Changelog

All notable changes to Hornero are documented here. The format follows Keep a Changelog. Plugin version is anchored in .claude-plugin/plugin.json and the root plugin.json.

[Unreleased]

[0.4.0] - 2026-10-27

Added

• claude plugin eval suite under evals/cases/: 25 routing cases (one per natural-language row of tests/dispatcher/routes.tsv, graded on the skill that fired) and 7 PM-lens artifact cases with an llm quality grader. validate-dispatcher.sh Check 4 keeps routes.tsv and the routing cases in sync.
• EVAL_HORNERO_NO_AUTO_EXTRACT=1 opt-out in hooks/auto-extract-lesson.sh, set by every eval case so the Stop hook's extra lesson turn does not alter graded output.
• .github/workflows/evals.yml: opt-in (workflow_dispatch plus weekly) eval run; needs the ANTHROPIC_API_KEY secret. ci.yml does not run evals.
• .claude-plugin/marketplace.json: /plugin marketplace add argen/hornero then /plugin install hornero@hornero.
• Portable root plugin.json (Agent Plugins standard) with Codex settings under extensions["com.openai"] (display name, default prompts, kickoff as the onboarding skill), and .agents/plugins/marketplace.json for the Codex repo marketplace.
• skills.sh install path (npx skills add argen/hornero); a local npx skills add run discovered all 59 skills.
• .claude/rules/skills.md, a path-scoped rule (skills/**, roles/**, resources/**) with the validate-skill, catalog, and validate-resource reminders.
• allowed-tools: Read, Grep, Glob, Bash(scripts/chronicle.sh *) on chronicle-recall, prep, and review.
• Stop hook timeout (30s) and statusMessage; _comment notes Claude Code's cap of 8 consecutive Stop-hook continuations.
• skills/{orchestrate,role,kickoff,catalog,audit}/SKILL.md as type: command skills (kickoff, catalog, audit, role set disable-model-invocation: true; orchestrate stays model-invocable because the dispatcher routes to it).
• .claude/skills -> ../skills so a session opened at the repo root discovers every skill without the plugin or install.sh; .gitignore un-ignores it.
• validate-skill.sh fails on unquoted name/description values containing : or starting with a YAML indicator, fails on descriptions over 1024 characters, and warns over 500 lines; covered by scripts/test-validate-skill.sh.
• scripts/check-md-links.sh (with test-check-md-links.sh) fails on broken relative Markdown links.
• smoke-check.sh runs claude plugin validate on a dereferenced copy of the repo (skipped when the CLI is absent) and the link check; CI installs the CLI and runs the new tests.
• This changelog.
• Hardening of the new checks: validate-skill.sh also rejects unquoted values with :<TAB>, a trailing :, a leading -/?/,/]/}, a # comment start, or an unclosed quote; check-md-links.sh ignores inline code spans and ~~~ fences, skips any URI scheme and accepts <angle> targets; smoke-check.sh validates a tracked-files-only dereferenced copy and no longer aborts on missing validator output; CI pins the Claude Code CLI to 2.1.311 with the auto-updater disabled.

Changed

• prd, positioning-statement, success-metrics, strategy-kernel, bet-portfolio and cross-functional-orchestration carry an Output discipline block: each finding names its framework and author, at least one trade-off is stated, and the response ends with one to three concrete next actions. Artifact eval cases went from 0.40-0.60 to 0.86-1.00.
• Artifact eval cases split the single quality.md grader into q-sources, q-distinct, q-tradeoff and q-actionable (plus q-decision for the launch plan), so a case passes at 0.8 with one criterion missing and shows which one.
• Per-role skills moved from roles/<role>/skills/<name>/ to skills/<name>/; every skill is now a real directory and role membership is the role: frontmatter key (38 skills carry role: product-manager). roles/<role>/ keeps ROLE.md and sources.md. All path references and relative links across the docs are rewritten.
• catalog.sh groups skills by role: and lists type: command skills under a new "Slash commands" section; smoke-check.sh step 7 now fails on any symlink under skills/ or a role: with no roles/<role>/ROLE.md.
• Slash commands are skills (type: command); the invocation syntax, argument handling, and route-declaration lines of the six commands that shadowed same-named skills (review, prep, ingest, hypothesis-track, stakeholder-update, customer-update) plus extract-lesson now live in each skill's Application section, with argument-hint: where they take arguments.
• validate-skill.sh exempts type: command skills from the six-section anatomy (they still need name, description, type).
• install.sh installs every file of each skill directory and uses flat skill names (no <role>- prefix); --dry-run counts and the DISPATCHER loading table follow the new layout.
• skills/competitive-teardown/SKILL.md split: the worked example moved to examples.md, bringing it under the 500-line warning.
• evals/judge.py and evals/README.md default judge models are now claude-sonnet-5-5, gpt-5 (unverified) and gemini-2.5-pro; the README no longer claims 7 roles.
• orchestrate and cross-functional-orchestration state that only product-manager-reviewer exists today and that the orchestrator covers unshipped lenses itself.
• catalog.sh and build-codex-agents.sh strip surrounding double quotes from frontmatter values.
• ROADMAP "Releases" section replaced by a pointer to this file.
• install.sh --codex installs skills to the shared ~/.agents/skills/ path (agents and hooks stay under ~/.codex/) and points to npx skills add argen/hornero as the lighter path.
• smoke-check.sh step 8 compares root plugin.json against .claude-plugin/plugin.json (name, version, extensions["com.openai"] present) and checks mcp.json server names against the Claude manifest (none declared today, so no mcp.json ships).
• README install sections lead with the marketplace flows; counts now read 59 skills (38 PM, 16 cross-role, 5 command skills; 12 slash-invocable) in the README and plugin descriptions.
• Descriptions of 15 skills in the finance, problem-framing, memory, discovery, and strategy clusters state what they produce, when to use them, and which sibling to use instead.
• ROADMAP Distribution: marketplace files ship; public Codex directory submission is blocked because plugins that ship hooks are ineligible.

Fixed

• Eval fixtures: the chronicle routing cases now seed a chronicle/ project and scripts/chronicle.sh via seed-chronicle.sh (run with --scaffold, which CI and the README commands now pass); the positioning-statement, saas-revenue-growth-metrics and saas-economics-efficiency-metrics routing cases carry their artifact or metrics inline; success-metrics description now covers critiquing an existing metric definition.
• Four skills (prep, hypothesis-track, stakeholder-update, customer-update), the /hypothesis-track command and the github resource card had unquoted description: values containing : , which is invalid YAML; Claude Code loaded them with empty metadata. Descriptions are now double-quoted.
• repository in both plugin manifests is a string, as claude plugin validate requires.
• Dead links to skills/multi-role-review and to roles that were never shipped (software-engineer, data-scientist, tester, designer, security-expert, marketing-expert) removed or reworded as plain "future role" prose; stale ../../ links in commands/ and three renamed-file links corrected.
• templates/settings.json hooks use the real Claude Code shape (matcher plus hooks[], tool_input.file_path read from stdin via jq); invented CLAUDE_TOOL_FILE_PATH, effortLevel and skipAutoPermissionPrompt keys removed.
• Generated Codex agent TOMLs no longer carry sandbox_mode (Codex ignores it in custom agents); they set model_reasoning_effort = "medium" instead. AGENTS.md uses agents.max_concurrent_threads_per_session.
• ROADMAP skill counts corrected to 38 product-manager and 16 cross-role skills.

Removed

• evals/judge.py, evals/score.sh, evals/tasks/ and evals/runs/ (bespoke human and LLM dual-scoring flow that was never run); evals/rubric.md now feeds the llm graders and evals/EVAL-REPORT.md is kept as history.
• .codex-plugin/plugin.json, superseded by the root plugin.json.
• commands/ and the .claude/commands symlink (superseded by skills), and scripts/build-plugin-skills.sh (no symlinks left to generate).
• install.sh no longer copies to ~/.claude/commands or ~/.codex/prompts; skills replace both. Claude Code users type /<name>. Codex has no disable-model-invocation and no prompts directory here, so a Codex user invokes a skill explicitly by typing $ or using /skills to mention it (for example $orchestrate), or implicitly by describing the task in natural language.

Upgrading

• install.sh now prunes files the previous install recorded in .hornero-manifest that the new install does not write (old ~/.claude/commands/*.md, ~/.codex/prompts/*.md, ~/.claude/skills/<role>-<skill>/), honouring --dry-run; run ./install.sh again to clean up. Files that are not hidden (dotfiles under skills/) are no longer installed.

[0.3.0] - 2026-05

PM skill expansion, conversational kickoff, automatic lesson extraction, a voice for the companion, present-focused docs.

• A voice for the companion — VOICE.md codifies the personality (calm, terse, plain-spoken, no preening) grounded in the rufous hornero's traits (patient layers, duet, ground-level walk, comfortable around humans). Six touchpoints rewritten to land the voice: SessionStart banner, /kickoff opener, dispatcher's declared one-liner, Stop-hook "nothing to capture" line, /catalog footer, README microcopy. CLAUDE.md + AGENTS.md reference VOICE.md on first dispatch. Artefact outputs (PRDs, OKRs, etc.) stay voice-neutral so they travel cleanly to stakeholders.
• Automatic lesson extraction — hooks/auto-extract-lesson.sh wires the Stop hook to fire the extract-lesson skill at every session end without the user having to type /extract-lesson. The agent self-filters via Chris MDP's signal-vs-noise discipline (most sessions stop in one line; only ones with real signal produce proposals). Re-stop loop prevented via stop_hook_active. Opt-out: HORNERO_NO_AUTO_EXTRACT=1. Plugin hook, install.sh recommendation, and templates/settings.json all updated.
• Docs reorientation — canonical docs (README, ROADMAP, methodology, ATTRIBUTION, DISPATCHER, CLAUDE, AGENTS, CONTRIBUTING, roles/README, onboarding, role-cheatsheet, adoption, commands/role) now lead with present capabilities rather than the journey. README opens with a "What you get" section; ROADMAP adopts Now / Next / Later + Releases.
• Conversational kickoff — /kickoff no longer asks one question and offers a 7-item bulleted menu; it opens with a 3-sentence intro and a single open invitation, then routes via the dispatcher. DISPATCHER.md gains Step 6b (ambiguous-but-rich → one clarifying question, never a menu). SessionStart banner names natural language first; slash commands as alternatives.
• PM skill expansion — 22 new PM skills in four commits, re-synthesised from Dean Peters' Product-Manager-Skills (CC BY-NC-SA 4.0 source; Hornero's re-synthesis is MIT). New themes: career/leadership (4), AI-PM (2), discovery extensions (3), outside-in (3), workshops + delivery (3), finance (7). Total PM skills 15 → 37; total skills 21 → 45.

[0.2.0]

• Antagonistic role-reviewer E2E (PM) — scripts/play-against.sh --check runs the PM positive case + PM negative control via expect against the real interactive claude CLI; transcripts persisted to tests/seeded/last-run/ on every run for spot-checking. Coarse regression net for catastrophic persona drift.
• CI self-breakage guard — .github/workflows/ci.yml runs validate-skill, validate-resource, test-validate-resource, and smoke-check on every PR. Opt-in local mirror via scripts/install-git-hooks.sh. Catches stale CATALOG.md, frontmatter contract violations, and skill↔resource cross-ref breaks before merge.
• Resources layer (PR 1 — foundation) — new top-level resources/<system>/RESOURCE.md cards describing external data sources (object model, MCP server, gotchas, fallback). Skills opt in via resources: [<slug>] frontmatter. Ships resources/linear/ and the linear-triage consuming skill; scripts/validate-resource.sh enforces anatomy; smoke-check.sh gains bidirectional skill↔resource cross-ref. Hornero stays usable when no MCP server is configured (graceful degradation via documented fallback paths). PR #24.
• Vendor-agnostic plugin parity — Hornero loads at full fidelity on both Claude Code and Codex from a single source-of-truth. .claude-plugin/plugin.json and .codex-plugin/plugin.json sibling manifests; reviewer subagents available as agents/<name>.md (Claude) and auto-generated .codex/agents/<name>.toml (Codex) via scripts/build-codex-agents.sh; install.sh --codex installs skills + agents + hooks + commands. PR #23.
• Plugin packaging (Claude Code) — .claude-plugin/plugin.json manifest, top-level skills/<name>/SKILL.md symlinks (auto-built by scripts/build-plugin-skills.sh), plugin-loadable session hooks. PRs #20, #21.
• Blog-alignment series (PRs #18–#19) — guardrail hooks in templates/settings.json, templates/sub-CLAUDE.md.template + playbooks/claude-md-layering.md, explore/edit separation made explicit in DISPATCHER.md, templates/MAP.md.template, /audit slash command + playbooks/skill-rot-review.md (quarterly cadence), playbooks/adoption.md (DRI + cross-functional working group + approval process).
• Memory layer v0 — project-local memory layer at memory/decisions/ shipped as the third Hornero primitive. Two new cross-role skills: memory-query and decision-record. playbooks/memory-loop.md walks the read-decide-write cycle end-to-end. Distinct from adr (architecture-shaped) and prd (feature-shaped) — decision-record captures cross-role posture-shaping calls. Opt-in per project; absent memory/decisions/ directory → no-op behaviour.
• Chronicle layer — machine-queryable record of decisions any future session or role can load as situational awareness. Influenced by Quorum's Chronicle module — same lifecycle, simpler substrate (Markdown-with-strict-frontmatter, deterministic harness, no embedder dependency). Two cross-role skills: chronicle-record (write side, calls scripts/chronicle.sh propose — LLM-free; human commits) and chronicle-recall (read side, byte-identical output for a given corpus). Deterministic harness at tests/chronicle/ backs the layer.