From 2aa95636ee04a01457f5986fcb8cf21535582c66 Mon Sep 17 00:00:00 2001 From: Bastien Chanot Date: Fri, 17 Jul 2026 22:54:12 +0200 Subject: [PATCH] =?UTF-8?q?chore(memory):=20BDR-074=20BDR-075=20EVAL-026?= =?UTF-8?q?=20LRN-136=20=E2=80=94=20config-protection=20removal=20+=20plan?= =?UTF-8?q?-challenge=20phase?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .claude/memory/decisions.md | 6 ++++++ .claude/memory/evals.md | 3 +++ .claude/memory/journal.md | 3 +++ .claude/memory/learnings.md | 3 +++ 4 files changed, 15 insertions(+) diff --git a/.claude/memory/decisions.md b/.claude/memory/decisions.md index 97e8f1f..8ed46d4 100644 --- a/.claude/memory/decisions.md +++ b/.claude/memory/decisions.md @@ -1046,3 +1046,9 @@ rules: - **Why**: /harden had a real scale (SKILL.md:435), /seo had NONE → every axis felt → two runs over identical code diverged, while /client-handover gates on 17/20. H2 sharpened it: once drift reports real change, a self-moving score is visibly noise. Same principle as engine-side cannibalisation grouping — never hand a model 1000 rows to add. - **Makes computable (was prose)**: "N/A is not a zero" (R2 on-page, I1 off-page) → axis excluded + weights renormalised, verified all-20 with 2 N/A → global 20.0. Prevalence: affected/sampled shift severity ONE step (≥50% escalate, single de-escalate). - **Files**: lib/seo-data/score.py (4818c61). + +### BDR-074 — Remove config-protection edit-block guardrail [accepted] (2026-07-17) +Deleted hooks/config-protection.sh + its settings.json PreToolUse registration + lib/tests/config-protection.test.sh. Hook blocked model Edit/Write on quality-gate files (settings.json, gitflow.sh, .githooks, doctor.sh, hooks, lib/tests, lint) via one-shot .claude/.config-edit-ok sentinel. Removed per user req — friction editing own config > guardrail value; user = human operator. Residual: gitflow pre-commit guard + Gitea branch protection still block direct code commits main/develop; only edit-time block gone. Alts rejected: warn-only (exit0+log), targeted relaxation. Supersedes any prior config-protection decision. + +### BDR-075 — Framework-wide 3-way adversarial plan-challenge phase [accepted] (2026-07-17) +After a plan/reflection elaborated + before execution, 3 fresh blind sub-agents (correctness/robustness/simplicity) attack it; main loop RE-THINKS every aspect a BLOCKER lands (named change or [deferred]) + re-challenges once if plan materially changed. Reusable lib/challenge-plan.md + new agents/plan-challenger.md (read-only, big-model per [[BDR-066]] — audit judgment, NOT sonnet). Fail-safe (never fail open: mute→retry→escalate), severity-driven (any single-lens BLOCKER=must-address, NOT consensus — lenses orthogonal), advisory into existing human gate. KIND tunes lenses: build-plan/proposals/fix-bundle. Wired 11 orchestrators: ship-feature/init-project/feat/bugfix + onboard/audit-delta/code-clean + seo/geo/harden/web-validate. Excluded (no real plan): hotfix/tour/analyze/client-handover/release-candidate/spec. Audit found 0 repo-owned plan-challengers pre-existing (only vendored gstack autoplan, sequential+unwired). See [[EVAL-026]]. diff --git a/.claude/memory/evals.md b/.claude/memory/evals.md index 73a677b..87905cc 100644 --- a/.claude/memory/evals.md +++ b/.claude/memory/evals.md @@ -248,3 +248,6 @@ rules: - **method**: each verifiable claim confronted DURING execution with a primary source or a live test — CrUX API metric list, web.dev, Search Console API reference, HEAD on data.commoncrawl.org, real curl on 2 live sites (zenquality Astro, lavageangels356 native PHP), 2 real repos. - **anomalies**: 7/7 of the verifiable claims were false or overstated (VSI exists / Off-page zero-data / stats drive weights / GSC Links API / SPA §0 flag / Twitter 403 / Common Crawl viable). 6 plan corrections mid-execution: I1 over-correction, I6 wrong framing, W1 wrong shape (verb vs extend), C1a false premise (grep already skips gitignore), C1b needless guard, B1 non-viable at 17.3 GB. The REAL corrected every time; re-reading the spec never did. - **action**: keep — see [[LRN-132]]. 4 features killed at measurement (B1/B2/B3 + W2 deferred) beat 4 false-signal features. The most trustworthy output of the session was the code NOT written. Method that worked: show/measure the real artifact before deciding, mirroring [[LRN-074]]'s watch-the-RED discipline applied to a plan. + +### EVAL-026 — 3-way plan challenge caught 4 BLOCKERs dogfooding own plan (2026-07-17) +Dogfood: 3 blind lenses attacked the v1 plan for the plan-challenge feature itself. Verdicts CONCERNS(4)/FATAL(6)/FATAL(4). Caught 4 distinct BLOCKERs a single pass would blend: (1) v1 unbuildable — targeted init-project (inline-load, no dispatch) + false "plan on disk" premise for feat/bugfix (only contract persists); (2) failed-open silently dropping a lens while claiming "challenged" (inverts verify-secure-loop "a mute verifier is NEVER a PASS"); (3) consensus-weighting buries lone L2 security finding (lenses orthogonal); (4) sonnet challengers violate [[BDR-066]] (audit judgment=big model). Synthesis REJECTED 1 false positive (allowed-tools-blocks-dispatch — ship-feature has same frontmatter + dispatches fine). Each lens found a DIFFERENT class of flaw → evidence 3-independent > 1-multilens. Action: hardened v2 (severity-driven + fail-safe + re-think loop) shipped. Method validated itself before build. diff --git a/.claude/memory/journal.md b/.claude/memory/journal.md index 0c27d3e..c8b2acb 100644 --- a/.claude/memory/journal.md +++ b/.claude/memory/journal.md @@ -402,3 +402,6 @@ rules: - seo/geo parity vs github.com/AgriciDaniel/claude-seo (11.5k★, MIT): full 20-point plan built from a 3-subagent inventory, then executed. Verdict cherry-pick-never-install ([[BDR-070]]). 21 commits: Phase 1 (I1-I8 integrity, markdown specs) MERGED to develop (02c7a6f, 8 commits); Phases 2-7 on bugfix/seo-geo-integrity UNMERGED (13 commits, human gate). `fetch.sh` 5→11 verbs (richresults via inspect, sitemap, rendercheck, linkgraph, cannibal, drift, score); seo-data test suite 85→167 pass, 0 fail. Dogfooded on 2 live sites (zenquality Astro + lavageangels356 native PHP) — the second caught 2 bugs Astro hid (image:loc counted as page, flat-URL family heuristic). - 4 features KILLED at measurement, not built: B1/B2 (Common Crawl edges = 17.3 GB, ref impl reads 2.9% and calls it a profile — [[BDR-071]]), B3 (GSC Links API doesn't exist), W2 (Bing OAuth swamp — [[BLK-017]]). 30/70 similarity refused (needs content extraction), Playwright refused (R2 [[BDR-072]]), defusedxml refused (DTD-reject keeps stdlib-only). The most trustworthy output was the code NOT written ([[EVAL-025]]). - BDR-070/071/072/073 + LRN-131/132/133 + BLK-017 + EVAL-025 capitalized; checked 14 TODO done (I1-I5,W1,W3,C1-C3,B3,R2,H1,H2), W2+R1 left unchecked (deferred/rejected). 2 learnings dropped as dup of [[LRN-074]] (grep/find gitignore + detector-proof). Red thread [[LRN-133]]: an omission must stay legible. Verification discipline [[LRN-131]]/[[LRN-132]]: WebSearch ≠ verification, subagent summary = claim not fact (7 disproven, 3 self-reproduced). +- Removed config-protection edit-block guardrail (full removal, user req) → feature/drop-config-protection (0e1b89c). Residual gitflow+Gitea guards only. [[BDR-074]] [[LRN-136]]. +- Built framework-wide 3-way plan-challenge phase → feature/plan-challenge-phase (6bfc054): lib/challenge-plan.md + agents/plan-challenger.md + 41-assertion lock, wired into 11 reflection orchestrators (build-plan/proposals/fix-bundle), excluded 6 no-plan skills. Full suite 16/16. [[BDR-075]]. +- Dogfooded the challenge on its own v1 plan: 3 blind lenses caught 4 BLOCKERs + rejected 1 false positive → hardened v2 shipped [[EVAL-026]]. Both branches finished into develop on user signal, NOT pushed. diff --git a/.claude/memory/learnings.md b/.claude/memory/learnings.md index 66c4b9e..c1a6dd1 100644 --- a/.claude/memory/learnings.md +++ b/.claude/memory/learnings.md @@ -1339,3 +1339,6 @@ rules: refuse it EVERYWHERE, not just where you look first. A fresh adversarial reviewer attacking diff A routinely surfaces a real hole in already-shipped code B — see [[EVAL-020]]. + +### LRN-136 — config-protection live state follows checked-out branch's symlinked settings.json (2026-07-17) +~/.claude/settings.json is a SYMLINK to the repo settings.json; Claude Code hot-reloads settings on change → the config-protection PreToolUse hook's active/inactive state tracks the CURRENT branch's settings.json. On feature/drop-config-protection (hook deregistered) a protected edit passed silently, sentinel unconsumed; after gitflow-switch to a branch off develop (hook still registered) the SAME class of edit was blocked. Apply: a change that removes a settings-registered hook is live only on that branch until merged; use the one-shot sentinel for protected edits on any branch that still registers it. ([[BDR-074]] context.)