Task 4 уточнён: нет миграций (только PHP), fast-fail активируется сразу после
деплоя. Добавлены конкретные gh workflow run команды для cleanup (Steps 3-4 из
sql-runner.yml) и верификации шторма + incidents алерта. Галочки [ ] оставлены
(задача контроллера, не агента).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Markdownlint added blank line before list in TL;DR v3.5 Changes section. Cosmetic, no semantic change. Captured to keep working tree clean for session handoff.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Dedicated brainstorm на S8 (последний acceptable residual после v3.4) — controller-writable markers BRANCH-SWITCH-CONFIRMED / RECOVERY-INTENT: заменены на AskUserQuestion-gate через расширение §4.5 answer parser. Тот же class fix что S5 v3.4 — перенос сигнала в side-channel который controller не контролирует.
V3.5 architectural shift:
- Old: controller wants git rebase → hook видит блокировку → controller пишет BRANCH-SWITCH-CONFIRMED в response → hook читает, allows. Самоодобрение через controller-written text.
- New: controller wants git rebase → hook блокирует → controller обязан AskUserQuestion('Выполнить git rebase main?') с 3 опциями → user picks → §4.5 parser распознаёт git-pattern, пишет approve_git_operation запись в askuser-decisions.jsonl с approved_action_pattern exact → controller retries → hook читает askuser-decisions, exact-match approval, consumes one-shot, allows.
Anti-forgery 4 защиты:
- askuser-decisions.jsonl в protected paths (~/.claude/runtime/*)
- Пишется только §4.5 parser при реальном AskUser-ответе
- Strict exact-match approved_action_pattern против tool_input.command
- 5-min time window + one-shot consumption (consumed flag)
Adversarial pre-check v4.3 — 6 attack-векторов проверены, все заблокированы.
5 spec edits:
- §3 preserved hooks таблица — enforce-branch-switch row переписан (no longer reads response markers, instead reads askuser-decisions)
- §4.5 — новый row + новое enum gate_interpretation: approve_git_operation
- §9 — S8 убран из acceptable residual list. Список residual теперь ПУСТ.
- §10.2 — askuser-decisions schema +consumed/consumed_at/consumed_by_tool_use_id + полный schema приведён
- §11 v3.5 history entry
Implementation budget: 21.5-31h (v3.4) → 22.5-32h (v3.5). +1.1h за: rewrite hook, extend parser, one-shot consumption, schema, spec edits.
Главный урок: класс controller-writable signals закрыт архитектурно через обе v3.4 (file side-channel) и v3.5 (askuser-decisions side-channel). Формула generalisable — любой controller-writable signal закрывается переносом write-канала в protected file который пишет отдельный gate-процесс при harness-driven событии.
Hard wall теперь полный hard. Acceptable residual list пуст.
Verify-sentinel: 1179/1179 vitest tools-only GREEN.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Handoff document for transition to new session. Contains:
- TL;DR of where we left off (spec v3.1 ready, impl not started)
- 4-step instructions for next session (read spec, writing-plans,
subagent-driven impl, post-impl tasks)
- Context (brain-retro #10 trigger, self-retrospect #2 confirmation,
user choice of Level 4)
- 7 design principles from spec section 2
- Architecture TL;DR (gate, 4 behaviors, baseline, deletes, preserved)
- All 4 spec versions in git with commits
- Cross-refs to L1+L2 plan, brain-retro, self-retrospect
- 5 open questions for writing-plans phase
Cannot write to memory/ path in this turn (memory-coverage hook
requires direct:memory-sync coverage, current turn has different
coverage). Memory entry can be added in next session via Skill or
manual annotation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Comprehensive analysis of v3 found ~24 minor issues (no critical bypasses).
V3.1 closes most via clarifications to prepare for writing-plans skill in
next session.
Additions:
- TL;DR at top — fast orientation for implementer
- 10.1 Function and registry references (nodeMatches source,
registry source docs/registry/nodes.yaml, SDD-skill impact,
coverage-hint to recovery resolution)
- 10.2 State file schemas (8 files: router-state, chain-state,
askuser-decisions, router-gate-decisions, subagent-inheritance,
coverage-hint, gate-errors, gate-config)
- 10.3 Test strategy: ~150 unit + 10-15 integration + 10-15 golden
snapshot + 5-7 smoke
- 10.4 Success metrics: quantitative (override drops to 0,
gate-decisions growing) + qualitative (lockout < 5/100,
correct% > 60%) + acceptance criteria
- 10.5 Rollback plan (3 levels: hook off / revert commits / v2 baseline)
- 10.6 Stages and parallelism: 6-9h wall-clock with SDD parallelism
vs 13.5-20h sequential
No architectural changes — v3.1 only clarifies what implementer needs
to know without making implicit decisions.
Spec versions in git:
- v1: 7a43c175
- v2: b510a758
- v3: b632bcba
- v3.1: this commit
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
5-task plan to close 3 enforcement gaps surfaced by brain-retro #10:
1. Narrow 'recovery' override scope (5→2 categories)
2. Narrow 'ремонт инфраструктуры' override scope (11→3)
3. Per-rate-window in enforce-override-limit (5/10min)
4. Lower classifier-match threshold 0.8→0.6 + inline router-skip
Driver: 679 override events on 2026-05-28 vs 12 baseline on 25.05.
User selected option B (Level 1+2) after brain-retro #10 analysis.
All 4 implementation tasks completed via subagent-driven workflow
(commits 09f6e332, 029dbe50, 2b23a1f2, 726c2121).
Final regression 1179/1179 GREEN. Plan saved post-implementation.
Also: cspell-words.txt += 'суппрессить' (project term used in plan).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Groups documentation produced during 2026-05-28 brain-retro session:
retro notes 8 (carryover) and 9, self-retrospect 1, sanity check JSON,
three Phase plans for router-hooks fixes. All implementation already
pushed in earlier commits — this commit groups artifact metadata.
Plus typo fixes in self-retrospect (agregatov, seryj) and cspell vocab
extensions for session-specific terms (PAMYATKA / procs / russian verbs).
Pure documentation. No code, no normative drift.
Investigation 2026-05-25: for tenant client1 (tenant_id=2) on prod liderra.ru:
- 205 leads at supplier (info@lkomega.ru, visit=rt) vs 160 deals on portal
- 82 leads lost (76 via 302-redirect from ValidationException, mostly
non-B-prefix projects: client.carmoney.ru, cashmotor.ru, etc.)
- 37 duplicate deals (CSV-recovered SupplierLead vid=null + later
webhook with real vid "create two Deals because supplier_lead_deliveries
locks on supplier_lead_id, not phone+project)
Three independent fixes, three plans, three deploys:
Phase 1 (low risk): Always JSON 422 for webhook ValidationException
Phase 2 (med risk, billing): merge webhook-after-CSV-recovered into
existing deal, no double-charge
Phase 3 (high risk, migration): accept non-B projects as platform=DIRECT
end-to-end (controller + 4 services + migration)
Phase 3 includes new LeadRouter fallback path: DIRECT-supplier_projects
match Liderra projects via signal_type+signal_identifier directly
(no project_supplier_links pivot required, since psl rows don't exist
for auto-created DIRECT supplier_projects).
Refs: docs/superpowers/specs/2026-05-25-supplier-webhook-reliability-design.md
v2.0 → v2.1 — 3 группы изменений (16 пунктов суммарно):
Группа 1 — решения принятые после v2.0, не внесённые:
- 1.1 Памятка classifier (4 паттерна: brainstorming / discovery-interview /
writing-plans / systematic-debugging). +flag prompt-enrichment-mode.
- 1.2 Reviewer как полноценный Claude Code subagent (tools=[Read,Grep,Glob,
Skill], model=opus). Новый файл .claude/agents/reviewer-agent.md.
+стоимость $240-1200/мес vs $40-80 direct API. Crash fallback на direct
API. Context bloat cap 10 соседних эпизодов.
- 1.3 Inheritance + 3 группы коротких prompt'ов (continuation/acknowledgment/
cancellation) + 30-минутный таймаут. +flag inheritance-mode. Новые поля
в schema v4.1: inherited_from_task_id, inheritance_age_minutes,
previous_direction_rejected, previous_task_id_rejected.
Группа 2 — edge cases:
- 2.1 Reviewer model явно opus в agent file.
- 2.2 Reviewer subagent crash → fallback direct API call.
- 2.3 Reviewer context bloat: max 10 episodes в agent system prompt.
- 2.4 Manual override приоритет №1 в prefilter (раньше inheritance).
- 2.5 Cancellation clears state + previous_task_id_rejected marker.
Группа 3 — мелкие упущения:
- 3.1 brain-retro SKILL.md description: раз в 1-2 недели (не sprint).
- 3.2 recommended_chain_id nullable для custom chains.
- 3.3 Embedding только для non-prefilter эпизодов.
- 3.4 PII filter wraps sanity-check comments.
- 3.5 requested_node fuzzy matching fallback.
- 3.6 Anchor word list inline initial.
- 3.7 Self-retrospect counter init в фазе 3 step 3.3.
- 3.8 Sanity-check answer file schema_version=1.
Cost rewrite: 720-1380 USD (v2.0) -> 1940-8200 USD (v2.1) на 6 месяцев
из-за reviewer subagent. Granular rollback через reviewer-mode=direct-api
возвращает к v2.0 ценам.
§21 новый — changelog v2.0 → v2.1 со всеми 16 пунктами и где правка.
Реализация — после закрытия Биллинга v2 Спек C.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3-task TDD-ish plan to create two new project agents:
- Task 1: .claude/agents/normative-sync.md (full Sonnet 4.6 system prompt)
- Task 2: .claude/agents/prod-deploy-validator.md (8 SSH checks + quirks 104-108)
- Task 3: First dry-run smoke test for both + capture lessons in memory
Spec: docs/superpowers/specs/2026-05-24-controller-offload-agents-design.md (71a5dd6).
Also: +2 cspell-words (маппинге, dogfooded).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3 дыры обнаружены при инспекции warn-only состояния сторожа 24.05.2026:
1. UTF-8 mojibake в state-файле сторожа (Windows Node stdin без setEncoding).
2. recommended_node не пишется в эпизоды наблюдателя (helper существует, не подключён).
3. chain_progress / chain_completed / recommended_chain — то же.
Без починки brain-retro новых метрик (domainHitRate / chainCompletionRate)
покажет пустоту, а Layer 2 эскалация на русских промптах работает по
испорченному тексту. Stage 3 enforce включать до фиксов нельзя.
Spec scoped к 3 файлам кода + ≤80 строкам нетто; subagent-driven
последовательно (3 Sonnet + closure Opus). Smoke на живой сессии.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Spec for two specialized AI agents to offload the main controller:
- #1 normative-sync: applies 4-file normative sync (Pravila/PSR/Tooling/CLAUDE.md)
after a completed task. ~20 invocations/week, saves ~70K controller tokens
per episode. Model: Sonnet 4.6.
- #2 prod-deploy-validator: 8-check pre-flight before liderra.ru deploy.
~5-7 invocations/week. Driven by 24.05 03:46 UTC 18-min portal incident
(quirk 107 — config:cache not under www-data). Model: Sonnet 4.6.
Based on brainstorming session 24.05 with measured frequencies from
MEMORY.md + CLAUDE.md §6 + push history 16-24.05.
Precedents: pest-parallel-debugger, rls-reviewer project agents.
Also: +7 cspell-words entries for the new spec.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pre-sharing-эпик legacy: ProcessWebhookJob + WebhookReceiveController +
POST /api/webhook/{token} + webhook_log/webhook_dedup_keys/tenants.webhook_token.
На проде 0 вызовов за всю историю. Не часть актуальной архитектуры каналов
(основной = шеринг crm.bp-gr.ru, резервный = CSV reconcile — оба уже на always-rub).
Удаление снимает блокер для Phase B Спека A (DROP COLUMN balance_leads),
закрывает публичный endpoint /api/webhook/{token}, убирает расхождение
биллинг-моделей в коде.
Approach: одним PR, одним релизом. Бэкап перед миграцией, 7д наблюдение.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
8 задач: baseline → таблица-замок supplier_lead_deliveries → раздача
по клиентам (LeadRouter DISTINCT ON) → удаление DuplicateDetector из
обоих джобов → замок insertOrIgnore → тесты (model-agnostic) → регрессия.
Вариант B. Заякорено на always-rub LedgerService (Спек A в origin/main).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Правило: дедуп — ответственность поставщика; Лидерра телефон не фильтрует,
берём за всё, что прислано. Защита от своих дублей — на уровне БД
(замок supplier_lead_deliveries, ключ по поставке, не по телефону).
Лимит шеринга = 3 разных клиента, одному клиенту одна копия.
Вариант B (железобетонный) утверждён на брейнсторме 23.05.2026.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Дмитрий