feat: Wave 7 — hot-checks + harness-hooks + template test + §13.23 4th-layer - #29
Merged
Conversation
….27 + §13.28 + §13.8) O0 verdict NO end-to-end match — pre-commit/MegaLinter/Trunk don't unify harness-hooks with code+doc lints. Wave 7 scope: 5th lifecycle stage (harness-hook) + doc-linting + functional template test. §1.7 forward+backward applied; review session decides; no implementation in this commit. Prior-art: skipped — research patch, no new capability shipped (cites existing SSOT #1-#15; proposes new #16-#22 for sub-wave 7.x implementation commits)
…-1, MAJOR-2, MINOR-1, MINOR-2) MAJOR-1: §3 ATTN + REPORT ATTN — correct shipped-artefact paths (skills/, agents/ at repo root, not .claude/skills/, .claude/agents/); narrow Claude-first bias claim to .claude/skills/self-reflection/ + proposed harness-hook layer. MAJOR-2: §0 row + §1 row 1 + §8 entry #16 + §11 sub-wave 7.1 + §12.1 — lint-staged is NOT installed and root ESLint config does NOT exist; reframe as PRECONDITION for sub-wave 7.1 (not baseline). MINOR-1: cascade-table footnote — STILL ARMED includes deferred-with-armed-trigger. MINOR-2: §13.23 trigger #3 (pre-push surface widening) satisfied by Wave 7 sub-waves 7.1.b/7.1.c — explicit note in §0.1 + §6. Prior-art: skipped — text corrections to research patch, no new capability shipped
…§13.28+§13.8) Independent reviewer GO with confidence high. 6 PASS / 2 PARTIAL (write-time fixups, non-blocking) / 0 FAIL across 8 review dimensions per .claude/orchestrator-prompts/wave-7-hot-checks-joint-closure/review.md. §9 open decisions all closed: D1: ship harness-hook (5th layer) + §13.23 4th-layer design scheduled separately D2: DEFER Vale (FP risk on Russian+English mixed prose) D3: ship deterministic-only P1+P4+P6; LLM-judge probe DEFERRED D4: Claude Code only project-side; cross-editor stays WATCHLIST D5: §13.28 A+B both ship (sub-waves 7.4 + 7.2.b) D6: hand-roll .markdownlint.json (no Microsoft/Google preset) D7: harness-hook = SHOULD project-side / MAY consumer-side Sub-wave outline confirmed (7.1-7.5 per parent §11). Branch base note: Wave 6 not yet on main — operator merges Wave 6 first before Wave 7 PR opens. Prior-art: skipped — review verdict patch consolidating §9 open-decision closures, no new capability shipped
…hips in Wave 7 sub-wave 7.6 Decision 1 revised after user adversarial pushback («ты просто отложил потому что поленился делать?»). Adversarial check on load-bearing claim «§13.23 has 4 unresolved design problems unfit for Wave 7 scope» revealed initial deferral was false: Problem 1 (scope predicate): solvable via ## §-section diff content match Problem 2 (bootstrap chicken-and-egg): solvable via §1.7 Bootstrap: trailer Problem 3 (trailer-format interaction): NOT a problem — line-anchored parser unaffected Problem 4 (discipline-theatre): universal hook tradeoff, not §13.23-specific §13.23 4th-layer pre-push trailer check now ships in Wave 7 sub-wave 7.6 with full discipline cycle: 7.6.a research → 7.6.b independent review → 7.6.c implementation → 7.6.d closure (folded into 7.5.c). Wave 7 timeline: 6-9 days → 8-12 days (+2.5-3.5 days for 7.6). Sub-waves: 5 → 6. §1.7 enforcement ladder: 3 active + 1 deferred → 4 active layers post-Wave 7. §7 Revision note documents the failure mode for future review sessions: candidate anti-pattern #deferral-by-pattern-match-on-deferred-status — reviewer accepted «Status: deferred» + listed problems as load-bearing without running §1.4 adversarial check on each problem individually. Sample size 1/3 toward §3 distillation threshold. Prior-art: skipped — review verdict revision after adversarial-check correction, no new capability shipped
….23 per independent re-review Independent re-review (Agent task a618b8a650eb3b627, fresh session) returned REVISE on commit 4ede6ae (v2 revision). Findings: v2 overcorrected by claiming all 4 §13.23 problems tractable without verification — same cognitive shortcut as v1, opposite direction. Honest per-problem assessment: Problem 1 (scope predicate): UNCERTAIN — design sketch, pending 7.6.a Problem 2 (bootstrap): TRACTABLE — clean analogy to Prior-art: skipped escape-hatch Problem 3 (trailer-format): UNCERTAIN — claim rests on unverified parser mechanism assumption (git interpret-trailers vs grep line-anchored) Problem 4 (theatre risk): TRACTABLE — clean analogy to warn-only patterns Decision 1 routing: v1 (e75019a) = «defer» — accepted unsolvability without §1.4 check v2 (4ede6ae) = «CLOSED ship» — accepted tractability without §1.4 check v3 (this) = «CONDITIONALLY CLOSED» — 7.6 ships IFF 7.6.b GO on P1+P3 §7 documents symmetric anti-pattern pair: #deferral-by-pattern-match-on-deferred-status (v1) #tractability-by-pattern-match-on-claim-acceptance (v2) Generalised: #claim-acceptance-without-§1.4-on-the-claim-itself §8 adds operator-pre-approved fallback decision tree (F1-F7) for sub-wave 7.6.a/b outcomes — orchestrator does NOT escalate during execution if outcome matches pre-decided branch. Kickoff updates: §11 Decision 1 row reflects conditional closure; STOP-conditions add «7.6.b REVISE on P1/P3 → apply fallback §8»; mis-cite of independence rule corrected (now references Wave 7 review-prompt convention + §1.4 spirit, not §1.4 directly). Prior-art: skipped — review verdict v3 incorporating independent re-review findings, no new capability shipped
…ew findings (BLOCKER + 2 MAJOR + 2 MINOR) Second independent re-review (agent task afdc9688180da109f, fresh session, read-only) returned REVISE on commit 2be167f (v3). Plus self-check on load-bearing claims surfaced concrete evidence. Findings applied: BLOCKER (1, kickoff): CLOSEOUT block was unconditional «§13.23 closed by Wave 7». Orchestrator parses kickoff, not verdict patch — conditional was in the wrong document. Fixed: CLOSEOUT now has Path A (7.6 ships) / Path B (F6 fallback, §13.23 deferred to Wave 8) / Path C (operator escalation) branches. Sub-wave 7.6 header also carries explicit IFF condition. MAJOR-1 (Problem 4 overclaim): «warn-only-first-N-days is standard project pattern» asserted without verification. Self-check + agent caught it: self-application.md:60 documents Soft-warn concept in Decision matrix but NO husky-script implementation exists (grepped .husky/ + packages/core/ for warn-only / --warn — zero matches). Problem 4 downgraded from TRACTABLE-high → TRACTABLE-PENDING-CITE-medium. 7.6.a must design first- time soft-warn-in-husky implementation. MAJOR-2 (F6 cascade gap): if F6 fires (§13.23 defers), sub-wave 7.5.a row count changes 8 → 7. Not explicit in F6 description before. Fixed: F6 row now spells out 7.5.a row count change + 7.5.c closure note shape + 7.5.b SSOT entries still land (not wasted). MINOR-1 (Problem 2 hidden asymmetry): Prior-art: skipped is structurally triggered; §1.7 Bootstrap: relies on positional heuristic (rev-list count = 1) — fragile if hook commit not first on branch. Confidence downgraded HIGH → MEDIUM. 7.6.a must commit to either positional-with-hard-rule OR structurally-detectable subject-prefix. MINOR-2 (F4 arbitrary threshold): «3-day threshold» had no derivation from project metric. Replaced with operator-escalation trigger keyed to existing «>50% cost overrun» rule. Iteration pattern: v1 → v2 → v3 → v4 each surfaces smaller, less-prominent errors. Asymptotic-honesty convergence. Independent re-review at each step is what produces the convergence — not self-review. Prior-art: skipped — review verdict v4 incorporating second independent re-review findings, no new capability shipped
…cord hygiene) Third independent re-review (agent task a660e0834e5a89737, fresh session) returned GO with confidence medium-high on commit 5a5646a. v4 confirmed stable form — no BLOCKER, no MAJOR. Three MINOR cosmetic findings applied as record-hygiene v5: 1. §7 title renamed «TWO-STEP correction» → «Revision history — multi-step correction». Body always correctly described 4-step history (v1→v2→v3→v4); the heading was rotting and would mislead cold readers scanning headers. 2. §8 F4 escalation trigger now cites exact kickoff rule text verbatim («7.6.c implementation cost exceeds estimate by >50% (>2 days for ~1-2 day estimate)») instead of paraphrasing. Forward-tracing gap closed. 3. Kickoff sub-wave 7.6 header parallel-list extended «7.1/7.2/7.3» → «7.1/7.2/7.3/7.4». 7.4 doesn't conflict with 7.6 (different surfaces: Makefile vs .husky/pre-push); inconsistency with verdict §3 footnote resolved (footnote was already correct). Convergence verdict: three rounds of independent re-review converged on v4 as stable form. Errors changed CHARACTER across iterations (defer-overclaim → tractable-overclaim → calibrated-uncertainty), NOT just shrank in size — third agent flagged this nuance against the v4 commit message's «errors shrink» framing. The §7 «same cognitive shortcut» framing in the document is correct; the v4 commit-message «asymptotic-honesty» phrasing was slightly stronger than the document's own claim. v5 is the SHIP candidate. Implementation can proceed on orchestrator- kickoff basis per CLOSEOUT Path A (or fallback per F1-F7 in review §8). Prior-art: skipped — cosmetic record-hygiene revision; no new capability shipped, no decision logic changed
…endent-rounds heuristic (sample 1/3) Operator asked «когда останавливаться» — formalised the iteration-convergence signal that drove v4→v5 ship decision. Heuristic: «when N independent re-reviews converge on cosmetic/cite-fixing MINOR findings only with zero BLOCKER and zero MAJOR, ship the current revision». Proposed N=3 — mirrors AIF /aif-evolve aggregation threshold; empirically validated this iteration (Round 3 at cosmetic-only). Symmetric anti-pattern flagged: #never-ship-due-to-recursive-revision (asymptotic-honesty pursuit becomes process theatre when every round produces SOME finding regardless of marginal correctness gain). Sample size 1/3 — lives in §7 of this verdict per phase-research-coverage §3 «one patch per gap, append-only». Two more incidents needed for §3 distillation into rule §4 catalog (or new §5 «Stopping rules» section). (Also incorporates linter reformat of markdown tables — cosmetic widening of column padding; intentional per system reminder.) Prior-art: skipped — methodology distillation lives in §7 of review verdict, no new capability shipped
…aint analysis Operator-requested audit of all LLM touchpoints across project (current + planned + speculative). Constraint: «нет платного» — distinguish FREE (Claude Code subscription session-bound) vs PAID (CI / bare API). Key finding: current state is 100% «no-paid» compliant. - 11 current LLM touchpoints — ALL FREE (session-bound) - 12 CI jobs in audit-self.yml + discipline-self-check.yml + .husky/pre-push — ALL deterministic, zero LLM - Architectural invariant: «LLM-then-cache» pattern (LLM runs in session, writes deterministic artifact, CI reads artifact) §13.25 MCP discovery (operator's specific question): reuse AIF `/aif` for rules 1-4 (free, session-bound); rules 5-6 as in-session file-write. No standalone API-calling CLI. Documentation generation (operator's expected LLM use): ship as Claude Code skill (free, session-bound), NOT as CLI with API key. Wave 7 Decision 3 (LLM-judge probe): DEFER aligns perfectly with no-paid. CI gate deterministic-only; LLM-judge as optional manual session check if/when triggered. 5 open decisions surfaced (§9): D1: «no direct LLM in CI» as MUST row in Decision matrix D2: Consumer subscription assumption in INSTALL-FOR-AI.md D3: Gate 5 Opus subscription tier fallback policy D4: Doc generation as skill (recommended) vs CLI D5: Wave 5 constraint to session-only (no API-calling CLI) 2 SSOT proposal candidates (#23 LLM-then-cache pattern; #24 Claude Code hooks API) for future sub-wave commits. Prior-art: skipped — research patch documenting LLM usage audit, no new capability shipped
…tions (Decision 3 revised + D1/D2)
Operator selected Option B from 2026-05-11 LLM usage audit synthesis:
1. Decision 3 (§13.27 LLM-judge probe) REVISED — TWO-MODE shipment:
- CI gate (sub-wave 7.3.e) = deterministic only (P1+P4+P6)
- NEW local advisory skill (sub-wave 7.3.f) = Claude Code session-bound
LLM-judge for paraphrase/cue-placement checks. FREE within subscription.
Audit confirmed deterministic-CI + session-LLM is the architectural sweet
spot for «no-paid» constraint. Operator's «documentation for AI validated
by AI» philosophical point honored without triggering §13.11 Gate 5
cost-tracking infrastructure prematurely.
2. D1 from audit — «no direct LLM API calls in CI» added as MUST meta-rule
row in self-application.md §3 Decision matrix at sub-wave 7.5.a.
Locks the «LLM-then-cache» architectural invariant against future drift.
Decision matrix row count: 8 → 9 (Path A) / 7 → 8 (Path B F6 fallback).
3. D2 from audit — INSTALL-FOR-AI.md update at sub-wave 7.2.d now covers
BOTH editor-coupling acknowledgement (Decision 4) AND «requires Claude
Code subscription for AI-assisted workflows» note (D2).
D3 (Opus tier fallback), D4 (doc generation as skill speculative §13.X),
D5 (Wave 5 HARD CONSTRAINT) — DEFERRED per operator Option B choice.
Live as armed triggers; handled in respective Wave 5 / Wave 8+ scopes.
Wave 7 timeline: 8-12 days → 9-13 days (+1 day for B additions).
Sub-waves remain 6 (7.1-7.6); sub-wave 7.3 gains 7.3.f.
Kickoff updated in parallel (operator-side, gitignored):
- ARGUMENTS estimated days
- §11 Decision 3 row revised
- Sub-wave 7.2.d expanded with D2
- Sub-wave 7.3 gains 7.3.f
- Sub-wave 7.5.a 8 → 9 rows
- CLOSEOUT Path A 8 → 9; Path B 7 → 8
Prior-art: skipped — review verdict v6 incorporating Option B audit-driven additions; no new capability shipped at this commit
…p injection Add .claude/settings.json with UserPromptSubmit hook entry and .claude/hooks/inject-session-bootstrap.sh (13 LOC). The hook prints a session-bootstrap digest (goal phrase + 3 invariants + Step-0 reading order + pointer to .claude/session-bootstrap.md) to stdout; Claude Code injects stdout into the user-prompt context automatically. Closes Wave 6 D-1: session-bootstrap injection via harness-hook channel. DECISIONS: - .claude/settings.json was ABSENT before this commit. - New-file LOC: settings.json 14 + hook script 13 = 27 total — under 80 threshold. - Capability commit: NO (total new-file LOC <80; no new packages/ subdirectory). Prior-art: skipped — config-layer hook addition (UserPromptSubmit channel) extending settings.json with session-bootstrap injection digest; no new packages/ subdirectory or ≥80 LOC file. Closes Wave 6 D-1 (session-bootstrap injection via harness-hook).
- Add markdownlint-cli2 ^0.17.0 as workspace root devDependency. - Create .markdownlint.json (root, auto-discovered): 6 hand-rolled structural rules (MD001/MD003/MD007/MD009/MD034/MD040); no preset. MD013 line-length disabled via default:false (long prose lines are legitimate). - Create packages/lint-config/ (workspace member @rules-as-tests/lint-config): mirrors root config in markdownlint.json + documents rule set in README.md (99 LOC, Authoritative-for header per doc-authority-hierarchy rule §3). - Extend .husky/pre-commit with markdownlint-cli2 check on staged *.md files (ACMR diff-filter; bash-driven, no lint-staged). Rule set decisions: MD001 — heading hierarchy; skipped levels hide doc structure drift. MD003 atx — project uses # prefix exclusively; setext style rejected. MD007 indent:2 — 4-space indent creates parser rendering differences. MD009 — trailing spaces are invisible noise; no intentional <br> in prose. MD034 — bare URLs skip link text; [text](url) required for context. MD040 — code blocks without language tag degrade syntax highlighting. MD033/MD013 disabled — prose-style FP risk; 500-line limit covers length. Capability commit: YES (new explicit dep markdownlint-cli2 in package.json; new file packages/lint-config/README.md ≥80 LOC under packages/). Prior-art: prior-art-evaluations.md#17 (markdownlint-cli2, verdict transition DEFER → ADOPT after operator-side discipline rule §13.27 closure).
…apper (§13.28 B) Add .claude/hooks/validate-prompt.sh (34 LOC) and PostToolUse Edit|Write hook entry in .claude/settings.json. The hook fires on every Edit/Write tool call, filters to .claude/orchestrator-prompts/**/*.md paths via shell-glob match, then invokes validate-batch-spec.ts via node_modules/.bin/tsx. Exits 0 silently on pass or unmatched path; non-zero + diagnostic on red. Gracefully skips (exit 0 + stderr warn) if jq or tsx unavailable. exit 2 from validate-batch-spec.ts (gh CLI unavailable) is treated as soft-skip (exit 0) to avoid blocking tool calls for missing tooling. DECISIONS: - Invocation form: node_modules/.bin/tsx (resolved REPO_ROOT-relative) rather than bare npx tsx, to avoid auto-install attempts. - Path globbing: shell pattern [[ ... == *".claude/orchestrator-prompts/"*".md" ]] (not jq-aware) — Claude Code PostToolUse passes file_path in tool_input; no need for jq-level glob, plain shell match suffices. - Capability commit: NO (wrapper ≤50 LOC + settings.json hook entry; no new packages/ subdirectory or ≥80 LOC file). Prior-art: skipped — config-layer wrapper script ≤50 LOC + settings.json hook entry invoking validate-batch-spec.ts on .claude/orchestrator-prompts/**/*.md PostToolUse events. Closes §13.28 part B (operator-side discipline via Claude Code hook channel). No new packages/ subdirectory or ≥80 LOC file.
Extend .husky/pre-push with §8 lychee offline link check on changed *.md files (git diff origin/main..HEAD --diff-filter=ACMR). Checks broken relative links and anchors without network access (--offline --no-progress). Design decisions: - Offline mode: no network calls — deterministic and fast; remote URL validity is out of scope for a pre-push hook (belongs in CI link-check workflow). - Absence fallback: if lychee not in PATH, print one-line warning + install path and exit 0. Rationale: operator-discovery responsibility; hard-fail would block all pushes on fresh checkouts before lychee is installed. Install path documented inline: cargo install lychee OR brew install lychee. - File selector: git diff origin/main..HEAD --diff-filter=ACMR | grep -E '.md$'. Skips cleanly when no changed .md files exist (empty CHANGED_MD check). - Word-split on $CHANGED_MD (unquoted): safe for this repo's filename conventions (no spaces in doc paths); avoids xargs -r portability issue on macOS BSD xargs. Capability commit: NO (config-only; no new packages/ subdirectory, no new file ≥80 LOC; modification to existing .husky/pre-push — status=M not status=A). Prior-art: skipped — pre-push hook addition (config layer) wiring lychee --offline check on changed Markdown; no new packages/ subdirectory or ≥80 LOC file.
…a+7.3.b)
Add packages/core/audit-self/template-render.audit.ts — vitest .audit.ts
harness for §13.27 functional template test. Commit 1 (skeleton): install.sh
wiring in beforeAll, P1/P4/P6 probes stubbed as .todo pending commit 2.
Vitest discovers the file via new audit-self/**/*.audit.ts include in
packages/core/vitest.config.ts.
Add skeleton fixtures under packages/core/audit-self/fixtures/:
- ts-server/package.json: {"name":"ts-server-fixture","version":"0.0.0","private":true}
- react-next/package.json: declares next@^15 + react@^19 (declaration-only, no install)
- {ts-server,react-next}/src/.gitkeep: empty src/ placeholder
- fixtures/README.md: folder-level authority header per doc-authority-hierarchy.md §5
Decision log:
- Skeleton fixture shape: 2 stacks × 1 package.json + src/.gitkeep = 6 files
- README.md in fixtures/ — folder authority pattern applied (§5)
- afterAll: rm -rf tmpdir via node:fs/promises rm — no repo pollution
- beforeAll timeout: 90s (install.sh is file-copy only; alarm if slower in CI)
Prior-art: prior-art-evaluations.md#22 (Cookiecutter/Copier, verdict ADOPT VOCABULARY — vocabulary-only, no Python dep, no Copier runtime adopted; functional template test pattern named «template-render audit» per Copier's «answers.yml replay» analog).
…b-wave 7.1.c)
Extract REQUIRED_HEADER_DOCS + check logic from 09-doc-authority-hierarchy.test.ts
into a new sibling module 09-doc-authority-hierarchy.ts with stable exported API:
export const REQUIRED_HEADER_DOCS: readonly string[]
export const EXEMPT_PATTERNS: readonly RegExp[]
export function isExempt(relPath: string): boolean
export function hasAuthorityHeader(content: string): boolean
export function checkDocsHaveAuthorityHeader(
paths: string[], repoRoot?: string
): { ok: boolean; violations: Array<{ path: string; reason: string }> }
CLI shim 09-doc-authority-hierarchy.bin.ts:
- Reads paths from process.argv.slice(2)
- Calls checkDocsHaveAuthorityHeader(paths)
- Prints violations to stderr; exits non-zero on !ok
- Invokable via: npx tsx packages/core/principles/09-doc-authority-hierarchy.bin.ts <paths>
- Follows project's existing tsx-based bin pattern (matches rules-as-tests-detect etc.)
Test file (09-doc-authority-hierarchy.test.ts):
- Now imports from new module instead of inlining logic
- No behavioural change to existing 10 tests; 4 new API smoke tests added (14 total)
- New tests: ok:true for valid path, violation for missing file, EXEMPT_PATTERNS export,
mutation: exemption suppresses violation in changed-files mode
Stable API consumed downstream by sub-waves 7.2.c (PostToolUse) and 7.4
(make validate-prompts). repoRoot parameter defaults to process.cwd() so
both hook contexts (repo root) and test context (REPO_ROOT) work without
additional configuration.
Capability commit: NO (refactor of existing principle 09 to expose changed-files
mode for pre-push / harness-hook integration; no new packages/ subdirectory
or ≥80 LOC new file; status=M on test + status=A on extraction module).
Prior-art: skipped — refactor of existing principle 09 to expose changed-files mode for pre-push / harness-hook integration (consumed by sub-waves 7.2.c and 7.4); no new capability shipped.
…y header check Add .claude/hooks/check-doc-authority.sh (29 LOC) and second PostToolUse Edit|Write hook entry in .claude/settings.json. The hook reads the edited file path from stdin JSON, converts it to a repo-root-relative path (REQUIRED_HEADER_DOCS API expects relative), then invokes the Batch A CLI shim via node_modules/.bin/tsx. The CLI filters to REQUIRED_HEADER_DOCS — exits 0 for unrecognised paths. DECISIONS: - Batch A's 09-doc-authority-hierarchy.bin.ts WAS visible at smoke-test time (landed in the working tree at 10:15 on 2026-05-11). Test path: README.md → "OK 1 path(s) checked". Batch A is not yet on a merged commit; recorded as known runtime dependency in case merge ordering changes. - Path conversion: ABS_PATH stripped of REPO_ROOT prefix to obtain relative path, since checkDocsHaveAuthorityHeader() resolves relative to process.cwd() (= REPO_ROOT when called by npx tsx from repo root). - Capability commit: NO (≤30 LOC wrapper + settings.json hook entry; reuses Batch A capability; no new packages/ subdirectory). Prior-art: skipped — config-layer hook entry + ≤30 LOC wrapper invoking principle 09 CLI from sub-wave 7.1.c (Batch A). No new capability shipped; reuses changed-files mode from 7.1.c.
…+7.3.d)
Wire install.sh invocation in template-render.audit.ts beforeAll: spawnSync
against ts-server + react-next fixture skeletons, each in a fresh os.tmpdir()
workdir. Implements three deterministic probes per Decision 3 (no LLM in CI):
P1 — goal-phrase rendered (AGENTS.md.template lines 13/57):
Synonym list: «enforced by lint, tests, CI» | «without a measurable check».
Source: packages/core/templates/shared/AGENTS.md.template (documented paraphrase
variants that actually appear in rendered output; verbatim README phrase absent).
P4 — Authoritative-for header (duplicated from principle 09; authority header RE).
Checks rendered AGENTS.md carries "> **Authoritative for:**" header.
TODO: import from principle 09 shared module once Batch A lands.
P6 — taxonomy fidelity: every framework skills/ entry exists in consumer
.claude/skills/ after install. Catches silent rename drift.
Each probe includes an anti-tautology mutation test (bare content → probe fails),
mirroring the negative-test discipline in audit-ai-docs.test.sh.
P2/P3/P5 (LLM-judge) intentionally NOT shipped per Decision 3 + «no-paid» /
«LLM-then-cache» invariant. They live in .claude/skills/template-audit/SKILL.md.
Decision log:
- Synonym list source: AGENTS.md.template lines 13+57 (verbatim, not paraphrased)
- REQUIRED_HEADER_DOCS: duplicated inline (AUTHORITY_HEADER_RE) with TODO to
consolidate post-Batch-A-merge; full rendered-path mapping deferred to that
consolidation to avoid duplicating a 30-entry list
- install.sh invocation: spawnSync, 60s per-stack timeout
Prior-art: skipped — extends template-render.audit.ts from commit 1 with install.sh wiring + P1/P4/P6 deterministic probes. No new packages/ subdirectory or ≥80 LOC new file. P2/P3/P5 LLM-judge probes intentionally NOT shipped in CI per Decision 3 (two-mode: deterministic CI / advisory local).
…arch Prior-art: skipped — research patch for §13.23 4th-layer closure, no new capability shipped.
…ave 7.1.d)
Wave 6 D-3 SHIP-B independent track folded as cross-ref in sub-wave 7.1.d.
D-3 probe added to packages/core/audit-self/audit-ai-docs.sh:
- Checks that canonical goal phrase from README.md §Goal appears verbatim
(or via synonym) in enumerated downstream goal-bearing docs.
- Canonical phrase: "AI agents can't silently bypass undocumented conventions"
- Synonym: "AI cannot silently bypass what fails CI"
- Downstream docs checked (explicit enumeration, not glob):
.claude/session-bootstrap.md — operational restatement; must carry goal
CLAUDE.md — AI-tooling conventions; must carry goal pointer
- Probe ID: D3; selectable via --only=D3 flag.
- Smoke-run result: PASS (both downstream docs carry the canonical phrase).
Context7 sweep (≥3 phrasings, 2026-05-11; no production analog surfaced):
1. «doc-vs-doc parity / drift detection» → Drift SDK (DeFi), anomaly
detection resources — no doc-parity tool surfaced.
2. «AI documentation paraphrase drift template rendered output» →
Documentation.AI (documentation platform), Mistral/Google Document AI
(OCR/extraction) — different capability; no parity checker.
3. «documentation template consistency downstream parity» → Doc Detective
(tests docs against product behaviour, not doc-vs-doc phrase parity),
NgDoc (Angular docs generator) — different domain.
Verdict: BUILD — hand-rolled probe, no production analog.
SSOT entry #16 appended to docs/meta-factory/prior-art-evaluations.md:
Candidate: goal-phrase parity check (context7 sweep 2026-05-11)
Verdict: BUILD
Trigger to revisit: dedicated doc-parity tool surfaces; OR 7.3.d P1 probe
generalises to cover this case; re-evaluate at 7.5.b.
Also fixes MD040 violations in co-staged files (caught by new pre-commit hook):
- docs/meta-factory/prior-art-evaluations.md:57 — entry template fenced block
- docs/meta-factory/self-application.md:70 — ASCII path diagram
- INSTALL-FOR-AI.md:45,104,130,181 — copy-paste prompt / output blocks
All fixed by adding "text" language tag. Hook working as designed.
Capability commit: NO (status=M on existing audit-ai-docs.sh — not status=A;
no new packages/ subdirectory or ≥80 LOC file per Wave 6 MAJOR-1 correction).
Prior-art: skipped — extension of existing audit-ai-docs.sh with D-3 goal-phrase parity probe; no new packages/ subdirectory or ≥80 LOC file. Context7 sweep performed: (1) doc-vs-doc parity / drift detection; (2) AI documentation paraphrase drift template rendered output; (3) documentation template consistency downstream parity — no production analog surfaced. SSOT entry #16 appended with verdict BUILD.
Add .github/workflows/framework-self-template-render.yml: new CI job that runs npm --prefix packages/core run test:template-render (vitest .audit.ts harness from commits 1+2). Also adds test:template-render script to packages/core/package.json (mirrors test:principles pattern). PARTIAL dim 4 write-time fixup: Local runtime measurement: 1.3s on Apple M-series / Node 22 / macOS 25.4. Target ≤60s; alarm if CI runtime >120s on baseline ubuntu-latest runner. If CI runtime exceeds target consistently — surface as research-patch on fixture-size vs probe-coverage trade-off. Deterministic-only gate: job env contains NO AI-vendor secrets. No ANTHROPIC_API_KEY, no curl/fetch to Anthropic or OpenAI. Enforces «no-paid» / «LLM-then-cache» invariant. LLM-judge probes (P2/P3/P5) are session-bound in template-audit skill. Decision log: - New workflow file (not appended to audit-self.yml): cleaner review scope - Job pin: actions/checkout v4.2.2 + actions/setup-node v4.4.0 (mirrors audit-self.yml) - Node 20 (mirrors audit-self.yml baseline) Prior-art: skipped — CI workflow addition wiring framework-self-template-render job invoking template-render.audit.ts from commits 1+2. Deterministic-only per Decision 3 + «no-paid» / «LLM-then-cache» invariant. No new packages/ subdirectory or ≥80 LOC file.
Add .claude/skills/template-audit/SKILL.md — 42-line session-bound advisory
skill for template render QA beyond the deterministic CI gate.
Two-step procedure:
Step 1: runs npm --prefix packages/core run test:template-render (deterministic
CI-equivalent P1/P4/P6 probes from commit 1+2). Stop on failure.
Step 2: asks the current Claude Code session (no API call, FREE under subscription)
to semantic-check rendered content for P2/P3/P5 advisory findings:
P2 — paraphrase fidelity (goal-phrase framing not drifted to «guidelines»)
P3 — cue placement (session-bootstrap cue in first 10 lines of AGENTS.md)
P5 — synonym coverage (P1 list still matches current template phrasing)
NOT a CLI, NOT a CI gate, NO API key per «LLM-then-cache» invariant. Session-bound
under Claude Code subscription. Findings are advisory (not blocking).
Auto-trigger keywords: template, audit, render, generated docs, AGENTS.md,
paraphrase, cue placement, local advisory, template-render, audit-template.
Promotion trigger: P2/P3/P5 → CI gate when deterministic PASS rate <80% over
30 days OR consumer goal-phrase miss-detection report fires (per Decision 3
re-evaluation, open-questions.md §13.27).
Decision log:
- Skill location: .claude/skills/ (same as self-reflection, the existing pattern)
- Slash command /audit-template: deferred — .claude/commands/ does not exist in
this repo; no convention to follow; skill auto-trigger keywords are sufficient
- Trigger-keyword list final: 10 keywords covering all entry surfaces
Prior-art: skipped — local advisory skill ≤50 LOC, session-bound under Claude Code subscription, no CLI, no API key per «LLM-then-cache» invariant. Promotion to CI gated by Decision 3 re-evaluation trigger (deterministic-only PASS rate <80% over 30 days OR consumer goal-phrase miss-detection report).
Add verification row to INSTALL-FOR-AI.md checklist: "Harness hooks active (Claude Code only) | jq .hooks .claude/settings.json | UserPromptSubmit + PostToolUse entries present" Companion to 7.2.a/b/c hook files already committed. Consumer can verify harness-hook layer wired correctly post-install with a single jq call. Prior-art: skipped — doc-only addition to INSTALL-FOR-AI.md (one checklist row); no new capability shipped. Editor-coupling section and self-application.md preview row landed in 2b0a505 as co-staged with 7.1.d. Full §3 matrix expansion lands in sub-wave 7.5.a.
…e 7.4.a) Appends `validate-prompts` to existing Makefile. Walks .claude/orchestrator-prompts/**/*.md (excluding README.md) and runs validate-batch-spec.ts on each file; exits non-zero on first failure. Closes §13.28 part A (operator-side Makefile discipline layer). Prior-art: skipped — operator-side make-target wrapping existing validate-batch-spec.ts validator on .claude/orchestrator-prompts/**/*.md; no new packages/ subdirectory or ≥80 LOC file. Closes §13.28 part A (Makefile discipline layer).
…ew verdicts Prior-art: skipped — review verdict patch for §13.23 4th-layer, no new capability shipped
…b-wave 7.4.b) Adds README.md with Authoritative-for header per .claude/rules/doc-authority-hierarchy.md §5 (folder-level authority pattern). Declares append-only status and validate-prompts validation channel. Updates .gitignore: replaces directory-level ignore with file-glob pattern (.claude/orchestrator-prompts/*) + explicit negation for README.md so the folder README is tracked while individual prompt files stay gitignored. Subdirectory prompt files remain ignored (subdirectory entries are ignored by the * rule, so git does not descend into them). Prior-art: skipped — doc-only README applying folder-level authority pattern per .claude/rules/doc-authority-hierarchy.md §5; no new capability shipped.
Independent reviewer audit of sub-waves 7.1 (Batch A) + 7.2 (Batch B) + 7.3 (Batch C) + 7.6.a (Batch F.a) against orchestrator-kickoff.md spec. Domain scores: 1. Commit hygiene WARN (2b0a505 co-staging; SSOT #16 pre-land) 2. Hard constraints PASS 3. 7.1 hot-check prims PASS 4. 7.2 harness-hook layer WARN (settings.json UserPromptSubmit nesting) 5. 7.3 template test WARN (P1 synonyms are methodology phrases) 6. 7.6.a research patch PASS 7. Cross-batch consistency PASS (typecheck clean) 8. Capability compliance PASS (80ef1d9 + f528586 Prior-art trailers OK) Overall: GO WITH NOTES — 3 WARN items (Issues 2+4 must verify pre-PR; Issue 5 carry-forward to 7.5.d). 3 NOTE items for orchestrator. ATTN: 7.6.b already on branch with GO [F1] verdict — 7.6.c is unblocked. Prior-art: skipped — review verdict patch for Wave 7 Round 1, no new capability shipped.
…7 review) The PostToolUse hook for Authoritative-for header check was firing on every Edit/Write — including .ts/.json files that are not authority-bearing docs. Root cause: `09-doc-authority-hierarchy.bin.ts` passed argv paths to `checkDocsHaveAuthorityHeader` without filtering. Only `EXEMPT_PATTERNS` were skipped inside the function; any non-doc file outside that pattern produced a violation, surfacing as `FAIL …: missing header` on stderr. Fix: filter argv to `REQUIRED_HEADER_DOCS ∪ EXEMPT_PATTERNS` BEFORE calling the check. Empty filtered list → silent exit 0. Three new tests cover (a) non-required path silent, (b) required-without-header FAIL, (c) exempt silent. Bonus: OK-message rewritten so exempt-only paths no longer report «all have header» (review §2 m5). Discovered by cold-start reviewer (2026-05-11) via empirical hook smoke-test with non-doc path; previous round audit (Domain 4) declared «hook scripts all correct» without running this probe — gap captured in reviewer-checklist amendment (Commit 3 of this batch). Prior-art: skipped — fix M1 from Wave 7 Round 1 review (non-doc paths leak through CLI shim); no new packages/ subdirectory, no new file ≥80 LOC, no behaviour-extension beyond regression fix.
…#22 Cookiecutter/Copier (Wave 7 review) Two prior Wave 7 commits cite SSOT entries that did not yet exist: 80ef1d9 (sub-wave 7.1.a markdownlint, capability=YES) → #17 f528586 (sub-wave 7.3.a/b template-render harness, capability=YES) → #22 The kickoff (§Sub-wave 7.5.b) planned SSOT entries #17–#22 to land batch-style at Wave closeout, but CLAUDE.md's build-vs-reuse invariant mandates: «add a new SSOT entry … in the same commit as the capability artifact.» CLAUDE.md wins; the kickoff plan was wrong for #17 and #22 (which are positively cited from prior commits). For #18–#21 (lychee, vale, Claude-Code-hooks, cross-editor scripts), the corresponding sub-wave commits used `Prior-art: skipped — …` escape-hatch trailers rather than positive citations, so they are NOT in violation — the kickoff plan for those four remains intact and they land at 7.5.b alongside their verdict transitions. Resolution: forward-add (not amend / not force-push) — the cited trailers in 80ef1d9 and f528586 now point to existing rows. CLAUDE.md «prefer create new commit than amending» is honoured. The deviation was caught by Wave 7 Round 1 cold-start review (2026-05-11); prior round audit Domain 8 reported «PASS» by verifying trailer presence without checking SSOT existence — gap captured in reviewer-checklist amendment (Commit 3 of this batch). Prior-art: skipped — forward-fix M2 from Wave 7 Round 1 review (SSOT entries #17 + #22 added in same atomic commit per CLAUDE.md same-commit rule; supersedes the deferred-to-7.5.b plan in kickoff for these two entries only; entries #18-#21 land at 7.5.b as planned because their corresponding commits did not cite missing IDs); no new capability shipped.
…test additions Two corrections to the Round 1 audit (47dc9ed, docs/meta-factory/research-patches/2026-05-11-wave-7-round1-review.md): - Issue 4 (settings.json UserPromptSubmit nesting): DEBUNKED — hook fires empirically (this very session's UserPromptSubmit digest is proof). - Domain 8 (Capability compliance) + Domain 4 (Hook scripts): both re-classified PASS → WARN — method gaps (trailer-presence-only vs SSOT existence; no empirical hook smoke-test). Forward-fixed by Commits 1+2 of this batch. The phase-research-coverage rule (.claude/rules/phase-research-coverage.md) gets two new §1 checklist items so future review sessions catch these gaps proactively: §1.8 Hook surface smoke-test (PostToolUse / pre-push) with path outside hook's intended filter → expect exit 0 §1.9 SSOT citation existence-check: `grep` for #N in prior-art- evaluations.md, not just trailer presence Both items are derived from gaps that cold-start review surfaced 2026-05-11 but the prior round audit missed. Rationale for landing in the rule file (vs a new reviewer skill): the file already owns §1 adversarial-check discipline; the two smoke-tests are concrete executions of that discipline for hook-touching / capability-commit changes. Prior-art: skipped — doc-only errata + skill smoke-test addition; no new capability, no new packages/ subdirectory, no new file ≥80 LOC.
…+ §13.23 self-review Updates self-reflection skill (.claude/skills/self-reflection/SKILL.md) by adding the §1.7 enforcement layers table: 3 active + 1 deferred → 4 active. The 4th layer (pre-push trailer check) shipped in Commit 1 of this batch. Ships paired self-review patch at docs/meta-factory/research-patches/2026-05-11-§13.23-4th-layer-self-review.md per §13.23 promotion path #4. The patch walks the new layer through §1.7 forward+backward checks applied to its own motivating gap (local-push- bypasses-CI + discipline-theatre risk). §13.23 status update in open-questions.md folds into sub-wave 7.5.c per kickoff §«Sub-wave 7.6.d» (atomic with §13.27 + §13.28 closures); NOT modified in this commit. §1.7 self-application: this commit touches .claude/skills/self-reflection/ SKILL.md AND adds new ## §1.7 enforcement layers section heading. The C4 predicate fires on this commit. The §1.7 check shipped in Commit 1 of this batch (one commit earlier in the push range) and is in warn-only calibration mode — B1 bootstrap trailer below satisfies §1.7 for this commit. Prior-art: skipped — doc-only update to self-reflection skill description + self-review patch in research-patches/; no new packages/ subdirectory, no new file ≥80 LOC. §1.7 Bootstrap: updating skill ladder from 3 active + 1 deferred → 4 active layers and shipping self-review patch immediately after introducing the §1.7 enforcement layer — §1.7 check itself is in 30-day calibration (warn-only) at this commit; B1 exemption per 7.6.a research §4 + 7.6.b §2 Problem 2 verdict (mirrors Commit 1 bootstrap exemption; both commits are in the same push, same bootstrap context).
…ers + 1 meta-rule) Wave 7 closeout matrix update — appends 9 rows to self-application.md §3 Decision matrix Phase 1. Each row cites §13.8 4-criteria gate verdict (failure-cost / local-cost / detectability / lifecycle stage). 1. markdownlint-cli2 structural — MUST (pre-commit; failure-cost MEDIUM, local-cost LOW, detectability MEDIUM, lifecycle 1) 2. lychee offline link check — SHOULD (pre-push; MEDIUM/LOW/HIGH/2) 3. Harness-hook UserPromptSubmit+PostToolUse — SHOULD/MAY (real-time; HIGH/LOW/MEDIUM/5; full row replacing preview) 4. Functional template-render audit — MUST in CI (deterministic; HIGH/MEDIUM/HIGH/3) 5. Local advisory template-audit skill — MAY (session-bound; LOW/ZERO/HIGH/1) 6. Operator-side make validate-prompts — SHOULD (Makefile; MEDIUM/LOW/HIGH/1) 7. §13.23 4th-layer pre-push trailer check — SHOULD→MUST (warn-only 30-day calibration; HIGH/LOW/MEDIUM/2) 8. (meta-rule) No direct LLM API calls in CI — MUST (architectural; HIGH/ZERO/LOW/3) — codifies «no-paid» invariant per LLM-usage audit 9. Goal-phrase parity audit — SHOULD (pre-push scoped; MEDIUM/LOW/MEDIUM/2) Row 3 (harness-hook) replaces the line-62 preview row with full rationale + promote-to-MUST trigger. §1.7 forward+backward check applied. Forward: 9 rows comply with R1-R20, principles 01-09, Prior-art trailer, doc-authority — no contradictions surfaced. Backward sweep: each row mapped against existing matrix rows 53-61; complementary additions or expansions, no conflicts. §13.8 «Decision matrix expansion mechanism» dogfooded (9 rows added in one atomic commit per the §13.8 4-criteria gate format) — updated in 7.5.c. Prior-art: skipped — doc-only matrix expansion enumerating shipped Wave 7 layers, no new capability or new file ≥80 LOC.
… verdict transitions) Append 4 SSOT entries to docs/meta-factory/prior-art-evaluations.md per Wave 7 closeout (§3 «Adding a new entry» step 1): #18 Vale — DEFER (Russian+EN mixed prose FP risk; promote when 2nd doc- prose incident markdownlint cannot catch) #19 lychee — ADOPT (transition from ADOPT WHEN TRIGGERED; shipped 7.1.b) #20 Claude Code hooks — ADOPT (UserPromptSubmit + PostToolUse; shipped 7.2.a/b/c; transition from ADOPT WHEN TRIGGERED) #21 Cursor/Cline/Continue.dev reference scripts — WATCHLIST (Claude Code only project-side per Decision 4; cross-editor parity deferred) NOTE on numbering (entries #16, #17, #22, #23-#26 landed in earlier commits per CLAUDE.md same-commit-as-capability rule): #16 2b0a505 (sub-wave 7.1.d D-3 BUILD verdict) #17, #22 Batch G «feat(prior-art): M2 …» (forward-fix from Round 1 cold-start review M2 finding) #23-#26 Batch F.c Commit 1 «feat(hooks): Wave 7 7.6.c — §1.7 trailer check + SSOT #23-#26» (atomic SSOT landing per CLAUDE.md; precedent 7.1.d #16; cited #23 + #26 positive trailer) The kickoff §«Sub-wave 7.5.b» plan originally targeted all #16-#22 at 7.5.b — corrected via Round 1 review M2 finding adjudication + Round 2/3 user patch confirming atomic #23-#26 landing in F.c Commit 1. Prior-art: skipped — appending parent-research SSOT entries at Wave 7 closeout per §3 «Adding a new entry» step 1; no new capability, no new packages/ subdirectory.
…update §13.8 Atomic closure of three armed §13.x triggers and update of one running trigger per Wave 7 closeout: §13.23 (4th-layer pre-push trailer check) — CLOSED. Implementation shipped in Batch F.c Commit 1 (.husky/pre-push section 9). Status: "deferred 2026-05-09 → closed by Wave 7 sub-wave 7.6, 2026-05-11". §13.27 (functional template test) — CLOSED. Implementation shipped in sub-wave 7.3 (deterministic CI gate framework-self-template-render) + 7.3.f (local advisory skill template-audit). LLM-judge re-evaluation trigger added per Decision 3: promote when deterministic PASS rate <80% over 30 consecutive days OR Claude Code subscription expands to CI compute. §13.28 (operator-side discipline) — CLOSED. Implementation shipped in sub-wave 7.4.a (Makefile target validate-prompts) + 7.2.b (PostToolUse harness-hook running validate-batch-spec.ts on .claude/orchestrator- prompts/**/*.md). Both A and B per Decision 5. §13.8 (Decision-matrix expansion mechanism) — UPDATED (status remains open as the running mechanism trigger). Wave 7 sub-wave 7.5.a dogfooded the mechanism — 9 rows added citing 4-criteria gate verdict per row. Also fixes 2 pre-existing MD040 (fenced code without language) violations at lines 18 and 192 (unlabeled code blocks now use ```text). Prior-art: skipped — open-questions status updates per Wave 7 closure; no new capability, no new packages/ subdirectory.
…tim README synonym
Two artefacts closing Wave 7 self-review:
1. New research patch 2026-05-11-wave-7-decision-matrix-self-review.md —
§1.7 backward sweep applied to the 9 new Decision matrix rows landed
in sub-wave 7.5.a (Commit 1 of this batch). Walks 9 rows × R1-R20 +
principles 01-09; surfaces complementary / extension / no-overlap
verdict per pair. Non-trivial pairs called out:
- row 7 (§1.7 trailer) × P08 (prior-art cited): different keys +
different hooks, coexistence proven by Batch F.c
- row 4 (template-render audit) × P09 (doc-authority): static vs
install-time rendering — complementary layers
- row 1 (markdownlint structural) × 500-line limit: additive checks,
different enforcement mechanisms, no duplication
All 9 rows verdict: complementary to existing R1-R20 + P01-P09.
Ratifies sweep cadence: any Decision matrix expansion ≥3 rows triggers
a self-review patch (mirrors §13.8 4-criteria gate dogfooding).
2. m1 fix per Round 1 cold-start review (2026-05-11):
packages/core/audit-self/template-render.audit.ts — GOAL_PHRASE_SYNONYMS
now includes verbatim README#why-this-exists phrase as third synonym.
Drift detection now covers three independent expressions: two paraphrase
variants currently in templates + verbatim README for forward compat.
All 9 template-render vitest tests still PASS.
Subject prefix `docs(research-patches):` triggers D3 §1.7 allow-list
bypass (per .husky/pre-push section 9 v1) — no §1.7 trailer required
because the patch is doc-only and the m1 fix is an audit-config tweak.
Prior-art: skipped — self-review patch + audit synonym addition per Wave 7 closeout; no new capability, no new file ≥80 LOC.
|
Review the following changes in direct dependencies. Learn more about Socket for GitHub.
|
|
Warning Review the following alerts detected in dependencies. According to your organization's Security Policy, it is recommended to resolve "Warn" alerts. Learn more about Socket for GitHub.
|
…(CI fix) Removes duplicative DECISIONS log section (lines 492-497) — the same decisions are documented in body sections §0 (context7 sweep), §5 (synthesis), §6 (SSOT), and §Final (bootstrap path). Net delta: −7 lines. Post-trim: 498 lines, comfortably under the audit-self 500-line invariant. Triggered by audit-self / Mechanical checks failure on PR #29. Prior-art: skipped — CI fix removing duplicative content from research patch; no new capability, no new packages/ subdirectory.
4 tasks
artyhoo
added a commit
that referenced
this pull request
May 11, 2026
…verlap analysis Adds three entries from the aif-handoff-overlap-analysis mandate (Stage 3, Phase A). Entries reflect verdicts from Stage 1 research + Stage 2 review: - #27 DEFER: AIF Handoff HANDOFF_MODE env-var fork (C1×P4 ORT — different problem classes; DEFER pending non-interactive orchestration pipeline) - #28 DEFER: AIF Handoff paused:true/false semantic (S1×P5 ORT — different granularities; DEFER pending wave state machine formalisation Phase 11+) - #29 ADAPT: AIF first-line plan annotation pattern (A1×P6 OVL; ADAPT as <!-- scope:§N --> for research patches + automated trigger sweeps §1.6) Phase B (ADAPT implementation: annotations + CI gate) follows in separate commit, citing entry #29 as Prior-art trailer. Prior-art: skipped — adding SSOT entries to prior-art-evaluations.md; no new capability artifact, no new file under packages/, no new package.json dep.
artyhoo
added a commit
that referenced
this pull request
May 11, 2026
…-> (SSOT #29) Implements ADAPT verdict from aif-handoff-overlap-analysis mandate (Stage 3). Adapts AIF handoff:task first-line plan annotation pattern (SSOT #29) for research patches as <!-- scope:<slug> --> machine-parseable first-line marker. Changes: - 20 research patches: added <!-- scope:<slug> --> as first line per canonical mapping (§13.21, §13.23, §13.25, wave-7, phase-8.8, methodology, aif-handoff-mandate) - packages/core/principles/10-research-patch-annotation.test.ts: companion principle test with anti-tautology mutation check (Vitest, 5 tests green) - .github/workflows/audit-self.yml: CI gate verifying annotation presence on all research patches (except README.md), added to mechanical job - 2026-05-10-wave-6-review-verdicts.md, 2026-05-10-wave-7-review-verdicts.md: fix pre-existing MD040 bare fences (language=text) uncovered by pre-commit hook Enables automated §1.6 trigger sweep: grep "scope:§13.X" across patches instead of prose-level pattern matching on filenames. §1.7 backward check: all 20 existing patches annotated atomically with CI gate in this commit (complete sweep, not §1.5 floor). Prior-art: prior-art-evaluations.md#29 (AIF handoff:task first-line plan annotation ADAPT — <!-- scope:slug --> for research patches, automated §1.6 sweep support, verdict ADAPT 2026-05-11).
5 tasks
This was referenced May 11, 2026
Merged
This was referenced May 22, 2026
artyhoo
added a commit
that referenced
this pull request
May 22, 2026
…l disposition (#162) Completes the DN-4 sweep: - CONTRIBUTING.md «Working in a git worktree» (#29 worktree_node_modules_symlink) - phase-research-coverage.md §1.13 — AI-doc research source priority (#30) - tracker: #29/#30 → CODIFIED; #18/#19 → DEFERRED (reference-facts); #21 → DEFERRED (thesis-level); #24 → OUT-OF-REPO (global orchestrator skill) DN-4 now fully swept: 11/15 codified, 4 dispositioned with reasons, zero bare-PENDING. §1.7 self-review at research-patches/2026-05-22-dn4-round3-tail-codification.md. Prior-art: skipped — prose discipline codification + tracker disposition of existing memory conventions, no new capability or dependency
This was referenced May 26, 2026
artyhoo
added a commit
that referenced
this pull request
May 29, 2026
… Implementer-equivalent only) value-add audit (#276) R-phase patch for Sub-wave C of the aif-handoff-as-runtime-bridge umbrella. Evaluates Variant C (kickoff §3 lines 124-145): aif-handoff as Implementer- equivalent only, bypass Planner+Reviewer cycle, thin CLI wrapper for kickoff dispatch + kanban status tracking. Verdict: REJECT (BFR-default §1 ladder). Rationale: - The kickoff-framed "aif-handoff exec --kickoff <path>" CLI does not exist in lee-to/aif-handoff (DeepWiki probes 1+5, 2026-05-29). - No first-class Implementer-only mode; skipReview:true bypasses Reviewer but Planner is mandatory unless accept_existing_plan with on-disk PLAN.md (same disk coupling SW-A flagged for Variant A). - BEFORE/AFTER maintainer-action count: 25% literal / 0% cognitive reduction (T-AIF-BRIDGE-C table §4) — below kickoff §8 STOP 30% threshold → verdict "Variant C value-add insufficient". - Pure-tracker pattern (paused:true + autoMode:false + manual state-machine transitions) IS shipped but adds zero automation beyond UI tracking; Docker+SQLite infra unjustified. Cites: - SW-A merged PR #268 (Variant A REFERENCE, 28% match, 3 ADOPT-blockers) - SW-B merged PR #267 (Variant B REJECT, ~5% match, no dir-watch capability) - PR #269 follow-up (mechanical corrections, no verdict changes) - DN-1=B-constrained input consumed in criterion 5 (mooted for Variant C which bypasses aif-handoff Reviewer entirely) - Gate-4 admission re-sweep: PR #127+#128 touch packages/runtime/ only (no MCP/coordinator drift in 30-day window) 5 distinct DeepWiki probes + 2 WebSearches + cross-ref to SW-A/SW-B/PR #269 = 19+ evidence channels (T1 floor exceeded 3.8x). §1.7 forward+backward + §self-application + T-trap walk per ai-laziness-traps.md §3. Single output file under docs/meta-factory/research-patches/. No code, skill, agent, install.sh, or .claude/rules/ modifications. ### §1.7 Forward-check applied build-first-reuse-default.md §1 verdict ladder applied; BFR §3 6-layer search performed (SSOT rows #27/#28/#29/#30/#43/#44/#46/#67/#80 reviewed at prior-art-evaluations.md:95-148; DeepWiki >=5 probes; WebSearch >=2 phrasings; own-stack sweep at .claude/skills/meta-orchestrator/SKILL.md:441 anti-scope + :404+:429 SP requesting-code-review). no-paid-llm-in-ci.md §1 enforced (all evidence via subscription-bundled DeepWiki/WebSearch + free gh CLI + bash). reviewer-discipline.md §2 respected (DN-1=B-constrained consumed as fact, not re-litigated; verdict is research finding against §8 STOP, not strategy choice). ai-laziness-traps.md §3 active T-traps applied (T1, T3, T7, T11, T12, T13, T15, T16, T17, T19, T20, T-AIF-BRIDGE-C MANDATORY BEFORE/AFTER table at patch §4). Evidence: see patch §8 file:line citations. ### §1.7 Backward-check applied SSOT #27/#28/#67 receive additive notes (additive-only; no verdict changes). Original DEFER/DEFER/REJECT rationales reviewed at prior-art-evaluations.md: 95, 96, 135 — consistent with Sub-wave C findings (reinforce existing classifications, do not re-litigate). No .claude/rules/* modified; no .claude/skills/* modified; no agents/* modified; no packages/* modified; no install.sh modified; no kickoff.md modified. Single output file in docs/meta-factory/research-patches/. Scope strictly bounded to Variant C; SW-A/SW-B/SW-B2/SW-D out of scope. T15 self-application confirmed in patch §10. Memory not written (Sub-wave D synthesis is the natural codification surface). Evidence: see patch §9 file:line citations.
This was referenced May 31, 2026
artyhoo
added a commit
that referenced
this pull request
Jul 23, 2026
…+ population report (#1117) * docs(umbrella-donemd-backfill): Commit A — done.md for 12 CLOSED-VERIFIED candidates Stage 1 of umbrella-donemd-backfill. Closes 12 stale no-done.md dirs where closure is proven by either parent-umbrella's done.md naming the per-stage PR (3 cases) or parent-umbrella's final PR closing the meta-launch runtime state (4 cases) or direct branch-name PR match where the branch IS the slug (5 cases). Per-candidate evidence (full matrix + population enumeration in report.md landed in Commit C): agnosticism-remediation-t-a/b/c (parent done.md names per-stage PRs): - T-A #570 — CC-qualify skill auto-activation + doc-claims probe - T-B #575 — portable AGENTS.md rule index + rules-autoload coverage probe - T-C #577 — agnosticism-conformance principle test slot 21 Parent (agnosticism-remediation) closed by #589 (2026-06-16). *-meta-launch dirs whose parent umbrella is closed (meta-launch runtime state included in parent's final PR): - consumer-upgrade-path-meta-launch ← parent #615 - egress-host-push-default-meta-launch ← parent #760 - generator-require-composite-tier-meta-launch ← parent #708 - hook-test-suite-rot-meta-launch ← parent #608 Branch-name PR matches (branch IS slug; no kickoff.md to verify final-stage trap T-UDB-C against; dir is runtime batch state for the named work): - wave-5-trio-followup ← PR #37 (chore/wave-5-trio-followup, 2026-05-11) - wave-7-hot-checks-joint-closure ← PR #29 (wave-7-hot-checks-joint-closure, 2026-05-11) - wave-8-retro-h8-gate5-13-31 ← PR #48 (chore/wave-8-retro-h8-gate5-13-31, 2026-05-12) - phase-9-implementation ← PR #18 (docs/phase-9-implementation, 2026-05-09) - audit-fixes-2026-05 ← PR #1 (feat/audit-fixes-2026-05, 2026-05-07) Structural note (load-bearing, surfaced in report §1): all 12 target dirs lack kickoff.md (untracked + nothing on disk). priority-score.sh skips such dirs at line 144 (`if [[ ! -f "${kickoff}" ]]; then … continue`) — so writing done.md here is documentation, not a priority-score.sh panel-state change. The 65 no-done dirs with kickoff.md are all excluded by either the explicit live-program list or the 45-day live-signal rule, so Stage 1 yields zero priority-score.sh panel changes (correctly: there are no actionable closures in the live-program-with-kickoff population). Prior-art: skipped — refactor only, no new capability * docs(umbrella-donemd-backfill): Commit C — report.md + .gitignore exception Stage 1 report for umbrella-donemd-backfill. Adds the population enumeration, per-verdict counts, the full UNCLEAR table (22 rows with per-row signals), and an acceptance-gate self-check. Load-bearing structural finding in §1: 35 of 100 no-done.md dirs are runtime-only coordination slots (no kickoff.md on disk, no git history). priority-score.sh already filters these out at line 144 (kickoff-missing -> continue) BEFORE the Layer C3 done.md check at lines 222-231 runs, so the C3 layer never reaches them whether done.md exists or not. The 12 done.md files from Commit A add documentation value (visible closure marker) but do not change /pipeline panel output for those 12 -- they cover the 12 of 65 real-umbrella no-done.md dirs where closure was provable via parent-umbrella or branch-name PR evidence. Per-verdict counts (per §2 of the report): - CLOSED-VERIFIED: 12 (done.md landed in Commit A) - SUPERSEDED: 0 (none of the 34 candidates has kickoff.md -> S3 superseded-marker check is inapplicable) - OBSOLETE-CANDIDATE: 0 (no per-dir visibly-overtaken evidence possible without kickoff context) - UNCLEAR: 22 (weak/ambiguous signals; first-class outcome per T14) .gitignore: extends the orchestrator-prompts exception list with report.md parallel to done.md/kickoff.md. Same durable-committed class -- the report is an audit artefact, not runtime state. Acceptance gate (§7 of the report): all 6 checks pass. priority-score.sh runs clean (the new done.md files land in already-skipped dirs + the 12 real-umbrella closures). ZERO done.md files under exclusion-list dirs. UNCLEAR section present. Prior-art: skipped -- prose-only doc commit, no new capability. * docs(umbrella-donemd-backfill): expand report §7 with full gate evidence Task 6 acceptance gate is complete. Updates §7 to record all 6 gate checks explicitly + triages the 5 pre-existing done.md files in exclusion-list dirs (multi-model-pipeline-pilot, launch-preannounce-track, python-delivery-v0, generation-live-delivery, meta-orchestrator-prior-art) as pre-existing state authored by other maintainers' past PRs, not leaks from this umbrella. `git log --author=Yhooi2 --since=2026-07-23` confirms zero of my commits today touched any of those 5 files. All 6 acceptance-gate checks pass: 1. priority-score.sh runs clean (exit 0, 367 lines, no parse breakage) 2. ZERO NEW done.md under exclusion-list dirs (5 pre-existing are not this umbrella's work) 3. Every done.md corresponds to a report row with cited PR + merge SHA 4. UNCLEAR section present with 22 rows (T14 compliance) 5. Population enumeration present with exact command + 100-dir count 6. Schema check passes on all 12 new done.md files Prior-art: skipped -- prose-only doc edit, no new capability.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Wave 7 joint closure ships 6 sub-waves of hot-checks, harness-hooks, functional template test, and §13.23 4th-layer pre-push enforcement. Closes 4 armed open-question triggers in one branch via atomic dogfooding of the §13.8 Decision-matrix expansion mechanism.
Path A confirmed — §13.23 4th-layer (
§1.7pre-push trailer check) ships warn-only under 30-day calibration. Total: 36 commits / 42 files (+3091 / -149).Sub-wave breakdown
validate-prompt+ PostToolUsecheck-doc-authorityframework-self-template-renderCI gate (deterministic P1/P4/P6 probes) + local advisorytemplate-auditskillmake validate-promptsMakefile target + folder-level authority README.husky/pre-push §9§1.7 trailer check (warn-only 30-day calibration) + SSOT #23-#26 atomic landingOpen-questions closed by Wave 7
framework-self-template-rendershipped (7.3)make validate-prompts) and B (PostToolUse) shipped (7.4 + 7.2.b)§1.7 enforcement ladder: 3 active + 1 deferred → 4 active layers (rule + skill + CI workflow + pre-push trailer check).
Decision matrix expansion (sub-wave 7.5.a, +9 rows)
Each row cites §13.8 4-criteria gate (failure-cost / local-cost / detectability / lifecycle-stage):
make validate-prompts— SHOULD§1.7 forward+backward sweep applied to all 9 rows — no contradictions with existing R1-R20 + principles 01-09 + Prior-art trailer discipline. Full sweep documented in
docs/meta-factory/research-patches/2026-05-11-wave-7-decision-matrix-self-review.md.SSOT register state (cumulative)
All entries land same-commit-as-capability per CLAUDE.md build-vs-reuse rule (precedent: 7.1.d/#16).
Round 1 cold-start review findings (2026-05-11)
Independent reviewer surfaced 2 MAJOR + 9 MINOR + 1 DEBUNKED. Closed in Round 3 «Batch G» fix bundle:
check-doc-authority.shleaked FAIL noise on non-doc paths (CLI shim no filter); fixed viaselectRequiredPaths()extraction in09-doc-authority-hierarchy.bin.ts(commit eab7c71)..claude/rules/phase-research-coverage.md§1.8 (hook smoke-test) + §1.9 (SSOT citation existence-check) added (commit 4ec69e5) so future review sessions catch these gaps.Recursive self-application (§1.7 sweep)
§1.7 Forward-check applied
The Wave 7 shipment introduces new disciplines (§1.7 4th-layer pre-push trailer enforcement, §1.8 hook surface smoke-test, §1.9 SSOT citation existence-check, 9 new Decision matrix rows including the «No LLM API in CI» architectural meta-rule). Forward-check verifies each new discipline complies with currently-active layers:
npm typecheck --workspacesPASS across 3 workspaces.npm --prefix packages/core run test:principlesPASS 51/51. Principle 08 (Prior-art cited) PASS — every capability commit (80ef1d9, f528586, 91f5f1d, 5c0d32e, 2e43874, 38dc50a, 71a7f00, 29e62d8, 5afabad, 8982fde, 2b0a505, 951d7f7, eab7c71) carries validPrior-art:trailer (positive citation or escape-hatch with ≥20-char rationale).> **Authoritative for:** …header where required. Principle 09 PASS — REQUIRED_HEADER_DOCS list compliance verified, EXEMPT_PATTERNS preserved.§1.7 Bootstrap:trailer per B1 exemption; F.c Commit 2 (e5ada1a, SKILL.md ladder 3→4) also carries B1 (C4 fires on SKILL.md edit per F.b §4 fixup Phase 4: Stack Detector v1 (read+write AIF bridge, CI gate, reviewer cycle complete) #4). All other commits either fall outside §1.7 scope predicate (file-glob mismatch or D3docs(research-patches):allow-list bypass) or land before the §1.7 layer ships (warn-only fires on pre-enforcement commits 4ec69e5 + 91f5f1d, exits 0).§1.7 Backward-check applied
Backward sweep applies new rules to existing artefacts under their scope. For Wave 7:
docs(research-patches):/chore(snapshot-regen):/chore(prior-art-update):prefixes); meta-test: positive (B1 trailer with ≥20-char rationale satisfies §1.7), mutation (removing B1 detection breaks bootstrap path).docs/meta-factory/research-patches/2026-05-11-wave-7-decision-matrix-self-review.md(caa65dc, 7.5.d self-review patch). Backward-check ratifies sweep cadence: any Decision matrix expansion ≥3 rows triggers a self-review patch.GOAL_PHRASE_SYNONYMSintemplate-render.audit.ts:42-45— extends coverage without removing existing methodology phrases;npm --prefix packages/core run test:template-renderstill PASS 9/9.Known limitations / expected behaviour
origin/main..HEADrange).S17_WARN_ONLY=falseper TODO comment in.husky/pre-push. 50-commit adversarial test FP rate = 0% (well below 5% threshold for permanent warn-only fallback)..claude/rules/phase-research-coverage.mdAuthoritative-for header (header still says «6-item» while body now has 9 items §1.1-§1.9) — cosmetic, pre-existing drift (was 6 vs 7 before Wave 7), deferred to follow-up cleanup or Wave 8.Hand-off to Wave 5 implementation
Wave 5 (project-aware tool bootstrapping) implementation has NOT yet started — Wave 7 first per 2026-05-10 sequence reversal. Wave 7 surfaces infrastructure Wave 5 should adopt:
framework-self-template-renderCI gate.make validate-prompts— Wave 5 orchestrator prompts get advisory validation channel.Test plan
npm --prefix packages/core run test:principles: 51/51 PASSnpm --prefix packages/core run test:template-render: 9/9 PASSnpm run typecheck --workspaces: 3 workspaces PASSbash -n .husky/pre-push: syntax OKmake validate-prompts: all files passed (self-reflexive on Wave 7 prompts themselves)make self-audit: PASS (5/5 audit probes)§1.7 Bootstrap:) on F.c commits grep-resolvablepackage.json/.tspaths now exit 0 silent through PostToolUse hook