Дмитрий
903aa70098
V3.3 — второй раунд adversarial audit от controller'а на свою же v3.2. Главное:
N1 fatal honest revision — мой v3.2-фикс S5 был самообманом. BLOCKED-protocol regex в subagent response = controller-written text → fake `{"status":"DONE"}` маркер trivially. §3.4 переписан: убран claim «closes S5», заменён weak heuristic (tool_use count > 0 + marker check). S5 → §9 acceptable residual класса controller-writable signals (как S8). Real-fix требует harness-level enforcement за пределами controller-writable spec'а.
12 правок без новых секций:
- §3.4 BLOCKED-protocol → weak heuristic + S5 residual (N1)
- §3.2.0 расширен: user-run smoke + PostToolUse semantics probe (N4, N10)
- §4 Поведение 1 stale-registry → mandatory AskUser (N2)
- §4.5 fail-CLOSE с конкретными keyword-примерами (N8)
- §4.7 first-option-position bias detection (N11)
- §5.1 file-watcher reset только на git commit exit 0 + lefthook GREEN (N3)
- §5.1 SKILL_BASH_ALLOW hardcoded mapping в gate-code, не Skill-controlled (N5)
- §8 +PostToolUse handler task + implementation order matrix (N7, N9)
- §9 N12 off-topic detection limits acknowledged + S5 acceptable residual
- §10.2 +schema для edited-files-<sess>.json (N6)
Implementation budget: 18-27h (v3.2) → 19.5-29h (v3.3). +1.5-2h за inline-фиксы.
Главный урок: controller-writable signals — fundamental class уязвимостей. Любой текстовый/JSON-маркер в response = fake-able trivially. Реальное закрытие требует out-of-band каналов. S5 и S8 — этого класса, оба unresolvable в controller-writable spec.
cspell-words.txt sync: +1 валидный термин (эскалируем).
Verify-sentinel: 1179/1179 vitest tools-only GREEN.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Observer infrastructure
Passive evidence-loop for the Лидерра «brain» per ADR-011.
Files
episodes-YYYY-MM.jsonl— append-only JSONL, one line per Stop-event. Schema v2 (schema_version: 2): the 5 mandatory fields +decision_provenance(who chose the node),environment(economy_level / model / post_compaction / session_turn / parallel_session),task_size,task_ref,prompt_signal, and anoutcomethat isunknownat write time (refined by/brain-retro). On an internal hook failure a minimalobserver_errormarker line is written instead of a silent skip. Written bytools/observer-stop-hook.mjsviatools/observer-transcript-parser.mjs.notes/YYYY-MM-DD-<slug>.md— optional MD notes for sessions with qualitative history.STATUS.md— auto-generated dashboard. Regenerated per-commit bytools/status-md-generator.mjs..read-counter.json— C3 observer-of-observer counter. Updated on Read of observer files.dashboard.html+dashboard.js+dashboard-core.js— Brain Dashboard: visualises the episode log over the automation-graph topology (4 views — Карта / Разбор / Лента / Агрегат). Runnpm run brain:dashboard, open the printed localhost URL.dashboard-core.jsis pure logic, unit-tested intools/brain-dashboard-core.test.mjs.
Lifecycle
- Write: every Stop-event appends one JSONL line, parsed from the session transcript (Stop-hook).
- Aggregate:
/brain-retroskill reads JSONL each sprint, proposes regulatory candidates. - Surface:
STATUS.mdshows controllers + monthly stats. - Self-prune: C3 warns if 54 weeks pass without any read of observer files.
Routing-tag discipline
When the user dictates a specific method/node (e.g. «запусти discovery-interview»), Claude must emit one line in its response:
<!-- routing: provenance=user_directed_method node=<chosen> counterfactual=<node Claude would have chosen autonomously> -->
The Stop-hook routing-gate (tools/observer-routing-detector.mjs + routingGateDecision) detects a dictated method; if the tag is missing it returns decision: block, so the turn cannot end without the tag. The gate fires at most once per turn (stop_hook_active guard). This makes decision_provenance reliable — factor analysis can separate a router error from a user-dictated one.
Privacy
PII filter (phone numbers, emails, tokens) is applied before every write — see tools/observer-pii-filter.mjs. gitleaks pre-push also scans observer files as part of full-history sweep.
Don't
- Don't edit
episodes-*.jsonlmanually — it's append-only. - Don't write outside
docs/observer/notes/for hand-curated notes. - Don't change
.read-counter.jsonmanually — it's maintained by hooks.
HK1 pre-check (Pravila ADR-010) — verified 2026-05-19
Before registering tools/observer-stop-hook.mjs on Stop event (Task B5), verified collision against 6-component economy/skill-discipline architecture:
- User-level
~/.claude/settings.jsonalready has Stop hook: agent-type Sonnet-4.6 economy compliance verifier (analyzes transcript for claim-without-evidence violations). - Project-level
.claude/settings.json— Stop slot empty.
Result: no overwrite. observer-stop-hook will be added as command-type entry in project-level Stop array. Project + user scopes are independent slots in Claude Code 2.x — both run on the same Stop event without conflict. The agent verifier (user scope) and the JSONL appender (project scope) have non-overlapping responsibilities.