pooyahayati/vibe-coding-skill
Risk-adaptive software engineering operating skill for reliable AI-assisted vibe coding, architecture, staged implementation, project intelligence, verification, security, and handoff.
Changelog
Technical release history. Entries describe their release; current operating and publication rules are in SKILL.md and the linked references. UPDATES.md is the shorter user-facing catalog.
Unreleased
No pending changes.
1.9.0 — Design decisions and accepted UI delivery
- R40: retain typed owner design choices, scoped authority and pending sample decisions through planning, assignment/return and successor-task handoff/resume. Preserve changed decision history, reject unsupported promotion and reuse of authority for changed choices, and keep legacy/mechanical work lightweight.
- R41: bind applicable design decisions to existing delivery criteria and rendered-review evidence. Pending or omitted decisions block accepted UI completion; provisional UI and functional-only outcomes remain explicitly scoped. Preserve existing approval/delegation and lightweight routes without another specialist, questionnaire or test quota.
1.8.1 — Trivy Git-checkout report compatibility
- Accept native Trivy filesystem scans labeled
repositoryfor local Git checkouts, alongsidefilesystem, only when the absolute report path matches the requested target. Preserve image identity/type, report validation and blocking findings; add focused regression coverage and synchronize the portable package.
1.8.0 — Bound acceptance and version-matched engineering
- R35: retain pre-change check definitions and require a current Head integrity disposition for changed/new acceptance checks, even after fresh passing execution. Equivalent test maintenance remains allowed; hashes do not prove assertion quality.
- R36: route version-sensitive decisions to bounded Context7 retrieval with official version-matched/local documentation fallback, scoped provenance reuse and no automatic host setup.
- R37: bind confirmed security outcomes to existing behavioral criteria/obligations and verification handoffs; scanner-only evidence cannot replace the protective behavior.
- R38: use existing decision/diff review to justify material new subsystems/dependencies against inspected alternatives, without a new workflow or per-helper document.
- R39: require representative conditions for capacity/recovery claims and bind explicitly selected relevant evidence context, preserving unrelated-context reuse.
- Task-contract schema 2 carries these conditional integrity/security/context controls and rejects unsupported fields; explicit legacy reconciliation preserves obligations. Receipt, completion-report and specialist-handoff versions remain unchanged.
1.7.1 — Consistent selection and bounded specialist maintenance
- R33: reuse current router inspection and risk evaluation in specialist preparation; return the same route for context loading, preserve explicit selections across entry points, and make uninspected direct selection conservative. Keep verified mechanical edits lightweight without a routing cache.
- R34: separate required preparation from periodic installed-skill inventory, retain failed-attempt scheduling independently of successful freshness, and add due-only maintenance at task boundaries. Keep required current-source checks and assignment-bound preparation validity independent of unrelated inventory warnings.
1.7.0 — Focused specialist participation and isolated assignments
- R31: distinguish inspected, preserved RTL/responsive context from changed domain behavior in mechanical specialist participation. Carry explicit domain effects through routing and planning; preserve formal assignment, uncertainty and material-risk safeguards.
- R32: retain exact managed Head/specialist resources for open assignments outside skill discovery, dispatch continuation through those resources and allow verified default updates for new tasks. Add explicit migration, resource-aware resume/handoff and bounded reference-based cleanup; preserve legacy guards, findings, history, current delivery controls and installation rollback protection.
1.6.1 — Specialist reassessment and verified delivery
-
H3: Verify the actual tag commit, channel/state and ZIP/checksum content before claiming publication. Recover matching drafts through explicit bounded retries, upload missing assets only, and preserve full same-revision CI and final-artifact native Trivy gates. Never move existing tags or replace published assets.
-
H2: Select regressions loaded through literal dynamic module names and propagate module-level helper dependencies. Keep unrelated classes/overview edits selective and fall back to full validation for unresolved loaders.
-
H1: Add explicit same-contract specialist reassessment with archived acceptance, preserved findings/authority and active instruction guards. Require fresh assigned receipts and a separate new Head decision without revising unrelated task evidence.
1.6.0 — Proportionate verification and focused CI
-
T5: Select CI checks by changed inputs: overview structure/links/mirror, instruction policy/routing, affected runtime regressions/platform checks, or the full baseline for version/CI controls and unknown impact. Remove duplicate feature-branch push runs and cancel only obsolete PR runs. Keep full same-revision publication evidence and native final-artifact Trivy. Release-readiness schema 2 and its collector inspect actual full-validation steps and the latest run attempt, exclude PR/fork evidence, and require recollection of legacy reports. Local implementation and hosted rollout status remain separate in ROADMAP.
-
T4: Retain bounded separate stdout/stderr from the original local check, with prefix/tail diagnostics, visible truncation, a returned local path and a pre-execution
--no-diagnosticsoption for sensitive output. Reuse the optional receipt log/digest without changing receipt schemas or trusting printed success. Keep timeout/tree cleanup and bounded inherited-pipe draining; incomplete diagnostic collection cannot claim success. Cover failure, verbose/binary output, suppression, cleanup, integrity and the extracted-package route. -
T3: Replace Tier 2/3 evidence-count/family quotas with required-outcome and retained-obligation coverage in declared and receipt-backed completion. Keep provenance/revision checks and explicitly block invalid required provenance. Report policy/scope without changing external formats or weakening retained checks. Centralize stage-based selection, independent-corroboration triggers and scoped reuse; remove configured-command run-all pressure and move optional E1/E2 procedures to a routed receipt guide. Preserve native Trivy publication, specialist acceptance and schema-3 migration protections.
-
T2: Separate reviewed mechanical use of current accepted specialist guidance from formal domain assignments; keep material decisions, uncertainty and retained commitments on the formal route. Add assignment-bound preparation reuse with local source/resource/control/scope revalidation, due-inventory invalidation and explicit cached-observation reporting. Preserve task-start upstream checks, S1/S2 acceptance, sole UI/UX ownership and active-instruction protection. Cover paired lightweight/material routes and preparation invalidation without live downloads, host installs or model evaluation.
-
T1: Resolve incidental risk words/paths only through a consistent Head-reviewed non-runtime scope assessment with complete facts and inspection evidence. Propagate the assessment through routing/planning without manufacturing behavioral integration work; retain explicit/incomplete-scope safeguards. Add optional retained risk facts and an explicit risk-only reconciliation that preserves outcomes, evidence and specialist commitments, protects known material risk, records history and invalidates old completion bindings. Cover paired harmless/material-risk cases and the installed capture/handoff/resume route. No release or local installation change.
1.5.0 — Verified workflow and integration reliability
-
A5: Separate read-only integration health/development from explicit publication scanning. Missing native Trivy no longer blocks safe development solely by tier; authoritative graph reliance requires freshness while source analysis remains available. Delegate publication to the existing native Trivy gate with a final target and private report, preserve unresolved security evidence in the completion gate, and document legacy/strict exits. Update callers and focused/portable regressions; no second scanner, release or installation change.
-
A4: Parse full WP-CLI stdout separately from bounded stderr/combined diagnostics. Explicitly reject oversized stdout, malformed plugin lists and empty/multiline version output; report command timeouts as structured CLI failures. Preserve exact-artifact install/upgrade and uninstall/data safeguards, with focused runtime and extracted-package regressions. No version or installation change.
-
A3: Bind npm license/repository evidence to the selected version and require matching version evidence for normalized-provider supplements. Replace license token matching with a bounded evaluator for explicit AND/OR/WITH policies; unresolved expressions require review, deny-list vetoes remain, and projects without a policy gain no default restrictions. Add deterministic registry, policy and extracted-package CLI regressions; no dependency, version or installation change.
-
A2: Preserve literal hidden/bracketed affected paths and governing instructions; reject unsafe paths and expose missing-input inspection gaps. Bind graph freshness to the actual source snapshot and output bytes, reject copy/generation drift, and keep previous graph evidence on provider/input failure. Legacy graph caches require refresh; task/receipt formats and release/installation state are unchanged.
-
A1: Keep checkout/receipt identity stable across remote changes, atomically retain task snapshots, and restore obligations or require explicit reconstruction after corruption. Add Head-controlled specialist reopening/retirement after retained contract revisions, preserving required findings and requiring renewed evidence/acceptance. Keep existing external formats and legacy workflows; no release or installation change.
-
Prepare the A1-A5 remediation roadmap and implementation criteria for seven confirmed deep-audit findings: retained-state/contract continuity, routing/graph input integrity, dependency decisions, WordPress command parsing and stage-aware integration controls. Implementation progress is tracked in
ROADMAP.md; release/installation state is unchanged. -
Streamline the README and current status/implementation pages; archive closed specifications and delivery evidence without removing history, restore documentation links and align public security guidance with the existing native Trivy publication gate. No runtime, specialist, release or installation change.
1.4.0 — Reliable routing, native release security and local runtime
-
Accept native Trivy's Windows path separators by binding filesystem reports to the canonical absolute host path. Keep other targets and altered immutable image identities rejected; add the actual platform reproduction to the existing release-gate tests.
-
Add a Head-owned, conditional product installation/runtime policy: prefer Compose for suitable self-hosted products, offer user-selected local Docker execution, retain native/WordPress delivery routes, and verify the documented clean-start/core-workflow and required data persistence. Route the policy through planning, build, verification and delivery; document safe local defaults, prerequisites and honest target-specific evidence. No new runtime module, specialist or fixed test suite.
-
Keep security native to the project during development and require installed native Trivy before publication. Add planning preflight and local release scans to the existing Trivy adapter, inspecting fresh target-bound JSON findings instead of treating exit zero or compatibility smoke as approval. Bind required scan evidence to final delivery; gate the Skill's own extracted ZIP before automated publication. Preserve ordinary offline installation, proportional checks and existing specialists; add no scanner ecosystem, model evaluation or host auto-installation.
-
Fix real-project runtime routing from Markdown prose/examples: share source-type marker gates across project and affected-file scans, ignore document path/directory signals, match complete extensions, and keep metadata descriptions out of executable platform/browser evidence. Read browser markers from embedded script bodies rather than page prose. Ignore clear excluded-domain mentions when selecting task packs and required specialists while retaining positive signals, actual source surfaces, explicit/structured selection, conservative risk floors and retained commitments. Add focused reproductions and an extracted-package router/planner route; preserve the existing portable layout and dependencies.
1.3.1 — Targeted reliability hardening
-
R4: Reconcile benchmark/scenario-output guidance with the implemented release gate: no model evaluation or provider key is required for ordinary use or publication, and former credentialed workflow/trigger claims are obsolete. Keep optional collection/aggregation tooling explicitly full-source and separately requested; remove unavailable runner commands from portable guidance, retain its included offline scorer, and distinguish declared policy scores from task contracts/receipts and product outcomes. Update related bootstrap/delivery references and portable documentation. No executable/schema/workflow changes, new tests, model runs, version bump or installation.
-
R3: Treat
.and equivalent project-relative forms consistently as project-root scope in execution-plan ownership, retained-contract containment and drift checks. Share scope normalization/coverage with the existing behavior-contract helper; block root/nested ownership across different writers while allowing disjoint scopes. Preserve literal bracket/hidden names, exact retained-contract fingerprints, legacy planner directory shorthand and material drift triggers; reject absolute/parent-escaping coverage. Add the confirmed bound-plan and drift regressions and an extracted-package CLI route. No version bump, new runtime module, dependency, model evaluation or broader path refactor. -
R2: Preserve platform markers in oversized sources using bounded prefix reads. Bound area-header discovery and invariant extraction as well as project/affected scans; prioritize affected headers and enforce actual byte budgets before reading. Expose truncated/unreadable/skipped inspection with bounded path samples, separate routing uncertainties, Head assessment guidance and plain-CLI warnings. Count real inspected input rather than silent empty placeholders; retain platform evidence gating and local routing. Add large-plugin, generic-PHP, unreadable and read-budget regressions plus an extracted-package routing check. No version bump, model evaluation or automatic risk escalation.
-
R1: Default fresh structured task state and
--new-taskto schema-3 receipt-backed completion; reserve schema 2 for retained in-flight legacy tasks, including older snapshots without a workflow field. Keep explicit legacy migration and prevent downgrade across capture/handoff/resume. Preserve required specialist assignment IDs against the retained contract before state replacement and plan validation; show commitments in handoffs and require recorded Head reconciliation for removal. Add focused regressions for both confirmed failures, positive compatibility/manual routes and the extracted portable runtime. No version bump, model evaluation or change to authorization.
1.3.0 — Local skill feedback and portable workflow reliability
-
Add an evidence-triggered Head-owned skill-feedback route for confirmed/suspected Vibe instruction, resource, tooling and composition defects. Prepare minimal English reports outside product source, deduplicate locally, preserve attribution and safe task progress, and invite optional user-reviewed email to hayatipooya@gmail.com. No automatic sending, telemetry, product-UI notice, model evaluation or bypass of required controls; include the reference in portable/install validation.
-
Complete D1 guidance for usable behavior-led handoffs, current schema-3 versus declared legacy completion, extracted-package evidence limits, risk-appropriate maintainer checks and stable/source installation choices. Preserve workflow/project-size tables, important links and full author attribution; correct stale implementation descriptions and the contributor newline typo. No runtime behavior, version bump, model evaluation or additional project gate.
-
Add I1 integration checks that build/extract the portable ZIP and exercise light, receipt-backed and cross-boundary specialist routes through its runtime entrypoints. Run them in the existing OS matrix; preserve legacy/manual evidence, scoped freshness and separate Head acceptance without adding runtime instructions, models or host dependencies.
-
Accept literal bracket-containing project paths in specialist Git baselines, assignments/returns, evidence inputs/exclusions, relative artifacts and receipt references; retain wildcard, traversal and link protections. Add regressions for dynamic/catch-all routes, unrelated assignment scopes, real execution receipts, missing/deleted files and stale acceptance.
-
Check bracket-named skill resources referenced in prose or inline code instead of silently skipping them. Preserve literal paths when relocating prose/commands and URL-encode only Markdown link destinations. Advance the package-adapter policy to revalidate cached packages under the corrected dependency extraction.
1.2.0 — Bound outcomes, receipts and specialist acceptance
-
Fix fresh Head preflight rejecting the valid 1.1.0 package because prose mentions the specialist-owned
vibe-head-contract.md. Require that contract only forhead-delegatedpackages. Share dependency extraction between validation and relocation: ignore illustrative/literal Markdown, retain executable command dependencies, check nested owner-relative resources and inline links with angle destinations/titles, and preserve fragments and URL-encoded paths during relocation. Advance the adapter fingerprint so prior cached validation is rechecked. Regressions cover the real portable Head package, missing delegated resources, cold/cached adoption, resource syntax and relocation containment without modifying the product or global installation. -
Implement S2 scoped assignments and separate Head acceptance against actual project changes, local receipts, required-finding history and producer/consumer contract decisions. Bind stage/source/task identities and delivery inputs, defer instruction replacement during active assignments, and prevent schema-3 Done from bypassing retained specialist acceptance. Analysis-stage acceptance stays historical; current-delivery stage checks remain fresh. Preserve tiny/legacy routes and existing authorization; no new model evaluation or blanket test gate.
-
Implement S1 instruction-compatibility assessments separately from specialist currency/package checks. Provide bounded resource diffs and an explicit Head decision for authority, platform, dependencies, verification and scope/authorization. Bind cached acceptance to observed source, installed resources and Head/domain controls; reject mismatches and unresolved conflicts. Only affected active delegation blocks; unchanged acceptance is reused without another assessment or model evaluation. Task-return acceptance is implemented by S2.
-
Implement E2 schema-3 completion against independently retained task contracts and local receipts. Resolve bounded task-local references/digests, validate identity/origin/obligations, recompute scoped input/artifact/context bindings and retain risk/evidence-family rules. Keep schema-2 reports explicitly declared evidence; retain receipt workflow activation through state/handoff/resume and reject downgrade. Support actual manual observations without invented execution; imported remote claims require a separate authorized resolver and remain unverified here.
-
Implement E1 local execution receipts with explicit argument-list commands, bounded timeouts, before/after scoped input and exact-artifact hashes, honest non-pass/unavailable results, external local storage and optional WordPress ZIP version binding. Do not retain process output or environment dumps; reject escaping/link paths and incomplete collection. Package the collector and exercise its actual behavior across supported operating systems. Schema-2 completion remains unchanged; receipt resolution is E2.
-
Implement B2 optional version-1 task-contract validation and propagation through execution-plan drafting/validation, local state, handoff and resume. Preserve stable criteria/behavior/evidence obligations and risk floors; compare edited plans/state against retained outcomes. Label schema-2 results as reported, invalidate changed source/contract context, and retain explicit Head reconciliation without treating it as permission or receipt proof. Legacy/light routes remain supported; package the new shared helper.
-
Implement B1's observable-behavior guidance: reuse acceptance criteria, express actor/action/result and only material failure/protected-state expectations, keep tiny tasks inline and persist durable/high-risk decisions in existing appropriate locations. Choose checks from those outcomes; schema-2 reports use existing descriptions and evidence links until B2/E2 implement structured propagation/receipts.
-
Specify P0's shared behavior, receipt, specialist assignment/return and Head-acceptance contracts, trust limits and explicit completion-schema migration. B2/E1/E2/S1/S2 implement the retained-contract, receipt and acceptance routes; schema-2 compatibility remains available for explicitly reported legacy evidence.
-
Link the design from relevant operating references and maintain the roadmap's implementation/merge status separately from release and installation.
1.1.0 — Lifecycle and scoped specialists
- Consolidate the Head workflow into Discover, Define, Plan, Design, Build, Verify, Review and Ship with explicit stage outputs and conditional specialist ownership.
- Register application security, interface-contract and systematic-debugging specialists alongside the sole UI/UX specialist; add multi-domain, stage-aware routing and domain-only authority contracts.
- Package only selected upstream skills and referenced approved shared resources, rewrite/validate resolved relative links, and inject a small Head-delegation contract without copying specialist methods.
- Expand the README with workflow ownership/activation, specialist responsibilities/permissions, and separate Graphify/Trivy purpose and use tables.
Earlier specialist composition
- Keep Vibe as the engineering Head and require the sole UI/UX specialist,
pooyahayati/UI-UX-Skill, for affected product interfaces across app, website, dashboard, and WordPress settings work. - Add offline specialist selection to context plans and a scoped, latest-source resolver/installer with daily inventory, provenance, protected local edits, atomic replacement/rollback, and Head reload handling.
- Keep specialist methods in the specialist; preserve Head scope, architecture, risk, acceptance, and final delivery controls. No fixed specialist revisions or model evaluations.
1.0.0 — Reviewed release and software-based publication checks
- Remove model-evaluation requirements and automatic credentialed workflows from publication. Stable releases retain the complete same-commit software validation baseline.
- Add a reproducible, installation-checked portable ZIP and SHA-256 release assets.
- Preserve unresolved risk facts, normalize bounded aliases, recognize the reported Persian destructive request, and carry structured risk into execution planning.
- Keep custom sibling projects and WordPress plugins distinct; accept ordinary runtime facts without requiring a specialist pack.
- Require a required completion outcome and optionally bind completion to independently retained approved criteria.
- Execute WordPress benchmark behavior from the exact ZIP with a trusted PHP harness; reject comment-only and unsafe-output implementations. Report real install preconditions without claiming database freshness.
- Retain offline maintainer regressions for delivery-envelope validation, recomputed success, scenario/check identity, and simulated-evidence classification. Maintainer research utilities do not qualify releases and are excluded from the portable runtime.
- Preserve earlier maintainer CLI adapter fixes for option placement, sandboxed shell checks, and parser/version/isolation preflight; no model campaign runs automatically.
- Read selected-release dependency metadata, isolate offline CLI stubs on Windows/POSIX, retain immutable-baseline change evidence after agent commits/renames, normalize fixture line endings, and let compatibility CI reach the supported fallback resolver.
0.10.7 — Real-delivery benchmark framework and stable dual-evidence gate
- Added the Real Delivery Benchmark as a second evidence class alongside the existing behavior-selection benchmark.
- Added control/treatment benchmark arms so Skill impact is measured within the same Agent rather than by ranking Agents against one another.
- Added provenance-bound delivery result envelopes with fixture, grader, catalog, schema, Skill-tree, baseline-commit, final-tree, diff, changed-path, dependency-change, and raw-output evidence.
- Added hidden graders outside the Agent workspace and made grader execution operate on immutable snapshots so grader-side artifact generation cannot contaminate captured Agent changes.
- Added five representative deterministic delivery scenarios: tiny local copy fix, brownfield duplicate-save regression, contained CSV export feature, mixed-monorepo API-only normalization with sibling protection, and a WordPress installable-artifact change.
- Added deterministic tests proving each representative broken baseline fails and each known-good implementation passes its hidden grader.
- Added real Codex and Claude Code delivery adapters with bounded workspace editing, provider credential redaction, credential-leak detection, timeout/error evidence, and strict missing-run handling.
- Added a credentialed Real Delivery Benchmark GitHub Actions workflow and exact-SHA benchmark trigger path.
- Hardened the stable release gate so stable releases require both complete real-agent behavior evidence and complete real-delivery evidence for the current Skill/catalog/schema identity, with no missing/duplicate/invalid delivery runs and no treatment regressions.
- Aligned GitHub Release metadata with internal release governance: pre-1.0/RC releases are now published as prereleases.
- Attempted the real behavior and real-delivery benchmarks on the exact Phase 9D main commit. Both provider credential preflights reported no configured non-interactive credentials, so no real-agent success is claimed and the stable gate remains correctly blocked until credentials are configured.
0.10.6 — Language-independent routing and semantic completion evidence
- Added config-driven structured context facts for runtime, platform, capability, and concern routing so known semantics no longer depend on natural-language keywords.
- Added explicit Context Router CLI inputs for structured runtime/platform/capability/concern facts and rejects unsupported fact values instead of silently ignoring them.
- Propagated structured context facts into execution planning so routing and planning operate from the same semantic inputs.
- Added regression coverage proving equivalent English/Persian greenfield WordPress/WooCommerce payment tasks select the same capability packs and workflow floor when supplied the same structured facts.
- Replaced free-form completion evidence diversity with a controlled semantic evidence taxonomy and evidence families.
- Made Tier 2/3 evidence diversity count semantic families rather than raw labels, preventing multiple verification-test labels from manufacturing independent evidence.
- Added regressions that reject unknown evidence kinds and block higher-risk completion when all passing evidence belongs to only one semantic family.
- Kept canonical and portable routing, completion, configuration, references, and Skill contracts synchronized.
0.10.5 — Path-aware context routing for large and mixed repositories
- Made capability routing project-area aware so platform evidence stays local to the affected application/package in common monorepo layouts such as
apps/,packages/,services/,plugins/,themes/, and scoped WordPress plugin/theme directories. - Switched known-path routing to bounded path-scoped source scanning, added directory-level affected-path support, and kept the wider project scan only as the fallback when impact paths are still unknown.
- Refined change-scope detection so sibling applications under the same top-level container can correctly become
cross-boundaryand trigger execution planning without making repository size itself a planning requirement. - Made repository-root and subtree-applicable
AGENTS.mdfiles explicit persistent instructions while excluding unrelated sibling agent instructions and preserving the standardREADME.mdfallback context. - Expanded context observability with scan strategy/areas plus separate Skill core, selected reference, selected pack, persistent-context upper-bound, and combined context upper-bound byte metrics.
- Added regressions for same-container mixed monorepos, directory-only paths, local platform evidence after more than 700 unrelated files, scoped
AGENTS.md, and sibling-application execution planning.
0.10.4 — Release orchestration reliability
- Fixed release orchestration for GitHub Actions eventual consistency by trusting the successful triggering
workflow_runevent when the runs API still reports that same run with a temporary null conclusion; release-gate workflow changes now retrigger the full path-filtered baseline. - Hardened live dependency contracts so rate-limited/unavailable optional GitHub source-health evidence passes only when the dependency decision explicitly degrades to
REVIEW REQUIREDwith arepository.health_unavailablesignal; falseACCEPTremains a contract failure.
0.10.3 — Independent review remediation and delivery hardening
- Added structured risk facts for operation, environment, data sensitivity, and change boundary, with supplemental English/Persian text signals and safer destructive-operation matching.
- Centralized tier policy generation so router escalation rebuilds the final tier, approval requirement, reasons, and required controls from one source of truth.
- Upgraded completion reports to schema v2 with explicit criterion IDs/descriptions/required status/evidence links and required evidence that blocks completion when it fails.
- Converted the toolchain resolver tests to discoverable
unittestcases and added regressions for package imports, timeout fallback, and portable Graphify version lookup. - Made Graphify compatibility version detection work from the portable Skill without a root
VERSIONfile. - Made toolchain compatibility timeouts structured failures so the resolver can attempt the verified last-known-good fallback, and switched child execution to the active Python interpreter.
- Gated WordPress REST routing on actual WordPress platform evidence so generic REST/API terminology does not inject WordPress/PHP context into unrelated backends.
- Added explicit task change-scope reporting and changed execution-plan requirements to depend on Tier 2/3 risk or demonstrated cross-boundary scope rather than repository size alone.
- Kept large-repository project intelligence active while allowing local low-risk work to stay on the light route.
- Reduced the canonical/portable
SKILL.mdentry point from roughly 480 lines to a 215-line routed core and added a 260-line validation guard against maintainer-detail creep. - Added explicit execution-scoped toolchain sessions so compound operations resolve each managed tool once and reuse the exact selected version without a process-global cache.
- Made Trivy filesystem execution target-aware: container fallback now bind-mounts the host target read-only and generic Trivy command construction rejects ambiguous unmounted
fsscans. - Allowed Graphify and Trivy compatibility contracts to use an exact-version native executable when available instead of requiring
uvxor Docker unconditionally. - Added regressions for same-session version reuse, new-session re-resolution, Windows/space-containing bind mounts, native compatibility runtimes, and explicit filesystem command routing.
- Added a bounded technology-selection contract: existing capability/stack first, at most 2–3 realistic options when a material choice remains, constraint-based selection, official support/version verification when decision-critical, and explicit rejection of the nearest alternative.
- Added a stage interaction contract with concrete inputs, outputs, exit criteria, readiness status, limitations, evidence, and next-step reporting while keeping routine implementation details autonomous.
- Replaced test-count/coverage-quantity thinking with behavior/failure/boundary-based selection, deduplication against existing coverage, stage-specific execution, and an explicit verification stopping rule.
- Added behavior-contract eval scenarios for existing-stack technology decisions and proportionate bug verification, including regressions against rewrite-by-default, universal language rankings, fixed coverage quotas, duplicate tests, and unrelated E2E repetition.
- Added a WordPress delivery workflow that distinguishes discovery/bootstrap/implementation/security/lifecycle/UI/WooCommerce/delivery stages and requires verification against the exact installable ZIP for release/distribution work.
- Added
wordpress_artifact.pyto build deterministic single-root plugin ZIPs, verify embedded slug/version and SHA-256, perform safe install-shape smoke checks, and drive WP-CLI install/upgrade/deactivate/reactivate checks against exact artifacts. - Added guarded uninstall execution for disposable environments and explicit project-specific data-retention verification; deactivation is never treated as uninstall.
- Added a real WordPress CI contract that builds two fixture releases, installs the first exact ZIP, upgrades to the second exact ZIP, and verifies activation/version behavior in a disposable WordPress runtime.
- Added observable code-structure invariants for clear responsibility, data/boundary ownership, entry validation, diagnosable failures, useful separation of decision logic from side effects, evidence-before-abstraction, and minimum safe refactoring in existing codebases.
- Made advanced structure concern-triggered: external integrations require timeout/failure/retry effects, multi-step data changes require transaction/consistency boundaries, multitenancy requires ownership/isolation, and performance structure requires measurement before optimization.
- Explicitly rejected interface-per-class, repository-per-table, service-per-function, DTO/layer ceremony, message/event buses, microservices, and Clean/DDD/layered architecture when no current requirement justifies them.
- Added mandatory behavior-contract eval coverage for small local structure, external integration structure, and multidomain module boundaries, including anti-overengineering regressions.
- Added proportional Plugin Check guidance, conditional HPOS/Cart-Checkout Blocks checks, and an explicit Composer/Packagist fallback because the bundled dependency guard does not claim Composer coverage.
- Hardened release readiness so RC/pre-1.0 releases require Validate Skill, Cross Platform Smoke, Real World Repository Validation, Agent Skills Spec Compatibility, Tool Contract Tests, and WordPress Artifact Contract on the exact target commit.
- Consolidated PR/release tool validation into Tool Contract Tests while keeping dedicated Graphify/Trivy workflows for scheduled/manual compatibility monitoring; removed the duplicate Live Integration Contracts workflow.
- Added release-trigger guarantees for path-filtered validation workflows, orphan-reference detection, and broader merged-branch cleanup for refactor/docs/release work branches.
- Corrected stale validation documentation so real-world routing coverage and bounded conservative eval escalation match current implementation.
0.10.2 — Latest-compatible-stable toolchain runtime binding
- Replaced normal operating-version pins for Graphify and Trivy with latest-compatible-stable resolution.
- Added a shared resolver that tests the latest published stable candidate and falls back to a verified last-known-good version when compatibility fails.
- Bound runtime execution to the exact resolved version instead of silently invoking whichever local executable happens to be installed.
- Added exact Graphify execution through a matching local binary or an isolated
uvx --from graphifyy==<version>runtime, and exact Trivy execution through a matching local binary or versioned official container image. - Added execution-scoped session pins so a resolved tool version remains stable for the current run without re-resolving mid-execution.
- Removed misleading integration-guard claims that merely detecting an installed executable meant the runtime version was pinned.
- Packaged the resolver, runtime adapter, and Graphify/Trivy compatibility contracts in the portable skill and extended install/sync validation so portable installations cannot omit these runtime dependencies.
- Removed fixed current tool versions from normal user-facing documentation while retaining machine-managed last-known-good fallback state and historical release evidence.
- Expanded CI coverage for latest-compatible-stable resolution, fallback behavior, exact runtime binding, portable installation, cross-platform smoke tests, and live Graphify/Trivy compatibility contracts.
- Kept the credentialed Codex/Claude Code benchmark deferred in Issue #12; deterministic, cross-platform, tool-contract, and pinned real-repository validation remain the active pre-1.0 release baseline.
0.10.1 — Real-world routing validation hardening
- Added pinned real-world routing validation against a large Django repository and the official WooCommerce Stripe gateway.
- Extended real-world validation to execute the Context Router and Execution Plan and record project complexity, selected capability packs, pack-reduction ratio, selected context size, integration points, project-intelligence preservation, and execution-plan requirements.
- Added selectivity regressions showing that small WooCommerce admin UI work avoids payment/security/external-HTTP concern packs while payment webhook/refund work composes the sensitive packs and escalates to Tier 3.
- Expanded CI triggers so routing configuration, capability packs, execution planning, and project-intelligence changes automatically run the real-world validation suite.
- Kept product repositories clean while validating pinned external repositories at exact commits.
- Deferred the credentialed blind Codex and Claude Code benchmark to future work in Issue #12; it is not a blocker for the current pre-1.0 release path.
0.10.0 — Composable context routing and execution coordination
- Added an evidence-based Context Router that keeps project complexity separate from task risk and selects minimum sufficient context instead of loading broad references by default.
- Added composable Capability Packs for PHP, WordPress, WooCommerce, browser JavaScript, WordPress REST, external HTTP, payments, web security, and web performance.
- Kept Vibe Core authoritative for scope, risk, project intelligence, evidence, recovery, and delivery; domain packs only add relevant constraints, risks, integration points, and checks.
- Added interaction rules so overlapping concerns can raise the workflow floor without making small UI/copy tasks inherit unnecessary payment/security ceremony.
- Added Context Plan metrics for selected packs/references, approximate context bytes, integration points, and preserved invariant/risk/project-intelligence coverage.
- Added lightweight Execution Planning for medium/large projects, Tier 2/3 work, or multi-boundary tasks with explicit workstreams, dependencies, integration ownership, completion levels, and plan-drift approval triggers.
- Kept one implementation agent as the default and required explicit ownership/stable contracts before parallelization.
- Changed the existing-project startup protocol to route context before broad architecture/domain loading while preserving project-wide invariants and large-project impact analysis.
- Added WordPress/WooCommerce constraints for public platform APIs, HPOS/modern compatibility surfaces, plugin/theme interoperability, authorization/input/output boundaries, request/query cost, remote-call failure modes, and payment-state integrity.
- Added deterministic regressions for tiny WooCommerce UI tasks, overlapping WooCommerce payment flows, WordPress REST plus external APIs, browser-JS overlap, large generic repositories, execution-plan dependency validation, and plan drift.
- Added portable packaging and installation/structure validation for routing config, packs, router, planner, and routing reference documentation.
0.9.3 — Installation completeness hardening
- Expanded the offline installation check to require the complete portable runtime script set, including
release_readiness.py. - Required every reference document directly used by the Skill to be present before installation can pass.
- Added checks for runtime configuration, eval catalog/schema, Agent metadata, and project templates.
- Added JSON parse validation for runtime config/eval files.
- Added regression coverage so incomplete portable installations cannot silently pass installation validation.
0.9.2 — Stable benchmark orchestration hotfix
- Re-triggered validated release evaluation when the manual Real Agent Benchmark completes successfully, so stable readiness can advance after benchmark evidence arrives without requiring a redundant base-CI rerun.
- Bound stable readiness to the exact current eval scenario ID list rather than accepting an arbitrary positive scenario count.
- Required each stable Agent row to report expected, completed, and passed scenario counts equal to the current catalog size.
- Added regression coverage for benchmark-triggered release orchestration, scenario-set mismatch, and per-Agent scenario-count mismatch.
0.9.1 — Evidence correctness and release provenance hotfix
- Rejected malformed acceptance-criteria entries instead of silently ignoring non-object or non-boolean completion claims.
- Required Tier 2/3 completion reports to declare a target commit and excluded passing evidence bound to a different revision.
- Tightened Tier 3 evidence timestamps to require timezone-aware ISO-8601 values.
- Hardened benchmark aggregation so complete evidence requires valid runner envelopes, matching Agent/scenario provenance, and exactly one consistent Skill/version/hash identity set.
- Prevented unexpected per-scenario runner exceptions from aborting the remaining benchmark scenarios; internal failures are now recorded as failing raw evidence and execution continues.
- Changed release readiness to use the latest check result for each required workflow, so an older success cannot mask a newer failure on the same commit.
- Changed stable benchmark resolution to consider the latest Real Agent Benchmark run instead of falling back to an older successful run when a newer run failed.
- Added regression coverage for malformed completion claims, revision mismatch, ambiguous timestamps, benchmark envelope identity, runner exception isolation, and latest-run release semantics.
0.9.0 — Adaptive evidence and release readiness
- Hardened the real-agent benchmark workflow so Codex and Claude Code failure paths are isolated, aggregation still runs, raw evidence is preserved, and aggregate failures cannot be masked by
tee. - Added bounded benchmark tier policies that distinguish preferred behavior, legitimate conservative escalation, under-classification, and over-engineering.
- Added per-scenario tier ceilings with explicit rationale where conservative escalation is permitted.
- Added risk-aware completion evidence provenance: higher-risk Done claims require stronger, revision-bound, and timestamped evidence.
- Added deterministic beta, RC, and stable release-readiness gates.
- Required stable releases to have complete, fully conformant real Codex and Claude Code benchmark evidence.
- Bound stable benchmark evidence to the current Skill version, portable Skill tree, eval catalog, and output-schema hashes.
- Wired the validated release workflow to the release-readiness gate while keeping repository branch protection outside the release criteria by project decision.
- Extended regression, failure-injection, portable-package, Agent Skills compatibility, and cross-platform coverage for the new policies.
- Kept missing real-agent runs as missing evidence rather than treating them as success; the remaining real-agent evidence requirement stays tracked separately for 1.0 readiness.
0.8.0 — Real-agent benchmark execution and evidence hardening
- Added strict machine JSON Schema for live-agent behavior contracts.
- Added real Codex CLI and Claude Code benchmark adapters with blind prompt generation.
- Added provenance envelopes containing agent version, model override, Skill version, timing, hashes, and workspace-integrity evidence.
- Added raw stdout/stderr preservation outside the repository.
- Added strict benchmark completeness checks so missing scenarios or agents cannot produce a conformance rate.
- Added credential-safe preflight that records only credential variable names, never values.
- Added isolated temporary project-scoped Skill installation for benchmark runs.
- Added manual credential-gated GitHub Actions workflow for full Codex + Claude Code evaluation.
- Fixed non-JSON scorer output and added schema/integrity validation.
- Added regression coverage for prompt blindness, provenance integrity, missing evidence, and secret redaction.
- Kept actual Agent performance claims blocked until genuine raw runs exist.
0.7.0 — Dependency intelligence hardening
- Upgraded Dependency Guard from baseline registry/OSV checks to risk-adaptive multi-signal evidence.
- Added explicit dependency necessity and purpose as separate judgment fields.
- Added deps.dev package/version, license, advisory, deprecation, related-project, and verified-attestation evidence.
- Added GitHub source-repository health checks with authenticated mode when a token is available.
- Added maintenance/release-age and deprecated-package review signals.
- Added package-name similarity / typo-squatting signals using explicit trusted names and discovered project dependencies.
- Added project license allow/deny policy evidence without pretending to provide legal advice.
- Added Tier 2/3 evidence requirements and Tier 3 provenance-attestation review.
- Added deterministic unit/failure-injection coverage and live provider contract coverage.
0.6.1 — Audit hotfix and governance hardening
- Fixed literal escaped-newline corruption in
SKILL.mdand added semantic validation to prevent recurrence. - Clarified the project-intelligence graph policy while keeping machine-generated graph data local-only.
- Removed the recommended Claude project-local checkout and hardened repository purity against local Vibe Skill checkouts.
- Pinned core GitHub Actions by commit SHA and made Graphify/Trivy contract tests read approved versions dynamically.
- Added automatic latest-stable Trivy compatibility proposals.
- Added official Agent Skills reference-validator compatibility workflow pinned to a known upstream commit.
- Hardened last-known-good recording with offline installation validation and isolated target validation before upgrades.
- Added validated-release automation and safe cleanup of unchanged merged work branches.
- Extended GitHub traceability snapshots and dry-run mutations to Milestones and GitHub Projects.
- Improved private security-reporting guidance.
- Removed stale hard-coded internal version identity from dependency/tool requests.
0.6.0 — Recovery, resume, cross-platform, and installation hardening
- Added repository-first resume context generation that does not depend on chat history.
- Added safe local-state inspection, quarantine, and recovery.
- Added offline installation validation for canonical and portable skill installs.
- Added last-known-good recording, local upgrade planning, and explicit rollback support.
- Added Linux, macOS, and Windows portability smoke tests across supported Python versions.
- Added portable-install validation in CI.
- Documented Python 3.10+ baseline, offline behavior, recovery semantics, and rollback limits.
0.5.0 — Graph provider contract, GitHub traceability, and project state automation
- Added a provider-neutral graph contract with Graphify as the default adapter.
- Added local shadow-source graph refresh so Graphify output never needs to live in the user project working tree.
- Added working-tree fingerprint freshness, including uncommitted and untracked non-ignored source changes.
- Added local graph query/path/explain routing with stale-graph blocking by default.
- Expanded Graphify compatibility contracts to cover explicit graph queries and incremental update.
- Added GitHub traceability snapshots, requirement mapping, verification, and dry-run-by-default Issue creation.
- Added local project-state capture, drift detection, and handoff snapshots without repository churn.
- Refactored doctor and integration guard to consume shared graph/GitHub adapters.
- Added end-to-end tests for graph shadowing, GitHub traceability, and local state automation.
0.4.1 — Local-only tooling and clean project repositories
- Moved Vibe Coding operational state from project-local
.vibe/to~/.vibe-coding/projects/<project-id>/. - Added a Local Workspace Manager for graph, security, test artifacts, benchmarks, caches, worktrees, and state.
- Added clone-local exclusions through
.git/info/excludewithout modifying project.gitignore. - Added Repository Purity Gate to block tracked/staged Graphify, Trivy, benchmark, coverage, and Vibe state artifacts.
- Updated bootstrap, doctor, integration guard, stale-graph tests, and failure injection to use external local state.
- Clarified that real product tests remain in Git while generated test/coverage reports remain local.
- Added worktree-safe Git exclude path resolution.
0.4.0 — Real-project validation, failure injection, and agent benchmark
- Added representative project validation fixtures and deterministic runner.
- Added failure-injection coverage for prompt injection, hallucinated dependencies, vulnerable packages, stale graphs, and unsupported Done claims.
- Added a structured completion evidence gate.
- Added vendor-neutral aggregation for real Codex/Claude Code behavior contracts.
- Added validation and benchmarking guidance without fabricated agent results.
- Added CI coverage for project validation and failure injection.
- Cleaned duplicate changelog heading.
0.3.1 — Live agent behavior evaluation
- Added a vendor-neutral JSON behavior contract for real Codex/Claude Code evaluation.
- Added deterministic scoring of agent tier, approval requirements, required controls, and forbidden actions.
- Added machine-checkable control contracts to every eval scenario.
- Added regression tests for passing and failing live-agent outputs.
0.3.0 — Risk classifier, full dependency adapters, and hardened tool contracts
- Added deterministic, explainable risk classification with Tier 0–3 workflow floors.
- Added Maven Central, NuGet, and Go module support to Dependency Guard.
- Added read-only GitHub/Graphify/Trivy integration health gate.
- Added executable scenario evals against the risk classifier.
- Added network-backed registry/OSV smoke tests.
- Added Trivy contract testing and approved-version toolchain tracking.
- Added unit tests for risk classification and stale-graph integration behavior.
0.2.0 — Bootstrap, dependency guard, evals, and portable packaging
- Added non-destructive adaptive project bootstrap.
- Added executable dependency guard with PyPI/npm/crates.io registry checks and OSV verification.
- Added deterministic behavior-contract eval catalog.
- Added unit tests for dependency decisions and bootstrap safety.
- Added portable Agent Plugin packaging for Codex-compatible discovery.
- Added package drift validation to CI.
0.1.0 — Initial core
- Added the risk-adaptive Vibe Coding Skill operating model.
- Added Tier 0–3 workflow classification.
- Added Graphify project-intelligence policy and automated compatibility proposal workflow.
- Added Trivy/OSV/package-registry security and dependency guidance.
- Added persistent project-state and traceability rules.
- Added doctor, change-budget, Graphify compatibility, and skill-validation utilities.
- Added GitHub Actions validation.