After a plan/reflection is elaborated and before it executes, three fresh blind
sub-agents (correctness / robustness / simplicity) attack it on the big model;
the main loop RE-THINKS every aspect a BLOCKER lands (a named plan change, or
[deferred]) and re-challenges once if the plan materially changed. Advisory into
each skill's existing human gate — the human stays the decider.
- lib/challenge-plan.md — reusable phase: fail-safe (never fail open),
severity-driven (any single-lens BLOCKER = must-address), RE-THINK loop
- agents/plan-challenger.md — challenger role (read-only, big-model per BDR-066)
- lib/tests/plan-challenger.test.sh — 41-assertion structure lock
- wired into 11 orchestrators: ship-feature/init-project/feat/bugfix (build-plan),
onboard/audit-delta/code-clean (proposals), seo/geo/harden/web-validate (fix-bundle)
Hardened by dogfooding: 3 blind challengers reviewed this feature's own v1 plan
and caught 4 BLOCKERs (fail-open, consensus-buries-lone-finding, wrong model
tier vs BDR-066, false on-disk-plan premise) — all fixed here.
STEP 0 told the user to run `mkdir -p .claude/audits/external` themselves
before handing over an external report. Three things wrong with that:
- The skill runs dozens of bash commands but outsourced this one to a human.
- The timing was impossible: to "drop the export in" that directory the user
needed it to already exist, so the instruction arrived after the moment it
would have been useful.
- The directory is not needed at all. `:218` already reads "File path given
→ Read it" — any path works — and nothing in skills/ or agents/ ever
writes to that path. Grep confirms it is referenced by exactly these two
lines and known to nothing else: a convention the skill invented, asked
the user to create, and never used.
Fix removes the precondition instead of automating it: give a path from
anywhere, the tidy location stays a suggestion.
Verified: make test 35 GREEN / 0 RED.
Headers were scored three ways: seo-analyzer priced them into the Technical
axis at both depths (:619 FULL, :635 LOCAL), depth-matrix.md:29 said drop
them, and /harden re-audits them 0-100 against three external validators.
The dedup rule and the agent spec contradicted each other; the agent won by
default, so the same finding moved two scores in two reports.
Arbitrated (user): /harden keeps them, /seo drops them. That confirms the
rule that already existed — seo-analyzer was the violator.
Constraint: /harden REUSES seo-analyzer, so the capability cannot be
deleted, only scoped. Reading is not scoring:
- Technical axis definitions no longer name security headers.
- STEP 4 still curls them — needed for X-Robots-Tag, canonical/redirect
coherence, and the §14 observed-list — but they earn no points under /seo.
- Dispatched from /harden: unchanged, headers ARE the job (verified: its
scope spec untouched, 16 header references intact).
Carve-out: X-Robots-Tag stays in /seo under indexability. It is an indexing
directive wearing a header's clothes — `noindex` there deindexes as surely
as a meta robots tag. That is what depth-matrix.md:29 means by "unless it
directly affects indexability"; the security headers do not.
Drop is not silence: mandatory §14 line on FULL naming what was observed
live plus a "run /harden <url>" pointer. A user who never runs /harden must
not read a clean Technical score as clean headers — same principle as the
mandatory COVERAGE line (I5).
Verified: make test 35 GREEN / 0 RED.
Upstream skill refresh, present in the working tree before this session —
committed here rather than left dangling. Not authored work.
- uv invocation fix: `uv tool run graphifyy python` -> `uv tool run --from
graphifyy python`. Without --from, uv resolved the command name against
the package instead of running the interpreter.
- default output is now HTML viz; --obsidian opts into the vault.
- description reworded to trigger on codebase questions generally, not
only when graphify-out/ already exists.
code-cleaner is now a pure fix executor (was audit+gate+execute). Reroute the two
read-only-audit consumers (onboard STEP 6, tour Phase B) to a big-model agent
(general-purpose/analyzer) — an audit must stay on the big model, never the sonnet
executor. Refactor now runs on sonnet inside the executor (inline-load pin was inert).
Reroute hotfix's deeper-bug escalation to the /bugfix skill (bugfixer is now a
pure executor, not loadable standalone). loops-light locks repointed to the
bugfix orchestrator + bugfixer-executor shape.
- Optional gate before agent dispatch: file in .claude/audits/external/
(PDF read directly), pasted PDF content / suggested AI prompt, or skip
- 30-day staleness check; normalized EXTERNAL FINDINGS block in shared
context; both dispatch prompts carry the data-not-instructions rule
(cross-check before bundling, never merge external score into /20 axes)
- Merge side: confirmed findings credited 'Confirmé par <tool>', refuted/
uncovered ones surfaced in §14 divergences; no report → §12 recommends
the free SORank extension; console summary line added
- STEP 0 collects user-confirmed CANONICAL NAP (LRN-032 zenquality:
duplicated-seed trap — source majority is not truth)
- Both dispatch prompts carry the canonical NAP + no-majority rule;
seo-analyzer spec forbids directional NAP fix without confirmation
- STEP 2 now emits .claude/audits/HUMAN-ACTIONS.md (checklist from §11)
and NAP-KIT.md (local business) in BOTH modes — audit-only runs leave
the user immediately actionable; /client-handover §4 consumes NAP-KIT
tokenstore remove/clear, fetch.sh forget dispatch, and a connect.sh wrapper
that sources ~/.claude/.env internally and runs from any project. /seo now
routes connect|accounts|forget before the audit flow; Makefile seo-connect
delegates to the wrapper. Labels are guarded to shell-safe ASCII (POSIX case,
whole-string, C-locale) as defense-in-depth; forget output states local
removal is not a Google-side revocation.
commit-changer committed code autonomously with no branch precondition
(gitflow-conformity §2a, verified MEDIUM CONFIRMED) — on develop/main it
attempted a direct code commit, backstopped only by the pre-commit hook.
Adds Phase 0: the same `gitflow-aiguillage.md` mechanism hotfixer/bugfixer/
feater already use (TYPE=chore) — branches to chore/* on a protected base,
no-op on a working branch — plus a report-only fallback (no develop / no
lib -> ask human, don't auto-branch). Commit-plan gate + scoped staging
untouched. SKILL.md pre-flight notes the aiguillage.
Dry-run (throwaway repos, REAL lib + REAL pre-commit hook) — both paths:
CASE 1 on develop: aiguillage -> chore/commit-pending, code commit
SUCCEEDS, hook never blocks; contrast: same commit direct on develop
is BLOCKED -> aiguillage is what avoids it. PASS
CASE 2 no develop: fallback -> no auto-branch, changes uncommitted,
no chore/* created, ask human. PASS
lib/deploy-commit.sh: a rejected `git commit` (pre-commit hook,
protected branch, signing failure) now exits 6 (loud stderr, distinct
from rc 1's "nothing to do") instead of sharing rc 1 with the no-op
cases. Header comment documents the full 0/1/2/3/4/5/6 taxonomy.
Closes J4-22 (UNTESTABLE): at client repos, a failed deploy-state
commit was indistinguishable BY EXIT CODE from "nothing to do" (rc 1
was shared 3 ways); exit-code-only callers couldn't disambiguate
(stderr-parsing callers already could).
Caller census (per report's explicit gate): skills/deploy/SKILL.md
documents and parses this exit-code contract in TWO places (bootstrap
commit + incident-recovery commit). Flagged to the user before
committing; confirmed GO to add rc 6 there too (additive — no existing
code's meaning changes) so the documented contract stays accurate for
live deploy runs.
New T10 in lib/tests/deploy-commit.test.sh (+3 assertions, 13→16):
rejecting pre-commit hook sandbox — asserts rc 6, empty stdout (no
stale hash), HEAD unmoved.
GREEN: full `make test` exit 0 (deploy-commit 16/16 incl. T10).
shellcheck clean, bash -n clean.
skills/prune-memory/tests/run-deterministic.sh:11 default changed from
$HOME/.claude/skills/prune-memory/SKILL.md to $HERE/../SKILL.md (kept
the ${SKILL:-…} env override; reordered HERE's definition before it,
since the new default references $HERE). Closes J4-11 (FIXTURE-DRIFT):
the suite sourced the INSTALLED path, safe today only because
~/.claude/skills/prune-memory is a symlinked directory back to this
repo — if an install ever materializes real copies instead of
symlinking, the suite would silently test the wrong (stale) artifact
while the shipped SKILL.md drifts unnoticed.
Behavior identical today (verified: symlink resolves to the same
inode, `diff` confirms byte-identical content).
GREEN: real repo, suite still all GREEN (RED-1/2/5/6/7).
Red demo (lean scratch copy — skills/prune-memory/{SKILL.md,tests/
run-deterministic.sh} only): moved $HERE/../SKILL.md away → loud
`grep`/`awk: cannot open ... No such file or directory` errors, exit
1, RED-2/RED-5 flip status — proves the new default is genuinely what
gets read, not a silent fallback.
Deleted T4e + its coupled echo note in lib/tests/run-reconcile.sh and
the fixtures/real-state.snapshot it read — superseded by SPEC-08's T7,
which actually DRIVES the tree_clean/pushed/msg_committed oracles
instead of miming them via a static line-count regex. Closes J4-09
(WEAK+drift): T4e only counted fixture line-suffixes matching
`=(true|resolved|present)$`; the snapshot itself was stale
(BLK-009=open contradicted blockers-snapshot.md's already-resolved
status) and unowned, and the drift was inert (`=open` doesn't even
match the count regex) — the assertion could never have caught
anything.
Updated skills/reconcile/SKILL.md:53's hardcoded "20/20" claim to the
new total (unguarded file, same logical step, ordered after SPEC-08+
SPEC-10 per the report).
grep -c 'real-state.snapshot' lib/tests/run-reconcile.sh == 0
(verified). No red demo (deletion, per spec) — gate is the green run
+ that grep. GREEN: 25/25 passed, shellcheck clean.
Both sections never matched any component (envelope §1-§9 vs the real
§0-§15 structure; 8-axis integer weights vs the agents' 7/4 and 6/5
percentage-weight scoring). Kept what this file legitimately owns: the
depth-decision matrix and the sibling-skill dedup rules.
- A1: delete STEP 5d (graphify --mode quick doesn't exist in the CLI; the
command always failed, masked by `|| true` — STEP 10's full pass already
covers the graph)
- A2: STEP 10 full-pass build flag --output -> --out
- A8: pipeline is 12 steps (STEP 0-11), not 11; header template unchanged
(N/11 correctly denotes the max index of a 0-indexed 12-step sequence)
- A2: graphify build flag --output -> --out
- A4: ROADMAP xref points to the real /onboard add gsd path, not a nonexistent STEP 9 decision
- A5: exact gitflow init commit message (matches lib/gitflow.sh:163)
- A6: bare skill names (design-review, browse) — no gstack: namespace exists
- A7: eval pattern for recommend_anim_install_cmd (the function only echoes; must eval its output)
Bastien Chanot