chore(memory): BDR-122 + LRN-219..221 + BLK-032 — ship-feature guardrails a11y split

This commit is contained in:
bchanot committed 2026-10-20 20:30:48 +02:00
1 parent 274396da6e
commit defe446a42
6 files changed
+156

No files matched your search

+5
View File
@@ -51,6 +51,7 @@ rules:
| BLK-029 | 2026-10-08 | Claude Code mods (2.1.294): alias → id resolution for a model set by a hook lags the Agent tool's (`sonnet` → `claude-sonnet-5`, 404); Agent tool schema refuses full ids | upstream |
| BLK-030 | 2026-10-15 | model-router kept routing best-tier phases to Fable out of usage credits: each new session + each hold lapse lost a turn to 429 `model_requires_usage_credits`; breaker reactive only | resolved |
| BLK-031 | 2026-10-16 | model-router: session starting on Fable w/o credits lost 1st turn + each hold lapse; probe skipped kept cur | resolved |
| BLK-032 | 2026-10-19 | Main loop pinned on Fable in long sessions: session model Fable 1M + downgrade cap 120k + prompt rule `plan` → inline code on Fable xhigh; user routing policy recorded for next /feat | open |
---
@@ -329,3 +330,7 @@ rules:
- **Solution**: c11cd3b drops `!call.moved` guard: model decision lands on (moved or kept) probed once if table id, unverified, not held. Fail → markDown 'probe', one re-decision → leaveDown → opus, stand-in effort +1. Cost ~1.3 s healthy / ≤8 s dead, 1st decision per process. Test `probe: a session model that never answers leaves for opus on the first step`.
- **Status**: resolved. Open: leaveDown follows classifier `wanted` (implement → sonnet), not opus; "context size unknown" likely document attachment → `promptSize` unknown.
- **Reference**: [[BLK-030]], [[LRN-218]], contract `.claude/tasks/contracts/2026-10-16-router-probe-session-model-1850.md`.
## BLK-032 — Main loop pinned on Fable in long sessions: session model Fable 1M + downgrade cap 120k + prompt rule `plan` → inline code on Fable xhigh [open]
- **Cause**: `~/.claude/settings.json` `"model": "claude-fable-5-1[1m]"` (every session starts on Fable); model-router never moves main above the session model, moves DOWN only under `downgradeMaxTokens` 120000 (LRN-204 cold read) → a long session (faunosteo R4.x, context > 120k) never reaches `implement`/sonnet; prompt rule `(plan|design|architecture…)` → phase plan = best/xhigh each turn ("plan §B.1.7"). No doctrine ceiling: CLAUDE.md allows inline work "doable in a few tool calls", BDR-115 "pin = default, never ceiling".
- **User policy 2026-10-19**: session model Opus 5.5; sonnet writes, opus decides what to write; Fable only for orchestrator reflection/plans and the important audits (security, functional), other judges opus. → next `/feat`: settings model, `judge` split, ceiling rule in CLAUDE.md, BDR superseding part of BDR-066 / BDR-115. Links [[BDR-066]], [[BDR-107]], [[BDR-115]], [[LRN-204]].
+7
View File
@@ -139,6 +139,7 @@ rules:
| BDR-119 | 2026-10-16 | impeccable = Build path for all UI; GATE-BLOCK swaps frontend-design for impeccable | accepted |
| BDR-120 | 2026-10-18 | Secure-by-default guardrails: always-on rule file + /guardrails stdlib scanner, passwords in api domain | accepted |
| BDR-121 | 2026-10-18 | Guardrails fix loop reuses bugfix circuit; scanner = contract oracle, baseline re-pass | accepted |
| BDR-122 | 2026-10-19 | Guardrails split: /a11y standalone on shared scan engine, injection + exposure domains, exposure proven live by read-only probe | accepted |
---
@@ -1462,3 +1463,9 @@ Branch feature/user-writing-web-rules, UNMERGED (human gate).
- **Why**: user ask. Proof by deterministic scanner, never executor's word; no new circuit to maintain.
- **Alternatives rejected**: guardrails edits code itself (skips fresh gates); chain to /bugfix per finding (one interview + contract per hit, too heavy for 60 labels); LLM verifier as sole proof (verdict without execution).
- **Refs**: feature/secure-by-default 73c25d7. Links [[BDR-120]].
## BDR-122 — Guardrails split: /a11y standalone on the shared scan engine, injection + exposure domains, exposure proven live by the read-only probe [accepted] (2026-10-19)
- **Decision**: a11y out of /guardrails → `skills/a11y` (own SKILL, Tab-only walk, report `.claude/audits/A11Y.md`, fix loop = bugfixer + verify-secure gates, scanner oracle A11Y-CLEAR / A11Y-NO-NEW). Shared engine `skills/guardrails/scripts/scanner.py` `Catalogue(name, domains, rule_domains, manual_rules)`, marker = name upper; a11y `scan.py` loads it via `importlib.import_module` after `sys.path` insert (no noqa). Guardrails = 7 domains: api, hosting, supabase, messaging, functions, injection (INJ-*, 6 rules), exposure (EXP-*, 9 rules). `rules/secure-by-default.md` six numbered domains; a11y doctrine → lazy `rules/a11y.md` (`paths:` globs), `rules/web-building.md` points at it. Probe exposure block sourcemap / error-body / fields: fail-closed (unobserved = SKIPPED or ERROR, never CLOSED), every plan path starts with `/`, same scheme+host only, production build target (dev-server OPEN = dev artefact). security-auditor gate keeps `--domain api,supabase`. INJ-SQL-RAW-UNSAFE = WARN (variable ≠ proof), INJ-SQL-CONCAT = FAIL.
- **Why**: user ask 2026-10-18 (keyboard/alt/labels as own skill; SQL/DB injection; browser-visible info). One engine = one CLI contract, two catalogues, no drift. Lazy a11y rule keeps CLAUDE.global.md at 320.
- **Alternatives rejected**: a11y as a `--domain` flag only (user wanted own report + walk); engine copied into skills/a11y (two CLIs drift); exposure probe against dev server (dev artefacts = false OPEN); semgrep-only (blind to live responses, policies).
- **Refs**: branch feature/a11y-split-inj-exp, contract `.claude/tasks/contracts/2026-10-18-guardrails-a11y-split-inj-exp-2210.md`; spec + plan purged at finish (`git show 988038c:docs/superpowers/...`). Supersedes the domain list of [[BDR-120]]; links [[BDR-121]].
+3
View File
@@ -626,3 +626,6 @@ rules:
- Passwords added on user question ("stockés en clair ?"): gap, rule had none. `checks/passwords.py` (api domain): clear column w/o hash lib, `==` or SQL compare, fast hash, log call, literal. Calibrated 7 projects: zero noise after htpasswd volume-mount FP. Suite 89/89, `make test` green.
- Live probe added (user: "est-ce qu'il attaque de l'exterieur ?"). `probe.py` read-only, `--authorized`, env secrets, same-host redirect only, budget-capped. 5 probes confirm a hole from outside (OPEN/CLOSED). Suite 16/16 on a localhost fixture. Classifier blocked edits/tests carrying exploit-language and the SSRF-shaped redirect fixture; full-file Write passed, so probe strings are neutral and interpretation lives in checks.md.
- Fix loop added (user: propose fixes via Fable, user accepts, re-pass proves). Reused bugfix circuit: contract criteria with scanner oracle (`--assert-absent`, GUARDRAILS-CLEAR) run by gates.sh, bugfixer executor, verify-secure gates, re-pass `--baseline` (GUARDRAILS-NO-NEW), 2 rounds max. Proposals on route plan (best tier = fable first). Model gate added, census locks updated.
## 2026-10-19
- Guardrails split shipped via /ship-feature (feature/a11y-split-inj-exp): /a11y skill on the shared engine, injection + exposure domains, probe exposure block, secure-by-default six domains. SDD 9 tasks, GATE 1 ECARTS(9)→CONFORME (fields CLOSED route, suppressions removed), GATE 2 PASS (7 LOW), opus review 0 Critical / 6 Important fixed. Routing finding: Fable pinned in long sessions (BLK-032), user policy recorded for a later /feat. Merge on user go.
+14
View File
@@ -210,6 +210,9 @@ rules:
| LRN-188 | 2026-09-30 | Contract oracle tied to code shape (`grep -A6` line window) breaks on the first refactor while the property still holds; assert the property over the whole unit | writing CHECK oracles |
| LRN-189 | 2026-10-06 | `cmd \| grep -q` under pipefail = SIGPIPE false negative (rc 141); `>/dev/null` is NOT a fix (GNU grep treats it like -q); portable = producer out of the pipeline | any bash under `set -o pipefail`, tests and prod |
| LRN-190 | 2026-10-06 | Contract oracles: join `\\` line continuations before regex; "lint clean" = no finding beyond base, not zero; never `rm -rf "$var"` inside a CHECK | writing CHECK: lines |
| LRN-219 | 2026-10-19 | Scanner regex over untrusted trees: bound every repeat, cap comment blankers, timing lock on adversarial file | any whole-file regex scanner |
| LRN-220 | 2026-10-19 | Live probe verdicts fail closed by construction: fixture route per negative branch, 3xx = SKIPPED, depth/size limit = SKIPPED never CLOSED | writing probes and their suites |
| LRN-221 | 2026-10-19 | Floor-guard suppressions: remove, never waive; contract CHECK asserts intent (symbol/behaviour), not one code spelling | suppressions, writing CHECK: lines |
---
@@ -1814,3 +1817,14 @@ Rule: when editing a doctrine file under structure locks, grep the test's lock s
- **Context**: model-router bugfix 2026-10-14/15, live scratch mod (`claude -p --plugin-dir`). `complete({model:'claude-opus-5-5', prompt:'ok', maxTokens:1})` resolved 1308 ms with OBJECT `{ isAnswered, text, usage }` (d.ts says string; maxTokens not honored, 43 output tokens). Same call on `claude-fable-5-1` without credits: no settle, no rejection after 900 s (client retries the 429). Plan r1 probed in a loop before deciding → BLOCKERs: unbounded spin on explicit alias, cur probed.
- **Apply**: availability check via `complete()` = race vs own timeout, treat timeout + rejection as down, any resolved shape as up; never read `isAnswered:false` as failure (unobserved). Decide first, probe only the move target, once, skip cur and held ids, one re-decision without probe. No hook signal on the probe (Esc must not mark model down); probe sleep counts against 10 s `turn.step` budget. Links [[BLK-030]], [[LRN-207]].
## LRN-219 — Scanner regex over untrusted trees: bound every repeat, cap every comment blanker, lock with a timing test on an adversarial file
- **Context**: guardrails split 2026-10-19. Challenge + 3 review rounds found quadratic COMMENT / env regexes, unbounded `[^>]*` and `\S+`. Fixes: DATA_ATTR `[\w-]{0,60}`, SCRIPT_SRC `[^>]{1,300}`, TRACE_MARKERS `[^)\n]{0,200}`, SQL_COMMENT `/\*[\s\S]{0,4000}?\*/`, MARKUP_COMMENT tempered `<!--(?:[^<]|<(?!!--)){0,4000}?-->`. Residual ~1.4 s on adversarial 200k file, accepted.
- **Apply**: whole-file regex = bounded repeats, tempered dot for nested delimiters, timing lock in the suite (`guardrails-scan.test.sh`), files > 1 MB listed by path (`Project.large`), `.map` head 4 KB only. Links [[BDR-122]].
## LRN-220 — Live probe verdicts fail closed by construction: every negative branch needs its own fixture route, a 3xx reaching a verdict = dropped redirect = SKIPPED, a depth or size limit hit = SKIPPED never CLOSED
- **Context**: same run. GATE 1 verifier: `fields` probe had OPEN + SKIPPED assertions, no CLOSED route. Opus review: three false verdicts (embedded `sourceMappingURL` string → OPEN; off-host map redirect dropped by `_SameHost` → 302 → CLOSED; sensitive key below depth 4 → CLOSED). Each = a code path no fixture route exercised.
- **Apply**: per probe, fixture routes for OPEN, CLOSED, unobservable (SKIPPED), dropped redirect, truncated, oversize; anchor marker comments at line start and take the last; walk bounded bodies fully (iterative). Links [[BDR-122]], [[LRN-219]].
## LRN-221 — Floor-guard suppressions: remove, never waive; contract oracles assert intent, not syntax
- **Context**: GATE 1 counted 6 floor findings: `# noqa: E402` ×2 (import after `sys.path.insert`), a `# shellcheck disable=SC2016` copied from another suite (dead: deleting it raised nothing), 3 plan quotes of the same lines. `importlib.import_module` after the insert needs no noqa. Then criterion 16's CHECK grepped `^import scanner` and went red on the legit fix.
- **Apply**: a suppression = first try the construct that needs none; test a copied `disable` by deleting it; a CHECK greps the symbol or behaviour (`scanner.Catalogue` loaded), never one spelling of the code. Links [[BDR-102]].