Changelog
Release history for 8-habit-ai-dev. This page summarizes notable changes; the authoritative sources are CHANGELOG.md (v2.3.0+), the GitHub releases page, and the git tag history.
Full detail for v2.3.0 and later lives in the root
CHANGELOG.md. This wiki page summarizes recent versions and keeps v2.2.0 and earlier for continuity.
Docs-only patch. Closes #194 via PR #195.
Documents a second spec-persistence deployment mode (project-orientation hub) as complement to v2.15.2's feature-spec mode. No new skill, no hook, no enforcement — pattern documentation only.
New guides/spec-digest-pattern.md (~180 lines) documents the project-root SPEC.md digest archetype: §1 Architecture (pointer), §2 Decisions snapshot (compact ADR digest table), §3 Live backlog, §4 Current state save point ("Read this section first after /clear or /compact"). Template paraphrased from a production artifact (scanopy/netbox-sit/SPEC.md, 153 lines) that independently arrived at this four-section shape after repeated /clear//compact flushing pain. ADR-013 receives an additive 2026-05-17 addendum clarifying its rejections (Alt-1 unified prd+design+tasks merge, Alt-4 always-on auto-write hook, CHANGELOG v2.15.0 /save-point skill rejection) cover the feature-spec mode specifically; the digest-layer-above-detail-files archetype is a different deployment mode those alternatives did not evaluate. /save-spec <slug> skill explicitly deferred pending ≥2 independent project adoption signal per working-with-pitimon "minimal additions, user-demand-driven" stance + PR #111 local-maximum lesson. Cross-links from guides/persistence-convention.md (the two modes are complementary) + README.md Use Cases row "Survive /clear and /compact". 8-habit-reviewer pre-commit dispatch caught nested-fence rendering issue (outer ```markdown collapsed \``bashto literal text); fixed to 4-backtick outer fence. Post-merge CI surfacedvalidate-content.shCheck 12b false-positive: link resolver flagged 8 template-internal placeholder links inside the code block as broken because the regex is not backtick-aware — template rewritten to use backticked filenames (``PLAYBOOK.md``, ``CONTRACTS.md` ``, etc) instead of bracket-paren link syntax with a note for the copy-paster. Same fix applied to Decisions table example row. Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 216 PASS / 0 FAIL / 1 WARN / 0 fitness breaches. Pattern: empirical-evidence-driven discipline addition — real-world artifact from another session revealed a deployment mode the plugin did not document; plan was revised twice (after reviewer + advisor) to match commitment level to evidence strength; deferred skill criterion makes the next escalation decision data-driven rather than aesthetic.
UX patch. Step 7 of /reflect now runs the 4-phase consolidation cycle automatically when count > 10 — no separate /reflect consolidate invocation needed. Closes #191 via PR #192.
No merges found → ~/.claude/lessons/INDEX.md updated automatically + 1-line summary. Deletions proposed → cycle stops for explicit user approval (In-the-Loop per ADR-002 preserved). Explicit /reflect consolidate argument retained for verbose output. DoD bullet updated to testable outcome. Pattern: PC² — H7 applied to H7 itself. Root cause was threshold 10 that mature repos cross quickly — auto-run is safe for the common case (additive INDEX update), gate preserved for the rare case (irreversible deletions). Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 217 PASS / 0 FAIL / 1 WARN / 0 fitness breaches.
Patch release. Doc-only addition responding to the industry move toward managed-agent runtime features (cross-session memory, self-evaluation against outcomes, built-in orchestration) from Claude Managed Agents, OpenAI Assistants, Bedrock Agents. Closes #188 via PR #189.
New guides/vendor-portability.md (~90 lines) structured around three principles: persist artifacts outside the vendor (repo = source of truth), treat managed memory as cache not source of truth, separate discipline (portable) from runtime (vendor-specific). Selection checklist framed as the /cross-verify Q14 "third alternative beyond the obvious options" exercise — managed vs. self-hosted is rarely binary, and hybrid with explicit persistence discipline is often the better answer. Habit mapping: H8 (Voice, architectural autonomy / Spirit) primary anchor, H1 (Proactive, prevent lock-in / Body), H4 (Win-Win, Emotional Bank Account / Heart — canonical framing per habits/h4-win-win.md), H7 (Sharpen the Saw, reproducibility = PC over P / Mind) — full Whole Person 4-dimension coverage. llms.txt indexed under Philosophy section. Plugin boundary preserved — workflow discipline lives here; enforcement hooks and compliance framework mapping live in pitimon/claude-governance. 8-habit-reviewer pre-commit dispatch caught a Commandment #13 violation: initial draft mislabeled Q14 as "External dependencies" but actual text per guides/cross-verification.md:45 is "third alternative beyond the obvious options" — corrected before merge, demonstrating the integrity discipline applies to its own writing. Pattern: discipline answer to a runtime trend — framework responds to vendor convergence not by replicating features but by naming the portability discipline that keeps users vendor-neutral. Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 217 PASS / 0 FAIL / 1 WARN / 0 fitness breaches.
Patch release. Doc-only fix closing a pair of same-shape drift gaps in guides/structured-output-protocol.md. Both stemmed from /design's SKILL_OUTPUT:design block being added without doc sync. Closes #153 via PR #186.
Producer entry: /design → SKILL_OUTPUT:design inserted between /requirements and /breakdown (workflow Step 2 placement). Schema matches skills/design/SKILL.md:128-142 exactly with all 7 keys; concrete example values match existing producer style. Consumer entries: Q14 (decision_count → third-alternative flag) + Q16 (sticky_decisions → WHY-not-captured flag); Q4 extended with design-block cross-check. All 5 SKILL_OUTPUT-consuming questions from skills/cross-verify/SKILL.md:35-41 now mirrored in the guide. Pattern: producer + consumer doc-sync-as-a-pair — strict-scope producer-only fix would have created the "confusion point" the issue cites (H4 + H1). Minimal scope expansion closes both halves in one PR. Originally surfaced by 8-habit-reviewer cross-verification of PR #152 (#151 attribution implementation). Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 217 PASS / 0 FAIL / 0 fitness breaches.
Patch release. Adds .github/workflows/link-check.yml — repo-wide markdown link validation using lychee. Triggers on PR + push to main when any **/*.md changes. Closes #172 via PR #184.
Scope: external HTTP/HTTPS URLs across all *.md outside docs/wiki/ (wiki covered by separate workflow). Excludes self-referential pitimon/8-habit-ai-dev/(blob|tree|raw)/main/ URLs (only resolve post-merge) and private cross-repo refs to pitimon/memforge + pitimon/devsecops-ai-team (workflow GITHUB_TOKEN is scoped to this repo only; cannot authenticate against other private repos). claude-governance is public and stays in scope. Internal markdown links remain validated by tests/validate-content.sh Check 12b — two-layer design (CI for external, shell for internal) covers full surface without duplication.
First CI run on PR #184 caught 3 real link-rot issues: README.md typo claud-mem-me → memforge (repo doesn't exist under typo'd name), ADR-005:137 dead EU AI Act Service Desk URL (EC restructured path; ADR-005 is Superseded per ADR-012, preserved as historical reference text with note), and pitimon/devsecops-ai-team/issues/467 404 due to private-repo CI token scope (added to exclude pattern with rationale). CONTRIBUTING.md gains "Link check (external URLs)" subsection under Testing Conventions documenting both workflows + the two-layer design. Pattern: CI gate that immediately proves its own value — link rot was already happening silently; the gate catches it at PR time vs when users report broken navigation. The 8-habit-reviewer recommendation from the 3-plugin integration audit is now enforced. Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 217 PASS / 0 FAIL / 0 fitness breaches; link-check CI run after fixes: PASS.
Patch release. First true bug-fix in the v2.15.x line — addresses the CRITICAL false-positive surfaced by v2.15.0's dogfood smoke test 9 days prior. Closes #167 via PR #182.
/consistency-check Pass 3 (Ambiguity) gains a "Backtick-context filter (required)" subsection — pre-strip backtick-quoted segments from each line before applying the [NEEDS CLARIFICATION] / TBD / TODO / ??? / XXX token match. Covers single-backtick inline spans and triple-backtick fenced blocks. Eliminates the docs/specs/consistency-check/prd.md:45 false-positive where the PRD legitimately mentions the token inside backticks as detection-target documentation. Option A from #167 (backtick-context filter) chosen over Option B (skip-marker comment): fewer escape hatches, principled, generalizes. Aligns analyzer runtime semantics with the validator-side whitelist that already exists for skills/consistency-check/ content (tests/validate-content.sh Check 12c) — the two-tier design (validator whitelist for skill prose + analyzer backtick-filter for spec artifacts) is now internally consistent. Reference.md known-limitation note removed (the limitation was the bug). F2 residual cleaned: tasks.md:48 Decision-D5/Decision-D9 → canonical Decision-5/Decision-9 per ADR-013, completing PR #169's earlier canonicalization. Validator Check 21 added asserting 3 contract signals — prevents future drift back to plain grep -nE semantics. Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 217 PASS / 0 FAIL / 0 fitness breaches. Pattern: bug fix of a feature shipped 9 days ago — distinct from v2.15.1/2/3 (content additions and convention imports).
Patch release. Content-only addition to guides/integrity-principles.md closing a verification-discipline gap surfaced during the v2.15.2 reflection. Two consecutive PR reviews (#174, #177) showed habit-attribution drift from gestalt pattern-matching, and a quote misattribution ("Magic" behavior at ADR-013 Alt-2 line 87 wrongly cited to Alt-4) propagated through 4 artifacts before reviewer catch. Closes #179 via PR #180.
New commandment #13: "Never paste a verbatim quote without grep-verifying its source." Scope: ADR citations, habit principle claims (H1-H8 attributions), scare-quoted external phrases, observation IDs, prior-conversation paraphrases presented as direct quotes. Title updated 12→13; mapping table row extended 5-7, 13 (Honesty) → H8 Find Voice / Spirit (conscience); 3 live cross-references synced (README architecture-tree comment, /review-ai load directive). Historical entries (v1.9.0 release line, v2.0.0 CHANGELOG delta, SELF-CHECK v1.9.0 improvements section) intentionally preserved as period record — editing them would itself violate commandment #4 (evidence) and #5 (paths). Dogfooding moment: drafting #13 caught its own meta-violation pre-commit — initial "magic behavior" (lowercase) self-corrected to "Magic" behavior (capitalized) to match ADR-013:87 exactly. Pattern: commandment growth driven by reflection-detected internal drift across ≥2 sessions — distinct from v2.15.1 upstream-import and v2.15.2 community-article convention-import. Validator state: validate-structure.sh 256/256 PASS.
Patch release. Convention-only addition to guides/persistence-convention.md importing a community-article save-point pattern (Thai-language article "ผมไม่เคยกลัว /clear กับ /compact") as a user-owned 4th file. Closes #176 via PR #177.
Three new sections in guides/persistence-convention.md: ## Current State File (Optional, User-Owned) (recommends docs/specs/<slug>/current-state.md with template doing-now / stuck-at / next / last-updated; user-owned, no plugin skill writes; frontmatter-exempt; /consistency-check explicitly excluded), ## Auto-Update Recipe (User-Side, Optional) (CLAUDE.md rule template the user adopts — plugin does NOT enforce per ADR-013 Alt-4: no-build philosophy + unintended file artifacts), and ## Attribution (credits the article + #176).
Solves the resume-after-context-loss problem (/clear, /compact, session crash) at task level — complements hooks/session-start.sh:83-115's step-level artifact-detection nudges (Issue #119, v2.7.0). Habit mapping: H5 + H3 for Current State File; H1 for Auto-Update Recipe. Honors PR #111's local-maximum lesson: no new skill, no new agent, no validator change, no DAG change. Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 0 fitness breaches. @8-habit-reviewer pre-merge: 17/17 PASS + 2 polish items both addressed in-PR. Pattern: community-article convention-only import — distinct shape from v2.15.1's upstream-skill extract; the article has no canonical PR/commit to cite, attribution is paraphrased title + issue link.
Patch release. Single-guide enhancement to guides/advisor-pattern.md importing three techniques from addyosmani/agent-skills — doubt-driven-development (MIT, PR #139, upstream 2026-05-07 — 27 days after our prior addyosmani audit in PR #111). Closes #173 via PR #174.
The new ## Disprove-Mode Disciplines section adds three labeled subsections: Anti-CLAIM-bias rule (H5) (pass ARTIFACT + CONTRACT only; hold the CLAIM back so the reviewer independently re-derives it), Iterative review 3-cycle cap (H3 + H7, conditional) (bounded loop with user escalation; single-shot pattern unchanged), and Adversarial prompt template (H1 + H5) (disprove-only output dispatched to a fresh subagent with no named role and read-only tools — not @8-habit-reviewer, whose 17-question process is fixed).
Honors PR #111's local-maximum lesson: no new skill, agent, or validator. Quick Reference table gains a second row for the adversarial pattern; See Also cites the MIT upstream. Rejected in-scope candidates explicit in #173 body: new /doubt-check skill, source-driven-development import, cross-model CLI escalation, agentskills.io frontmatter migration (ADR-007 NO-GO holds). Validator state: validate-structure.sh 256/256 PASS, validate-content.sh 214/214 PASS pre-release-meta. Pattern: post-audit delta — evaluate upstream methodology innovations in isolation when they publish after a periodic audit.
Minor release adding /consistency-check (the 18th skill) and an opt-in --persist <slug> argument to /requirements, /design, /breakdown. Inspired by github/spec-kit /analyze, adapted to our discipline-not-enforcement philosophy (#165).
The new /consistency-check skill reads persisted docs/specs/<slug>/{prd,design,tasks}.md files and runs 5 detection passes — Coverage, Drift, Ambiguity, Underspec, Inconsistency — emitting severity-graded findings (CRITICAL/HIGH/MEDIUM/LOW) with file:line citations. Hybrid evaluation: deterministic when artifacts include FR-NNN/Decision-N/Task #N ID markers (recommended), LLM semantic with explicit warning when absent. Read-only by design — emits findings, never blocks. Boundary preserved: enforcement on persisted artifacts still belongs in pitimon/claude-governance.
The persistence half is fully backward compatible: without --persist, all three modified skills behave byte-identically to v2.14.3. With --persist <slug>, they additionally write outputs to docs/specs/<slug>/{prd,design,tasks}.md with YAML frontmatter, while preserving conversation SKILL_OUTPUT blocks (/cross-verify auto-detect unaffected). Conflict policy: AskUserQuestion prompt → fallback to numbered variant in non-interactive contexts. Slug validation regex ^[a-z0-9][a-z0-9-]{1,63}$ prevents path traversal.
Self-applied dogfood: this release's own prd.md, design.md, tasks.md, and ADR-013 live in docs/specs/consistency-check/. Run /consistency-check consistency-check against the dogfood as smoke test (manual procedure documented in CONTRIBUTING.md). Skills count: 17 → 18. See ADR-013 for design rationale, 5 alternatives considered, flag-style argument convention attestation, and hybrid pass evaluation strategy.
Patch release closing #163 — three small cleanups left in v2.14.2's wake plus applying the validator's own 800-line rule to itself.
-
ADR-012 metadata closure —
SELF-CHECK.mdreframed two lines describing the deleted EU AI Act research + mapping files as if they still existed;docs/adr/ADR-012-eu-ai-act-migration-completion.mdstatus header upgraded with**Implementation**:field naming commited65b97(v2.14.2 release) and the metadata-closure date -
.gitignore— created with/deep-project/and/.claude/entries to gate against accidentalgit add .of cross-plugin checkouts and Claude Code session artifacts -
tests/validate-content.shtrim — 831 → 793 lines via comment consolidation across Check 15 (EU AI Act stub explainer), Check 19 sub-check rationales, F2 + F3 explainers; logic untouched, 10 checks preserved, PASS count = 205
Pattern: validator self-discipline — when a fitness function applies to the rest of the codebase, it applies to the validator too.
Plugin-boundary correction. The EU AI Act compliance toolkit (skill + reference + research + mapping guide) migrated to pitimon/claude-governance v3.1.0 on 2026-05-02 per memory observation #233270 (2026-04-07): 8-habit-ai-dev = workflow discipline; claude-governance = compliance enforcement + framework mappings. EU AI Act compliance is a framework mapping, not a workflow step. Original placement here was a boundary error.
-
Removed:
skills/eu-ai-act-check/reference.md,docs/research/eu-ai-act-obligations.md,guides/eu-ai-act-mapping.md(now canonical inclaude-governancev3.1.0) -
Stub:
skills/eu-ai-act-check/SKILL.mdrewritten as a redirect to the canonical location with install + invocation examples; preserves NOT LEGAL ADVICE disclaimer; skill name remains valid in the catalog so existing cross-references resolve - ADR: ADR-005 marked Superseded by ADR-012 (this migration completion)
-
Validator: Check 15 in
tests/validate-content.shrewritten to assert post-migration state (stub correctness + deleted-file negative assertions + ADR-012 presence + ADR-005 Superseded marker) - Cross-refs: reframed in RESOLVER, using-8-habits, design Step 5, ai-dev-log, session-start hook, README skill table
Wiki Architecture/FAQ/Home/Installation/Skills-Reference/Workflow-Overview pages and the README "EU AI Act ready" badge are deferred to a follow-up doc-only PR (precedent: pitimon/claude-governance PR #25 + #26 — local Markdown formatter rewrites tables on every Edit, producing 140+ lines of unrelated noise).
Coordination: this is the second half of the migration. The first half shipped in pitimon/claude-governance v3.1.0 PR #26.
Patch release closing #157 — external QA found validate-content.sh Check 19A was passing on the README badge URL releases/tag/v2.14.0 instead of asserting the section header ## What's New in v2.14.0. Same bug class as #124 (CHANGELOG pointer-fallback) and #141 (SELF-CHECK.md body drift) — capability-level recurrence on a sibling surface.
-
README.md backfill — restored missing
## What's New in v2.14.0block, fixed broken TOC anchor (#whats-new-in-v220→#whats-new-in-v2141, broken since v2.3.0), backfilled 4 missing skills in the architecture file tree (calibrate, using-8-habits, eu-ai-act-check, ai-dev-log) so the tree matches the declared 17-skill count. -
tests/validate-content.shCheck 19 sub-check G — anchored grep^## What's New in v${current_version}in README.md. Closes the badge-URL false-positive class permanently. Mirrors the sub-checks E + F mechanism from PR #144. -
Check 20 hardening — pin literal
Find → Fix → Re-Verifyloop name + assert exactly 5 numbered steps inskills/review-ai/SKILL.mdVerification Phase. Closes the v2.14.0 self-disclosed follow-up (passes-on-3-string-matches risk).
Pattern: same shape as v2.11.1 and v2.13.1 — QA surfaces drift class on a sibling surface → fix is fitness function, not checklist. Check 19 now covers README + CHANGELOG + wiki + SELF-CHECK + README-section-header — five anchored assertions, all co-located.
Fitness receipts: validate-structure.sh 246/0, validate-content.sh 206/0/1 WARN (was 203; +3 new assertions), test-skill-graph.sh 57/0, test-verbosity-hook.sh 19/0.
External QA report by @itarunp-apple — ran the plugin's own 8-habit-reviewer agent against v2.14.0 install per the framework's intended self-discipline workflow.
Closes #157.
Minor release closing milestone #15 — three workflow-discipline imports from Toh Framework (an "AI-Orchestration Driven Development" framework for solo SaaS builders). Cross-pollination filtered through plugin-boundary: workflow discipline here, project-state persistence routed to claude-governance.
-
SKILL_OUTPUT attribution lines (#151, PR #152) —
[/<skill>] COMPLETE SKILL_OUTPUT:<type>directly above each HTML comment in the 4 emitter skills. Status markersCOMPLETE/PARTIAL/FAILED(text-only). New Check 22 invalidate-structure.sh;cross-verifyparser unaffected. Inspired by Toh's Agent Announcement format. -
Argument-driven smart-routing for
/using-8-habits(#149, PR #154) —/using-8-habits "<intent>"returns ≤3 ranked skills + reasoning + alternatives + one direct question, instead of the full narrative tree. Reads~/.claude/habit-profile.mdfor verbosity and recent~/.claude/lessons/for context. Activates existingargument-hintfrontmatter — no new skill file. Inspired by Toh's/tohSmart Command (reshape: extend rather than wrap). -
/review-aiVerification Phase (#150, PR #155) — Find → Fix → Re-Verify loop: list CRITICAL/HIGH, apply fix, re-run review, cite evidence per finding, refuse to emitpass: trueunless all CRITICAL closed. Output ends with a Verification Table. Plugin boundary: section header reads "guidance only — NOT a hook"; new Check 20 invalidate-content.shenforces three-anchor boundary qualifier. Inspired by Toh's Test → Fix → Loop adapted as discipline guidance, not automated enforcement.
Pattern: external-framework cross-pollination kept tight by the boundary rule — 3 of 10 Toh ideas imported, 7 rejected with reason (multi-agent builders, "vibe" command, design profiles, component registry, multi-IDE adapters, 7-file project memory). Companion proposal in claude-governance (#24) for the persistence layer.
Fitness receipts: validate-structure.sh 246/0 (+1), validate-content.sh 201/0/1 WARN (+3), test-skill-graph.sh 57/0, test-verbosity-hook.sh 19/0.
Closes #149, #150, #151.
Patch release closing the three-PR arc for #141 — SELF-CHECK.md body drift (header said v2.13.0 but footer said Previous: 2.7.1 and per-release list ended at v2.8.0, skipping 6 releases).
-
One-time catch-up (PR #142) — SELF-CHECK.md footer updated to
Previous: 2.12.0; added 6 missing rows (v2.9.0 through v2.13.0). Plugin opens with "H8 Modeling: Follow the process always, no shortcuts when unwatched" — contradicted by 6 consecutive silent releases. -
Convention correction (PR #143) — CONTRIBUTING.md § Version Bumping "Version lives in 3 files" → "4 files", adds
SELF-CHECK.mdheader to the list (convention enforced since #106 but CONTRIBUTING.md never caught up). -
CI invariant (PR #144) —
tests/validate-content.shCheck 19 sub-checks E + F: footer must matchgit tag -l "v2.*" | sort -Vpredecessor ofplugin.json.version(E); every v2.x tag must have a matching^- v<x.y.z>:row in SELF-CHECK.md (F). Prevents recurrence of the drift class. -
.github/workflows/validate.yml—fetch-tags: true+fetch-depth: 0added toactions/checkout@v4so CI can read tag history.
Fitness receipts: validate-structure.sh 245/0, validate-content.sh 198/0/1 WARN (was 196 + 2 net new pass-able assertions — same hardening shape as v2.11.1 drift guard), test-skill-graph.sh 57/0, test-verbosity-hook.sh 19/0.
Closes #141.
Minor release making the plugin discoverable from non-Claude agent platforms — three linked PRs from the 2026-04-22 /research session on garrytan/gbrain. No breaking changes.
-
skills/RESOLVER.md(#135, PR #139) — flat phrase-to-skill dispatcher covering all 17 skills in 3 sections (Workflow / Assessment / Meta), ≤3 triggers each. Fills the phrase→path lookup gap; Check 20 enforces bidirectional coverage (directory ↔ RESOLVER row). -
llms.txt+AGENTS.mdat repo root (#136, PR #140) — cross-agent entry points for Codex / Cursor / Windsurf / Aider / Continue / LLM-based fetchers.llms.txtfollows llmstxt.org convention;AGENTS.mdis the non-Claude operating protocol. Check 21 enforces both files exist + 4× pointer integrity toskills/RESOLVER.mdandCLAUDE.md. -
README "Design Principle" section (#137, PR #138) — cites Garry Tan's 2026 essay "Thin Harness, Fat Skills" as external validation of the bounded-hook + fat-skills pattern already enforced by
hooks/session-start.sh(≤300 tokens). -
ADR-010 (Flat Skill Dispatcher) and ADR-011 (Cross-Agent Discoverability) — 6 options considered each; ADR-011 records the design-time empirical finding that
obra/superpowers-skills/.../references/(cited in #136's issue body) was already HTTP 404 — AGENTS.md cites upstream tool patterns by name only. - Pattern extracted: "Bidirectional Validator for Canonical Cross-References" — forward + reverse coverage invariants as the unit-test analog for documentation integrity. Check 12 / 20 / 21 share this shape.
Fitness receipts: validate-structure.sh 245/0, validate-content.sh 196/0/1 WARN, test-skill-graph.sh 57/0, test-verbosity-hook.sh 19/0. End-to-end cross-agent chain: llms.txt → AGENTS.md → skills/RESOLVER.md → individual SKILL.md.
Closes #135, #136, #137.
Minor release adding a new Evidence Standard obligation to /research and clarifying research-verifier scope (#133). Guidance-only — no automation, no hook.
-
/researchEvidence Standard — code-symbol verdicts matching/remove|dead|unused|transitional|safe to (drop|remove)/imust cite a grep-check showing consumer usage across source directories; declaration-site citations (e.g.package.json:6) do not establish liveness. Closes a false-positive class surfaced by memforgeneo4j-driveraudit (the canonical Bolt client for Memgraph — brand-name mismatch passed Deep-mode with pristine citations). -
research-verifierscope —description:rewritten to "citation-integrity verification agent"; new## Limit of Verificationsection defines in-scope (citation accuracy) vs. out-of-scope (semantic correctness of conclusions). Verifier emitsSEMANTIC-EVIDENCE-MISSINGflag on code-symbol verdict rows lacking liveness evidence but does not perform the grep itself. -
Trigger regex scope —
"revisit"intentionally excluded; over-triggers on follow-up-style verdicts. Hard-removal verdicts are the load-bearing class.
Fitness receipts: validate-structure.sh, validate-content.sh, test-skill-graph.sh, test-verbosity-hook.sh all green.
Closes #133.
Patch release hardening validate-content.sh Check 19 against recurring CHANGELOG drift (#124, PR #131). Post-v2.11.0 /cross-verify exposed the same drift class slipping CI twice (v2.9.0 + v2.11.0) through a pointer-fallback loophole.
-
3 new FAIL-severity assertions in Check 19:
CHANGELOG.md ^## v<ver>present, wiki^## v<ver>present (no pointer fallback), wiki badgelatest-v<ver>-bluematch. -
Backfilled missing v2.9.0 + v2.11.0 entries in root
CHANGELOG.mdand v2.11.0 in wikiChangelog.md.
Fitness receipts: validate-structure.sh 243/0, validate-content.sh 185/0, test-skill-graph.sh 57/0, test-verbosity-hook.sh 19/0.
Closes #124.
Close the final gap in the /requirements → /design → /breakdown → /review-ai structured-output-block handoff chain, and upgrade 20 wiki pages to a professional template. Both PRs merged within 3 minutes of each other.
-
SKILL_OUTPUT:designstructured block (#128 / PR #129) — closes the last cross-skill handoff gap;/cross-verifyQ4/Q14/Q16 now auto-populate from the design block. -
Tech-stack decisions as formal concerns —
/designStep 2 +/researchStep 1 surface language/framework as explicit outputs. -
Scope validation via
SKILL_OUTPUT:requirements—/designconsumes the requirements block so scope drift between decisions and success criteria is flaggable. -
H8 Whole Person in
/designcheckpoint — Body/Mind/Heart/Spirit prompts. -
Wiki redesign (#127 / PR #130) — new
Architecture.md(4-layer plugin design), newMaturity-Model.md(4-level adaptive guidance),Home.mdrewritten as hero landing page,Skills-Reference.mdexpanded 13 → 17 with quick-select matrix,Workflow-Overview.mdMermaid diagram, all 8 Step pages with> [!IMPORTANT]checkpoints,_Sidebar.mdreorganized with Concepts section. 18 files, 291 insertions, 53 deletions.
Fitness receipts: validate-structure.sh 243/0, validate-content.sh 183/0, test-skill-graph.sh PASS, test-verbosity-hook.sh PASS.
Refactor the 3 largest skills into SKILL.md + reference.md + examples.md triads per ADR-009. Creates headroom below the F3 word-budget fitness ceiling so future feature additions don't break the validator. Pattern sourced from external research (shanraisshan/claude-code-best-practice), filtered through plugin-boundary audit.
-
ADR-009 — codifies Load-directive mechanism (
${CLAUDE_PLUGIN_ROOT}inline paths), naming convention, and when-to-apply threshold (SKILL >1500 words). Reuses existing Check 8 for sibling existence — no new fitness function needed for that. -
F6 (Check 9b) — soft word-budget warning for
reference.md/examples.md(5000-word soft limit). -
using-8-habitssplit: SKILL 1990 → 1094 words;reference.mdholds the full 17-skill inventory + cross-plugin composition tables;examples.mdholds the password-reset onboarding walkthrough. -
eu-ai-act-checksplit: SKILL 1989 → 908 words;reference.mdholds the full 9-obligation checklist (25 MUST / 27 SHOULD / 8 COULD items) with article/paragraph references. Noexamples.md(pure-reference skill). -
calibratesplit: SKILL 1774 → 1161 words;reference.mdholds the scoring rubric + profile-write procedure;examples.mdholds 4 sample profiles (one per maturity level). -
Content validator triad-awareness — Checks 15 and 18 in
tests/validate-content.shnow search the SKILL + reference + examples triad as a unit, so content moved to sibling files still satisfies anti-drift and tier-count assertions. -
Rejected from this release — 27-event hook catalog (referred to
pitimon/claude-governance#22), Idea B (auto-loaduser-invocable: falsehabits), Idea D (parallel/cross-verifydispatch), F3 WARN→FAIL upgrade. Rejection rationale documented in rootCHANGELOG.md.
Fitness receipts: validate-structure.sh 243/0, test-skill-graph.sh PASS, validate-content.sh 183/0 with 0 fitness breaches. Release shipped via Issue #125 / PR #126 in a single clean forward pass.
Three features inspired by comparison research against piercelamb/deep-project. Cross-verified (14/17), advisor-reviewed, 8-habit QA passed (13/17 → 15/17).
-
Interview protocol for
/requirements(#118) — newguides/templates/interview-protocol.mdgives structured conversation scaffolding (Quick/Standard/Deep depth) for discovering requirements before EARS criteria. Replaces the "ask the user 5 questions" default with a better-shaped discovery flow. -
Workflow step awareness in session-start hook (#119) —
hooks/session-start.shnow surfaces a workflow step cue so partial chains can resume across sessions without per-skill rework. -
Machine-readable structured output blocks (#120) — new
guides/structured-output-protocol.mddefines<!-- SKILL_OUTPUT:... END_SKILL_OUTPUT -->HTML comment blocks at the end of/requirements,/breakdown, and/review-ai. Enables/cross-verifyto auto-check scope alignment (task_count vs ears_count ≤ 3× ratio) and review coverage, reducing manual re-reading of prior-step artifacts.
Production patterns from Anthropic's Claude Code internals (reverse-engineered in "Claude Code from Source") adapted into 4 existing skills as workflow guidance.
-
/build-briefcontext compression awareness (#114) — step 6 "Context survival" for briefs that survive the 4-layer compression pipeline -
/designsticky latch principle (#116) — step 5 "Sticky decisions" with rework-level classification table -
/reflectlesson consolidation (#113) — Step 7 +/reflect consolidateargument with 4-phase dream-inspired cycle -
/breakdownfork agent pattern (#115) — step 5 "Token-efficient parallel design" with ~90% cache hit guidance
Small post-milestone patch adding two disciplines to /review-ai after a cost/benefit audit against addyosmani/agent-skills (MIT). Scope deliberately minimal — only one of six candidate mechanics was imported.
-
/review-aiPerformance axis (#110, PR #111) — fourth review category flagging N+1 queries, unbounded loops, missing pagination, unindexed queries, and memory leaks; samefile:lineevidence standard as the other axes - Review-tests-first directive — new Process step 2 directs the reviewer to read new/changed tests before judging implementation
-
Rejections preserved in PR #111 body —
guides/anti-rationalization.md,guides/red-flags.md,guides/google-engineering-principles.md,/cross-verifyQ18, and a cross-plugin hard-gate spec were all evaluated and rejected as duplicative of existing features or out-of-scope
Closes the /calibrate feature loop by making skills read ~/.claude/habit-profile.md via a session-start hook.
-
Hook-based verbosity adaptation (#96) —
hooks/session-start.shemits a per-level directive into session context; 16 existing skills auto-adapt with zero file changes -
guides/verbosity-adaptation.md— canonical per-level rules with 5 skill-archetype examples -
tests/test-verbosity-hook.sh— 12-assertion regression coverage for all 8 hook branches + HABIT_QUIET opt-out + ≤300-token budget check - 4 validators in CI, 482 total assertions (up from 3 / 470)
- Milestone v2.7.0 CLOSED — Hermes-inspired feature loop (v2.6.0 + v2.7.0) complete
-
/reflectQ6 Skill Effectiveness signal (#92) — 6th retro question captures "most useful" and "least useful/confusing" skill -
SKILL-EFFECTIVENESS.md(repo root) — maintainer-curated trend tracker; H7 applied to the plugin itself -
guides/templates/lesson-template.md— new## Skill effectivenesssection for consistent Q6 capture -
Fix: SIGPIPE flake in
validate-content.shF3 extractor — replacedsed | headwith pipe-safe awk
-
/calibrateskill + habit-profile schema v1 (#90, ADR-008) — 5-7 question self-assessment writing~/.claude/habit-profile.mdwith dominant-level scoring -
guides/habit-profile-schema.md— public schema contract (YAML + markdown body, versioned viaschema-version) -
Persistent reflection artifacts (#88) —
/reflectwrites lessons to~/.claude/lessons/;/researchand/build-briefread them before starting work -
guides/habit-nudges.md+ ADR-007 — nudge spec (hook delegated to claude-governance) + agentskills.io NO-GO decision -
Skill count: 16 → 17 (
/calibrateadded)
-
tests/test-skill-graph.sh— DAG validator forprev-skill/next-skillchains (#79): cycles, dangling refs, symmetric edges, orphans -
hooks/pre-commit.sh.example— template running/review-aion staged files (opt-in, not auto-installed) -
Bidirectional wiki ↔ skills linking (#81) — each workflow skill has a
## Further Readingsection linking to its wiki page - CI now runs 3 validators; 443 total assertions
Same-day correction after comparing /brainstorm to superpowers:brainstorming.
-
Removed
/brainstormskill (breaking) — superpowers' 500+ line hard-gate discipline suite is a better fit;/researchnow references it for fuzzy problem statements -
HABIT_QUIET=1opt-out forhooks/session-start.sh— users who internalize the workflow can silence the reminder -
"Core 5" tier in
/using-8-habits— 80/20 reality acknowledgment - ADR-006 — audience-honesty + opt-out + "check peer plugins before building parity" lesson
-
/brainstorm(Step 0a, later removed in v2.4.1) — 5 Whys, alternative framings, hidden assumptions -
EARS-notation in
/requirements— 5 structured acceptance criteria templates from Rolls-Royce (Mavin et al. 2009) -
/using-8-habits— onboarding meta-skill with decision tree and complete walkthrough example - Validators: +52 assertions, anti-drift check ensures meta-skill references every directory skill
- Fix:
validate-structure.shregex allowing digits in skill names
Flagship blue-ocean feature: first Claude Code plugin with explicit EU AI Act compliance toolkit, shipped ~4 months before 2 August 2026 enforcement.
-
/eu-ai-act-check— 9-obligation tiered checklist (25 MUST + 27 SHOULD + 8 COULD) covering Articles 9-15 -
/ai-dev-log+scripts/generate-ai-dev-log.sh— AI-assisted development log fromgit log+ Co-Authored-By trailers (4 modes: markdown/json/summary/out) -
/designStep 5 — Article 14 human-oversight 5-capability checkpoint (Understand / Automation bias / Interpret / Override / Stop button) -
docs/research/eu-ai-act-obligations.md— primary-source research with Verified Quotes for all 7 articles -
ADR-005 + Plugin Boundary section in
CLAUDE.md— documents complementary relationship withpitimon/claude-governance - Version file convention corrected to 4 files (added
SELF-CHECK.md) - Fix:
validate-structure.shSIGPIPE race replacedsed | headwith awk
- README overhauled with a professional 8-Habit aligned template
- Hero tagline reframed around pain-point + benefit
- Quick Start split into install + use blocks
- 7-Step Workflow diagram simplified
- GitHub repo description and topics updated
- Wiki infrastructure introduced (
docs/wiki/, sync Action, link check) — see ADR-004
- Smart QA integration with 8-Habit framework
- README internal link and skill cross-reference validation
-
validate-content.shfitness function improvements
- Three accepted ADRs drive architecture:
- Three architecture fitness functions (skill complexity, content depth, cross-reference integrity)
-
validate-content.shadded alongsidevalidate-structure.sh
- Initial 7-step workflow and 8-habit skill set
- Cross-verify agent (
8-habit-reviewer) - External QA review v1.0.0 scored 9.5/10 (EXCELLENT) — Issue #1
This plugin follows semantic versioning:
- Major — breaking change to skill interfaces, skill removal, or workflow restructuring
- Minor — new skills, new habits content, backward-compatible additions
- Patch — documentation fixes, typo corrections, clarifications
Version is tracked in four files that must bump together (enforced by tests/validate-structure.sh):
.claude-plugin/plugin.json.claude-plugin/marketplace.json-
README.md(badge + footer) -
SELF-CHECK.mdheader