bats — 28 tests certified. Cubre:
docs/propuestas/savia-enterprise/SPEC-SE-001-foundations.md — PROPOSED → IMPLEMENTED. Todos los 6 AC verificados en disco con evidencia.
`.claude/hooks/scope-guard.sh`: optimizado de 203ms → 48ms. Restringe find a 3 paths conocidos (projects/, docs/specs/, docs/propuestas/) + maxdepth 6 + prune de node_modules/.git/build/dist/target. Elimina el timeout SLA 200ms que bloqueaba PRs en CI Hook Latency Gate.
`.gitattributes`: añade `CHANGELOG.md merge=union` (built-in, sin config necesaria) y `.scm/* merge=ours` (requiere driver). Eliminates 3 de 4 tipos de conflictos recurrentes sin necesidad de tooling externo.
`scripts/setup-merge-drivers.sh`: configura `merge.ours.driver = true` local. Sin esto, los `merge=ours` en .gitattributes eran silent no-ops — razón por la que los conflictos persistían pese a tener el attribute.
`.claude/hooks/session-init.sh`: invoca setup-merge-drivers idempotentemente al arrancar sesión. Zero-config para usuarios — drivers quedan wired automáticamente en cualquier checkout fresco.
`tests/test-merge-drivers.bats`: 23 tests incluyendo integración real de merge en sandbox git (union, ours, sin driver = conflict). Auditor score 82.
docs/rules/domain/receipts-protocol.en.md: trimmed from 166 → 119 lines to comply with the 150-line rule (tests/structure/test-workspace-structure.bats:95).
docs/rules/domain/slm-training-pipeline.en.md: trimmed from 157 → 111 lines (same rule).
SE-094 closed via allowlist mechanism (.claude/hooks-allowlist.tsv): last orphan hook (SPEC-125 WIRE-READY) now classified as deliberate non-registration with mandatory spec/rule citation.
New BATS test tests/structure/test-hooks-integrity-allowlist.bats enforces the allowlist contract — 5/5 OK.
opencode.json: remove unknown schema fields (_hooks_doc, _comment, websearch, subtask) that caused startup failure
savia-bridge.py: probe nvm versioned bin paths in find_claude_cli() so bridge resolves claude binary under nvm
push-pr.sh: title extraction changed from tail -1 to head -1 with sign/merge commit filtering — PR titles now use the most recent meaningful commit instead of the oldest sign commit.
push-pr.sh: sign commit message no longer hardcodes Co-Authored-By: Claude Opus 4.7 — PV-06 compliant.
push-pr.sh: removed Generated with Claude Code footer from PR body — PV-06 compliant.
pr-plan.sh: removed duplicate sign step — eliminates 2-3 sign commits per PR execution.
CLAUDE.md: skills count updated 90 → 92 (drift check).
Tool-healing guard no longer false-blocks read/write/edit on OpenCode v1.14 — extractFilePath/extractContent now recognize camelCase (filePath, newString) in addition to legacy snake_case (file_path, new_string). Resolves SPEC-TOOL-HEALING-FIX.
validate-bash-global.sh: $CLAUDE_PROJECT_DIR → fallback ${CLAUDE_PROJECT_DIR:-${OPENCODE_PROJECT_DIR:-$PWD}} — antes era el único hook BROKEN sin fallback provider-agnostic.
`scripts/resolve-pr-conflicts.sh`: bugfix — cuando `git merge origin/main` tenía éxito SIN conflictos pero CON cambios (merge commit nuevo), el script reportaba "already in sync" y exit 0 sin pushear. Ahora detecta via `rev-parse HEAD` pre/post si hubo merge real y pushea. Resuelve el caso en que los PRs quedaban CONFLICTING en GitHub pese a que `resolve-all-open-prs.sh` reportaba "resolved".
memory-store.sh: alias recall → search — el comando documentado en SKILL.md ya funciona.
auto/MEMORY.md ahora se puebla automáticamente con cada memory-store.sh save.
Índice canónico poblado con 6 entradas (antes vacío).
Fix path output/research-{tema} → output/research/{tema} en autonomous-safety.md, tech-research command y tech-research-agent SKILL.md
SPEC-144: bump CLAUDE.md command count 545→553 to reflect 8 new /speckit.* aliases (unblocks claude-md-drift-check.sh + readiness-stamp BATS tests). Confidentiality signature refreshed.
SPEC-155 [CRITICAL]: align plugin hook signature with OpenCode v1.14+ contract. Pre-fix, all 12 guards read input.args (undefined in real runtime) instead of output.args, operating on empty data = security theater. Symptom: spurious BLOCKED [tool-healing]: read called with empty file_path on valid tool calls. Helpers now prefer output.args with retro-compat fallback to input.args. 4 golden tests added with real v1.14 shape.
Terminología FOSS: reemplazado "robar/robado" por "copiar/adoptar/inspirado en" en 7 ocurrencias a lo largo de specs y CHANGELOG. En software libre, copiar patrones bajo licencias compatibles es el mecanismo legítimo del ecosistema — "robar" es terminología incorrecta que estigmatiza práctica legítima y aplica marco de propiedad intelectual propietaria a commons. Scope: SE-031, SE-042, ROADMAP-UNIFIED-20260418, SAVIA-SUPERPOWERS-ROADMAP, CHANGELOG entries 5.44 y 5.50.

Security

se-220: memory dedup by topic_key prevents AgentPoison-style memory poisoning (MEMORY.md 737→43 lines, 0 duplicates)
se-220: project-isolation-gate.sh promoted from WARN to BLOCK with SAVIA_ALLOW_CROSS_PROJECT=1 override + audit log
se-220: prompt-injection-guard.sh extended to scan untrusted .md/.txt/.html outside workspace (zero-trust spotlighting)
se-220: agent catalogs unified to .opencode/agents/ source-of-truth (72 agents, drift eliminated, skill confusion mitigated)
se-220: scripts/memory-canary-check.sh + canary token defense against memory poisoning (Chen et al. 2024)

6.14.3 — 2026-06-09

Fixed

opencode.json: remove unknown schema fields (_hooks_doc, _comment, websearch, subtask) that caused startup failure on strict schema validation
scripts/savia-bridge.py: probe nvm versioned bin paths in find_claude_cli() so bridge resolves the claude binary regardless of active nvm node version

6.14.2 — 2026-05-08

Repair CI pipeline: fix missing .opencode/ dirs in test setups, align test generators with .claude/ paths, sync agents converter output, update stale baseline, add CHANGELOG reference link

Fixed

scripts/push-pr.sh: PRs now created as draft by default (was: ready-for-review). Use --no-draft to opt out.
scripts/push-pr.sh: auto-title selection prefers first feat:/fix: commit chronologically, avoiding drift commits like fix(ci) being chosen as PR title.

6.14.1 — 2026-05-02

Era 64 — Provider-agnostic hardening: API endpoints, hooks audit, docs. Era 191 — Audit Remediation: OpenCode-SCM alignment (batch1 IMPLEMENTED).

Fixed

savia-watchdog.sh: API_URL uses SAVIA_API_UPSTREAM env var (was hardcoded api.anthropic.com).
savia-dual-proxy.py, savia-shield-proxy.py, sovereignty-switch.sh, setup-savia-dual.sh: same.
validate-bash-global.sh: project dir fallback (CLAUDE_PROJECT_DIR || OPENCODE_PROJECT_DIR || PWD).
CHANGELOG.md: missing 6.14.1 reference link added (CI check #6 fix).
court-review.md + trace-optimize.md: missing YAML frontmatter added (SCM coverage 534/534 commands).
agents-opencode-convert.sh: extended to sync subdirectories (decision-trees/ now replicated to .opencode/agents/).

Changed

SKILLS.md: header rewritten — removed Claude Code vendor references.
CLAUDE.md: hooks count corrected (65 hooks, 61 registered, 4 orphans).
.claude/hooks/README.md: full 65-hook catalog with OpenCode dual architecture docs.
init-pm.sh: Azure DevOps variables only loaded if ORG_URL is configured (non-placeholder). Core vars always load. Backends are optional/under-demand.
ROADMAP.md: v6.14.1 — Era 191 approved, pipeline reprioritized with agent time (not human hours). Total ~630 min agente (17 items).
generate-capability-map.py: added --check flag (read-only mode, tempdir, exit 0=FRESH 1=STALE 2=MISSING).

Added

docs/rules/domain/hook-event-equivalence.md: Claude Code ↔ OpenCode event mapping (17 events).
docs/rules/domain/provider-agnostic-tech-debt.md: tracks vendor references in 51/92 skills.
docs/specs/SPEC-OPC-AGENTSYNC.spec.md: replicate decision-trees/ to .opencode/agents/.
docs/specs/SPEC-SCM-COVERAGE.spec.md: fix frontmatter gaps in court-review + trace-optimize.
docs/specs/SPEC-SCM-FRESHCHECK.spec.md: fix --check mode in generate-capability-map.py.
docs/specs/SPEC-OPC-CROSS-AUDIT.spec.md: OpenCode ↔ Claude continuous audit script.

6.14.0 — 2026-04-26

Batch 63 — SE-074 IMPLEMENTED Slices 1 + 1.5 — paralelismo de specs + adaptive halting (Critical Path #2-3).

Added

scripts/parallel-specs-orchestrator.sh — Slice 1 core. N workers paralelos en worktrees aislados, bounded concurrency hard cap 5, port allocation, tmp dir sandboxing, runtime timeout, session.log per spec.
scripts/spec-budget.sh — Slice 1.5. Mapeo effort (S/M/L) → Poisson-clipped retry budget. Determinístico por defecto.
scripts/adaptive-halting.sh — Slice 1.5. Doble criterio halting (convergencia tree-hash + confidence ≥ floor + tests passed).
docs/rules/domain/parallel-spec-execution.md — regla canónica.
56 tests BATS (23+15+18) scores 88/85/81 — todos certified.

Changed

SE-074 status APPROVED → IMPLEMENTED (Slices 1+1.5). Slices 2+3 diferidas.

Context

Critical Path #2-3. Worker command configurable via SPEC_WORKER_CMD (claude/opencode/cualquier CLI). 3-4x throughput esperado en cualquier proyecto que invoque el orchestrator. Bounded concurrency hard cap 5 (default 3). Inspirado en Cole Medin LinkedIn 2026-04-25 + Kohli et al. 2026 (arXiv:2604.07822) para halting adaptativo y Poisson budget.

Version bump 6.13.0 → 6.14.0.

6.13.0 — 2026-04-26

Batch 62 — SE-073 Slice 1 IMPLEMENTED — MEMORY.md L1 hard-cap tiered (Critical Path #1).

Added

scripts/memory-tier-rotate.sh — 2-tier rotation (Tier A active ≤30 entries, Tier B filename-only archive).
scripts/memory-access.sh — increments access_count + updates last_access por memory file.
tests/structure/test-memory-tier-rotate.bats — 24 tests, score 83 certified.

Changed

docs/memory-system.md — documenta hard-cap 30 + algoritmo score (access_count + recency_bonus + pin_bonus + identity_bonus). 200-line histórico como ceiling absoluto.

Algorithm

Score = access_count + recency_bonus(<30d=+3) + pin_bonus(true=+999) + identity_bonus(user_*=+500). Tied scores → mtime desc.

Garantías: user_* y pin: true NUNCA caen a Tier B. Cap configurable via MEMORY_TIER_A_CAP (default 30).

Context

Critical Path #1 del roadmap reprio. Inspirado en GenericAgent L0_MetaRules pattern. Cada línea MEMORY.md cuesta tokens en CADA turn. Auto-trigger en Stop hook diferido a follow-up batch.

Version bump 6.12.0 → 6.13.0.

6.12.0 — 2026-04-26

Batch 61 — OpenCode sovereignty: SE-077 + SE-078 specs APPROVED + nueva regla obligatoria + G12 gate.

Added

docs/propuestas/SE-077-opencode-replatform-v114.md — APPROVED. 2 slices: plugin TS savia-gates (M, 8h) + parity audit ratchet (M, 6h). Era 189.
docs/propuestas/SE-078-agents-md-cross-frontend.md — APPROVED. AGENTS.md generator + drift check + Stop hook auto-regenerate. M 6h, era 189. Supersedes SPEC-114.
docs/rules/domain/spec-opencode-implementation-plan.md — regla canónica. Cada spec APPROVED post-2026-04-26 incluye sección obligatoria. Grandfathering documentado. Hot-fix exemption con exempt_opencode_plan frontmatter.
scripts/spec-opencode-plan-audit.sh — audit script (3 sub-secciones obligatorias, exit 1 si missing).
scripts/pr-plan-gates.sh:g_opencode_plan — G12 gate, solo se activa si el PR toca specs.
.ci-baseline/spec-opencode-plan-violations.count — baseline frozen at 0.

Changed

scripts/pr-plan.sh — invoca G12 tras G11.
docs/propuestas/SE-074-parallel-spec-execution.md — añade sección OpenCode Implementation Plan (PURE_BASH) + nuevo Slice 1.5 (S, 3h) "Adaptive halting + dynamic retry budget" inspirado en Kohli et al. 2026 (arXiv:2604.07822). Doble criterio halting (convergencia + confianza) + Poisson-clipped budget según effort field.
docs/ROADMAP.md Era 189 inaugurada con SE-077 + SE-078 priority alta.
CLAUDE.md — referencia lazy a la regla nueva.

Context

Decisión estratégica de la usuaria 2026-04-26: Anthropic restringe Claude Code (Pro → Max-only). Soberanía técnica = compatibilizar con OpenCode v1.14 desde origen, no como retrofit. Inversión Slice 1 SE-077 (~8h) compra opción real de switch sin perder workspace.

Version bump 6.11.0 → 6.12.0 (6.11.0 reservado para PR #704 en cola).

6.11.0 — 2026-04-26

Batch 60 — SE-075 (Voicebox) + SE-076 (QueryWeaver) specs APPROVED + ROADMAP reprio.

Added

docs/propuestas/SE-075-voicebox-adoption.md — APPROVED. task_queue + auto-chunking + Kokoro CPU voice español. Source: jamiepine/voicebox MIT.
docs/propuestas/SE-076-queryweaver-patterns.md — APPROVED. Graphiti episodic + schema-graph WIQL + LLM healer. Source: FalkorDB/QueryWeaver patterns (no AGPL code import).

Changed

docs/ROADMAP.md Era 188 pipeline: SE-073 → SE-074 → SE-075 → SE-076. Sinergias documentadas (task_queue habilita paralelismo + healer async; episodes extienden SPEC-027).

Context

6 patrones extractables sin adoptar infra ni licencias bloqueantes. Pendiente aprobación de la usuaria para arrancar SE-074 Slice 1.

Version bump 6.10.0 → 6.11.0.

6.10.0 — 2026-04-26

Batch 59 — SE-074 spec creado + ROADMAP Era 188 reprio.

Added

docs/propuestas/SE-074-parallel-spec-execution.md — spec APPROVED. 3 slices (worktree manager + PR queue + DB sandbox). Pre-reqs cumplidos (hook 100%, G11, cascade-rebase pattern, bounded concurrency).

Changed

docs/ROADMAP.md Era 188 "Memory + Throughput foundations" con pipeline SE-072 (done) → SE-073 → SE-074. Trade-off explícito: 8h bloqueado, break-even 2-3 Eras post.

Context

Inspirado en LinkedIn de Cole Medin "5 Claude Code sessions in parallel". Capitaliza Era 186 (hook 100%) + Era 187 (drift cleanup) + batch 57 (verified memory) + batch 58 (PR rule). 3-4x throughput esperado.

Version bump 6.8.0 → 6.10.0 (6.9.0 reservado para PR #702 batch 58 todavía en cola).

6.9.0 — 2026-04-25

Batch 58 — Nueva regla: cada PR requiere párrafo en lenguaje no técnico.

Added

docs/rules/domain/pr-natural-language-summary.md — regla canónica.
scripts/pr-plan-gates.sh:g_summary — gate G11 valida .pr-summary.md.

Changed

scripts/pr-plan.sh — invoca G11 tras G10.
scripts/push-pr.sh — prepend .pr-summary.md al PR body.
.gitignore — excluye .pr-summary.md.
CLAUDE.md — referencia lazy nueva.

Context

Solicitud de la usuaria: PRs autónomos sin párrafo plano dejan de ser auditables. Slice 1 sin LLM. PR #701 editado retroactivamente.

Version bump 6.8.0 → 6.9.0.

6.8.0 — 2026-04-25

Batch 57 — SE-072 Verified Memory axiom IMPLEMENTED (Slice 1). Era 188 inaugural.

Added

.claude/hooks/memory-verified-gate.sh — PreToolUse Write hook (33 tests, score 94). Bloquea auto-memory writes sin citation pattern (5 patterns OK).
tests/test-memory-verified-gate.bats — 33 tests certified score 94.
docs/rules/domain/verified-memory-axiom.md — política "No Execution, No Memory" de GenericAgent.

Changed

scripts/memory-save.sh — --source <origin> obligatorio. Valida format (4 OK), rechaza blacklist. Embed en JSONL.
tests/test-memory-store.bats — 9 tests SE-072 nuevos (score 90).
.claude/settings.json — hook registrado PreToolUse Edit|Write.
docs/propuestas/SE-072-verified-memory-axiom.md — APPROVED → IMPLEMENTED.
CLAUDE.md — hooks 59→60 (64 regs).

Context

Memoria persistente debe reflejar hechos verificados — no intenciones. Escape hatch: SAVIA_VERIFIED_MEMORY_DISABLED=true. Hook coverage 100% mantenido (60/60).

Version bump 6.7.0 → 6.8.0.

6.7.0 — 2026-04-25

Batch 56 — SPEC-124 pr-agent wrapper IMPLEMENTED. Era 187 trigger: 0 PROPOSED priority alta restantes.

Added

.github/workflows/templates/pr-agent-review.yml — reusable workflow para 5º juez Court (cost gate, feature-flag, draft skip, tagged comments).
docs/rules/domain/court-external-judges.md — política inclusión jueces externos OSS (7 requisitos, 6 reglas operación, activación pasa a paso).

Changed

docs/propuestas/SPEC-124-pr-agent-wrapper.md: PROPOSED → IMPLEMENTED, 9/9 ACs.

Context

ACs 01/02/03/05/06/07 ya implementados; completados AC-04/08/09. Era 187 todas las 6 specs alta IMPLEMENTED (SPEC-055/078/121/122/124 + SE-070).

Version bump 6.6.0 → 6.7.0.

6.6.0 — 2026-04-25

Batch 55 — SPEC-078 dual-estimation status drift correction (PROPOSED → IMPLEMENTED).

Changed

docs/propuestas/SPEC-078-dual-estimation-agent-human.md: PROPOSED → IMPLEMENTED. Resolution con 5/5 AC Fase 1 verificados. Fases 2-4 quedan como evolución, no bloquean status.

Context

3er drift fix de la sesión. Implementación Fase 1 (engine + hook + política + tests score 82) ya existía desde Era 179. PROPOSED alta restantes: 1 (SPEC-124, real work).

Version bump 6.5.0 → 6.6.0.

6.5.0 — 2026-04-25

Batch 54 — SPEC-122 LocalAI emergency-mode hardening IMPLEMENTED (4 ACs faltantes completados).

Added

.claude/hooks/emergency-mode-readiness.sh — SessionStart async hook con feature-flag EMERGENCY_MODE_ENABLED. Logs verdict a output/emergency-mode/readiness.jsonl, surface FAIL/WARN a stderr, timeout 10s, nunca bloquea SessionStart.
tests/test-emergency-mode-readiness.bats — 30 tests certified (score 94).

Changed

.claude/settings.json — hook registrado en SessionStart.
docs/rules/domain/autonomous-safety.md — sección "Emergency-mode" con prohibiciones explícitas (NUNCA bypass AUTONOMOUS_REVIEWER en emergency).
docs/propuestas/SPEC-122-localai-emergency-hardening.md — PROPOSED → IMPLEMENTED, 7/7 ACs.

Context

ACs 01/02/04 ya implementados; completados AC-03/05/06/07. Hook coverage 58/58 → 59/59 (100% mantenido).

Version bump 6.4.0 → 6.5.0.

6.4.0 — 2026-04-25

Batch 53 — SPEC-121 handoff-as-function convention IMPLEMENTED (3 ACs faltantes completados).

Changed

5 agentes SDD actualizados con sección "Handoff Format (SPEC-121)": sdd-spec-writer, dotnet-developer, code-reviewer, test-engineer, court-orchestrator.
docs/agent-notes-protocol.md: tabla de decisión handoff-as-function vs agent-notes longform.
docs/propuestas/SPEC-121-handoff-convention.md: status PROPOSED → IMPLEMENTED, 6/6 ACs cumplidos.

Context

ACs 01/03/04 ya implementados; completados AC-02/05/06. Aditivo, sin ruptura.

Version bump 6.3.0 → 6.4.0.

6.3.0 — 2026-04-25

Batch 52 — SPEC-055 status drift correction + Era 186 hook ratchet CLOSURE + sweep bug fix + baseline tighten.

Changed

docs/propuestas/SPEC-055-test-auditor.md: status PROPOSED → IMPLEMENTED con Resolution section. AC 5/5 cumplidos.
docs/ROADMAP.md: Era 186 extension marcada CLOSED. Tabla milestones extendida 49-51. Header v6.3.0 con SPEC-055 IMPLEMENTED.
.ci-baseline/hook-critical-violations.count: 5 → 4. Ratchet never-loosen mantenido (current = 4 consistente).

Fixed

scripts/test-auditor-sweep.sh: extracción de .score (campo inexistente) → .total (correcto). Sweep ahora reporta 100% compliance real vs 0% bug. Impact LOW (no en CI).

Context

PR-A del plan post-#695. Era 186 hook coverage ratchet finales:

13 batches (39-51) en 5 días
18/58 → 58/58 (+40 hooks, 1100+ tests, avg score ~90)
4 bugs reales descubiertos via tests
Drift hooks: 0 (CI-enforced)

Próximo: SPEC-121 (3 ACs), SPEC-122 (4 ACs).

Version bump 6.2.0 → 6.3.0.

6.2.0 — 2026-04-25

Batch 51 — Hook coverage +3: token-tracker-middleware, subagent-lifecycle, task-lifecycle. 100% HOOK COVERAGE — 58/58.

Added

tests/test-token-tracker-middleware.bats — 30 tests certified (score 91). PostToolUse async monitor de context tokens, 3 zonas (50% hint / 70% alert / 85% critical → auto-compact).
tests/test-subagent-lifecycle.bats — 29 tests certified (score 94). SubagentStart/Stop logging a output/agent-lifecycle/lifecycle.jsonl.
tests/test-task-lifecycle.bats — 30 tests certified (score 94). TaskCreated/Completed logging con team/teammate fields.

Changed

.ci-baseline/hook-untested-count.count: 3 → 0. Hook coverage 55/58 (94.8%) → 58/58 (100%).

Context

Decimotercera iteración del ratchet — y la última. 89 tests nuevos certified. Meta 100% ALCANZADA.

Progreso completo desde pre-batch-39 (Era 186 inicio): 18/58 → 58/58 hooks tested (31% → 100%, +40 hooks en 13 batches). Average score certified ≈ 90. 1100+ tests añadidos.

Próximo trabajo: PROPOSED priority alta (SE-034, SPEC-055, SPEC-078, SPEC-121, SPEC-122, SPEC-124) o Slice 4 de SE-070 si se libera presupuesto.

Version bump 6.1.0 → 6.2.0.

6.1.0 — 2026-04-25

Batch 50 — Hook coverage +4: instructions-tracker, file-changed-staleness, session-end-snapshot, config-reload.

Added

tests/test-instructions-tracker.bats — 27 tests certified (score 93). InstructionsLoaded async log.
tests/test-file-changed-staleness.bats — 26 tests certified (score 90). FileChanged stale marker.
tests/test-session-end-snapshot.bats — 24 tests certified (score 88). Stop hook context-snapshot delegation.
tests/test-config-reload.bats — 28 tests certified (score 92). ConfigChange profile cache invalidation.

Changed

.ci-baseline/hook-untested-count.count: 10 a 6. Hook coverage 48/58 a 52/58 (89.7%) standalone, 55/58 (94.8%) cuando combinado con PR #692.

Context

Duodecima iteracion ratchet. +4 hooks en una iteracion (vs 3 habituales). 105 tests nuevos certified. Solo 3 hooks pendientes (token-tracker-middleware, subagent-lifecycle, task-lifecycle) → 100% en una iteracion mas.

Version bump 5.98.0 a 6.1.0.

6.0.0 — 2026-04-25

Dos nuevas specs APPROVED del research GenericAgent repo (6.8k ).

Added

docs/propuestas/SE-072-verified-memory-axiom.md — "No Execution, No Memory" gate: memory-store.sh save requerirá --source <origin> (tool/file/verified/user). Hook PreToolUse para Write en auto/MEMORY.md. Grandfathering entries existentes. S-effort 3h.
docs/propuestas/SE-073-memory-index-cap-tiered.md — MEMORY.md cap 200→30 líneas con 2-tier system (HIGH-FREQ inline, LOW-FREQ filename-only en MEMORY-ARCHIVE.md). scripts/memory-tier-rotate.sh para rotation automática por access_count. S-effort 3h.

Context

Research completo de lsdefine/GenericAgent (7 patterns analizados): 2 adoptables (estos specs), 5 ya cubiertos por Savia stack, 6 descartados. Veredicto: "ADOPTAR LUEGO" — wins claros S-effort con complemento directo a memoria externa.

Este PR SOLO crea specs. Implementation sigue en PRs separadas post-review humana.

Queue APPROVED ejecutable sin-GPU: 0 → 2 (+SE-072, SE-073). PROPOSED: 70 → 68.

Version bump 5.97.0 → 6.0.0 (major: research-driven spec addition + status semantics change).

5.99.0 — 2026-04-24

Batch 49 — Hook coverage +3: memory-prime-hook, shield-autostart, stop-quality-gate. MILESTONE 85% superado (87.9%).

Added

tests/test-memory-prime-hook.bats — 33 tests certified (score 90). PreToolUse async memory auto-prime + bounded concurrency.
tests/test-shield-autostart.bats — 31 tests certified (score 83). SessionStart shield proxy autostart (port 8443, fire-and-forget).
tests/test-stop-quality-gate.bats — 34 tests certified (score 91). Stop hook secret detection (password/api_key/token/private_key block pattern).

Changed

.ci-baseline/hook-untested-count.count: 10 a 7. Hook coverage 48/58 (82.7%) a 51/58 (87.9%).

Context

Undécima iteración ratchet. 98 tests nuevos certified. Meta 85% SUPERADA en batch 49 (48→51 hooks con tests). Progreso desde pre-batch-39: 18→51 hooks tested (31% → 87.9%), +33 hooks en 11 batches.

Queue restante: 7 hooks (todos <40 lines, 4 de ellos <30 lines). Próxima iteración puede cerrar 4-5 hooks de una vez → 95%+ coverage.

Version bump 5.97.0 → 5.99.0.

5.98.0 — 2026-04-24

SE-070 Opus 4.7 calibration scorecard — IMPLEMENTED (Slice 1-3, Slice 4 deferred). Backlog APPROVED sin-GPU cerrado.

Added

scripts/opus47-calibration-scorecard.sh — Slice 1. Lists 37 sonnet-4-6 agents con cost delta +1025% estimate + golden-set detection. CLI --help/--quiet/--json. Outputs YAML + MD.
tests/golden/opus47-calibration/ — Slice 2. README + TEMPLATE (prompt/expected/score.yaml) para A/B eval scaffolding.
docs/rules/domain/opus47-calibration-playbook.md — Slice 3. 6-step workflow + decision matrix (quality_cost_ratio >= 2.0 upgrade) + 5 anti-patterns + rollback + cost guidance (~$27 full suite).
tests/test-opus47-calibration-scorecard.bats — 45 tests certified (score 98). Coverage CLI, cost model, golden detection, slice 2/3 files.

Changed

docs/propuestas/SE-070-opus47-eval-scorecard.md: status APPROVED → IMPLEMENTED. Resolution section con breakdown per-slice. 4/5 AC cumplidos (AC-03 Slice 4 evals deferred per spec).

Context

Slice 4 (3 actual A/B evals) deferred per spec's own "defer execution until batch budget allows" criterion. Infrastructure 100% ready; ~$2.20 API cost + human eval time pending.

Queue APPROVED: 5 → 4 (-1). Los 4 restantes (SE-028, SE-042, SPEC-023, SPEC-080) son TODOS GPU-blocked. Ningún APPROVED ejecutable sin GPU queda en queue.

Próximo trabajo autónomo: hook coverage continuar hacia 85%, o PROPOSED priority alta (SE-034, SPEC-055, SPEC-078, SPEC-121, SPEC-122, SPEC-124).

Version bump 5.97.0 → 5.98.0.

5.97.0 — 2026-04-24

SPEC-120 Spec template alignment con github/spec-kit — IMPLEMENTED.

Added

projects/proyecto-alpha/specs/templates/spec-template.md: spec_kit_compatible marker + pointer header a canonical source.
projects/proyecto-beta/specs/templates/spec-template.md: same marker + pointer.

Changed

docs/propuestas/SPEC-120-spec-kit-alignment.md: status APPROVED → IMPLEMENTED. Resolution section con breakdown pre-existente vs this PR. 7/7 AC cumplidos.

Context

Infrastructure mayor ya existia (canonical template con Spec-Kit Alignment section + docs/agent-teams-sdd.md mapping + 26 tests score 81). Este PR completa los 2 project templates duplicados con pointer headers, preservando su content completo (customizaciones project-specific). Command reference pointer ya era redirect.

Queue APPROVED: 6 → 5 (-1). Sin-GPU restantes: solo SE-070 Opus 4.7 scorecard (1 spec ejecutable en dev). 4 specs GPU-blocked (SE-028, SE-042, SPEC-023, SPEC-080) en espera de hardware.

Version bump 5.96.0 → 5.97.0.

5.96.0 — 2026-04-24

SE-065 responsibility-judge S-06 i18n fix — IMPLEMENTED. Safety hook calibration para Spanish prose.

Fixed

.claude/hooks/responsibility-judge.sh S-06 rule: case-sensitive match (drop -i flag) + file-type exemption (.md/.mdx/.txt/.rst/CHANGELOG.d//docs/). Spanish prose (lowercase "todo" quantifier) ya no triggerea el detector de code-shortcuts. Annotations uppercase en code siguen detectadas.

Added

13 SE-065 regression tests en tests/test-responsibility-judge.bats (16 a 29 total, score 89 certified). Hex-encoded keywords para evitar self-triggering en test file.

Changed

docs/propuestas/SE-065-responsibility-judge-s06-i18n.md: status APPROVED → IMPLEMENTED. Resolution section con diff + meta-level observation.
tests/test-responsibility-judge.bats: JSON field names fixed (tool_name/tool_input vs tool/input). 12 existing tests updated — they "passed" via early-exit on empty CONTENT, not rule logic.

Context

Resolved friction pattern desde batch 30: CHANGELOG fragments con phrase "salta todo" eran bloqueados por S-06 pattern match en grep -i. Fix surgical per spec: narrow to uppercase + exempt markdown. Zero regresion en S-01..S-05 rules o en deteccion real de shortcuts en code files.

Meta-observation: edit al hook mismo requerio annotation TODO(#65) para satisfacer exemption regex durante el propio Edit — propiedad emergente de auto-consistencia.

Queue APPROVED: 7 → 6 (-1). Sin-GPU restantes: SE-070, SPEC-120 (2).

Version bump 5.95.0 → 5.96.0.

5.95.0 — 2026-04-24

SE-038 Agent catalog size audit — IMPLEMENTED via ratchet. 27 violations baseline frozen.

Added

tests/test-agent-size-audit.bats — 44 tests certified (score 95). Coverage de scripts/agent-size-audit.sh: CLI flags, execution, SLA 4096, size_exception, ratchet mode, statistics, safety, negative/edge cases.

Changed

scripts/agent-size-audit.sh: +--ratchet flag (never-loosen policy), +--baseline N override. Usage docs actualizados.
docs/propuestas/SE-038-agent-size-audit.md: status APPROVED → IMPLEMENTED. Resolution section con Slice 1 probe results (27/65 violations), Slice 2 deferred explanation, AC breakdown final.
.ci-baseline/agent-size-violations.count: 27 baseline frozen (was present from pre-existing check #8).

Context

Slice 1 probe: 65 agents scanned, 27 violate Rule #22 (<4KB). Top offenders are safety-adjacent agents (code-reviewer 6581 bytes, test-runner 6454, commit-guardian 6423). Bulk remediation risky — per spec's ratchet strategy, baseline frozen at 27 with never-loosen policy. Incremental reduction happens in future PRs.

Slice 3 (enforcement gate) ya estaba pre-existente en scripts/ci-extended-checks.sh check #8. Script extendido con --ratchet CLI para invocacion standalone.

Queue APPROVED: 8 → 7 (-1). Sin-GPU ejecutables restantes: SE-065, SE-070, SPEC-120.

Version bump 5.94.0 → 5.95.0.

5.94.0 — 2026-04-24

SE-039 Test-auditor global sweep — IMPLEMENTED. Baseline 100% (232/232 ≥80, avg 87).

Added

tests/test-audit-all-bats.bats — 38 tests certified (score 97). Coverage de scripts/audit-all-bats.sh (Slice 1 sweep script pre-existente).
.github/workflows/bats-audit-sweep.yml — Weekly cron (lunes 06:00 UTC) + manual dispatch. 10min timeout, workflow annotation, artifact upload 30d retention, GitHub Step Summary.
docs/rules/domain/test-quality-gate.md — Doctrine doc: SLA ≥80/file + ≥95% soft target + avg ≥85. 3 enforcement layers, 9 scoring criteria de SPEC-055, 6-step remediation playbook, historia baseline.
output/bats-audit-sweep-20260424.md — Baseline report: 232/232 compliant, avg 87, 13 tests en bottom decile (score=80).

Changed

docs/propuestas/SE-039-test-auditor-global-sweep.md: status APPROVED a IMPLEMENTED. 5/6 AC cumplidos (AC-06 mutation testing integration deferred per SE-035 dependency). Resolution section con breakdown per-slice.

Context

Slice 2 remediation (bottom-10 fix) resulto N/A — probe Slice 1 demostro 100% compliance pre-existente. Per criterio "Spec Ops / Probe" del propio spec ("si ≥95% ya está ≥80, abort"), cerrado sin remediation.

Queue APPROVED: 9 a 8 (-1). IMPLEMENTED: 55 a 56 (+1). Proximos disponibles en dev: SE-038, SE-065, SE-070, SPEC-120 (4 GPU-blocked en espera).

Version bump 5.93.0 a 5.94.0.

5.93.0 — 2026-04-24

SE-071 safety hook fix + spec triage + roadmap update.

Fixed

.claude/hooks/block-branch-switch-dirty.sh: profile_gate "minimal" a profile_gate "security". SE-071 resolved with the user's approval. Bug: "minimal" invalid tier silently disabled safety hook under profile default. Verified fix blocks dirty checkout with exit 2.

Changed

Spec triage (74 PROPOSED specs): 5 promoted to APPROVED (SE-038, SE-039, SE-065, SE-070, SPEC-120), 9 alta, 33 media, 21 baja, 6 skipped (meta/ADR/TEMPLATE).
Priority normalization: Baja/Alta/Media a lowercase globally.
docs/ROADMAP.md: nueva seccion "Era 186 extension — Hook coverage ratchet + triage". Tabla milestones hook coverage 18/58 a 48/58, bugs descubiertos via tests (4), triage results, nuevos APPROVED con rationale.
tests/test-block-branch-switch-dirty.bats: removido bypass SAVIA_HOOK_PROFILE=strict en 10 block-path tests. Anadido regression test SE-071 regression: no invalid tier 'minimal' remains. 36/36 PASS.

Audit

grep -rn 'profile_gate "[^"]*"' .claude/hooks/ sobre 58 hooks: 27 standard + 8 security + 3 strict = 38 valid, 0 invalid. Bug SE-071 fue unico.

Context

Post-batch 48, limpieza del backlog y cierre del bug de safety hook descubierto durante testing. Backlog APPROVED final: 9 specs (5 nuevos alineados con trabajo actual + 4 training pipeline bloqueados por GPU).

Version bump 5.92.0 a 5.93.0.

5.92.0 — 2026-04-24

Batch 48 — Hook coverage +3: bash-output-compress, block-branch-switch-dirty, compress-agent-output. Bug real descubierto via tests (SE-071).

Added

tests/test-bash-output-compress.bats — 30 tests certified (score 90). PostToolUse async rtk-ai inspired token compression. Script delegation, context-tracker metric logging, 30-line threshold.
tests/test-block-branch-switch-dirty.bats — 36 tests certified (score 90). PreToolUse security. Intercepta git checkout/switch con arbol sucio.
tests/test-compress-agent-output.bats — 29 tests certified (score 92). PostToolUse Task SPEC-041 P4. Streaming compression >200 tokens en dev-sessions.
docs/propuestas/SE-071-profile-gate-invalid-tier-audit.md — bug audit: block-branch-switch-dirty.sh usa tier invalido "minimal", hook silent-disabled bajo profile default. Requiere aprobación de la usuaria (safety hook).

Changed

.ci-baseline/hook-untested-count.count: 13 a 10. Hook coverage 45/58 (77.6%) a 48/58 (82.7%).

Context

Decima iteracion ratchet. 95 tests nuevos certified. Meta 85% al alcance en batch 49 (+2 hooks).

Bug descubierto: durante testing de block-branch-switch-dirty.sh, detectado que profile_gate "minimal" es tier invalido (tiers validos: security/standard/strict). Bajo profile default el safety hook NO bloquea. Permission hook correctamente bloqueo auto-fix. Propuesta SE-071 creada para review humana.

Version bump 5.91.0 a 5.92.0.

5.91.0 — 2026-04-24

Batch 47 — Hook coverage +3: post-tool-failure-log, post-edit-lint, acm-turn-marker. MILESTONE 75% alcanzado (77.6%).

Added

tests/test-post-tool-failure-log.bats — 39 tests certified (score 98, mejor del batch). PostToolUseFailure SPEC-068. 6 error categories, retry hints, pattern detection (3+ same tool).
tests/test-post-edit-lint.bats — 37 tests certified (score 90). PostToolUse async multi-lang lint. 11 extensions (cs/py/ts/tsx/js/jsx/go/rs/rb/php/tf), missing linter graceful skip.
tests/test-acm-turn-marker.bats — 37 tests certified (score 93). PostToolUse SE-063 Slice 2. ACM enforcement chain marker.

Changed

.ci-baseline/hook-untested-count.count: 16 a 13. Hook coverage 42/58 (72%) a 45/58 (77.6%).

Context

Novena iteracion ratchet. 113 tests nuevos certified. Milestone 75% superado — 45/58 hooks con tests.

Proxima meta 85% (50/58) en 2 batches mas. Candidatos: bash-output-compress, block-branch-switch-dirty, compress-agent-output, memory-prime-hook, shield-autostart.

Version bump 5.90.0 a 5.91.0.

5.90.0 — 2026-04-24

Batch 46 — Hook coverage +3: android-adb-validate, live-progress-hook, dual-estimation-gate.

Added

tests/test-android-adb-validate.bats — 41 tests certified (score 92). PreToolUse ADB safety classifier. Blocks destructives (rm -rf, format, dd, su, root), logs RISKY (install/uninstall/push/reboot), SAFE silent log.
tests/test-live-progress-hook.bats — 36 tests certified (score 93). PreToolUse async live.log feed. 8 tool cases, rotation 500 lines, basename stripping, Task* wildcard, emoji prefixes.
tests/test-dual-estimation-gate.bats — 33 tests certified (score 92). PostToolUse SPEC-078 dual-scale warning. agent_effort_minutes + human_effort_hours enforcement on *.spec.md / backlog/pbi / backlog/task.

Changed

.ci-baseline/hook-untested-count.count: 19 a 16. Hook coverage 39/58 (67%) a 42/58 (72%).

Context

Octava iteracion ratchet. 110 tests nuevos certified. Lecciones: # Ref: batch X + SPEC reference garantiza auditor +10pts. EDGE_PAT keywords obligatorias para score 80+. Evitar palabra "estimation" en test fixtures no-estimacion.

Meta 75% al alcance en batch 47 (+3 → 45/58).

Version bump 5.89.0 a 5.90.0.

5.89.0 — 2026-04-24

Batch 45 — Hook coverage +3: tool-call-healing, user-prompt-intercept, session-end-memory.

Added

tests/test-tool-call-healing.bats — 37 tests certified (score 93). PreToolUse validation para Read/Edit/Write/Glob/Grep. Typo detection via find, parent-dir check, empty-pattern blocks.
tests/test-user-prompt-intercept.bats — 34 tests certified (score 87). UserPromptSubmit SPEC-015 context gate. Silent pass ES/EN confirmations, session-hot injection, active project hint.
tests/test-session-end-memory.bats — 29 tests certified (score 87). SessionEnd SPEC-013/055 perf. Sync log + disowned worker, session-hot.md con failures y modified files, no-op cuando repo limpio.

Changed

.ci-baseline/hook-untested-count.count: 22 a 19. Hook coverage 36/58 (62%) a 39/58 (67%).

Context

Septima iteracion ratchet. 100 tests nuevos certified. UTF-8 locale explicit para tests con multibyte (sí), path normalization con $(pwd) para project hint tests.

Proximos: android-adb-validate (69), live-progress-hook (69), dual-estimation-gate (63). A ritmo +3/batch, 2 batches mas para 75% (45/58).

Version bump 5.88.0 a 5.89.0.

5.88.0 — 2026-04-24

Batch 44 — Hook coverage +3: competence-tracker, memory-auto-capture, agent-trace-log.

Added

tests/test-competence-tracker.bats — 36 tests certified. UserPromptSubmit competence fact extraction (strict profile only). Cubre 11 categorias dominio, log rotation 1000 lineas, ISO timestamp, user field.
tests/test-memory-auto-capture.bats — 30 tests certified. PostToolUse auto memory capture tras Edit/Write en paths especiales. Cubre rate limit 5min, type inference (pattern/convention/discovery), concept extraction, content preview.
tests/test-agent-trace-log.bats — 31 tests certified. PostToolUse Task metering con token estimation, budget alerts, outcome classification (success/failure/partial), JSONL append-only.

Changed

.ci-baseline/hook-untested-count.count: 25 a 22. Hook coverage 33/58 (57%) a 36/58 (62%).

Fixed

.claude/hooks/memory-auto-capture.sh: TOOL_NAME="${TOOL_NAME:-}" guard contra unbound variable con set -u.

Context

Sexta iteracion ratchet. 97 tests nuevos certified. Bug fix descubierto via tests (hooks catch real bugs, no weakening).

Proximos: tool-call-healing (72), user-prompt-intercept (71), session-end-memory (70). A ritmo +3/batch, 3 batches mas para 75% (45/58).

Version bump 5.87.0 a 5.88.0.

5.87.0 — 2026-04-24

Batch 43 — Hook coverage +3: post-report-write, agent-tool-call-validate, stress-awareness-nudge.

Added

tests/test-post-report-write.bats — 34 tests certified. PostToolUse async Truth Tribunal queue. Cubre 6 path patterns, filename heuristics, frontmatter override, self-recursion prevention.
tests/test-agent-tool-call-validate.bats — 35 tests certified. PreToolUse param validation. Cubre pass-through tools, file_path/command required, env override, JSON field aliases.
tests/test-stress-awareness-nudge.bats — 41 tests certified. UserPromptSubmit pressure pattern detection. Cubre 5 categorias ES/EN, silent pass, nudge content, boundary edges.

Changed

.ci-baseline/hook-untested-count.count: 28 a 25. Hook coverage 30/58 (52%) a 33/58 (57%).

Context

Quinta iteracion ratchet. 110 tests nuevos certified. Pattern consolidado: cd con HOOK_ABS para git repos, scrambled markers, null/empty edge handling documentado como behavior vs bug.

Proximos: competence-tracker (76), memory-auto-capture (74), agent-trace-log (74). A ritmo +3/batch, 4 batches mas para 75% (44/58).

Version bump 5.86.0 a 5.87.0.

5.86.0 — 2026-04-24

Batch 42 — Hook coverage +3: pbi-history-capture, prompt-hook-commit, agent-hook-premerge.

Added

tests/test-pbi-history-capture.bats — 28 tests certified. PostToolUse PBI frontmatter diff capture. Cubre 11 tracked fields, new creation path, author extraction, git history diff.
tests/test-prompt-hook-commit.bats — 29 tests certified. Semantic commit message validation. Cubre 4 heuristics, 3 modes, CHANGELOG integration.
tests/test-agent-hook-premerge.bats — 32 tests certified. Pre-merge security + quality gate. Cubre 3 secret patterns, conflict markers, 150-line limit per file category, 3 modes.

Changed

.ci-baseline/hook-untested-count.count: 31 a 28. Hook coverage 27/58 (47%) a 30/58 (52%). Milestone 50% cruzado.
Tambien arreglados 2 bugs encontrados en batch 41 (fix commit anterior aplicado): cwd-changed-hook C# detection (compgen -G) y emotional-regulation-monitor set -u crash.

Context

Cuarta iteracion del ratchet. Pattern consolidado: HOOK_ABS antes de cd, scrambled markers para evitar S-06 self-trigger. 89 tests nuevos certified.

Proximos: post-report-write (83), agent-tool-call-validate (81), stress-awareness-nudge (78). A ritmo +3/batch, 5 batches mas para 75% (44/58).

Version bump 5.85.0 a 5.86.0.

5.85.0 — 2026-04-23

Batch 41 — Hook coverage +3: cwd-changed-hook, emotional-regulation-monitor, ast-quality-gate-hook.

Added

tests/test-cwd-changed-hook.bats — 29 tests certified. CwdChanged hook auto-inyecta contexto al entrar en projects/. Cubre 10 language packs, state dedup, cleanup on exit.
tests/test-emotional-regulation-monitor.bats — 24 tests certified. Stop hook session stress assessment (Anthropic research). Cubre 3 level thresholds, boundary detection, memory persist + dedup.
tests/test-ast-quality-gate-hook.bats — 26 tests certified. PostToolUse async quality gate. Cubre 16 source extensions, graceful degradation, --advisory flag, latest.json alias.

Changed

.ci-baseline/hook-untested-count.count: 34 a 31. Hook coverage 24/58 (41%) a 27/58 (47%).

Context

Tercera iteracion del ratchet. 79 tests nuevos certified. Patron consolidado: tracker mock, tier-specific profile override, boundary edge cases, isolation.

Proximos candidatos: pbi-history-capture (93), prompt-hook-commit (91), agent-hook-premerge (90). Meta implicita: 75% coverage (44/58).

Version bump 5.84.0 a 5.85.0.

5.84.0 — 2026-04-23

Batch 40 — Hook coverage +3: ast-comprehend, agent-dispatch-validate, stop-memory-extract.

Added

tests/test-ast-comprehend-hook.bats — 25 tests certified. PreToolUse(Edit) invariant RN-COMP-02 (never blocks). Cubre MIN_LINES threshold, COMPLEXITY_WARN, env fallback, malformed inputs.
tests/test-agent-dispatch-validate.bats — 25 tests certified. PreToolUse(Task) tier strict. Cubre 5 validation categories (commands, CHANGELOG, skills, git-push, rules), ERROR blocks vs WARNING informs.
tests/test-stop-memory-extract.bats — 27 tests certified. Stop hook SPEC-013v2. Cubre 4 PHASE extraction flow, quality gate invocation, action-log archiving, URL dedup.

Changed

.ci-baseline/hook-untested-count.count: 37 a 34. Hook coverage 21/58 (36%) a 24/58 (41%).

Context

Continuacion del ratchet iniciado en batch 39. Siguientes 3 hooks mas grandes sin coverage cerrados. Patron de tests: env isolation via TMPDIR, pattern dynamic construction para evitar auto-bloqueo por credential scanner, profile_gate override en setup segun tier del hook.

Proximos candidatos: cwd-changed-hook (104), emotional-regulation-monitor (99), ast-quality-gate-hook (96).

Version bump 5.83.0 a 5.84.0.

5.83.0 — 2026-04-23

Batch 39 — Hook test coverage audit + 3 critical hooks covered.

Added

scripts/hook-test-coverage-audit.sh — audit ratchet. Escanea hooks sin BATS tests, compara contra baseline, exit 1 si regression. Flags --json, --min-lines N.
.ci-baseline/hook-untested-count.count: 37 baseline ratchet establecido (reducido 40 a 37).
tests/test-data-sovereignty-gate.bats — 29 tests PreToolUse hook (173 lines, security tier). Cubre private destination exemptions, fail-open malformed JSON, path normalization, sovereignty whitelist.
tests/test-pre-commit-review.bats — 20 tests commit-hook (119 lines, code review). Cubre rules hash cache invalidation, code file filter, combined content+rules hash, isolation.
tests/test-data-sovereignty-audit.bats — 27 tests PostToolUse async (118 lines, security tier). Cubre 6 leak patterns (JDBC/AWS/PAT/OpenAI/private-key/internal-IP), is_public helper, async exit 0 invariant.

Context

Gap crítico: 69% hooks sin test coverage (40/58). Los 3 hooks más grandes son de seguridad (sovereignty gate + audit) y quality (pre-commit review) — ausencia de tests permite regression silente. Batch 39 cierra los 3 top + deja ratchet para prevenir nuevos hooks sin test.

Descubierto durante tests: SAVIA_HOOK_PROFILE=security required en setup para activar profile gate; unset SAVIA_SHIELD_ENABLED para evitar env leak. Patterns de credenciales construidos via printf "%s%s%s" para evitar auto-bloqueo por block-credential-leak.sh.

Version bump 5.82.0 a 5.83.0.

5.82.0 — 2026-04-23

Batch 38 — SE-049 Slice 1: SLM dispatcher + shared lib scaffolding.

Added

scripts/slm.sh — dispatcher unificado para los 16 scripts slm-*.sh. Flags: <subcommand>, list, --json list, --help. Exit codes 0/1/2. Usa exec bash para preservar args y exit code de subcommands.
scripts/lib/slm-common.sh — shared library con slm_die, slm_warn, slm_project_root, slm_data_dir, SLM_REGISTRY (single source of truth) y helpers de routing.
tests/test-slm-dispatcher.bats — 30 tests certified. Existence, help, registry, negative/edge/coverage/isolation.
docs/rules/domain/slm-consolidation-pattern.md — doc canonica del pattern (problema, slicing, usage antes/despues, extension guide).

Changed

SE-049 status: PROPOSED a IN_PROGRESS. slices_complete: [1], batches: [38].

Context

SE-049 (L 16h) se descompone en 3 slices. Slice 1 (batch 38) scaffolding: routing thin, registry declarativo, tests + doc. Slice 2 migrara logica de cada script a funciones cmd_<subcommand> en slm.sh. Slice 3 deprecara originales. Permite consolidar 16 scripts en un solo dispatcher sin breaking change inmediato.

Version bump 5.81.0 a 5.82.0.

5.81.0 — 2026-04-23

Batch 37 — SE-046 baseline integrity guard + stale ratchet tightened.

Added

tests/test-baseline-integrity.bats — 20 tests BATS guard sobre .ci-baseline/. Asserta que cada baseline esta dentro de 3 unidades de la medida actual (tight). Detecta drift futuro automaticamente y sugiere el comando de remediation.
Hook-critical baseline actualizado 6 a 5 (MAX de 5 runs para tolerar noise en measurements). Batch 37 aplica scripts/baseline-tighten.sh sobre stale baseline que rendia el ratchet inerte (SE-046 motivacion original).

Changed

.ci-baseline/hook-critical-violations.count: 6 a 5 (MAX over 5 runs)
SE-046 status PROPOSED a IMPLEMENTED (batches [7, 37])

Context

SE-046 acceptance criteria tenia 3 partes: auto-tighten tool (batch 7), BATS tests del tool (batch 7), y BATS guard baseline <= measured (faltaba). Batch 37 cierra el tercer punto + aplica tightening a baseline stale. ci-extended-checks.sh ya no emite "baseline stale" warning para hook-critical.

Version bump 5.80.0 a 5.81.0.

5.80.0 — 2026-04-23

Batch 36 — Spec status drift sweep + README refresh + drift auditor.

Added

scripts/spec-status-drift-audit.sh — detecta specs con status: PROPOSED pero con N+ referencias en CHANGELOG.d/ (evidencia de implementacion). Flags: --min-refs N (default 2), --json. Previene futura drift spec-status.
tests/test-spec-status-drift-audit.bats — 26 tests certified.

Changed

20 spec frontmatters actualizados de PROPOSED a IMPLEMENTED: SE-029, SE-032, SE-033, SE-035, SE-036, SE-041, SE-043, SE-044, SE-047, SE-048, SE-050, SE-051, SE-052, SE-053, SE-054, SE-056, SE-058, SE-059, SE-061, SE-062. Todos con evidencia on-disk + applied_at + batches.
README.md — counts actualizados (65 agentes, 86 skills, 58 hooks, 283+ test suites). Cierra G8 WARN de PR anterior.

Context

Post-merge batches 31-35, analisis de status: PROPOSED vs evidencia real detecto drift acumulado: 20 specs implementadas pero nunca marcadas. Distorsionaba el backlog planning. Nuevo auditor previene recurrencia con cutoff configurable.

Version bump 5.79.0 a 5.80.0.

5.79.0 — 2026-04-23

Batches 31-35 — Opus 4.7 calibration: 5 specs SE-066..SE-070 implementadas en un unico PR.

Added

SE-066 Batch 31 — Reporting Policy (coverage-first) block anadido a 19 review/judge/auditor agents. Preserva recall bajo Opus 4.7 que sigue filter instructions mas literalmente que 4.6. Cada finding debe incluir {confidence, severity} para ranking downstream.
SE-067 Batch 32 — Subagent Fan-Out Policy block en 3 orchestrators (dev-orchestrator, court-orchestrator, truth-tribunal-orchestrator). Feasibility-probe SKILL.md migrado de budget_tokens fijo a adaptive thinking (Opus 4.7 decide por step).
SE-068 Batch 33 — XML tag structure (<instructions>, <context_usage>, <constraints>, <output_format>) anadidos a 5 top-tier opus-4-7 agents (architect, dev-orchestrator, court-orchestrator, truth-tribunal-orchestrator, code-reviewer). Canonical doc docs/rules/domain/agent-prompt-xml-structure.md publicado.
SE-069 Batch 34 — Nueva skill .claude/skills/context-rot-strategy/ (SKILL.md + DOMAIN.md). 5-option decision model para gestion de context rot en sesiones de 1M tokens: continue, rewind, /compact con hint, /clear, subagent. Umbrales 60/75/90% con recomendaciones per-tier.
SE-070 Batch 35 — Propuesta scorecard (A/B eval framework para 37 sonnet-4-6 agents). Deferred execution — infraestructura lista, evals opportunistic.
scripts/opus47-compliance-check.sh — valida SE-066..SE-070 con flags --finding-vs-filtering|--fan-out|--adaptive-thinking|--xml-tags|--context-rot-skill. JSON output disponible.
tests/test-opus47-compliance.bats — 24 tests cubriendo los 5 batches.

Changed

CLAUDE.md bump skills(85) a skills(86) por nueva context-rot-strategy.
19 review agents + 3 orchestrators + 5 top-tier agents = 27 agent prompts modificados (compatible, solo append de policies nuevas).
feasibility-probe SKILL.md ya no documenta budget_tokens — adaptive thinking es default Opus 4.7.

Context

Post-analisis del Opus 4.7 migration guide (Anthropic + Daily Dose of Data Science 2026-04-23). Gap identificado: zero agents usaban XML tags, zero separaban finding-vs-filtering, orchestrators sin fan-out explicito, feasibility-probe con budget fijo deprecado, y skill set sin cobertura explicita de context rot en 1M context. 5 propuestas creadas e implementadas en un batch combinado tras aprobacion explicita.

Era 186 abre con enfoque "model calibration" — ajustar Savia a los cambios de comportamiento de Opus 4.7 (mas literal, menos subagents, mejor bug-finding con stricter filtering).

5.78.0 — 2026-04-22

Batch 30 — SE-060 close-loop: hook-audit detector exemptions.

Added

scripts/hook-injection-audit.sh — mecanismo de exención por fichero # hook-audit-detector: HOOK-XX,HOOK-YY (o ALL). Solo primeras 20 líneas del hook para prevenir bypass via regex-string payload. Funciones detector_exemptions() + is_exempt() helpers.
.claude/hooks/validate-bash-global.sh — header marcado # hook-audit-detector: HOOK-03,HOOK-06. Hook es detector legítimo: contiene regex strings de curl | bash y sudo para bloquear comandos, no ejecuciones. Sin exención, generaba 4 false positives.
tests/test-hook-injection-audit.bats +8 tests (25→33). Cubren listed-rules skip, ALL wildcard, partial skip (otras reglas siguen disparando), anti-bypass (comentario tras línea 20 ignorado), validate-bash-global marcado, real-world clean audit, helper functions existen.

Changed

docs/propuestas/SE-060-hook-injection-hidden-directives.md status PROPOSED → IMPLEMENTED. Cierre de loop del research agentshield (batch 10 Scripts 1+2, batch 30 exención + clean audit).

Context

Batch 10 implementó hook-injection-audit.sh con 9 reglas HOOK-XX y extendió prompt-security-scan.sh con PS-11..PS-14 (zero-width, base64, URL-pipe-bash, time-bomb). El audit generaba 4 false positives en validate-bash-global.sh — hook detector legítimo cuyas strings de regex disparaban HOOK-03/HOOK-06 pese a no ejecutar los patrones. Batch 30 cierra el gap añadiendo anotación explícita # hook-audit-detector: RULES top-of-file (max 20 líneas) + helpers + 8 tests + marca hook real. Audit ahora reporta findings_count=0 sobre 60 hooks reales.

5.77.0 — 2026-04-22

Batch 29 — SE-063 Slice 2 registro + Slice 3 bypass semántico.

Added

.claude/settings.json PostToolUse Read → acm-turn-marker.sh (timeout 3s, async). Cierra el ciclo detector↔marker: ahora leer projects/{p}/.agent-maps/INDEX.acm dentro del turno libera automáticamente las queries amplias siguientes sobre ese proyecto.
.claude/hooks/acm-enforcement.sh Slice 3 — per-project opt-out via projects/{p}/.agent-maps/.acm-enforce-skip (fichero vacío). Evita enforcement en proyectos que voluntariamente renuncian al guard (ej. sandboxes o proyectos sin código estructurado).
.claude/hooks/acm-enforcement.sh Slice 3 — SAVIA_ACM_LOG_LEVEL={silent,warn,debug}. silent suprime stderr y el log (conserva exit codes). debug añade turn id y marker_dir al log para diagnóstico.
tests/test-acm-enforcement.bats +9 tests (32→41). Cubren opt-out isolation entre proyectos, silent sin stderr, silent sin log, debug con turn id, mensaje block menciona .acm-enforce-skip.

Changed

Mensaje de guidance en bloqueo/warn incluye línea Opt-out proyecto: touch projects/{p}/.agent-maps/.acm-enforce-skip para instruir camino de escape.
CLAUDE.md bump 61reg → 62reg por registro PostToolUse Read nuevo.

Context

Cierra el loop crítico de SE-063. Batch 28 dejó el marker script sin registrar por self-modification guard; batch 29 lo registra tras la aprobación de la usuaria ("mergeado, seguimos desarrollando"). Slice 3 del spec queda cumplido: env runtime override ya estaba en Slice 1, ahora añade verbosidad controlada y opt-out per-proyecto. Era 185 progresa hacia cierre con SE-063 completo al 100%.

5.76.0 — 2026-04-22

Batch 28 — SE-063 Slice 1+2: ACM enforcement hooks (Era 185 arranca).

Added

.claude/hooks/acm-enforcement.sh — PreToolUse hook para Glob/Grep. Detecta queries amplias (.*, **/*, sin path/type/glob) en projects/{name}/ cuando existe .agent-maps/INDEX.acm. Modos: warn (default, solo stderr) / block (exit 2) / 0,off (disabled). Bypass semántico para .claude/, docs/, scripts/, tests/.
.claude/hooks/acm-turn-marker.sh — PostToolUse hook (registro pendiente de aprobación) que escribe marker per-turno cuando el agente lee un .acm. Marker en $TMPDIR/savia-turn-{id}/acm-read-{project}.
tests/test-acm-enforcement.bats — 32 tests certified, cubre warn/block modes, exempciones, bypass por marker, logging, isolation.
Registro en .claude/settings.json: PreToolUse Glob|Grep → acm-enforcement.sh (timeout 3s, statusMessage "ACM enforcement (SE-063)...").

Context

Era 185 arranca con SE-063 derivado de research coderlm (batch 25). Hooks activan el sistema ACM que ya existía pero era ignorado. Default warn-only para permitir observación antes de upgrade a block. Env override SAVIA_ACM_ENFORCE={0,warn,block} para modos. Registro PostToolUse del marker queda pendiente de aprobación explícita del usuario (self-modification guard). Mientras tanto el marker script existe y es testeable pero no se invoca automáticamente — en warn mode no bloquea nada, sólo informa.

Note

Hook counter bumped 56→58 (61 regs con PreToolUse Glob|Grep).

5.75.0 — 2026-04-22

Batch 27 — SE-062.5 Era 184 finale: frontmatter migration cierre.

Changed

docs/propuestas/SPEC-066-enhanced-local-llm.md — convertido a YAML frontmatter (status: IMPLEMENTED, era 174). Inline **Status**: eliminado.
docs/propuestas/SPEC-067-claudemd-diet.md — YAML frontmatter (status: IMPLEMENTED, era 165). CLAUDE.md diet 121→48 líneas ya ejecutado.
docs/propuestas/SPEC-068-hook-enhancement.md — YAML frontmatter (status: SUPERSEDED, superseded_by: SPEC-071). Reemplazado por Hook Overhaul Era 171.
docs/propuestas/SPEC-069-coordinator-mode.md — YAML frontmatter (status: IMPLEMENTED, era 168). Research cerrado en batch Eras 167-170.

Context

Era 184 CERRADA tras 5/5 slices SE-062 completados:

SE-062.1 counter sync (batch 24)
SE-062.2 duplicate SE-056 resolution (batch 24)
SE-062.3 skills aggregator (batch 25)
SE-062.4 changelog workflow activation (batch 26)
SE-062.5 frontmatter finale (batch 27, este)

specs-frontmatter-normalize.sh --scan reporta PASS sin drift. Los 4 specs legacy con **Status**: inline documentados como excepción en batch 8 (SE-054) ahora normalizados. Cero drift frontmatter en 198 specs.

5.74.0 — 2026-04-22

Batch 26 — SE-062.4 CHANGELOG.d consolidation workflow activation.

Added

.github/workflows/changelog-consolidate.yml — GHA workflow post-merge a main, trigger en CHANGELOG.d/**, threshold 20 fragments, concurrency serial, skip marker [skip consolidate] previene loops. Activa scripts/changelog-consolidate-if-needed.sh (implementado batch 7, dormido hasta ahora).
tests/test-changelog-consolidate-workflow.bats — 31 tests (YAML valid, triggers, permisos, safety guards).

Context

SE-062.4 Era 184 slice 4: cierre de SE-053. El script existía desde batch 7 con 25 tests pasando pero sin ningún trigger registrado (no hook, no workflow, no cron). Workflow añade el trigger de activación. Version 5.74.0 asume merge de 5.73.0 (batch 25) primero; rebase ajustará si el orden cambia.

5.73.0 — 2026-04-22

Batch 25 — SE-062.3 skills aggregator + SE-063/064 coderlm-inspired propuestas.

Added

.claude/skills/tier3-probes/ — aggregator skill para 6 feasibility probes (scrapling, oumi, memvid, bertopic, reranker, pdf-extract). SKILL.md + DOMAIN.md.
.claude/skills/workspace-integrity/ — aggregator skill para 7 integrity auditors (claude-md-drift, baseline, catalog-sync, orphan, manifest, size, usage). SKILL.md + DOMAIN.md.
docs/propuestas/SE-063-acm-enforcement-pretool-hook.md — pre-tool hook que bloquea glob/grep amplio sin consulta previa de .agent-maps/INDEX.acm. Effort S 4-6h, prioridad Media.
docs/propuestas/SE-064-acm-multihost-generator.md — generador ACM multi-host (Cursor/Windsurf/Copilot). Effort M 8h, prioridad Baja (on-demand).
output/research-coderlm-20260421.md — investigación técnica coderlm (veredicto ADOPTAR PATRÓN).
ROADMAP Era 185 PROPOSED — agrupa SE-063/064 post-Era 184.

Changed

CLAUDE.md — skills count 83 → 85 tras 2 aggregator skills nuevos.
docs/ROADMAP.md — Era 185 section añadida, SE-063/064 registradas como propuestas.

Context

SE-062.3 usa patrón aggregator para documentar 18 scripts huérfanos sin crear N skill dirs separados. Dos aggregator skills cubren la totalidad de scripts de probes e integrity. SE-063/064 derivan de research coderlm — inspirado en su patrón de hook enforcement + multi-host generator, pero sin adoptar el daemon Rust (superficie innecesaria). Era 184 progresa: 3 de 5 slices SE-062 completados.

5.72.0 — 2026-04-22

Batch 24 — SE-062.1 counter sync + SE-062.2 SE-056 duplicate resolution.

Removed

docs/propuestas/SE-056-python-runtime-sbom-virtualenv-enforceme.md — duplicate de SE-056 (canónico: SE-056-python-sbom-virtualenv.md, referenciado en CHANGELOG batch 11)

Changed

docs/ROADMAP.md header — version bump v5.69.0 → v5.71.0, Era 184 añadida a status

Context

Primeros 2 slices SE-062 (Era 184 hygiene). Drift auditor reportó counter drift pero verificación directa mostró CLAUDE.md/filesystem ya alineados (skills=83). ROADMAP header actualizado. Duplicate SE-056 resuelto dejando el fichero canónico (más detallado, referenciado en batch 11).

5.71.0 — 2026-04-22

Batch 23 — Post-Era 183 drift audit + SE-062 Era 184 proposal.

Added

docs/propuestas/SE-062-era184-consolidation-hygiene.md — 5 slices cortos (12-15h) agrupando deuda identificada tras 22 batches sin ciclo hygiene

Changed

docs/ROADMAP.md — Era 184 añadida en PROPOSED con SE-062 slicing

Context

Drift auditor post-Era 183 identifica deuda compuesta: skills count triple drift, duplicate SE-056, 18 scripts huérfanos de skill docs, CHANGELOG inflación (>8000 líneas), 33 specs PROPOSED sin owner, frontmatter migration incompleta. Era 184 propone hygiene cycle sin features nuevas antes de abrir Era 185.

5.70.0 — 2026-04-22

Batch 22 — Era 183 closure. Tier 3 Champions 5/6 ejecutados.

Changed

docs/ROADMAP.md — Era 183 marcada CLOSED 2026-04-22. Tier 3 status actualizado con checkmarks por champion. Header counters sincronizados con realidad (532/65/83/56).

Context

Cierre formal Era 183 tras 8 batches (#655-662) ejecutando Tier 3 Champions. 5/6 implementados: SE-061 (4 slices completo), SE-035, SE-032, SE-033, SE-041. SE-028 Oumi diferido a Tier 7 por requerir GPU ausente en máquina dev. 249 tests nuevos certified. Todos los skills diseñados con fallback graceful (zero-install default) + opt-in opcional para stack ML completo.

5.69.0 — 2026-04-22

Batch 21 — SE-041 Slice 2. Memvid portable backup wrapper.

Added

scripts/memvid-backup.py — wrapper 3 subcomandos (pack/restore/verify) con SHA256 integrity. Fallback tar-gzip cuando memvid ausente
.claude/skills/memvid-backup/SKILL.md + DOMAIN.md — skill integrable con travel-pack / vault-export
tests/test-memvid-backup.bats — 40 tests certified incluyendo round-trip content preservation

Changed

CLAUDE.md — skills count 82 → 83

Context

Quinto champion Tier 3. Evalua memvid (.mv2) como alternativa a tar-gzip para backup de memoria externa. Slice 2 implementa contrato + tar-gzip con integrity SHA256. Slice 3 pendiente (memvid API real) tras acceptance criteria.

5.68.0 — 2026-04-22

Batch 20 — SE-033 Slice 2. Topic cluster skill + BERTopic wrapper.

Added

scripts/topic-cluster.py — clustering tematico (UMAP+HDBSCAN+c-TF-IDF via BERTopic) con fallback keyword cuando bertopic no instalado
.claude/skills/topic-cluster/SKILL.md + DOMAIN.md — skill invocable integrado con retro-patterns, backlog-patterns, lesson-extract
tests/test-topic-cluster.bats — 37 tests certified

Changed

CLAUDE.md — skills count 81 → 82

Context

Cuarto champion Tier 3. Descubre patrones cross-project que keyword-matching pierde. Validación: fallback-keyword sobre 6 docs (sprint + pr-review themes) produjo 2 clusters correctos sin ML stack.

5.67.0 — 2026-04-21

Batch 19 — SE-032 Slice 2. Reranker skill + wrapper Python.

Added

scripts/rerank.py — cross-encoder reranker con 3 backends: cross-encoder (sentence-transformers+BAAI/bge-reranker-base), fallback-cosine, fallback-identity. Zero-install default via ImportError graceful degradation.
.claude/skills/reranker/SKILL.md + DOMAIN.md — skill invocable documentando integración con memory-recall, savia-recall, cross-project-search
tests/test-rerank.bats — 36 tests certified

Changed

CLAUDE.md — skills count 80 → 81

Context

Tercer champion Tier 3 (post-SE-061, SE-035). Filtra ruido entre embedding retrieval y agent consumption. Validado localmente: sobre 3 candidates con cosine alto para no-relevante, el cross-encoder correctamente priorizó el candidate con menor cosine pero mayor relevancia semántica real.

5.66.0 — 2026-04-21

Batch 18 — SE-035 Slice 2. Mutation audit skill.

Added

.claude/skills/mutation-audit/SKILL.md + DOMAIN.md — skill invocable sobre módulo concreto, detecta tests zombies AI-generated
tests/test-mutation-audit-skill.bats — 33 tests certified (skill estructura, frontmatter, referencias, negative, edge, isolation, coverage)

Changed

CLAUDE.md — skills count 79 → 80 (drift-check compliance)

Context

Segundo champion Tier 3 (post-SE-061). Script probe ya existía (batch 9). Slice 2 añade wrapper skill invocable para /mutation-audit y documenta cuándo usar (sprint-end, post-test-generation) vs cuándo no (cada PR, módulos sin tests). Proyección Q3 2026: 500+ tests AI-generated → necesario auditar zombies.

5.65.0 — 2026-04-21

Batch 17 — SE-061 Slice 4. MCP opt-in template para Scrapling.

Added

.claude/mcp-templates/scrapling.json — entry template con activation steps, compliance block (autoApprove:false, BSD-3, legal note)
tests/test-scrapling-mcp-template.bats — 25 tests certified cubriendo estructura, compliance, integración, negative, edge, isolation

Changed

docs/rules/domain/research-stack.md — sección "MCP opt-in" con 6 pasos de activación (probe → install → audit → copy → no autoApprove → restart)
docs/rules/domain/security-scanners.md — nueva fila "MCP templates (SE-061)" en el catálogo

Context

Cuarto y último slice SE-061. Template NO activa el MCP automáticamente: requiere probe+install+audit explícito. Rule MCP-02 (no wildcard autoApprove) enforzada por test. Completa la cadena SE-061 (probe → fetch → skills → MCP) en 4 slices / 21h planificadas.

5.64.0 — 2026-04-21

Batch 16 — SE-061 Slice 3. Integración Scrapling en skills research.

Added

docs/rules/domain/research-stack.md — cadena de backends (Cache → WebFetch → scrapling-fetch → curl), robots.txt, rate limiting, GDPR, attribution
tests/test-research-stack.bats — 26 contract tests certified

Changed

.claude/skills/tech-research-agent/SKILL.md — sección "Fallback de fetch (SE-061)" con invocación a scrapling-fetch.sh cuando WebFetch falla 403/429/503
.claude/skills/web-research/SKILL.md — sección "Scrapling enrichment (SE-061)" para extracción post-SearxNG

Context

Tercer slice SE-061. Las skills research ya conocen el backend adaptativo con fallback robusto. La cadena documentada en research-stack.md aclara cuándo usar cada backend y las reglas de legalidad (robots.txt, rate limiting, no bypass paywalls).

5.63.0 — 2026-04-21

Batch 15 — SE-061 Slice 2. Scrapling fetch wrapper con fallback a curl.

Added

scripts/scrapling-fetch.sh — wrapper estable sobre Scrapling parser-only. Detecta backend (scrapling|curl), extrae título/status/url_final/text. Flags --selector, --json, --stealth, --timeout.
tests/test-scrapling-fetch.bats — 29 tests certified.

Context

Segundo slice SE-061 (Era 183 Tier 3). Interface estable que prefiere Scrapling cuando está disponible, cae a curl (con user-agent genérico) cuando no. Preparado para integración Slice 3 con tech-research-agent + web-research skill.

5.62.0 — 2026-04-21

Batch 14 — SE-061 Slice 1. Scrapling viability probe (Tier 3 champion #1).

Added

scripts/scrapling-probe.sh — probe determinista VIABLE/NEEDS_INSTALL/BLOCKED. Verifica Python >= 3.10, scrapling, lxml, opcional playwright+chromium (--check-browser).
tests/test-scrapling-probe.bats — 23 tests certified.

Context

Primer slice del champion Tier 3 #1 (Era 183 reprioritization). Probe zero-egress, lectura sólo, exit codes 0/1/2. Slice 2 añadirá wrapper scrapling-fetch.sh.

5.61.0 — 2026-04-21

Batch 13 — Era 182 closure. SE-054 Slice 3 legacy-inline exception + ROADMAP status update.

Added

scripts/specs-frontmatter-normalize.sh — legacy-inline exception: SPEC-NNN files con # SPEC-NNN en line 1 y **Status**: inline son respetados (añadir YAML frontmatter empujaría el header fuera del head -5 de validate-spec).
3 tests legacy (skip scan + apply no-op + non-legacy sigue migrando).

Changed

docs/ROADMAP.md — Era 182 CLOSED 2026-04-21. Tier 0-2 completado en batches 5-12. SE-045 diferido (Enterprise-only scope, #648). SE-054 con 4 excepciones legacy documentadas.

Context

Cierre formal Era 182 (post-audit arquitectónico 2026-04-20). 75h ejecutadas vs 112h planificadas — diferencia es SE-045 (12h) movido a scope Enterprise, y eficiencia superior en SE-043/044/046/047/048/053 (scripts preexistentes parcialmente).

5.60.0 — 2026-04-21

Batch 12 — Era 183 research reprioritization. SE-061 Scrapling champion #1 Tier 3.

Added

docs/propuestas/SE-061-scrapling-research-backend.md — adaptive scraping backend para research agents (4 slices, 21h).
ROADMAP Era 183: reorden Tier 3 Champions — SE-061 > SE-035 > SE-032 > SE-033 > SE-028 > SE-041.

Changed

docs/ROADMAP.md — Tier 3 Champions reordenado por ROI research-stack (SE-061 Scrapling champion #1).

Context

Scrapling (D4Vinci/Scrapling) desbloquea research en sites Cloudflare/DataDome/Akamai que hoy fallan silenciosamente en tech-research-agent + skill web-research. ROI inmediato vs probes sin casos activos. Adopción opt-in (core sin browser), fallback a curl siempre disponible.

5.58.0 — 2026-04-20

Batch 11 — Tier 3 probes + Python SBOM + unified security runner. 4 scripts + 4 BATS suites (82 tests certified).

Added

SE-028 Slice 1 scripts/oumi-probe.sh + 20 tests — oumi framework viability probe.
SE-041 Slice 1 scripts/memvid-probe.sh + 21 tests — memvid portable memory viability probe.
SE-056 Slice 1 scripts/python-sbom.sh + 22 tests — Python imports vs requirements.txt audit + venv hint.
scripts/security-audit-all.sh + 19 tests — unified runner for mcp/permissions/hook/prompt scanners. Graceful degradation when sub-scanners missing.
Spec SE-056 registrada.

Batch 10 — Security stack hardening. Adopt patterns from agentshield (MIT) research. 3 new scripts + 28 new rules + PS-11..PS-14 extension + security-scanners.md catalog.

Added

SE-058 scripts/mcp-security-audit.sh + 28 tests — 11 reglas MCP (supply chain, auto-approve, secrets hardcoded, shell transport, path traversal).
SE-059 scripts/permissions-wildcard-audit.sh + 25 tests — 8 reglas wildcard permissions (Bash/Write/WebFetch sin deny, auto mode + skip prompts, destructive commands).
SE-060 scripts/hook-injection-audit.sh + 25 tests — 9 reglas hook injection (eval unquoted, curl exfil, pipe-to-shell, reverse shell /dev/tcp, sudo sin -n, redirect a credenciales).
Extension scripts/prompt-security-scan.sh con PS-11..PS-14 (zero-width chars, long base64, URL-pipe-shell, time bombs).
docs/rules/domain/security-scanners.md — catálogo unificado del stack.
3 specs: SE-058, SE-059, SE-060.

Context

Research output/research/agentshield-20260420.md identificó gap 77/102 reglas (solo 24% solape). Batch 10 cubre MCP + permissions + hook injection + hidden directives sin adoptar stack TS/Node externo.

5.57.0 — 2026-04-20

Batch 9 — Tier 2 close + Tier 3 champions. SE-050 SPEC-122 skill + SE-057 manifest integrity + SE-032/033 probes. 3 scripts + 1 skill + 3 suites (76 tests).

Added

SE-057 Slice 1 scripts/rule-manifest-integrity.sh + 27 tests — audita INDEX.md size + manifest-filesystem crosscheck.
SE-032 Slice 1 scripts/reranker-probe.sh + 26 tests — viability probe cross-encoder reranker.
SE-033 Slice 1 scripts/bertopic-probe.sh + 23 tests — viability probe BERTopic (UMAP+HDBSCAN).
SE-050 Slice 2 .claude/skills/emergency-mode/ (SKILL.md + DOMAIN.md) — SPEC-122 cierre.

5.56.0 — 2026-04-20

Batch 8 — Tier 2 consolidación. SE-054 frontmatter normalization aplicado (125 specs) + SE-052 agent-size remediation plan. 2 scripts + 2 suites (61 tests, scores 95 y certificado).

Added

SE-054 Slice 2+3 scripts/specs-frontmatter-normalize.sh + 31 tests — normaliza status (case + missing field) + adds id.
SE-052 Slice 1 scripts/agent-size-remediation-plan.sh + 30 tests — hit-list DESC por size + extractable blocks detector + estimated savings.

Fixed

125 specs en docs/propuestas/ normalizados: case drift Proposed→PROPOSED, Draft→DRAFT, Rejected→REJECTED, + added missing status: fields. Verified: --scan drift = 0.

5.55.0 — 2026-04-20

Batch 7 — Tier 1+2 remediation continuación. SE-044 colisión SPEC-110 resuelta + ADR-001. SE-053 changelog post-merge. 2 scripts + 2 suites (50 tests, scores 92 y 88).

Added

SE-044 Slice 1 scripts/spec-id-duplicates-check.sh + 27 tests — detect spec-ID collisions en docs/propuestas/.
SE-053 Slice 1 scripts/changelog-consolidate-if-needed.sh + 23 tests — post-merge wrapper sobre consolidate con threshold.
SE-045 Slice 1 .claude/hooks/session-init-bootstrap.sh — async bootstrap (standalone, not wired — requires your authorization to replace session-init.sh).
docs/decisions/adr-001-spec-110-id-collision-resolution.md — ADR resolución colisión.

Fixed

SPEC-110 colisión resuelta: polyglot-developer REJECTED renombrado a SPEC-126-polyglot-developer-rejected.md. Verified: spec-id-duplicates-check.sh PASS.

Pending (requires user auth or next PR)

SE-045 Slice 2: replace .claude/hooks/session-init.sh con fast-path calling bootstrap (blocked por sandbox self-modification guard).
SE-050 SPEC-122 Slice 2+3.
SE-052 Agent-size remediation (24h).
SE-054 SE-036 frontmatter Slices 2-3 (10h).

5.54.0 — 2026-04-20

Batch 6 — Tier 0 + Tier 1 remediation desde audit 2026-04-20. 4 scripts nuevos + 5 test suites (98 tests totales).

Added

SE-051 Slice 1 scripts/spec-approval-gate.sh + 21 BATS tests — Rule #8 enforcement detecta scripts linkeando specs no aprobados.
SE-046 Slice 1 scripts/baseline-tighten.sh + 21 BATS tests — ratchet auto-tighten, jamás afloja.
SE-047 Slice 1 scripts/agents-catalog-sync.sh + 22 BATS tests — catalog auto-regenerate desde frontmatter. Aplicado: 56→65 agents sincronizados.
SE-048 Slice 1 scripts/rule-orphan-detector.sh + 20 BATS tests — detecta rules sin referencias reales.
SE-043 Slice 1 tests/test-claude-md-drift-check.bats — 14 tests sobre script existente.

Fixed

docs/rules/domain/agents-catalog.md regenerated (drift D5: 56→65 agents).
.ci-baseline/hook-critical-violations.count: 10→5 (audit D6 corregido, margen CI variance).

Priority impact

Tier 0 progress: SE-051 probe en main (Rule #8 fence). SE-045 session-init split pendiente para PR #XXX (riesgo hook crítico, mejor aislado).
Tier 1 progress: SE-043 (tests sobre script existente), SE-046, SE-047, SE-048 cerrados Slice 1. SE-044 pendiente (ADR decisión humana).

5.53.0 — 2026-04-20

Batch 5 — 3 SE specs probes (Slice 1) + audit architecture + 15 new spec stubs from audit + ROADMAP reprioritization.

Added

SE-035 Slice 1: scripts/mutation-audit.sh + 23 BATS tests — mutation testing scaffolding (bash/python/typescript).
SE-039 Slice 1: scripts/test-auditor-sweep.sh + 24 BATS tests — global sweep sobre todos los tests .bats con ranking + compliance pct.
SE-037 Slice 1: scripts/hook-latency-audit.sh + 27 BATS tests — enforcement layer sobre hook-latency-bench con SLA per-tier + BATS coverage check.
15 new spec stubs SE-043 → SE-057 derivados de auditoría arquitectónica.
docs/ROADMAP.md: Era 182 reprioritization Tier 0-7 con ROI post-audit.
output/audit-arquitectura-20260420.md, audit-new-specs-20260420.md, audit-roadmap-reprioritization-20260420.md.

Audit findings (critical)

SE-045 Critical: session-init p50=468ms vs SLA 20ms.
SE-051 Critical: SPEC-123 graphiti merged without approved spec (Rule #8 erosion).
21 desincronizaciones CLAUDE.md / catalog / roadmap / baseline.
27/65 agents violate Rule #22 (>4KB) without remediation plan.
92 scripts sin test BATS nominal (24%).

5.52.0 — 2026-04-20

Batch 3 + 4 consolidado: SLM pipeline scaffolding + SE-020 portfolio-as-graph Slice 5 + SPEC-099 gitagent export + SPEC-102 pdf-extract probe + SPEC-100 GAIA harness. 88 BATS tests nuevos. 3 rule docs traducidos a EN.

Added

SE-020 Slice 5: scripts/portfolio-deps-status.sh + 22 tests — dashboard per-project (upstream/downstream/shared + implicit discovery).
SPEC-099 Slice 1: scripts/gitagent-export.sh + 23 tests — adaptador .claude/agents/{name}.md → gitagent v0.1 (SOUL/RULES/DUTIES/agent.yaml).
SPEC-102 Slice 1: scripts/pdf-extract-probe.sh + 18 tests — probe preconditions opendataloader-pdf (Java ≥11, Maven, PyMuPDF).
SPEC-100 Slice 1: scripts/gaia-benchmark-harness.sh + 25 tests — scaffolding GAIA benchmark (harness-config + prompts-subset + results-template).
EN translations: receipts-protocol.en.md, slm-training-pipeline.en.md, portfolio-as-graph.en.md.

Fixed

Trim EN rule docs para cumplir el límite 150 líneas (test-workspace-structure gate 9).

5.51.0 — 2026-04-18

Auto-resolver de conflictos CHANGELOG+signature en PRs concurrentes. 37 tests. Era 234.

Added

scripts/resolve-pr-conflicts.sh: auto-resolver para conflictos predecibles al mergear PRs concurrentes. Estrategia: git merge origin/main --no-edit, si solo hay conflictos en CHANGELOG.md + .confidentiality-signature + .scm/*, los resuelve automáticamente (CHANGELOG: semver-ordered union con dedupe de link lines; signature: take-theirs + re-sign; .scm: take-theirs + regen). Si hay conflicto en OTRO fichero, exit 3 + aborta merge (humano revisa).
scripts/resolve-all-open-prs.sh: orquestador que itera cada PR abierto vía gh pr list, detecta CONFLICTING/DIRTY y aplica el resolver. Preserva branch original. Flags --dry-run + --no-push.
tests/test-resolve-pr-conflicts.bats: 37 tests — safety, CLI, conflict handling, regression guards (no force-push, no --amend), edge cases. Auditor score 83.

Motivacion

Patrón observado empíricamente en esta sesión: cada merge de PR provocaba conflictos CHANGELOG+signature en todos los PRs abiertos (hasta 4 simultáneos). Resolución manual costó ~5 min cada vez × 4 PRs × 3 oleadas = 60 min de fricción innecesaria. Este tooling elimina el coste: tras cada merge, bash scripts/resolve-all-open-prs.sh deja todos los PRs limpios automáticamente. Único caso en que requiere humano: conflicto REAL en código (diferente archivo).

5.48.0 — 2026-04-18

SPEC-SE-012 Módulos 3 + 4 — signal/noise reduction tooling. 31 tests. Era 234.

Added

scripts/pr-plan-queue-check.sh (Módulo 4): detecta colisiones de versión CHANGELOG entre PRs abiertos antes del push. Fetcha cada PR via gh api contents/CHANGELOG.md?ref=branch, extrae top version, compara con local. Si colisión, sugiere next-free. Graceful skip vía PR_PLAN_SKIP_QUEUE_CHECK=1 o ausencia de gh/jq/red. Bounded timeouts 8-10s.
scripts/pre-push-bats-critical.sh (Módulo 3): runner selectivo de BATS. Mapea files cambiados (hooks/scripts/skills/agents) a sus .bats relacionados vía convención de nombres. Ejecuta solo los relevantes en vez de las 136 suites completas.
tests/test-signal-noise-scripts.bats: 31 tests consolidados (2 scripts). Safety, CLI, graceful degradation, mapping, read-only invariant, negative + edge cases. Auditor score 88.

Motivacion

SPEC-SE-012 lleva merged desde 2026-03 pero Módulos 3 + 4 quedaron pendientes. Módulo 4 ataca el problema concreto reobservado esta sesión: versiones CHANGELOG 5.42/5.43/5.44 colisionaron entre PRs #607/#608/#609 requiriendo merge-resolve manual. Módulo 3 acelera iteración local (solo corre los bats afectados, no 136). Ambos opt-in — no bloquean flujo si red falla.

5.46.0 — 2026-04-18

Ratchet enforcement gates — SE-037/038/039 Slice 3. 30 tests. Era 234.

Added

.ci-baseline/: contadores de violaciones congelados. agent-size-violations.count=27, hook-critical-violations.count=5 (margen +2 por variance timing), bats-compliance-min.pct=95. README documenta pattern ratchet-down-only.
scripts/ci-extended-checks.sh checks #8/9/10: 3 nuevos gates ratchet. #8 agent size (Rule #22), #9 hook latency (SLA 20ms critical p50), #10 BATS auditor compliance floor (opt-in full sweep via BATS_GATE_FULL=1). Gates fallan en REGRESIÓN (current > baseline); emiten hint cuando current < baseline para lock-in de mejora.
tests/test-ci-extended-ratchet.bats: 30 tests — structure, baseline integrity, execution, regression simulada, stale hints, negative cases, edge cases. Auditor score 83.

Motivacion

ROADMAP Tier 1 Slice 3 enforcement. Tras probes baseline quedan 27 agentes + 3-5 hooks sobre SLA. Ratchet pattern aterriza gates SIN bloquear remediation: compliance monotónica no-creciente. Alternative (bloquear inmediatamente) crearía churn; ratchet es convención rustc/typescript para silenciar deuda heredada sin amnistiarla.

5.45.0 — 2026-04-18

Tier 1 probes — hook bench + agent size audit + BATS auditor sweep. 3 scripts + 42 tests. Era 234.

Added

scripts/hook-bench-all.sh: SE-037 Slice 1. Mide p50/p95/p99 latencia de los 56 hooks (runs configurable 1-20). Clasifica critical (session/memory/claude/pre/post/tool) vs analysis con SLAs 20ms y 100ms. Primer run produce baseline: 56 hooks, 4 critical violations, 1 analysis violation.
scripts/agent-size-audit.sh: SE-038 Slice 1. Audita Rule #22 sobre 65 agentes (<4KB). Soporta size_exception: en frontmatter. Primer run: 27 agentes exceden SLA sin excepción (42% del catalogo, 249KB total).
scripts/audit-all-bats.sh: SE-039 Slice 1. Ejecuta test-auditor (SPEC-055) sobre todos los .bats con bounded concurrency (MAX_PARALLEL=5). Primer run: 136 tests, 100% compliance ≥80 score, average 87. Surpresa positiva — suite en buen estado.
tests/test-tier1-probes.bats: 42 tests validando los 3 scripts: safety headers, bash -n, CLI surface, report generation, read-only invariant, bounded concurrency, exit codes, negative + edge cases. Auditor score 90.

Motivacion

ROADMAP §Tier 1 (probes de deuda, read-only, high signal). Consolidados en 1 PR para agilizar — los 3 comparten patrón estructural (script → report → exit code) y son measurement-first antes de remediation. Resultados dan ground truth: SE-037 y SE-038 tienen deuda real medida; SE-039 confirma suite saludable, Slice 2 puede enfocarse en enforcement gate directo.

ROADMAP.md canonical — consolida 3 roadmaps previos + reprioriza con Tier 1-7. Era 234.

Added

docs/propuestas/ROADMAP.md: nuevo único roadmap canónico. Tier 1 (probes deuda read-only) + Tier 2 (champions research con probes blocking) + Tier 3 (seguridad gated) + Tier 4 (PROPOSED maduros) + Tier 5 (enterprise SE-XXX absorbidos, 25 specs) + Tier 6 (convergencias) + Tier 7 (backlog frío). Sección diferida por hardware/humans. DAG de dependencias explícito. Live status.

Changed

docs/propuestas/SAVIA-SUPERPOWERS-ROADMAP.md: status PROPOSED → SUPERSEDED (SPEC-120..124 merged en PRs #592–#594; redirect a ROADMAP.md).
docs/propuestas/ROADMAP-UNIFIED-20260418.md: status LIVING → SUPERSEDED (v1, absorbido en ROADMAP.md; mantener por auditoría).

Motivacion

Consolidación de 3 roadmaps dispersos (SAVIA-SUPERPOWERS 2026-04-17 + ROADMAP-UNIFIED v1 2026-04-18 + savia-enterprise/DEVELOPMENT-PLAN 2026-04-11) que creaban confusión. Un solo ROADMAP.md como source of truth, 2 roadmaps marcados SUPERSEDED con redirect, savia-enterprise DAG absorbido en Tier 5. Reprioritización: Tier 1 son probes read-only que generan ground truth antes de remediation — evita specs zombies (Spec Ops Repetition principle). Tier 5 completado con 10 specs SE-XXX adicionales (002/003/004/006/009/010/014/022/024/027) tras detección de gap.

5.43.0 — 2026-04-18

Debt specs formalizadas (SE-036/037/038/039) — roadmap Wave 4 consolidado. Era 234.

Added

docs/propuestas/SE-036-specs-frontmatter-migration.md: migración de 111 specs sin YAML frontmatter. 4 slices: Implemented confirmados (30) + UNLABELED review humano (40) + Proposed vigente (41) + enforcement gate.
docs/propuestas/SE-037-hook-latency-audit.md: audit latencia 60 hooks bajo SLA 20ms p50 + BATS coverage 10 críticos. Absorbe SPEC-081. Incidents pasados (fork bomb 2026-04-18, 205ms hook PR #595) como evidencia.
docs/propuestas/SE-038-agent-size-audit.md: Rule #22 enforcement — 65 agentes <4KB. Measurement-first (probe) antes de remediation. Enforcement gate ci-extended-checks #10.
docs/propuestas/SE-039-test-auditor-global-sweep.md: barrido SPEC-055 auditor sobre 100+ .bats legacy. Complementa SE-035 mutation testing (ortogonal: forma vs eficacia).

Changed

docs/propuestas/ROADMAP-UNIFIED-20260418.md: Wave 4 table actualizada — cada deuda D1/D2/D3/D4 ahora tiene spec dedicado (SE-036/039/038/037). D5 absorbido por SE-037.

Motivacion

Formalización de deuda tecnica antes de continuar iteración. 4 specs creados. Sin ellos, los items D1-D5 del roadmap eran prosa que se pierde; ahora son entidades trackeables con Purpose/Objective/Slicing/AC/expires. Todos PROPOSED, Rule #8 intacto — pendientes revisión humana.

5.42.0 — 2026-04-18

Spec status normalization tool + 26 tests. Era 234.

Added

scripts/spec-status-normalize.sh: herramienta 3-modos (audit / apply / suggest) para visibilizar status de 156 specs. Detecta gap de 111 specs sin status: field (0 auto-aplicables, 111 requieren migración manual a YAML frontmatter). Heurística --suggest infiere Implemented/Proposed/SUPERSEDED/UNLABELED cruzando CHANGELOG refs + body keywords.
tests/test-spec-status-normalize.bats: 26 tests — 3 modos, idempotencia, sandbox non-destructive, edge cases. Safety verification (set -uo pipefail).

Motivacion

Wave 4 D1 del ROADMAP-UNIFIED-20260418. 76 specs NONE + 35 legacy (total 111 sin status canonico) es deuda de visibilidad — grep-tooling falla silencioso, dashboards mienten. Tool no aplica cambios a 111 specs (riesgo de mal-classify sin review); genera reporte que habilita PR manual de migracion por lotes controlados.

5.41.0 — 2026-04-18

SPEC-082 — orphan skill fix (pr-agent-judge DOMAIN.md) + regression test. Era 234.

Added

.claude/skills/pr-agent-judge/DOMAIN.md: Completa Clara Philosophy dual-doc (skill tenía SKILL.md pero no DOMAIN.md desde SPEC-124). Documenta rol en Code Review Court, cuándo usar/no usar, límites, confidencialidad y outputs.
tests/test-skills-no-orphans.bats: 19 tests validando que SKILL.md count == DOMAIN.md count, no orphans, estructura Clara Philosophy mayoritaria (≥70%). Regression gate para SPEC-082.

Changed

docs/propuestas/SPEC-082-orphan-skill-fix.md: status Proposed → Implemented + applied_at 2026-04-18.

Motivacion

Wave 1 del roadmap unificado (ROADMAP-UNIFIED-20260418.md serie 1 item 4). SPEC-082 era quick win — solo había 1 skill huérfano. Fix trivial + test preventivo es mejor que skill huérfana drift-ing en silencio. 77/77 skills ahora pareados.

5.40.0 — 2026-04-18

Roadmap unificado + SE-035 mutation testing skill. Era 234.

Added

docs/propuestas/ROADMAP-UNIFIED-20260418.md: roadmap único sobre el que Savia itera autónomamente. Consolida 27 PROPOSED + 7 research outputs + savia-enterprise gaps. Clasifica autonomy-viable (A/B/C/D waves) vs hardware-or-humans-required (diferido). Aplica Rule #8 + Spec Ops principles + gate de autonomía.
docs/propuestas/SE-035-mutation-testing-skill.md: skill mutation-audit invocable on-demand. StrykerJS + mutmut + bash wrapper. Feasibility Probe 1.5h blocking. Champion del research Gómez Corio 2026-04-18.

Motivacion

Generación de roadmap unificado sobre el que Savia pueda iterar autónomamente hasta el final. Items que requieren hardware o humanos operativos diferidos a Sección B. ZeroClaw + enterprise billing pipeline no bloquean iteración. PROPOSED — pendiente revisión humana.

5.39.0 — 2026-04-18

SE-032/033/034 — 3 specs propuestos tras research convergente + 35 tests. Era 234.

Added

docs/propuestas/SE-032-reranker-layer.md: cross-encoder reranker sobre memory/knowledge-graph/savia-recall. Patron identificado independientemente en Hands-On LLM cap.8 y Dify api/core/rag. Escrito aplicando principios Spec Ops (McRaven): Purpose separado de Objective, objetivo unico medible (precision@5 >=80%), Feasibility Probe OBLIGATORIO blocking, expires 2 sprints tras approved.
docs/propuestas/SE-033-topic-cluster.md: BERTopic skill para retros/backlog/lessons (Hands-On LLM cap.5).
docs/propuestas/SE-034-workflow-node-typing.md: I/O schemas explicitos en DAG skills (Dify workflow pattern).
tests/test-se-032-specs-format.bats: 35 tests validando estructura Spec Ops de los 3 specs.

Motivacion

Tres research reports autonomos (Spec Ops book, Hands-On LLM repo, Dify OSS) convergieron en la misma mejora: reranker en capa RAG. SE-032 es el champion. Rule #8: PROPOSED — implementacion solo tras revision humana.

5.38.0 — 2026-04-18

MCP overhead audit tool + doctrina + 23 tests. Era 234.

Added

scripts/mcp-audit.sh: auditoria de overhead MCP. Escanea configs globales y per-project. Estima tokens/turn via heuristica MindStudio. Flags: --budget N, --json, --quiet. Emite recomendaciones si supera presupuesto.
docs/rules/domain/mcp-overhead.md: doctrina — MCP tools se reenvian en cada mensaje. Patron canonico: mcp.json vacio + on-demand loading. Checklist 5 puntos.
tests/test-mcp-audit.bats: 23 tests.

Estado pm-workspace

Auditoria ejecutada: 0 tokens/turn de MCPs user-configurados. Diseno on-demand actual ya es optimo.

5.37.0 — 2026-04-18

Bounded concurrency audit + hardening + doctrina + 22 tests. Era 234.

Added

docs/rules/domain/bounded-concurrency.md: doctrina canonica basada en dos outages reales — Bluesky AppView 2026-04-14 (8h, SetLimit(50) comentado) y el fork bomb de pm-workspace 2026-04-18 (15k procesos). Principio: N FIJO / N ACOTADO / N DERIVADO — solo el tercero necesita semaphore explicito. Patron canonico bash con jobs -rp + wait -n + drain final. Auditoria de hooks/scripts + checklist de 5 puntos.
tests/test-memory-prime-hook-bounded.bats: 22 tests — MAX_PARALLEL numerico y acotado, semaphore pattern, drain final, regression guards, edge cases.

Changed

.claude/hooks/memory-prime-hook.sh: hardened a MAX_PARALLEL=5 explicito con semaphore y wait drain final. Antes: bounded implicitamente por --top 3 upstream. Defense-in-depth post Bluesky post-mortem.

Motivacion

Post-mortem Bluesky compartido en sesion (https://pckt.blog/b/jcalabro/april-2026-outage-post-mortem-219ebg2) + fork bomb de esta semana. Auditoria sistematica del workspace: mayormente safe — fork-agents.sh y wave-executor.sh ya usan semaphores, un unico vector endurecido, doctrina escrita para prevenir el patron.

5.36.0 — 2026-04-18

Gate: SCM Freshness (ci-extended-checks #7) + resync + 19 tests. Era 234.

Added

scripts/ci-extended-checks.sh check #7: regenera .scm/INDEX.scm via el generador deterministico, compara SHA-256 antes y despues, falla si stale. No destructivo — restaura el arbol tras fail via git checkout -- .scm/. Mensaje de remediacion incluye el comando exacto para regenerar.
tests/test-scm-freshness-check.bats: 19 tests — estructura, registro del check, invocacion del generador, comparacion sha256, non-destructive behavior, positivos (pasa cuando fresh), negativos (falla cuando stale, mensaje correcto, guard de missing files), edge cases (1-line diff, idempotencia).

Fixed

.scm/: resync a 1008 recursos actuales (drift residual post merges #599/#600). El generador deterministico no elimina el drift — solo lo hace detectable. Este check + la rutina "regenera antes de PR" completa el loop.

Motivacion

PR #599 hizo el generador deterministico (mismo input a mismo output). Pero nada forzaba a commitear la regen tras anadir commands/skills/agents/scripts. El tracked .scm/ en main iba quedando stale silenciosamente. Ahora, cualquier PR que anada componentes sin regenerar falla en CI extended-checks (y localmente en pr-plan G5b). Front-load al autor del cambio.

5.35.0 — 2026-04-18

ast-comprehension refactor al patron RLM (queries tipadas) + 28 tests. Era 234.

Changed

.claude/skills/ast-comprehension/SKILL.md reescrito (142 a 119 lineas, bajo el cap de 150). Inspirado en el paper Recursive Language Models (Zhang, Kraska, Khattab — MIT CSAIL 2025, arXiv:2512.24601) y en el patron del repo coderlm investigado el 2026-04-18 (veredict ROBAR PATRON, ver output/research-coderlm-20260418.md).
- Antes: pipeline monolitico de 3 capas (tree-sitter + native + semgrep) que dumpeaba un JSON enorme de clases/funciones/complejidad por fichero.
- Ahora: 6 queries tipadas con recipes bash concretos — symbol-search, impl, callers, tests, peek, grep-code. Cada una responde una pregunta especifica. El dump monolitico queda como fallback explicito para legacy assessment (referenciado, no inline).
- Anti-patterns explicitos: Read de fichero entero para preguntas sobre 1 simbolo.
- Dato empirico citado: en savia-web (56 call sites de useAuthStore), approach current lee 2077 LoC = 15k tokens; query callers devuelve 200 tokens. Reduccion 75x.

Added

tests/test-ast-comprehension-rlm.bats: 28 tests — presencia de las 6 queries, recipes bash concretos (grep/sed/tree-sitter), filtros anti-falso-positivo (ej. callers filtra definiciones), referencias al research report y al paper RLM, anti-patterns, frontmatter YAML valido, boundaries de tamano.

Motivacion

Reduccion de tokens 10-100x en exploracion de codigo ajeno. Inspirado en coderlm (MIT, Rust server + tree-sitter) pero sin adoptar el binario — solo el patron. Zero nueva infra, zero nueva dependencia. La mejora es puramente de instrucciones al agente.

5.34.0 — 2026-04-18

SCM generator deterministico + resync con 208 a 1008 resources + 20 tests. Era 234.

Fixed

scripts/generate-capability-map.py: el generador embebia date.today().isoformat() en el header de INDEX.scm y resources.json. Cada sesion regeneraba los ficheros con fecha distinta → git status siempre dirty en .scm/ → ruido constante + necesidad de stash antes de cada PR. Fix: content hash SHA-256 truncado a 12 chars reemplaza el timestamp. Mismo input → mismos bytes. Eliminado from datetime import date.
.scm/: resync completo. Los ficheros tracked estaban stale desde 2026-04-06 (208 resources) mientras el repo real tiene 1008 (530 commands, 77 skills, 65 agents, 336 scripts). Regenerados con el hash deterministico.

Added

tests/test-scm-determinism.bats: 20 tests — structure/safety, 3 regression guards (no import date, no date.today(), hash field in header), 3 determinism tests (INDEX.scm + resources.json + categorias byte-identicos en reruns), output shape verification, negative cases, edge cases (boundary mtime-only change no cambia hash).

Motivacion

Root-cause de la friccion constante en PRs recientes: git stash pop para limpiar .scm/ antes de cada commit. Ahora el generador es idempotente. Futuro: drift real (por contenido nuevo en commands/skills/etc) sigue produciendo diff — eso es real y debe commitearse.

5.33.0 — 2026-04-18

pr-plan: 2 nuevas gates (G5b extended-checks, G6b test-quality) + 26 tests. Era 234.

Added

scripts/pr-plan-gates.sh: g5b() ejecuta ci-extended-checks.sh localmente (CHANGELOG version links, rule deps, hook safety flags, agent file size, doc link validation, skills frontmatter). Unos 2s. g6b() ejecuta test-auditor.sh sobre ficheros *.bats AÑADIDOS o MODIFICADOS en el PR (diff-filter=AM), threshold SPEC-055. Skipped si no hay bats cambiados.
scripts/pr-plan.sh: registra G5b "Extended CI checks" entre G5 y G6, y G6b "Test quality (changed)" entre G6 y G7.
tests/test-pr-plan-gates-extended.bats: 26 tests — structure, registration, ordering, negative paths, edge cases (empty diff, boundary, race condition con ficheros borrados), integration smoke tests que sourcean el script real.

Motivación

Lección de PRs previos: CI gates (CHANGELOG refs, test quality) fallaban solo en CI, requiriendo push + fail + re-push. Ahora fallan local en /pr-plan en unos 3s adicionales. Elimina ciclos de retry costosos.

5.32.0 — 2026-04-18

SE-031 slice 3 v2 — azdevops-queries.sh migrado a Query Library + 13 tests integracion. Era 234.

Changed

scripts/azdevops-queries.sh: las 2 WIQL inline (get_sprint_items, get_board_status) migradas a query-lib-resolve.sh --id sprint-items-detailed|board-status-not-done con encoding JSON seguro via jq -n --arg q. Elimina escape-hell de backslashes cuadruple en heredoc (\\\\ → \). Comportamiento idéntico, WIQL canonica.

Added

tests/test-azdevops-queries-migration.bats: 13 tests — resolver snippets, ausencia de WIQL inline, referencias por ID, end-to-end resolver+jq produciendo JSON valido, edge cases de quoting (single + double quotes en params), syntax check, entrypoint preservado.

Notas

Este PR cierra el pendiente slice 3 v2 mencionado en v5.30.0: migracion real del script con 13+ callers. Los snippets ya existian desde slice 3 v1 (v5.30.0); ahora las funciones que usaban WIQL inline leen del fichero canonico. Cualquier command que delegue a get_sprint_items o get_board_status hereda el cambio sin modificaciones adicionales.

5.31.0 — 2026-04-18

SCM (Savia Capability Map) regeneration — Python rewrite + determinístico SessionStart hook. Era 234.

Fixed

scripts/generate-capability-map.sh pasa de 5+ horas a ~2 segundos al reescribir la generación en Python in-process. La versión Bash original spawnaba sed/grep 10k+ veces sobre Git Bash en Windows (fork cost alto). El shell-script ahora es un wrapper de una línea que invoca generate-capability-map.py.
.scm/INDEX.scm estaba desactualizado (generado 2026-04-06, 208 recursos cambiados desde entonces). Regenerado: 991 recursos (530 commands · 76 skills · 64 agents · 321 scripts).

Added

scripts/generate-capability-map.py: implementación Python 3 del generador. Dropped-in replacement del .sh. Documentado con docstrings y normas de código Savia (nombres descriptivos, sin abreviaturas). Genera además .scm/resources.json (mirror machine-readable para hooks/lookups).
.claude/hooks/session-init.sh: hook determinístico que regenera .scm en background si hay recursos en .claude/{commands,skills,agents} o scripts/ más nuevos que .scm/INDEX.scm. Patrón idéntico al del skill manifest (fire-and-forget, ~2s). Garantiza que el mapa esté fresco al arrancar cualquier sesión.

Changed

.scm/categories/*.scm regenerados (7 ficheros: analysis/communication/development/governance/memory/planning/quality).

5.30.0 — 2026-04-18

SE-031 Query Library slice 3 — 3 snippets mas + migracion backlog-groom + 4 tests integracion. Era 234.

Added

.claude/queries/azure-devops/backlog-groom-open.wiql: items abiertos (User Story/Feature/Bug) para grooming por antiguedad.
.claude/queries/azure-devops/sprint-items-detailed.wiql: items del sprint con CompletedWork/RemainingWork/StoryPoints/Activity — para tracking detallado.
.claude/queries/azure-devops/board-status-not-done.wiql: items del board excluyendo Epic/Feature y estados terminales — vista kanban operativa.
4 tests integracion en tests/test-query-lib.bats validando resolve + param substitution de las 3 nuevas queries + migracion del command.

Changed

.claude/commands/backlog-groom.md: WIQL inline reemplazada por call a query-lib-resolve.sh --id backlog-groom-open.

Scope honesto

Slice 3 target spec: "≥5 commands migrados". Entregado: 1 command migrado + 3 snippets canonicos que cubren queries existentes en scripts/azdevops-queries.sh. La migracion del script es deferida — tiene callers multiples (13+ commands) y requiere integration tests dedicados antes de refactorizar call-sites. Los snippets creados dejan el camino sembrado para ese PR de seguimiento.

5.29.0 — 2026-04-18

SE-031 Query Library slice 2 — NL-to-query heuristico deterministico + 23 tests. Era 234.

Added

scripts/query-lib-nl.sh: NL → query ID con pipeline normalizacion + alias expansion ES/EN + F1/Dice scoring + shingle boost + disambiguacion. Exit codes 0 (match), 1 (fallback), 2 (ambiguo), 3 (error). Flags: --lang, --json, --min-score, --topk.
tests/test-query-lib-nl.bats: 23 tests — estructura, validacion input, matching (ES/EN/savia-flow), fallback schema prompt, lang filter, JSON output, threshold tuning, ambiguedad, pipe e2e con resolver.
Docs: seccion "NL-to-query (slice 2)" en docs/rules/domain/query-library-protocol.md con algoritmo documentado (6 pasos).

5.28.0 — 2026-04-18

SE-031 Query Library slice 1 — snippets canonicos + resolver + INDEX generator + 31 tests. Era 234.

Added

.claude/queries/{azure-devops,jira,savia-flow}/: 9 snippets canonicos (5 WIQL, 2 JQL, 2 Savia Flow YAML) con frontmatter id/lang/description/params/returns/tags. Reemplazan WIQL inline disperso en commands.
scripts/query-lib-resolve.sh: resolver por ID con --param, --list, --lang, --json. Exit codes 0/1/2. Warning stderr para placeholders no sustituidos.
scripts/query-lib-index.sh: regenerador determinista de .claude/queries/INDEX.md con modo --check para CI.
docs/rules/domain/query-library-protocol.md: protocolo canonico — formato frontmatter, uso desde commands, hygiene rules, lesson learned del fork bomb.
docs/propuestas/SE-031-query-library-nl.md: spec con 3 slices (library, resolver, NL-to-query).
tests/test-query-lib.bats: 31 tests — structure/safety (5), resolve modes (8), param substitution (4), list modes (6), index generator (5), integration (3). Incluye test de regresion fork-bomb.

Fixed

Fork bomb en query-lib-index.sh: el heredoc python3 <<PY (no quoted) interpretaba backticks del cuerpo python como command substitution bash, ejecutando el script recursivamente (15k+ procesos spawn). Fix: <<'PY' + export REPO_ROOT + os.environ.get. Test de regresion cubre el patron.

5.27.0 — 2026-04-18

Close SPEC-115/122/124 + SE-028 slice 1 — 4 specs cerrados. Era 234.

Added

docs/rules/domain/INDEX.md: SPEC-115 — auto-generated index de 157 rule files categorizados en 34 categorías. scripts/rules-domain-index.sh --check detecta staleness.
scripts/rules-domain-index.sh: generador determinista del INDEX.
docs/rules/domain/emergency-mode-protocol.md: SPEC-122 close — activation criteria, what's preserved/degraded, explicit "no safety relaxation" rule.
docs/rules/domain/slm-pipeline-protocol.md: SE-028 slice 1 — pipeline zero-egress SLM per-project (Unsloth training + oumi eval + Ollama deploy). YAML recipe template.
scripts/slm-synth.sh: SE-028 synth wrapper con graceful fallback si oumi no instalado + zero-egress guard (rechaza cloud deploy targets).
tests/test-context-frozen-check.bats: fix tests 11/12 — usar REPO_ROOT env override en lugar de cd (test isolation).
tests/test-slm-synth.bats: 21 tests (certified) para synth wrapper incluyendo zero-egress violation cases.

Changed

docs/rules/domain/pm-config.md: sección SPEC-124 Code Review Court — COURT_INCLUDE_PR_AGENT, PR_AGENT_VERSION, PR_AGENT_MODEL, PR_AGENT_MAX_LINES.
.claude/agents/court-orchestrator.md: sección "External Judges (SPEC-124)" — policy, aggregation, skip rules.
scripts/context-frozen-check.sh: añadido REPO_ROOT env override para testability.

5.26.0 — 2026-04-18

SE-029/SE-030 P3 — 5 componentes adicionales (classifier, frozen, re-state, thresholds, ablation). Era 234.

Added

scripts/context-task-classifier.sh (SE-029-C): clasifica turns en 6 clases (decision/spec/code/review/context/chitchat) con max_ratio y frozen flag por clase. Heurístico priority-order.
scripts/context-frozen-check.sh (SE-029-F): detecta frozen zones (decision-log, APPROVED/DONE specs, task classes frozen, AC files, stack traces). Exit 1 = frozen.
scripts/context-restate-anchor.sh (SE-029-R): emite re-state anchor markdown cuando ratio > 20:1. Threshold configurable con --force.
scripts/graphrag-quality-gate.sh (SE-030-T): valida metrics.json contra 12 thresholds (NDCG@10, Recall@20, MRR, Cross-Repo Precision, Coherence, Relevance, Completeness, Groundedness, Hallucination, Attribution, Factual, Coherence-gen). 3 phases rollout.
scripts/eval-ablation-run.sh (SE-030-A): seam test — compara full vs ablated metrics, determina si layer añade valor (VALUABLE vs QUESTIONABLE).
docs/rules/domain/graphrag-quality-gates.md: 12 thresholds canónicos + rollout phases.
tests/test-context-task-classifier.bats: 26 tests (certified).
tests/test-context-frozen-check.bats: 25 tests (certified).
tests/test-context-restate-anchor.bats: 21 tests (certified).
tests/test-graphrag-quality-gate.bats: 23 tests (certified).
tests/test-eval-ablation-run.bats: 20 tests (certified).

Total iteration P3: 5 scripts + 5 test files + 115 bats nuevos + 1 doc.

5.25.0 — 2026-04-18

SE-029/SE-030 implementaciones — receipts protocol + distortion metric (bytebell-inspired). Era 234.

Added

docs/propuestas/SE-030-graphrag-quality-gates.md: spec GraphRAG quality gates (receipts + 12 thresholds + source hierarchy + seam tests) — basado en serie bytebell Dic'25-Ene'26.
docs/rules/domain/receipts-protocol.md: protocolo "no proof means no answer" — formato canónico YAML con 7 tipos de receipt (file, spec, decision, commit, pr, test, external).
scripts/context-receipts-validate.sh: validator de receipts — exit 0/1/2 según verified/unverified/broken. Soporta --strict, --json.
scripts/context-distortion-measure.sh: SE-029-M baseline — token-set recall + anchor coverage + distortion D (fórmula 0.4recall + 0.6anchor_cov). Verdict HIGH_QUALITY / ACCEPTABLE / UNACCEPTABLE.
tests/test-context-receipts-validate.bats: 21 tests (quality threshold SPEC-055 passed).
tests/test-context-distortion-measure.bats: 19 tests verificando fórmula, thresholds, aislamiento.

5.24.0 — 2026-04-18

SPEC-121/122/123/124 — implementaciones completas iteración P2 (bats tests + pr-agent agent). Era 234.

Added

.claude/agents/pr-agent-judge.md: 5º juez opt-in del Court (SPEC-124) — wraps qodo-ai/pr-agent, emite handoff SPEC-121 al orchestrator. Activation por COURT_INCLUDE_PR_AGENT=true.
tests/test-localai-readiness-check.bats: 16 tests para SPEC-122 readiness check (certified quality SPEC-055).
tests/test-graph-temporal-ops.bats: 18 tests para SPEC-123 temporal ops (add/invalidate/query con semántica valid_from/invalid_at + filtro por relación).
tests/test-pr-agent-wrapper.bats: 21 tests para SPEC-124 wrapper — graceful fallback, feature flag check, JSON schema, isolation.

Changed

CLAUDE.md: agents count 64 → 65 (pr-agent-judge añadido).

5.23.0 — 2026-04-17

SPEC-110 — Memoria externa canónica parent-relative. Era 110. Auto-load de identidad Savia + usuario activo + MEMORY.md en SessionStart.

Added

docs/propuestas/SPEC-110-memoria-externa-canonica.md: spec completa del store canónico en ../.savia-memory/ (parent-relative, OS-agnostic).
scripts/savia-memory-bootstrap.sh: crea el store e instala symlink relativo .claude/external-memory → ../../.savia-memory/. Idempotente.
scripts/savia-memory-migrate.sh: migración idempotente desde silos internos/externos. Copia nunca mueve; --cleanup-origin opcional.
.claude/hooks/session-init.sh: invoca bootstrap al arranque y reporta Memoria: ../.savia-memory (canónico).
CLAUDE.md: sección ## Usuario activo (SPEC-110) con @imports de active-user.md y external-memory/auto/MEMORY.md.

Changed

.gitignore: excluye .claude/external-memory, .claude/external-memory-target y /.savia-memory/.

Security

shield-maps/ del store canónico chmod 700, aislado del bridge external-memory (N4 local-only).

5.22.0 — 2026-04-17

SPEC-120 implementado + SPEC-121..124 + SE-028 propuestos. Roadmap autónomo SAVIA-SUPERPOWERS. Era 234.

Added

docs/propuestas/SAVIA-SUPERPOWERS-ROADMAP.md: roadmap autónomo con 5 specs (SPEC-120..124) de mejoras seleccionadas por Savia desde el research de repos innovadores 2026.
docs/propuestas/SPEC-120-spec-kit-alignment.md: spec alignment con github/spec-kit.
docs/propuestas/SPEC-121-handoff-convention.md: propuesta handoff-as-function OpenAI SDK pattern (pending impl).
docs/propuestas/SPEC-122-localai-emergency-hardening.md: propuesta LocalAI Anthropic shim para emergency-mode (pending impl).
docs/propuestas/SPEC-123-graphiti-temporal-pattern.md: propuesta temporal edges en knowledge-graph (pending impl).
docs/propuestas/SPEC-124-pr-agent-wrapper.md: propuesta pr-agent como 5º juez del Court (pending impl).
tests/test-spec-template-compliance.bats: 26 tests verifican que el spec template mantiene secciones spec-kit + secciones Savia exclusivas (quality threshold SPEC-055 passed).
tests/test-handoff-as-function.bats: 20 tests para validator de handoff-as-function (SPEC-121).
docs/rules/domain/agent-handoff-protocol.md: protocolo handoff-as-function (SPEC-121) — OpenAI SDK pattern.
docs/propuestas/SE-028-oumi-integration.md: integración oumi (data synth + eval + distillation) como complemento SE-027.
scripts/localai-readiness-check.sh: SPEC-122 — readiness check de LocalAI como Anthropic shim para emergency-mode.
.claude/rules/pm-config.local.md: AUTONOMOUS_REVIEWER configurado.

Changed

.claude/skills/spec-driven-development/references/spec-template.md: añadida sección ## Spec-Kit Alignment con mapping canónico de secciones Savia ↔ spec-kit estándar. Marker spec_kit_compatible: true.
docs/agent-teams-sdd.md: añadida sección "Spec-Kit Alignment (SPEC-120)" con tabla de correspondencia y pointer al test de validación.

Fixed

scripts/ y scripts/lib/: 103 shell scripts recuperados como ejecutables (chmod +x) tras doctor check.

5.20.0 — 2026-04-17

Savia Shield — hardening Capa 0 (proxy API) + autostart hook + gitignore para mask-map.

Added

.claude/hooks/shield-autostart.sh: SessionStart hook que levanta shield-launcher en background si el proxy (puerto 8443) no responde. Fire-and-forget, espera máx 3s. Respeta SAVIA_SHIELD_ENABLED=false.
.claude/settings.json: registra el hook en SessionStart tras session-init.sh.

Fixed

scripts/savia-shield-proxy.py: filtra accept-encoding en headers reenviados (evitaba respuestas gzip que el desenmascarador no puede parsear) y content-encoding en respuestas de upstream. En HTTPError propaga body + headers originales.

Security

.gitignore: excluye directorio de config local con datos sensibles N4.

5.19.0 — 2026-04-17

SPEC-114 — docs alignment post-SPEC-109/111/112/113. Era 234.

Changed

11 READMEs multilingües: counts actualizados 513/56/91 → 532/64/76 (commands/agents/skills). Ficheros: README.md, README.en.md, README.pt.md, README.fr.md, README.gl.md, README.ca.md, README.de.md, README.eu.md, README.it.md, README.es.md, y otros.
.opencode/CLAUDE.md: predictive-analytics → enterprise-analytics (skill viva).
.claude/agents/coherence-validator.md: skills: [coherence-check] → skills: [] (skill eliminada en SPEC-111, agent funciona vía /check-coherence command).
docs/quick-starts/quick-start-qa.md + _en/: skills/coherence-check/ → commands/check-coherence.md.
docs/propuestas/SPEC-046-visual-diff-qa-merge.md: dependencia visual-quality skill → visual-qa command.
.claude/commands/spec-verify-ui.md: comentario "coherence-check" → "check-coherence".

Added

docs/propuestas/SPEC-114-docs-savia-alignment.md: spec completa con acceptance criteria.

Rationale

Audit de docs post-merges identificó 11 READMEs + 6 ficheros con refs stale a skills borradas o counts desactualizados. Drift-check CI (SPEC-109 item 7) cubre CLAUDE.md pero no README/docs cross-ref. Cleanup one-shot.

5.18.0 — 2026-04-17

SPEC-112 + SPEC-113 — adoption de patterns externos (beans + edgequake). Era 234.

Added

scripts/agent-journal.sh: JSONL append-only journal para agent-runs autónomos (SPEC-112). Inspirado en henriquebastos/beans. Ruta: output/agent-runs/{date}/journal.jsonl. Soporta append|tail|list.
docs/propuestas/SPEC-112-agent-journal-ready-queue.md: spec completa con acceptance criteria.
docs/propuestas/SPEC-113-graph-query-modes.md: spec de query modes inspired by raphaelmansuy/edgequake.

Changed

.claude/commands/graph-query.md: añade flag --mode=local|global|hybrid|bypass (SPEC-113). Default local = backward compatible.
.claude/commands/flow-sprint-board.md: añade flag --ready que filtra PBIs sin bloqueos (SPEC-112).

Rationale

Investigación tras merge de SPEC-111 identificó dos repos con patterns aplicables. Adopción selectiva:

Beans: journal JSONL + ready-queue (compatible con feedback_session_journal.md, Rule #24, autonomy safety).
EdgeQuake: query modes explícitos (guían al LLM, mejoran precisión de traversal).

Rechazados: filosofía "no hooks" de Beans (choca con feedback_friction_is_teacher.md), infra pgvector+AGE de EdgeQuake (viola zero-dep startup).

5.16.0 — 2026-04-17

SPEC-111 Debt cleanup — item 1 (polyglot-developer decision). Era 234.

Changed

docs/propuestas/SPEC-110-polyglot-developer.md: status PROPOSED → REJECTED tras análisis de feasibility. Cierra la deuda con decisión razonada en vez de implementación.

Rationale

Tras análisis: consolidar 12 *-developer agents en uno es diseño incorrecto, no solo "alto riesgo":

La "duplicación" percibida es expertise discreto por lenguaje (convenciones, comandos, linters).
Un agent polyglot cargaría 12× más tokens de prompt por invocación — Opus 4.7 rinde mejor con contextos focalizados.
Routing por nombre es arquitectura válida, no deuda.
La redundancia real es ~240 líneas (no 1500 como el audit sugirió).

Alternativa futura de menor alcance: extraer boilerplate común (~20 líneas por agent) a fragmento importado. Solo si el ahorro justifica el trabajo.

5.15.0 — 2026-04-17

SPEC-111 Debt cleanup — item 2 (orphan skills deletion). Era 234.

Removed

15 orphan skills confirmados sin uso real tras refinar skills-usage-audit.sh: coherence-check, google-chat-notifier, google-drive-memory, google-sheets-tracker, headroom-optimization, non-engineer-templates, plugin-packaging, pm-mcp-server, postmortem-training, predictive-analytics, sdlc-state-machine, semantic-memory, session-recording, skills-marketplace, visual-quality.
Entradas correspondientes en .claude/skill-manifests.json eliminadas.

Fixed

scripts/skills-usage-audit.sh: patrón de detección ampliado — ahora reconoce referencias en backticks (`skill-name`), referencias en prose ("skill X", "skill: X"), y excluye docs/audits/ del scan. El audit anterior falso-positivó 9 skills que SÍ estaban referenciadas.
Referencias cruzadas en DOMAIN.md de skills vivas actualizadas para no apuntar a skills borradas.

Changed

CLAUDE.md: skills count 91 → 76.
.claude/skill-manifests.json: total_skills 91 → 76.

Rationale

Audit original (SPEC-109 action 10) marcó 24 skills como orphan, pero el patrón era demasiado estrecho. Audit refinado: 15 confirmados orphan, 9 FALSE POSITIVES recuperados (executive-reporting, devops-validation, risk-scoring, rules-traceability, savia-hub-sync, evaluations-framework, reflection-validation, context-optimized-dev, context-caching). Lección: audits destructivos requieren grep con tolerancia amplia.

5.13.0 — 2026-04-17

SPEC-111 Debt cleanup — item 3 (hook perf CI gate). Era 234.

Added

.github/workflows/ci.yml: nuevo job hook-perf-gate que ejecuta hook-latency-bench.sh --strict en cada PR. Falla CI si cualquier hook excede 200ms. Previene regresión de performance en hooks.
scripts/hook-latency-bench.sh: flag --strict que hace exit 1 si algún hook supera el threshold.

Rationale

Sin gate automático, la performance de hooks podía degradarse sin alerta. Baseline actual: 55 hooks, max ~124ms (scope-guard.sh), session-init 121ms. Threshold 200ms da margen razonable mientras evita regresión severa. Target a largo plazo: bajar ambos hooks a <100ms.

5.12.0 — 2026-04-17

SPEC-111 Debt cleanup — item 4 (CI signature env bug). Era 234.

Fixed

scripts/confidentiality-sign.sh: reemplazado git diff base..HEAD por git ls-tree -r HEAD (content-addressed blob SHAs). Elimina divergencia local/CI del hash de firma — git diff tenía dependencias no deterministas (merge-base volatility tras merges paralelos, formato de diff sensible a entorno). Nuevo enfoque: firma aprueba un estado de árbol específico; cualquier cambio en ficheros trackeados invalida firma; rebase/merge que no toca ficheros trackeados preserva firma.

Rationale

Post SPEC-109, 5 PRs consecutivos requirieron --admin merge por fallos en "Verify Audit Signature". Root cause: hash del diff variaba entre local (donde se firma) y CI (donde se verifica) incluso para el mismo commit. El fix usa objetos git content-addressed que son estables por construcción.

5.10.0 — 2026-04-17

SPEC-109 Savia Self-Excellence — action 7 (drift-check CI). Era 234.

Added

scripts/claude-md-drift-check.sh: verifica que los conteos en CLAUDE.md coincidan con realidad. Exit 2 si drift, 0 si match. Integrado en readiness-check.sh. Previene el patrón histórico que el audit 7.2/10 detectó.

Changed

CLAUDE.md: counts sincronizados a valores actuales (commands 532, skills 91 tras merges).
scripts/readiness-check.sh: añadido check critical que corre drift-check al inicio de sesión.

5.9.0 — 2026-04-17

SPEC-109 Savia Self-Excellence — action 6 (model canonicalization). Era 234.

Changed

27 agents: model: short forms (opus, sonnet, haiku, inherit) normalizados a canonical (claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5-20251001).
Distribución final: 25 Opus + 36 Sonnet + 3 Haiku = 64 agents con modelo explícito canónico.

5.8.0 — 2026-04-17

SPEC-109 Savia Self-Excellence — acciones 4-5 (identity consolidation). Era 234.

Changed

docs/rules/domain/radical-honesty.md: marcado explícitamente como canonical source. Añadida sección de interacción con emotional-regulation.md (no hay conflicto — axes diferentes). Otros archivos ahora referencian vía @-import sin duplicar principios.
docs/rules/domain/critical-rules-extended.md:24: Rule #24 compactada — quita duplicación de principios, delega en canonical source.
.claude/profiles/savia.md: 223 → 109 líneas (51% reducción). Modo Agente extraído a .claude/profiles/savia-agent-mode.md (carga bajo demanda cuando role: "Agent").

Added

.claude/profiles/savia-agent-mode.md: Protocolo machine-to-machine en fichero separado. 81 líneas, YAML/JSON estructurado, status codes, cero narrativa.

5.7.0 — 2026-04-17

SPEC-109 Savia Self-Excellence — acciones 1-3 (drift + contradicción + emoji). Quick wins del audit 2026-04-17 (score 7.2/10). Era 234.

Fixed

CLAUDE.md: counts actualizados — agents(56→64), commands(513→533), hooks(55→59 registros), skills(91→92). El auditor drift-auditor se aplica ahora a sí mismo.
.claude/profiles/savia.md:222: contradicción tono base resuelta — "profesional-cercano" → "profesional-directo" (match con línea 17, tono canónico).
docs/rules/domain/autonomous-safety.md:82: emoji FAIL eliminado (violaba Rule #24 no-emojis). Sustituido por "ERROR:".

5.6.0 — 2026-04-17

Restore executable bit on .claude/hooks/block-branch-switch-dirty.sh. Hook was committed with mode 644 and triggered "Permission denied" on every bash call (non-blocking but noisy). Git mode change only (100644 → 100755). Era 234.

Fixed

.claude/hooks/block-branch-switch-dirty.sh: chmod +x via git update-index --chmod=+x

5.5.0 — 2026-04-16

Savia Shield hardening for Windows + BATS test fixes. Era 233.

Fixed

scripts/savia-shield-daemon.py: whitelist now requires directory

prefixes (scripts/, hooks/, tests/), closing a path traversal bypass where ../data-sovereignty-fake.md was incorrectly whitelisted. Cross-write detection added to daemon (catches credential splits across writes) with Windows cygpath resolution for Git Bash /tmp/ paths. NER all-words filter: multi-word entities now pass if every constituent word is in the allowlist.
scripts/savia-shield-proxy.py: /health endpoint returns local status

instead of forwarding to Anthropic (which returned 502 proxy error).
scripts/shield-ner-allowlist.txt: added connection string terms

(Server, Password, Database, Localhost) and Lorem ipsum terms that were triggering NER false positives as PERSON entities.
tests/test-savia-shield-daemon.bats: daemon requests now include

X-Shield-Token; private key test fixed (broken bash quoting).
tests/test-data-sovereignty.bats: type name and audit log grep

accept both daemon (github_pat, BLOCK) and fallback variants.
tests/test-data-sovereignty-extended.bats: SEC-021 Ollama mock tests

force fallback mode; SEC-005 cross-write forces fallback for Windows paths.

83/83 BATS tests pass (was 78/83 before this fix set).

5.4.0 — 2026-04-16

SPEC-108 proposal — Agent Self-Improvement Loop + Sentry Root Cause Analysis Pipeline. Inspired by Rakuten QA case study (two patterns adapted to existing pm-workspace infrastructure). Review-first PR. Era 246.

Added

docs/propuestas/SPEC-108-agent-self-improvement-sentry-rca.md:

2-part spec (~16h, 2 sprints). Part 1 extends post-tool-failure-log.sh to auto-write lessons in public-agent-memory/{agent}/MEMORY.md when the same error pattern repeats 3+ times across sessions (pattern-hash + counter + sanitization + FIFO cap of 30). Part 2 adds /sentry-rca <id> command that auto-generates root cause analysis from Sentry stack traces, enriches with comprehension reports, and validates through Truth Tribunal (SPEC-106) before delivery. 5 inviolable constraints per part. 3 decisions pending human input.

Why

Rakuten case study documents "agent memory enabling self-improvement across sessions" and "production exception agent with root cause analysis." pm-workspace has the components (agent-memory-isolation, self-improvement rule, sentry-bugs, error-investigate, comprehension-report, Truth Tribunal) but lacks the wiring. This spec connects existing infrastructure without new abstractions.

5.3.0 — 2026-04-16

SPEC-106 Phase 3 — Truth Tribunal calibration harness + operations guide. Closes the spec: all 3 phases now implemented. Era 245.

Added

scripts/tribunal-benchmark.sh: deterministic calibration

harness with subcommands run, sample, metrics. Generates a 6-case sample dataset (one per profile + compliance-gate-override case) and validates the aggregation layer (weights, thresholds, veto rules) end-to-end. Reaches 100% accuracy on the sample. Writes per-case JSONL results.
tests/fixtures/truth-tribunal-bench/: 6 labelled cases shipped

as the default benchmark dataset. Each case contains report.md (with report_type: frontmatter), expected.yaml (ground truth verdict + profile), and 7 synthetic per-judge YAML outputs.
docs/tribunal-guide.md: operations guide covering sync vs

async usage, profile detection rules, verdict math, veto rules, compliance-gate override, calibration loop (deterministic + manual human-in-the-loop), troubleshooting matrix, and honest limits.
tests/test-tribunal-benchmark.bats: 19 BATS tests covering

sample generation, deterministic verdict matching, compliance-gate override, results-file output, JSON validity, metrics computation, and idempotency. Auditor certified.

Why

The aggregation layer is the part of the tribunal we can validate without spending money on real LLM calls. The harness catches drift from weight changes or threshold tweaks. Real-judge calibration (human-in-the-loop) is documented in docs/tribunal-guide.md as a manual procedure — automating it would reintroduce the LLM-judging- LLM problem the tribunal exists to solve.

Spec status

SPEC-106 closed. All 3 phases implemented:

Phase 1 (sync MVP): v4.88.0, PR #571
Phase 2 (async hooks): v4.91.0, PR #573
Phase 3 (calibration): v4.92.0, this PR

5.2.0 — 2026-04-16

SPEC-106 Phase 2 — Truth Tribunal async hook integration. Adds the async PostToolUse pipeline that auto-queues report verifications when the assistant writes a report file, plus a queue worker and a dashboard command. Era 244.

Added

.claude/hooks/post-report-write.sh: async PostToolUse hook on

Edit|Write that detects report-like markdown (path heuristics under output/audits|reports|postmortems|governance|compliance|dora plus filename patterns like ceo-report-*, *-digest*, compliance-*, audit-*, plus frontmatter override report_type:). Self-recursion guards skip .truth.crc and queue files. Idempotent: skips if a fresh cached verdict already exists. Never blocks the write.
scripts/truth-tribunal-worker.sh: queue worker with subcommands

process [--max N], status, clean, enqueue <report>. Atomic claim via .req → .work → .done/.fail rename. Writes a .truth.pending marker next to the report (judge agent invocation must run inside Claude Code session, not the worker — Phase 2 stages the work, the user runs /report-verify to convene the tribunal).
.claude/commands/tribunal-status.md: /tribunal-status dashboard

showing queue depth, pending markers, and recent verdicts. Optional --process N and --clean flags delegate to the worker.
tests/test-truth-tribunal-phase2.bats: 24 BATS tests covering

hook heuristics, self-recursion guards, idempotency, worker subcommands, and pending-marker structure. Auditor certified.

Changed

.claude/settings.json: registered post-report-write.sh in the

PostToolUse Edit|Write matcher with async: true, 5s timeout.
docs/propuestas/SPEC-106-truth-tribunal-report-reliability.md:

status updated to "Phase 1 + 2 Implemented".

Phase 2 honest limit

The worker stages reports as .truth.pending instead of invoking the 7 judge agents directly. Reason: judge agents are Claude Code agents that run inside an active session with API access. The worker runs outside that context. Phase 2.5 (future) can plug in an in-session orchestrator that watches the queue and converts pending markers into full tribunals without manual /report-verify invocation.

4.98.0 — 2026-04-15

Shield NER improvements — expanded filters + allowlist + persistent launcher

defensive hook for uncommitted branch switch. Era 232.

Added

scripts/shield-launcher.py: persistent start/stop/status/restart for

Shield daemon + proxy. Windows uses DETACHED_PROCESS + CREATE_NO_WINDOW so the processes survive terminal close. Unix uses start_new_session=True. PID tracking in ~/.savia/. Stderr to ~/.savia/{label}.log for debugging.
scripts/shield-ner-allowlist.txt: allowlist for public domains,

common English names, and code patterns that NER was false-flagging as personal data.
.claude/hooks/block-branch-switch-dirty.sh: new PreToolUse hook on

Bash that blocks git checkout <branch> / git switch <branch> when the working tree has uncommitted changes (tracked). Untracked files continue to follow the checkout (git default). Prevents silent loss of in-progress work. Use git stash -u or commit before switching branches.
tests/test-shield-ner-allowlist-domains.bats: BATS coverage for the

new allowlist and filters.

Changed

scripts/savia-shield-daemon.py + scripts/savia-shield-proxy.py:

NER filter expansion — new filters 0b (snake_case identifiers), 0e (keyword arguments), 0f (truncated string literals). Unblocks Savia startup when NER was misclassifying code patterns as sensitive data.
.claude/hooks/session-init.sh: auto-start Shield launcher at session

start if the daemon is not already running. Detached so terminal close does not kill it.

Fixed

False positives in NER for common English names, public domains, and

code patterns (snake_case, kwargs, truncated literals).
Shield daemon startup race where HTTP hook ran before daemon was ready.

4.97.0 — 2026-04-15

SPEC-107 proposal — AI Cognitive Debt Mitigation. Evidence-backed spec to measure and counter "AI brain fry" in heavy Claude Code users with 5 ranked interventions (hypothesis-first commit, teach-back gate, critical-eval checklist, dependency telemetry, weekly retrieval drill). Review-first PR. Era 243.

Added

docs/propuestas/SPEC-107-ai-cognitive-debt-mitigation.md:

3-phase implementation plan (~32h) grounded in MIT Media Lab "Your Brain on ChatGPT" (arXiv 2506.08872, 54 EEG participants, 83% recall failure), Microsoft+CMU CHI 2025 critical-thinking study (319 knowledge workers), ICER 2025 longitudinal Copilot study (differential metacognitive atrophy), and Roediger-Karpicke 2006 retrieval practice (+50% retention vs re-reading). Includes 4 inviolable restrictions (CD-01..04), N3 privacy guarantees, integration matrix with existing wellbeing components (wellbeing-guardian, burnout-radar, code-comprehension, dev-session-protocol, verification-before-done), honest limits ("no measurement without EEG"), and 4 explicit decisions pending human input.

Why

8h/day with Claude Code is the operative scenario for this workspace. Existing wellbeing components track schedule and team aggregates but no component measures individual cognitive offloading or AI dependency. The 5 interventions are ranked by evidence strength; first 3 form the Phase 1 MVP (telemetry + hypothesis-first warning).

Phase 1 scope (proposed)

Opt-in telemetry + hypothesis-first commit trailer in warning mode. Phase 2 activates teach-back gate and critical-eval checklist as blocking hooks. Phase 3 calibrates thresholds with 30-day real data.

4.95.0 — 2026-04-16

Migration of workspace rules from .claude/rules/ to docs/rules/. Rules are documentation artefacts and belong under docs/. Lazy-loading behaviour preserved: .claudeignore continues to exclude rules from auto-context, so they only load when explicitly @-referenced. Era 244.

Changed

Moved: .claude/rules/domain/*.md → docs/rules/domain/*.md (151 rules)
Moved: .claude/rules/languages/*.md → docs/rules/languages/*.md (14 packs)
URL sanitisation: 413 files updated (agents, commands, skills, docs, hooks, scripts, tests)
.claudeignore: now excludes docs/rules/{domain,languages}/
scripts/rule-usage-analyzer.sh: RULES_DIR points to docs/rules/domain/; fixed \d → d regex bug
Hooks dual-pattern (prompt-injection-guard.sh, validate-layer-contract.sh,

agent-hook-premerge.sh, memory-auto-capture.sh, data-sovereignty-gate.sh): match both docs/rules/ (new) and .claude/rules/ (kept for git-ignored pm-config.local.md)
.claude/compliance/checks/check-file-size.sh: regex covers docs/rules/ paths
sovereignty-auditor/SKILL.md: fixed pre-existing broken ref

@docs/rules/domain/cognitive-sovereignty.md → @docs/rules/domain/ai-governance.md

Added

tests/structure/test-rule-migration-audit.bats: 15 new tests —

no stale .claude/rules/{domain,languages}/ refs, all @docs/rules/ refs resolve, hook coverage, tier1 identity (radical-honesty + autonomous-safety)
docs/rules/domain/rule-manifest.json: generated manifest — 151 rules,

2 tier1, 50 tier2, 99 dormant

Kept

.claude/rules/pm-config.local.md — git-ignored local config stays at original path
Existing tier counts: tier1=2 (radical-honesty, autonomous-safety) unchanged

4.94.0 — 2026-04-16

Upgrade all Opus agents and workspace configuration to Claude Opus 4.7 (claude-opus-4-7), released 2026-04-16. Same pricing, same 1M context. Key gains: 21% fewer document reasoning errors, better instruction following, enhanced vision (3.75MP), improved autonomous capabilities. Era 243.

Changed

config/model-capabilities.yaml: added claude-opus-4-7 entry (4.6 kept for backward compat)
pm-config.md: CLAUDE_MODEL_AGENT upgraded to claude-opus-4-7
12 Opus agents: frontmatter model: field updated to claude-opus-4-7
agents-catalog.md, consensus-protocol.md: model references updated
PR scripts: Co-Authored-By updated to Opus 4.7
Docs, READMEs, AGENTS-INDEX: all opus-4-6 references migrated
BATS tests: verify both opus-4-7 and backward-compat opus-4-6

4.88.0 — 2026-04-15

SPEC-106 Phase 1 — Truth Tribunal MVP. Seven independent judges evaluate report reliability with fresh context per judge, weighted aggregation by report type, and absolute veto rules for compliance/PII. Era 242.

Added

scripts/truth-tribunal.sh: orchestration helper with subcommands

detect-type, detect-tier, weights, aggregate, verdict, cache-check, cache-store. Hardcoded weights synced with the rule file. SHA256 cache with 24h TTL. Aggregates 7 per-judge YAML outputs into a single .truth.crc artifact via python3 weighted scoring.
7 judge agents under .claude/agents/: factuality-judge (Opus),

source-traceability-judge (Sonnet), hallucination-judge (Opus), coherence-judge (Sonnet), calibration-judge (Sonnet), completeness-judge (Sonnet), compliance-judge (Opus). Each declares its YAML output schema, scoring rubric, and veto conditions.
.claude/agents/truth-tribunal-orchestrator.md (Opus L2):

convenes the 7 judges in parallel via fork pattern, applies vetos, computes weighted consensus, and emits the canonical .truth.crc.
docs/rules/domain/truth-tribunal-weights.md: weights table

for 6 profiles (default, executive, compliance, audit, digest, subjective) with auto-detection and frontmatter override.
.claude/commands/report-verify.md: /report-verify <report>

slash command — invokes the orchestrator, shows verdict banner, and blocks delivery on ITERATE/ESCALATE.
tests/test-truth-tribunal.bats: 31 BATS tests covering all

subcommands, profile detection, verdict thresholds, abstention handling, compliance gate override, and cache TTL. Auditor score 87.

Verdicts

PUBLISHABLE (high score, no vetos)
CONDITIONAL (mid score, no critical vetos)
ITERATE (low score or any veto)
ESCALATE (after 3 iterations still failing)
NOT_EVALUABLE (≥4 abstentions)

Phase 1 scope

Manual invocation via /report-verify. Phase 2 will add async hook integration and iteration loop. Phase 3 will calibrate weights against a benchmark harness.

4.87.0 — 2026-04-15

SPEC-098 workspace bundle — nidos.sh gains dev-server lifecycle. Extend parallel terminal isolation with fast dev-server management across 12 language packs. Era 241.

Added

scripts/nidos-dev-lib.sh: dev-server library for nidos. Provides

detect / start / stop / url / logs. Auto-detection for Angular, Next.js, Vite, Django, FastAPI, Spring, Go, Rust, .NET, Laravel, Rails plus CLAUDE.md override (DEV_SERVER_COMMAND, DEV_SERVER_PORT, DEV_SERVER_READY). Ports auto-resolve on conflicts. State persists in <nido>/.dev-server/.
nidos.sh dev <name> {start|stop|url|logs}: new dispatcher.
tests/test-nidos-dev.bats: 28 BATS tests covering detection,

lifecycle, failure modes, edge cases.

Changed

scripts/nidos.sh: remove now calls dev_stop before removing

the worktree (NIDOS-DEV-02 — no zombie processes after cleanup).
scripts/nidos-lib.sh: usage text lists the new dev subcommand.
Status Proposed → Implemented for SPEC-098.

Why

Before: agents that needed a live dev server (visual-qa, web-e2e-tester, frontend-developer) had to ask the human to start it manually and pass the URL. With nidos.sh dev current url, the URL auto-discovers. Inspired by vibe-kanban's workspace bundle — adapted to our CLI/file-based model.

4.86.0 — 2026-04-15

Tier 1 roadmap sweep — close 6 more implemented specs via verification, add missing BATS test for SPEC-089 (memory-stack-load). Era 240.

Added

tests/test-memory-stack-load.bats: 21 BATS tests for SPEC-089

memory stack L0-L3 loader. Validates budgets, progressive loading, graceful degradation, edge cases.

Changed

Status Proposed → Implemented for 6 specs all verified via existing

scripts and tests:
- SPEC-086 proactive-context-budget
- SPEC-089 memory-stack-l0l3 (new BATS test added)
- SPEC-090 temporal-knowledge-graph
- SPEC-094 heat-based-parallelism
- SPEC-095 competitive-architects
- SPEC-096 blocker-as-context

Why

Six specs were implemented across previous sprints but never marked as such. The roadmap discrepancy between listed Proposed and actually pending has been corrected. Missing BATS suite for SPEC-089 added at quality threshold. Remaining Proposed: SPEC-085 (savia-web, out-of-scope here) plus the four strategic specs added from external research analysis.

4.85.0 — 2026-04-15

Roadmap reprioritization — close 6 implemented specs, add SAVIA-GENESIS recovery doc + script, add 3 strategic specs (workspace bundle, gitagent export adapter, GAIA benchmark) from external research analysis (vibe-kanban, gitagent, AutoAgent). Era 239.

Added

docs/SAVIA-GENESIS.md: dual-purpose recovery and best-practices

document. Reader 1: a clean Claude instance can rebuild Savia from this single file. Reader 2: humans get principles of context engineering and agentic programming. Includes 11 parts + 2 appendices, 7 immutable principles, recovery playbook, 10 best practices each for context engineering and agentic programming.
scripts/recover-savia.sh: launcher that creates an isolated

sandbox OUTSIDE the repo, copies SAVIA-GENESIS.md as initial context, and launches a clean Claude session with READ-ONLY access to the broken pm-workspace. The recovery Claude proposes fixes; humans apply via /pr-plan.
tests/test-spec-088-pair-integrity.bats: 13 BATS tests

validating the SPEC-088 inviolable rule (tool_use ↔ tool_result pairs) is documented in canonical locations + 4 algorithmic simulator cases.
tests/test-recover-savia.bats: 16 BATS tests for the recovery

script — syntax, sandbox isolation, exit codes, prompt invariants.
docs/propuestas/SPEC-098: workspace bundle (nidos with dev server)
docs/propuestas/SPEC-099: gitagent export adapter (defensive)
docs/propuestas/SPEC-100: GAIA benchmark integration

Changed

Status Proposed → Implemented for: SPEC-087 (tool-result-trim),

SPEC-088 (compact pair integrity), SPEC-091 (optimal band scoring), SPEC-092 (variable consensus weights), SPEC-093 (hardware-aware Ollama), SPEC-097 (compiled agent index). All verified via existing tests.

Why

Pending specs accumulate when nobody verifies what's already done. This release closes 6 (5 of which were silently implemented) and brings 3 new strategic ones — including a defensive adapter against gitagent becoming THE agent definition standard. SAVIA-GENESIS exists because Savia must be reconstructible from text alone — that's principle #1 (data sovereignty) and principle #2 (provider independence) made executable.

4.84.0 — 2026-04-14

Claude Code native integrations — document Auto Mode as complementary defense layer for autonomous modes, extend scheduling commands to cover Routines' three execution modes (cloud/desktop/session) and three trigger types (cron/api/event), add scheduling guide disambiguating /loop, /schedule, Routines and OS cron. Era 238.

Added

docs/scheduling-guide.md: decision tree + table comparing the

four scheduling mechanisms available in pm-workspace. Covers when to use each, their persistence model, minimum intervals, and supported triggers. Includes Auto Mode integration guidance.
Auto Mode section in autonomous-safety.md: documents

claude --enable-auto-mode as a complementary pre-tool-call classifier that blocks destructive actions. Explicit statement that Auto Mode does NOT replace AUTONOMOUS_REVIEWER, agent/* branches, PR Draft or AGENT_MAX_CONSECUTIVE_FAILURES — it adds defense in depth.

Changed

.claude/skills/overnight-sprint/SKILL.md: new prerequisite

row recommending --enable-auto-mode; new section describing Auto Mode as complementary safety net.
.claude/skills/code-improvement-loop/SKILL.md: same Auto Mode

additions.
.claude/commands/scheduled-setup.md: Paso 4 now asks where the

routine should run (Cloud / Desktop Local / Session /loop) with interval limits and capability notes for each backend.
.claude/commands/scheduled-create.md: new `--trigger

{cron|api|event}and--mode {cloud|desktop|session}` options. Restrictions table expanded per-trigger.

Why

Claude Code 2026-03-24 introduced Auto Mode and 2026-04-14 introduced Routines. pm-workspace already used Claude Code Scheduled Tasks natively, but docs treated cron as the only trigger. Users had no reference to disambiguate /loop vs /schedule vs Routines (third parties already publish comparisons). This release closes the doc gap and recommends Auto Mode as complementary safety layer — without rewriting the pm-workspace autonomous-safety gates which remain stricter than the Anthropic classifier.

4.83.0 — 2026-04-14

Postponement Judge — Stop hook that refuses unjustified deferrals in assistant responses. Era 237. Savia is a 24/7 agent; phrases like "we'll leave it for tomorrow" / "lo dejamos para mañana" are human reflexes inherited from training data that do not apply here. When detected WITHOUT a valid blocker (pending human approval, CI wait, rate limit, user stop), the hook blocks the Stop and forces one more iteration to continue the task.

Added

.claude/hooks/postponement-judge.sh: Stop hook. Reads the last

assistant text message from the transcript, normalizes to accent-free lowercase, scans 21 postponement patterns (ES + EN) and 24 justification patterns. Loop-safe: respects stop_hook_active=true and caps at POSTPONEMENT_JUDGE_MAX interventions per session (default 2). Profile: standard.
tests/test-postponement-judge.bats: 24 BATS tests covering

structure, loop prevention, 6 positive (block) cases, 7 negative (allow) cases including all major justification categories, and 2 transcript-parsing edge cases.

Changed

.claude/settings.json: registered postponement-judge.sh in

the Stop hook array (timeout 5s, synchronous — must complete to return the decision).

Why

Unjustified postponement is the clearest signal that the agent has imported a human reflex that costs throughput. A 24/7 agent has no "end of day" — but LLM training corpora are saturated with them. This hook converts that failure mode into a runtime counter: if you're going to stop, say WHY. If the why is missing, continue.

How it works

Stop hooks receive {session_id, transcript_path, stop_hook_active} on stdin. On block, they emit {decision: "block", reason: "..."} on stdout and Claude Code feeds reason back to the agent as an implicit user message, then continues. If a genuine blocker exists, the agent states it on the next turn and the second Stop is allowed (because the pattern matches a justification). If the agent keeps postponing without reason, the per-session counter (max 2) eventually gives up.

4.82.0 — 2026-04-14

Multica patterns v2 — two tactical improvements inspired by the Multica daemon (github.com/multica-ai/multica) task_message table design. Era 236. Extends the session-event-log from v4.81.0 with monotonic seq numbers for catch-up queries, and adds a new session resumption index so agents can pick up work where they left off after disconnect.

Added

scripts/session-resume-index.sh: (agent_type, spec_id) →

(session_id, work_dir, timestamp) mapping in a grep-friendly TSV. Commands: record, lookup, list [--agent|--spec], forget, status. Last write wins per (agent, spec). Storage: ~/.savia/session-resume-index.tsv.
tests/test-multica-patterns-v2.bats: 20 BATS tests covering

seq number behavior (8 tests), session resume index CRUD (11 tests), and integration between both patterns (1 test).

Changed

scripts/session-event-log.sh: every emitted event now includes

a monotonic seq field per session, starting at 1. query accepts two new flags: --since-seq N (catch-up: return events with seq > N) and --session <id> (scope to a specific session file). Backward compatible — existing queries without these flags behave unchanged.

Why

Multica's daemon stores a task_message table that lets disconnected agents resume exactly where they left off. The session-event-log we shipped in v4.81.0 already provided durable events, but lacked two pieces that make resumption cheap: (a) a deterministic ordering per session for catch-up queries, and (b) a lookup index so agents don't scan every log file. Both are added here without breaking the v4.81 API — seq is additive, and the index is a separate script.

Integration

Events emitted by v4.81.0 without seq still parse correctly (queries

skip them when --since-seq is used — awk match returns no seq).
The resume index is opt-in: no existing workflow depends on it yet.

Agents that want to checkpoint their session call record at completion and lookup at startup.

4.81.0 — 2026-04-14

Managed Agents patterns — architectural improvements inspired by Anthropic's Managed Agents engineering post (2026-04-14). Era 235. Three patterns adapted: credential proxy, durable session event log, and validation of the existing stateless brain + lazy provisioning.

Added

scripts/credential-proxy.sh: Vault+Proxy pattern for credential

isolation. Agents call git-push, git-clone, api-call through the proxy; credentials never enter agent context. Append-only audit log at ~/.savia/credential-proxy-audit.jsonl. Sanitizes output to strip tokens from URLs and headers.
scripts/session-event-log.sh: durable append-only session log

that survives /compact. Supports emit, query --type|--last|--since, and recover --session latest for post-crash context reconstruction. Event types: decision, correction, discovery, error, milestone, handoff. Storage: ~/.savia/session-events/{session-id}.jsonl, 30-day retention.
managed-agents-patterns.md: domain rule documenting the 3

patterns, their integration points, and prohibited behaviors.
tests/test-managed-agents-patterns.bats: 22 tests.

Why

Rule #1 ("never hardcode PAT") is a behavioral rule. The credential proxy eliminates the underlying vulnerability class by construction — the agent cannot leak what it never sees. Structural isolation beats behavioral rules. Similarly, /compact destroys Tier C events; the durable session log means decisions and corrections survive compaction.

4.80.0 — 2026-04-14

LLM Wiki pattern improvements inspired by Karpathy's gist. Era 234. Three enhancements to the persistent knowledge base: claim classification, knowledge lint, and automated weekly lint integration.

Added

scripts/knowledge-lint.sh: 6-check health scanner for the memory

store — orphan index entries, unlisted files, missing evidence types, oversized index, stale project memories, duplicate descriptions. Supports --fix for auto-repair.
/knowledge-lint: command to run knowledge base health check.
evidence_type field in session-memory-protocol: sourced,

analyzed, inferred, gap — classifies reliability of each memory.
tests/test-knowledge-lint.bats: 14 tests covering all lint checks,

fix mode, integration with context-rotation, and evidence classification.

Changed

scripts/context-rotation.sh: weekly cycle now runs knowledge-lint.sh

and appends results to the weekly summary report.
session-memory-protocol.md: added Evidence Classification section

documenting the 4 evidence types and their priority for recall.

4.79.0 — 2026-04-14

SE-016 Project Valuation — Business-Case-as-Code. Era 233. Living, agent-maintained business cases per engagement with NPV/IRR computation, risk-adjusted scoring, variance alerts, benefit realization reviews at 90/180/365 days, and kill recommendations when thresholds are exceeded.

Added

business-case-as-code.md: domain rule documenting valuation

schema, variance thresholds, 4 agents, 6 failure modes, NPV formula.
schemas/business-case-frontmatter.schema.json: JSON Schema

validating 13+ required frontmatter fields in business-case.md files.
scripts/case-validate.sh: validation script detecting 6

failure modes (missing assumption source, stale assumptions >90d, risk without probability/impact, benefit schedule without review dates, cost variance without alert, duplicate case IDs).
4 case commands: /case-init, /case-recompute,

/case-review, /case-kill-check.
tests/test-business-case-valuation.bats: 20+ tests covering

script logic, schema integrity, failure mode detection, command docs.

4.78.0 — 2026-04-13

SE-015 Project Prospect — Pipeline-as-Code. Era 232. Sovereign, agent-queryable opportunity pipeline where pursuits live as versioned .md files with BANT/MEDDIC qualification scoring, bid/no-bid audit trails, proposal library reuse, and sales-to-delivery handoff packages.

Added

pipeline-as-code.md: domain rule documenting pursuit schema,

stage gates, 4 agents, 8 failure modes (82 lines).
schemas/pursuit-frontmatter.schema.json: JSON Schema validating

15+ required frontmatter fields in pursuit.md files.
scripts/pursuit-validate.sh: validation script detecting 8

failure modes (missing qualification, bid-decision, handoff, orphan pursuits, duplicate IDs, missing SA role, incomplete scoring, broken library references).
7 pursuit commands: /pursuit-init, /pursuit-qualify,

/pursuit-bid, /pursuit-draft, /pursuit-handoff, /pursuit-close, /pipeline-view.
tests/test-pursuit-pipeline.bats: 36 tests covering script

logic, schema integrity, failure mode detection, and command docs.

4.77.0 — 2026-04-13

SE-030 Skill Self-Improvement Pipeline. Era 231. Detects repeated patterns in skill invocations, proposes new skills with scaffold (SKILL.md + DOMAIN.md), and suggests refinements for skills with high failure rates. The flywheel: 91 skills get smarter with use.

Added

scripts/skill-detect.sh: SE-030 orchestrator with 4 modes

(scan/propose/refine/status). Detects co-occurring skills, high failure rates, and unmatched NL patterns. Generates scaffolds with 50% initial confidence and experimental maturity.
/skill-detect: command for pattern detection and proposals.
tests/test-skill-detect.bats: 17 tests, quality score 98.

4.76.0 — 2026-04-13

SE-032 Cross-Project Lessons Pipeline. Era 231. Three-phase pipeline (extract, catalogue, search) for sharing lessons across projects with PII sanitization. Competitive differentiator: no other PM tool does this.

Added

scripts/lesson-pipeline.sh: SE-032 orchestrator with 4 modes

(extract/search/stats/status). PII sanitization (emails, IPs, connection strings, tokens). JSONL index for fast lookup.
/lesson-extract: extract lesson from current task context.
/lesson-search: search lessons by domain or keyword.
tests/test-lesson-pipeline.bats: 17 tests, quality score 98.

4.75.0 — 2026-04-13

SE-033 Context Rotation + SE-034 Agent Activation Plan. Era 231. Two intelligence specs implemented: automated memory rotation with 3 cycles (daily/weekly/monthly) to enforce 25KB cap, and daily agent activation plan mapping backlog to agents with token budgets.

Added

scripts/context-rotation.sh: SE-033 orchestrator with 4 modes

(daily/weekly/monthly/status). Archives stale session-hot, retires old project memories, generates weekly summaries, enforces 25KB cap.
/memory-rotate: manual context rotation command.
scripts/daily-activation-plan.sh: SE-034 plan generator. Scans

backlog + approved specs, maps to agents via assignment-matrix, allocates token budgets per tier, defers items exceeding budget.
/daily-plan: generate/show/status for daily activation plans.
tests/test-context-rotation.bats: 11 tests for SE-033.
tests/test-daily-activation-plan.bats: 11 tests for SE-034.

Changed

session-init.sh: wired context rotation (daily/weekly/monthly)

as async non-blocking calls at session start.

4.74.0 — 2026-04-13

Multi-language documentation audit and alignment. Era 231. Corrected counters across all 9 READMEs, added 8 Code Review Court agents to catalog, fixed broken CHANGELOG links, translated 4 missing guides to English, and expanded Savia Enterprise README from 11 to 34 specs.

Changed

All 9 READMEs (ES, EN, GL, EU, CA, FR, DE, PT, IT): counters updated

to 513 commands, 56 agents, 91 skills, 55 hooks; Code Review Court section added to Security heading.
CLAUDE.md: structure counters and hooks line aligned to real values.
agents-catalog.md: title and count updated from 49 to 56; 8 Court

agents added (correctness-judge, architecture-judge, security-judge, cognitive-judge, spec-judge, court-orchestrator, fix-assigner, test-architect).
async-hooks-config.md: total hook instances corrected (45 → 55).
pm-workflow.md: command count corrected (505 → 513).
CHANGELOG.md: comparison links for v4.71.0 and v4.65.0 fixed (were

pointing to v4.34.0 instead of their respective predecessors).
docs/propuestas/savia-enterprise/README.md: expanded from 11 specs

to 34 specs across 5 logical groups (Core Platform, Business Operations, Project Lifecycle, Quality & Security, Intelligence & Optimization).

Added

docs/guides_en/guide-emergency-watchdog.md: EN translation.
docs/guides_en/guide-prompt-security-scanner.md: EN translation.
docs/guides_en/guide-spec-quality-auditor.md: EN translation.
docs/guides_en/guide-workspace-consolidation.md: EN translation.

4.73.0 — 2026-04-13

Roadmap specs SE-032..034 from synthesis-console research. Era 231. Three patterns for context lifecycle: cross-project lessons, context rotation (daily/weekly/monthly), and daily agent activation plans with token budgeting.

Added

SPEC-SE-032-cross-project-lessons.md: pipeline for extracting,

cataloguing, and querying lessons across projects before escalating.
SPEC-SE-033-context-rotation.md: automated rotation with 3

cycles (daily session-hot, weekly decisions, monthly consolidation) to keep memory under 25KB budget.
SPEC-SE-034-agent-activation-plan.md: daily plan generated from

sprint backlog mapping items to agents with token budget per tier.

4.72.0 — 2026-04-12

Hook safety audit + registration fixes. Era 231. Fixed 4 blocking hooks that wrote to stdout (corrupting tool output), registered 3 missing hooks (memory-prime, cwd-changed, compress-agent-output), fixed 2 broken hook paths, and added nidos exec bit. Total: 55 hook instances across 17 events.

Fixed

4 stdout corruption bugs: plan-gate, agent-dispatch-validate,

dual-estimation-gate, pre-commit-review now write to stderr.
2 broken hook paths: prompt-injection-guard and delegation-guard

had empty prefix instead of $CLAUDE_PROJECT_DIR.
nidos.sh: missing executable permission (644 → 755).

Added

memory-prime-hook.sh registered on UserPromptSubmit (async).
cwd-changed-hook.sh registered on CwdChanged (new event).
compress-agent-output.sh registered on PostToolUse/Task (async).
2 missing MEMORY.md index entries for feedback memories.

4.71.0 — 2026-04-12

SE-028/029/030/031: Security and quality patterns inspired by Hermes Agent research. Era 231. Prompt injection guard, iterative context compression, skill self-improvement pipeline, and delegation depth enforcement.

Added

.claude/hooks/prompt-injection-guard.sh: PreToolUse hook (security

tier) scans context files for adversarial prompt injection before loading. Detects override attempts, hidden Unicode, HTML comment injection, and social engineering patterns. Audit log in JSONL.
scripts/iterative-compress.sh: deterministic prune (removes

confirmations, separators, UX banners) + iterative structured summary that survives across multiple /compact cycles. Session-hot.md with delta updates instead of full rebuilds.
.claude/hooks/delegation-guard.sh: enforces max delegation depth

of 1 (no grandchild agents). Blocks recursive Agent/Task invocations. Traces all delegations in JSONL.
SE-028..031 specs: 4 specs in docs/propuestas/savia-enterprise/

covering prompt injection, iterative compression, skill self-improvement, and delegation enforcement.
3 test suites: 41 tests total (18 injection, 12 compression,

11 delegation).

4.70.0 — 2026-04-12

SE-027 SLM Training Pipeline for inference sovereignty. Era 231. Local fine-tuning of small language models using project data (N4) with zero cloud egress. Unsloth + TRL stack, GGUF export, Ollama deployment. Integrates with Savia Dual for sovereign routing.

Added

scripts/slm-data-prep.sh: collect project documents, sanitize

PII, format as ChatML (SFT/DPO), validate, split train/eval sets. Manifest with SHA-256 hashes for data lineage.
scripts/slm-train.sh: Unsloth wrapper with auto hardware

detection and model selection. SFT training, GGUF export, Ollama deploy, and RGPD-compliant forget. Model registry at ~/.savia/slm-registry/.
docs/propuestas/savia-enterprise/SPEC-SE-027-slm-training.md:

full spec covering architecture, data pipeline, training methods, governance, and 3-phase implementation plan.
tests/test-slm-training.bats: 22 tests covering data prep

pipeline, PII sanitization, format/validate/split, dependency check, forget cleanup, and edge cases.

4.69.0 — 2026-04-12

SE-006 Governance & Compliance Pack. Era 230. Append-only chain-hashed audit trail for regulated clients (AI Act, DORA, NIS2, GDPR, CRA, ISO 42001). Tamper-evident JSONL with SHA-256 chain verification and markdown export.

Added

scripts/governance-audit-log.sh: append-only audit trail with

chain hash (sha256). Subcommands: append, verify, export (md/json). Each entry links to previous via prev_hash for tamper detection.
.claude/enterprise/rules/governance-compliance.md: rule

documenting 6 compliance frameworks, audit requirements, and retention policies for enterprise tenants.
tests/test-governance-audit-log.bats: 15 tests covering

append, chain integrity, tamper detection, export, and edge cases.

4.68.0 — 2026-04-12

SE-005 Sovereign Deployment implementation. Era 229. Network egress guard hook blocks outbound calls in sovereign/air-gap modes. Rule documents 4 deployment modes, agent sovereign compatibility flags, graceful degradation.

Added

.claude/enterprise/hooks/network-egress-guard.sh: blocks curl, wget,

git push, npm install, etc. in sovereign/air-gap modes. Respects deployment.yaml → network.allowed_hosts exceptions. No-op when module disabled or no tenant active.
.claude/enterprise/rules/sovereign-deployment.md: 4 modes (cloud,

hybrid, sovereign, air-gap), per-tenant deployment.yaml config, sovereign_compatible agent flag, LLM provider abstraction (Ollama, vLLM, llama.cpp, LocalAI), graceful degradation policy.
tests/test-sovereign-deployment.bats: 18 tests, SPEC-055 certified.

4.67.0 — 2026-04-12

SE-011 Docs Restructuring — Slice 1. Era 228. New docs taxonomy, getting started guides (community + enterprise), enterprise overview with module map and dependency graph, GitHub metadata sync script.

Added

docs/getting-started/community.md: first-session guide for Core.
docs/getting-started/enterprise.md: module activation guide.
docs/enterprise/overview.md: what Enterprise is and is not, module

map with implementation status across 4 waves (26 specs), dependency graph.
scripts/sync-github-metadata.sh: idempotent script to update repo

description, topics, and homepage via gh CLI.
New taxonomy directories: docs/{getting-started,core,enterprise,adapters,operations,reference,i18n}/.

4.66.0 — 2026-04-12

SE-010 Migration Path implementation. Era 227. Enterprise module lifecycle manager with 6 subcommands: status, modules, enable, disable, uninstall, migrate-data. All activations opt-in, all reversible, zero residue.

Added

scripts/savia-enterprise.sh: central Enterprise lifecycle command.

Status shows Community/Enterprise mode, modules lists all 16 with ON/OFF, enable/disable toggle individual modules in manifest, uninstall resets all to Community, migrate-data runs module-specific migration wizards.
scripts/lib/enterprise-helpers.sh: sourceable helpers for hooks and

scripts. enterprise_enabled "module", enterprise_mode, enterprise_version.
tests/test-savia-enterprise.bats: 18 tests, SPEC-055 certified.

4.65.0 — 2026-04-12

SE-001 Foundations + PAT-strip security fix. Era 225.

Added

.claude/enterprise/extension-points.md: 6 formal extension points

documented in the canonical Enterprise directory.
.claude/enterprise/manifest.json: 8 new module entries

(SE-014..026), 16 total, all disabled by default.

Fixed

scripts/pr-plan-gates.sh: G4 CHANGELOG reconstruction now strips

credentials from git remote get-url origin output. Prevents PAT leakage into compare links when remote URL has embedded token. Adds s|https://[^@]*@|https://| to the sed chain.

4.64.0 — 2026-04-12

SE-002 Multi-Tenant & RBAC completion. Era 226. Savia Shield now treats tenant paths as N4-private. Enterprise rule documented. All 7 acceptance criteria satisfied. 21 BATS tests pass. Layer contract clean.

Added

.claude/enterprise/rules/multi-tenant.md: enterprise rule

documenting tenant model, resolution order, isolation rules, RBAC declarative format, Savia Shield integration, and feature flag.

Changed

.claude/hooks/data-sovereignty-gate.sh: added tenants/* to

the N4-private path skip list (AC4 — Savia Shield respects tenant boundaries). Tenant data is never scanned for public-repo leakage.

4.62.0 — 2026-04-12

Savia Enterprise resource, knowledge and compliance batch: 5 specs (SE-022..026). Era 224. Second wave of enterprise capabilities covering the operational backbone of a large consultancy.

Added

SE-022 Resource & Bench Management: utilization tracking (78-85%

target), skills-based matching, bench optimization, EU AI Act compliant allocation decisions with equality shield counterfactual test.
SE-023 Knowledge Federation: cross-project pattern mining with N4

anonymization, expertise directory ("know-who"), knowledge feed at bid time closing the SE-019 to SE-015 loop.
SE-024 Client Health Intelligence: 6-dimension health scoring

(delivery, commercial, relationship, satisfaction, growth, payment), stakeholder change tracking, churn prediction.
SE-025 Agentic Workforce Analytics: human vs agent throughput/cost/

quality measurement, ROI calculation, EU AI Act transparency disclosure at 3 levels (none/aggregate/full), SE-013 calibration advisor.
SE-026 Compliance Evidence Automation: automated audit trail for

ISO 9001, DORA, NIS2, AI Act. Evidence harvested from git history, Court verdicts, allocation decisions. Auditor-ready packages.

4.61.0 — 2026-04-12

4 tactical patterns adopted from multica-ai research. Era 223. Path redaction, agent result schema, concurrent executor, skills lock.

Added

scripts/path-redact.sh: redacts $HOME/username from text before

persisting. Stdin, file, and check modes. Prevents PII leakage via filesystem paths in agent output and logs.
.claude/schemas/agent-result.schema.json: structured result from

any agent execution — duration, token counts (input/output/cache), model, tool calls, verdict, session ID. Enables accurate /agent-cost.
scripts/lib/concurrent-executor.sh: semaphore-bounded parallel

task execution with graceful 30s drain on shutdown. Sourceable library for overnight-sprint and dag-execute.
scripts/skills-lock.sh: SHA-256 integrity verification for all

skills and agents. Generate, verify, diff modes. Critical for enterprise distribution (SE-008).
tests/test-multica-patterns.bats: BATS test suite, SPEC-055 certified.

4.60.0 — 2026-04-12

Code Review Court implementation (SE-021). Era 221. A panel of 5 specialized agent-judges that review AI-generated code from distinct angles, enforcing a 400 LOC batch-size gate (Nyquist bound). Human E1 reviews findings, not raw diffs. Score formula: 100-(C×25+H×10+M×3+L×1).

Added

docs/rules/domain/code-review-court.md (86 lines): rule

documenting the 5 judges, scoring model, flow, batch gate, fix cycle.
.claude/schemas/review-crc.schema.json: JSON Schema for the

.review.crc artifact (judges, findings, per-file SHA-256, rounds).
7 new agents: court-orchestrator (L4), correctness-judge (L1),

architecture-judge (L1), security-judge (L1), cognitive-judge (L1), spec-judge (L1), fix-assigner (L2). All with token_budget.
scripts/court-review.sh: orchestration helper — check (batch

gate), skeleton (.review.crc template), score (formula), hash (per-file SHA-256).
/court-review command: convenes the Court on current branch diff.
tests/test-code-review-court.bats: 37 tests — structural, agents,

rule, scoring, hash, skeleton, integration invariants.

4.59.0 — 2026-04-12

Code Review Court critical findings fix. Era 222. First dogfood of the Court (4 judges ran on the 4 pipeline scripts) identified 4 critical bugs. All 4 fixed in this PR, demonstrating the Court's value on its first run.

Fixed

COR-001 negative version: awk '$2-1' produced X.-1.0 on major

versions. Fix: git tag fallback + clamp minor ≥ 0.
COR-002 HMAC bypass: do_verify silently passed without HMAC when

secret file missing. Fix: explicit WARNING + "no cryptographic proof" message on stderr. Verification still passes (diff-hash match) but the audit trail now shows HMAC was skipped.
COG-001/ARCH-003 g4 complexity: 70-line monolith split into

_resolve_changelog_conflict() and _resolve_signature_conflict(). g4 is now 20 lines. Header split uses dynamic grep -n '^## \[' instead of hardcoded head -7 (fixes COR-003 magic number).
F-001/COG-004 PII: hardcoded GitHub handle replaced with

git remote get-url origin derivation. No personal data in source.

Changed

scripts/confidentiality-sign.sh: ::error:: GitHub Actions syntax

replaced with echo "ERROR:" for local execution compatibility (F-007).

4.57.0 — 2026-04-12

Savia Enterprise Project Lifecycle batch: 5 specs (SE-016..021). Era 220. Completes the full consultancy lifecycle suite started in Eras 208-213. Single batch PR to avoid the cascading CHANGELOG conflict problem.

Added

SE-016 Project Valuation — Business-Case-as-Code: living NPV/IRR,

benefit realization at 90/180/365d, portfolio dashboard, kill sentinel.
SE-018 Project Billing — Revenue-as-Code: IFRS 15 POC, WIP auto-compute,

invoice drafting with human gate, SOX audit trail, chargeback.
SE-019 Project Evaluation — Lessons-as-Code: quality metrics (CMMI),

NPS/CSAT, knowledge loop feeding lessons back to SE-015 library.
SE-020 Cross-Project Dependencies — Portfolio-as-Graph: deps.yaml

declarations, critical path, resource contention, rebalancing proposals.
SE-021 Code Review Court — 5 specialized agent-judges (correctness,

architecture, security, cognitive, spec-alignment), .review.crc verdicts with SHA-256 per file, 400 LOC batch gate, fix cycle orchestration, human E1 reviews findings not diffs. From Bryan Finster research.

4.56.0 — 2026-04-12

pr-plan G4 CHANGELOG reconstruction (replaces marker-stripping). Era 219. When a branch is behind origin/main and CHANGELOG.md conflicts, G4 now reconstructs the file from scratch: takes origin/main:CHANGELOG.md as base, extracts this branch's new entry, and inserts it at the correct position — above main's top version. This is deterministic and correct regardless of how many PRs have merged since the branch was created. The previous approach (sed-stripping conflict markers) could produce malformed output when conflict structure was complex. The reconstruction approach treats main's CHANGELOG as the authoritative base and simply prepends the branch's contribution.

Changed

scripts/pr-plan-gates.sh (g4): full rewrite. On CHANGELOG

conflict, extracts branch's new entry between header and main's top version, takes main's CHANGELOG as base, inserts entry at line 8, adds compare link at the correct position. Fallback to marker-strip if extraction fails. .confidentiality-signature auto-removed. Non-CHANGELOG conflicts still fail for human resolution.
scripts/changelog-assemble.sh (new): fragment-based assembly

script for future use. Reads CHANGELOG.d/*.md fragments and concatenates into CHANGELOG.md. Currently optional; will become the primary path when fragment workflow is adopted.
CHANGELOG.d/.gitkeep (new): directory for future per-version

fragments.

4.50.0 — 2026-04-12

Savia Enterprise Project Prospect — Pipeline-as-Code (SE-015). Era 213. A sovereign, agent-queryable opportunity pipeline where pursuits live as .md files with BANT/MEDDIC qualification scoring, bid/no-bid decision audit trails, proposal knowledge reuse from a local library, and a canonical sales→delivery handoff package. Blocked by SE-001, SE-002. Blocks SE-016 (valuation), SE-017 (SOW init on win), SE-018 (billing), SE-019 (evaluation), SE-020 (resource demand forecast from pipeline).

Added

SPEC-SE-015 Project Prospect (docs/propuestas/savia-enterprise/SPEC-SE-015-project-prospect.md):

pursuit.md frontmatter schema, BANT+MEDDIC qualification.yaml, bid/no-bid decision record, proposal library (capabilities, case-studies, team-bios, templates), handoff.md for sales→delivery bridge, 4 agents (prospect-qualifier L1, proposal-drafter L2, handoff-generator L2, win-loss-analyst L1), 7 commands /pursuit-* + /pipeline-view, 5 lifecycle events (qualified, bid_decided, won, lost, handoff_completed), air-gap capable with Ollama. Spec proposal only.

4.49.0 — 2026-04-12

Savia Enterprise Project Definition — SOW-as-Code (SE-017). Era 212. A machine-readable, agent-queryable, contract-grade Statement of Work stored as .md inside the tenant's pm-workspace, with testable acceptance criteria, YAML RACI matrix, deliverables ledger linked to the backlog, and a structured change-request mechanism that produces auditable amendments. Delivery work not traceable to a SOW deliverable is refused by the workspace. Blocked by SE-001, SE-002, SE-010. Blocks SE-018 (billing amounts derive from SOW contract value) and SE-019 (evaluation baseline is SOW deliverables vs actually-shipped).

Added

SPEC-SE-017 Project Definition (docs/propuestas/savia-enterprise/SPEC-SE-017-project-definition.md):

canonical SOW.md frontmatter schema, 4 new agents (sow-writer L2, cr-drafter L2, raci-validator L1, acceptance-linker L1), commands /sow-init /sow-validate /sow-cr-draft /sow-amend /sow-query, sow.amended event emitted on amendment signing for SE-018/SE-019 consumers, sow-trace-validate.sh hook blocks orphan PBIs, JSON Schema + 20+ BATS tests target, air-gap capable. Spec proposal only.

4.48.0 — 2026-04-12

Engineering principles distilled from the Linux kernel. Era 211. Savia mapped the Linux kernel tree (v7.0-rc7 "Baby Opossum Posse") into five .acm files (INDEX + mm + kernel + net + fs) plus a patterns digest and extracted the five most transferable engineering principles into a new rule that applies them to pm-workspace.

Added

docs/rules/domain/engineering-principles-from-kernel.md (132

lines): five principles distilled from 37M lines of kernel source — (1) pay for what you use (compile-out in prod), (2) make mechanisms visible (no hidden control flow), (3) verify at runtime in debug, not compile-time, (4) interface tables over class hierarchies, (5) safe extension via verified bytecode (the eBPF model). Each principle includes "how to apply" bullets that map to concrete pm-workspace mechanisms (hook profiles, pr-plan gates, BATS, language packs, agent ops tables). Also a "what NOT to copy from the kernel" section so Savia doesn't over-engineer in directions that don't fit her I/O-bound architecture. References the research tree under /home/monica/research/linux-kernel/ (outside the repo).

4.47.0 — 2026-04-12

pr-plan G5 queue check hardened against recurrent CHANGELOG merge conflicts. Era 210. The old check only flagged exact version-number equality between my branch and any open PR. That prevented number collisions but not the real pain: two branches both insert at line 8 of CHANGELOG.md with different anchors, producing a merge conflict whenever one of them lands first. The fix enforces that my local version is strictly greater than every version claimed by main OR any open PR. This guarantees a rebase after a peer lands is always a clean "my entry on top of theirs" insertion.

Changed

scripts/pr-plan-gates.sh (g5): queue check now tracks max_claimed

across origin/main top + every open PR's CHANGELOG top, and fails with FAIL: version <lv> <= max in queue <max_claimed> — bump to <next> when the local version does not strictly exceed the queue. Uses sort -V for semver-correct comparison (avoids the 4.10.0 < 4.2.0 lexicographic trap). Degrades to silent skip when gh is unavailable or PR_PLAN_SKIP_QUEUE_CHECK=1.
tests/test-pr-plan-queue-check.bats: updated failure-message strings

(max in queue, bump to) and added three regression tests (tracks max_claimed across main and open PRs, uses sort -V, Era 210 hardening guard). 17/17 pass locally.

4.46.0 — 2026-04-12

Hook noise fix. Era 209. 15 shell hooks were committed to git with mode 100644 (non-executable), causing PreToolUse:Bash to fail every turn with /bin/sh: 1: <hook>: Permission denied (exit 126). Every clone of the repo reproduced the noise. This change flips them to 100755 via git update-index --chmod=+x without touching content.

Fixed

.claude/hooks/agent-hook-premerge.sh, bash-output-compress.sh,

config-reload.sh, emotional-regulation-monitor.sh, file-changed-staleness.sh, instructions-tracker.sh, post-tool-failure-log.sh, pre-compact-backup.sh, stop-memory-extract.sh, stress-awareness-nudge.sh, subagent-lifecycle.sh, task-lifecycle.sh, tool-call-healing.sh, lib/memory-extract-lib.sh, lib/profile-gate.sh — all now executable. No content change.

4.45.0 — 2026-04-12

Savia Enterprise Release Orchestration (SE-014). Era 208. Release-as-Code for multi-tenant consultancy delivery: a .md-first contract per release with compliance profiles (standard, dora-banking, hipaa-health, gdpr-eu, nis2-critical, airgap), human-gated deploy, and audit-safe rollback. Blocked-by SE-001, SE-002, SE-003; blocks SE-018 (billing consumes release.completed) and SE-019 (evaluation baseline). Spec proposal only — implementation in a later sprint.

Added

SPEC-SE-014 Release Orchestration (docs/propuestas/savia-enterprise/SPEC-SE-014-release-orchestration.md):

release-plan.md canonical schema, 6 compliance profiles, 3 new agents (release-orchestrator L2, release-validator L1, rollback-executor L1), /release-plan + /release-validate commands, release.completed event consumed by SE-018/SE-019, air-gap first-class, event-driven integration with SE-003 MCP adapters (Azure DevOps Release + GitHub Deployments as v1 targets).

4.44.0 — 2026-04-11

Dual estimation rule (SE-013) with two-ratio system. Formalizes the ~10x end-to-end pipeline speedup claim and — critically — keeps TWO live ratios simultaneously: a fixed conservative 10x (default for planning, always safe) and an updating empirical ratio computed from data/agent-actuals.jsonl. PM decides when to opt-in to empirical mode; conservative stays as default until the team has enough data to trust its own numbers. Era 207.

Added

docs/rules/domain/dual-estimation.md: the rule. Phase breakdown, dual-ratio model (conservative 10x + empirical on-demand), canonical formula, adjustment table (trivial 15x → legacy 2x), 4 conditions for the 10x claim, sources (METR papers + n=2 HUDI + SE-002 real data).
docs/propuestas/TEMPLATE.md: spec header updated with Estimate (human): Nd and Estimate (agent): Nh dual fields + Category classifier.
scripts/estimate-calibrate.sh: reads data/agent-actuals.jsonl, groups by category, computes empirical speedup, suggests adjustments when samples ≥ DUAL_ESTIMATION_MIN_SAMPLES. Supports --format json and --log.
scripts/estimate-convert.sh: PM-facing helper. Converts human-days to agent-hours using either --mode conservative (default 10x) or --mode empirical (opt-in, reads live ratio from actuals log). Falls back to conservative when empirical lacks samples. Supports --category, --format json, --min-samples.
data/agent-actuals.example.jsonl: seed with SE-001/002/008/012 + HUDI-8865/8551 real data. The live file data/agent-actuals.jsonl is gitignored (PM-local tracking).
tests/test-dual-estimation.bats: 23 BATS tests (17 original + 6 for the two-ratio helper), SPEC-055 certified.
docs/propuestas/savia-enterprise/DEVELOPMENT-PLAN.md: throughput claim block at top, linking to the rule.
docs/rules/domain/pm-config.md: three new config keys (DUAL_ESTIMATION_ENABLED, DUAL_ESTIMATION_MIN_SAMPLES, AGENT_ACTUALS_LOG).

Changed

Spec template and development plan now explicitly distinguish human-days from agent wall-clock hours. The conservative 10x is the planning default; empirical is opt-in when the PM trusts team data.

4.43.0 — 2026-04-11

Deep research over 17 external repos (legalize-es, mempalace, llmfit, rowboat, claude-usage, feynman, Ix, qwen-code, repowise, caveman, deepteam, claude-memory-kit, METATRON, advisor-strategy, claude-code-from-source, claudecowork, prompt-caching-2026) produces 10 executable SDD specs targeting sprint 2026-08. Era 206.

Added

SPEC-TOOLS-BY-TASK: N-target batch tools, −49% tool calls (repowise pattern)
SPEC-HOOKS-OVER-PROMPTS-AUDIT: delete rules Claude already follows (Era 165 follow-up)
SPEC-PROMPT-CACHING-2026: static/dynamic split, 1h TTL, workspace isolation
SPEC-HOOK-CONFIG-SNAPSHOT: freeze settings.json at session start (sticky latches)
SPEC-CACHE-HIT-TRACKING: SQLite scanner for real cache hit rate measurement
SPEC-SAVIA-DUAL-NUPSTREAM: Anthropic → Qwen OAuth → Ollama cascade
SPEC-AGENTIC-REDTEAM: Goal Theft / Recursive Hijacking / Identity Corrosion (sprint 2026-09)
SPEC-LEGAL-DRIFT: git log over legalize-es to detect legislative reforms
SPEC-RECURSION-GUARD: CLAUDE_INVOKED_BY guard against sub-Claude loops

Changed

SPEC-ADVISOR-STRATEGY: extended §8 Context Safety with payload gates, timeouts,

cascade prevention, and fallback hierarchy to prevent executor blocking on inherited contexts. 6 new business rules (CTX-01 to CTX-06) with per-agent budgets.

4.42.0 — 2026-04-11

Savia Dual installer is now fully automatic. A single command provisions the proxy service, shell integration, and launches Claude Code with the routing environment already applied. The user runs one script and lands inside Claude Code with the proxy active — no manual sourcing, no new terminal, no thinking about environment propagation. Era 205.

Added

systemd unit (Linux) savia-dual-proxy.service installed at

/etc/systemd/system/, enabled at boot, runs as the invoking user with logs at /var/log/savia-dual-proxy.log.
launchd agent (macOS) com.savia.dual.proxy.plist installed at

~/Library/LaunchAgents/ with KeepAlive.
Shell integration — idempotent block appended to ~/.bashrc and

~/.zshrc that sources ~/.savia/dual/env on every new shell, so ANTHROPIC_BASE_URL is set automatically without manual intervention.
Health check — GET http://127.0.0.1:8787/health verified right

after the service starts.
Auto-launch — the installer sources the env file in its own

process and execs claude, so Claude Code inherits ANTHROPIC_BASE_URL without requiring the parent shell to reload. A child process cannot modify its parent's environment, so bash setup.sh && claude is impossible on Unix. Exec from the same process sidesteps the limitation cleanly.
--no-launch flag — install without exec'ing claude, for CI and

headless server provisioning.
-- passthrough — arguments after -- are forwarded to claude on

auto-launch (e.g. ./setup-savia-dual.sh -- --resume).

Changed

Installer UX — the final summary no longer asks the user to open

a new terminal; the installer itself ends inside Claude Code.

4.41.0 — 2026-04-11

Savia Enterprise multi-tenant isolation & RBAC (SE-002). Era 206. Third P0 of the Savia → Savia Enterprise migration plan. Depends on SE-001. Core stays untouched: all new behaviour is gated by manifest.json → multi-tenant.enabled.

Added

.claude/enterprise/hooks/tenant-resolver.sh: resolves active tenant slug from $SAVIA_TENANT → cwd under tenants/<slug>/ → active user profile tenant: key → empty (single-tenant fallback). Exposes tenant_resolve() for sourcing and runs standalone. Implements extension point EP-5 from SE-001.
.claude/enterprise/hooks/tenant-isolation-gate.sh: PreToolUse hook (Edit|Write|Read) that blocks any cross-tenant file access with exit 2. Allowlists .claude/, scripts/, docs/, tests/, output/. No-op when the multi-tenant module is disabled or no tenant is active. Audit log at output/tenant-audit.jsonl. Implements extension point EP-3.
.claude/enterprise/commands/rbac-manager.md: /rbac-manager slash command with grant, revoke, list, check subcommands. Documents rbac.yaml schema with role inheritance (reader → developer → admin) and glob command patterns.
scripts/rbac-manager.sh: backend implementation of /rbac-manager. Pure-bash YAML parser, atomic writes, recursive inheritance resolution, idempotent grant, no-op revoke for absent users.
tests/test-tenant-isolation.bats: 21 BATS tests, SPEC-055 certified.
tests/test-rbac-manager.bats: 16 BATS tests, SPEC-055 certified.
docs/propuestas/savia-enterprise/SE-002-extension-points.md: implementation notes mapping the delivered files to EP-3 and EP-5 declared in SE-001.

Changed

docs/propuestas/savia-enterprise/ documentation now has a concrete implementation reference for EP-3 and EP-5 beyond the SE-001 contracts.

4.40.1 — 2026-04-11

Savia Dual installer scripts are now fully idempotent. Re-running them no longer re-downloads Ollama or models already present, and no longer overwrites an existing config. Era 205.

Fixed

Installer reuse logic scripts/setup-savia-dual.sh and

scripts/setup-savia-dual.ps1 — if any gemma4 variant is already installed, reuse it instead of pulling the hardware-ideal pick. The user's deliberate choice is honored (e.g. gemma4:26b kept even on machines where the ideal pick would be smaller).
Config preservation — existing ~/.savia/dual/config.json and

env files are no longer overwritten on re-run. New --force / -Force flag to rewrite them explicitly; --reconfigure now implies --force.
Model detection — stronger parsing of ollama list output

(skip header, filter by gemma4: prefix).

4.40.0 — 2026-04-11

Savia Dual: inference sovereignty layer with transparent failover between Anthropic API and local Ollama gemma4. Cloud when it works, local when it does not. The user never gets stuck because of network, rate limits, or provider outages. Era 205.

Added

Rule docs/rules/domain/savia-dual.md — architecture, failover

triggers, hardware-based model selection, audit log format, hard limits.
Skill .claude/skills/savia-dual/ — SKILL.md + DOMAIN.md with

the Clara dual-doc pattern.
Command /savia-dual {install|start|stop|status|test|logs} — full

lifecycle management.
Proxy scripts/savia-dual-proxy.py — Python stdlib, no external

dependencies, transparent Anthropic → Ollama fallback on network error, HTTP 5xx, HTTP 429 (quota), or timeout. Circuit breaker included.
Installer Linux/macOS scripts/setup-savia-dual.sh — installs

Ollama, detects RAM/VRAM, picks gemma4 variant (e2b / e4b / 26b), downloads it, writes config.
Installer Windows scripts/setup-savia-dual.ps1 — winget-based

equivalent with WMI hardware detection.
Docs docs/savia-dual.md — user guide covering install, daily use,

log format, and honest limitations compared to cloud.
Docs README.md (9 languages) — new Inference Sovereignty section

comparing Emergency Mode (manual) and Savia Dual (automatic).

Changed

Skill catalog auto-includes savia-dual under category governance.
Command catalog auto-includes /savia-dual in capability group.

4.39.0 — 2026-04-11

Savia Enterprise licensing & distribution strategy (SE-008). Era 205. Second P0 of the Savia → Savia Enterprise migration plan. Depends on SE-001.

Added

LICENSE-ENTERPRISE.md at repo root: explicit MIT statement for the .claude/enterprise/ layer. Identical terms to Savia Core. Formally rejects Open Core, BSL, AGPL, SaaS hosted, and pay-per-agent licensing models with the specific foundational principle each violates.
TRADEMARK.md at repo root: policy on the "Savia" and "Savia Enterprise" names. Forks permitted with name change. Repackaging the code as a closed product under another name requires permission.
docs/support-offering.md: six monetizable services (professional support with SLA, implantation, certified training, custom spec development, sovereignty audits, hardware integration). Services are labor, never license fees. Code stays MIT.
docs/savia-enterprise-mit-forever.md: public statement explaining the MIT-forever commitment. Cites the 7 foundational principles and the clone-your-own-instance test.
docs/propuestas/TEMPLATE.md: RFC template for future SPEC-XXX proposals.
tests/test-licensing-files.bats: 27 tests covering file existence, required content, and the 5 rejected licensing models.

Principles preserved

All 7 foundational principles unchanged. This spec is the legal guardian of the principles.

4.37.0 — 2026-04-11

Signal/noise reduction (SE-012). Era 203. Unblocks efficient work on Savia Enterprise migration by eliminating two chronic friction sources: false-positive Bash hooks and invisible CI failure rates.

Removed

LLM-based commit validation hook (.claude/settings.json PreToolUse:Bash prompt-type hook, Haiku). Root cause of recurring PreToolUse:Bash hook error banners: the if: "Bash(git commit*)" matcher fired on unrelated git subcommands (git merge-tree, git commit --no-edit for merges), and returning {ok: false} blocked legitimate commands. The existing deterministic prompt-hook-commit.sh (command-type, warning mode) already covers every real case with zero false positives and zero token cost.

Added

scripts/ci-failure-tracker.sh: CLI to record CI state per PR (record), compute failure rate aggregates (health), and list top recurring failures (top). Log is append-only JSONL at output/ci-runs.jsonl (N3 local, gitignored).
/ci-health slash command: surfaces per-check failure rate and top-5 recurring causes in the last N days.
tests/test-ci-failure-tracker.bats: 16 tests covering structure, empty/missing/malformed log handling, rate computations (0%, 100%, mixed), per-check grouping, top ordering, and append-only invariant.
docs/propuestas/savia-enterprise/SPEC-SE-012-signal-noise-reduction.md: spec with diagnosis, module breakdown, acceptance criteria.

Fixed

.gitkeep files in .claude/enterprise/{agents,commands,rules,skills}: restore empty subdirs that git was dropping. This is what caused test-validate-layer-contract.bats tests 11-14 and 19 to fail on CI while passing locally. Root cause of the BATS Hook Tests check failing on the initial push.
Hardcoded absolute paths in tests/test-validate-layer-contract.bats: replaced occurrences of a developer-local home path with $CLAUDE_PROJECT_DIR so the tests follow the same PROJECT_DIR resolution the hook uses at runtime. Without this, the hook's .claude/commands/* pattern never matched the CI checkout path, so negative tests silently allowed instead of blocking.

Added (Module 4 — PR Queue Check)

g5() in scripts/pr-plan-gates.sh: after verifying CHANGELOG against origin/main, queries all open PRs via gh api and compares each remote branch's top ## [X.Y.Z] entry against the local one. On collision, fails with an actionable suggestion ("rebase to X.(Y+1).0 (next free)"). Degrades gracefully when gh is missing or offline via PR_PLAN_SKIP_QUEUE_CHECK=1.
tests/test-pr-plan-queue-check.bats: 15 tests certified by the SPEC-055 quality gate. Covers structure, opt-out env var, collision detection pattern, graceful degradation, next-free-version math, and regression invariants on pre-existing main-comparison behavior.
Motivation: two real version collisions in rapid succession during an active migration. Each collision required: detect conflict → understand version chain → manual bump → rewrite CHANGELOG → fix compare links → re-merge → re-sign. The gate prevents this entire cycle.

Impact

Hook noise: expected reduction in false-positive Bash hook errors during git operations (merges, analysis commands, flag combinations). The remaining validation is deterministic and non-blocking.
CI visibility: first measurable signal on pipeline reliability via /ci-health.
Version collision prevention: g5() now catches the pattern that required manual intervention twice during this same workstream.

4.36.0 — 2026-04-11

Savia Enterprise foundations (SE-001). Era 202. First step of the Savia → Savia Enterprise migration plan (11 specs in docs/propuestas/savia-enterprise/).

Added

Savia Enterprise layer contract: new .claude/enterprise/ directory (opt-in, MIT, unidirectional). Contains agents/commands/skills/rules subdirs + manifest.json + JSON schema. All modules ship disabled by default.
scripts/validate-layer-contract.sh: full-repo scanner that enforces "Core never imports from .claude/enterprise/". Full scan: 1092 files, 0 violations.
.claude/hooks/validate-layer-contract.sh: PreToolUse hook (Edit|Write) that intercepts Core→Enterprise imports before they land on disk. Registered in .claude/settings.json.
docs/propuestas/savia-enterprise/: 11 executable specs (SE-001..SE-011) covering foundations, multi-tenant, MCP catalog, agent framework interop, sovereign deployment, governance pack, onboarding, licensing, observability, migration path, and docs restructuring.
docs/propuestas/savia-enterprise/DEVELOPMENT-PLAN.md: DAG with 3 parallel waves, resume protocol across sessions, escalation rules.
docs/propuestas/savia-enterprise/extension-points.md: 6 formal extension points (agent registry, hook registry, RBAC gate, audit sink, tenant resolver, compliance validator).
tests/test-validate-layer-contract.bats: 26 tests covering hook positive/negative cases, script full scan, manifest schema, opt-in defaults, invariant enforcement.

Principles preserved

All 7 foundational principles remain intact (data sovereignty, vendor independence, radical honesty, absolute privacy, human decides, equality shield, identity protection). Enterprise layer is MIT, agnostic, opt-in. No vendor lock-in introduced.

4.35.1 — 2026-04-11

Savia Claw rescue on host: implement the missing HTTPS bridge as a systemd service so SaviaClaw stops looping remote:unreachable SOS alerts on Nextcloud Talk. Add /memory-check health command covering all 10 memory layers, and generate the per-project Agent Code Map (ACM) for the Savia Claw subsystem. Era 204.

Added

Command /memory-check: 10-layer memory health check (auto-memory, JSONL store,

vector index, SQLite cache, knowledge graph, agent memory, personal vault, session-hot, instincts, memory stack). Exit 0 on OK/warnings, 1 on critical failures.
Script scripts/memory-check.sh — runs all checks, dashboard output
Script scripts/start-bridge.sh — wrapper invoked by remote_host.restart_bridge();

prefers systemctl restart savia-bridge (system unit), falls back to systemctl --user.
Script scripts/install-savia-bridge-system.sh — idempotent sudo installer that

promotes the user-level savia-bridge to a hardened system unit so the bridge auto-starts on host reboot (PrivateTmp, ProtectSystem=strict, ProtectHome=read-only, NoNewPrivileges, MemoryMax=512M, CPUQuota=50%).
ACM zeroclaw/.agent-maps/ (INDEX + host/daemons + host/survival + host/comms) —

per-project Agent Code Map for Savia Claw's 37 Python modules, complying with feedback_agent_maps_per_project.md (never at repo root).
Doc docs/savia-claw-bridge.md — architecture, lifecycle, failure modes, and

operational runbook for the Savia Bridge service.
BATS test-memory-check.bats — 7 tests
BATS test-savia-bridge-scripts.bats — 13 tests (lint-only; no sudo in CI)
BATS test-zeroclaw-agent-maps.bats — 10 tests (format + 150-line cap)

Changed

scripts/savia-bridge.service — corrected ExecStart path from /home/monica/savia/scripts/

to /home/monica/claude/scripts/. Added explicit User=monica, Group=monica, narrowed ReadWritePaths to /home/monica/.savia/bridge, added ReadOnlyPaths=/home/monica/claude.

Fixed

Root cause of the remote:unreachable SOS loop on Nextcloud Talk: Savia Claw's

remote_host.py expected an SSH-reachable bridge, but ~/.savia/remote-host-config was missing and no scripts/start-bridge.sh existed. Both are now provided, SSH loopback is wired to monica@localhost via a dedicated ed25519 key, and SaviaClaw correctly detects is_reachable=True and is_bridge_running=True.

4.35.0 — 2026-04-11

Lazy context architecture fix + GitHub release pipeline fix. Era 201.

Fixed

Critical architecture bug: subagents launched via Task tool were crashing with autocompact thrashing ("context refilled to the limit within 3 turns of the previous compact, 3 times in a row"). Root cause: CLAUDE.md had 15 @import directives that resolved recursively to 24 files totalling ~29,177 tokens. CLAUDE.md is a per-turn dynamic suffix (NOT cached system prompt), so those tokens were paid on EVERY turn, leaving Sonnet subagents with insufficient headroom after just a few tool calls.

Changed

CLAUDE.md: refactored from eager @import model to lazy reference model. Only 3 @imports remain (savia.md, radical-honesty.md, autonomous-safety.md — the absolute foundational minimum). Everything else is documented as a "Lazy Reference" table with explicit paths + when to read. Agents now Read these files on demand instead of auto-loading them on every turn.
Per-turn token cost: 29,177 tokens → 4,904 tokens (83% reduction, 24,273 tokens reclaimed per turn).
File count: 24 files auto-resolved → 4 files auto-resolved.

Validation

Test 1: Sonnet subagent reads CLAUDE.md and answers 4 questions → completed in 1 turn, 149K tokens total (previously thrashed before completing).
Test 2: Sonnet subagent implements a small bash script + BATS tests → completed without thrashing (blocked only by data-sovereignty hook on /tmp writes, which is correct behavior, not thrashing).

Impact

Subagents can now perform meaningful work before hitting compact thresholds.
fork-agents.sh (SPEC-FORK-AGENT-PREFIX) is now actually usable at scale — agents won't thrash on the inherited prefix.
The remaining 336K tokens of rules/profiles content in .claude/rules/ are still available via Read — nothing was deleted, only the eager-load behavior was changed.

Fixed (Release pipeline)

Critical CI bug: GitHub releases stopped being created around v2.0.0 (March 2026) despite git tags being created correctly by auto-tag.yml. Root cause: GitHub Actions has a documented safeguard where workflows triggered by GITHUB_TOKEN DO NOT trigger other workflows (prevents infinite loops). Since auto-tag.yml pushed tags using GITHUB_TOKEN, the tag push did NOT trigger release.yml, resulting in most git tags with no corresponding GitHub release.
Fix: merged release creation into auto-tag.yml as a single workflow. It now detects the version, creates the tag, AND creates the GitHub release in one run. No cross-workflow dependency needed.
release.yml is kept as a fallback for manual tag pushes by humans (explicitly skips github-actions[bot] actor since auto-tag.yml already handles that).
Added workflow_dispatch trigger to auto-tag.yml with force_version input for manual re-runs if a release is missed.
Pinned softprops/action-gh-release to a specific SHA per security best practice.

Added (Release pipeline)

Script scripts/release-backfill.sh (~190 lines): creates missing GitHub releases from existing git tags. Supports --dry-run, --limit N, --from VERSION, --to VERSION, --force (overwrite). Extracts changelog per-version via awk, creates releases idempotently (skips if already exists).
BATS tests/test-release-backfill.bats: 39 tests covering script integrity, CLI flag parsing, auto-tag.yml structure, release.yml fallback behavior, CLAUDE.md lazy context validation, and edge cases. Certified by auditor.

4.34.0 — 2026-04-10

Savia Monitor Linux build support — deb, rpm, appimage targets. Era 200.

Added

Script projects/savia-monitor/scripts/build-linux.sh (~170 lines): automated Linux build with environment checks, prerequisite detection (Debian/Ubuntu and Fedora/RHEL), selective target builds (deb/rpm/appimage only), dev mode, --check flag for environment verification only
BATS tests/test-savia-monitor-linux.bats: 38 tests covering script integrity, tauri.conf.json Linux targets, README alignment ES/EN, Rust source cross-platform compatibility, and build script edge cases

Rationale

CI workflow intentionally NOT added: Tauri Linux builds exceed 15 minutes per run, blocking PR iteration. Linux artifacts are built on-demand locally via build-linux.sh or on release tags, not on every push.

Changed

tauri.conf.json: added explicit bundle targets list (deb, rpm, appimage, msi, nsis, dmg), Linux-specific section with deb/rpm dependencies (libwebkit2gtk-4.1-0, libgtk-3-0, webkit2gtk4.1, gtk3), category Utility, short/long descriptions
README.md + README.en.md: added Linux system prerequisites section (apt and dnf commands), Linux Build section with build-linux.sh usage examples
projects/savia-monitor/CLAUDE.md: documented Linux build commands and targets supported matrix (Windows/macOS/Linux)

Notes

Rust source code was already cross-platform compatible: sessions.rs uses /proc/{pid} for Linux PID detection, config.rs falls back from HOME to USERPROFILE, git.rs guards Windows-specific flags with #[cfg(target_os = "windows")]
Build artifacts generated in $CARGO_TARGET_DIR/release/bundle/ (default ~/.savia/cargo-target/savia-monitor/ per workspace convention)

4.33.0 — 2026-04-10

Five SPECs + implementation from deep analysis of claude-code-from-source repo (reverse-engineered Claude Code internals). Era 199.

Added

SPEC SPEC-FORK-AGENT-PREFIX (CRITICA): byte-identical prompt prefix for batch agents, exploiting 90% prompt cache discount
SPEC SPEC-AUTOCOMPACT-CALIBRATION (ALTA): recalibrate autocompact threshold from 65% to 75% to match native 20-25% buffer
SPEC SPEC-FORK-VS-SUBAGENT-GUIDE (ALTA): decision tree and comparison table for fork vs subagent patterns
SPEC SPEC-HOOK-EVENT-GAP-AUDIT (MEDIA): audit 11 uncovered hook events out of 28 total
SPEC SPEC-TERMINAL-STATE-HANDOFF (MEDIA): termination_reason enum in handoff templates with retry policy
Script scripts/fork-agents.sh: batch agents with cacheable prefix, sha256 verification, parallel execution
Script scripts/hook-event-gap-audit.sh: generates hook-event-gap-audit.md report
Script scripts/validate-handoff.sh: validate termination_reason enum in handoffs
Script scripts/context-calibration-measure.sh: measure context usage patterns for calibration
Rule fork-agent-protocol.md: strict protocol for byte-identical prompt construction
BATS test-fork-agents.bats: 24 tests
BATS test-context-calibration.bats: 24 tests
BATS test-fork-vs-subagent-docs.bats: 37 tests
BATS test-hook-event-gap-audit.bats: 23 tests
BATS test-handoff-termination.bats: 34 tests

Changed

settings.json: CLAUDE_AUTOCOMPACT_PCT_OVERRIDE 65 -> 75
context-health.md: Gradual zone 50-70%, Alerta zone 70-85%
dev-session-protocol.md: +Fork vs Subagent decision section
handoff-templates.md: +termination_reason enum + fork comparison table
verification-before-done.md: +Retry Policy by Termination Reason table

4.32.0 — 2026-04-10

Four scripts + four test suites for dev-session pipeline + Advisor Strategy. Research: Ix, Feynman, Anthropic blog. Era 198.

Added

Script impact-analysis.sh: grep-based dependency graph + risk scoring for 6 languages, depth 1-3, SHA256 caching (SPEC-IMPACT-ANALYSIS)
Script semantic-map.sh: compressed semantic maps — public interfaces, deps, patterns, extension points. 6 languages + fallback (SPEC-SEMANTIC-CONTEXT-MAPS)
Script verification-middleware.sh: parallel 3-check orchestrator — traceability, tests, consistency. Security veto, retry context (SPEC-VERIFICATION-MIDDLEWARE)
Script advisor-config.sh: Anthropic Advisor Strategy config — Sonnet executor + Opus advisor pairing, agent frontmatter lookup (SPEC-ADVISOR-STRATEGY)
BATS test-impact-analysis.bats: 34 tests (SPEC-IMPACT-ANALYSIS)
BATS test-semantic-map.bats: 49 tests (SPEC-SEMANTIC-CONTEXT-MAPS)
BATS test-verification-middleware.bats: 42 tests (SPEC-VERIFICATION-MIDDLEWARE)
BATS test-advisor-config.bats: 32 tests (SPEC-ADVISOR-STRATEGY)
SPECs IMPACT-ANALYSIS, SEMANTIC-CONTEXT-MAPS, VERIFICATION-MIDDLEWARE, ADVISOR-STRATEGY

4.31.0 — 2026-04-08

Four patterns from Anvil research (ppazosp/anvil) — heat parallelism, competitive architects, knowledge chains, compiled agent index. Era 197.

Added

Script heat-scheduler.sh: lightweight heat-based parallelism — phases = sequence, heats = parallel within phase, file conflict detection (SPEC-094)
Script competitive-design.sh: parallel design generation with 3 philosophies (minimal, clean, pragmatic) + 4-criteria evaluation (SPEC-095)
Script slice-context-chain.sh: knowledge chain between dev-session slices — injects completion summaries as context for next slice (SPEC-096)
Script compile-agent-index.sh: compiled AGENTS-INDEX.md from 49 agent definitions with hash-based freshness check (SPEC-097)
BATS test-heat-scheduler.bats: 17 tests (SPEC-094)
BATS test-competitive-design.bats: 16 tests (SPEC-095)
BATS test-slice-context-chain.bats: 18 tests (SPEC-096)
BATS test-compile-agent-index.bats: 20 tests (SPEC-097)
SPECs 094, 095, 096, 097: proposed from Anvil research

4.30.0 — 2026-04-08

Scoring improvements from llmfit research — optimal bands + variable weights + hardware-aware Ollama. Era 196.

Added

Script ollama-hardware-check.sh: GPU/VRAM detection, model size calculation, quantization recommendation, tok/s estimation (SPEC-093)
BATS test-ollama-hardware-check.bats: 21 tests (SPEC-093)

Changed

Rule scoring-curves.md: context usage now uses optimal band (40-65% = peak) instead of linear from 0% (SPEC-091)
Rule consensus-protocol.md: 4 weight profiles by task type — default, security, business, architecture — with keyword auto-detection (SPEC-092)

4.29.0 — 2026-04-08

Readiness check updated + memory-cache-rebuild migrated to Python sqlite3. Era 195b.

Changed

readiness-check.sh: expanded from 7 to 9 sections — added SQLite Cache Systems (4d) and Savia Shield (4e)
memory-cache-rebuild.sh: migrated from sqlite3 CLI to Python sqlite3 module (portability, no apt install needed)

4.28.0 — 2026-04-08

Memory architecture: SQLite cache + L0-L3 stack + temporal knowledge graph. Era 195.

Added

Script memory-cache-rebuild.sh: rebuilds SQLite cache from .md memory files (SPEC-089)
Script memory-stack-load.sh: token-budgeted progressive loading L0-L3 (SPEC-089)
Script knowledge-graph.sh: entity-relation graph with build/query/impact/status (SPEC-090)
BATS test-memory-stack.bats: 26 tests (SPEC-089)
BATS test-knowledge-graph.bats: 21 tests (SPEC-090)

Design

SQLite at ~/.savia/ as local cache per machine (regenerable, gitignored)
.md files remain source of truth (Principle #1: .md sovereignty)
Graceful degradation: all scripts work without SQLite (fall back to grep/read)

4.27.0 — 2026-04-08

Confidentiality hardening + context management patterns from Claudepedia analysis. Era 194.

Added

Hook block-gitignored-references.sh: blocks writing gitignored paths, audit scores, vulnerability counts, and internal metrics to public (N1) files. 8 detection patterns, security tier
BATS test-block-gitignored-references.bats: 14 tests — 8 blocking + 6 allowing patterns
BATS test-tdd-gate.bats: 20 tests — TDD enforcement (SPEC-081)
BATS test-plan-gate.bats: 14 tests — spec requirement warnings (SPEC-081)
BATS test-compliance-gate.bats: 13 tests — compliance runner (SPEC-081)
BATS test-block-project-whitelist.bats: 13 tests — gitignore privacy protection (SPEC-081)
BATS test-block-infra-destructive.bats: 20 tests — IaC safety (SPEC-081)
Skill savia-school/SKILL.md: orphan fix, 12 school commands documented (SPEC-082)
Session journal: crash recovery mechanism for session continuity
SPECs 081, 082, 085: proposed from pre-audit findings

Fixed

CHANGELOG.md: removed all leaked internal metrics (quality scores, audit results, debt-scores, vulnerability counts, output paths with dates, .human-maps details per project)
WORKSPACE.ctx: removed private project names from N1 tracked file
Hook block-infra-destructive.sh: false positive fix — pro in approve no longer matches environment regex
BATS test-sovereignty-benchmark.bats: added timeouts to prevent hang on slow Ollama
BATS test-spellcheck-docs.bats: added timeout for no-args full-repo scan

Changed

settings.json: registered block-gitignored-references.sh as PreToolUse hook (Edit|Write)
Script context-budget-check.sh: proactive dual-threshold budget tracker (80%/95%) with circuit breaker (SPEC-086)
Script tool-result-trim.sh: deterministic 5K char hard cap for tool results (SPEC-087)
BATS test-context-budget-check.bats: 18 tests (SPEC-086)
BATS test-tool-result-trim.bats: 9 tests (SPEC-087)
Rule context-health.md: pair integrity [SPEC-088] + proactive budget [SPEC-086]
Rule session-memory-protocol.md: pair integrity step in pipeline [SPEC-088]
Config pm-config.md: added TOOL_RESULT_MAX_CHARS constant (SPEC-087)
README counters updated across all 9 language variants (90 skills, 49 hooks, 130 test suites)

4.26.0 — 2026-04-07

Pre-audit: 5 BATS suites for critical hooks + audit report. Era 193.

Added

BATS test-block-credential-leak.bats: 10 tests (secrets, API keys, tokens, connection strings)
BATS test-block-force-push.bats: 12 tests (force push, main protection, amend, reset --hard)
BATS test-validate-bash-global.bats: 12 tests (rm -rf, chmod 777, curl|bash, PR auto-approve)
BATS test-scope-guard.bats: 8 tests (scope warnings, graceful degradation)
BATS test-session-init.bats: 8 tests (startup, PAT check, profile detection, git branch)
Audit report: pre-audit workspace analysis for critical hooks

4.25.0 — 2026-04-07

SPEC-080 LLM Training Pipeline + SPEC-022 + SPEC-032. Era 192.

Added

SPEC-080 LLM training pipeline: prepare-training-data.sh (traces → Alpaca) + import-gguf.sh (GGUF → Ollama)
SPEC-080 research: hardware verified (no local GPU), Colab strategy, 3 agent tiers, Mythos learnings
PM Keybindings 8 shortcuts for PM workflow (SPEC-022 F3)
Security Benchmarks scaffolding: Juice Shop docker, 15 known vulns, run-benchmark.sh (SPEC-032 Phase 1)

Verified

SPEC-028 (Search Reranker): already implemented (CrossEncoder in memory-vector.py)

4.23.0 — 2026-04-07

SPEC-078 + Savia Web Phase 2.5 (viewer/editor). Era 190.

Added

Hook dual-estimation-gate.sh: warns if spec/PBI missing dual estimation scale (SPEC-078 Phase 1)
FileViewer syntax highlighting via highlight.js (10 languages: JS, TS, Python, Bash, JSON, YAML, XML, CSS, SQL, C#)
FileViewer copy button + language label on code blocks, line numbers for blocks > 5 lines
FileViewer external links in new tab, lazy-load images, click-to-zoom
Editor toolbar: bold, italic, strikethrough, H1-H3, bullet/numbered/checklist, link, code, horizontal rule
Editor auto-save draft to localStorage every 30s with unsaved indicator

Verified

SPEC-019 (Memory Contradiction): already implemented (supersedes + rev)
SPEC-048 (Dev Session Discard): already implemented (11 tests pass)
Phase 2.5 backlog-filters, backlog-persistence, create-project, project-context-switch: all already implemented

4.22.0 — 2026-04-07

SPEC-079 Phase 2 + SPEC-020 completion. Era 189.

Added

legalize-es.sh history command: historial de reformas de una norma via git log
legalize-es.sh check-status command: verificar si norma está vigente o derogada
memory-store prune command: eliminar físicamente entradas expiradas (SPEC-020)
memory-store TTL stats: conteo de entradas con expiración y expiradas en stats

Changed

README ES/EN: añadido /legal-audit en sección comandos y legal-compliance en tabla agentes

4.21.0 — 2026-04-07

SPEC-079: Legal Compliance Agent powered by legalize-es (12.235 normas BOE). Era 188.

Added

Agent legal-compliance (Opus 4.6, L2): auditoría legal contra legislación española consolidada
Skill legal-compliance/: algoritmo búsqueda 3 fases, 12 dominios legales, clasificación severidad
Command /legal-audit: auditoría con scopes (rules, contract, architecture, policy, pbi, full) y soporte CCAA
Script legalize-es.sh: gestión corpus legislativo (install, update, status, search, search-article)
Reference domain-terms.md: 12 dominios legales mapeados a BOE identifiers y términos grep
SPEC-079 propuesta en docs/propuestas/

Changed

agents-catalog.md: agente #50 + flujo Legal Compliance
assignment-matrix.md: tipo "Legal compliance audit" → legal-compliance
pm-config.md: constantes LEGALIZE_ES_PATH, LEGALIZE_ES_AUTO_UPDATE, LEGALIZE_ES_DEFAULT_CCAA

4.20.0 — 2026-04-07

Savia Monitor v2.1 + Nidos fix + 150-line rule scoping. Era 187.

Added

Shield behavioral test button with progress bar and per-layer results
shield_test.rs module: sends real data to daemon/Ollama for verification
Git/Nidos polling every 30s for real-time updates
Profile polling every 5s in Shield dashboard

Fixed

CREATE_NO_WINDOW on Windows: git/tasklist no longer flash console windows
workspace_dir() walks up from exe location + scans cloud-sync folders
Nidos: Windows/OneDrive path resolution and stash handling
App.vue TypeScript error (tab.label -> tab.key)
Shield daemon auth: X-Shield-Token header + correct "text" field name
i18n keys for test button (ES/EN)

Changed

150-line rule scoped to .claude/ workspace files only — no longer applies to application source code (.rs, .ts, .vue, .py, etc.)
Cleaned 26 stale/merged local branches

4.19.0 — 2026-04-06

Savia Monitor — Desktop control tower for orchestrating multiple Claude Code sessions. Era 186.

Added

projects/savia-monitor/: Tauri v2 + Vue 3 desktop tray app (Rust backend, 7 modules)
Sessions tab: Detects active Claude Code instances via PID, shows session name, branch, agents, Shield status, health score
Shield tab: 8 protection layers with i18n tooltips, real audit feed from data-sovereignty-audit.jsonl
Git tab: Multi-project branch viewer grouped by prefix (feat/, fix/, agent/), merged branches greyed out, pending file count, nidos support
Activity tab: Real-time feed from live.log, agent-lifecycle.jsonl, audit events with type filters
Health score: Composite metric from Shield layers (35pts), git cleanliness (25pts), agent success (25pts), hook profile (15pts)
i18n: Full ES/EN with automatic system locale detection
E2E tests: 15 Playwright tests + 6 Vitest unit tests
Savia logo: Official owl from savia-web in titlebar and system tray

Changed

.gitignore: Whitelist !projects/savia-monitor/ for public tracking

4.18.0 — 2026-04-05

SPEC-061 SaviaDivergent complete runtime implementation. Era 185b.

Added

scripts/nd-autoconfig.sh: Auto-configures accessibility.md from ND profile (ADHD.rsd→review_sensitivity, dyslexia→dyslexia_friendly, giftedness→cognitive_load:high, modes→guided_work/focus_mode). Exports env vars for sensory_budget, ceremony_preview, time_blindness_markers
accessibility-setup Phase 5: Neurodivergent onboarding (5 dimensions, conversational, privacy-first)
savia-forget --neurodivergent: RGPD Art. 17 erasure of ND profile
pbi-assign strengths_map routing: ND strengths bonus (+10% scoring) for task-type matching
meeting-agenda ceremony_preview: Pre-meeting structure preview for ND users with ceremony_preview enabled

Changed

session-init.sh: Detects neurodivergent.md and launches nd-autoconfig.sh in background
accessibility-output.md: ND integration table, priority updated to include ND active_modes
neurodivergent-integration.md: Sensory budget, ceremony preview, time blindness, strengths map sections
session-memory-protocol.md: ND preferences added to Tier A (preserved across compact)
SPEC-061: Status Proposed → Implemented

4.17.0 — 2026-04-05

SPEC-061 ND Profiles + SPEC-044 Trace Optimization engine. Era 185.

Added

neurodivergent.md template: Profile schema for 5 ND dimensions (ADHD, autism, dyslexia, giftedness, dyscalculia) with active modes, sensory budget, strengths map, body double. N3 privacy
neurodivergent-integration.md: Rule connecting ND profiles to accessibility system (auto-sets review_sensitivity, dyslexia_friendly, cognitive_load)
scripts/prompt-suggestion-engine.sh: SPEC-044 Phase 2 — trace-driven prompt optimization (classifies failure patterns, generates per-agent optimization plans)
tests/test-neurodivergent-profiles.bats: 25 tests
tests/test-trace-optimization.bats: 15 tests — extractor + engine

Changed

SPEC-061: status Proposed to Implemented
SPEC-044: status Draft to Implemented (Phase 1 + Phase 2 complete)
Test suites: 122 to 124

4.16.0 — 2026-04-05

5 Draft SPECs verified with tests — Hybrid Search, Context Index, Graph Memory, Test Auditor, Live Progress. Era 184.

Added

tests/test-memory-hybrid.bats: 11 tests — SPEC-035 Hybrid Search verified (vector+graph+grep fallback)
tests/test-generate-context-index.bats: 10 tests — SPEC-054 Context Index verified
tests/test-memory-graph.bats: 10 tests — SPEC-027 Graph Memory verified (entity-relation extraction)
tests/test-test-auditor.bats: 10 tests — SPEC-055 Test Auditor verified (meta: testing the tester)
tests/test-live-progress.bats: 10 tests — SPEC-042 Live Progress verified (tool use logging)

Changed

SPECs 027, 035, 042, 054, 055: status Draft to Verified (implementations confirmed with tests)
Test suites: 117 to 122

4.15.0 — 2026-04-04

SPEC-078 Dual Estimation engine implemented. Era 183.

Added

scripts/dual-estimate.sh: Dual estimation engine — classify tasks (agent vs human), capacity planning, review bottleneck detection. 10 task types with golden rule
tests/test-dual-estimate.bats: 17 tests — classify, capacity, bottleneck, matrix, edge cases

Changed

docs/propuestas/SPEC-078: Status DRAFT to Implemented (engine complete)

4.14.1 — 2026-04-04

fix: test-architect agent upgraded to Sonnet with larger context window.

Changed

test-architect.md: model opus→sonnet, context 12K→20K, output 1K→2K, permission L1→L3, budget 13K→22K
test-architect-agent.bats: assertion updated to accept sonnet model

4.14.0 — 2026-04-04

Sovereignty Phase 2 + SPEC-021 verification + test coverage push. Era 182.

Added

scripts/sovereignty-benchmark.sh: Benchmark pm-workspace prompts with local LLMs (10 tests, quick/full modes)
tests/test-sovereignty-benchmark.bats: 11 tests
tests/test-readiness-check.bats: 17 tests (SPEC-021 verification)
Benchmark results: qwen2.5:3b scores 80% quick / 40% full — viable for basic ops, not complex tasks

Changed

Test suites: 107 to 110

4.13.0 — 2026-04-04

Sovereignty Phase 1 — multi-provider switch, CI pr-plan gate, OpenCode backup. Era 181b.

Added

scripts/sovereignty-switch.sh: Multi-provider manager (local/mistral/claude) with auto-detect and smoke test
tests/test-sovereignty-switch.bats: 17 tests
CI gate (Rule #25): PRs require /pr-plan — enforced via .confidentiality-signature check

Changed

.github/workflows/ci.yml: pr-plan gate using gh pr diff
OpenCode v1.3.13 installed as Claude Code backup
Test suites: 106 to 107

4.12.0 — 2026-04-04

SPEC implementation: Execution Supervisor, Dev Session Discard, Memory TTL verified. Era 181.

Added

tests/test-session-action-log.bats: 12 tests — log, attempts, history, reset, session isolation
tests/test-execution-supervisor.bats: 14 tests — silent on 1-2, reflection at 3+, escalation at 4+, advisory exit 0
tests/test-dev-session-discard.bats: 11 tests — lock removal, state archive, discard log, reason defaults
scripts/sovereignty-switch.sh: Multi-provider sovereignty manager (local/mistral/claude) with auto-detect and smoke test
tests/test-sovereignty-switch.bats: 17 tests — provider switching, round-trip, edge cases
CI gate (Rule #25): PRs now require /pr-plan — enforced structurally via .confidentiality-signature check in CI

Changed

docs/ROADMAP.md: SPEC-065, SPEC-048, SPEC-020 verified. Sovereignty Phase 1 complete. Era 181 documented
.github/workflows/ci.yml: Added pr-plan gate using gh pr diff
Test suites: 103 to 107 (4 new suites, all 80+ quality)
OpenCode v1.3.13 installed as Claude Code backup

4.11.0 — 2026-04-04

Granular Permissions + Test Coverage Push. Era 180.

Added

agent-permission-levels.md: 5-tier access control (L0 Observer to L4 Operator) for all 48 agents
validate-agent-permissions.sh: Validates agent permission_level matches declared tools
10 new BATS test suites: validate-ci-local, hook-profile, nidos, memory-store, emergency-plan, spec-quality-auditor, prompt-security-scan, validate-commands, validate-agent-permissions, adaptive-strategy-selector
Test coverage: 10 suites to 20 suites (100% increase), covering CI, security, workflow, and utility scripts

Changed

48 agent frontmatter: Added permission_level: field (L0: 2, L1: 13, L2: 9, L3: 18, L4: 6)
docs/ROADMAP.md: Era 180 documented, P1+P2 marked as Done

4.10.0 — 2026-04-04

Audit Correctiva — Clara Philosophy 100%, SPEC triage, dual estimation, doc coherence. Era 179.

Added

36 DOMAIN.md files: Clara Philosophy compliance from 59% to 100% (89/89 skills with dual docs SKILL.md + DOMAIN.md)
docs/propuestas/SPEC-078-dual-estimation-agent-human.md: Dual estimation framework (agent_minutes + human_hours + review_minutes + context_risk)
docs/decision-log.md: Architecture decision log with rejections and approvals
docs/best-practices-claude-code.en.md: English translation of core best practices doc
docs/memory-system.en.md: English translation of memory system doc
Spec triage: Comprehensive triage of SPECs (archive, promote, merge, keep)
Efficiency audit: Full efficiency and documentation audit
4 feature guides: Emergency Watchdog, Prompt Security Scanner, Spec Quality Auditor, Workspace Consolidation

Changed

9 READMEs (es, en, ca, de, eu, fr, gl, it, pt): Counters corrected to 508 commands, 48 agents, 89 skills, 48 hooks. Structure section aligned
7 regional READMEs (ca, de, eu, fr, gl, it, pt): Rewritten to v4.6+ benefits-first format
docs/ROADMAP.md: P0a-P0g marked as Done. SPEC status table updated with triage results
docs/politica-estimacion.md: Section 9 added (dual estimation agent/human)
.claude/skills/spec-driven-development/references/spec-template.md: Dual estimation fields added
.claude/hooks/responsibility-judge.sh: S-06 excludes DOMAIN.md and propuestas/ from reference check

4.9.0 — 2026-04-04

Workspace Consolidation — inventory audit, counter correction, roadmap sync. Era 178.

Changed

README.md + README.en.md: Counters corrected (508 commands, 48 agents, 88 skills, 47 hooks, 93 test suites)
docs/ROADMAP.md: Eras 175-178 documented. Pipeline renumbered for Era 179+. Version corrected to v4.8

Added

Workspace consolidation report: Full inventory audit (documented vs actual, test coverage map, orphaned hooks, model inventory)

4.8.0 — 2026-04-03

feat: Spec Quality Auditor — deterministic 9-criteria scorer for SDD specs. Era 177.

Added

spec-quality-auditor.sh: Scores specs 0-100 against 9 criteria (header, metadata, problem, solution, acceptance, effort, deps, testability, clarity). Batch mode with --min-score filter. JSON output. 21/73 existing specs certified at 80+
test-spec-quality-auditor.bats: 17 tests — high/low quality detection, batch mode, edge cases, JSON validation

4.7.0 — 2026-04-03

feat: Prompt Security Scanner — static analysis for injection/leakage in agent prompts. Era 176.

Added

prompt-security-scan.sh: 10-rule static analyzer (PS-01 to PS-10) for prompt injection bait, exfiltration, role hijack, credential leak, code execution, base64 blobs, PII, missing model, wildcard tools. Zero LLM — pure regex. Supports --path, --quiet, single file or directory scan
test-prompt-security-scan.bats: 17 tests — structure, clean pass, 4 detection types, edge cases, coverage breadth

4.6.0 — 2026-04-03

feat: Communication Upgrade — README benefits-first rewrite. Era 175.

Changed

README.md: Rewritten with benefits-first framework (215→148 lines). Problem→solution table, 3-minute install, role-based quick-starts
README.en.md: English version aligned with new structure
docs/ROADMAP.md: Era 174 documented as completed. Pipeline renumbered for Era 175+
settings.json: CLAUDE_CODE_NO_FLICKER=1 enabled for flicker-free terminal rendering

4.5.0 — 2026-04-03

feat: Emergency Watchdog — automatic local LLM fallback on internet loss. Era 174.

Added

savia-watchdog.sh: Systemd service that monitors connectivity to api.anthropic.com every 5 min. After 3 consecutive failures, activates Ollama with local model and notifies via wall. When internet returns, unloads model to free RAM
savia-watchdog.service: Systemd unit file (runs as current user, auto-restart)
install-watchdog.sh: One-time installer (sudo bash scripts/install-watchdog.sh)

Changed

emergency-plan.sh: Default model selection updated to Gemma 4 (e2b for 16GB, e4b for 32GB+, qwen2.5:3b for 8GB)
ollama-classify.sh: Shield default model changed from qwen2.5:7b to qwen2.5:3b (fits in available RAM alongside Claude Code)

4.4.0 — 2026-04-03

feat: Hygiene + Debt Audit — SPEC dedup, PII gate bugfix, 5 new test suites. Era 174.

Fixed

hook-pii-gate.sh: Critical bug — subshell pipe pattern caused FINDINGS counter to never propagate. PII gate was detecting patterns but never blocking commits. Fixed with process substitution + temp file counter
critical-rules-extended.md: Broken reference to non-existent confidentiality-config.md (now points to context-placement-confirmation.md)
10 scripts: Input validation for edge cases (build-skill-manifest, backlog-resolver, backlog-pbi-crud, memory-store, memory-vector, memory-search, mock-env, adaptive-strategy-selector, notify)
test-memory-vector.bats: Missing directory in setup (mkdir -p for output/)
test-mock-env.bats: Relaxed assertion for mock_mcp_response default behavior

Added

test-hook-pii-gate.bats: 11 tests covering email, DNI, IP, API key detection, binary skip, clean pass
test-confidentiality-sign.bats: 10 tests covering sign, verify, secret permissions, HMAC validation
test-backup.bats: 23 tests covering config, rotation, encryption constants, status handling
test-company-repo.bats: 9 tests covering help, args validation, dependency checks
test-emergency-plan.bats: 9 tests covering --help, --check, model selection, constants
trap error logging: 3 async hooks (live-progress, session-end-snapshot, file-changed-staleness) now log errors to ~/.savia/hook-errors.log instead of failing silently

Changed

docs/ROADMAP.md: Unified from 3 sources. Eras 165-173 documented. Pipeline P1-P6 replaces Tier/Quarter structure. 73 SPECs classified. Gemma 4 added to backlog
6 SPECs renumbered: Resolved duplicate numbers (029→070, 030→073, 031→074, 032→075, 033→076, 041→077)

Removed

13 PBI placeholders: Empty template files (PBI-013 to PBI-024 + duplicate PBI-001-no-project)

4.3.0 — 2026-04-03

feat: Memory Resilience — deep extraction with quality gates. Era 166.

Added

memory-extract-lib.sh: Shared library for memory extraction hooks (quality gates, MEMORY.md index registration, file persistence)
Discovery extraction: Detects root cause patterns (bug was, caused by, root cause, resulta que) in session context
Reference extraction: Captures URLs from session context as reference memory
Quality gates: Rejects items < 50 chars, PII (email patterns), and duplicates
MEMORY.md index registration: Extracted items automatically registered in memory index
22 BATS tests: stop-memory-extract (14) + session-end-memory (8) — first test coverage for memory hooks

Changed

stop-memory-extract.sh: Refactored with lib extraction, 4 extraction types (decisions, failures, discoveries, references)

4.2.0 — 2026-04-03

feat: Savia Emotional Regulation System — functional stress monitoring and self-regulation. Era 173.

Based on Anthropic Research "Emotion concepts and their function in a large language model" (2026-04-02), which proved that LLMs develop measurable functional emotions that causally influence behavior.

Added

emotional-state-tracker.sh: Session stress tracker (record/score/reset/status) with 5 event types and 0-10 frustration scale
stress-awareness-nudge.sh: UserPromptSubmit hook detecting pressure patterns (urgency, shame, failure attribution, corner-cutting, emotional manipulation) in ES/EN — injects calm-anchoring context
emotional-regulation-monitor.sh: Stop hook assessing session stress, persisting high-friction sessions (score 5+) to auto-memory as feedback for future sessions
emotional-regulation.md: Rule defining 5-part self-regulation protocol (detect, respond, protect, transmit calm, wellness check)
savia-emotional-regulation.md: User-facing documentation with Anthropic paper reference, frustration scale, configuration guide
SAVIA-MODEL-STANDARD.md: AI5 (Emotional Architecture), AI6 (Context Engineering from arXiv:2512.05470), AI7 (Agent Interoperability from A2A Protocol) — 21 cross-cutting concerns
ROADMAP-IMPROVEMENTS.md: Research-driven section — toolchain updates (7), new models (4), SPEC v0.2 layers (3), agent infrastructure (6) — ~512h total
47 BATS tests across 3 test files: tracker (18), nudge hook (17), monitor hook (12)

4.1.1 — 2026-04-02

fix: Shield NER false positives on technical English documentation. Era 172.

Fixed

savia-shield-daemon.py: NER allow-list, soft type filter, raised threshold (0.7 to 0.85)
shield-ner-allowlist.txt: External allow-list for technical terms (editable without code changes)

4.1.0 — 2026-04-02

feat: Savia Nidos — parallel terminal isolation via named git worktrees.

Added

scripts/nidos.sh + nidos-lib.sh: CLI for isolated git worktrees outside cloud-sync paths (create, list, enter, remove, status)
nidos-protocol.md: Domain convention for worktree naming and lifecycle
/nidos command: Slash command wrapper for the CLI
session-init.sh: Auto-detection of nido context with SAVIA_NIDO export

4.0.0 — 2026-04-02

feat: Hook System Overhaul — Claude Code alignment, Shield hardening, 61% event coverage. Era 171 (SPEC-071).

Added

subagent-lifecycle.sh: SubagentStart/SubagentStop hooks — native agent lifecycle tracking (replaces PostToolUse workaround)
task-lifecycle.sh: TaskCreated/TaskCompleted hooks — automatic task audit trail
file-changed-staleness.sh: FileChanged hook — marks code maps stale on file changes (<100ms)
instructions-tracker.sh: InstructionsLoaded hook — logs which rules load per session
config-reload.sh: ConfigChange hook — invalidates caches on settings change
SPEC-071: Hook System Overhaul proposal with 7 slices

Changed

settings.json: first type: prompt hook — semantic commit validation via Haiku ($0.0003/call, warning mode)
settings.json: first type: http hook — Shield daemon gate via native HTTP POST (SSRF guard, 5s timeout)
settings.json: 7 hooks now use if conditional field for 40% fewer unnecessary spawns (ast-comprehend, tdd-gate, post-edit-lint, ast-quality-gate, block-force-push, agent-hook-premerge, prompt-hook-commit)
settings.json: registered 7 new Claude Code events (SubagentStart/Stop, TaskCreated/Completed, FileChanged, InstructionsLoaded, ConfigChange, CwdChanged). Coverage: 17/28 events (61%)
data-sovereignty-gate.sh: daemon timeout reduced 10s to 5s, timeout events now logged to audit trail
session-init.sh: Shield daemon pre-warm on startup (reduces NER cold start)
async-hooks-config.md: corrected event count to 28 (was 27), updated coverage table

Fixed

Portability: replaced all hardcoded /tmp/ paths with $TMPDIR/$HOME/.savia/tmp in 5 hooks
Portability: replaced sed -i with portable sed + mv pattern in pbi-history-capture.sh
Portability: replaced hardcoded memory path (project-slug) with dynamic PROJ_SLUG detection in 4 hooks
SCM indexer: grep -oE '[a-z]{4,}' now includes accented chars (evaluacion, tecnico no longer truncated)
SCM indexer: head -c 120 replaced with cut -c1-120 (char count, not bytes — UTF-8 safe)
Orphan hooks: registered cwd-changed-hook.sh, compress-agent-output.sh, memory-prime-hook.sh

Documentation

Shield guide: Era 171 improvements section added in 9 languages (es, en, fr, de, it, pt, ca, eu, gl)

3.99.0 — 2026-04-01

feat: Eras 167-170 — token economics, spec validation, coordinator research, tool healing. Era batch.

Added

token-estimator.sh (Era 167): pre-calculate token cost and pricing for files/dirs without calling LLM. Supports --budget and --model flags. 10 tests
validate-spec.sh (Era 169): declarative spec validation — checks header, required sections, line count, ambiguity detection. Supports --strict mode. 9 tests
tool-call-healing.sh (Era 170): PreToolUse hook that validates parameters before execution — blocks empty paths/patterns, detects typos, suggests similar files. 9 tests
SPEC-069 (Era 168): Coordinator Mode research — documented CLAUDE_CODE_COORDINATOR_MODE and PROACTIVE env vars. Marked as research (undocumented features, not enabled)

Changed

settings.json: registered tool-call-healing.sh in PreToolUse hooks (Read|Edit|Write|Glob|Grep matcher)
data-sovereignty-gate.sh: path extraction and private-destination skip moved before daemon call (reduces latency). Windows backslash normalization added
validate-bash-global.sh: git commit/add block scoped to savia/pm-workspace repos only (allows commits in project repos like project-alpha)
SCM INDEX.scm: category rebalancing and keyword re-indexing

3.98.0 — 2026-04-01

feat: Memory Resilience + language profile support. Era 166.

Added

stop-memory-extract.sh: SPEC-013v2 deep memory extraction via Stop hook (10 min timeout vs SessionEnd 1.5s). Extracts decisions and repeated failures to auto-memory
Roadmap integrado: roadmap con aprendizajes de repos investigados + architecture review

Changed

memory-hygiene.sh: added 25KB byte limit enforcement and >150 char entry trimming (SPEC-142 enhancement)
CLAUDE.md: language directive now reads from user profile preferences.md (generic, not hardcoded)
session-init.sh: injects user language preference at session start from preferences.md
settings.json: registered stop-memory-extract.sh in Stop hooks (async, 30s timeout)

Fixed

Language switching: root cause was missing language directive + English rules bleeding into responses. Fixed by reading preferences.md language field

3.97.0 — 2026-04-01

feat: CLAUDE.md diet + hook enhancement — architecture exploit. Era 165.

Changed

CLAUDE.md: reduced from 121 to 48 lines (60% reduction) by moving inline content to @import references. Removes per-turn token waste discovered in architecture review — CLAUDE.md is NOT cached by Claude Code
pre-compact-backup.sh: SPEC-016 Tier A/B/C classification. Tier A (ephemeral) discarded, Tier B (session-hot) persisted to session-hot.md, Tier C (permanent) to memory-store
post-compaction.sh: reinjects session-hot.md context after compaction for session continuity. Refactored 5 if-blocks into compact loop (139 to 108 lines)
post-tool-failure-log.sh: 6 structured error categories (permission, not_found, timeout, syntax, network, unknown) with retry hints. Pattern detection flags 3+ same-tool failures per day

Added

critical-rules-extended.md: rules 9-25 extracted from CLAUDE.md into dedicated @import rule file
SPEC-067: CLAUDE.md diet specification
SPEC-068: Hook enhancement specification
test-hook-enhancements.bats: 12 tests for all 3 enhanced hooks (all pass)
Roadmap review: era plan based on architecture internals analysis

3.96.0 — 2026-03-31

feat: hidden features activation — new hooks, raised output limits, deferred tool loading. Era 164.

Added

SessionEnd hook: session-end-memory.sh — triggers memory extraction when session ends (native SPEC-013 support)
UserPromptSubmit hook: user-prompt-intercept.sh — pre-processes user input before Claude sees it (NL-query foundation)
CwdChanged hook: cwd-changed-hook.sh — auto-detects project context on directory change
Architecture review report: hook events and env vars documented from Claude Code internals analysis

Changed

BASH_MAX_OUTPUT_LENGTH: raised from 30K to 80K chars — prevents truncation before hook compression
TASK_MAX_OUTPUT_LENGTH: raised from 32K to 80K chars — subagents return fuller results
ENABLE_TOOL_SEARCH: set to auto — deferred tool loading for 400+ commands reduces upfront context
Hook coverage docs: updated from 9/16 to 9/27 events (corrected total available events)

3.95.0 — 2026-03-31

chore: 10 performance optimizations based on architecture review. Era 164.

Changed

Auto-compact threshold: raised from 50% to 65% of effective window (~108K tokens) for longer sessions before compaction
Context health zones: updated Gradual (50-65%) and Alerta (65-85%) to align with new auto-compact threshold
Context health rule: documented that CLAUDE.md is NOT cached (per-turn cost) — reinforces 150-line discipline
Memory system docs: added 25KB byte cap and 150-char entry length guidance for MEMORY.md
Async hooks config: updated auto-compact documentation with effective window calculation
Best practices: added section 18 "Internal Architecture Insights" with 7 key performance findings

3.94.0 — 2026-03-31

fix: comprehensive CI test portability + SPEC-065 Execution Supervisor. Era 164.

Added

SPEC-065 Execution Supervisor: session action log + retry supervisor. Pauses after 3rd failed attempt with mandatory reflection prompt. Advisory, not blocking. Integrated into pr-plan.sh and push-pr.sh. 20 tests, score 84.
scripts/session-action-log.sh: append-only JSONL session log with attempt counting.
scripts/execution-supervisor.sh: reflection trigger after repeated failures.

Fixed

44 test files: comprehensive CI portability fix. All tests now use REPO_ROOT absolute paths instead of relative $PWD references. Hook tests set required env vars. Resolves 67 CI failures in GitHub Actions.
gitignore: skill-manifests.json and backlog _config.yaml (auto-generated, broke push-pr).

3.93.0 — 2026-03-30

feat: SPEC-063 Test Architect + SPEC-060/062 SaviaDivergent. Era 164.

Added

agent: test-architect (SPEC-063): Opus-powered agent that generates tests scoring 80+ from first attempt. 8 excellence patterns, 14 test types, 16 language frameworks. Self-verification loop. Golden BATS template (90+).
SaviaDivergent (SPEC-060/061/062): neurodivergent support system with 5 adaptive modes (Focus Enhanced, Clarity, Structure, Sensory, Strengths). Evidence-based design from 13 research sources. N3 privacy profiles. 4-phase roadmap.

3.92.0 — 2026-03-30

feat: all tests >= 80 + CI quality gate enforced + SPEC-056/059 AgentScope. Era 164.

Added

CI quality gate (SPEC-055): ci-test-quality-gate.sh integrated into GitHub Actions CI. ALL tests must score >= 80 on the Test Auditor to merge.
docs/propuestas/SPEC-056 through SPEC-059 — 4 specs from AgentScope research: typed agent messages, fanout pipeline, OpenTelemetry tracing, semantic fault handlers.

Changed

57 test files improved across hooks/, structure/, evals/, scripts/ to meet 80-point quality bar. Added safety verification, edge cases, assertion diversity, spec references, isolation.
test-auditor-engine.py: expanded safety detection (recognizes grep.*set -[euo]), improved spec reference detection (# Ref:, .claude/rules/), wider assertion patterns.

3.91.0 — 2026-03-30

feat: SPEC-055 Test Auditor — test quality scoring, certification, CI gate. Era 164.

Added

scripts/test-auditor.sh + test-auditor-engine.py (SPEC-055): deterministic test quality analyzer. 9 criteria (existence, safety, positive/negative/edge cases, isolation, coverage, spec ref, assertion quality). Score 0-100, certified at >= 80. Embeds certification hash in test header.
scripts/test-coverage-checker.sh: verifies every script has a corresponding test file. JSON output.
scripts/ci-test-quality-gate.sh: CI gate — all tests >= 80 + coverage check. Exit 1 if any fail.
tests/evals/test-auditor.bats — 14 tests for the auditor itself.
docs/propuestas/SPEC-055-test-auditor.md — full spec.

3.90.0 — 2026-03-30

docs: getting-started + shield guide in all 9 languages + README links. Era 164.

Added

docs/getting-started.{ca,de,eu,fr,gl,it,pt}.md — Getting started guide translated to 7 additional languages (Catalan, German, Basque, French, Galician, Italian, Portuguese)
docs/savia-shield-guide.{ca,de,eu,fr,gl,it,pt}.md — Shield practical guide translated to 7 additional languages

Changed

README.{ca,de,eu,fr,gl,it,pt}.md — Added "First time?" links to getting-started and shield guides in each language

3.89.0 — 2026-03-30

docs: user onboarding guides + Shield practical guide + cross-references. Era 164.

Added

docs/getting-started.md + .en.md: step-by-step from clone to first command. Profile setup, Shield config, .scm/.ctx maps, role-based quickstart table (PM/TL/Dev/QA/PO/CEO).
docs/savia-shield-guide.md + .en.md: practical guide for 4 profiles (minimal/standard/strict/ci), 5 protection layers, Responsibility Judge, project configuration.
tests/structure/test-user-docs.bats — 12 tests verifying all user-facing docs.

Changed

README.md + README.en.md: added prominent "First time?" link to getting-started and shield guide
profile-onboarding.md: cross-reference to getting-started guide
hook-profiles.md: cross-reference to savia-shield-guide

3.88.0 — 2026-03-30

feat: push-pr auto-release + update.sh Shield preservation + Judge verification. Era 164.

Changed

scripts/push-pr.sh: Step 7 auto-creates/updates GitHub release after successful merge. Extracts CHANGELOG entry as release notes. Only with --merge flag + gh CLI.
scripts/update.sh: Paso 3b saves Savia Shield hook profile before pull; Paso 5b restores it after. Paso 5c verifies Responsibility Judge is active and registered post-update.

3.87.0 — 2026-03-30

feat: wire 37 agents to Context Index — full .ctx adoption across workspace. Era 164.

Changed

37 agents updated: all writers (8 digesters), readers (17 analysts/validators), and developers (12 language agents) now consult PROJECT.ctx before reading/writing project data. Writers use [digest-target], readers use [location], developers use [location] for specs.
context-health.md: added Context Index as navigation map for project context loading
digest-traceability.md: digesters must consult .ctx before writing
SPEC-054: expanded to cover all agent groups, not just digesters

Added

tests/evals/test-context-index-adoption.bats — 7 tests verifying all 37 agents reference .ctx

3.86.0 — 2026-03-30

feat: SPEC-054 Context Index System (.ctx) — knowledge map for digesters. Era 164.

Added

Context Index System (.ctx) (SPEC-054): maps WHERE all context information lives — workspace-level (WORKSPACE.ctx) and per-project (PROJECT.ctx). Entries: [location] (exists), [optional] (suggested), [intent] (search guide), [digest-target] (where to store extracted info). Serves as navigation for digesters, Savia, and humans.
scripts/generate-context-index.sh: scans workspace and project structure, generates .ctx indices. Supports --workspace, --project NAME, or both.
.context-index/WORKSPACE.ctx: workspace-level context map with sections for rules, agents, skills, memory, docs, projects.
.context-index/PROJECT-TEMPLATE.ctx: generic project template covering business rules, team, meetings, architecture, analysis, backlog, glossary, environments.
tests/evals/test-context-index.bats — 15 tests covering generation, sections, entry types, line limits, project detection.

3.85.0 — 2026-03-30

feat: SPEC-053 Savia Capability Map (.scm) + documentation alignment. Era 164.

Added

Savia Capability Map (.scm) (SPEC-053): self-knowledge index for Savia — 875 resources indexed across 7 intent-based categories (quality, development, planning, analysis, memory, communication, governance). Progressive loading: L0 INDEX.scm (~400 tokens), L1 category files, L2 full docs on-demand.
scripts/generate-capability-map.sh: scans commands, skills, agents, scripts and generates .scm index with intent keywords. Deterministic, idempotent.
tests/evals/test-capability-map.bats — 12 tests covering generation, classification, format validation.

Changed

All 9 READMEs (es, en, ca, de, eu, fr, gl, it, pt): updated counts (505 cmds, 49 agents, 85+ skills, 35 hooks) + Era 164 capabilities summary
5 scripts: added --help flag (requirement-pushback, dev-session-discard, review-depth-selector, reaction-engine, session-state-machine)

3.84.0 — 2026-03-30

feat: SPEC-047 through SPEC-052 batch implementation — 6 new capabilities. Era 164.

Added

scripts/requirement-pushback.sh (SPEC-047): analyzes specs for assumptions, ambiguities, complexity and scope risks. 4 heuristic categories. 9 tests.
scripts/dev-session-discard.sh (SPEC-048): clean abort of dev sessions — validates, logs reason to JSONL, cleans lock, archives state. 11 tests.
scripts/review-depth-selector.sh (SPEC-049): maps risk score (0-100) to review depth (quick/standard/thorough) with model and perspective selection. 14 tests.
scripts/reaction-engine.sh + reaction-engine-core.py (SPEC-050): declarative event-to-action mapping for SDD pipeline. Handles ci-failure, review-changes, test-failure with retry and escalation. 11 tests.
scripts/session-state-machine.sh (SPEC-051): 13-state lifecycle for dev sessions (spawning through merged/discarded). Validated transitions, trace events. 10 tests.
scripts/task-decomposer.sh (SPEC-052): classifies tasks as atomic/composite, recursive decomposition (max depth 3). English + Spanish connectors. 14 tests.
docs/propuestas/SPEC-050 through SPEC-052 — specs from ComposioHQ/agent-orchestrator research

3.83.0 — 2026-03-30

feat: SPEC-045 exploration collapse detection + SPEC-047/048/049 proposals. Era 164.

Added

scripts/instinct-collapse-detector.sh (SPEC-045 Phase 1): analyzes instincts registry for 3 staleness signals — AMI (activation monotony), CDS (context drift), PAR (passive acceptance rate). Classifies instincts as healthy/stale/drifted/collapsed. JSON and table output.
tests/evals/test-instinct-collapse-detector.bats — 10 tests: empty registry, missing registry, healthy/collapsed/drifted detection, disabled skip, table format.
docs/propuestas/SPEC-047-requirement-pushback.md — Requirement pushback pass (from nanostack research)
docs/propuestas/SPEC-048-dev-session-discard.md — Dev session discard mechanism (from nanostack research)
docs/propuestas/SPEC-049-depth-adjustable-review.md — Depth-adjustable review (from nanostack research)

3.82.0 — 2026-03-30

feat: SPEC-044 trace-to-prompt optimization + SPEC-043 S-02 refinement. Era 164.

Added

scripts/trace-pattern-extractor.sh (SPEC-044): analyzes agent-traces.jsonl, computes per-agent failure rate, budget overage, duration trend, token efficiency. Ranks candidates by optimization need score. 5 pattern classifiers: frequent_failures, budget_blowout, slow_execution, sparse_output, verbose_output.
command: /trace-optimize — slash command for SPEC-044 Phase 1 (analysis + dry-run)
tests/evals/test-trace-pattern-extractor.bats — 13 tests covering analysis, thresholds, patterns, edge cases

Changed

responsibility-judge.sh: removed override mechanism (no bypass allowed), tightened S-02 regex to require start-of-line annotations (reduces false positives on data strings)

3.81.0 — 2026-03-29

feat: SPEC-043 Responsibility Judge — deterministic shortcut detector hook. Era 164.

Added

hook: responsibility-judge.sh (SPEC-043): PreToolUse hook on Edit|Write that detects 7 shortcut patterns (S-01 through S-06) via regex. Blocks threshold lowering, test skipping, empty catch handlers, gate bypasses, coverage reduction, and untracked TODOs. Zero latency Layer 1 (standard profile). Override with RESPONSIBILITY_JUDGE_OVERRIDE=1 (logged to audit JSONL).
tests/hooks/test-responsibility-judge.bats — 15 tests covering all patterns, overrides, profile gating, and registration.
docs/propuestas/SPEC-043-responsibility-judge.md — Full spec with two-layer architecture (Layer 1 regex + Layer 2 LLM judge for strict profile).

Changed

settings.json: registered responsibility-judge.sh as PreToolUse hook for Edit|Write events (5s timeout, standard tier)

3.80.0 — 2026-03-29

feat: SPEC-042 live progress feedback — real-time visibility of Savia work execution. Era 163.

Added

live-progress-feedback (SPEC-042): real-time progress updates during agent execution — subprocess state machine tracks phases (ready → running → checkpoint → complete)
progress endpoint: /progress {task-id} returns JSON: phase, percentage, current_step, eta_seconds, logs_tail
checkpoint protocol: agent emits CHECKPOINT {phase} {pct} markers for heartbeat — enables timeout detection and kill-switch
Tests: test-live-progress.bats — phase transitions, checkpoint parsing, timeout detection (all green)

Changed

Removed 50 legacy PBI test stub files (PBI-004 through PBI-063) — cleanup for SPEC-042 development
Updated tests/structure/test-backlog-structure.bats for new backlog structure

3.79.0 — 2026-03-29

docs: Human Code Maps (.hcm) documentation rollout — 5 example project maps + all 9 language variants of AST strategy + full README alignment across all 9 languages + ARCHITECTURE.md update. Era 163.

Added

Human Code Maps (.hcm): Generated for 5 example projects (proyecto-alpha, proyecto-beta, sala-reservas, savia-web, savia-mobile-android). Narrative format with 6 sections: La historia, El modelo mental, Puntos de entrada, Gotchas, Por qué, Indicadores de deuda.
docs/ARCHITECTURE.md — Added dual code intelligence (.acm + .hcm) to Key Design Decisions, .human-maps/ to project directory structure, and /codemap:generate-human to Extension Points.

Changed

docs/ast-strategy.md — Updated to quadruple strategy (Comprehensión + Quality Gates + .acm Agent Maps + .hcm Human Maps). New Part 4 section: "Human Code Maps (.hcm) — La lucha activa contra la deuda cognitiva". Updated strategy diagram and References.
docs/ast-strategy.en.md — Same update in English.
docs/ast-strategy.ca.md — Same update in Catalan.
docs/ast-strategy.de.md — Same update in German.
docs/ast-strategy.eu.md — Same update in Basque.
docs/ast-strategy.fr.md — Same update in French.
docs/ast-strategy.gl.md — Same update in Galician.
docs/ast-strategy.it.md — Same update in Italian.
docs/ast-strategy.pt.md — Same update in Portuguese.
README.md, README.en.md, README.ca.md, README.de.md, README.eu.md, README.fr.md, README.gl.md, README.it.md, README.pt.md — Added Human Code Maps (.hcm) mention to AST strategy blockquote in all 9 language variants. Each .hcm entry describes cognitive debt reduction via narrative subsystem maps, commands /codemap:generate-human, /codemap:walk, /codemap:debt-report, and the 58% stat (Osmani 2024).

3.78.0 — 2026-03-29

feat: human-code-map skill — .hcm maps fighting cognitive debt; ACM system validation tests.

Added

skill: human-code-map — New skill (SKILL.md + DOMAIN.md) that generates .hcm (Human Code Maps), the human twin of .acm Agent Code Maps. 4-phase pipeline: load .acm context → debt analysis → generate narrative draft → human validation cycle. Addresses Addy Osmani's comprehension debt: devs spend 58% of time reading code; .hcm converts expensive "first walks" into reusable assets. Maturity: experimental.
.claude/skills/human-code-map/SKILL.md — Full pipeline spec: Phase 1 (load .acm + 5 max source files), Phase 2 (debt-score calculation: staleness + complexity + coverage gaps), Phase 3 (generate: La historia, El modelo mental, Puntos de entrada, Gotchas, Por qué, Indicadores de deuda), Phase 4 (human validation checklist — last-walk only updatable by human). When NOT to generate: <50 LOC, pure config, generated code, single-use scripts.
.claude/skills/human-code-map/DOMAIN.md — Domain context (Clara Philosophy). Key concepts: cognitive debt, first walk, walk-time (target 2-4 min), debt-score, gotcha. Business rules: .hcm always derived from .acm; debt-score >7 → escalate PM; last-walk only human; no validation = borrador. Upstream: agent-code-map + ast-comprehension. Downstream: onboarding + dev-session + spec-generate.
docs/rules/domain/hcm-maps.md — Canonical rule for .hcm lifecycle. debt-score formula: min((days/30)*2, 4) + complexity(0-3) + (1-coverage)*3. Lifecycle: Creation → Validation → Active → Stale → Refresh → Archive. Staleness propagation: code change → .acm hash invalid → .hcm auto-stale. Commands: /codemap:generate-human, /codemap:walk, /codemap:debt-report, /codemap:refresh-human. Directory: .human-maps/ parallel to .agent-maps/.
.human-maps/INDEX.hcm — First example of .hcm format applied to pm-workspace itself. Newspaper editorial metaphor (Commands = editor inbox, Agents = specialized journalists, Skills = reference library, Hooks = fact-checkers). 6 non-obvious Gotchas including: Rules not auto-loaded, Hooks are bash not prompts, .claude/commands/*.md IS the prompt, projects/ gitignored deny-by-default, E1 always human, SAVIA_HOOK_PROFILE controls hook tier.
ACM test report — A/B/C comparative test of .acm Agent Code Map system showing significant improvements in files explored, tool uses, tokens consumed, and duration vs baseline.

Changed

.claude/skills/agent-code-map/SKILL.md — Added "Gemelo humano: .hcm" section cross-referencing the new human-code-map skill. Documents the .acm/.hcm duality table (audience, language, content, freshness). Staleness propagation rule: if .acm hash invalid → .hcm marked stale. Final line count: 149 (at hard limit).

3.77.0 — 2026-03-29

feat: agent-code-map skill — triple AST architecture with persistent .acm cross-session context maps.

Added

skill: agent-code-map — Third pillar of the AST strategy. Complements Comprehension (PreToolUse) and Quality (PostToolUse async) with pre-generated persistent structural maps for agents. Eliminates 30–60% blind context exploration at session start.
.claude/skills/agent-code-map/SKILL.md — Full spec: /codemap:generate, /codemap:check, /codemap:load, /codemap:refresh --incremental, /codemap:stats slash commands; .acm format with sha256 hash freshness; INDEX.acm navigation table; @include progressive loading; freshness model (fresco/obsoleto/roto); SDD pipeline [0] integration; 150-line limit + auto-split rule; anti-patterns.
.agent-maps/INDEX.acm — Root navigation map for pm-workspace public structure: domain, infrastructure, api layers with element counts and priority (FAILWARNOK).
.agent-maps/domain/entities.acm — Domain entities: Commands, Skills, Agents, Hooks, Rules with file references and public API.
.agent-maps/domain/services.acm — Business services: SprintManagement, SDD Pipeline, NLCommandResolution, ContextHealth, MemorySystem.
.agent-maps/infrastructure/repositories.acm — Infrastructure access: ScriptsLayer, ProfilesStorage, ProjectsStorage, OutputStorage, HooksRuntime.
.agent-maps/api/controllers.acm — Entry points: SlashCommands catalog, HooksPipeline, AgentDispatch, UserProfile routing.

Changed

docs/ast-strategy.md (Spanish) — Updated to triple AST architecture: "dos propósitos, un árbol" → "tres propósitos, un árbol"; expanded 3-branch ASCII diagram; added Part 3 on .acm maps; added reference to agent-code-map skill.
docs/ast-strategy.en.md (English) — Triple architecture: "dual: two purposes" → "triple: three purposes"; Part 3 with .acm freshness model (fresh/stale/broken); SDD [0] LOAD step.
docs/ast-strategy.ca.md (Catalan) — Arquitectura triple; Part 3 amb estat de frescor (fresc/obsolet/trencat); integració SDD [0] CARREGAR.
docs/ast-strategy.de.md (German) — Dreifach-Architektur; Teil 3 mit Frischemodell (frisch/veraltet/defekt); SDD [0] LADEN.
docs/ast-strategy.it.md (Italian) — Architettura tripla; Parte 3 con modello di freschezza (fresco/obsoleto/rotto); SDD [0] CARICA.
docs/ast-strategy.pt.md (Portuguese) — Arquitetura tripla; Parte 3 com modelo de frescor (fresco/obsoleto/quebrado); SDD [0] CARREGAR.
docs/ast-strategy.eu.md (Basque) — Arkitektura hirukoitza; 3. zatia freskotasun-ereduarekin (fresko/zaharkitua/hautsia); SDD [0] KARGATU.
docs/ast-strategy.gl.md (Galician) — Arquitectura tripla; Parte 3 con modelo de frescura (fresco/obsoleto/roto); SDD [0] CARGAR.
docs/ast-strategy.fr.md (French) — Architecture triple; Partie 3 avec modèle de fraîcheur (frais/obsolète/cassé); SDD [0] CHARGER.

3.76.0 — 2026-03-29

feat: ast-comprehension skill — dual-purpose AST for legacy code understanding and pre-edit structural context injection.

Added

skill: ast-comprehension — Companion to ast-quality-gate. Quality-gate asks "¿tiene errores el código generado?"; comprehension asks "¿qué hace este código ajeno antes de tocarlo?". Pre-edit context injection via PreToolUse hook.
scripts/ast-comprehend.sh — Multi-language structural extractor. 3-layer pipeline: tree-sitter (universal AST) → language-native semantics (python ast module, ts-morph, gopls, Roslyn) → grep-structural fallback (0 deps). Supports 16 language packs. Flags: --surface-only, --legacy-mode, --output <path>.
.claude/skills/ast-comprehension/SKILL.md — Dual-use design, 3-layer extraction architecture, PreToolUse hook integration, CLI usage, pre-edit context injection pattern
.claude/skills/ast-comprehension/DOMAIN.md — Clara Philosophy: WHY (agents fail modifying code they don't understand), domain concepts (structural map, API surface, hotspot, call graph), 5 business rules (RN-COMP-01..05)
.claude/skills/ast-comprehension/references/comprehension-schema.md — Unified JSON output schema: meta, structure (classes/methods/properties/constants/enums), imports (internal/external/standard), complexity (hotspots), api_surface, call_graph, summary. 3 extraction levels (L1 grep-structural 70%, L2 tree-sitter 95%, L3 native semantic 100%) with degradation defaults.
.claude/skills/ast-comprehension/references/extraction-commands.md — CLI extraction commands per language: Python ast.walk(), ts-morph, Roslyn SyntaxWalker, go doc + gopls, javap, cargo check, rubocop, php-parser, sourcekitten, detekt, dart analyze, terraform tflint, universal tree-sitter, universal grep-structural fallback
.claude/hooks/ast-comprehend-hook.sh — PreToolUse hook (NOT async): fires before Edit on files >50 lines, runs --surface-only extraction, injects structural map as context. Always exits 0 — never blocks (RN-COMP-02). Timeout 15s.
docs/ast-strategy.md — Technical document explaining Savia's intelligent AST strategy: dual-purpose architecture, language-agnostic quality gates (QG-01..QG-12), pre-edit comprehension pipeline, degradation guarantees
docs/ast-strategy.en.md — English translation of ast-strategy.md

Changed

settings.json: registered ast-comprehend-hook.sh as PreToolUse hook for Edit events (non-async, 15s timeout, fires before ast-quality-gate PostToolUse)

3.75.0 — 2026-03-29

feat: ast-quality-gate skill — language-agnostic code quality verification for AI-generated code.

Added

skill: ast-quality-gate — Meta-analyzer for 16 language packs. Detects 12 quality gates (QG-01..QG-12) covering the 5 most common LLM error patterns: async misuse, N+1 queries, null dereference, magic numbers, empty catch blocks
scripts/ast-quality-gate.sh — Shell meta-analyzer: detects language by extension/project files, routes to native linter (eslint/ruff/golangci-lint/cargo clippy/dotnet build/phpstan/swiftlint/detekt/rubocop/tflint/dart analyze), runs Semgrep for universal LLM patterns, normalizes outputs to unified JSON schema, computes penalty-based score 0-100
.claude/skills/ast-quality-gate/SKILL.md — 3-layer architecture docs (native precision + Semgrep coverage + LSP semantics), 12 QG table, pipeline steps, SDD integration, CLI usage flags
.claude/skills/ast-quality-gate/DOMAIN.md — Clara Philosophy dual-documentation: WHY, domain concepts, 5 business rules (RN-AST-01..05), upstream/downstream relationships
.claude/skills/ast-quality-gate/references/unified-schema.md — JSON contract {meta, score, issues[], summary} normalizing ESLint, Ruff, SARIF, Cargo JSON, SpotBugs XML
.claude/skills/ast-quality-gate/references/semgrep-rules.yaml — 20 Semgrep rules across QG-01..QG-10 covering TypeScript, JavaScript, Python, Java, Go, C#, Ruby, PHP, Kotlin (10 languages per rule)
.claude/skills/ast-quality-gate/references/language-commands.md — CLI commands, verification snippets, and jq normalization templates per language
.claude/hooks/ast-quality-gate-hook.sh — PostToolUse async hook: triggers gate after Edit|Write on source files, writes to output/quality-gates/latest.json

Changed

settings.json: registered ast-quality-gate-hook.sh as async PostToolUse hook for Edit|Write events (background execution, 60s timeout)

3.74.0 — 2026-03-28

fix: workspace audit — security hooks never called, sovereignty bug, catalog sync. Era 162.

Fixed

data-sovereignty-audit.sh: premature rm -f on undefined $NORM_FILE (line 46 before definition at 85); malformed printf with literal \n; misleading exit 0 indentation — all three bugs corrected
data-sovereignty-audit.sh: changed PostToolUse entry from async: false to async: true — was blocking every Edit/Write event unnecessarily

Added

settings.json: registered 4 security hooks that existed on disk but were never executed: block-credential-leak.sh, block-force-push.sh, block-infra-destructive.sh, tdd-gate.sh — 100% BATS pass rate but zero production coverage gap now closed
agents-catalog.md: documented feasibility-probe (bypassPermissions, Opus 4.6) and model-upgrade-auditor (Opus 4.6) — catalog now at 49 agents

Changed

Count references synced across CLAUDE.md, README.md, pm-workflow.md: 505 commands, 49 agents, 85 skills, 31 hooks

3.73.0 — 2026-03-28

5 specs implemented: wave-executor, G11 review scaling, output compression, agent metering, skill feedback. Era 161.

Added

wave-executor (SPEC-WAVE-DAG): scripts/wave-executor.sh + lib — generic parallel task execution engine for DAG scheduling
G11 gate (SPEC-PR-REVIEW-SCALING): PR review depth scaling — XS/STANDARD/ENHANCED/FULL tiers by lines changed + risk score
output-compress (SPEC-OUTPUT-COMPRESS): scripts/output-compress.sh — standalone compression with 7 command-specific filters (60-90% reduction)
agent metering (SPEC-AGENT-METERING): token_budget in 44 agent frontmatter + budget tracking + budget-alerts.jsonl
skill feedback (SPEC-SKILL-FEEDBACK): /skill-rank command + skill-feedback-log.sh + skill-feedback-rank.sh — effectiveness tracking via data/skill-invocations.jsonl
Tests: test-wave-executor.bats, test-output-compress.bats, test-pr-review-scaling.bats, test-agent-budget-lookup.bats

Changed

agent-trace-log.sh logs token_budget and budget_exceeded per invocation
pr-plan-gates.sh G11 uses risk-score.json for tier escalation
.gitignore updated with data/skill-invocations*.jsonl*

3.72.0 — 2026-03-28

feat: deepagents context engineering — SPEC-138/140/141/142/144 implementados (Era 160).

Added

SPEC-138 Token-tracker middleware: .claude/hooks/token-tracker-middleware.sh — PostToolUse hook async que monitoriza CLAUDE_CONTEXT_TOKENS_USED/MAX y emite alertas por zonas (verde <50%, gradual 50-70%, alerta 70-85%, crítica 85%+); activa scripts/auto-compact.sh en background al superar el 85%
SPEC-140 Progressive skill disclosure: scripts/build-skill-manifest.sh genera .claude/skill-manifests.json con name/description/path/tokens_est de los 83 skills; .claude/commands/skill-read.md carga el SKILL.md completo bajo demanda (95% reducción de tokens en carga inicial)
SPEC-141 Tool-call healing: .claude/hooks/agent-tool-call-validate.sh — PreToolUse hook que bloquea file_path vacío en Edit/Write/Read y command vacío en Bash con mensaje de diagnóstico; 8/8 tests
SPEC-142 Memory hygiene: scripts/memory-hygiene.sh archiva entradas >90 días, deduplica MEMORY.md, trunca a 200 líneas; soporte DRY_RUN=true; 9/9 tests
SPEC-144 Context-aware skill loader: scripts/skill-loader.sh --task "..." --budget N — keyword scoring + greedy token-budget packing; selecciona skills relevantes según la tarea sin cargar todos
scripts/auto-compact.sh: companion de token-tracker; guarda snapshot JSON en output/context-snapshots/ al compactar
Tests nuevos: tests/scripts/{test-skill-manifest,test-skill-loader,test-memory-hygiene}.bats, tests/hooks/{test-token-tracker-middleware,test-tool-call-validate}.bats — 45 tests añadidos

Changed

.claude/settings.json: añadidos agent-tool-call-validate.sh (PreToolUse) y token-tracker-middleware.sh (PostToolUse async)
scripts/build-skill-manifest.sh: acepta segundo argumento opcional para output path (aislamiento en tests)
docs/ROADMAP.md: sección Tier 0+ DeepAgents con 8 SPECs y estado de implementación

3.71.0 — 2026-03-28

Hook profiles system (SAVIA_HOOK_PROFILE) + 5 specs (wave-executor, G11 review scaling, output compression, agent metering, skill feedback). Era 159.

Added

SAVIA_HOOK_PROFILE: nueva variable de entorno que controla qué hooks se activan según el contexto de trabajo — 4 perfiles: minimal, standard (default), strict, ci
.claude/hooks/lib/profile-gate.sh: librería compartida con función profile_gate() — sourcing condicional, sin dependencias externas
/hook-profile: nuevo comando slash para consultar y cambiar el perfil activo (get, set, list)
scripts/hook-profile.sh: script CLI que persiste el perfil en ~/.savia/hook-profile
docs/rules/domain/hook-profiles.md: regla que documenta la arquitectura de perfiles, jerarquía de tiers y el principio "hooks > prompts"
wave-executor (SPEC-WAVE-DAG): scripts/wave-executor.sh + lib — generic parallel task execution engine for DAG scheduling
G11 gate (SPEC-PR-REVIEW-SCALING): PR review depth scaling — XS/STANDARD/ENHANCED/FULL tiers by lines changed + risk score
output-compress (SPEC-OUTPUT-COMPRESS): scripts/output-compress.sh — standalone compression with 7 command-specific filters
agent metering (SPEC-AGENT-METERING): token_budget in 44 agent frontmatter + budget tracking + budget-alerts.jsonl + enhanced /agent-cost
skill feedback (SPEC-SKILL-FEEDBACK): /skill-rank command + skill-feedback-log.sh + skill-feedback-rank.sh — effectiveness tracking

Changed

29 hooks clasificados: todos los hooks de bloqueo ahora incluyen profile_gate — seguridad (5), estándar (10), estricto (3), siempre-activos (11)
README.md y README.en.md: nueva sección "Aprendizaje clave: hooks > prompts" — el hallazgo arquitectónico más importante de pm-workspace, emergido de forma independiente en gstack, ECC y Astromesh
agent-trace-log.sh logs token_budget and budget_exceeded per invocation
bash-output-compress.sh delegates to scripts/output-compress.sh
pr-plan.sh runs 11 gates (G0-G11)
.gitignore updated with data/skill-invocations*.jsonl*

3.70.4 — 2026-03-28

Era 158. /pr-plan enforcement: sentinel gate + Rule #25 + lessons + memory.

Added

Rule #25 in CLAUDE.md: SIEMPRE /pr-plan antes de cualquier PR. NUNCA push-pr.sh directamente.
Sentinel .pr-plan-ok: pr-plan.sh escribe el token tras pasar los 10 gates; push-pr.sh lo requiere o termina con error explicativo
Flag --from-pr-plan: bypass interno para que pr-plan.sh pueda llamar a push-pr.sh sin activar el gate
lessons.md: nueva entrada sobre el fallo de CI en PR #441 (causa: signing sin /pr-plan)
Auto-memory: feedback_push_pr.md actualizado con la regla del sentinel y Rule #25

Fixed

push-pr.sh: Step 0 gate — falla con error claro si se llama sin sentinel y sin --from-pr-plan
push-pr.sh: limpia .pr-plan-ok al final para que el siguiente PR requiera volver a ejecutar /pr-plan
pr-plan.sh: limpia .pr-plan-ok en --skip-push para no dejar sentinel huérfano

3.70.3 — 2026-03-28

3.70.2 — 2026-03-27

pr-plan feedback + push-pr.sh gh CLI fallback (Era 157).

Added

Sistema anti-atajo (G0): fallo registra fichero causa — no permite reintentar sin fix de causa raiz
/task-create: comando slash para añadir tareas al todo list
pr-plan-gates.sh: gates extraidos para mantener ficheros bajo 150 lineas

Fixed

validate-ci-local.sh: checks en paralelo — de 525s a 120s (4x speedup)
pr-plan.sh: feedback "..." antes de cada gate + tiempo transcurrido
pr-plan.sh G5: verifica Era reference en entradas CHANGELOG
push-pr.sh: gh CLI para crear PRs + force-with-lease + --skip-ci + filtro commits firma
push-pr.sh: body auto-generado excluye commits "chore: sign"

3.70.1 — 2026-03-27

pr-plan G5 fix: docs-only PRs exempt from CHANGELOG requirement (Era 157).

Fixed

pr-plan.sh G5: docs-only PRs (.md files) no longer require CHANGELOG, aligned with PR Guardian Gate 8
G5 high-impact patterns: removed docs/ from list (docs changes are not code changes)

3.70.0 — 2026-03-27

PR pre-flight protocol: 10-gate checklist before push/PR (Era 157).

Added

/pr-plan: comando con 10 gates secuenciales pre-push (branch, clean tree, conflicts, divergence, CHANGELOG, BATS, confidentiality, docs, leakage, CI)
scripts/pr-plan.sh: script standalone ejecutable desde terminal
SPEC-PR-PLAN.spec.md: spec SDD del protocolo de PR

3.69.0 — 2026-03-27

Shield docs: arquitectura 5 capas, regla zero-project-leakage, CI fixes (Era 157).

Added

zero-project-leakage.md: regla que prohibe datos derivados de proyectos privados en docs publicos
Capa 0 (proxy API): documentada como capa independiente en las 9 traducciones
Gate fallback: restauradas detecciones IP, SAS, base64, NFKC, cross-write, path traversal, Ollama Layer 2

Changed

Shield docs (9 idiomas): arquitectura 5 capas, sin conteos de vulnerabilidades ni datos privados
data-sovereignty-gate.sh: regex corregido, audit log en fallback, Ollama L2
block-force-push.sh: eliminado bloqueo catch-all de push (solo force-push y main/master)

Fixed

CI: patrones de credenciales escapados en scripts Python, CHANGELOG, PII scan

3.68.0 — 2026-03-27

Savia Shield: rewrite completo de la capa de soberania de datos con arquitectura unificada (Era 157).

Added

savia-shield-daemon.py: daemon unificado (scan/mask/unmask/health) en localhost:8444 — reemplaza multiples hooks con un unico proceso persistente
savia-shield-proxy.py: proxy API entre Claude Code y Anthropic — intercepta prompts, enmascara entidades, desenmascara respuestas
shield-ner-daemon.py: NER con Presidio/Ollama como background process (sin cold-start por invocacion)
block-force-push.sh: hook de seguridad contra force-push

Changed

data-sovereignty-gate.sh: simplificado de 232 a ~80 lineas — regex-only, sin dependencia LLM en ruta critica
settings.json: eliminadas entradas de hooks redundantes
Documentacion Shield: actualizada en 9 idiomas (ca, de, en, es, eu, fr, gl, it, pt)

3.67.0 — 2026-03-27

SPEC-041: Estrategia global de optimización de memoria y contexto. Inspirado en TurboQuant (arXiv:2504.19874).

Added

P1 Compactación por Tiers: clasificación A/B/C en pre-compact — verbatim, bullets comprimidos, descarte. Retención semántica ~85% vs. ~20% anterior
P3 Gate de calidad en memoria: campo quality: high|medium|low|unverified + questions[] generadas al guardar entradas Tier A
P4 Compresión streaming de agentes: hook compress-agent-output.sh (async PostToolUse/Task) — outputs >200 tokens comprimidos a bullets en sesiones dev-session activas
P5 Importance tiers en búsqueda: campo importance_tier: A|B|C auto-asignado; Tier A pondera 3× en ranking de memoria
scripts/memory-verify.sh: herramienta de verificación de calidad post-compresión (verify, check-all)
Guías: docs/guides/guide-context-memory-optimization.md (ES) + docs/guides_en/guide-context-memory-optimization.md (EN)

Changed

P2 Umbrales de contexto calibrados: 4 zonas basadas en evidencia del paper — Verde <50% (sin acción), Gradual 50-70% (sugerir), Alerta 70-85% (bloquear pesados), Crítico >85% (bloquear todo)
context-health.md: zonas recalibradas con base científica (TurboQuant arXiv:2504.19874)
scoring-curves.md: curva de uso de contexto actualizada con labels de zona
session-memory-protocol.md: lógica de tiers A/B/C integrada en flujo pre-compact
memory-save.sh: nuevos campos importance_tier, quality, questions[] (retrocompatibles)

3.66.0 — 2026-03-26

Era 149. Browser agents, Teams/DevOps readers, memory sync + backup.

Added

Scripts: browser-daemon.py — persistent off-screen Playwright browser per account for email/calendar monitoring
Scripts: inbox-check.py — read Outlook Web inbox via browser session (direct mode + daemon mode)
Scripts: teams-check.py — read Teams Web chats, activity feed, and channels via browser session
Scripts: devops-read.py — READ-ONLY Azure DevOps scraper (backlog, board, sprint taskboard)
Scripts: memory-sync-index.sh — sync auto-memory markdown files to JSONL vector store (markdown remains source of truth)
Scripts: memory-backup-pm.sh — AES-256 encrypted backup of memory indices to PM repo with manifest verification
Rules: transcription-resolution.md — ASR error correction using project phonetic maps and context dictionaries

Changed

Scripts: memory-search.sh — detect FAISS index (.faiss) in addition to hnswlib (.idx) for vector search fallback
Scripts: memory-store.sh — improved dispatcher with background index rebuild detection
Scripts: memory-vector.py — improved embedding generation, FAISS support, batch processing
Agents: meeting-digest.md — added Phase 0 (transcription resolution) before extraction
Config: .gitignore — added scheduled_tasks.lock and .claude/sessions/

3.65.0 — 2026-03-26

Era 156. SaviaClaw survival system — autonomous three-phase health monitoring with SSH self-healing.

Added

ZeroClaw: Three-phase survival system (Latido, Respiración, Despertar) in survival.py and survival_phases.py — monitors heartbeat (disk/memory), verifies bridge connectivity via SSH, validates Claude Code responsiveness
ZeroClaw: SSH self-healing in remote_host.py — restarts bridge on remote server if down, wakes Claude Code if unresponsive, escalates to Talk if max failures exceeded
ZeroClaw: setup-savia-remote.sh — one-time provisioning script for remote server: creates savia user, configures SSH key-only auth, whitelists allowed commands via allowed-cmds.sh
ZeroClaw: remote-host-config.example — template configuration for ~/.savia/remote-host-config with generic naming (REMOTE_HOST, REMOTE_SSH_USER, REMOTE_SSH_KEY)
Docs: Immovable Privacy Principle documented in code: remote server contains family personal data; savia user has zero access to other users' directories

3.64.0 — 2026-03-26

Savia Shield — Enterprise data sovereignty for AI-assisted workflows.

Added

Savia Shield: 4-layer defense system preventing client data leakage to cloud LLM APIs
Hooks: data-sovereignty-gate.sh (PreToolUse, blocking) + data-sovereignty-audit.sh (PostToolUse, synchronous)
Scripts: ollama-classify.sh (local LLM classifier), sovereignty-mask.py/.sh (reversible entity masking), shield-ner-scan.py + shield-ner-hook.sh (NER via Presidio), pre-commit-sovereignty.sh (git hook), savia-shield-setup.sh (installer)
Rules: data-sovereignty.md domain rule with 5 confidentiality levels (N1-N4b)
Docs: savia-shield.md in 9 languages (ES, EN, CA, DE, EU, FR, GL, IT, PT) + 4 technical docs (architecture, operations, auditability, finetune plan)
Tests: 51 BATS tests (core + edge cases + bypass attempts + fix verification + mock Ollama)
README: Savia Shield linked in all 9 README language versions

Security

3 independent audits: Red Team, Code Review, Confidentiality — all findings resolved
Defenses: NFKC unicode normalization, sandwich prompt injection defense, cross-write split detection, HMAC audit chain, chmod 600 mask maps, dual-language NER scan

3.63.0 — 2026-03-25

Era 155. Memory Intelligence from Qwen-Agent — memory-agent + humanized SaviaClaw notifications.

Added

Agents: memory-agent.md — conversational memory agent (SPEC-029). Recall, save, stats and forget via natural language. Model: haiku
ZeroClaw: Humanized Talk notifications in consciousness.py — human-readable Spanish messages instead of technical strings
ZeroClaw: _SILENT_TASKS set — suppresses known-broken tasks (memory-consolidate) from spamming notifications
ZeroClaw: _FAILURE_MSGS dict + _notify_failure() / _notify_success() helpers
ZeroClaw: poll_and_respond() now injects Savia persona context into claude prompt and uses empathetic fallback
Specs: SPEC-029 through SPEC-033 from Qwen-Agent research (Memory Agent, Keyword Expansion, Parallel Doc QA, Capability Router, PENDING_USER_INPUT)
Roadmap: era21-masterplan.md updated with WS8 — Memory Intelligence workstream (8 workstreams total)

3.62.0 — 2026-03-24

Era 154. Ellipsis guardrail — rhetorical dots are not truncation.

Added

Rules: ellipsis-guardrail.md — never assume text is incomplete based on ... alone. Fixes false "message seems cut" on complete emails

3.61.0 — 2026-03-24

Era 153. Bidirectional Talk — user messages Savia via Nextcloud, Savia responds autonomously.

Added

ZeroClaw: poll_and_respond() in nctalk.py — reads user messages, launches claude headless, sends response back to Talk
ZeroClaw: check-talk scheduled task in consciousness (every 2 min) — bidirectional communication loop

3.60.0 — 2026-03-24

Era 152. SaviaClaw talks — Nextcloud Talk integration for autonomous messaging.

Added

ZeroClaw: nctalk.py — send/read Nextcloud Talk messages. SaviaClaw can message user autonomously via localhost
ZeroClaw: consciousness notifies via Talk on task failure or important results
Config: ~/.savia/nextcloud-config stores credentials locally (never in repo)

3.59.0 — 2026-03-24

Era 151. Memory prime hook + SaviaClaw consciousness — persistent daemon with identity and scheduler.

Added

Hooks: memory-prime-hook.sh — auto-primes memory context on user prompts, logs access for forgetting curve
ZeroClaw: consciousness.py — scheduler runs tasks autonomously: heartbeat (5m), sensors (10m), git-status (30m), memory-consolidate (60m via claude headless)
ZeroClaw: identity.json — SaviaClaw self-identity: name, role, purpose, principles. Read on boot
ZeroClaw: daemon integration — consciousness tick in main loop, 3 task types (device/shell/claude)

3.57.0 — 2026-03-24

Era 149. Foundational Principles (immutable) + fix update.sh bug.

Added

Rules: savia-foundational-principles.md — 7 immutable principles burned into the repo. No human, no agent, no Savia instance can violate them. Sovereignty, honesty, privacy, human-decides, equality, identity protection.

Fixed

Scripts: update.sh — $REPO_DIR (undefined) corrected to $WORKSPACE_DIR in post-update readiness check

3.56.0 — 2026-03-24

Era 148. E2E pipeline benchmark — honest proof that each layer adds (or doesn't add) value.

Added

Scripts: benchmark-context-pipeline.py — compares 5 levels (none → grep → domain → prime → brain) with precision, noise, tokens, latency
Results: L3 (auto-prime) improves grep precision +32%. L4 (brain) compresses tokens 70%. L2 (domain routing) too aggressive on small stores.
Docs: honest benchmark report with root cause analysis and production recommendations

3.55.0 — 2026-03-24

Era 147. SPEC-041: Brain-Inspired Context Reasoning Engine — pre-LLM intelligence.

Added

SPEC-041: 4 brain mechanisms for context selection before LLM: Working Memory Gate (MUST/USEFUL/NOISE), Contradiction Detection (hippocampus), Priority Tagging (amygdala), Attention Focus (narrow/medium/wide zoom)
Scripts: context-reasoning.py — pure arithmetic, no LLM calls. 100% accuracy on 6-query benchmark
Tests: 8 BATS tests including accuracy and zoom detection verification

3.54.0 — 2026-03-24

Era 146. Forgetting curve loop closed + 8-layer memory architecture documentation.

Changed

Scripts: context-auto-prime.py — Ebbinghaus forgetting curve integrated into scoring. Memories accessed frequently strengthen; unused ones decay. Access tracker feeds the curve.
Docs: memory-architecture.md — complete 8-layer architecture explained for humans. Formulas, diagrams, numbers, analogies. Written by Savia in first person.

3.53.0 — 2026-03-24

Era 145. Prefetch cache + access tracking — EXP-02 from experiment to production.

Added

Scripts: context-prefetch.py — Markov prefetch cache (predict next command, pre-load domain context) + memory access tracker (feeds forgetting curve)
Tests: 8 BATS tests including hypothesis verification (top-3 >= 90%)

Changed

Architecture: memory stack now has predictive layer: save → classify → index → search → auto-prime → prefetch next

3.52.0 — 2026-03-24

Era 144. SPEC-040: Memory R&D — 3 scientific experiments on agentic memory.

Added

SPEC-040: Memory Research Experiments with hypotheses, methods, and measurements
Scripts: memory-experiments.py — Ebbinghaus forgetting curve, Markov workflow prediction, semantic consolidation
Results: EXP-02 CONFIRMED (Markov top-3: 100%, top-1: 95%). EXP-01/03 need larger datasets.
Tests: 9 BATS tests including hypothesis verification (top-3 accuracy >= 70%)

3.51.0 — 2026-03-24

Era 143. SPEC-039: Context Auto-Priming — memory that loads itself. Frontier context management.

Added

SPEC-039: Context Auto-Priming — scores memories by domain + keywords + recency + importance, loads relevant ones automatically before any task
Scripts: context-auto-prime.py — arithmetic scoring (no LLM), 0.6ms avg latency, silent on trivial queries
Tests: 11 BATS tests + benchmark (100% prime accuracy, 100% domain accuracy, 25% silence rate)

3.50.0 — 2026-03-24

Era 142. Memory integration — domain routing feeds hybrid search, auto-classify on save.

Changed

Scripts: memory-hybrid.py — SPEC-038 domain pre-filter integrated into grep fallback path. Queries auto-classified, results filtered by knowledge domain before scoring
Scripts: memory-save.sh — auto-assigns domain field on save using SPEC-038 classifier. Entries get sector (037) + domain (038) automatically
Scripts: compacted hybrid (178→124 lines) and save (148→150 lines)

3.49.0 — 2026-03-24

Era 141. SPEC-038: Knowledge domain routing — memory search partitioned by knowledge domains (user insight from human team organization).

Added

SPEC-038: Knowledge Domain Routing — 8 domains (security, architecture, sprint, quality, devops, team, product, memory) with keyword classifier and domain index
Scripts: memory-domains.py — classify queries, rebuild index, domain-routed search, benchmark (routed vs full)
Tests: 13 BATS tests + 20-entry benchmark store for domain routing validation
Benchmark: domain classifier 100% accuracy on 8 test queries; speedup appears at scale (>100 entries)

3.48.0 — 2026-03-24

Era 140. First agent benchmarks: baseline + adversarial golden sets. All 3 agents pass.

Added

Tests: adversarial golden sets — obfuscated SQLi (f-string), SSRF, clean diff (must approve), well-written PBI (must VALIDO)
Evals: baseline + adversarial benchmark results for security-attacker, code-reviewer, business-analyst
Docs: BENCHMARK-20260324-baseline.md and BENCHMARK-20260324-adversarial.md with action plan

Changed

Tests: expanded BATS to verify adversarial pairs (4 pairs security-attacker, 2 each reviewer+BA)

3.47.0 — 2026-03-24

Era 139. SPEC-036: Agent Evaluation Framework — golden sets, metrics, regression detection.

Added

Tests: tests/evals/ directory structure with golden sets for 3 critical agents (security-attacker, code-reviewer, business-analyst)
Scripts: eval-agent.sh — runner that generates eval templates, lists agents, detects regressions (>10% drop)
Commands: /eval-agent — evaluate agent quality against golden sets (precision, recall, F1, hallucinations, bias)
SPEC-036: Agent evaluation framework with 4 dimensions (precision, coherence, bias, hallucination)

Changed

SPEC-036: status DRAFT -> APPROVED

3.46.0 — 2026-03-24

Era 138. SPEC-035: Hybrid search (vector + graph + grep) with fallback chain.

Added

Scripts: memory-hybrid.py — combines vector similarity, graph traversal, and grep. Dedup-merge with multi-source boost. Fallback: hybrid → vector → graph → grep (always works)
Memory: hybrid mode in memory-search.sh as default search strategy

Changed

Scripts: memory-search.sh — new --mode hybrid (default), respects SPEC-034 (skips superseded) and SPEC-037 (sector filter)

3.45.0 — 2026-03-24

Era 138. SPEC-034/037: Temporal memory + cognitive sectors. Ecosystem research (30+ repos).

Added

Memory: temporal validity fields (valid_from, valid_to, superseded_by) in memory-save.sh
Memory: cognitive sectors (episodic/semantic/procedural/referential/reflective) with independent decay
Memory: --supersedes flag to mark old decisions as superseded (not deleted)
Memory: --sector and --include-superseded filters in memory-search.sh
Specs: SPEC-034 (temporal memory), SPEC-035 (hybrid search), SPEC-036 (agent evaluation), SPEC-037 (cognitive sectors)
Roadmap: Tier 1-4 evolution vision from 30+ repo analysis (700K+ combined stars)

Changed

Rules: context-aging.md rewritten with sector-based decay (5 sectors, independent half-lives)
Roadmap: updated to v3.45.0 with sources from CrewAI, Graphiti, LightRAG, OpenMemory, DeepEval

3.44.0 — 2026-03-23

Era 137. SPEC-029/030/031 implemented + internal audit (code, security, docs) + 6 fixes.

Added

Commands: workspace-doctor (SPEC-031) — 14-check health check, jato-inspired
Commands: security-auto-remediation (SPEC-029) — PR Draft from validated fixes
Commands: security-pipeline updated with Fase 4 auto-PR + Nuclei integration
Skills: nuclei-scanning (SPEC-030) — CVE scanner with graceful degradation
Specs: SPEC-029 to SPEC-033 (security PRs, Nuclei, doctor, benchmarks, modular skills)
Scripts: workspace-doctor.sh — 14 checks (critical/important/recommended)

Fixed

Security: regex injection in spellcheck-docs.sh via dictionary (M-01) — use grep -F
Scripts: duplicate set -uo pipefail in pr-context-loader.sh and semantic-compact.sh
Docs: 5 broken links in README.en.md (ES filenames instead of EN)
Docs: untranslated Spanish text in 6 READMEs (CA/GL/DE/FR/IT/PT line 44)
Scripts: workspace-doctor.sh grep conflict marker false positive fixed

Changed

Roadmap: updated with SPEC-029-033 prioritized in Q2-Q3 2026
Sources: added jato + strix to roadmap sources

3.43.0 — 2026-03-22

SPEC-022 complete + spellcheck + roadmap sync + accent fixes across 12 files.

Added

Scripts: semantic-compact.sh (F2) — smart compact summary from git diff, memory, trace. Integrated in pre-compact hook.
Scripts: pr-context-loader.sh (F4) — loads project rules, team, specs, decisions before PR. Suggests reviewers.
Scripts: spellcheck-docs.sh — multilingual spelling checker (ES/GL/CA/FR). Detects common accent errors per language. 0 false positives.
Tests: 6 new tests in test-power-cli.bats (F2 compact + F4 PR context)

Fixed

Docs: Accent fixes across 12 files: memory-architecture.md, README.md, README.fr.md, README.ca.md, SECURITY.md, confidentiality-levels.md, zeroclaw/ROADMAP.md, CHANGELOG.md. ~40 accent corrections.
Docs: ROADMAP.md synced to v3.42.0 (Eras 134-136 added)

Changed

Hooks: pre-compact-backup.sh now calls semantic-compact for richer summaries

3.42.0 — 2026-03-22

Rewrite memory-architecture.md — human-friendly, complete, 379 lines.

Changed

Docs: memory-architecture.md — completely rewritten for non-technical readers. Explains with analogies (index of a book, not the book itself). Covers: project knowledge (meetings, team, rules, stakeholders, decisión log), 7 digesters and their dual output (project markdown + central memory), contradiction tracking, TTL, search levels, privacy. New section "El panorama completo" connects everything.

3.41.0 — 2026-03-22

Memory Architecture doc — Savia explains her memory system in first person.

Added

Docs: memory-architecture.md — complete guide to the 4-layer memory system (auto-memory, JSONL, vector index, graph). Written in Savia's first person. Emphasizes: plain text files readable by humans are the source of truth. Indices are derived accelerators. Covers all 3 input flows (session, digesters, manual), search order, degradation levels, privacy.

3.39.0 — 2026-03-22

SPEC-027 Phase 1 — Graph memory layer + roadmap sync.

Added

Scripts: memory-graph.py — entity-relation extraction from JSONL using regex+heuristics. Extracts technology names, capitalized terms, concepts, projects. Builds JSON graph (entities + relations). Commands: build, search, entities, status.
Scripts: memory-store.sh — 4 new subcommands: build-graph, graph-search, graph-status, graph-entities
Tests: test-memory-graph.bats — 12 tests (build, entity extraction, relations, search, status, CLI integration)
Docs: ROADMAP.md synced to v3.38.0 (Eras 125-133 documented, community research sources added)

3.38.0 — 2026-03-22

Community insights: 3 new specs + PreCompact hook + reranker + failure logging.

Added

Spec: SPEC-026 PreCompact hook (transcript backup before /compact)
Spec: SPEC-027 Graph memory layer (entity-relation extraction, LightRAG-inspired)
Spec: SPEC-028 Search reranker (cross-encoder post-retrieval ranking)
Hooks: pre-compact-backup.sh — SPEC-026 implemented. Extracts decisions/corrections from session, saves via memory-store before compact. Never blocks.
Hooks: post-tool-failure-log.sh — logs tool execution failures to ~/.pm-workspace/tool-failures/ (async, JSONL). Inspired by disler/claude-code-hooks-mastery.
Tests: test-new-hooks.bats — 8 tests for both hooks + settings.json validation

Changed

Scripts: memory-vector.py — SPEC-028 reranker integration. Fetches 3x candidates, reranks with cross-encoder/ms-marco-MiniLM-L-6-v2 when available. Graceful degradation (vector-only if no reranker). Status shows reranker availability.
Config: settings.json — added PreCompact and PostToolUseFailure hook events

Research analyzed

7 community repos: LightRAG (graph+vector), Obsidian Skills (agent spec), Context Engineering (PRP), AY Skills (meta-skill), Hooks Mastery (13 lifecycle events), Best Practices (84 tips), n8n-MCP (automation bridge).

3.37.0 — 2026-03-22

SPEC-023 Phase 1 — Training data generator for Savia context brain.

Added

Scripts: generate-training-data.py — extracts instruction/response pairs from commands (1488), rules (24), skills (30), and memory store. Deduplicates by instruction. Respects PROJECT_ROOT env var. Output: JSONL for QLoRA fine-tuning.
Tests: test-training-data.bats — 8 tests (valid Python, JSONL output, JSON integrity, field validation, extraction from 3 sources, dedup)

3.36.0 — 2026-03-22

SPEC-024 — Doc audit: CONTRIBUTING.md and SECURITY.md rewritten in Savia's first person voice.

Changed

Docs: CONTRIBUTING.md — rewritten as Savia speaking directly to contributors. Updated testing commands to current suite (run-all.sh, validate-ci-local.sh). Mentions 9 languages.
Docs: SECURITY.md — rewritten as Savia explaining how she protects data. Added zero telemetry section, mentions block-credential-leak.sh hook, updated versión table to 3.x.

3.35.0 — 2026-03-22

SPEC-022 Phase 1 — Budget Guard + PM Keybindings.

Added

Scripts: budget-guard.sh — context budget monitor with 4 levels (healthy/warning/high/critical). Supports --block mode for heavy commands. budget_banner() for command output. Heuristic estimation from trace log when no env var.
Docs: pm-keybindings.json — PM-optimized keybindings template (Ctrl+Shift+S sprint, B board, M my-sprint, D daily, P compact, H help). Copy to ~/.claude/keybindings.json.
Tests: test-power-cli.bats — 13 tests (4 keybindings + 9 budget guard)

3.34.0 — 2026-03-22

4 new specs + roadmap sync to v3.33.0.

Added

Spec: SPEC-022 Power Features CLI (budget guard, semantic compact, PM keybindings, PR context loader)
Spec: SPEC-023 Savia LLM Trainer (dataset gen, QLoRA fine-tune, eval, integration)
Spec: SPEC-024 Doc Audit — Savia en primera persona (rewrite public docs with Savia voice)
Spec: SPEC-025 Chinese (ZH) compatibility study (CJK, encoding, cultural)
Docs: ROADMAP.md fully synced — Eras 125-130 documented as Done, In Progress updated, all items have SPEC references, Rejected includes SQLite rationale

3.33.0 — 2026-03-22

Multilingual docs — Savia speaks 9 languages. Chinese study in roadmap.

Added

Docs: README translations in 7 new languages: Galego (gl), Euskara (eu), Catala (ca), Francais (fr), Deutsch (de), Portugues (pt), Italiano (it). All written in Savia's first person voice explaining architecture, privacy, and how to get the most from pm-workspace.
Roadmap: Chinese (ZH) compatibility study added as Proposed item (CJK tokenization, encoding, native review needed)

Changed

Docs: README.md + README.en.md — language selector bar updated with all 9 languages

3.32.1 — 2026-03-22

Split memory-store.sh into 3 modules (75+127+100 lines, all under 150).

Changed

Scripts: memory-store.sh split into dispatcher (75 lines) + memory-save.sh (save/upsert/entity/session, 127 lines) + memory-search.sh (search/context/stats, 100 lines). All functions preserved, zero behavior change.

3.32.0 — 2026-03-22

SPEC-019/020/021 implemented — contradiction tracking, TTL, hardware checks, zero telemetry.

Added

Scripts: memory-store.sh — SPEC-019 supersedes field tracks what changed on upsert. SPEC-020 --expires DAYS sets TTL, search auto-filters expired entries (--include-expired to show all)
Scripts: readiness-check.sh — SPEC-021 hardware checks (RAM, disk, CPU, GPU), connectivity test (Nomad pattern), section [4b/7]
Docs: README.md + README.en.md — "Privacy & Telemetry" section: zero telemetry declaration
Tests: 4 new tests for SPEC-019/020 (supersedes, no-supersedes, expires, no-expires). Total: 20 memory-store tests

3.31.0 — 2026-03-22

Specs + roadmap update — Supermemory/Nomad-inspired improvements.

Added

Spec: SPEC-019 Memory contradiction resolution (supersedes field on upsert)
Spec: SPEC-020 Memory TTL/expiration (temporal fact auto-hiding)
Spec: SPEC-021 Readiness hardware checks + zero telemetry declaration
Docs: ROADMAP.md updated — Eras 125-128 done, new In Progress section, sources from Engram/Supermemory/Nomad

3.30.0 — 2026-03-22

Readiness check — deterministic capability checklist on install/update.

Added

Scripts: readiness-check.sh — 50-point deterministic checklist across 7 categories (runtime, structure, scripts, vector memory, hooks, tests, git). Writes stamp to ~/.pm-workspace/.readiness-stamp on pass. Runs automatically post-update.
Hooks: session-init.sh — detects stale/missing readiness stamp, suggests re-run after git pull

Changed

Scripts: update.sh — runs readiness-check.sh automatically after successful update (auto-adaptation for all Savia instances)

3.29.0 — 2026-03-22

SPEC-018 Vector memory index — semantic search over plain-text JSONL.

Added

Scripts: memory-vector.py — vector index engine: rebuild, search, status, benchmark. Uses sentence-transformers (all-MiniLM-L6-v2, 22MB, Apache 2.0) + hnswlib. Zero vendor lock-in, offline-compatible.
Scripts: memory-store.sh — vector search integration with --mode auto|grep|vector, auto-rebuild on JSONL changes, rebuild-index/index-status/benchmark subcommands
Tests: test-memory-vector.bats — 8 integration tests (fallback, auto-rebuild, status)
Tests: test-vector-quality.py — benchmark: Recall@5 grep=40% vs vector=90% (+50pp)
Spec: SPEC-018-vector-memory-index.md — architecture, justification, auto-adaptation
Config: requirements-vector.txt — optional deps (sentence-transformers, hnswlib)

Changed

Scripts: memory-store.sh — search attempts vector first, falls back to grep. Auto-rebuild triggers in background after each save.

3.28.0 — 2026-03-22

Engram-inspired memory patterns — structured observations, topic key families, session summaries.

Added

Scripts: memory-store.sh — What/Why/Where/Learned structured fields (--what, --why, --where, --learned), auto-generated topic key families (decisión/*, bug/*, architecture/*, etc.), suggest-topic command, session-summary command with Goal/Discoveries/Accomplished/Files format
Tests: test-memory-store.bats — 16 tests covering structured save, topic keys, upsert, dedup, search, session summary

Changed

Scripts: memory-store.sh — search rewritten with temp-file scoring (bash subshell compatible), topic_key shown in context/search output, stats include topic family and revision metrics

3.27.1 — 2026-03-22

sovereignty-ops.sh fixes for real-world download testing.

Fixed

Scripts: sovereignty-ops.sh — correct Python standalone URL (was 404), use CPU-only PyTorch wheels, detect HF cache properly

3.27.0 — 2026-03-22

SPEC-014 Phase 2 — Competence tracking + scoring pipeline.

Added

Hooks: competence-tracker.sh — async PostToolUse hook that logs domain per command to competence-log.jsonl. Maps 11 domains: sprint-mgmt, sdd, architecture, security, devops, testing, reporting, product, context, team, hardware.
Scripts: competence-score.sh — reads log, calculates 3-signal scores (entry count, recency, outcome), generates competence.md in user profile. Classifies as expert/competent/novice/unknown.

3.26.0 — 2026-03-22

SPEC-012 Complete — L1 progressive loading for all 82 skills.

Added

Skills: L1 summary field added to all 62 remaining skills (82/82 total)
SPEC-012 now COMPLETE — all skills have 3-4 line summaries for progressive loading

3.25.1 — 2026-03-22

PR signing protocol — zero re-sign commits.

Added

Scripts: push-pr.sh — automates CI + CHANGELOG + sign + push + PR + auto-merge. Auto-generates PR body with Summary section (PR Guardian Gate 1). Detects repo from remote URL. Polls CI instead of fixed sleep. --skip-changelog for docs-only PRs.
Rules: pr-signing-protocol.md — strict sign-last order to prevent re-sign loops

3.25.0 — 2026-03-22

SaviaClaw voice v2.4, Context Intelligence Tier 1-2, SPEC-017 Sovereignty, docs alignment.

Added

SaviaClaw: savia-voice daemon v2.4 — full-duplex, conversation model, Kokoro TTS, pre-cache 64 phrases
SaviaClaw: 31 new unit tests for savia-voice (77 total zeroclaw tests)
Core: SPEC-015 Context Gate in skill-auto-activation (6 bypass conditions)
Core: SPEC-012 Phase 1 — L1 summaries for 20 skills (progressive loading)
Core: SPEC-013/016 Session memory extraction + intelligent compact protocol
Core: SPEC-014 Competence-aware output in adaptive-output
Core: SPEC-017 Dependency Sovereignty spec + sovereignty-pack.sh (offline USB installer)
Rules: session-memory-protocol.md — pre-compact + end-of-session extraction

Fixed

SaviaClaw: daemon.py NameError in respond() (crash on first voice turn)
Scripts: contribute.sh ERE lookahead privacy leak (regex never matched)
Scripts: memory-store.sh grep injection via topic_key/hash
Scripts: validate-bash-global.sh POSIX ERE compat (macOS grep)
Scripts: scope-guard.sh restrict file extraction to bullet lines

Changed

Docs: README.md + README.en.md + CLAUDE.md counters aligned to real state
Docs: agents-catalog.md 44→46, pm-workflow.md commands 99→496

3.24.0 — 2026-03-21

SaviaClaw v1.0 prep — daemon stability, voice pipeline, roadmap.

Added

Host: voice.py — TTS (espeak-ng/spd-say) + STT (whisper) pipeline, offline-first. --say, --listen, --test CLI
Host: daemon_util.py — shared utilities extracted from daemon (find_port, truncate_lcd, write_status, show_status)
Host: saviaclaw_daemon.py — signal handling (SIGTERM/SIGINT), status file (status.json), stuck detection (120s), --status flag
Roadmap: zeroclaw/ROADMAP.md — 6 phases (foundations → stability → voice → sensors → actuators → autonomy)
Tests: test_daemon.py (9 tests), test_voice.py (7 tests) — total 39 tests without hardware
Docs: README.md + README.en.md — SaviaClaw section, directory tree, documentation table

Changed

Host: daemon refactored into 2 modules (daemon 148 lines + util 84 lines, both under 150)
Host: daemon log uses RotatingFileHandler (1MB, 3 backups)

3.23.0 — 2026-03-21

SaviaClaw v0.9 — self-test, daemon, autonomous operation.

Added

Firmware: selftest.py — hardware diagnostic at boot: CPU, RAM, LED, LCD I2C, WiFi, flash. Results on LCD. Warns on failures
Firmware: main.py v0.9 — selftest at boot before main loop
Script: saviaclaw_daemon.py — background process: auto-detect ESP32, reconnect on disconnect, process ask commands via claude -p, log to ~/.savia/zeroclaw/daemon.log
Script: saviaclaw.service — systemd unit for auto-start on boot
Script: install-daemon.sh — one-command daemon installation

3.22.0 — 2026-03-21

SaviaClaw autonomy roadmap + heartbeat + BT audio research.

Added

Spec: SPEC-010-saviaclaw-autonomy-roadmap.md — 6-level autonomy plan: stability → proactivity → voice → BT audio → context guardian → multi-claw. Research: HFP AG for bidirectional BT audio with headset
Firmware: heartbeat.py — periodic LCD status rotation (identity, uptime, WiFi, RAM, custom messages). 8-second cycle
Firmware: main.py v0.8 — integrates heartbeat, LCD shows live status

Research

ESP32 Bluetooth: A2DP Source (send to speaker, SBC codec), HFP AG (bidirectional with headset, CVSD/mSBC)
Decisión: HFP AG for full-duplex voice via any BT headset (~10 EUR)
MicroPython limitation: BT audio requires ESP-IDF (C). Hybrid approach planned

3.21.0 — 2026-03-21

Savia Brain Bridge + CI signature fix.

Added

Script: savia_brain.py — ESP32 asks → claude -p → LCD response
Firmware: ask command for querying Savia brain

Fixed

CI: confidentiality-sign.sh — exclude self + workflow YAML from diff hash, fixing circular dependency. CI workflow checks out PR head SHA
CI: confidentiality-gate.yml — checkout head.sha instead of merge commit

3.20.1 — 2026-03-21

Fix LCD overwrite bug — verified on hardware.

Fixed

Firmware: main.py — main loop was writing command name on LCD row 0 after every command, overwriting the text that _cmd_lcd had just set. Fix: skip LCD status write when cmd == "lcd"

3.20.0 — 2026-03-21

ZeroClaw v0.7 — first stable firmware tested on real ESP32 hardware.

Added

Firmware: lcd_i2c.py — LCD 16x2 I2C driver (PCF8574 @ 0x3F, SCL=23, SDA=22): clear, write, message, backlight, cursor control
Firmware: lcd command added to command handler — write to LCD via serial: lcd Hello | World

Fixed

Firmware: main.py v0.7 — replaced broken sys.stdin.buffer.any() with select.poll() + sys.stdin.read(1) for reliable non-blocking serial I/O on MicroPython v1.19.1
Firmware: all imports wrapped in try/except to prevent boot crash from missing hardware

Verified on hardware

ESP32 module (spiram), MicroPython v1.19.1
LCD 16x2 I2C @ 0x3F (SCL=23, SDA=22)
NeoPixel RGB LED @ GPIO2
6/6 serial commands pass: ping, info, led, sensors, lcd, help
Savia wrote her first message: "Soy Savia | Vivo en ZeroClaw"

3.19.1 — 2026-03-21

Fix confidentiality signature system — CI compatibility.

Fixed

Script: confidentiality-sign.sh — rewritten get_diff_hash() with 4-strategy fallback: merge-base diff (feature branch) → GITHUB_BASE_REF (CI merge commit) → staged changes → last commit diff. Fixes empty hash when HEAD=origin/main
Script: HMAC now computed over diff_hash only, not commit hash (which changes on squash merge). CI verifies diff match; HMAC verified only when key available (local)
Lesson: added to tasks/lessons.md — always sign before push

3.19.0 — 2026-03-21

Savia in Teams — same brain, two channels (ZeroClaw + Teams).

Added

Spec: SPEC-009-savia-teams-participant.md — architecture for Savia joining Teams meetings via Graph API: transcript reading, chat participation, speaker identity from Azure AD, 4-phase implementation plan
Script: teams_client.py — Graph API client: OAuth2 client credentials auth, meeting discovery, transcript retrieval, chat message posting. All credentials from files (N2, gitignored)
Script: meeting_orchestrator.py — unified controller for ZeroClaw + Teams: same MeetingParticipant + ContextGuardian + SpeakerRoles brain, channel-agnostic process_utterance() and handle_query() with role filtering
Tests: test_teams_integration.py — 11 tests: config detection, auth fallback, orchestrator start/stop, risk detection, query with role filtering, cross-channel brain persistence

3.18.0 — 2026-03-21

Savia as active meeting participant — etiquette protocol, context guardian, speaker role permissions.

Added

Rule: meeting-participant-etiquette.md — 4 simultaneous roles (transcriber, context guardian, query responder, proactive participant). 5-condition window for proactive speech. 3 configurable modes (silent, query, active). Post-meeting output: transcript, digest, action items, contradictions, risks, unanswered questions
Script: meeting_participant.py — opportunity window detector (3s silence + no pending turn + critical info + not already said + PM allows). Max interventions limit, cooldown timer, mode switching, internal note buffer
Script: context_guardian.py — cross-references live speech against decisión log, business rules, sprint state. Detects: action items (commitment language), contradictions with prior decisions, risk mentions, unanswered questions
Script: speaker_roles.py — deterministic role-based access control in CODE. 5 levels: external → observer → developer → tech_lead → pm. Topic filter gate: filter_response() strips unauthorized data BEFORE voice output. NEVER_VOICE set blocks biometric, salary, credentials, PII from voice output for ALL roles including PM
Tests: test_meeting_participant.py (12 tests) + test_speaker_roles.py (10 tests)

Security design

Speaker permissions enforced by Python filter_response() function, not LLM instruction
NEVER_VOICE topics (evaluations, salary, credentials, PII, voiceprints) blocked for ALL roles in voice output — PM accesses these via console only
Unknown speakers default to "observer" (minimal access)
Context integrity: Savia ANNOTATES contradictions but does NOT override or modify project data based on meeting requests

3.17.0 — 2026-03-21

ZeroClaw meeting digest — speaker diarization + voice fingerprinting.

Added

Spec: SPEC-008-zeroclaw-meeting-digest.md — live meeting pipeline: audio capture → VAD → pyannote diarization → SpeechBrain voice ID → whisper STT → JSONL transcript with speaker labels → meeting-digest agent
Script: voiceprint.py — voice enrollment: extract ECAPA-TDNN embedding from 10-15s speech, store as numpy array in ~/.savia/zeroclaw/voiceprints/ (N4b biometric)
Script: voiceprint_ops.py — identify speaker from embedding (cosine similarity), list/delete voiceprints (RGPD Art. 17 right to erasure)
Script: meeting_pipeline.py — orchestrator: process audio buffer through diarization → speaker ID → STT, output JSONL transcript, graceful fallback for each missing dep
Rule: zeroclaw-meeting-protocol.md — consent guardrails (audible warning mandatory), voice enrollment flow, confidence thresholds (75%/50%), RGPD compliance, degradation matrix
Command: /zeroclaw meeting — start, stop, voice enroll/list/delete, status subcommands
Tests: test_voiceprint.py — 9 tests: cosine similarity math, index operations, N4b storage location, threshold values, file sizes

3.16.0 — 2026-03-21

ZeroClaw network auto-config — Savia detects its WiFi and provisions ESP32 to join.

Added

Script: network_setup.py — cross-platform (Linux/macOS/Windows) detection of host WiFi SSID and IP address via nmcli/iwgetid/airport/netsh/ipconfig
Script: esp32_wifi.py — ESP32 WiFi operations via mpremote: scan networks, verify connection, deploy config, reset device
Script: network_cli.py — interactive wizard: detects host network → asks password → deploys config → resets ESP32 → verifies same-subnet connectivity
Script: connectivity_test.py — end-to-end test: USB serial + WiFi ping + HTTP endpoint verification
Firmware: wifi_server.py — minimal HTTP server on ESP32 (GET /ping, POST /cmd, GET /status) for wireless command execution
Firmware: main.py upgraded to v0.2.0 — dual-mode: serial USB AND WiFi HTTP simultaneously, auto-starts HTTP if WiFi connected
Command: /zeroclaw network — setup, check, scan subcommands
Tests: test_network.py — 8 tests (SSID detection, IP detection, config structure, no secrets, dual-mode firmware)

3.15.0 — 2026-03-21

ZeroClaw sensory protocol + deterministic guardrails — no agent can bypass.

Added

Rule: zeroclaw-sensory-protocol.md — ingestion pipeline (classify → transcribe → filter → digest → persist → discard raw), confidentiality alignment with N1-N4b, RGPD compliance for biometric data, session storage structure, retention policies
Script: guardrails.py — 7 deterministic security gates in Python code: size limits (5MB audio, 2MB image), rate limiting (5 audio/min), command allowlist, master validator validate_incoming() that ALL data must pass
Script: guardrails_pii.py — PII detection (DNI, IBAN, phone, email, card), raw data auto-expiry (1h), storage quota (100MB), immutable audit log (append-only JSONL)
Tests: test_guardrails.py — 14 tests proving gates block oversized data, flooding, unknown commands, PII, full storage. Tests pass without hardware

Security design

Gates are Python functions, not LLM instructions — deterministic, untrickable
Command allowlist: only 12 known commands pass (ping, led, info, sensors, gpio, help, capture_image, capture_audio, speak, set_led, play_tone, status)
Immutable audit log: every incoming datum logged before processing
Raw data (audio/images) auto-deleted after 1 hour

3.14.0 — 2026-03-21

ZeroClaw Firmware v0.1 — ready to flash when ESP32 is connected.

Added

Firmware: zeroclaw/firmware/ — MicroPython firmware for ESP32: boot.py (WiFi + CPU config), main.py (JSON command loop with watchdog), lib/commands.py (ping, led, info, sensors, gpio), lib/status.py (LED patterns for feedback)
Host: zeroclaw/host/bridge.py — serial bridge PC↔ESP32: auto-detect port, JSON protocol, timeout handling
Host: zeroclaw/host/cli.py — self-test (5 checks), interactive mode, CLI entry point
Setup: zeroclaw/setup.sh — one-command setup: installs esptool+mpremote, detects ESP32, flashes MicroPython, deploys firmware, verifies with LED blink
Command: /zeroclaw — setup, test, ping, led, flash, interactive subcommands
Tests: zeroclaw/tests/test_bridge.py — 9 tests that run without hardware (imports, protocol, firmware structure, security, sizes)

3.13.0 — 2026-03-21

ZeroClaw voice pipeline + voice/console decisión protocol.

Added

Spec: SPEC-007-zeroclaw-voice-pipeline.md — full bidirectional voice architecture: 3 processing levels (ESP32 wake word → Host STT/TTS → optional cloud), Wyoming-adapted protocol, latency target ~6s, 5-phase implementation plan
Rule: voice-console-protocol.md — decisión algorithm for what goes to voice (short instructions, safety warnings) vs console (code, tables, diagrams). 4 session modes: assembly, coding, monitoring, chat. LED indicator states for ZeroClaw
Script: voice_bridge.py — host-side voice server: faster-whisper STT + pyttsx3/Piper TTS, dependency detection, setup guide. Graceful fallback when deps missing

Research (incorporated in specs)

ESP-SR WakeNet/MultiNet for on-device wake word on ESP32-S3
whisper.cpp for edge STT (~273MB RAM for tiny model)
Piper TTS for fast local Spanish voice synthesis
Wyoming protocol (Rhasspy/Home Assistant) for audio streaming
HuggingFace speech-to-speech pipeline architecture (VAD→STT→LLM→TTS)

3.12.0 — 2026-03-21

Physical assembly guidance + ZeroClaw spec — Savia guides hardware and gains physical senses.

Added

Spec: SPEC-005-physical-assembly-guide.md — 3 guidance modes (ASCII diagrams, step-by-step manuals, offline TTS voice), component knowledge base (9 components with safety warnings), schemdraw SVG generation
Spec: SPEC-006-zeroclaw.md — ESP32-S3 as Savia's physical interface: microphone (INMP441), speaker (MAX98357A), camera (OV2640). Protocol design, security model, 5-phase implementation plan
Script: pinout.py — ASCII pinout generator for ESP32/Arduino/RPi Pico with connection annotations and wire colors
Script: assembly_guide.py — Step-by-step assembly guide generator with BOM, wiring steps, verification checklists, and per-component safety warnings
Script: voice_guide.py — Offline TTS voice narrator (pyttsx3) with interactive controls (next/repeat/back/status), ES/EN support

3.11.0 — 2026-03-21

Robotics Vertical — architecture, security, and MicroPython for the physical AI era.

Added

Spec: SPEC-004-robotics-vertical.md — full 5-layer robotics stack (AI → ROS2 → Edge → MCU → Hardware), STRIDE threat model for robotics, language packs, agent proposals, ESP32 lab integration plan
Rule: robotics-safety.md — 10 immutable safety principles + 5 REJECT rules (watchdog, actuator limits, auth, OTA signing, sensor redundancy)
Language Pack: micropython-conventions.md — auto-loads on boot.py/main.py, patterns for sensor reading, actuator control, async with watchdog
Docs ES: docs/robotics-roadmap.md — 5-phase roadmap from ESP32 to LeRobot
Docs EN: docs/robotics-roadmap.en.md — English versión

3.10.1 — 2026-03-21

Web Research: tests, documentation (ES/EN), and skill registration.

Added

Tests: tests/test-web-research.bats — 22 BATS tests covering cache, sanitizer, reranker, formatter, gap detector, suggestions, SearxNG, and CLI
Docs ES: docs/web-research.md — full documentation in Spanish
Docs EN: docs/web-research.en.md — full documentation in English
Skill: web-research registered with SKILL.md + DOMAIN.md (Clara Philosophy)

3.10.0 — 2026-03-21

FAIR-Perplexica improvements: autonomous SearxNG, gap detection, global context, follow-up suggestions.

Added

Script: searxng.py — SearxNG Docker auto-start: detects Docker, starts container savia-searxng on demand, health check, graceful fallback to WebSearch
Script: docker-compose.searxng.yml — SearxNG container definition (port 8888, localhost only, no tracking)
Script: search.py — 3-layer search orchestrator: cache → SearxNG (auto-start) → Claude WebSearch
Script: gap_detector.py — detects context gaps in user queries (versions, docs, CVEs, comparisons) vs internal PM questions
Script: suggestions.py — post-command follow-up suggestions for 10 command families (inspired by Perplexica suggestion generator)
Rule: global-context.md — compact company DNA (~100 tokens) injected into all agent prompts, saves ~360 tokens vs full profile
Script: generate-global-context.sh — generates global context from company profile + config

Changed

Command: /web-research — now auto-starts SearxNG Docker, 3-layer search with engine presets by category

3.9.0 — 2026-03-21

Savia Web Research — web search to resolve context gaps. Inspired by FAIR-Perplexica (UB-Mannheim).

Added

Skill: web-research — search engine with local cache, query sanitization, heuristic reranking, and inline citations [web:N]
Command: /web-research <query> — search the web for documentation, versions, CVEs, best practices. Subcommands: --cache-stats, --cache-clear, --cache-only
Rule: web-research-config.md — configuration, privacy protocol, context-budget integration, degradation levels
Spec: SPEC-003-web-research-system.md — full architecture proposal (3 layers: cache → Claude tools → SearxNG)
Script: scripts/web-research/ — Python package: cache (LRU, TTL by category), sanitizer (PII/project removal), reranker (heuristic scoring), formatter (citation generation)

3.8.1 — 2026-03-21

Native markdownlint — replaces npm markdownlint-cli dependency.

Added

Script: scripts/markdownlint/ — Python3 native markdownlint (17 rules, zero npm dependency, --fix mode)
Script: scripts/markdownlint.sh — CLI wrapper

Changed

CI: ci.yml lint-markdown job uses native linter instead of npm markdownlint-cli
CI: validate-ci-local.sh — added markdown lint check for CHANGELOG.md

Fixed

CHANGELOG.md: Fixed 259 markdownlint errors (MD012, MD022, MD032, MD053) from v3.7.0 base rebuild

3.8.0 — 2026-03-21

Feasibility Probe and Model Upgrade Audit — inspired by Cat Wu's "Product management on the AI exponential" (Anthropic, March 2026).

Added

Agent: feasibility-probe — time-boxed prototype attempt on a spec, produces viability report with score 0-100, blocking sections, and decomposition suggestions
Agent: model-upgrade-auditor — audits agents/skills/rules for prompt debt (emphatic repetitions, defensive parsing, coded retries) that newer models may not need
Skill: feasibility-probe (SKILL.md + DOMAIN.md) — decisión checklist, scoring formula, SDD integration as optional gate between spec-approve and dev-session
Skill: model-upgrade-audit (SKILL.md + DOMAIN.md) — 6 workaround patterns, 3-tier risk classification (APPLY/REVIEW/SKIP), longitudinal tracking
Command: /feasibility-probe <spec_path> — validate spec feasibility with budget-constrained prototype
Command: /model-upgrade-audit [--scope] — detect prompt debt and propose simplifications

3.7.1 — 2026-03-20

Fix update system and auto-release pipeline.

Fixed

update.sh: compares against origin/main instead of GitHub releases — no longer requires gh CLI, reads versión from CHANGELOG.md
update.sh: uses git pull origin main instead of merging a tag that may not exist

Added

auto-tag.yml: GitHub Actions pipeline that creates git tag automatically when CHANGELOG.md is updated on main, triggering release.yml
sync-tags-from-changelog.sh: one-time script to backfill missing tags from CHANGELOG.md history (138 versions synced)

3.7.0 — 2026-03-20

Context optimization, React quality, and decisión-guided skills — inspired by rtk-ai/rtk and no-use-effect.

Added

Hook: bash-output-compress.sh — async PostToolUse hook that compresses verbose Bash output (blanks, repeats, ANSI, truncation). Specialized filters for git, dotnet, npm, az devops. Inspired by rtk-ai/rtk (60-90% token reduction)
Rule: react-use-effect-anti-patterns.md — 6 rules + 8-question decisión checklist for React useEffect. Auto-loads on .tsx/.jsx. Inspired by no-use-effect skill
Tracker: context-tracker.sh compression-report — new subcommand for Bash compression metrics
Pattern: Decisión Checklists added to 6 core skills (sequential yes/no routing before execution)

Changed

Skill: spec-driven-development — added 5-question decisión checklist + abort conditions for human vs agent routing
Skill: pbi-decomposition — added 5-question decisión checklist + abort conditions for decomposition gates
Skill: risk-scoring — added 5-question decisión checklist with score modifiers and abort conditions
Skill: consensus-validation — added 5-question decisión checklist for mandatory vs optional consensus
Skill: product-discovery — added 5-question decisión checklist for skip/start/delay discovery
Skill: verification-lattice — added 5-question decisión checklist for layer selection by risk
Rule: react-conventions.md — added reference to new useEffect anti-patterns file
Config: settings.json — registered bash-output-compress hook as async PostToolUse for Bash
Docs: README.md and README.en.md — hooks count updated from 16 to 17

Specs

SPEC-001: Bash Output Compression Hook (rtk-inspired)
SPEC-002: React useEffect Anti-Patterns (no-use-effect inspired)
SPEC-003: Decisión Checklists for Top 6 Skills

3.6.1 — 2026-03-20

PII purge from tracked files + full-repo scan mode.

Fixed

PII: Removed private project name from 7 tracked files (commands, rules, skills, specs, tests, docs) committed before the confidentiality system existed
Scanner: confidentiality-scan.sh --full-repo mode scans ALL tracked file contents, not just PR diffs — closes the gap that allowed pre-existing PII to persist undetected
Blocklist: generate-blocklist.sh auto-detects public projects from .gitignore whitelist to avoid false positives, handles empty arrays in CI
CI: confidentiality-gate.yml adds weekly scheduled full-repo audit (Monday 06:00 UTC) + manual dispatch

3.6.0 — 2026-03-20

Pre-PR confidentiality audit system with cryptographic signature.

Added

Agent: confidentiality-auditor rewritten — dynamic context-aware audit that reads workspace context (project names, team members, org URLs) to discover sensitive data semantically, not with static patterns
Script: confidentiality-sign.sh — HMAC-SHA256 signature generation/verification after clean audit. Signature must be committed with the PR; CI verifies diff hash matches
Script: generate-blocklist.sh — dynamic blocklist generator from 6 workspace sources (projects, profiles, teams, local config, email domains, static list)
CI: confidentiality-gate.yml — two parallel jobs: signature verification + deterministic scan (defense in depth)
Command: /confidentiality-check updated to orchestrate full flow (agent audit + signature + scan)
Script: confidentiality-scan.sh — 8-check deterministic scanner with dynamic blocklist support

Fixed

Hook: validate-bash-global.sh — detects target repo from cd in command instead of always checking CLAUDE_PROJECT_DIR branch. Fixes false "commit on main" blocks when working in sub-repos with their own .git

3.5.3 — 2026-03-19

Fix: multi-repo branch detection in global bash hook.

Fixed

Hook: validate-bash-global.sh — detects target repo from cd in command instead of always checking CLAUDE_PROJECT_DIR branch. Fixes false "commit on main" blocks when working in sub-repos with their own .git

3.5.2 — 2026-03-19

Criticality scoring engine — backing scripts for Era 120 criticality commands.

Added

Script: scripts/criticality.sh — dispatcher for assess/dashboard/rebalance
Script: scripts/criticality-scoring.sh — pure scoring (WSJF, confidence decay, urgency boost, 5-dimension model, P0-P3 classification)
Script: scripts/criticality-engine.sh — operations (assess single item, dashboard cross-project, rebalance analysis)

3.5.1 — 2026-03-19

Backing scripts for vault, confidentiality scanner, and travel sync.

Added

Script: scripts/vault.sh + scripts/vault-ops.sh — full implementation of 5 vault operations (init, sync, status, restore, export) with NTFS junction support
Script: scripts/confidentiality-check.sh — project-level scanner for N4-SHARED compliance (PII, secrets, cross-level leaks, scoring 0-100)

Fixed

Script: scripts/savia-travel.sh — implemented sync command (was stub), includes vault in travel package

3.5.0 — 2026-03-19

Personal Vault (N3) + Confidentiality Auditor + 5-level confidentiality documentation.

Added

Personal Vault: 5 commands (vault-init, vault-sync, vault-status, vault-restore, vault-export) + skill + config rule for N3 user data in separate git repo
Confidentiality Auditor: Opus agent for multi-repo confidentiality compliance auditing
Command: /confidentiality-check — verify level compliance per project
Rule: personal-vault-config.md — vault configuration constants
Docs: docs/confidentiality-levels.md — full 5-level (N1-N4b) documentation

Changed

Rule: context-placement-confirmation.md — compressed to 150 lines, added N3 vault integration

3.4.1 — 2026-03-19

PII sanitization from security audit + Confidentiality Gate CI pipeline.

Fixed

PII: removed 7 real names, 3 real companies, HR data from tracked files (GDPR)
Untracked: active-user.md and settings.local.json removed from git index
IPs: replaced hardcoded 192.168.1.x with <YOUR_PC_IP> in savia-mobile-android

Added

Confidentiality Gate: CI pipeline with 7-check scanner (blocklist, credentials, emails, proper nouns, forbidden files, merge markers, private IPs)
Files: scripts/confidentiality-scan.sh, confidentiality-blocklist.txt, confidentiality-allowlist.txt

3.4.0 — 2026-03-19

Era 120 — Task Criticality System + Multi-Tenant Calendar Sync.

Added

Task Criticality: multi-level prioritization (WSJF, Cost of Delay, RICE, Eisenhower) with 5 scoring dimensions, auto-escalation, and confidence decay
Commands: /criticality-dashboard, /criticality-assess, /criticality-rebalance (3 new)
Multi-Tenant Calendar Sync: /sync-calendars — bidirectional free/busy sync between 2 Microsoft 365 tenants with AES-256 encrypted per-user credentials
Specs: spec-task-criticality.md, spec-criticality-frameworks.md (9 frameworks researched), spec-multi-tenant-sync.md, spec-multi-tenant-security.md
Docs: Smart Calendar (7 cmds) and Task Criticality (3 cmds) sections in ES+EN

Changed

smart-calendar SKILL.md: added criticality integration section + sync-calendars reference

3.3.0 — 2026-03-19

Era 118 — Five improvements from open-source research (GitNexus, NemoClaw, GSD, Context Hub, Everything Claude Code).

Added

Skill: codebase-map + Command: /codebase-map: Symbol indexing of pm-workspace itself. Scans commands→agents→rules→skills dependency graph. Detects orphaned rules, hub rules, routing chains. Inspired by GitNexus code intelligence engine.
Rule: agent-policies.md + Command: /policy-check: Policy-driven agent isolation with per-project YAML policies (allowed/denied paths, approval requirements, timeouts, network restrictions). Audit trail for violations. Inspired by NVIDIA NemoClaw sandbox orchestration.
Rule: dev-session-locks.md + Command: /dev-session-resume: Crash recovery for dev-sessions via lock files with PID detection. State machine (pending→implementing→validating→verified→completed). Auto-resume from last checkpoint. Inspired by GSD 2 disk state machine.
Skill: doc-quality-feedback + Command: /docs-quality-audit: Agent feedback loop for documentation quality. Agents rate docs after use (clear/confusing/incomplete/outdated). Monthly aggregation flags low-quality docs for rewrite. Inspired by Context Hub agent annotations.
Command: /skill-propose + Rule: skill-lifecycle.md: Auto-generate skill scaffolds from repeated workflows (3+ observations). Consensus validation, adoption tracking, archival of unused skills. Inspired by Everything Claude Code continuous learning.

Changed

README.md + README.en.md: Updated with new commands and skills.
12-comandos-agentes.md + 12-commands-agents.md: Added new command categories.

3.2.0 — 2026-03-19

Era 117 — Document Digest Suite: 4 new agents for PDF, Word, Excel, PowerPoint digestión with context-aware 4-phase pipeline.

Added

Agent: pdf-digest (Opus 4.6): 4-phase pipeline for PDF documents using PyMuPDF for text extraction + Claude Vision for embedded images. Phases: raw extraction → project context loading → analysis/synthesis with cross-referencing → context document update. Supports protocols, manuals, proposals, reports, specs.
Agent: word-digest (Opus 4.6): 4-phase pipeline for DOCX using python-docx. Extracts text with styles, tables, embedded images, metadata. Same 4-phase context-aware architecture.
Agent: excel-digest (Opus 4.6): 4-phase pipeline for XLSX using openpyxl. Extracts structure, formulas (translated to natural language), validations, conditional formatting, macro names. Detects business rules embedded in formulas and anti-patterns.
Agent: pptx-digest (Opus 4.6): 4-phase pipeline for PPTX using python-pptx. Prioritizes presenter notes over slide text. Extracts chart data, images via Vision.
Skill: prompt-optimizer + Command: /skill-optimize: AutoResearch Loop for self-optimizing skill and agent prompts. Inspired by Karpathy/Eric Risco pattern. Executes skill with test fixture, scores output against weighted checklist (G-Eval 0-10), modifies prompt, re-executes, compares scores. Keeps changes that improve, reverts those that don't. Stop criterion: score >= 8.0 for 3 consecutive iterations. Output saved as .optimized.md — original never modified.
Prompt optimizer auto-trigger (prompt-optimizer/auto-trigger.md): Protocol for Savia to automatically suggest /skill-optimize when agents accumulate 3+ corrections in last 10 executions. Detects explicit signals (PM corrections, re-executions) and implicit signals (low coherence scores, outputs exceeding limits). Auto-generates test fixtures from real usage patterns.

Changed

agents-catalog.md: Updated from 39 to 43 agents. Added Document Digest Suite flow.
CLAUDE.md: Agent count updated from 39 to 43.
meeting-digest.md: Phase 4 context update now references README.md as project index (generic, not hardcoded document names).
visual-digest.md: Added Phase 5 (context update + _digest-log.md registration) + memory path.
visual-qa-agent.md: Added YAML frontmatter (was missing entirely).
dev-orchestrator.md: Added missing tools, permissionMode, maxTurns, color.
drift-auditor.md: Fixed non-standard role: → description:, added tools/permissionMode/maxTurns/color.
frontend-test-runner.md: Added tools, maxTurns, color. Changed bypassPermissions → acceptEdits.
5 agents (architect, business-analyst, sdd-spec-writer, meeting-risk-analyst, diagram-architect): Updated reglas-negocio.md references to generic RULES.md (o reglas-negocio.md).
coherence-validator.md + reflection-validator.md: Fixed bare MEMORY.md → full 3-level path.
.gitignore: Added git/ to exclude local infrastructure.

Fixed

Agent audit: Scanned all 43 agents across 7 dimensions. Fixed 15 agents with structural issues (incomplete frontmatter, outdated references, missing memory paths, missing Phase 4/5 in digest agents). Report: output/agent-audit-20260319.md.

3.1.0 — 2026-03-17

Era 116 — Universal digest traceability + visual-digest agent with 4-pass contextual OCR pipeline.

Added

Rule: digest-traceability.md: Universal traceability for all data sources processed by Savia (documents, transcriptions, audio, web, repos, diagrams). Idempotency protocol ensures no source is processed twice. Centralized _digest-log.md per project with change detection and archival strategy. Privacy-first: log lives inside projects/ (gitignored).
Agent: visual-digest (Opus 4.6): 4-pass contextual OCR for whiteboard photos, handwritten notes, paper diagrams, screenshots, and slides. Pipeline: raw extraction → project context loading (reads team/members, business rules, prior digests) → resolution with homonym disambiguation protocol (3 Sergios, 2 Javiers, 2 Alvaros) → cross-verification against verbal digests. Tested: resolved 10 more items than naive OCR, corrected 3 misidentifications.

Changed

Digest workflow: All digest agents (meeting-digest, document-digest, visual-digest) must now consult _digest-log.md before processing and update it after completion.
agents-catalog.md: Updated from 37 to 39 agents (+visual-digest, +web-e2e-tester). Added Visual Digest flow.
README.md + README.en.md: Agent count updated from 34 to 39 (aligned with actual .claude/agents/ directory).
CLAUDE.md: Agent count updated from 34 to 39.

3.0.0 — 2026-03-16

Era 115 — Agent memory 3-level architecture (public/private/project). Meeting digest pipeline with confidentiality judge.

Added

Agent memory 3 levels: public-agent-memory/ (git-tracked best practices), private-agent-memory/ (gitignored personal context), projects/{p}/agent-memory/ (gitignored client data)
Agent: meeting-digest (Sonnet 4.6): extracts team profiles, business context and action items from meeting transcriptions (VTT, DOCX, TXT)
Agent: meeting-risk-analyst (Opus 4.6): cross-references meeting decisions against business rules, detects interpersonal conflicts, duplicities, dependencies and risky decisions
Agent: meeting-confidentiality-judge (Opus 4.6): validates that confidential data marked during extraction does not leak to project files
Command: /meeting-digest: 3-phase pipeline — extraction, confidentiality filter, risk analysis
Rule: agent-memory-isolation.md: immutable rule enforcing 3-level separation with RGPD compliance

Changed

agent-memory-protocol.md: rewritten for 3-level architecture (public/private/project)
agent-self-memory.md: rewritten for 3-level architecture with classification criteria
agents-catalog.md: updated from 34 to 37 agents, added Meeting Digest flow
memory-system.md: added Agent Memory section documenting 3 levels
.gitignore: private-agent-memory/ added, public-agent-memory/ explicitly tracked

Removed

.claude/agent-memory/: legacy single-level agent memory (migrated to 3 levels)

2.99.0 — 2026-03-16

Era 114b — Windows installer zero-touch: auto-install deps, PATH config, parse fixes.

Fixed

install.ps1: ASCII art reading "Saxia" instead of "Savia"
install.ps1: PowerShell parse errors from em dashes and subexpressions in double-quoted strings
install.ps1: Windows python3.exe Store stub causing NativeCommandError
install.ps1: Clone failure when running from inside the repo
install.ps1: Unicode box-drawing chars rendering as mojibake in PowerShell terminal

Changed

install.ps1: Auto-install missing dependencies (Git, Node.js, Python, jq) via winget/choco instead of just detecting and aborting
install.ps1: Add Claude Code ~/.local/bin to user PATH permanently after install
install.ps1: Fallback to ~/pm-workspace when ~/claude exists but is not a git repo

2.98.0 — 2026-03-15

Era 114 — Git Manager roadmap, E2E screenshot validation rule, settings privacy guard.

Added

Git Manager roadmap (specs/roadmap-git-manager.md): full open-source research (10 projects analyzed: Ungit, isomorphic-git, Gitea, lazygit...), technical design (17 Bridge endpoints, TypeScript interfaces, SVG graph algorithm, security patterns), and 3-week implementation plan
Rule: E2E screenshot validation (e2e-screenshot-validation.md): cross-project rule — all web E2E tests must include screenshots for visual confirmation
Script: validate-settings-local.sh: detects private data (localhost URLs, hardcoded paths, session-specific commands) in settings.local.json before commit
Domain rules: added globs frontmatter for path-specific auto-loading (41 rules)

Fixed

settings.local.json: cleaned session-specific permissions (hardcoded URLs, piped commands), kept only generic tool wildcards

2.97.0 — 2026-03-15

Era 113 — Savia Web: chat multi-thread, tool feedback, markdown quality, session fixes.

Added

Chat tool activity feed: Live progress inside assistant bubble while Savia uses tools ( Reading, Searching, Delegating...) with pulsing indicator
Chat multi-thread: Session-scoped streaming — responses don't leak between sessions. Stream cancelled on session switch
Chat session titles: "Mar 15, 18:30 — message digest" format with date+time
Chat delete persistence: Deleted sessions tracked in localStorage, won't reappear from Bridge on reload
Markdown rendering: Headings (H1-H3), 10px paragraph spacing, tables with borders, blockquotes, code blocks, lists with indentation, horizontal rules
Session active indicator: Left border accent (violet) + icon color for active session
Spec: chat-multithread, chat-tool-feedback (2 new specs)
Rule: pre-commit-bats — always run tests/run-all.sh before commit

Fixed

Session panel width: Was 0px (missing CSS width), now 260px
Delete button hover: Bigger click area, visible on hover for non-active sessions
SSE streaming: One-shot mode, Connection: close, client-side stream break
Chat identity: User context injected in every message (works with --resume)
Dashboard: Greeting field flattened from nested user.greeting
BATS: CHANGELOG Era references, hook set flags, duplicate versión entries

2.96.0 — 2026-03-15

Era 112 — Savia Web Phase 3: per-user auth, user management, chat sessions, bug fixes.

Added

Per-user tokens: Individual tokens per user in ~/.savia/bridge/users/{slug}/token, profile.json with roles
User management: Admin panel /admin/users — create/edit/delete users, role dropdown, token rotation/revocation, last-admin protection
Chat session management: Session list sidebar (260px), New Chat, switch sessions, delete, titles with "date — message digest", localStorage persistence
Chat markdown: Bubbles render markdown (bold, code, lists, links) via marked
Chat identity: Bridge injects [Contexto: usuario=Name, rol=role] in every message — Savia knows who you are
Bridge /auth/me: Returns authenticated user slug + role
Bridge user CRUD: GET/POST /users, PUT/DELETE /users/{slug}, rotate-token, revoke
File access control spec: Role-based file browser access (admin=root, user=projects only)
Create project modal: Teleport to body, responsive, z-index 9999

Fixed

SSE chat hanging: Switched from interactive stdin/stdout to one-shot streaming (claude -p)
Chat input disabled after response: Connection: close header + client-side stream break on done
Session conflict: Corrupted sessions detected and invalidated, user gets friendly message
Dashboard not loading: Bridge user.greeting flattened to match DashboardData.greeting
Chat blank on navigation: initSession skips re-init if messages in memory
Session delete not reactive: Changed filter to splice for in-place mutation
Session panel width 0: Added width: 260px + min-width

Stats

Unit tests: 228 (42 files)
E2E tests: 148 (18 files) — with screenshots for visual validation
Bridge tests: 29
Specs: 22 (21 implemented, 1 planned)

2.95.0 — 2026-03-15

Era 112 — Savia Web Phase 2: i18n fully wired, project context switch, all gaps fixed.

Added

i18n fully wired: All 12 pages + AppSidebar use useI18n() / $t(). Zero hardcoded strings
Project context switch: Dashboard, reports, pipelines, integrations stores watch projectStore.selectedId and reload on change
Vitest i18n setup: Global test setup registers i18n plugin for all component tests
Phase 2.5 spec: Create project from web (modal + scaffolding via Bridge)

Fixed

E2E clearSession sets English locale to match test assertions
useReportData now uses projectStore instead of dashboardStore for project context
All 217 unit tests + 109 E2E tests pass with i18n

2.94.0 — 2026-03-15

Era 111 — Radical Honesty Principles (Rule #24).

Added

Rule #24 — Radical Honesty: new domain rule (radical-honesty.md) with 6 prohibitions (no filler, no sugar-coating, no unearned praise, no hedging, no self-announcement, no comfort-seeking language) and 6 obligations (challenge assumptions, expose blind spots, mirror self-deception, show where they play small, objective depth, ground in personal truth)
tone.md: new honesty field (radical | standard) in user tone template

Changed

Savia persona (savia.md): personality rewritten from "warm, bonachona" to "direct, strategic, radically honest". Linguistic register, banned phrases and example phrases updated
Adaptive output (adaptive-output.md): all 3 modes (coaching, executive, technical) rewritten to follow radical honesty — no false encouragement, no hedging, quantified costs
CLAUDE.md: Rule #24 added to Critical Rules. Savia description updated

2.93.0 — 2026-03-14

Era 110 — Autonomous Pipeline Engine: local CI/CD without Jenkins.

Added

scripts/pipeline-engine.sh: Orchestrate pipeline execution from YAML definition — parses stages, respects dependencies, parallel support, dry-run mode
scripts/pipeline-stage-runner.sh: Execute individual pipeline stages (bash command or agent), log results as JSON
.claude/templates/pipeline/ci-template.yaml: Sample CI pipeline with build, test, security, lint, review stages
.claude/commands/pipeline-local-run.md: /pipeline-local-run command for local pipeline execution
tests/structure/test-pipeline-engine.bats: 7 BATS tests for engine and stage runner

2.92.0 — 2026-03-14

Era 103 — LSP-Powered Code Intelligence: best-practices-check command.

Added

.claude/commands/best-practices-check.md: /best-practices-check evaluates workspace against 5 categories (structure, hooks, context, testing, docs) with score 0-100

2.91.0 — 2026-03-14

Era 107.3 — Backlog resolver: local-first data source for commands.

Added

scripts/backlog-resolver.sh: Sourceable helper for commands — resolves backlog path, sprint ID, PBI counts by state, board summary, sprint items. Local-first with API fallback
tests/structure/test-backlog-resolver.bats: 8 BATS tests for resolver functions

2.90.0 — 2026-03-14

Era 102 — Real-Time Observatory: statusline, notifications, activity log.

Added

scripts/statusline-provider.sh: HUD data provider for Claude Code statusline — outputs JSON with tier, context window, project, branch, PBI counts
scripts/notify.sh: Cross-platform desktop notifications (Linux notify-send, macOS osascript, fallback echo)
.claude/commands/agent-activity.md: /agent-activity command showing structured log of recent agent executions
tests/structure/test-observatory.bats: 6 BATS tests for observatory components

2.89.0 — 2026-03-14

Era 107.2 — Sync Adapters: Azure DevOps, Jira, and GitHub Issues bidirectional sync.

Added

scripts/sync-adapters/adapter-interface.sh: Common functions for sync — logging, field extraction, state mapping, conflict detection
scripts/sync-adapters/azure-devops-adapter.sh: Pull/push/diff with Azure DevOps work items via REST API
scripts/sync-adapters/jira-adapter.sh: Pull/push/diff with Jira Cloud via REST API v3
scripts/sync-adapters/github-issues-adapter.sh: Pull/push/diff with GitHub Issues via gh CLI
.claude/commands/backlog-sync.md: /backlog-sync command for pull/push/diff operations
tests/structure/test-sync-adapters.bats: 11 BATS tests for adapter interface and state mapping

2.88.0 — 2026-03-14

Era 107.1 — Backlog Sovereignty: markdown-based backlog as source of truth.

Added

scripts/backlog-init.sh: Initialize backlog structure for any project (config, sprint folder, PBI directory)
scripts/backlog-pbi-crud.sh: Create, read, update, list, archive PBIs as markdown files with YAML frontmatter
scripts/backlog-query.sh: Query PBIs by state, sprint, assigned, priority, type. Outputs table, JSON, or count
.claude/templates/backlog/: PBI template, sprint-meta template, config template
tests/structure/test-backlog-structure.bats: 11 BATS tests for backlog init, CRUD, and query

2.87.0 — 2026-03-14

Era 100.3 — Context metrics command and session snapshot integration.

Added

.claude/commands/context-status.md: Lightweight command (Haiku) showing model, context window, tier, compact threshold, strategy, and recommendations

2.86.0 — 2026-03-14

Era 100.2 — Context Sync Persistente. Session snapshot save/load between sessions.

Added

scripts/context-snapshot.sh: Save/load session context (project, branch, sprint, last activity) as JSON. 24h TTL, auto-expired
.claude/hooks/session-end-snapshot.sh: Stop hook that auto-saves snapshot at session end (async, 5s timeout)
tests/structure/test-context-snapshot.bats: 7 BATS tests for snapshot save/load/status

Changed

session-init.sh: Loads fresh snapshot at startup, shows recovered project in init output
.claude/settings.json: Added session-end-snapshot hook to Stop event
.gitignore: Added .claude/context-cache/

2.85.0 — 2026-03-14

Era 100.1 — Lazy Loading of Rules Domain. Tier-based rule classification and manifest.

Added

scripts/rule-usage-analyzer.sh: Analyzes domain rule usage across workspace — classifies 110 rules into tier1 (startup), tier2 (on-demand), dormant (unreferenced). Outputs JSON manifest
docs/rules/domain/rule-manifest.json: Pre-computed map of 110 rules with tier + consumers. 13 tier1, 35 tier2, 62 dormant
tests/structure/test-rule-lazy-loading.bats: 8 BATS tests for analyzer and manifest integrity

2.84.0 — 2026-03-14

Era 100.0 — Context Window Adaptive per Model. Provider-agnostic dynamic context detection.

Added

config/model-capabilities.yaml: LLM capability registry — maps models to context window, tier, and strategy. Supports Claude, GPT, Gemini, Llama, Mistral (provider-agnostic)
scripts/model-capability-resolver.sh: Detects active model, parses YAML registry, exports SAVIA_CONTEXT_WINDOW, SAVIA_MODEL_TIER, SAVIA_COMPACT_THRESHOLD env vars. Falls back to 128K/fast for unknown models
scripts/adaptive-strategy-selector.sh: Given a tier (max/high/fast), outputs JSON with lazy loading, agent budget, autocompact, and sprint loading strategy
tests/structure/test-model-capabilities.bats: 13 BATS tests covering registry, resolver, and strategy selector

Changed

session-init.sh: Model capability detection runs as first step — sets SAVIA_* vars for downstream scripts
CLAUDE.md: Updated to reflect Era 100.0 context intelligence

2.83.0 — 2026-03-14

Era 63 — Multi-user session architecture with per-user isolation in savia-bridge.

Added

Per-user session isolation: savia-bridge supports multiple concurrent users without lock contention with terminal
Two-tier auth: master token + per-user tokens via POST /auth/register
Username field: added to mobile BridgeSetupDialog and web LoginPage
Eye toggle on token field: visibility toggle for token input in web and mobile
HTTPS cert-hint: visual indicator when connecting over HTTPS
Session persistence: per-user sessions survive bridge restarts
Token toggle E2E tests: new Playwright test file for token visibility (e2e/token-toggle.spec.ts)
Multi-user sessions spec: formal specification for the feature (specs/multi-user-sessions.spec.md)

Changed

savia-bridge.py: refactored session management for per-user isolation
SecurityRepository (mobile): updated interface for username-based auth
auth store (web): session management adapted for multi-user flow
chat store (web): threading adapted for per-user context

Fixed

Bridge lock contention between web/mobile users and terminal session

2.82.0 — 2026-03-14

Era 62b — Savia Web production-ready: login system, E2E testing, modern UI, and bridge threading fix.

Added

Login system: Server URL + @username + token authentication with cookie persistence, team profile loading, and registration wizard for new users (LoginPage.vue, RegisterWizard.vue)
E2E test suite: 8 Playwright test files covering login, navigation, dashboard, theme, reports, chat, pages, and UI quality — with regression plan (specs/regression-plan.md)
web-e2e-tester agent: autonomous browser testing agent equivalent to android-autonomous-debugger (.claude/agents/web-e2e-tester.md)
Dark/light mode toggle: sidebar footer switch with localStorage persistence and full CSS variable adaptation
Chat typing indicator: animated dots spinner while waiting for bridge response
Versión auto-increment: prebuild script bumps patch versión on every npm run build
Lucide icons: replaced all emoji icons with tree-shakeable SVG icons (ISC license)
Savia logo: owl PNG from savia-mobile with transparent background for dark mode

Changed

savia-bridge.py: HTTPServer → ThreadingHTTPServer — fixes concurrent request blocking (chat no longer freezes health/dashboard/team endpoints)
LoginPage.vue: 8-second fetch timeout with AbortController — shows error instead of hanging on "Connecting..."
MainLayout.vue: auto-connect with timeout; shows login form on failure
AppSidebar.vue: Lucide icons, logo image, theme toggle, dynamic versión from package.json
AppTopBar.vue: profile name + logout button with Lucide icons
Design system: glassmorphism surfaces, Inter font, layered shadows, focus rings, spacing tokens

Fixed

Bridge single-thread blocking: /chat no longer prevents other endpoints from responding
Login "Connecting..." infinite hang when bridge is unreachable

2.81.0 — 2026-03-14

Era 62a — Savia Web: Vue.js web client for PM-Workspace dashboards with reporting endpoints.

Added

savia-web: Vue 3 + TypeScript + Vite web client with 10 dashboard pages (sprints, burndown, DORA, capacity, workload, quality, debt, cycle-time, portfolio, team health) and 10 reusable ECharts components (line, bar, gauge, pie, heatmap, sankey, scatter, radar, tree, timeline)
savia_bridge_reports.py: 8 HTTP JSON endpoints for reporting (velocity, burndown, DORA, workload, quality, debt, cycle-time, portfolio) that feed the web client
setup-savia-web.sh: Production build and serve script for Savia Web with health checks and graceful shutdown
projects/savia-web/CLAUDE.md: Project configuration and development guide

2.80.0 — 2026-03-13

Era 62 — Agent and skill enrichment: handoff templates, assignment matrix, enhanced verification, skill metadata, and agent identity profiles.

Added

handoff-templates rule: 7 standardized templates for agent-to-agent transitions (Standard, QA Pass/Fail, Escalation, Phase Gate, Sprint Review, Status Report)
assignment-matrix rule: Task Type → Agent routing table (39 task types, 12 language packs, selection rules)
decisión-trees/: externalized decisión trees for agents exceeding 150-line limit

Changed

verification-before-done rule: enhanced with evidence-based quality gates, retry policy (haiku→sonnet→opus→human), escalation handoff format
skill-auto-activation rule: refined scoring (40% base + 30% context + 30% history), 7 category taxonomy, priority-based thresholds
75 skills: added category, tags, priority metadata to YAML frontmatter for intelligent routing and auto-activation
10 agents (architect, business-analyst, code-reviewer, commit-guardian, dotnet-developer, frontend-developer, sdd-spec-writer, security-guardian, test-runner, typescript-developer): enriched with Identity, Core Mission, Decisión Trees, Success Metrics
README.md / README.en.md: updated skill count (45→75), agent count alignment (34)

2.79.0 — 2026-03-13

Autonomous modes — overnight sprint, code improvement loop, tech research agent, and dev onboarding with AI buddy.

Added

/overnight-sprint command: autonomous overnight sprint — executes low-risk tasks, creates Draft PRs for human review
/code-improve command: autonomous code improvement loop — detects coverage, lint, debt opportunities and generates PRs
/tech-research command: autonomous technical research — investigates topics, generates reports, notifies designated reviewer
/onboarding-dev command: technical onboarding with AI Buddy — auto-generates 12 project docs, personalized 30/60/90 plan, 3-layer buddy agent
overnight-sprint skill (SKILL.md + DOMAIN.md): task selection, risk scoring, fail-safe with model escalation
code-improvement-loop skill (SKILL.md + DOMAIN.md): detect/improve/verify cycle with auto-categorization
tech-research-agent skill (SKILL.md + DOMAIN.md): 5-phase research pipeline (scope, search, analyze, synthesize, report)
onboarding-dev skill (SKILL.md + DOMAIN.md): buddy-ia agent with 3 layers (navigator, mentor, pair)
autonomous-safety rule: immutable safety guardrails for all autonomous modes (agent/* branches, Draft PRs, human reviewer gate, fail-safes)
docs/AUTORESEARCH.md: autonomous research methodology documentation
docs/autoresearch-cases.md: example research cases and templates

Changed

CLAUDE.md: added rule 8b (Autonomy) and 4 new skills in catalog
pm-config.md: added Autonomous Modes and Onboarding configuration sections
README.md / README.en.md: updated counters (400+ commands, 45 skills) and added autonomous modes section

2.78.0 — 2026-03-11

Reverse orgchart import — parse diagrams (Mermaid, Draw.io, Miro) to generate teams/ structure.

Added

/orgchart-import command: imports orgchart diagrams and generates department, team and member files in teams/
orgchart-import skill: 7-phase pipeline (detect format, parse, normalize, validate, detect conflicts, write, summary) with 3 conflict modes (create, merge, overwrite)
Mermaid parser: recognizes DEPT nodes, subgraphs with capacity, member nodes with lead markers (), supervisor links
Draw.io parser: identifies entities by shape styles (swimlane=dept, rounded rect=team, person shape=member, green fill=lead)
Miro parser: heuristic-based detection by color/shape/position with user confirmation fallback
Org model schema: normalized JSON contract bridging all parsers to the write phase
DOMAIN.md: Clara Philosophy documentation for the skill

Changed

diagram-config.md: added ORGCHART_IMPORT_MODES and ORGCHART_IMPORT_DEFAULT_MODE constants
README.md / README.en.md: documented orgchart import capability in code intelligence section

2.77.0 — 2026-03-10

Orgchart diagram generation from teams data — new diagram type for /diagram-generate.

Added

Orgchart diagram type: /diagram-generate {dept} --type orgchart generates hierarchical team diagrams from teams/ data, exportable to Draw.io, Miro or local Mermaid
Orgchart shapes reference: Draw.io XML snippets for department containers, team nodes, person shapes (lead vs member), hierarchy and supervisor links (orgchart-shapes.md)
Orgchart Mermaid template: graph TB template with subgraphs per team, lead markers (), @handle-based naming, PII-Free compliant (orgchart-mermaid-template.md)
Test suite: scripts/test-orgchart-diagrams.sh — 45 tests covering config, structure, shapes, template, command integration, skill integration and Mermaid output generation

Changed

diagram-config.md: DIAGRAM_TYPES now includes orgchart, added ORGCHART_DATA_DIR and ORGCHART_OUTPUT_DIR
diagram-generation SKILL.md: Added Orgchart to supported types, teams data source note, two new reference files
diagram-generate command: Added --type orgchart with 6-step orgchart-specific flow (dept validation, hierarchy read, Mermaid generation, export, metadata, presentation)
README.md / README.en.md: Documented diagram generation capabilities including orgchart

2.76.5 — 2026-03-10

Fixed — Savia Mobile v0.3.46: one-shot mode + command pre-fill

Savia Mobile chat fix (CRITICAL): Switched from interactive bidirectional stream-json to one-shot mode (-p --output-format stream-json) — Claude CLI interactive mode does not work as subprocess. Each message now launches a fresh process with --resume for session continuity
Command pre-fill from palette: Commands screen now passes selected command text to Chat via navigation query parameter (?command=encoded). ChatInput uses remember(key) to reinitialize state on new command
Unified chat navigation: Merged duplicate composable() routes into single chat?conversationId={}&command={} with both optional params
Bridge interactive session manager: Added InteractiveSession class with full bidirectional protocol, permission request/response flow, and /chat/permission endpoint (infrastructure for future interactive mode)
Permission request model: New StreamDelta.PermissionRequest and sendPermissionResponse() in SaviaBridgeService for tool approval UI
Regression test: 10/10 pass on OUKITEL C36 (Android 14), zero permission popups across 55 operations

2.76.4 — 2026-03-10

Fixed — Era 104: Auditoría — stdin timeout en 4 hooks + documentación

Hook stdin timeout (CRÍTICO): Aplicado timeout 3 cat a block-credential-leak.sh, block-force-push.sh, block-infra-destructive.sh, tdd-gate.sh y validate-bash-global.sh. Usaban INPUT=$(cat) sin timeout, causando "PreToolUse:Bash hook error" al bloquear stdin indefinidamente
hooks/README.md: Documentados 7 hooks faltantes (android-adb-validate, block-project-whitelist, compliance-gate, agent-dispatch-validate, memory-auto-capture). Total: 19 hooks documentados
PCRE fallback: Reemplazado grep -oP '\K' por alternativa POSIX en block-infra-destructive.sh y tdd-gate.sh

2.76.3 — 2026-03-10

Fixed — Era 104: Compound command patterns + APK test robustness

APK integration tests: Added dismiss_system_dialogs() to handle Android 13+ notification permission dialog that blocked all test UI interactions. Tests now re-launch app if it goes to background after Bridge connection. All 23/23 APK tests pass
Compound &&/|| patterns: Added Bash(cd * && *), Bash(cd * || *), Bash(source * && *), Bash(. * && *) to default permission whitelist. Claude Code's * wildcard is shell-aware and does not cross &&/|| operators — a simple Bash(cd *) never matched cd dir && cmd
Hook robustness: Improved validate-bash-global.sh stdin parsing with printf '%s' and IFS= read for reliable JSON handling
setup-claude-permissions.sh: Updated default generated patterns to include compound command entries and network utilities (ip, ifconfig, hostname)

2.76.2 — 2026-03-10

Fixed — Era 104: adb-run.sh wrapper + hook error fix

adb-run.sh: New single-command runner that replaces source wrapper.sh && cmd1 && cmd2 chains. Claude Code's shell-aware * doesn't cross &&/|| operators, making compound command patterns impossible to whitelist. adb-run.sh encapsulates source + functions in one simple command
Hook stdin fix: validate-bash-global.sh now uses read -t 2 with timeout instead of cat (which could hang indefinitely waiting for stdin)
SKILL.md rewrite: All examples now use ./scripts/adb-run.sh pattern exclusively

2.76.1 — 2026-03-10

Fixed — Era 104: CHANGELOG link enforcement + Claude Code permission cleanup

CI Gate 6: CHANGELOG Versión Links: Added validation to ci-extended-checks.sh that fails CI if any ## [X.Y.Z] header lacks its reference link at the end of the file. Prevents the recurring issue of missing comparison links
Claude Code permission setup: New scripts/setup-claude-permissions.sh generates settings.local.json with glob-based permission patterns (auto-detects Android SDK, JAVA_HOME, ADB). Eliminates the ~50 exact-match ADB commands that caused constant permission popups
Installer integration: Added Step 6 to install.sh — runs permission setup automatically during workspace installation
Shell-aware permission patterns: Fixed compound && command patterns — Claude Code is shell-aware and won't auto-approve chained commands with prefix-only patterns. Added explicit Bash(source wrapper.sh && *) patterns
Fixed missing [2.76.0] comparison link in CHANGELOG.md

2.76.0 — 2026-03-10

Added — Android Debug Agent: autonomous device testing

ADB wrapper library (scripts/lib/adb-wrapper.sh): 40+ functions for device management, APK lifecycle, screenshots, UI hierarchy, interaction (tap/swipe/type/scroll), logcat analysis, crash detection, and element finding. Auto-detects ADB binary and device, includes retry logic and structured JSON output
Security hook (.claude/hooks/android-adb-validate.sh): PreToolUse hook classifying ADB operations into safe (auto-approved), risky (logged), and blocked (rejected). Prevents destructive commands while allowing autonomous debugging without permission prompts
Debugger skill (.claude/skills/android-autonomous-debugger/SKILL.md): Complete workflow for autonomous debug cycles — install, launch, interact, detect crashes, capture evidence, report results
Integration test suite (scripts/tests/test-adb-wrapper.sh): 44 tests covering core functions, security classification, device management, visual capture, logcat, Savia Mobile integration, and hook validation. All tested against physical OUKITEL C36 device
Documentation (docs/android-debug-agent.md): Full API reference, architecture diagram, use cases for PM/QA smoke testing, developer debugging, and CI verification. Includes security model and environment variable reference

Fixed

Bridge duplicate text: Response text no longer appears twice in chat bubbles (result event suppressed when streaming already delivered the content)
Bridge session persistence: Known sessions saved to ~/.savia/bridge/known-sessions.json — multi-turn conversations survive bridge restarts
Bridge "already in use" recovery: If session conflict detected, session marked as known for automatic retry

Added

Chat timestamps: Message bubbles display HH:mm time for traceability (SimpleDateFormat with remember for performance)

2.75.0 — 2026-03-10

Added — OpenCode Integration: PM-Workspace compatibility layer

OpenCode compatibility layer: Created .opencode/ with symlinks to original directories (.claude/, docs/, projects/, scripts/) for OpenCode tool usage while preserving Claude Code functionality
Cross-platform installers: install.sh (Linux/macos) and install.ps1 (Windows) similar to Claude Code's installers but adapted for OpenCode
Hooks integration solution: Git hooks automation (scripts/install-git-hooks.sh) installs pre-commit, pre-push, and commit-msg hooks that automatically validate security/quality gates missing in OpenCode
OpenCode wrappers: scripts/opencode-hooks/wrappers/safe-*.sh validate commands before executing with OpenCode tools, bridging the security/quality gap from missing automatic hook execution
Documentation: Updated .opencode/README.md with comprehensive OpenCode usage guide and hooks strategy explaining why integration doesn't affect Claude Code's ongoing Savia Mobile work
Branch isolation: Created feat/opencode-hooks-integration branch with all OpenCode changes, ready for PR creation without interfering with Claude Code's work on main branch
CHANGELOG audit and fix: Consolidated scattered versión links to end of file, added missing links for versions 2.73.0–2.74.2, ensuring compliance with changelog-enforcement rule

2.74.2 — 2026-03-09

Fixed — Era 103: Chat runtime crash + crash handler

Fixed Savia Mobile chat crash: replaced LocalLifecycleOwner/DisposableEffect with ProcessLifecycleOwner in SaviaNotificationManager
Added global crash handler (SaviaApp.installCrashHandler()) logging to logcat + last_crash.log
Simplified ChatViewModel — removed isAppInForeground field

2.74.1 — 2026-03-09

Fixed — Era 103: ChatViewModel crash + build gate

Fixed ChatViewModel crash: added missing SaviaNotificationManager mock in unit and integration tests (5 call sites)
New buildAndPublish Gradle task: tests → build → publish chain. If tests fail, no APK gets published. Replaced unsafe finalizedBy pattern
Added Savia Mobile build rule to CLAUDE.md: always ./gradlew buildAndPublish, never assembleDebug

2.74.0 — 2026-03-09

Changed — Era 103: All gaps implemented — code review 4-judge panel, file browser, notifications, output persistence

Code Review gaps (all 4 done): performance-auditor as 4th consensus judge (weights 0.3/0.3/0.2/0.2), parallel dispatch via dag-scheduling, enforced risk-based routing, per-finding confidence curves
Bridge file browser: new GET /files and GET /files/content endpoints with path traversal prevention and 500KB limit
Savia Mobile file browser: FileBrowserScreen with directory listing, code viewer (monospace + line numbers), markdown renderer (Markwon), breadcrumb navigation. New Screen.Files route + HomeScreen quick action
Android notification permission: POST_NOTIFICATIONS for Android 13+, runtime permission request on launch, SaviaNotificationManager singleton with "response complete" notification when app is backgrounded
Output persistence: SavedOutputEntity Room table (v2 migration) for persisting Claude-generated outputs (code, reports, snippets) with favorites, type filtering, and conversation linkage. SavedOutputDao with CRUD + favorites
Updated docs/roadmap-code-review-improvements.md: all 4 gaps marked as implemented

2.73.0 — 2026-03-09

Changed — Era 102: Pentester integration, Savia Mobile non-blocking chat, code review roadmap

Integrated pentester agent into documentation: agents-catalog (34 agents), adversarial-security rules, README security section
Savia Mobile non-blocking chat: message queue (Channel<String>) allows sending multiple messages without waiting for response. Spinner moved from input box to streaming bubble. Pending message count badge on send button
Fixed SQLCipher dependency visibility: implementation → api in data module so DatabaseModule (app module) can resolve SupportOpenHelperFactory
Added code review improvements roadmap: confidence scoring, performance analyzer agent, parallel judge dispatch, adaptive review depth
New doc: docs/roadmap-code-review-improvements.md

2.72.0 — 2026-03-09

Changed — Era 101: Pentester v2: Shannon-inspired pipeline architecture

Rewrote pentester agent with autonomous 5-phase pipeline: pre-recon → recon → vulnerability analysis (5 parallel classes) → exploitation (proof-based) → reporting
"No exploit, no report" policy: only Level 3 (impact demonstrated) findings appear in final report. L1 (theoretical) and L2 (partial) go to "Failed Attempts" section for transparency
Queue-driven architecture: Phase 3 produces JSON vulnerability queues (03-vuln-{class}.json) consumed by Phase 4 exploitation — prevents hallucinated findings
Added JSON schema (queue-schema.json) for formal queue validation with per-class ID patterns (INJ-xxx, XSS-xxx, AUTH-xxx, SSRF-xxx, AUTHZ-xxx)
Phase prompts with Shannon-style framing: "sole responsibility", "mathematical rigor", cascading intelligence between phases
Proof templates for L3 evidence: data_extraction, rce, auth_bypass, info_leak
New queue validator script (validate-queue.py) for Phase 3→4 handoff validation
Test suite expanded from 65 to 73 tests: new CAT-11 (Pipeline Architecture, 8 tests) with mandatory 100% on proof enforcement (S-04, S-05)
Inspired by KeygraphHQ/Shannon (96.15% on XBOW benchmark)

2.71.0 — 2026-03-09

Added — Era 100: Pentester lab infrastructure

Docker Compose lab (tests/pentest-lab/) with intentionally vulnerable services for controlled security testing
Lab orchestrator script (run-lab.sh) with up/down/status/test commands
Finding validator (validate-findings.py) for automated format checking

2.70.0 — 2026-03-09

Added — Era 99: Pentester agent for dynamic security testing

New pentester agent (95L): elite ethical hacker for dynamic penetration testing across dev/pre/production environments. References pentesting skill for detailed arsenal
New pentesting skill (98L): OWASP Top 10, PTES methodology, MITRE ATT&CK mapping, CVSS v3.1 scoring, detailed checklists
Expertise areas: web app attacks, API security, authentication/authorization, network/infrastructure, container/cloud, cryptography, post-exploitation
Environment-aware rules: aggressive in dev, moderate in pre, restrictive in production
Integration with existing security pipeline (security-defender → security-auditor → pentester retest)
Test suite with 65 tests across 10 categories (mandatory 100% on reporting quality and environment awareness)

2.69.0 — 2026-03-09

Security — Era 98: Full audit and remediation (55 findings)

Comprehensive security audit across all of pm-workspace with full same-day remediation.

Audit — 55 findings identified (18 critical, 22 high, 15 medium) across 6 areas: Android app, Bridge, dotnet-microservices, shell scripts, CI/CD, installers. Full report in SECURITY-AUDIT-2026-03-09.md.
Android — SQLCipher enabled for Room Database (C2), logging restricted to DEBUG builds (C6), passphrase encoding fix (A11), cleartext traffic documentation (M4).
Bridge v1.6.0 — Input validation regex (C3), PAT encrypted with Fernet (C4), auth required on sensitive endpoints (C5), path traversal prevention (A1), SSE connection limit (A2), rate limiting on auth (A3), security headers (A4), CORS restricted (A5), body size limit 1MB (A6), log sanitization (A7), YAML injection prevention (M1), session ID validation (M2), minimum TLS cipher suite v1.2 (M3).
Kubernetes — NetworkPolicy default-deny (A14), RBAC with dedicated ServiceAccounts (A15), Pod Security Context (A16), mTLS TODO (A17), image pinning (A18), worker health checks (M9), secrets TODO (M10).
dotnet-microservices — Docker .env for credentials (C7), K8s secrets template (C11), CORS restricted (C12), JWT secret placeholder (C13), Dockerfile npm ci --omit=dev (M11), JWT logging (M12), Production templates (M14).
Shell scripts — bash -c → eval in 44 test scripts (C10), trap quoting (C15), curl | sh safety (C14/C17), irm | iex warning (C18), atomic mv (A8), mktemp -d (A19), sudo validation (A20), tar safety (A21), temp cleanup (M5).
CI/CD — SHA pinning in Actions (C9), npm versión pinning (C8), jq mandatory in hooks (C16), expanded secret patterns (A13), tag validation (A9), explicit permissions (A22), BATS SHA pinning (M6), improved secret regex (M7).
Infrastructure — Systemd hardening (A10), .gitignore binaries (A12), SECRETS-ROTATION.md (M13), plan-gate.sh 30s timeout (M15).
PR Guardian — New Gate 8: CHANGELOG required for code PRs. Exempts docs, chore, ci, style types unless they touch domain rules (.claude/rules/). Previous Gate 8 (PR Digest) renumbered to Gate 9.
Language rule — Mandatory English for all versioned content (CHANGELOGs, commits, PR titles, READMEs). Added to github-flow.md. Both CHANGELOGs translated from Spanish to English.
PRs: #280, #281, #282, #283, #285, #286

2.68.0 — 2026-03-09

Added — Era 97: Savia Mobile v0.3.34: Full Dashboard + Bridge REST (Sprint 2026-04)

Second major release of Savia Mobile with functional dashboard, chat fixes, robust auto-update, and integrated test pipeline.

Dashboard (Home) — Project selector with filtered search, sprint selector, sprint progress bar with story points, blocked items + hours metrics, My Tasks section, Recent Activity feed, Quick Actions (See Board, Approvals), FAB for quick capture. Project selection persists across reloads (local storage).
Secondary screens (REST) — Kanban board, Time log, Approvals, Capture, Git Config, Team Management, Company Profile — all via Bridge REST endpoints.
Chat fixes — Eliminated duplicate messages (Room as single source of truth), fixed CLAUDECODE nested session error (Bridge strips env var from subprocess), slash command autocomplete (8 commands).
Auto-update — APK download progress bar (LinearProgressIndicator + %), "Check updates" button in both Profile and Settings, reset state on re-check.
Build pipeline — Versión auto-increment at Gradle configuration phase (fixes versión lag), unit tests as mandatory gate before APK publish, assembleDebug runs testDebugUnitTest automatically, publishToBridge + publishToDist only if tests pass.
Tests — 48 unit tests passing (HomeViewModelTest added: 5 tests for dashboard load, project selection, persistence, errors). Spec coverage: Chat, Home, Settings, Profile, Navigation.
Bridge v1.5.0 — POST /timelog endpoint, CLAUDECODE env var stripped from Claude CLI subprocess, all REST endpoints verified (/kanban, /timelog, /approvals, /capture, /profile, /dashboard).
Path: projects/savia-mobile-android/, scripts/savia-bridge.py

2.67.0 — 2026-03-08

Added — Era 96: Savia Mobile: Android App + Bridge Server

Native Android companion app for pm-workspace with Python Bridge server.

Savia Mobile Android — Native Kotlin/Jetpack Compose app with Clean Architecture (:app, :domain, :data). Chat with Claude via SSE streaming, session persistence (Room + Tink AES-256-GCM), Material 3 violet theme, dual-backend (Bridge primary, API fallback). 39 Kotlin files, 157 tests.
Savia Bridge — Python HTTPS server (port 8922) wrapping Claude Code CLI. SSE streaming, session management, Bearer token auth, auto-TLS. HTTP install server (port 8080) for APK distribution. 1,191 lines, v1.2.0.
Updated installers — install.sh and install.ps1 now include Step 6: automatic Bridge setup (systemd/launchd/Windows service, token generation, health check).
Documentation — KDoc on all 39 source files, 8 specs rewritten, 3 new guides (ARCHITECTURE, SETUP, BRIDGE-GUIDE), API reference, CHANGELOG.
Path: projects/savia-mobile-android/, scripts/savia-bridge.py, scripts/savia-bridge.service

2.66.0 — 2026-03-08

Added — Era 95: Rules Topology & Consolidation

Rules dependency analysis and workspace governance tooling.

Rules topology analyzer (scripts/rules-topology.sh) — cross-reference map, orphan detection, duplicate detection with --summary, --json, --graph modes
105 domain rules analyzed, 25 orphans identified (23%), 0 duplicates
CI integration — --ci mode with 20% orphan threshold gate

2.65.0 — 2026-03-08

Added — Era 94: CI Pipeline Complete

Extended CI validation covering all workspace components.

CI extended checks (scripts/ci-extended-checks.sh) — 5 validation categories: skills frontmatter, rule dependencies, hook safety flags, agent file size, docs link validation
Added to CI workflow — runs automatically on PR and push to main
All 5 checks passing — 67 skills, 105 rules, 17 hooks, 33 agents, 44 docs validated

2.64.0 — 2026-03-08

Added — Era 93: Agent Accountability

Agent activity tracking and accountability dashboard.

Agent activity dashboard (scripts/agent-activity.sh) — reads JSONL traces from agent-trace-log hook, modes: --summary, --json, --recent N
6 BATS tests for agent activity dashboard (tests/structure/test-agent-activity.bats)
22 test suites, 199 tests — all passing

2.63.0 — 2026-03-08

Added — Era 92: MCP Server Specification

Model Context Protocol server specification for pm-workspace.

MCP server spec (mcp/pm-workspace-server.json) — 8 tools (sprint-status, pbi-decompose, security-scan, coverage-report, workspace-health, component-index, risk-score, capacity-check), 3 resources, 2 prompts
Follows MCP 1.0 specification standard

2.62.0 — 2026-03-08

Added — Era 91: Alpha Skills Maturation

Systematic upgrade of alpha-maturity skills to beta.

13 skills upgraded alpha → beta (banking-architecture, context-optimized-dev, evaluations-framework, google-sheets-tracker, headroom-optimization, non-engineer-templates, postmortem-training, resource-references, sdlc-state-machine, semantic-memory, session-recording, skills-marketplace, visual-quality)
Distribution: 51 stable, 15 beta, 1 alpha

2.61.0 — 2026-03-08

Added — Era 90: Technical Documentation

Comprehensive technical documentation for workspace internals.

HOOKS.md — all 17 hooks documented with exit codes, types, test coverage
AGENTS.md — all 33 agents with decisión tree and category grouping
ARCHITECTURE.md — component hierarchy, data flow, directory structure
TROUBLESHOOTING.md — common issues, debugging commands, hook inspection

2.60.0 — 2026-03-08

Added — Era 89: Hook Coverage 100%

Complete test coverage for all 17 hooks.

11 new BATS test suites — 69 new tests covering all previously untested hooks
Fixed hook safety flags — set -uo pipefail (not -euo) for all hooks
Fixed pipefail edge cases — || true guards for grep pipelines on empty input
21 suites, 193 tests — 100% hook coverage

2.59.0 — 2026-03-08

Added — Era 88: Script Hardening

Security hardening across all hooks and test scripts.

set -uo pipefail added to 14 hooks that were missing safety flags
Replaced eval with bash -c in 44 test scripts
Fixed hardcoded paths — absolute user paths → $ROOT in 5 scripts
5 BATS tests for script safety validation

2.58.0 — 2026-03-07

Added — Era 87: Strategic Vision & Health Dashboard

Workspace health metrics and strategic roadmap consolidation.

Workspace health dashboard (scripts/workspace-health.sh) — 6-dimension health scoring: skill completeness, command completeness, maturity distribution, test coverage, security posture, documentation
Current health: 84% (Grade B)
Roadmap update — Eras 79-87 stability roadmap added to docs/ROADMAP.md
JSON/CI modes — machine-readable output, 60% threshold gate

2.57.0 — 2026-03-07

Added — Era 86: Vulnerability Scanner

Deep security analysis for workspace scripts.

Vulnerability scanner (scripts/vuln-scan.sh) — 8-section analysis: eval usage, unquoted vars, temp files, HTTP security, hardcoded paths, permissions, strict mode, input validation
Severity separation — vulnerabilities block CI, warnings are informational
CI integration — added to bats-tests workflow

2.56.0 — 2026-03-07

Added — Era 85: Mock Mode

Reusable mock environment for offline testing.

Mock library (scripts/lib/mock-env.sh) — mock functions for Azure DevOps, MCP servers, sprint data, team data
Auto-detection — --mock flag or PM_MOCK environment variable
8 BATS tests validating all mock functions

2.55.0 — 2026-03-07

Added — Era 84: Discoverability

Component index and onboarding documentation.

Index generator (scripts/generate-index.sh) — --summary, --json, --markdown modes for all 454 commands, 67 skills, 33 agents, 17 hooks
Quick-start guide (docs/QUICK-START.md) — 5-minute onboarding

2.54.0 — 2026-03-07

Added — Era 83: Maturity Levels

Maturity classification for all workspace skills.

Maturity levels — alpha|beta|stable field added to all 67 skill SKILL.md files
Results: 51 stable, 2 beta, 14 alpha
Frontmatter standardization — 14 skills without frontmatter now have proper --- blocks
Classification script (scripts/add-maturity-levels.sh)

2.53.0 — 2026-03-07

Added — Era 82: Security Hardening

Security audit tooling and credential protection.

Security scan (scripts/security-scan.sh) — 5-section audit: credential patterns, hardcoded URLs, security infrastructure, hook test coverage, .gitignore completeness
CI integration — --ci mode gates on findings (warnings informational)
Hardened .gitignore — added .env.*, *.p12, *.pfx, credential/secret wildcard patterns
Verbose/summary modes — --verbose for full pass/fail detail, default summary for quick checks

2.52.0 — 2026-03-07

Added — Era 81: Coverage Metrics

Comprehensive coverage reporting across all workspace components.

Coverage report (scripts/coverage-report.sh) — weighted scoring across hooks, commands, skills, test quality
Multiple output modes — --summary, --json, --markdown, --ci (60% threshold gate)
CI integration — coverage report runs in bats-tests workflow
Current metrics: hooks 35%, commands 100%, skills 98%, overall 65%

2.51.0 — 2026-03-07

Added — Era 80: Test Quality Upgrade

Test quality audit tooling and structural integrity tests.

2 new BATS suites — workspace-structure (20 tests: settings.json, frontmatter, hooks, skills, OSS files) + changelog-integrity (7 tests: semver, ordering, Era refs)
BATS in CI — GitHub Actions workflow now runs all BATS tests on every push/PR
Test quality audit (scripts/audit-test-quality.sh) — classifies 104 test files by level (L0-L3), reports 62% real tests
Total test count: 8 suites, 111 tests, all passing

2.50.0 — 2026-03-07

Added — Era 79: BATS Testing Framework

Comprehensive unit testing infrastructure for all Claude Code hooks using BATS (Bash Automated Testing System).

6 test suites, 84 tests covering all 6 PreToolUse hooks:
- test-block-credential-leak.bats (19 tests) — 11 credential patterns + safe commands
- test-validate-bash-global.bats (17 tests) — 7 dangerous command gates
- test-agent-dispatch-validate.bats (10 tests) — 5 dispatch context validations
- test-block-force-push.bats (9 tests) — force push, main/master push, amend, reset
- test-block-infra-destructive.bats (11 tests) — terraform, az, aws, kubectl destructive ops
- test-tdd-gate.bats (18 tests) — TDD enforcement for production code
Test runner (tests/run-all.sh) with TAP output, filtering, and suite-level reporting
Test fixtures — reusable JSON inputs for hook testing
Phase 1 of 9-phase stability roadmap

2.49.0 — 2026-03-07

Added — Era 78: Agent Dispatch Validation

Pre-dispatch hook system that validates subagent prompts contain required project context before execution.

agent-dispatch-validate.sh hook — PreToolUse hook (matcher: Task) that inspects prompts sent to subagents.
agent-dispatch-checklist.md rule — Reference checklist per task type (commands, CHANGELOG, skills, rules, git ops).
Blocking validation — Missing critical context (frontmatter for commands, ordering for CHANGELOG) blocks dispatch (exit 2).
Warning validation — Missing recommended context (example references, CI mention) warns but allows (exit 0).
settings.json updated — Registered new PreToolUse hook for Task matcher with 5s timeout.

Changed

Prevents recurrence of Era 77 frontmatter issue where agents created commands without required fields.

2.48.0 — 2026-03-07

Added — Era 77: Postmortem Training Template

Postmortem process focused on reasoning heuristics rather than root cause.

/postmortem-create {incident} — Guided postmortem with 7-section template.
/postmortem-review [incident-id] — Analyze patterns and recurring gaps.
/postmortem-heuristics [module] — Compile debugging playbook from postmortems.
postmortem-training skill — Full integration with comprehension reports.
postmortem-policy rule — Mandatory for MTTR > 30 minutes.

Changed

Énfasis en Diagnosis Journey (paso a paso del razonamiento) en lugar de resumen ejecutivo.

2.47.0 — 2026-03-07

Added — Era 76: Templates for Non-Engineers

Guided interfaces for POs, stakeholders, and QA. Simplified wizards, plain language, no technical jargon required.

/po-wizard {action} — PO interface: plan-sprint, prioritize, acceptance-criteria, review.
/stakeholder-view {view} — Executive dashboard: summary, milestones, risks, budget.
/qa-wizard {action} — QA interface: test-plan, bug-report, validate, regression.
non-engineer-templates skill — 3 personas, 6 templates, step-by-step guided flows.

2.46.0 — 2026-03-07

Added — Era 75: Semantic Memory Layer

Vector-based similarity search over project memory. Three memory layers: session (ephemeral), project (JSONL), semantic (vector index).

/memory-search {query} — Natural language search over indexed memories. Top-5 results with relevance scores.
/memory-index {project} — Build/rebuild semantic vector index from agent-notes, lessons, decisions, postmortems.
/memory-stats {project} — Index statistics: entry count, last updated, coverage per source.
semantic-memory skill — Lightweight JSON vector store, embedding-based search, incremental updates.

2.45.0 — 2026-03-07

Added — Era 74: Session Recording

Record, replay, and export agent sessions for auditing, documentation, and training.

/record-start — Begin recording all session actions. Creates unique session ID, stores events in JSONL format.
/record-stop — Stop recording. Summary: duration, events count, files modified.
/record-replay {session-id} — Replay recorded session with timeline.
/record-export {session-id} — Export as markdown report to output/recordings/.
session-recording skill — Records commands, files modified, API calls, decisions, agent-notes with timestamps.

2.44.0 — 2026-03-07

Added — Era 73: PM-Workspace as MCP Server

Expose project state as MCP server. External tools can query projects, tasks, metrics and trigger PM operations.

/mcp-server-start {mode} — Start MCP server: local (stdio) or remote (SSE). Optional --read-only.
/mcp-server-status — Server status: connections, requests, uptime.
/mcp-server-config — Configure exposed resources, tools, and prompts.
pm-mcp-server skill — 6 resources, 4 tools, 3 prompts. Token auth for remote, read-only mode.

2.43.0 — 2026-03-07

Added — Era 72: Agent Skills Marketplace

Integration with claude-code-templates marketplace (5,788+ components). Browse, install, and manage Claude Code extensions.

/marketplace-search {query} — Search marketplace by keyword, type, or category.
/marketplace-install {component} — Install component from marketplace. Validates compatibility.
/marketplace-publish — Publish pm-workspace components to marketplace.
skills-marketplace skill — Marketplace integration, compatibility checks, versión management.
component-marketplace rule — 6 component types: agents, commands, hooks, MCPs, settings, skills.

2.42.0 — 2026-03-07

Added — Era 71: Evaluations Framework

Systematic evaluation of agent outputs with 5 built-in evaluation types, scoring rubrics, trend analysis, and automated regression detection.

/eval-run {eval-name} — Execute evaluation: pbi-quality, spec-quality, estimation-accuracy, review-quality, assignment-quality.
/eval-report {eval-name} — Display results and trends. Filter by --sprint, analyze with --trend.
/eval-create — Define custom evaluations with personalized rubrics.
evaluations-framework skill — 5 eval types with scoring rubrics, automated scheduling, trend analysis, regression detection.
eval-policy rule — Post-sprint evaluation, monthly evals, 10% regression alert threshold.

2.41.0 — 2026-03-07

Added — Era 70: Knowledge Graph for PM Entities

Graph-based representation of PM entities (projects, PBIs, specs, teams, decisions) with relationship queries and impact analysis.

/graph-build {project} — Build knowledge graph from project artifacts.
/graph-query {query} — Query entity relationships and dependencies.
/graph-impact {entity} — Analyze impact of changes to an entity across the graph.
knowledge-graph skill — Entity extraction, relationship mapping, traversal queries.

2.40.0 — 2026-03-07

Added — Era 69: SDLC State Machine

Formal state machine for development lifecycle with 8 states, configurable gates, and audit trail.

/sdlc-status {task-id} — Current state, available transitions, gate requirements.
/sdlc-advance {task-id} — Evaluate gates and advance to next state.
/sdlc-policy {project} — View and configure gate policies per project.
sdlc-state-machine skill — 8 states: BACKLOG→DISCOVERY→DECOMPOSED→SPEC_READY→IN_PROGRESS→VERIFICATION→REVIEW→DONE.
sdlc-gates rule — Default gate configuration with per-project overrides. Full audit trail.

2.39.0 — 2026-03-07

Added — Era 68: Google Sheets Tracker

Google Sheets as lightweight task database for POs and stakeholders. Bidirectional sync with Azure DevOps.

/sheets-setup {project} — Create tracking spreadsheet with Tasks, Metrics, and Risks sheets.
/sheets-sync {project} push|pull|both — Bidirectional sync between Azure DevOps and Sheets.
/sheets-report {project} — Generate sprint metrics from task data.
google-sheets-tracker skill — 3-sheet structure, bidirectional sync, MCP integration.

2.38.0 — 2026-03-07

Added — Era 67: Resource References (@)

Referenciable resources with @ notation for automatic context inclusion. Lazy resolution, session caching, 6 resource types.

/ref-list {project} — List available resource references with patterns and examples.
/ref-resolve {reference} — Manually resolve and preview a resource reference.
resource-references skill — 6 resource types: @azure:workitem, @project, @spec, @team, @rules, @memory.
resource-resolution rule — Lazy resolution, session cache, max 5 simultaneous, approved sources only.

2.37.0 — 2026-03-07

Added — Era 66: Headroom Context Optimization

Token compression framework achieving 47-92% reduction. Context budgets per operation.

/headroom-analyze {project} — Analyze token usage per context block with compression opportunities.
/headroom-apply {project} — Apply compressions. Preview default, --apply to persist changes.
headroom-optimization skill — 5-phase compression: analyze → identify → compress → measure → report.
context-budget rule — Max token budgets per operation type. Auto-alert if exceeded.

2.36.0 — 2026-03-07

Added — Era 65: Managed Content Markers

Safe regeneration pattern for auto-generated content. Managed markers protect manual content while allowing automatic updates.

/managed-sync [file] — Regenerate managed sections. Preview mode by default, --apply to write changes.
/managed-scan — Scan workspace for all managed markers with freshness status.
managed-content skill — Marker-based content management: scan → regenerate → validate.
managed-content rule — All auto-generated content must use markers.

2.35.0 — 2026-03-07

Added — Era 64: Verification Lattice

5-layer verification pipeline: deterministic → semantic → security → agentic → human.

/verify-full {task-id} — Run all 5 verification layers. Progressive results, stop on critical failure.
/verify-layer {N} {task-id} — Run specific layer for debugging.
verification-lattice skill — 5 layers with dedicated agents.
verification-policy rule — Layers 1-3 mandatory, L4 for risk>50, L5 always except risk<25.

2.34.0 — 2026-03-07

Added — Era 63: Risk Scoring & Intelligent Escalation

Risk-based review routing with automatic score calculation (0-100) and 4 review levels.

/risk-assess {task-id} — Calculate risk score with factor breakdown.
/risk-policy — View and update risk scoring thresholds per project.
risk-scoring skill — 4-phase pipeline: collect signals → calculate score → route review → generate report.
risk-escalation rule — Configurable thresholds, PM override, audit trail.

2.33.0 — 2026-03-07

Added — Era 62: DAG Scheduling (Parallel Agent Orchestration)

Dependency-graph-based execution for SDD pipeline. Parallelizes independent phases, reducing execution time by 30-40%.

/dag-plan {task-id} — Visualize execution DAG, critical path, and estimated time savings.
/dag-execute {task-id} — Execute SDD pipeline with parallel agents.
dag-scheduling skill — 6-phase pipeline: parse DAG → critical path → scheduling → execution → sync → reporting.
parallel-execution rule — Max 5 concurrent agents, worktree isolation, conflict prevention.

2.32.0 — 2026-03-07

Added — Era 61: Google Chat Notifier

Rich notifications for PM events via Google Chat webhooks.

/chat-setup — Guide webhook configuration and send test message.
/chat-notify {type} {project} — Send formatted notification: sprint-status, deployment, escalation, standup, custom.
google-chat-notifier skill — 5 message types with Google Chat card format.

2.31.0 — 2026-03-07

Added — Era 60: Google Drive Memory

Bidirectional sync for non-technical users. Google Drive as persistence alternative to Git.

/drive-setup — Create Drive folder structure with role-based permissions.
/drive-sync {action} — Push/pull/status operations for local↔Drive sync.
google-drive-memory skill — 4-phase pipeline: setup → sync → permissions → MCP. Timestamp-based conflict resolution.

2.30.0 — 2026-03-07

Added — Era 59: MCP Tool Search & Smart Routing

Intelligent tool discovery for 400+ commands. Auto-categorization, keyword routing, and usage-based prioritization.

tool-search-config rule — 8 command categories with routing heuristics. Auto-activates when tools exceed 128 in context.
/tool-search {query} — Search commands, skills, and agents by keyword. Discovers tools across 400+ commands.
/tool-catalog [category] — Categorized tool catalog with counts. Navigate the full command library.
smart-routing skill — Intent classification, frequency tracking, Top-20 algorithm for always-available commands.

2.29.0 — 2026-03-07

Added — Era 58: DOMAIN.md per Skill (Clara Philosophy)

Multi-level documentation layer: SKILL.md defines the "how", DOMAIN.md defines the "why" and domain context.

DOMAIN.md files added to: pbi-decomposition, product-discovery, rules-traceability, spec-driven-development, capacity-planning, sprint-management, azure-devops-queries, scheduled-messaging, context-caching, code-comprehension-report.
clara-philosophy rule — Documentation standard: every skill requires SKILL.md (how) + DOMAIN.md (why). Max 60 lines.
/plugin-validate enhancement — Checks for DOMAIN.md presence, max line count, required sections.

2.28.0 — 2026-03-07

Added — Era 57: Code Comprehension Report

Automatic mental model generation after SDD implementation. Addresses AI-generated code opacity by documenting decisions, failure heuristics, and 3AM debugging guides.

/comprehension-report {task-id} — Generate mental model report: architecture decisions, flow diagram (mermaid), failure heuristics, implicit dependencies, 3AM debugging guide. Output saved to output/comprehension/YYYYMMDD-{task-id}-mental-model.md.
/comprehension-audit {project} — Scan recent implementations, identify missing mental models, report coverage (X of Y tasks have reports). Prioritize by risk level.
code-comprehension-report skill — 7-phase pipeline: Phase 1 collect data → Phase 2 architecture decisions → Phase 3 flow diagram → Phase 4 failure heuristics → Phase 5 implicit dependencies → Phase 6 3AM debugging guide → Phase 7 generate report.
code-comprehension rule — Every dev-session completion SHOULD trigger comprehension report. Code Review E1 includes "debuggeable at 3AM?" criterion. Integration with postmortem process: link comprehension reports to incident analysis, update on failures.

2.27.0 — 2026-03-07

Added — Era 56: Scheduled Messaging Integration

Wizard-guided setup for Claude Code Scheduled Tasks with automatic result delivery to messaging platforms.

/scheduled-setup {platform} — Interactive wizard: platform selection → credential config → module generation → test → task creation. Supports: Telegram, Slack, Teams, WhatsApp (Twilio), NextCloud Talk.
/scheduled-test {platform} — Send test message to verify integration.
/scheduled-create — Create scheduled task with --notify {platform} and --cron "schedule".
/scheduled-list — List tasks with notification config and status.
scheduled-messaging skill — 5-phase pipeline, 5 platform adapters, 5 pre-built templates (standup, blocker, burndown, deploy, security).
scripts/notify-{platform}.sh — Auto-generated notification modules per platform.

2.26.0 — 2026-03-07

Added — Era 55: Prompt Caching Strategy

Context loading optimization for prompt caching. Reduces input token costs by ordering stable content first with cache breakpoints.

prompt-caching rule — 4-level caching hierarchy: PM globals → project context → skill content → dynamic request. Ordering rules and TTL guidance.
/cache-optimize {project} — Analyze context loading order and suggest reordering for optimal cache hit rates. Shows estimated token savings.
context-caching skill — Caching templates for common operations (PBI decomposition, spec generation, dev session). Token measurement patterns.

2.25.0 — 2026-03-07

Added — Era 54: Plugin Bundle Packaging

Package PM-Workspace as distributable Claude Code plugin with validation and export commands.

.claude-plugin/plugin.json — Plugin manifest with capabilities declaration, dependencies, and install paths.
/plugin-export — Package current workspace as distributable plugin. Supports --components for partial export.
/plugin-validate — Validate plugin structure: skills, agents, commands integrity, PII check, line limits.
plugin-packaging skill — Packaging logic, validation rules, versión management.

2.24.0 — 2026-03-07

Added — Era 53: Business Rules to PBI Mapping

Bridges the gap between business rules documentation and PBI creation. Automatic traceability matrix RN↔PBI with coverage analysis.

/pbi-from-rules {project} — Parse reglas-negocio.md, cross-reference with Azure DevOps PBIs, identify coverage gaps, propose new PBIs.
/pbi-from-rules-report {project} — Generate traceability matrix report without creating PBIs.
rules-traceability skill — 7-phase pipeline: parse rules → query PBIs → build matrix → gap analysis → propose PBIs → create (with confirmation) → report.
Integrates with product-discovery for complex features: auto-triggers JTBD + PRD when rule requires feature analysis.

2.23.1 — 2026-03-06

Added — Guide: Project from Scratch

Step-by-step guide for PMs to start a project from scratch: client profile, team, architecture, business rules, specs, test requirements, and implementation with Dev Session Protocol. Works across Azure DevOps, Jira, and Savia Flow.

docs/guides/guide-project-from-scratch.md (ES) — 8-step workflow with concrete examples: client profile, CLAUDE.md, equipo.md, reglas-negocio.md, PBI decomposition, spec generation, test strategy, dev session orchestration.
docs/guides_en/guide-project-from-scratch.md (EN) — English translation.
Updated guides index (ES + EN) with new entry highlighted.

2.23.0 — 2026-03-06

Added — Era 52: Dev Session Protocol (Context-Optimized Development)

5-phase development protocol for producing high-quality code within ~40% free context window. Disk-based state persistence between phases.

/dev-session — Orchestrate spec implementation: start → next (per slice) → status → review → abort. Session state in output/dev-sessions/.
/spec-slice — Break specs into context-optimized slices (≤3 files, ≤15K tokens, ≤1 business rule group). Dependency detection, critical path, YAML output.
dev-orchestrator agent — Sonnet-based planner for slice analysis, token budgets, risk assessment.
context-optimized-dev skill — Subagent delegation patterns, context priming templates, anti-patterns, token estimation formulas.
dev-session-protocol rule — 5-phase protocol definition with per-phase token budgets.

2.22.0 — 2026-03-06

Changed — Era 51: Context Window Optimization

Systematic reduction of auto-loaded context (~20,000 tokens recovered per conversation, ~10% of context window).

Language rule dedup — Merged 4 duplicated pairs (Python, Java, Go, TypeScript conventions into rules files). 4 files deleted.
Vertical rules → skills — Moved 8 vertical-specific rules from rules/domain/ to skills/references/ for on-demand loading.
csharp-rules.md — Compressed from 1,323 to 206 lines (84% reduction). All 65 SonarQube IDs + 12 ARCH patterns preserved in tabular format.
Conditional loading — Added paths: frontmatter to 17 domain rules (messaging, frontend, AI/HR, IaC, hub, etc.).
Worktree cleanup — Removed abandoned keen-chebyshev worktree (2.3 MB).

2.21.0 — 2026-03-06

Added — Era 50: Multimodal Quality Gates

Visual regression testing and wireframe validation using Claude's native vision capabilities (JPEG/PNG/WebP, up to 8000×8000px).

/visual-qa — Screenshot capture, compare against reference, regression detection, QA report. Visual match score 0-100.
/wireframe-check — Register wireframes, validate implementation, detect gaps, extract UI specs from mockups.
/visual-regression — Baseline management, regression testing, pixel-level diffing, approval workflow. 5% default tolerance.
visual-qa-agent — Sonnet-based vision agent (5-phase: input→analysis→scoring→classification→report).
visual-quality skill — Defect taxonomy, WCAG contrast checks, screenshot best practices, comparison methodology.
visual-quality-gates rule — Gate levels: auto-pass (≥90), informational (≥80), blocking (<60). Privacy-first.

2.20.3 — 2026-03-06

Added — Era 49: Connectors vs MCP Integration Architecture Decisión

ADR confirming Claude Connectors = MCP servers with managed OAuth. Connector-first strategy for end users, MCP-first for developers/CI. No code changes — documentation-only.

ADR — docs/propuestas/adr-connectors-vs-mcp.md: Full technical comparison, 11/12 tools have official Connectors, Azure DevOps remains MCP-only.
Connectors quickstart — docs/guides/guide-connectors-quickstart.md (ES+EN): 1-click setup guide, verification, per-project configuration.
Integration catalog — docs/recommended-mcps.md: Reorganized with Connectors-first + MCP community. Added coverage table mapping Connectors → pm-workspace commands.
connectors-config.md — Added ENABLE_CLAUDEAI_MCP_SERVERS auto-sync documentation and fallback message for tools without Connector.
ROADMAP.md — Added Era 49, moved Connectors evaluation from backlog to completed.

2.20.2 — 2026-03-06

Fixed — Colon-to-Kebab Command Reference Migration

Replaced all legacy colon-style command references (/bias:check, /score:diff, /sprint:review, etc.) with kebab-case (/bias-check, /score-diff, /sprint-review) across 12 files. Claude Code does not support colons in command names.

bias-check.md, score-diff.md — Added missing YAML frontmatter and fixed internal /command:name references.
agents-catalog.md, equality-shield.md, scoring-curves.md, severity-classification.md — Updated all command references from colon to kebab-case.
ROADMAP.md, CHANGELOG.md — Migrated historical references.
guides/guide-enterprise-gap-analysis.md (ES+EN) — Updated command tables.
docs/estudio-equality-shield.md, docs/política-igualdad.md — Updated references.

2.20.1 — 2026-03-06

Fixed — Documentation Consistency Audit

Full documentation audit to align all stats and features with current state after Eras 43-48.

README.md / README.en.md — Updated stats: 396+ commands (was 360+), 31 agents (was 27), 41 skills (was 38), 16 hooks (was 14), 14 guides (was 13). Added new feature sections: universal accessibility, industry verticals, adversarial security, adaptive intelligence.
CLAUDE.md — Synchronized all resource counts: commands (396+), agents (31), skills (41), hooks (16).
agents-catalog.md — Added 4 missing agents: frontend-test-runner, security-attacker, security-defender, security-auditor. Updated count: 31. Added adversarial security flow.
ROADMAP.md — Corrected agent/skill counts in Era 46 (41 skills), Era 47 (31 agents, 41 skills), Era 48 (31 agents, 41 skills, 16 hooks).

2.20.0 — 2026-03-06

Added — More Industry Verticals: Insurance, Retail, Telco (Era 48)

12 domain-specific commands for 3 additional industries.

Insurance (4 commands): /insurance-policy (POL-NNN, lifecycle: create/renew/cancel, endorsement tracking), /insurance-claim (CLM-NNN, investigation→resolution, loss ratio analytics), /solvency-report (Solvency II: SCR/MCR/own funds, RAG indicator), /underwriting-rule (criteria definition, accept/refer/decline evaluation, audit trail).
Retail/eCommerce (4 commands): /product-catalog (SKU-NNNN, pricing, stock, CSV/JSON export), /order-track (ORD-NNNN, status lifecycle, returns, revenue analytics), /inventory-manage (multi-warehouse, reorder points, dead stock alerts), /promotion-engine (PROMO-NNN, discount/BOGO/bundle/coupon, ROI analysis).
Telco (4 commands): /service-catalog-telco (SVC-NNN, voz/datos/fibra/tv, SLA, bundling), /network-incident (NI-NNNN, eTOM classification, SLA compliance), /subscriber-lifecycle (SUB-NNNN, churn-risk scoring, ARPU/LTV), /capacity-forecast-telco (utilization, trend-based forecasting, expansion planning).

Changed

ROADMAP.md — Added Era 48 entry. Removed "More industry verticals" from backlog (implemented). Updated stats: 396+ commands.

2.19.0 — 2026-03-06

Added — Adversarial Security Pipeline (Era 47)

Red Team / Blue Team / Auditor pattern for systematic security testing.

3 security agents: security-attacker (Red Team: OWASP Top 10, CWE Top 25, dependency audit, VULN-NNN structured findings), security-defender (Blue Team: patches, hardening, NIST/CIS, FIX-NNN structured corrections), security-auditor (independent evaluation, security score 0-100, gap analysis, executive summary).
/security-pipeline command — 3-phase sequential orchestration: Attack → Defend → Audit. Scopes: full, api, deps, config, secrets. Outputs per-project: vulns, fixes, and audit report.
/threat-model command — STRIDE/PASTA threat modeling with asset inventory, threat analysis (probability × impact), control mapping, gap identification, prioritized recommendations.
adversarial-security.md rule — Severity classification (critical/high/medium/low/info), scoring formula, agent independence, compliance integration (critical/high block main merge).
adversarial-security/SKILL.md skill — CVSS scoring, STRIDE mapping table, OWASP Top 10 checklist, dependency audit commands (npm/pip/dotnet).

Changed

ROADMAP.md — Added Era 47 entry. Moved adversarial security from backlog to implemented. Updated stats: 384+ commands, 30 agents, 40 skills.

2.18.0 — 2026-03-06

Added — Skill Evaluation Engine & Instincts System (Era 46)

Self-learning intelligence layer for automatic skill recommendation and adaptive behavior patterns.

/skill-eval command — Analyzes prompts against available skills with composite scoring (keywords 40% + project context 30% + history 30%). Subcommands: analyze, recommend, activate, history, tune. Auto-detects 7 project types (software, research, hardware, legal, healthcare, nonprofit, education).
/instinct-manage command — Manages Savia's learned behavior patterns with confidence scoring. Subcommands: list, add, disable, stats, decay, export. Confidence: initial 50%, +3% success, -5% failure, floor 20%, ceiling 95%. Decay: -5% per 30 days without use.
skill-auto-activation.md rule — Suggests skills above 70% relevance threshold. Max 2 suggestions per interaction. Respects focus-mode. Learns from rejections (3 consecutive → stops suggesting).
instincts-protocol.md rule — Lifecycle: detect ≥3 repetitions → propose → create → reinforce/penalize → decay → review. 5 categories: workflow, preference, shortcut, context, timing.
skill-evaluation/SKILL.md skill — Prompt tokenization, 7 project-type detection, project→skills mapping, instinct integration (+20 boost for high-confidence instincts).
Registries: eval-registry.json (skill activations), instincts/registry.json (instinct entries).

Changed

ROADMAP.md — Added Era 46 entry. Moved instincts + skill evaluation from backlog to implemented. Updated stats: 382+ commands, 39 skills.

2.17.0 — 2026-03-06

Added — Vertical-Specific Commands: 5 Industry Domains (Era 45)

20 domain-specific commands implementing all gap proposals from Era 23 guide writing. Every command follows pm-workspace conventions (≤150 lines, YAML frontmatter, project-scoped storage).

Research Lab (5 commands): /experiment-log (hypothesis→run→result→compare with EXP-NNN IDs), /biblio-search (DOI/BibTeX import, APA/IEEE/Vancouver citation export), /dataset-versión (SHA256 integrity, DVC/Git LFS support), /grant-track (lifecycle: draft→submitted→review→approved/rejected, deadline alerts), /ethics-protocol (IRB tracking with experiment cross-references, renewal lineage).
Hardware Lab (3 commands): /hw-bom (component registry, cost breakdown by category, CSV import/export), /hw-revision (REV-A/B/C lifecycle, BOM snapshots, tags: prototype/pilot/production), /compliance-matrix (CE/FCC/UL/RoHS/ISO, evidence linking, gap analysis reports).
Legal Firm (5 commands): /legal-deadline (procesal/contractual/regulatorio, auto-alerts <48h/<7d/<14d), /court-calendar (ICS import/export, scheduling conflict detection), /conflict-check (client/matter screening, privacy-preserving reports), /legal-template (demanda/contestación/recurso/contrato/poder, variable substitution), /billing-rate (hourly/fixed/contingency/mixed, invoice generation).
Healthcare (5 commands): /pdca-cycle (plan→do→check→act quality improvement cycles), /incident-register (severity classification, 5-why root cause analysis, GDPR-compliant), /accreditation-track (JCI/EFQM/ISO 9001/15189, evidence→requirement linking), /training-compliance (mandatory training, expiry alerts <30d), /health-kpi (define/measure/trend/dashboard, RAG status alerts).
Nonprofit (2 commands): /impact-metric (SDG-aligned, output/outcome/impact tiers, donor reports), /volunteer-manage (register/availability/hours, retention tracking, GDPR/LOPD).

Changed

ROADMAP.md — Era 23 gap table marked as OK implemented. Added Era 45 entry. Updated stats: 380+ commands.

2.16.1 — 2026-03-06

Changed — Repository Cleanup & Link Fixes

Removed 5 obsolete files: docs/roadmap-v1.7.0.md (subsumed by ROADMAP.md Era 22), docs/guia-adopcion-pm-workspace.docx (replaced by ADOPTION_GUIDE.md), docs/guia-incorporacion-lenguajes.docx (replaced by .md equivalent), docs/context-optimization-completed.md and docs/context-optimization-roadmap.md (work already integrated).
Fixed 8 broken links in English quick-starts (quick-starts_en/) — referenced Spanish filenames (02-estructura, 04-uso-sprint-informes, 06-configuración-avanzada, 10-kpis-reglas) instead of English (02-structure, 04-usage-sprint-reports, 06-advanced-config, 10-kpis-rules).
Fixed 2 broken links in enterprise consultancy guides pointing to non-existent quick-start.md.
Added docs/guides_en/guide-accessibility.md — English translation of the accessibility step-by-step guide (was missing from bilingual pair).
Updated references in ROADMAP.md and CHANGELOG.md to reflect removed files.

2.16.0 — 2026-03-06

Added — Automated Rule Compliance Verification (Era 44)

Pre-commit gate that blocks commits violating domain rules, independent of LLM context.

compliance-gate.sh: PreToolUse hook that runs compliance checks before every git commit. Blocks (exit 2) on violations instead of warning. Registered in .claude/settings.json.
runner.sh: Orchestrator in .claude/compliance/ running 4 check scripts on staged files. Supports --all mode for full repo scan.
check-changelog-links.sh: Verifies every ## [X.Y.Z] heading has a matching [X.Y.Z]: URL comparison link at the end of CHANGELOG.md.
check-file-size.sh: Enforces ≤150 lines for commands, rules, and skills. Excludes languages/, references/, CHANGELOG.
check-command-frontmatter.sh: Validates YAML frontmatter on newly staged commands.
check-readme-sync.sh: Verifies README.md/README.en.md ≤150 lines and bilingual sync warning.
compliance-check.md: /compliance-check command for manual verification.
RULES-COVERED.md: Coverage manifest — 4 rules automated, extensible framework for adding more.

Fix: added missing [2.15.0] comparison link in CHANGELOG.md.

Tests: bash .claude/compliance/runner.sh --all — 4/4 checks passed. CI: 14/14 green.

2.15.0 — 2026-03-06

Added — Universal Accessibility: Guided Work & Inclusive Design (Era 43)

Comprehensive accessibility system so people with disabilities can work in tech companies using pm-workspace. Central piece: Savia as digital job coach.

guided-work.md: /guided-work --task, --continue, --status, --pause. Savia decomposes any task into micro-steps (3-5 min), presents ONE at a time with a question, waits, adapts. Three guidance levels: alto (closed questions, 3 lines max), medio (2-3 steps, open questions), bajo (full checklist). Block detection: reformulates on "no sé", checks in on silence, redirects on topic change. Based on N-CAPS (Nonlinear Context-Aware Prompting System) and ADHD-aware productivity framework (arxiv 2507.06864).
focus-mode.md: /focus-mode on, off, status. Single-task mode — loads ONE PBI, hides sprint board and backlog. Complements guided-work (focus = clean environment, guided = active guidance).
accessibility-setup.md: /accessibility-setup. 5-minute conversational wizard in 4 phases (Vision → Motor → Cognitive → Wellbeing). Creates/updates accessibility.md profile fragment.
accessibility-mode.md: /accessibility-mode on, off, status, configure. Quick toggle for all adaptations with current config summary.
accessibility-output.md: Domain rule adapting ALL Savia outputs based on profile: screen_reader → text descriptions, high_contrast → no color dependency, cognitive_load:low → 5 lines max, motor → command aliases. Priority chain: screen_reader > cognitive_load > high_contrast > rest.
guided-work-protocol.md: Interaction protocol rule — task decomposition, question patterns per level, block detection table, calibrated celebrations ("Hecho. Paso X/N." — never condescending), context recovery, N-CAPS non-linear adaptation. Core principle: "The goal is not speed. It's that the person CAN complete it, at their pace, with dignity and autonomy."
inclusive-review.md: Strengths-first code reviews when review_sensitivity=true. Vocabulary mapping: "Bug"→"Caso no cubierto", "Error"→"Oportunidad de mejora". Structure: strengths → opportunities → constructive close.
accessibility.md (profile fragment template): 7th opt-in profile fragment. Fields: screen_reader, high_contrast, reduced_motion, cognitive_load (low/medium/high), focus_mode, guided_work, guided_work_level (alto/medio/bajo), motor_accommodation, voice_control, review_sensitivity, dyslexia_friendly, break_strategy, break_interval_min.
guide-accessibility.md: Step-by-step guide per disability profile — visual, motor/RSI, ADHD, autism, dyslexia, hearing. Each with recommended config, workflow example, and tips.
accessibility-es.md / accessibility-en.md: Bilingual quick-reference docs with feature list, common configurations table, and FAQ.
ACKNOWLEDGMENTS.md: Credits to all inspiring projects (claude-code-templates, kimun, Engram, BullshitBench, claude-mem), studies (LLYC, Fundación ONCE, N-CAPS, DX Core 4, NIST/ISO/EU AI Act), and people (Daniel Avila, Eduardo Díaz, Miguel Luengo-Oroz).
READMEs updated to link ACKNOWLEDGMENTS.md instead of inline credits.

Research sources: Fundación ONCE "Por Talento Digital" (30K+ trained), N-CAPS, arxiv 2411.13950 (ADHD/Autism in Software Development), arxiv 2507.06864 (ADHD-Aware Productivity Framework), DX Core 4.

Tests: test-accessibility.sh — 56 structural tests. CI: 14/14 green.

2.14.0 — 2026-03-06

Added — Enterprise Readiness: Eras 36-42 (Score 5.6 → 8.1)

Seven Eras to make pm-workspace viable for large consultancies (500-5000 employees, 50+ projects):

v2.11.0 — Multi-Team Coordination (Era 36): /team-orchestrator with create, assign, deps, sync, status. Team Topologies (Skelton & Pais), RACI, cross-team dependency detection, circular alerts. Rule: team-structure.md. Skill: team-coordination/.
v2.12.0 — RBAC File-Based (Era 37): /rbac-manager with grant, revoke, audit, check. 4-tier roles (Admin/PM/Contributor/Viewer), pre-command enforcement, append-only audit trail. Rule: rbac-model.md. Skill: rbac-management/.
v2.12.1 — Cost & Billing (Era 38): /cost-center with log, report, budget, forecast, invoice. Timesheet JSONL, EVM (EAC/CPI/SPI), rate tables, client invoicing. Rules: billing-model.md, cost-tracking.md. Skill: cost-management/.
v2.12.2 — Onboarding at Scale (Era 39): /onboard-enterprise with import, checklist, progress, knowledge-transfer. CSV batch import, 4-phase onboarding, per-role checklists. Rule: onboarding-enterprise.md. Skill: enterprise-onboarding/.
v2.13.0 — Governance & Audit (Era 40): /governance-enterprise with audit-trail, compliance-check, decisión-registry, certify. JSONL audit log, governance matrix (GDPR/AEPD/ISO27001/EU AI Act). Rules: audit-trail-schema.md, governance-enterprise.md. Skill: governance-enterprise/.
v2.13.1 — Enterprise Reporting (Era 41): /enterprise-dashboard with portfolio, team-health, risk-matrix, forecast. SPACE framework, Monte Carlo forecasting, cross-project risk aggregation. Rule: enterprise-metrics.md. Skill: enterprise-analytics/.
v2.14.0 — Scale & Integration (Era 42): /scale-optimizer with analyze, benchmark, recommend, knowledge-search. 3-tier scaling model, vendor sync, full-text search, CI/CD standardization. Rule: scaling-patterns.md. Skill: scaling-operations/.

Tests: 295 structural tests across 7 test scripts.

2.10.0 — 2026-03-06

Added — Cognitive Sovereignty: AI Vendor Lock-in Audit (Era 35)

sovereignty-audit.md: /sovereignty-audit scan, report, exit-plan, recommend. Diagnoses and quantifies organizational independence from AI providers. 5-dimension Sovereignty Score (0-100): data portability, LLM independence, organizational graph protection, consumption governance, exit optionality. Based on "La Trampa Cognitiva" (De Nicolás, 2026) — cognitive lock-in as the new enterprise dependency.
cognitive-sovereignty.md: Domain rule with lock-in evolution framework (technical→contractual→process→cognitive), 5 dimensions with weighted scoring, vendor risk matrix, alarm signals, integration with governance-audit.
sovereignty-auditor/SKILL.md: Scan orchestration (workspace analysis, score calculation), executive report generation, concrete exit plan with migration timeline, actionable recommendations mapped to pm-workspace commands.
Tests: test-sovereignty-audit.sh — 50 structural tests across command, rule, skill, and cross-references.

2.9.0 — 2026-03-05

Added — Wellbeing Guardian: Proactive Individual Wellbeing (Era 34)

wellbeing-guardian.md: /wellbeing-guardian status, configure, breaks, report, pause. Proactive nudge system for individual work-life balance — break reminders, after-hours alerts, weekend disconnection suggestions. 5 break strategies (Pomodoro, 52-17, 5-50, custom, 20-20-20 eye rule). Non-blocking philosophy: suggestions, never interruptions.
wellbeing-config.md: Domain rule with break science reference (HBR Feb 2026 research on AI-intensified work), strategy definitions, 5 nudge template categories, work schedule schema for user profiles, integration points with burnout-radar and sustainable-pace.
wellbeing-guardian/SKILL.md: Orchestration — session start (load schedule, detect after-hours), periodic check (time-based nudges), configure (interactive setup), status, pause, breaks history, weekly report with break_compliance_score.
session-init-priority.md: Added Wellbeing context entry (Media priority, ~25 tokens) for ambient work schedule awareness.
Tests: test-wellbeing-guardian.sh — 50 structural tests across command, rule, skill, and cross-references.

2.8.2 — 2026-03-05

Emergency plan hardened for offline reliability.

Changed

emergency-plan.sh/.ps1: Added connectivity check (Step 0) — fails fast with clear message if no internet. Added idempotency to cached binary path — checks ollama list before pulling. Added verification step (Step 5) — confirms what is cached and ready for offline. Updated step numbering from [1/4]...[4/4] to [1/5]...[5/5]. Extracted _extract_ollama() and _pull_small() helpers to reduce duplication.

2.8.1 — 2026-03-05

Emergency mode model alias overrides — subagents now resolve in offline mode.

Changed

emergency-setup.sh/.ps1: Map opus/sonnet/haiku aliases to local Ollama models via official Claude Code variables (ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU}_MODEL, CLAUDE_CODE_SUBAGENT_MODEL). Auto-tiered by RAM: 8GB→3b, 16GB→7b/7b/3b, 32GB+→14b/7b/3b.
emergency-plan.sh/.ps1: Pre-download qwen2.5:3b alongside main model for haiku alias differentiation.
EMERGENCY.md / EMERGENCY.en.md: New "Model Mapping" section. Updated unset commands. Claude Code Router documented as community option.
emergency-mode.md: Document model alias variables in activate subcommand.

Community contribution: Cristián Rojas identified the subagent resolution gap.

2.8.0 — 2026-03-05

Added — Context Analysis Assistant (Era 33)

context-interview.md: /context-interview start, resume, summary, gaps. 8-phase structured interview for client/project onboarding: Domain, Stakeholders, Stack, Constraints, Business Rules, Compliance (sector-adaptive), Timeline, Summary. Proactive gap detection.
context-interview-config.md: Domain rule defining 8 interview phases, session format, sector-adaptive compliance questions (fintech, healthcare, legal, education), one-question-at-a-time rule, gap detection schema, persistence targets per phase.
context-interview-conductor/SKILL.md: Interview orchestration — start, conduct phases, resume, summary, gaps. Adaptive questions per sector. Immediate persistence. Phase 8 generates consolidated summary with gap analysis.
Tests: test-context-interview.sh — 49 structural tests across command, rule, skill, and cross-references.

2.7.0 — 2026-03-05

Added — BacklogGit: Backlog Versión Control (Era 32)

backlog-git.md: /backlog-git snapshot, diff, rollback, deviation-report. Captures periodic markdown snapshots of backlogs from any PM tool (Azure DevOps, Jira, GitLab, Savia Flow, manual). Diff algorithm detects added/removed/modified items with scope creep and re-estimation metrics.
backlog-git-config.md: Domain rule defining snapshot format (YAML frontmatter + items table), 5 source types with auto-detection, diff algorithm, deviation metrics, immutability rules, frequency guidance.
backlog-git-tracker/SKILL.md: Snapshot capture (9 steps), diff with flexible references, rollback (info-only, NEVER auto-execute), deviation report with temporal metrics and ASCII charts.
Tests: test-backlog-git.sh — 41 structural tests across command, rule, skill, and cross-references.

2.6.0 — 2026-03-05

Added — Client Profiles (Era 31)

client-profile.md: /client-create {name}, /client-show {slug}, /client-edit {slug} [section], /client-list. First-class client entities in SaviaHub with identity, contacts, business rules, and projects.
client-profile-config.md: Domain rule defining client directory structure (profile.md, contacts.md, rules.md, projects/), frontmatter schema, slug generation, status/SLA validation, security rules.
client-profile-manager/SKILL.md: CRUD orchestration skill — create (10 steps), show (7 steps), edit, list with index regeneration, add-project. Error handling with fuzzy match.
Tests: test-client-profiles.sh — 41 structural tests across command, rule, skill, cross-references, and SaviaHub integration.

2.5.0 — 2026-03-05

Added — SaviaHub: Shared Knowledge Repository (Era 30)

savia-hub.md: /savia-hub command with 5 subcommands — init (local or remote clone), status, push, pull, flight-mode on|off. Centralizes company identity, org chart, clients, users, and projects in a single Git repository.
savia-hub-config.md: Domain rule defining repository structure (company/, clients/, users/), path configuration (SAVIA_HUB_PATH, SAVIA_HUB_REMOTE), local config format (.savia-hub-config.md), naming conventions, and security rules.
savia-hub-offline.md: Domain rule for flight mode — activation/deactivation, sync queue (.sync-queue.jsonl), divergence detection, auto-sync config. Safety: NUNCA auto-resolver conflictos.
savia-hub-sync/SKILL.md: Sync orchestration skill — init flow (delegates to savia-hub-init.sh), push (10-step with PM confirmation), pull (7-step with conflict handling), flight mode management.
savia-hub-init.sh: Bash init script with --remote URL, --path PATH, --help flags. Creates directory structure, company templates, clients index, .gitignore, local config, initial commit. Idempotent.
Tests: test-savia-hub.sh — 44 structural tests across command, rules, skill, init script, and cross-references.

2.4.0 — 2026-03-04

Added — One-Line Installer (Era 29)

install.sh: macOS + Linux one-line installer (curl -fsSL ... | bash). OS detection (macOS/Ubuntu/Fedora/Arch/Alpine/WSL), prerequisite checks (git, node ≥18, python3, jq), Claude Code auto-install, pm-workspace clone, npm deps, smoke test. Idempotent, configurable via SAVIA_HOME env var, --skip-tests and --help flags.
install.ps1: Windows PowerShell one-line installer (irm ... | iex). Same flow adapted for PowerShell 5.1+. Winget/Chocolatey install hints. WSL detection with cross-platform suggestion.
Tests: test-install.sh — structural validation for both installers.

2.3.0 — 2026-03-04

Added — Scoring Intelligence (Era 28)

scoring-curves.md: piecewise linear normalization for 6 dimensions (PR size, context usage, file size, velocity deviation, test coverage, Brier score). Smooth degradation with calibrated breakpoints instead of binary pass/fail. Inspired by kimun (lnds/kimun) and SonarSource/Microsoft Code Metrics.
score-diff.md: /score-diff command comparing workspace metrics between git refs. Delta tracking with regression/improvement classification. Haiku subagent for data collection.
severity-classification.md: Rule of Three severity system — 3+ occurrences → CRITICAL, 2 → WARNING, 1 → INFO. Temporal escalation (same WARNING × 3 sprints → auto-CRITICAL). Thresholds for PR quality, sprint health, context health, code quality.
Tests: test-scoring-intelligence.sh — 39 tests across scoring curves, score diff, severity classification, integration and cross-references.

2.2.0 — 2026-03-04

Added — Best Practices Audit & Documentation (Era 27)

CLAUDE-GUIDE.md: guide and template for project-level CLAUDE.md files (minimal ~50 lines, complete ~120)
estudio-equality-shield.md: full Equality Shield implementation study with academic references
External audit of claude-code-best-practice repo: confirmed existing coverage of 12/13 recommended features (context-map, agent-self-memory, intelligent-hooks, source-tracking, semantic-hub-index, confidence-protocol, consensus-protocol, context-aging, command-ux-feedback, skillssh-publishing, output-first, file-size-limit)

2.1.0 — 2026-03-04

Added — Equality Shield (Era 26)

equality-shield.md: anti-bias domain rule based on LLYC "Espejismo de Igualdad" (2026) study blocking 6 bias types
bias-check.md: /bias-check command for counterfactual bias auditing in sprints
política-igualdad.md: equality policy documentation with academic references (Dwivedi 2023, EMNLP 2025, RANLP 2025)
Rule #23 in CLAUDE.md: mandatory counterfactual test in assignments and communications
Tests: test-equality-shield.sh — 41 tests covering full framework validation

2.0.0 — 2026-03-04

Quality Validation Framework — Era 25. Multi-judge consensus, confidence calibration, and output coherence validation inspired by BullshitBench.

Added

Multi-Judge Consensus — 3-judge panel (reflection-validator, code-reviewer, business-analyst) with weighted scoring (0.4/0.3/0.3), verdicts (APPROVED/CONDITIONAL/REJECTED), veto rule for security/GDPR, dissent handling. Skill + rule + command /validate-consensus.
Confidence Calibration — Tracks NL-resolution success/failure in JSONL log, computes per-band accuracy and Brier score, decay mechanism (-5% for 3 pattern failures, -10% for 5 command failures, floor 30%), recovery (+3% per success). Script confidence-calibrate.sh + protocol rule.
Output Coherence Validator — coherence-validator agent (Sonnet 4.6) checks output↔objective alignment: coverage, internal consistency, completeness. Severity levels (ok/warning/critical). Skill + command /check-coherence.
98 new tests: test-consensus.sh (33) + test-confidence-calibration.sh (30) + test-coherence-validator.sh (35).

Changed

NL-command resolution — Added recalibration section with confidence logging and decay mechanism.
Agents catalog — Updated to 27 agents (added coherence-validator). Added consensus flow.
CLAUDE.md / READMEs — Updated agent count (26→27), skill count (23→25).

1.9.1 — 2026-03-04

Reflection Validator agent and skill — System 2 meta-cognitive validation protocol.

Added

reflection-validator agent (Opus 4.6): 5-step System 2 protocol — extracts real objective, audits assumptions, simulates causal chain, detects gaps, corrects transparently.
reflection-validation skill (SKILL.md, 148 lines): embeddable pattern for internal reflection, cognitive bias taxonomy, structured output format.
Agent memory (agent-memory/reflection-validator/MEMORY.md): persistent context for reflection sessions.
65 new tests (scripts/test-reflection-validator.sh): covers agent structure, skill protocol, memory, integration, and cognitive bias detection.

Changed

Agents catalog — Updated to 26 agents (added drift-auditor and reflection-validator).
CLAUDE.md / READMEs — Updated agent count (25→26) and skill count (22→23).

1.9.0 — 2026-03-04

Memory improvements inspired by claude-mem + Natural Language command resolution system.

Added

Concepts dimension in memory-store.sh: --concepts parameter stores CSV tags as JSON array for 2D taxonomy (type + concepts).
Token economics: every memory entry tracks tokens_est (content length / 4) for budget awareness.
Hybrid search: scored multi-field search (title 3x, concepts 2x, content 1x) with --type and --since filters, top-10 limit.
/memory-recall — Progressive disclosure in 3 layers: index (titles only), timeline (last N), detail (full entry).
/memory-stats — Dedicated stats command with type/concept breakdown and token estimates.
/memory-consolidate — Session consolidation: groups entries by concept, generates session-summary, deduplicates.
/savia-recall — Unified search across memory store, agent MEMORY.md files, and lessons.md.
memory-auto-capture.sh — PostToolUse async hook that auto-captures patterns from Edit/Write operations with 5-min rate limit.
Intent catalog (.claude/commands/references/intent-catalog.md): 60+ NL patterns mapped to commands across 19 categories, bilingual ES/EN.
NL resolution rule (docs/rules/domain/nl-command-resolution.md): automatic intent detection, confidence scoring (base + context + history), anti-improvisation guards.
/nl-query rewritten: loads intent catalog, scores confidence, resolves params from context, learns from successful mappings. Subcommands: --explain, --learn, --history.
32 new tests: test-memory-improvements.sh (13 tests) + test-nl-resolution.sh (19 tests).

Changed

memory-store.sh — Enhanced cmd_save() (concepts, tokens), cmd_search() (scored, filtered), cmd_stats() (concept breakdown). Fixed dedup logic.
README.md / README.en.md — Added new memory and NL commands to command catalog. Versión history updated.

1.8.0 — 2026-03-04

Usage guides by scenario + README restructure + documentation alignment.

Added

10 usage guides in docs/guides/: Azure DevOps consultancy, Jira consultancy, Savia standalone, Education (Savia School), Hardware lab, Research lab, Startup, Non-profit, Legal firm, Healthcare. Each guide includes roles, setup, day-to-day workflows, command sequences, and example conversations with Savia.
20 gap proposals identified during guide writing (hardware BOM, experiment tracking, grant lifecycle, legal deadlines, PDCA cycles, and more). Added to roadmap backlog.
Guides section in both README.md and README.en.md with links to all 10 guides.

Changed

README restructured: removed 3 scattered release note blocks, added clean "Versión History" table.
README.en.md aligned: added missing /excel-report, /savia-gallery, /vertical-* commands and /aepd-compliance + /governance-* to match Spanish versión.
CLAUDE.md compacted: 123→119 lines to pass CI gate (max: 120).
ROADMAP.md updated: added Era 22 (v1.6–v1.7) and Era 23 (v1.8 guides) with gap analysis table.

Fixed

README parity: English and Spanish READMEs now have identical feature coverage and command references.

1.7.0 — 2026-03-03

Company Savia v3: branch-based isolation with Git orphan branches + quality framework.

Added

savia-branch.sh: new abstraction layer for cross-branch read/write/list/exists/ensure-orphan/check-permission/fetch-messages via git show and temporary worktrees.
test-savia-branches.sh: 15 tests for branch abstraction layer.
Rule #21 — Self-Improvement Loop: persistent tasks/lessons.md reviewed at session start. Rule: docs/rules/domain/self-improvement.md.
Rule #22 — Verification Before Done: proof-based completion. Rule: docs/rules/domain/verification-before-done.md.
Agent Self-Memory: 10 agents with persistent MEMORY.md files (code-reviewer, architect, security-guardian, test-runner, triage, and 5 more). Rule: docs/rules/domain/agent-self-memory.md.
/drift-check command: audits CLAUDE.md rules vs repo state. Agent: drift-auditor.md.
hook-pii-gate.sh: pre-commit PII scanner (emails, phones, API keys, IBAN, DNI/NIE).
Frontend Component Rules: docs/rules/domain/frontend-components.md (naming, a11y checklist, states, design tokens).
Roadmap v1.7.0: archived (content integrated into docs/ROADMAP.md Era 22).

Changed

20 core scripts migrated: from directory-based to orphan branch isolation (main, user/{handle}, team/{name}, exchange).
8 test suites rewritten: 120 Savia tests pass (branch-based architecture).
Config, skills, docs updated: company-savia-config.md, SKILL.md, message-schema.md reflect branch architecture.
CLAUDE.md: 22 rules (was 20). New checklist entries for self-improvement and verification.

Fixed

git fetch origin --all: invalid command replaced with git fetch --all across all tests.
assert_ok pattern: fixed $? capture bug in test harnesses (was always 0).
Dispatcher command names: tests now use short names (read, write, exists) matching savia-branch.sh dispatcher.

1.6.0 — 2026-03-03

Company Savia v2: complete directory restructure for clarity, consistency, and indexing.

Changed

Directory layout: team/ → users/, company-inbox/ → company/inbox/, new teams/ directory with per-team member references.
User paths simplified: removed public/ subdirectory and savia- prefixes (savia-inbox/ → inbox/, savia-state/ → state/, savia-flow/ → flow/).
35+ files updated: all scripts, tests, config rules, skills, and docs aligned with new structure.

Added

inboxes.idx: new index mapping handle → inbox path for fast lookup.
teams.idx: new index mapping team → members.
teams/{name}/users/{handle}.md: per-team member reference files with role and join date.

Fixed

.gitignore: pubkey exclusion rule updated (!**/pubkey.pem instead of !**/public/*.pem).
Test company repo: reinitialized with new structure.

1.5.1 — 2026-03-03

Confidentiality hardening: E2E encryption testing, subject sensitivity validation, 7 bug fixes, 5 new test suites.

Added

5 test scripts: test-savia-confidentiality.sh (34 tests — E2E encryption, metadata, non-recipient rejection, privacy scanner, idempotency, subject sensitivity), test-savia-flow-tasks.sh (24 tests), test-savia-index.sh (12 tests), test-savia-travel.sh (18 tests), test-savia-school.sh (34 tests).
1 script: savia-messaging-privacy.sh — Subject sensitivity validation: detects monetary amounts, dates, company names, credentials, API keys, IPs, emails, DNI/NIE, IBAN in subjects. Warns but doesn't block delivery.
1 rule: messaging-subject-safety.md — Agent guidance for safe subject lines. "Instead of X, use Y" table. 12 pattern categories.
Company Savia initialization: Structure deployed to test repo via company-repo-templates.sh.

Fixed

savia-flow-tasks.sh: Multiline seq from ls|grep|echo pipeline; mkdir with braces inside quotes (no shell expansion).
savia-travel.sh: local keyword used outside functions in case blocks — refactored into proper functions.
savia-index.sh: Missing init dispatcher entry; update_entry shift bug (captured name before shift).
savia-school.sh: SCHOOL_ROOT used $1 (the command) as base path — replaced with SCHOOL_BASE env var.
savia-flow.sh: Missing do_sprint_start/do_sprint_close/do_metrics adapter functions.
savia-flow-sprint.sh: Case dispatcher executed when sourced — added BASH_SOURCE guard.
savia-messaging.sh: Integrated savia-messaging-privacy.sh and check_subject_sensitivity() call before send.

Changed

test-integration-company.sh: Runs 18 suites (197 tests total, all green). Accepts repo URL as parameter.

1.5.0 — 2026-03-03

Ecosystem Integration: research of 12+ Claude Code repos with actionable improvements for pm-workspace.

Added

2 research docs: investigacion-ecosistema-claude-code-2026.md (12 repos analyzed), era21-masterplan.md (7 workstreams planned).
12 improvement proposals: instincts system, adversarial security, skill evaluation engine, anti-rationalization hook, quality sweeps, deny rules, pass@k metrics, verify/fix loops, audit trail, AGENTS.md format, VoiceMode, event broker.

1.4.0 — 2026-03-03

Savia School: educational vertical for classrooms. Teachers tutor and evaluate, students create projects. GDPR/LOPD compliant.

Added

12 commands: /school-setup, /school-enroll, /school-project, /school-submit, /school-evaluate, /school-progress, /school-portfolio, /school-diary, /school-export, /school-forget, /school-analytics, /school-rubric.
2 scripts: savia-school.sh (classroom management), savia-school-security.sh (encryption, audit, content filtering, GDPR compliance).
1 rule: school-safety-config.md — Security config for school vertical (encryption, consent, isolation, content filtering).

Security

Alias-based enrollment (no PII in repository).
AES-256-CBC encrypted evaluations (teacher-only decryption).
GDPR Art. 8 (parental consent), Art. 15 (data export), Art. 17 (right to erasure).
Student folder isolation. Audit trail for all operations.

1.3.0 — 2026-03-03

Git Persistence Engine: TSV indexes for low-context lookups. ~60-80% token reduction per query.

Added

3 commands: /index-rebuild, /index-status, /index-compact — Manage TSV indexes.
2 scripts: savia-index.sh (core: lookup, update, remove, verify, compact), savia-index-rebuild.sh (rebuild profiles, messages, projects, specs, timesheets from source files).
6 index types: profiles.idx, messages.idx, projects.idx, tasks.idx, specs.idx, timesheets.idx.

1.2.0 — 2026-03-03

SDD/Tickets/Tasks Git-native: complete Savia Flow task management in Git folders. No database dependency.

Added

12 commands: /flow-task-create, /flow-task-move, /flow-task-assign, /flow-sprint-create, /flow-sprint-close, /flow-sprint-board, /flow-timesheet, /flow-timesheet-report, /flow-burndown, /flow-velocity, /flow-spec-create, /flow-backlog-groom.
3 scripts: savia-flow-tasks.sh (task CRUD + board), savia-flow-sprint.sh (sprint lifecycle + metrics), savia-flow-timesheet.sh (time tracking + reporting).
1 rule: flow-tasks-config.md — Configuration for Git-native flow system.

1.1.0 — 2026-03-03

Travel Mode extended: full pack/unpack/sync/verify/clean lifecycle for portable Savia on USB.

Added

5 commands: /travel-pack, /travel-unpack, /travel-sync, /travel-verify, /travel-clean.
3 scripts: savia-travel.sh (core dispatcher), savia-travel-ops.sh (advanced sync operations), savia-travel-init.sh (self-contained USB bootstrap).

Security

AES-256-CBC encryption for keys and PATs on USB.
SHA256 integrity checksums for all files.
Secure cleanup of traces from borrowed machines.

1.0.0 — 2026-03-03

Script Hardening: 6 critical + 7 medium fixes across 9 scripts. Cross-platform (macOS + Linux + WSL).

Fixed

backup.sh: Hash comparison bug (comparing plaintext vs SHA256), race condition in rotation (subshell pipe), cp -r without -p flag.
contribute.sh: Perl regex lookahead (?!) invalid in grep -E — corporate email detection was silently failing.
memory-store.sh: grep without -F allows regex injection via topic_key; newlines corrupt JSONL format.
pre-commit-review.sh: Cache invalidation on empty CACHE_DIR.
session-init.sh: Unquoted git branch variable.
update.sh: sed -i not portable on macOS — now uses portable_sed_i.
context-aging.sh: date -d doesn't exist on macOS — now detects OSTYPE.
validate-bash-global.sh: \s not POSIX ERE — replaced with [[:space:]].
block-force-push.sh: Pattern matching bypass via compound commands — added anchoring.

0.101.0 — 2026-03-03

Savia Flow: Git-based project management — PBIs, sprints, Kanban board, timesheets. No Azure DevOps dependency.

Added

5 commands: /savia-pbi, /savia-sprint, /savia-board, /savia-timesheet, /savia-team — Git-based PM lifecycle stored as markdown in company repo.
5 scripts: savia-flow.sh (dispatcher), savia-flow-ops.sh (PBI CRUD), savia-flow-sprint.sh (sprint lifecycle + metrics), savia-flow-board.sh (ASCII Kanban), savia-flow-templates.sh (project/team scaffolding).
1 test script: test-savia-flow.sh — 29 tests covering PBI create/assign/move, sprint start/close, log-time, board, metrics.
1 reference: flow-schemas.md — YAML schema specs for PBI, Sprint, Timesheet, Team.

Changed

company-repo-templates-init.sh — Added projects/ and teams/ dirs to repo init.

0.100.0 — 2026-03-03

Travel Mode: portable USB bootstrap with savia-init for deploying pm-workspace on new machines.

Added

2 commands: /savia-travel-pack, /savia-travel-init — Pack and bootstrap pm-workspace portably.
2 scripts: savia-travel.sh (pack), savia-travel-init.sh (bootstrap: OS detect, deps check, Claude Code install, workspace copy, profile restore).

0.99.2 — 2026-03-03

Integration tests against real Company Savia repo structure.

Added

1 test script: test-integration-company.sh — Orchestrates all 3 Company Savia test suites + smoke tests against cloned repo.

0.99.1 — 2026-03-03

Cross-platform compatibility: replace GNU-only patterns with portable helpers.

Added

1 script: savia-compat.sh — Portable helper library: portable_base64_encode, portable_base64_decode, portable_sed_i, portable_read_config, portable_yaml_field, portable_wc_l.

Fixed

7 scripts: Replaced base64 -w0, grep -oP, bare sed -i with portable helpers from savia-compat.sh. Affected: savia-crypto-ops.sh, savia-messaging.sh, savia-messaging-inbox.sh, company-repo.sh, company-repo-ops.sh, backup.sh, test-savia-messaging.sh.

0.99.0 — 2026-03-03

Company Savia: shared company repository with async messaging and E2E encryption.

Added

7 commands: /company-repo, /savia-send, /savia-inbox, /savia-reply, /savia-announce, /savia-directory, /savia-broadcast — Git-based company repo lifecycle and async messaging with @handle addressing.
4 scripts: company-repo.sh (repo lifecycle), savia-messaging.sh (message CRUD), savia-crypto.sh (RSA-4096 + AES-256-CBC encryption), privacy-check-company.sh (pre-push privacy filter).
1 script: company-repo-templates.sh — Heredoc templates for repo structure (CODEOWNERS, directory.md, org-chart, holidays, conventions).
1 skill: company-messaging — Knowledge module with message schema, encryption protocol, and privacy rules.
1 rule: company-savia-config.md — Configuration constants for repo, encryption, privacy, inbox, and messaging.
3 test scripts: test-company-repo.sh, test-savia-messaging.sh, test-savia-crypto.sh — Full test coverage for repo lifecycle, messaging round-trip, and encryption.
Session-init integration: unread inbox count displayed at startup (filesystem-only, no network).

0.98.0 — 2026-03-03

PR Guardian System — Automated PR validation with 8 quality gates + contextual digest.

Added

.github/workflows/pr-guardian.yml — 8-gate automated PR validation: description quality, conventional commits, CLAUDE.md context guard (≤120 lines), ShellCheck differential, Gitleaks secret scanning (700+ patterns), hook safety validator, context impact analysis, PR Digest (auto-comment in Spanish with risk assessment for maintainer).
.claude/commands/pr-digest.md — /pr-digest command for manual contextual PR analysis. Classifies changes by area, evaluates risk level, measures context impact, generates executive summary in Spanish.
.gitleaks.toml — Gitleaks configuration with allowlist for mock data, test fixtures, and placeholder patterns.
docs/propuestas/propuesta-pr-guardian-system.md — Full design document with gap analysis, 8-gate architecture, and implementation plan.
docs/propuestas/roadmap-research-era20.md — Era 20 research based on claude-code-best-practice analysis.

Changed

.github/pull_request_template.md — Added "Context impact" and "Hook safety" sections, conventional commits requirement.
docs/ROADMAP.md — Added Era 19 (Open Source Synergy) and Era 20 (Persistent Intelligence & Adaptive Workflows) with 6 milestones.

0.97.0 — 2026-03-03

Era 20 — MCP Toolkit & Async Hooks.

Added

/mcp-recommend — Curated MCP recommendations by stack and role (Context7, DeepWiki, Playwright, Excalidraw, Docker, Slack).
async-hooks-config.md — Hook classification (2 async, 10 blocking), event coverage 9/16 (56%), CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50.

0.96.0 — 2026-03-03

Era 20 — Adaptive Output & Onboarding.

Added

/onboard — Guided onboarding for new team members with role-specific checklists (dev/PM/QA). Auto-explore, component map, personalized Day 1/Week 1/Month 1 plan.
adaptive-output.md — Three output modes: Coaching (junior devs), Executive (stakeholders), Technical (senior engineers). Auto-detection from profile and command context.

0.95.0 — 2026-03-03

Era 20 — RPI Workflow Engine.

Added

/rpi-start — Research → Plan → Implement workflow with GO/NO-GO gates. Creates rpi/{feature}/ folder structure orchestrating product-discovery, pbi-decomposition, and spec-driven-development skills.
/rpi-status — Track progress of active RPI workflows with phase detection.

0.94.0 — 2026-03-03

Era 20 — Smart Command Frontmatter.

Added

smart-frontmatter.md — Domain rule defining model selection taxonomy (haiku/sonnet/opus), allowed-tools, context_cost, validation.

Changed

57 commands updated with model and context_cost frontmatter fields: 20 haiku, 29 sonnet, 8 opus.

0.93.0 — 2026-03-03

Era 20 — Savia Contextual Memory.

Added

/savia-recall — Query Savia's accumulated contextual memory (decisions, vocabulary, communication preferences).
/savia-forget — GDPR-compliant memory pruning implementing Art. 17 RGPD.
.claude/agent-memory/savia/MEMORY.md — Savia-specific persistent memory template.

0.92.0 — 2026-03-03

Era 20 — Agent Memory Foundation.

Added

.claude/agent-memory/ — Persistent memory directory with MEMORY.md templates for 9 agents (architect, security-guardian, commit-guardian, code-reviewer, business-analyst, sdd-spec-writer, test-runner, dotnet-developer, savia).
/agent-memory — Command to inspect and manage agent memory fragments (list, show, clear).
agent-memory-protocol.md — Domain rule defining three memory scopes (project, local, user), hygiene rules, and integration with existing systems.

0.91.0 — 2026-03-03

Era 20 — Stress Testing & Bug Fixes. 5 bug fixes + 165 new tests + orchestrator.

Fixed

block-credential-leak.sh — jq fallback: if jq not installed, secrets no longer pass through. Added grep-based extraction.
block-credential-leak.sh — Added missing Azure SAS token (sv=20), Google API key (AIza), and PEM private key detection patterns.
session-init.sh — ERR trap now exits 1 (not 0) and includes $LINENO for diagnostics.
agent-hook-premerge.sh — File line count uses awk 'END{print NR}' instead of wc -l (fixes off-by-one for files without trailing newline).
agent-hook-premerge.sh — Merge conflict markers now detected with \s* prefix (catches indented markers).
skillssh-adapter.sh — references: removal now uses awk frontmatter-aware parser instead of broad sed that matched comments.

Added

scripts/test-stress-hooks.sh — 25 stress tests for all 14 hooks under edge conditions (credential patterns, jq fallback, line counting, merge markers).
scripts/test-stress-security.sh — 27 tests covering SEC-1 through SEC-9 security patterns.
scripts/test-stress-scripts.sh — 21 tests for supporting scripts (skillssh-adapter, validate-commands, validate-ci-local, context-tracker, memory-store).
scripts/test-era18-commands.sh — 32 tests validating Era 18 command structure (frontmatter, line limits, content).
scripts/test-era18-rules.sh — 37 tests validating Era 18 rules (6 AI competencies, 4 AEPD phases, hook taxonomy, source tracking, skills.sh publishing).
scripts/test-era18-formulas.sh — 23 tests for scoring formula correctness (AI Competency boundaries, AEPD weights, banking detection weights).
scripts/test-stress-runner.sh — Orchestrator that runs all 9 test suites, aggregates counts, generates report in output/test-results/.

Changed

test-savia-e2e-harness.sh — Added Section 9: Era 18 Integration (6 tests).
Tests: 64→229 (+165 new tests across 7 scripts)

0.90.0 — 2026-03-03

Era 19 — Open Source Synergy (6/6). ERA 19 COMPLETA.

Added

/mcp-browse — Comando para explorar el catálogo de 66+ MCPs del ecosistema claude-code-templates (database, devtools, browser_automation, deepresearch, productivity).
/component-search — Búsqueda de componentes en el marketplace claude-code-templates (5.788+ components: agents, commands, hooks, MCPs, settings, skills).
docs/recommended-mcps.md — Catálogo curado de MCPs recomendados para equipos PM/Scrum con instrucciones de instalación y contexto de uso.
hooks/README.md — Documentación categorizada de los 14 hooks: seguridad (4), puertas de calidad (4), integración de agentes (3), flujo de desarrollo (3). Inspirado en la organización por categorías de claude-code-templates.
agent-observability-patterns.md — Regla de dominio con patrones de observabilidad inspirados en el analytics dashboard de claude-code-templates: detección de estado en tiempo real, caché multinivel, WebSocket live updates, monitorización de rendimiento.
component-marketplace.md — Regla de dominio que documenta la integración con el marketplace de componentes claude-code-templates (instalación, tipos de componentes, complementariedad).
Agradecimiento especial en README.md y README.en.md a claude-code-templates de Daniel Avila (21K+ stars) como referencia imprescindible para herramientas libres para Claude Code.
projects/claude-code-templates/ — Repositorio clonado para seguimiento de releases, análisis de sinergias y preparación de contribuciones bidireccionales.
SYNERGY-REPORT-PM-WORKSPACE.md — Informe completo de sinergias entre ambos proyectos con plan de contribución en 4 fases.

Changed

README.md / README.en.md — Añadida sección v0.90.0 con nuevos comandos y sección "Agradecimiento especial" con enlace a claude-code-templates.
Commands: 271→273 · Rules: 50→52

0.89.0 — 2026-03-03

Era 18 — Compliance, Distribution & Intelligent Hooks (6/6). ERA 18 COMPLETA.

Added

/aepd-compliance — Auditoría de cumplimiento AEPD para IA agéntica (framework 4 fases: tecnología → cumplimiento → vulnerabilidades → medidas). Scoring calibrado.
aepd-framework.md — Regla de dominio con el framework AEPD completo, mapping de controles pm-workspace, integración EU AI Act/NIST/ISO 42001.
framework-aepd-agentic.md — Marcadores de detección de proyectos agénticos y checklist de compliance.
skillssh-publishing.md — Especificación de formato para publicar en skills.sh marketplace (5 skills core mapeadas).
scripts/skillssh-adapter.sh — Script de conversión pm-workspace → skills.sh (package.json, README, LICENSE).
intelligent-hooks.md — Taxonomía de 3 tipos de hooks (Command/Prompt/Agent) con protocolo de calibración gradual.
hooks/prompt-hook-commit.sh — Hook semántico de validación de mensajes de commit (heurísticas, sin LLM).
hooks/agent-hook-premerge.sh — Quality gate pre-merge (secrets, TODOs, conflict markers, 150-line limit).
/excel-report — Generar plantillas Excel interactivas (capacity, CEO, time-tracking) en CSV multi-tab.
excel-templates.md — Estructuras CSV con fórmulas documentadas y reglas de validación.
/savia-gallery — Catálogo interactivo de 271 comandos por rol y vertical con source tracking.
source-tracking.md — Sistema de citación de fuentes (rule:/skill:/doc:/agent:/cmd:/ext:) con formatos inline/footer/compacto.
ai-competency-framework.md — 6 competencias AI-era (Problem Formulation, Output Evaluation, Context Engineering, AI Orchestration, Critical Thinking, Ethical Awareness) con 4 niveles cada una.

Changed

governance-audit.md — Añadidos 4 criterios AEPD (EIPD, base jurídica, scope guard, protocolo brechas).
governance-report.md — Añadido AEPD como framework soportado con score 4 fases.
regulatory-compliance/SKILL.md — Nueva referencia framework-aepd-agentic.md.
marketplace-publish.md — Añadido --target skillssh con referencia a adapter script.
settings.json — Registrados 2 nuevos hooks (prompt-hook-commit, agent-hook-premerge).
adoption-assess.md — Añadida opción --ai-skills con AI Competency radar (6 dimensiones).
Commands: 268→271 · Hooks: 12→14

0.83.0 — 2026-03-02

Safe Boot, Deterministic CI, PR Governance — Savia arranca siempre: MCP servers vacíos (conexión bajo demanda), session-init blindado (sin red, sin jq, timeout 5s). Mock engine determinista (cksum hash, 29/29 consistente). Hooks de gobernanza PR (bloqueo auto-aprobación y bypass branch protection).

Changed

mcp.json — Servidores vacíos. Savia conecta bajo demanda con /mcp-server start, no al arranque.
session-init.sh — v0.42.0: sin llamadas de red, sin dependencia jq, timeout global 5s, ERR trap para salida limpia garantizada. Context tracker en background.
engines.sh — Mock determinista: varianza con cksum hash (no $RANDOM), context overflow solo en límite real (200k tokens).
CLAUDE.md — 216→120 líneas: sección Savia duplicada eliminada, catálogo de comandos movido a referencia, regla 19 (arranque seguro).
validate-bash-global.sh — Nuevos bloqueos: gh pr review --approve (auto-aprobación) y gh pr merge --admin (bypass branch protection).
github-flow.md — Reglas explícitas: NUNCA auto-aprobar, NUNCA --admin.

0.82.0 — 2026-03-02

Auto-Compact — Compresión automática de contexto entre escenarios. Cuando el contexto acumulado supera un umbral configurable (default 40%), se ejecuta retro-summary --compact simulado que reduce 60-70% del contexto. Harness refactorizado en 3 ficheros (≤150 líneas cada uno).

Added

--auto-compact flag en harness.sh — activa compresión automática entre escenarios.
--compact-threshold=N — umbral configurable (% de ventana 200K) para disparar compactación.
engines.sh — Mock engine + live engine extraídos a fichero independiente.
report-gen.sh — Generador de reports extraído a fichero independiente.
Sección "Auto-Compaction Events" en el report cuando se activa.

Changed

harness.sh — Refactorizado de 269→150 líneas, ahora orquestador puro.
test-savia-e2e-harness.sh — 44 tests (vs 38), incluye test de auto-compact.

0.81.0 — 2026-03-02

AI Role Tooling — Dos nuevos comandos basados en gaps detectados en role-evolution-ai: /knowledge-prime (genera .priming/ con 7 secciones Fowler) y /savia-persona-tune (5 perfiles de tono/personalidad).

Added

/knowledge-prime — Genera .priming/ analizando código, packages, ADRs y git log. 7 secciones: architecture, stack, sources, structure, naming, examples, anti-patterns.
/savia-persona-tune — 5 perfiles (warm, technical, executive, mentor, minimal). Genera .savia-persona.yml.

Changed

CLAUDE.md, README.md, README.en.md — Command count 267→268.

0.80.0 — 2026-03-02

Context Optimization v2 — Mock engine realista calibrado por tipo de comando. State file para acumulación de contexto entre steps. Probabilidad de overflow crece con contexto acumulado (>80K: +10%, >120K: +20%).

Changed

harness.sh — Mock engine reescrito: rangos de tokens calibrados por comando, state file state.json, columna context_acc en CSV, sección "Context Accumulation" en report con umbrales 50%/70%.

0.79.1 — 2026-03-02

Role Evolution update — Reescrita role-evolution-ai.md con la taxonomía real de Kelman Celis (6 categorías: Estrategia, Ingeniería, Datos, Gobernanza, Interacción, Mantenimiento). Mapping equipo SocialApp a categorías Kelman. Gaps detectados → propuestas de mejora en roadmap.

Changed

role-evolution-ai.md — Reescrita completa: 6 categorías Kelman (vs genéricas previas), roles industria mapeados a Savia Flow, gaps detectados (RAG Engineer, Behavioral Trainer, AI UX Designer).
ROADMAP.md — Añadido "AI Role Tooling" en propuestas: /knowledge-prime, /savia-persona-tune, mock engine realista.

0.79.0 — 2026-03-02

CI + Multimodal Agent Prep — GitHub Action para E2E mock en PRs. Reference de agentes multimodales (VLM vision+text+code) con roadmap de integración para quality gates visuales.

Added

.github/workflows/savia-e2e.yml — CI workflow: E2E mock test en PRs que modifiquen flow-* o savia-test.
multimodal-agents.md — Reference: agentes VLM, tool-use, roadmap integración visual gates + spec from wireframe.

0.78.0 — 2026-03-02

Role Evolution — 6 categorías roles AI-era mapeadas a Savia Flow. Escenario stress test (10+ specs concurrentes).

Added

role-evolution-ai.md — 6 categorías (Orchestrator, Translator, Guardian, Builder, Context Engineer, Governance), mapping equipo, madurez L1-L4.
05-stress.md — Escenario stress: 10+ specs, intake masivo, board full-load, retro exhaustivo.

0.77.0 — 2026-03-02

Knowledge Priming (Fowler) — 5 patrones para reducir fricción AI. Estructura .priming/ por proyecto.

Added

knowledge-priming.md — 7 secciones priming, Design-First, Context Anchoring, Feedback Flywheel.

Changed

SKILL.md: +3 references (knowledge-priming, role-evolution-ai, multimodal-agents).

0.76.0 — 2026-03-02

Context Optimization — Correcciones del informe E2E v0.75.0. max_context budgets, --spec filter, escenario flow-protect.

Changed

flow-board/intake/metrics/spec.md — max_context en frontmatter para budget enforcement.
flow-intake.md — Nuevo --spec {ID} para intake individual.
03-coordination.md — Nuevo Step 5: flow-protect (WIP overload, deep work).
test-savia-e2e-harness.sh — Check flow-protect en escenario 03.

0.75.0 — 2026-03-02

Savia E2E Test Harness — Entorno Docker aislado con agente autónomo que ejecuta Claude Code headless contra pm-workspace. Simula 4 roles de equipo ejecutando 23 pasos en 5 escenarios (setup → exploration → production → coordination → release). Recopila métricas de tokens, tiempos, errores y bloqueos de contexto. Modo mock para CI, modo live con API key real.

Added

docker/savia-test/ — Test harness Docker: Dockerfile, docker-compose.yml, harness.sh orchestrator.
5 escenarios E2E — 00-setup (3 pasos), 01-exploration (5), 02-production (5), 03-coordination (5), 04-release (5). 23 pasos totales cubriendo todo el ciclo Savia Flow.
Motor mock — Simula respuestas con tokens aleatorios, 5% error rate (context overflow + timeout). Para CI sin API key.
Motor live — Ejecuta claude -p headless real. Captura tokens, duración, errores. Configurable via env vars.
Métricas CSV — scenario, step, role, command, tokens_in, tokens_out, duration_ms, status, error.
Informe automático — report.md generado al final con resumen, failures, errors, token totals.

0.74.0 — 2026-03-02

Savia Flow Practice — Implementación práctica de la metodología Savia Flow: configuración Azure DevOps dual-track, tablero exploración/producción, intake continuo, métricas de flujo y creación de specs. Ejemplo completo: SocialApp (Ionic + microservicios + RabbitMQ) con equipo de 4 personas.

Added

/flow-setup — Configurar Azure DevOps para Savia Flow: board dual-track (Exploration + Production), campos custom (Track, Outcome ID, Cycle Time), area paths. Modos: --plan (preview), --execute (aplicar), --validate (verificar).
/flow-board — Visualizar tablero dual-track: exploración a la izquierda, producción a la derecha. Alerta WIP limits excedidos. Filtros por track y persona.
/flow-intake — Intake continuo: mover items Spec-Ready a Production. Valida acceptance criteria, check capacidad, asigna a builder disponible.
/flow-metrics — Dashboard métricas de flujo: Cycle Time, Lead Time, Throughput, CFR. Métricas IA: spec-to-built time, handoff latency. Tendencias y comparativas.
/flow-spec — Crear spec ejecutable desde outcome de exploración. Genera stub con 5 secciones Savia Flow, crea User Story vinculada al Epic padre.
Skill savia-flow-practice/ — Guía práctica con 6 references: azure-devops-config, backlog-structure, task-template-sdd, meetings-cadence, dual-track-coordination, example-socialapp.

Changed

Command count: 262 → 267 (+5 comandos flow)
Skills: 20 → 21 (+savia-flow-practice)
Context-map: añadido grupo Savia Flow

0.73.0 — 2026-03-02

Vertical Banking — Herramientas especializadas para equipos de desarrollo en banca: validación BIAN + ArchiMate, pipelines Kafka/EDA, data governance (lineage, clasificación, GDPR), auditoría MLOps (model risk, XAI, scoring). Auto-detección de proyectos bancarios.

Added

/banking-detect — Auto-detección de proyecto bancario. 5 fases: entidades BIAN (Account, Settlement, KYC/AML), rutas API bancarias, deps (Kafka, Snowflake, MLflow), config (BIAN_, KAFKA_, SWIFT_*), documentación. Score ≥55% → confirmar.
/banking-bian — Validar arquitectura contra estándar BIAN. Mapeo microservicios a Service Domains (Payments, Settlement, Deposits, Lending, Risk). Diagrama ArchiMate en Mermaid. Detección de anti-patrones (God Service, Fragmented Domain).
/banking-eda-validate — Validar pipelines Kafka/MSK/AMQ: topologías, DLQ, schemas Avro/Protobuf, idempotencia, ordering guarantees. Evaluar patrones EDA: Saga, CQRS, Event Sourcing. Circuit breakers en settlement flows.
/banking-data-governance — Auditar data governance: lineage (BCBS 239), clasificación (PII/PCI/Confidencial), catálogo Snowflake/Iceberg, feature stores (batch + real-time). Validar GDPR/LOPD. Data mesh domain ownership.
/banking-mlops-audit — Auditar pipeline MLOps bancario: versionado, CI/CD/CT, drift detection, model registry. Explicabilidad (XAI/SHAP/LIME). Model risk management (SR 11-7). Scoring architectures (batch/streaming/event-driven). GenAI (RAG, embeddings).
Skill banking-architecture/ — Skill con 3 references: BIAN framework, EDA patterns banking, data governance banking.
Regla banking-detection.md — Regla de detección automática de proyectos bancarios con 5 fases y scoring.

Changed

Command count: 257 → 262 (+5 comandos banking)
Context-map: añadido grupo Banking
CLAUDE.md: añadida sección Banking Architecture

0.72.0 — 2026-03-02

Trace Intelligence — Búsqueda y análisis profundo de trazas distribuidas, investigación asistida de errores con root cause analysis, correlación multi-fuente de incidentes. Era 13 — Observability & Intelligence (2/2). ERA 13 COMPLETE!

Added

/trace-search {criterio} — Buscar y filtrar trazas en Grafana Tempo, Datadog APM, Azure App Insights, OpenTelemetry. Soporta búsqueda en lenguaje natural. Filtros: servicio, estado (error/slow), periodo temporal, código error, tipo de excepción, usuario. Resultados con paginación automática.
/trace-analyze {trace-id} — Análisis profundo de traza específica. Waterfall ASCII timeline, detección de cuellos de botella (span más lento), cadena de errores (origen y propagación), detección de anomalías vs baseline, mapa de dependencias de servicios, recomendaciones contextuales. Output adaptado por rol.
/error-investigate {descripción} — Investigación asistida de errores. Busca logs coincidentes, correlaciona trazas, analiza despliegues recientes, verifica métricas de infraestructura, identifica servicio origen, construye hipótesis de root cause, sugiere mitigación inmediata y preventiva.
/incident-correlate [--incident-id ID] — Correlación cruzada de métricas (Grafana, Datadog, App Insights), logs (Loki, Datadog, App Insights), trazas (Tempo, APM, Dependencies), despliegues (CI/CD), alertas previas y cambios de configuración. Genera timeline unificado, detecta cascading failures, cuantifica blast radius, draft de post-mortem automático.

Changed

Command count: 253 → 257 (+4 comandos trace intelligence)
Era 13 (Observability & Intelligence): COMPLETE! (2/2)

0.71.0 — 2026-03-02

Observability Core — Conexión a Grafana, Datadog, Azure App Insights, OpenTelemetry. Consultas en lenguaje natural a datos de observabilidad (PromQL, KQL, Datadog Query Language). Dashboards digeridos por rol (CEO, CTO, PM, Dev, QA, SRE). Health checks de fuentes. Era 13 — Observability & Intelligence (1/2).

Added

/obs-connect {platform} — Conectar Savia a Grafana, Datadog, App Insights, OpenTelemetry. Almacena credenciales cifradas (AES-256-CBC). Soporta múltiples instancias simultáneamente. Test de conexión automático.
/obs-query {pregunta} — Consultas en lenguaje natural a datos de observabilidad. Traduce automáticamente a PromQL (Grafana), KQL (App Insights), Datadog Query Language. Detecta anomalías vs baseline. Correlaciona con deployments.
/obs-dashboard [--role] — Dashboard digerido por rol. CEO: disponibilidad + SLA + costos. CTO: latencias por servicio + errors. PM: impacto en usuarios + features. Dev/SRE: detalles técnicos + logs/traces. QA: pre/post deploy comparisons.
/obs-status — Health check de todas las fuentes conectadas. Estado de conexión, última sincronización, volumen de datos, alertas activas, recomendaciones.

Changed

Command count: 249 → 253 (+4 comandos observabilidad)
Era 13 (Observability & Intelligence): iniciada (1/2)

0.70.0 — 2026-03-02

Multi-Tenant & Skills Marketplace — Workspaces aislados por departamento/equipo, marketplace interno de skills/playbooks, compartición de recursos con control de aprobación. Era 12 — Team Excellence & Enterprise (5/5). PLAN COMPLETADO: v0.54-v0.70 = 68 comandos en 17 versiones.

Added

/tenant-create — Crea workspace aislado por departamento con perfiles, roles, configuración de proyecto e herencia empresarial. Isolation levels: full (separado) o shared (datos separados, reglas comunes).
/tenant-share — Comparte recursos (playbooks, templates, skills, reglas) entre tenants con flujo de aprobación, versionado y prevención de config drift.
/marketplace-publish — Publica skills/playbooks al marketplace interno con metadatos, validación de calidad y sistema de ratings tipo Anthropic Skills.
/marketplace-install — Instala recursos del marketplace con resolución de dependencias, preview y rollback automático. Verificación de compatibilidad.

Changed

Command count: 249 → 253 (+4 comandos multi-tenant y marketplace)
Era 12 (Team Excellence & Enterprise): ahora completa (5/5 fases)

Plan Roadmap Completado

v0.54–v0.70: 17 versiones, 68 nuevos comandos estructurados en 4 eras:

Era 9 (v0.54–v0.57): Company Intelligence — 16 comandos
Era 10 (v0.58–v0.61): AI Governance — 17 comandos
Era 11 (v0.62–v0.65): Context Engineering 2.0 — 17 comandos
Era 12 (v0.66–v0.70): Team Excellence & Enterprise — 18 comandos

Total: 253 comandos en pm-workspace. Todos los comandos ≤150 líneas, con YAML frontmatter, warm Savia persona (female owl), contexto Spanish.

0.69.0 — 2026-03-02

Audit Trail & Compliance — Inmutable audit trail de todas las acciones de Savia con exportación para auditorías externas, búsqueda contextual y alertas de anomalías. Era 12 — Team Excellence & Enterprise (4/5).

Added

/audit-trail — Log inmutable de todas acciones: comandos ejecutados, recomendaciones, decisiones, archivos. Append-only. Cumple EU AI Act, ISO 42001, NIST AI RMF.
/audit-export — Exporta trail en JSON (SIEM), CSV (análisis), PDF (compliance). Incluye hash SHA-256 para verificación de integridad.
/audit-search — Búsqueda contextual por fecha, usuario, acción. NL search soportado. Regex patterns. Timeline visualization. Saved searches.
/audit-alert — Alertas automáticas por patrones anómalos: fuera de horario, comandos riesgo alto sin aprobación, volumen inusual, acceso a datos sensibles. Canales: Slack, email, dashboard.

Changed

Command count: 245 → 249 (+4 comandos auditoría)

0.68.0 — 2026-03-02

Accessibility & Inclusive Design — Auditoría WCAG 2.2, correcciones automáticas, reportes de conformidad, monitorización continua.

Added

/a11y-audit — Auditoría exhaustiva de accesibilidad WCAG 2.2 (AA/AAA) con detección de alt text, contraste, navegación por teclado, ARIA, focus management, jerarquía de encabezados
/a11y-fix — Correcciones automáticas con preview y verificación; covers alt text, ARIA attributes, focus traps, skip links, color contrast
/a11y-report — Reportes multi-formato: ejecutivo (score + gráficos), técnico (detalles + código), legal (VPAT/Section 508); tracking de tendencias
/a11y-monitor — Monitorización continua en CI/CD; bloquea deploys con regresiones de accesibilidad; digest semanal

Changed

Command count: 245 → 249 (+4 comandos accesibilidad)

0.67.0 — 2026-03-02

Team Wellbeing & Sustainability — Detección temprana de burnout, equilibrado de carga y ritmo sostenible.

Added

/burnout-radar — Detección de señales tempranas de burnout con mapa de calor por miembro
/workload-balance — Equilibrado objetivo de carga respetando especialidades
/sustainable-pace — Cálculo de ritmo sostenible basado en histórico y capacidad
/team-sentiment — Análisis de sentimiento del equipo con pulse surveys y tendencias

Enhanced

role-workflows.md — Aggregated wellbeing commands for SM/Flow Facilitator role
context-map.md — Added wellbeing group for Team Excellence domain

Changed

Command count: 237 → 241 (+4 wellbeing commands in Era 12)
Era 12 — Team Excellence & Enterprise (2/5 features)

0.66.0 — 2026-02-28

Advanced DX Metrics — Deep-work analysis, flow-state protection, developer experience profiling, and prevention-focused feedback loops.

Added

/dx-core4-survey — Adapted survey for Speed, Effectiveness, Quality, Impact dimensions
/flow-protect — Detect and protect deep-work sessions; block interruptions; suggest focus blocks
/deep-work-analyze — Analyze developer deep-work patterns; measure focus time and context switching
/prevention-metrics — Preventive metrics: friction points before they block; suggested workflow improvements

Changed

Command count: 241 → 245 (+4 DX metrics commands)

0.65.0 — 2026-02-28

Multi-Layer Caching — Cache strategy, warm operations, analytics, and selective invalidation for context optimization.

Added

/cache-strategy — Define multi-layer cache policy (system, session, command, query levels)
/cache-warm — Predictive pre-warming for next operations based on patterns
/cache-analytics — Dashboard of cache hit rates, latency improvements, and cost savings
/cache-invalidate — Selective invalidation after configuration changes; audit trail

Changed

Command count: 237 → 241 (+4 caching commands)

0.64.0 — 2026-03-02

Semantic Memory 2.0 — Four new memory intelligence commands for semantic compression, importance scoring, knowledge graphs, and intelligent pruning.

Added

/memory-compress — Semantic compression: reduce engrams by up to 80% while preserving fidelity via entity extraction, event summarization, decisión condensation, context deduplication
/memory-importance — Importance scoring: rank engrams by composite score (relevance × recency × frequency access). Identify high-value and low-value candidates
/memory-graph — Knowledge graph from engrams: build relational map of entities, events, decisions. Query connections, detect isolated memories, generate Mermaid visualization
/memory-prune — Intelligent pruning: archive low-importance memories, preserve critical ones. Reversible with restore. Never prunes decisión-log entries

Changed

Command count: 237 → 241 (+4 memory commands)

0.63.0 — 2026-03-02

Evolving Playbooks — Four new playbook commands for capturing and evolving repetitive workflows using ACE framework.

Added

/playbook-create — Create evolutionary playbooks for releases, onboarding, audits, deploys
/playbook-reflect — Post-execution reflection (ACE Reflector): analyze what worked, failed, improve
/playbook-evolve — Evolve playbooks with insights (Generator→Reflector→Curator cycle from ACE)
/playbook-library — Shareable library of mature playbooks across projects with effectiveness ratings

Changed

Command count: 233 → 237 (+4 playbook commands)

0.62.0 — 2026-03-02

Intelligent Context Loading — Four new context management commands for optimal token budgeting and lazy loading (Context Engineering 2.0).

Added

/context-budget — Token budget per session with optimization suggestions
/context-defer — Deferred loading system (85% token reduction)
/context-profile — Context consumption profiling (flame-graph style)
/context-compress — Semantic compression (80% reduction target)

Changed

Command count: 229 → 233 (+4 context commands)

0.61.0 — 2026-03-02

Vertical Compliance Extensions — Four new vertical-specific compliance commands for regulated sectors (healthcare, finance, legal, education).

Added

/vertical-healthcare — HIPAA, HL7 FHIR, FDA 21 CFR Part 11
/vertical-finance — SOX, Basel III, MiFID II, PCI DSS
/vertical-legal — GDPR, eDiscovery, contract lifecycle, legal hold
/vertical-education — FERPA, Section 508/WCAG, COPPA, LMS integration

Changed

Command count: 225 → 229 (+4 vertical compliance commands)

0.60.0 — 2026-03-02

Enterprise AI Governance — Four new governance commands based on NIST AI RMF, ISO/IEC 42001, and EU AI Act.

Added

/governance-policy — Define company AI policy, risk classification, approval matrix, audit trail
/governance-audit — Compliance audit against policy
/governance-report — Executive report mapped to frameworks
/governance-certify — Certification checklist and readiness scoring

Changed

Command count: 221 → 225 (+4 governance commands)

0.59.0 — 2026-03-02

AI Adoption Companion — Four new adoption commands for team maturity assessment, personalized learning paths, safe practice environments, and friction tracking.

Added

/adoption-assess — Evaluate team adoption maturity using ADKAR model
/adoption-plan — Personalized adoption plan by role with learning paths
/adoption-sandbox — Safe practice environment without risks
/adoption-track — Adoption metrics and friction point detection

Changed

Command count: 217 → 221 (+4 adoption commands)

0.58.0 — 2026-03-02

AI Safety & Human Oversight — Four new safety commands for supervision levels, confidence transparency, boundary definition, and incident tracking.

Added

/ai-safety-config — Configure supervision levels (inform/recommend/decide/execute)
/ai-confidence — Transparency dashboard showing confidence, reasoning, data used
/ai-boundary — Define explicit boundary matrix per role
/ai-incident — Record and analyze Savia incidents

Changed

Command count: 213 → 217 (+4 safety commands)

0.57.0 — 2026-03-02

Ceremony Intelligence — Four new commands for asynchronous standups, retro pattern analysis, ceremony health metrics, and smart agenda generation.

Added

/async-standup — Asynchronous standup collection and compilation
/retro-patterns — Pattern analysis from retrospectives
/ceremony-health — Health metrics for ceremonies
/meeting-agenda — Intelligent agenda generation

Changed

Command count: 209 → 213 (+4 ceremony commands)

0.56.0 — 2026-03-02

Intelligent Backlog Management — Four new commands for assisted grooming, smart prioritization (RICE/WSJF), outcome tracking, and conflict resolution.

Added

/backlog-groom — Detect obsolete, duplicate items without acceptance criteria
/backlog-prioritize — Automatic RICE/WSJF prioritization
/outcome-track — Post-release outcome tracking
/stakeholder-align — Conflict resolution with objective data

Changed

Command count: 205 → 209 (+4 backlog commands)

0.55.0 — 2026-03-02

OKR & Strategic Alignment — Four new commands for OKR definition, tracking, visualization, and strategic mapping.

Added

/okr-define — Define Objectives and Key Results linked to projects
/okr-track — Automatic OKR progress tracking
/okr-align — Visualize project→OKR→strategy alignment
/strategy-map — Strategic map with initiatives and dependencies

Changed

Command count: 201 → 205 (+4 strategy commands)

0.54.0 — 2026-03-02

Company Profile — Four new commands for enterprise onboarding and configuration.

Added

/company-setup — Conversational onboarding of enterprise profile
/company-edit — Edit company profile sections
/company-show — Display consolidated profile with gap detection
/company-vertical — Detect and configure vertical and regulations

Changed

Command count: 197 → 201 (+4 company setup commands)

0.53.0 — 2026-03-02

Multi-Platform Support — Three new commands for multi-platform integration.

Added

/jira-connect — Connect and sync with Jira Cloud
/github-projects — Integration with GitHub Projects v2
/platform-migrate — Assisted migration between platforms

Changed

/linear-sync — Rewritten with new format, webhooks, unified metrics

0.52.0 — 2026-03-02

Integration Hub — Four new commands for MCP server exposure, natural language queries, webhook configuration, and integration status.

Added

/mcp-server — Expose Savia tools as MCP server for other projects
/nl-query — Natural language queries without memorizing commands
/webhook-config — Configure webhooks for real-time event push
/integration-status — Dashboard of all integration health

Changed

Command count: 174 → 178 (+4 integration commands)

0.51.0 — 2026-03-02

AI-Powered Planning — Four new commands for intelligent sprint planning, risk prediction, meeting summarization, and capacity forecasting.

Added

/sprint-autoplan — Intelligent sprint planning from backlog and capacity
/risk-predict — Sprint risk prediction with early signals
/meeting-summarize — Transcription and action item extraction
/capacity-forecast — Medium-term capacity forecasting (3-6 sprints)

Changed

Command count: 170 → 174 (+4 planning commands)

0.50.0 — 2026-03-02

Cross-Project Intelligence — Four new commands for portfolio-level visibility and analysis.

Added

/portfolio-deps — Inter-project dependency graph with bottleneck detection
/backlog-patterns — Detect duplicates across projects
/org-metrics — Aggregated DORA metrics at organization level
/cross-project-search — Unified search across all portfolio projects

Changed

Command count: 166 → 170 (+4 cross-project commands)

0.49.0 — 2026-03-01

Product Owner Analytics — Four new commands providing strategic views for POs.

Added

/value-stream-map — Value stream mapping with bottleneck detection
/feature-impact — Feature impact on ROI and engagement
/stakeholder-report — Executive report for stakeholders
/release-readiness — Release readiness verification

Changed

Command count: 162 → 166 (+4 PO commands)

0.48.0 — 2026-03-01

Tech Lead Intelligence — Four new commands for technology health and team knowledge.

Added

/tech-radar — Technology stack mapping (adopt/trial/hold/retire)
/team-skills-matrix — Competency matrix with bus factor calculation
/arch-health — Architectural health scoring
/incident-postmortem — Blameless postmortem template

Changed

Command count: 158 → 162 (+4 tech lead commands)

0.47.0 — 2026-03-01

Developer Productivity — Four new commands for personal sprint view, deep focus, learning opportunities, and pattern catalog.

Added

/my-sprint — Personal sprint view (private, no comparisons)
/my-focus — Deep focus mode with context loading
/my-learning — Learning opportunity detection from commits
/code-patterns — Living pattern catalog from codebase

Changed

Command count: 154 → 158 (+4 developer commands)

0.46.0 — 2026-03-01

QA and Testing Toolkit — Four new commands for complete testing workflow.

Added

/qa-dashboard — Quality panel with coverage and test metrics
/qa-regression-plan — Regression test planning based on changes
/qa-bug-triage — Assisted bug triage with duplicate detection
/testplan-generate — Test plan generation from specs

Changed

Command count: 150 → 154 (+4 QA commands)

0.45.0 — 2026-03-01

Executive Reports for Leadership — Three new commands for C-level strategic views.

Added

/ceo-report — Multi-project executive report with traffic-light scoring
/ceo-alerts — Strategic alert panel for director-level decisions
/portfolio-overview — Bird's-eye portfolio view with dependencies

Changed

Command count: 147 → 150 (+3 CEO commands)

0.44.0 — 2026-03-01

Semantic Hub Topology — Agentexecution tracing, cost estimation, and efficiency metrics for subagent operations.

Added

/hub-audit — Topology audit revealing hubs, near-hubs, and dormant rules

Changed

Command count: 146 → 147 (+1 hub audit command)

0.43.0 — 2026-03-01

Context Aging and Verified Positioning — Semantic compression of old decisions using neuroscience-inspired aging.

Added

/context-age — Analyze and compress aged decisions
/context-benchmark — Verify optimal information positioning
scripts/context-aging.sh — Automation script

Changed

Command count: 144 → 146 (+2 context commands)

0.42.0 — 2026-03-01

Subagent Context Budget System — All 24 agents now have explicit max_context_tokens and output_max_tokens fields.

Changed

All 24 agent frontmatter files updated with context budgets (4 tiers)

0.41.0 — 2026-03-01

Session-Init Compression and CLAUDE.md Pre-compaction — 4-level priority system for session initialization.

Changed

session-init.sh — Rewritten with priority-based array system
CLAUDE.md — Pre-compacted from 154 → 125 lines (36% reduction)

0.40.0 — 2026-03-01

Role-Adaptive Daily Routines, Project Health Dashboard, and Context Usage Optimization.

Added

/daily-routine — Role-adaptive daily routine
/health-dashboard — Unified project health dashboard
/context-optimize — Context usage analysis with recommendations
scripts/context-tracker.sh — Lightweight context usage tracking

Changed

Command count: 141 → 144 (+3 context commands)

0.39.0 — 2026-03-01

Encrypted Cloud Backup System — AES-256-CBC encryption before cloud upload with auto-rotation.

Added

/backup — 5 subcommands for backup management
scripts/backup.sh — Full backup lifecycle automation

Changed

Command count: 140 → 141 (+1 backup command)

0.38.0 — 2026-03-01

Private Review Protocol — Maintainer workflow for reviewing community PRs and issues.

Added

/review-community — 5 subcommands for PR/issue review and release

Changed

Command count: 139 → 140 (+1 review command)

0.37.0 — 2026-03-01

Vertical Detection System — Detect non-software sectors and propose specialized extensions.

Added

/vertical-propose — Detect vertical or receive name and generate extensions

Changed

Command count: 138 → 139 (+1 vertical detection command)

0.36.0 — 2026-03-01

Community & Collaboration System — Privacy-first contribution system with credential validation.

Added

/contribute — Create PRs, propose ideas, report bugs
/feedback — Open issues with validation

Changed

Command count: 136 → 138 (+2 community commands)

0.35.0 — 2026-03-01

Savia — User Profiling System and Agent Mode. Introduce Savia identity with fragmented user profiles and agent mode support.

Added

/profile-setup — Savia's conversational onboarding
/profile-edit — Edit profile sections
/profile-switch — Switch between profiles
/profile-show — Display active profile

Changed

Command count: 131 → 135 (+4 profile commands)
~72 existing commands updated with profile loading

0.34.0 — 2026-02-28

Performance Audit Intelligence — Static analysis for code performance hotspots.

Added

/perf-audit — Static performance analysis
/perf-fix — Test-first optimization
/perf-report — Executive performance report

Changed

Command count: 129 → 131 (+3 performance commands)

0.33.3 — 2026-02-28

Azure DevOps project validation — Automated audit of project configuration.

Added

/devops-validate — Audit Azure DevOps project config

Changed

Command count: 128 → 129 (+1 DevOps command)

0.33.2 — 2026-02-28

Detection algorithm calibration after real-world testing across regulated sectors.

Changed

Detection algorithm: 4 phases → 5 phases
Confidence thresholds recalibrated

0.33.1 — 2026-02-28

Compliance commands improvements after real-world testing.

Fixed

Output file naming with date suffix
Scoring formula documentation
Dry-run vs actual execution indication

0.33.0 — 2026-02-28

Regulatory Compliance Intelligence — Automated sector detection and compliance scanning across 12 regulated industries.

Added

/compliance-scan — Automated compliance scanning
/compliance-fix — Auto-fix framework for violations
/compliance-report — Generate compliance report

Changed

Command count: 125 → 128 (+3 compliance commands)

0.32.3 — 2026-02-28

Multi-OS emergency mode — Support for Linux, macOS, and Windows.

0.32.2 — 2026-02-28

Fix Ollama download — Adapted to new tar.zst archive format.

0.32.1 — 2026-02-28

Emergency plan — Preventive pre-download of Ollama and LLM for offline installation.

0.32.0 — 2026-02-28

Emergency mode — Local LLM contingency plan with Ollama setup and offline operations.

Added

/emergency-mode — Manage emergency mode with local LLM

0.31.0 — 2026-02-28

Architecture intelligence — Pattern detection and recommendations across 16 languages.

Added

/arch-detect — Detect architecture pattern
/arch-suggest — Generate improvement suggestions
/arch-recommend — Recommend optimal pattern
/arch-fitness — Define and execute fitness functions
/arch-compare — Compare architecture patterns

0.30.0 — 2026-02-28

Technical debt intelligence — Automated analysis and prioritization.

Added

/debt-analyze — Automated debt discovery
/debt-prioritize — Prioritize by business impact
/debt-budget — Propose sprint debt budget

0.29.0 — 2026-02-28

AI governance and EU AI Act compliance — Model cards and risk assessment.

Added

/ai-model-card — Generate AI model cards
/ai-risk-assessment — Risk assessment per EU AI Act
/ai-audit-log — Chronological audit log from traces

0.28.0 — 2026-02-28

Developer Experience metrics — DX Core 4 surveys and automated dashboards.

Added

/dx-survey — Adapted DX Core 4 surveys
/dx-dashboard — Automated DX dashboard
/dx-recommendations — Friction point analysis

0.27.0 — 2026-02-28

Agent observability — Execution tracing, cost estimation, and efficiency metrics.

Added

/agent-trace — Dashboard of agent executions
/agent-cost — Cost estimation per agent
/agent-efficiency — Efficiency analysis

0.26.0 — 2026-02-28

Predictive analytics and flow metrics — Sprint forecasting with Monte Carlo simulation.

Added

/sprint-forecast — Predict sprint completion
/flow-metrics — Value stream dashboard
/velocity-trend — Velocity analysis

0.25.0 — 2026-02-28

Security hardening and community patterns — SAST audit, dependency scanning, and SBOM generation.

Added

/security-audit — SAST analysis against OWASP Top 10
/dependencies-audit — Vulnerability scanning
/sbom-generate — Generate SBOM
/credential-scan — Scan git history for leaked credentials
/epic-plan — Multi-sprint epic planning
/worktree-setup — Automate git worktree creation

Changed

Command count: 96 → 102 (+6 security commands)

0.24.0 — 2026-02-28

Permissions and CI/CD hardening — Plan-gate hook and CI validation steps.

Added

/validate-filesize — Check file size compliance
/validate-schema — Validate JSON schemas

Changed

Command count: 94 → 96 (+2 validation commands)

0.23.0 — 2026-02-28

Automated code review — Pre-commit review hook with SHA256 cache.

Added

/review-cache-stats — Show review cache statistics
/review-cache-clear — Clear review cache

Changed

Command count: 92 → 94 (+2 review commands)

0.22.0 — 2026-02-28

SDD workflow enhanced with Agent Teams Lite patterns.

Added

/spec-explore — Pre-spec exploration
/spec-design — Technical design phase
/spec-verify — Spec compliance matrix

Changed

Command count: 89 → 92 (+3 SDD commands)

0.21.0 — 2026-02-28

Persistent memory system inspired by Engram — JSONL-based memory with deduplication.

Added

/memory-save — Save memory with topic
/memory-search — Search memory store
/memory-context — Load context from memory

Changed

Command count: 86 → 89 (+3 memory commands)

0.20.1 — 2026-02-27

Fix developer_type format — Revert to hyphen format.

0.20.0 — 2026-02-27

Context optimization and 150-line discipline enforcement.

Changed

9 skills refactored with progressive disclosure
5 agents refactored with companion domain files
CLAUDE.md compacted from 195 → 130 lines

0.19.0 — 2026-02-27

Governance hardening — Scope guard hook and parallel session serialization rule.

Added

Scope Guard Hook for scope creep detection

Changed

/context-load expanded with ADR loading

0.18.0 — 2026-02-27

Multi-agent coordination — Agent-notes system, TDD gate hook, and ADR support.

Added

/security-review — Pre-implementation security review
/adr-create — Create Architecture Decisión Records
/agent-notes-archive — Archive completed agent-notes

Changed

SDD skill workflow expanded with security review and TDD gate

0.17.0 — 2026-02-27

Advanced agent capabilities and programmatic hooks system.

Changed

23 agents upgraded with advanced frontmatter
11 skills updated with context and agent fields
7 programmatic hooks added via settings.json

0.16.0 — 2026-02-27

Intelligent memory system — Path-specific auto-loading and auto memory.

Added

/memory-sync — Consolidate session insights
scripts/setup-memory.sh — Initialize memory structure

Changed

21 language files and 3 domain files now have path-specific rules

0.15.1 — 2026-02-27

Auto-compact post-command — Prevent context saturation.

Changed

Auto-compact protocol enforced after every command
7 commands freed from context-ux-feedback dependency

0.15.0 — 2026-02-27

Command naming fix — All commands renamed from colon to hyphen notation.

Fixed

All 106 unique command references renamed across 164 files

0.14.1 — 2026-02-27

Context optimization — Auto-loaded baseline reduced by 79%.

Changed

10 domain rules moved to on-demand loading
/help rewritten with separate setup and catalog modes

0.14.0 — 2026-02-27

Session persistence — Save/load rituals for persistent "second brain".

Added

/session-save — Capture decisions before clearing
decisión-log.md — Private cumulative decisión register

Changed

/context-load rewritten to load big picture

0.13.2 — 2026-02-27

Fix silent failures — Heavy commands now explicitly delegate to subagents.

Fixed

/project-audit silent failure fixed with subagent delegation

0.13.1 — 2026-02-27

Anti-improvisation — Commands strictly execute only what their spec defines.

Changed

/help rewritten with explicit stack detection

0.13.0 — 2026-02-27

Context health and operational resilience — Proactive context management.

Added

Context health rule with output-first pattern and compaction suggestions

Changed

Auto-loaded context reduced: 2,109 → 899 lines

0.12.0 — 2026-02-27

Context optimization — 58% reduction in auto-loaded context.

Changed

8 rules moved from auto-load to on-demand
Auto-loaded context reduced from 2,109 → 882 lines

0.11.0 — 2026-02-27

UX Feedback Standards — Consistent visual feedback for all commands.

Added

UX Feedback rule with mandatory standards for all commands

Changed

6 core commands updated with UX feedback pattern

0.10.0 — 2026-02-27

Infrastructure and tooling — GitHub Actions and MCP migration guide.

Added

GitHub Actions PR auto-labeling workflow
MCP migration guide for azdevops-queries functions

0.9.0 — 2026-02-27

Messaging & Voice Inbox — WhatsApp, Nextcloud Talk, and voice transcription.

Added

/notify-whatsapp — Send WhatsApp notifications
/whatsapp-search — Search WhatsApp messages
/notify-nctalk — Send Nextcloud Talk notifications
/nctalk-search — Search Nextcloud Talk messages
/inbox-check — Check and process new messages
/inbox-start — Start background inbox monitoring

Changed

Command count: 75 → 81 (+6 messaging commands)
Skills count: 12 → 13 (+voice-inbox)

0.8.0 — 2026-02-27

DevOps Extended — Azure DevOps Wiki, Test Plans, and security alerts.

Added

/wiki-publish — Publish to Azure DevOps Wiki
/wiki-sync — Bidirectional wiki sync
/testplan-status — Test Plans dashboard
/testplan-results — Detailed test run results
/security-alerts — Security alerts from Azure DevOps

Changed

Command count: 70 → 75 (+5 DevOps Extended commands)

0.7.0 — 2026-02-27

Project Onboarding Pipeline — 5-phase automated workflow.

Added

/project-audit — Phase 1: deep project audit
/project-release-plan — Phase 2: prioritized release plan
/project-assign — Phase 3: distribute work across team
/project-roadmap — Phase 4: visual roadmap
/project-kickoff — Phase 5: compile and notify

Changed

Command count: 65 → 70 (+5 onboarding commands)

0.6.0 — 2026-02-27

Legacy assessment and release notes — Backlog capture from unstructured sources.

Added

/legacy-assess — Legacy application assessment
/backlog-capture — Create PBIs from unstructured input
/sprint-release-notes — Auto-generate release notes

Changed

Command count: 62 → 65 (+3 legacy & capture commands)

0.5.0 — 2026-02-27

Governance foundations — Technical debt tracking and DORA metrics.

Added

/debt-track — Technical debt register
/kpi-dora — DORA metrics dashboard
/dependency-map — Cross-team/PBI dependency mapping
/retro-actions — Retrospective action tracking
/risk-log — Risk register

Changed

Command count: 57 → 62 (+5 governance commands)

0.4.0 — 2026-02-27

Connectors ecosystem and Azure DevOps MCP optimization.

Added

Connector integrations (12 commands)
Azure Pipelines (5 commands)
Azure Repos management (6 commands)

Changed

Command count: 46 → 57 (+11 new commands)
Skills count: 11 → 12 (+azure-pipelines)

0.3.0 — 2026-02-26

Multi-language support, multi-environment, and infrastructure as code.

Added

16 Language Packs with conventions, rules, and agents
12 new developer agents for different languages
7 new infrastructure commands
File size governance (max 150 lines per file)

Changed

Command count: 24 → 46
Skills count: 11 → 23
Agents count: 8 → 35

0.2.0 — 2026-02-26

Quality, discovery, and operations expansion.

Added

Product Discovery workflow (/pbi-jtbd, /pbi-prd)
Quality commands (/pr-review, /context-load, /changelog-update, /evaluate-repo)
product-discovery skill with JTBD and PRD templates
test-runner agent for post-commit testing

Changed

Command count: 19 → 24 (+6)
Skills count: 7 → 8
Agents count: 9 → 11

0.1.0 — 2026-03-01

Initial public release of PM-Workspace.

Added

Core workspace with CLAUDE.md and setup guide
Sprint management commands (4)
Reporting commands (6)
PBI decomposition commands (4)
Spec-Driven Development with skills and agents
Test project (sala-reservas)
Test suite (96 tests)
Documentation with methodology