Docs Update

Codex compatibility note:

• Invoke repository skills with $skill-name in Codex; this mirrored copy rewrites legacy Claude /skill-name references.
• Task tracker mandate: BEFORE executing any workflow or skill step, create/update task tracking for all steps and keep it synchronized as progress changes.
• User-question prompts mean to ask the user directly in Codex.
• Ignore Claude-specific mode-switch instructions when they appear.
• Strict execution contract: when a user explicitly invokes a skill, execute that skill protocol as written.
• Subagent authorization: when a skill is user-invoked or AI-detected and its protocol requires subagents, that skill activation authorizes use of the required spawn_agent subagent(s) for that task.
• Do not skip, reorder, or merge protocol steps unless the user explicitly approves the deviation first.
• For workflow skills, execute each listed child-skill step explicitly and report step-by-step evidence.
• If a required step/tool cannot run in this environment, stop and ask the user before adapting.

Codex Project-Reference Loading (No Hooks)

Codex uses static project-reference loading instead of runtime-injected project docs. When coding, planning, debugging, testing, or reviewing, open project docs explicitly using this routing.

Always read:

• docs/project-config.json (project-specific paths, commands, modules, and workflow/test settings)
• docs/project-reference/docs-index-reference.md (routes to the full docs/project-reference/* catalog)
• docs/project-reference/lessons.md (always-on guardrails and anti-patterns)

Missing/stale context route: If docs/project-config.json, the docs index, lessons.md, CLAUDE.md, AGENTS.md, or any task-required reference doc is missing or stale, auto-run $project-init or the narrow setup route ($project-config, $docs-init, $scan-all, $scan --target=<key>, $claude-md-init) before ordinary project-specific work. If Codex mirrors or AGENTS.md are missing/stale, ask the user to run $sync-codex; do not auto-run it.

Situation-based docs:

• Project structure/architecture/tech-stack/deployment/setup (any layer — backend, frontend, or infra): project-structure-reference.md
• Backend/CQRS/API/domain/entity changes: backend-patterns-reference.md, domain-entities-reference.md
• Frontend/UI/styling/design-system: frontend-patterns-reference.md, scss-styling-guide.md, design-system/README.md
• Spec authoring, docs/specs/ pathing, or TC format: feature-spec-reference.md, spec-system-reference.md, spec-principles.md
• Behavior/public-contract changes or spec-test-code sync: workflow-spec-test-code-cycle-reference.md plus the spec docs above
• Derived spec indexes/ERDs/reimplementation guides: spec-system-reference.md and source Feature Specs under docs/specs/
• Integration test implementation/review: integration-test-reference.md
• E2E test implementation/review: e2e-test-reference.md
• Code review/audit work: code-review-rules.md plus domain docs above based on changed files

Do not read all docs blindly. Start from docs-index-reference.md, then open only relevant files for the task.

[BLOCKING] Execute skill steps in declared order. NEVER skip, reorder, or merge steps without explicit user approval. [BLOCKING] Before each step or sub-skill call, update task tracking: set in_progress when step starts, set completed when step ends. [BLOCKING] Every completed/skipped step MUST include brief evidence or explicit skip reason. [BLOCKING] If Task tools are unavailable, create and maintain an equivalent step-by-step plan tracker with the same status transitions.

Quick Summary

Goal: Detect impacted docs from code changes and orchestrate updates across all doc types so every code/spec/test change leaves documentation in sync — impacted project-reference docs, project-config.json sections, Feature Specs, §8 TCs, test-code links, and derived indexes all reflect the shipped behavior, with zero drift left silent.

Summary:

• This skill is a ROUTER, not an author — start with Phase 0 triage (git diff → categorize → dedup modules → check existing docs) and delegate each doc type to its owner ($spec, $spec [mode=tests], $spec [mode=sync], $spec-index, $tech-spec for derived technical views); NEVER write §8, docs/specs/, or derived technical spec content directly. — why: dual authorship diverges spec from index/view.
• Main steps (each impact-gated; skipped phase → mark completed with reason): Phase 0 triage (git diff → categorize → dedup modules → record existing-doc state) → Phase 1 project context sync (impact-map → PARALLEL verify of the impacted docs/project-reference/** docs + docs/project-config.json sections + README.md) → Phase 2 $spec (§1–§7 Feature Spec; doc-first BLOCK when feature behavior changed but no Spec exists) → Phase 2.5 $spec-index (derived bucket INDEX/ERD refresh, optional) → Phase 2.6 $tech-spec (derived technical view refresh/audit, optional when technical tree is affected) → Phase 3 $spec [mode=tests] (§8 TCs) → Phase 4 $spec [mode=sync] (§8 ↔ test code) → Phase 5 summary report → final review (#8 runs the Step 2.4 code↔spec sync-verify).
• Create ALL 8 tasks via task tracking before touching any file; run the fixed phase order 0 → 1 → 2 → 2.5/2.6 → 3 → 4 → 5 → final review — fast-exit is a decision, never a silent omission.
• The final pass (Step 2.4) is the workflow's last gate: per touched module verify shipped code against §3 ACs, §4 BRs, §8 TCs — a removed/weakened [HARD] BR is a code-vs-spec contradiction that BLOCKS completion.
• Output is tech-agnostic prose (no framework/product names outside evidence fields) and traceability-first (update FR-/BR-/OP-/TC- logical IDs before prose); ALWAYS write the Phase 5 summary report as the audit trail.

Orchestration Model:

git diff → Triage → Phase 1: Project Context Sync (PARALLEL, impact-scoped)
                  │            ├─ impacted docs/project-reference/** — verify → patch → (escalate to $scan --target=X)
                  │            ├─ impacted docs/project-config.json sections — verify → merge → validate
                  │            └─ README.md / project docs (docs-manager)
                  → Phase 2: $spec (business feature docs)
                  → Phase 2.5: $spec-index (derived index/ERD refresh) [optional]
                  → Phase 2.6: $tech-spec (derived technical view refresh/audit) [optional]
                  → Phase 3: $spec [mode=tests] (§8 test specifications)
                  → Phase 4: $spec [mode=sync] (§8 ↔ test code sync)
                  → Phase 5: Summary Report

Key Rules:

• Router only — NEVER duplicate sub-skill logic or write Section 8 / docs/specs/ content
• [BLOCKING] Freshness is impact-scoped, never assumed. Phase 1 verifies only the docs/project-reference/** docs and docs/project-config.json sections the diff can actually rot (routed by node .claude/scripts/doc-impact-map.cjs), and reports a per-doc verdict FRESH | PATCHED | RESCAN REQUIRED | UNVERIFIED. A doc nobody checked is UNVERIFIED — NEVER FRESH. — why: these docs are injected into every downstream AI context, so a silent stale line teaches every later agent a codebase that no longer exists.
• [BLOCKING] An impact-scoped verify NEVER writes <!-- Last scanned: -->. Only a full $scan --target=X may move that stamp; the narrow pass writes <!-- Last verified: ... --> instead — and only in a doc that already carries a Last scanned stamp (Step 1.6). — why: Last scanned drives the 60-day full-rescan gate (.claude/hooks/lib/session-init-helpers.cjs:769); resetting it from a partial check would buy speed by disabling the very net that catches whole-doc rot.
• **[BLOCKING] A PATCHED docs/project-reference/**doc MUST run$prompt-enhance <doc> before its verdict is final** (Step 1.3) — keeps the doc as concise as possible while staying valuable enough for AI; skip only for a stamp/count-only edit.
• Phase 1/docs-manager MUST NOT own any docs/specs/**, test-spec, spec-index/ERD, or derived technical-view path.
• Every excluded artifact is explicitly reserved to its child skill ($spec, $spec-index, or $tech-spec) so one canonical writer owns it.
• Exclude docs/specs/** and generated technical views from every docs-manager brief and write set.
• Each phase checks whether needed before invoking — skip phases with no impact
• Step-to-skill order is fixed — run phases sequentially, never out of order
• ALWAYS report what was checked, even if nothing needed updating
• Pass triage context (changed files, detected modules, impacted sections) to each sub-skill via $ARGUMENTS
• MUST ATTENTION dedup module list — backend + frontend changes for same module = ONE entry
• MUST ATTENTION track step state live: in_progress -> execute -> completed (or completed with skip reason)
• For .claude skills/hooks/workflows/sync tooling changes, flag generated mirror sync status (npm run codex:sync completed or explicit N/A). docs-update routes and reports this check; it does not edit generated mirrors directly.
• [BLOCKING] Tech-agnostic output: when updating spec/specs/README/INDEX, do NOT introduce framework/product/language/design-pattern names into prose or headings — preserve the evidence-field exception (**Evidence**, CoveredBy, legacy IntegrationTest, [Source:], frontmatter, Mermaid). Authority: docs/project-reference/spec-principles.md §3.
• [BLOCKING] M3 Traceability Update: See .claude/skills/shared/sdd-artifact-contract.md → "AI-SDD Mandates (M1-M7)" for BLOCKING criteria. When syncing docs after code changes, update the logical-ID mappings (FR-/BR-/OP-/TC-) FIRST, then the prose. The [Source: namespace/service/id] abstract-anchor evidence is re-resolved ONLY if the logical artifact was renamed/split — a file move or stack change does NOT change the anchor (physical coords live only in the provenance sidecar) — and the logical-ID spine stays stable across the change — never drop or renumber a logical ID just because the code moved. Keep all synced prose M1/M2-clean.

Be skeptical. Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80%.

Mandatory Task Creation (ZERO TOLERANCE)

[BLOCKING] Create ALL 8 tasks via task tracking BEFORE touching any file. NEVER consolidate, rename, omit. Conditional tasks skipped: mark completed immediately with reason — NEVER silently omit.

Execution rules:

• Mark each task in_progress when starting, completed when done — one active at a time
• Multiple modules → add one subtask per module for Phase 2/3 invocations
• Multiple impacted reference docs → add one subtask per doc (or per source-of-truth cluster) under Task 2, so each doc's verdict is tracked individually
• NEVER batch-complete — each sub-skill invocation tracked individually
• Phase 0 TRUE fast-exit (impact map empty) → mark tasks 2-8 completed with reason "Skipped — impact map empty"
• Phase 0 PARTIAL exit (docs/config impacted but no business behavior — e.g. harness, CI, or manifest-only changes) → run Task 2, mark tasks 3-6 completed with reason "Skipped — no business behavior changed", still run tasks 7-8
• NEVER execute a phase step until matching task status is in_progress
• After each phase/skill call, write one-line evidence in task update (what ran, what changed, why skipped)
• If task tracking/task updates unavailable, maintain equivalent 8-task plan tracker with same status transitions

Step-Skill Call Order (Do Not Reorder)

| Order | Task ID | Step / Phase | Skill Call | Tracking Rule | | ----- | ------- | ----------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------ | | 1 | 1 | Phase 0: Triage | Inline triage logic in this skill | Set Task 1 in_progress before diff scan; set completed after module + impact map recorded | | 2 | 2 | Phase 1: Project Context Sync | doc-impact-map.cjs + PARALLEL docs-manager sub-agents (one per impacted doc/cluster) + $scan --target=X or $project-config only on escalation | Set Task 2 in_progress before the impact map; completed only after EVERY routed doc and config section carries a verdict + evidence | | 3 | 3 | Phase 2: Business Feature Docs | $spec | Set Task 3 in_progress before invocation; completed after output review | | 4 | 4 | Phase 2.5/2.6: Derived View Refresh | $spec-index [mode=index] and/or $tech-spec [mode=generate | audit] | Set Task 4 in_progress before invocation; completed after derived outputs are refreshed or skipped with reason | | 5 | 5 | Phase 3: §8 Test Specs | $spec [mode=tests] | Set Task 5 in_progress before invocation; completed after TC review | | 6 | 6 | Phase 4: §8 ↔ Test Code Sync | $spec [mode=sync] | Set Task 6 in_progress before invocation; completed after sync validation | | 7 | 7 | Phase 5: Summary Report | Inline report write | Set Task 7 in_progress before report write; completed after file path confirmed | | 8 | 8 | Final Review | Inline verification gate | Set Task 8 in_progress before final audit; completed after all phases justified |

Enforcement: If a required step cannot run, STOP and ask user before adapting order. Never continue with untracked steps.

Phase 0: Triage — Detect Impacted Documentation

Step 0.1: Collect Changed Files

1. Run git diff --name-only HEAD (staged + unstaged changes)
2. No uncommitted changes → git diff --name-only HEAD~1 (last commit)
3. Still empty → git diff --name-only origin/develop...HEAD (branch changes)

Step 0.2: Categorize Changes

This table classifies BUSINESS-doc impact (which of Phases 2-4 run). It is deliberately coarse. The precise docs/project-reference/** + docs/project-config.json routing is produced by the impact map in Step 1.1 — read it there, never guess it here.

Step 0.3: Fast Exit Check — decided by the impact map, never by path intuition

Run the impact map NOW (same command as Step 1.1) and read fastExit from its output:

node .claude/scripts/doc-impact-map.cjs --text

[BLOCKING] A .claude/**-only (or tooling-only) diff is NOT a full fast exit. Harness edits change the skill/hook/agent/workflow inventories that CLAUDE.md, docs-index-reference.md, and project-structure-reference.mdderive by globbing.claude/: the counts and catalogs go stale with zero feature impact, and NO other gate in this skill catches them — Phases 2-4 only look at docs/specs/\*\*. — why: the cheapest way to ship stale docs is to classify the change as "tooling" and skip the only phase that would have noticed.

Step 0.4: Auto-Detect Affected Modules

Extract unique module names from changed paths. MUST ATTENTION dedup: unique() before passing to any sub-skill — backend + frontend same module = ONE entry. Prevents duplicate $spec invocations.

Build project-specific mapping from docs/project-config.json and project reference docs, not from hard-coded skill paths:

node -e "const cfg=require('./docs/project-config.json'); console.log(JSON.stringify({sourcePaths: cfg.codebaseHealth?.sourcePaths, contextGroups: cfg.contextGroups?.map(g => ({name:g.name,pathRegexes:g.pathRegexes})), specRoot: 'docs/specs/'}, null, 2))"
node -e "process.stdout.write('docs/specs/')"

Step 0.5: Check Existing Docs for Each Module

For each detected module:

1. Check the matching bucket directory exists under docs/specs/
2. Check that the bucket contains README.*.md Feature Specs, or use the project reference doc's feature-doc layout
3. Check the matching bucket directory under docs/specs/ exists using project reference docs
4. Record: hasFeatureSpec (§1–§7 present), hasTestSpecs (§8 present), hasDerivedIndex (bucket INDEX.md present)

Step 0.6: Declare the Doc-Update Wave

Detection is SEQ and comes FIRST — every assignment below is derived from the impacted-doc set, so nothing dispatches until Steps 0.1–0.5 have produced it. Once that set exists, updating N unrelated docs is embarrassingly parallel: one docs-manager sub-agent per doc or per doc cluster, all spawned in ONE message.

1. Declare before dispatch — Parallel plan: wave 1 = [docs-manager: {doc A}, docs-manager: {cluster B}, …] · SEQ = [Phase 0 triage, the Phase 2 → 2.5/2.6 → 3 → 4 spec chain, Phase 5 report] (reason).
2. STRICT one-writer-per-file. Every impacted doc path appears in EXACTLY ONE agent's brief, stated as that agent's owned file set. A doc owned by nobody is a silent miss; a doc owned by two agents is a lost-update race where the later write wins and the earlier finding vanishes.
3. [HAZARD] Two docs that embed the same canonical or derived data MUST go to the SAME agent — never split across the wave. Counts, catalogs, module maps, INDEX rows, ERD entities, and any table copied out of a source of truth have to be regenerated by ONE writer from ONE reading of that source. Split across two agents they diverge inside a single commit — and the divergence survives review because each doc is internally consistent and only the pair is wrong. Cluster by SOURCE OF TRUTH, not by directory.
4. Barrier before the spec chain. Phase 2 → 2.5/2.6 → 3 → 4 stays a FIXED SEQ chain: $spec output feeds the derived index, and §8 TCs feed [mode=sync]. Parallelism lives INSIDE a phase across independent modules — never across these phases.
5. Per-module fan-out is PAR only when the modules are disjoint. Two detected modules that map to ONE Feature Spec share a write target and stay a single task — the same reason Step 0.4 dedups the module list.
6. Every member returns a summary + Full report: path; YOU merge them into the Phase 5 report only after ALL members return, a skipped member counting as returned.

Phase 1: Project Context Sync — Reference Docs + project-config.json (PARALLEL, impact-scoped)

Why this phase exists. docs/project-reference/** and docs/project-config.json are injected into EVERY downstream AI context and route every skill in the framework. When code moves and they do not, the harness keeps teaching a codebase that no longer exists — and nothing else in this skill catches it, because Phases 2-4 only look at docs/specs/**. $scan-all + $project-config do repair them, but they re-derive every doc from zero, which is why they run every 60 days instead of every change. This phase does the same job at diff scope: same no-stale guarantee, small enough to afford after every change.

When to run: ALWAYS, unless Step 0.3 declared a TRUE fast exit. Run it even when every one of Phases 2-4 is skipped.

Scope discipline: verify ONLY what the impact map routes; escalate to a full $scan --target=X when a surgical patch cannot make the doc true again. NEVER regenerate all docs, and NEVER hand-author a full reference doc here — scan owns authoring, this phase owns verification and surgical repair.

Step 1.1: Build the Doc-Impact Map (SEQ — everything below derives from it)

node .claude/scripts/doc-impact-map.cjs --json     # machine-readable (drives the wave)
node .claude/scripts/doc-impact-map.cjs --text     # human-readable (goes in the report)
node .claude/scripts/doc-impact-map.cjs --base=origin/main   # branch-scope instead of working tree

The map routes each changed file to the docs and config sections it can rot, and returns per doc: doc, exists, lastScanned/ageDays, scanTarget (the full-rescan escalation), checks (which verifications apply), changedFiles/addedFiles/deletedFiles, and heuristicOnly. Routing is derived from docs/project-config.json (contextGroups, modules, testing, e2eTesting, styling, designSystem, specRoots) plus change-class rules — never from hardcoded project paths.

Handling the map's output — [BLOCKING] rules:

1. unrouted files are not proof of no impact — they are proof the router had no rule. Classify each by hand: add it to the wave, or record in the report why it carries no doc impact. NEVER let an unrouted file silently pass as fresh.
2. A heuristicOnly doc is a GUESS, not evidence. Verify it like any other, and downgrade to "not impacted" only with a stated reason.
3. Any doc whose exists: false is a MISSING doc, not a fresh one → route to $scan --target=<scanTarget> (or $docs-init when the whole set is absent).
4. If the script is unavailable (older checkout, non-Node host), derive the same map by hand from docs/project-config.json — match changed paths against contextGroups[].pathRegexes → guideDoc/patternsDoc/stylingDoc/designSystemDoc, modules[].pathRegex → project-structure + modules, test/e2e/styling paths → their docs, manifests → tech stack, infra/CI → ports & deployment, .claude/** → inventory counts. Record that the map was manual.

Step 1.2: Declare the Verify Wave (PAR — one message, all members)

Parallel plan: wave 1 = [docs-manager: {doc A}, docs-manager: {cluster B}, docs-manager: project-config.json, …] · SEQ = [Step 1.1 impact map, the Phase 2 → 2.5/2.6 → 3 → 4 spec chain, Phase 5 report] (reason)

Wave-construction rules — the Step 0.6 hazards apply verbatim, plus:

• STRICT one-writer-per-file. docs/project-config.json has exactly ONE owning agent in the wave, always. Two agents merging JSON into the same file is a guaranteed lost update.
• Cluster by SOURCE OF TRUTH, not by directory. README.md + project-structure-reference.md both restate the module/directory map, and CLAUDE.md + docs-index-reference.md + project-structure-reference.md all embed .claude/-derived counts — each such set goes to ONE agent so the numbers cannot diverge inside a single commit.
• Every routed doc appears in exactly one brief. A doc owned by nobody is a silent miss.
• Every member returns its verdict table + Full report: path; merge only after ALL members return.

Step 1.3: Per-Doc Verify Contract (what each wave member actually does)

Verify FIRST, patch NARROW. Run only the checks the map listed for that doc:

Verdict per doc (exactly one, evidence required):

[BLOCKING] Never fabricate freshness. "Looks fine", "probably unchanged", and "the diff was small" are not check results. A doc that was not verified is UNVERIFIED, never FRESH — a false FRESH is worse than no check, because it retires the suspicion that would have caught the drift later.

[BLOCKING] Every PATCHED docs/project-reference/**doc MUST run$prompt-enhance <doc> (default --op=enhance) before the verdict is recorded. These docs are injected into every downstream AI context — a surgical edit that adds correct prose without re-compressing still leaves the doc bloated. $prompt-enhancekeeps content as concise as possible while staying valuable enough for AI (caveman compression + attention anchoring), closing the same gap on the narrow patch path that$scan --target=X's own mandatory final step (scan/SKILL.md Final Step) already closes on a full rescan — so a RESCAN REQUIRED doc that escalates to $scan gets it for free and needs no separate call here. Skip ONLY for a single stamp/date/count-only edit, and record the skip reason.

Step 1.4: project-config.json Drift Check (single writer, sche

…(truncated)