feat(e7): live-сессии пилота — реальные LLM, run-процесс, Docker-образы, фиксы лимитов речи и баланса (TASK-53…TASK-56) #8

Merged

geekiot merged 16 commits from feat/pilot-e7 into main 2026-08-12 22:52:26 +05:00

Conversation 0 Commits 16 Files changed 54 +2093 -116

geekiot commented 2026-08-11 11:27:47 +05:00

Owner

Copy link

Контекст

Эпик E7 (пилот на реальных LLM): цель — прогнать live-сессии на GigaChat, собрать метрики NFR-010 и подготовить продакшен-развёртывание через Docker Compose.

Что входит

• Live-процесс game-core (TASK-53): session runner, python -m game_core, активация сессии, wakeup/start-now, восстановление запланированных сессий из журнала (D-81), управление жизненным циклом соединения с БД.
• LLM-агент (TASK-53): живой игрок на GigaChat/Kimi через LangChain, промпты знакомства и голосований; KIMI_BASE_URL из env; GIGACHAT_VERIFY_SSL_CERTS для проверки TLS.
• Контент: каталог темпераментов — 20 записей (D-94, D-95).
• Docker (TASK-56): Dockerfile и wiring в compose для game-core, api и frontend (замена заглушек); frontend — standalone-сборка NextJS, api — uv-образ; healthcheck'и и depends_on у всех сервисов.
• Фиксы по пилотам:
• RULE-094/D-111: реплики длиннее 6 предложений обрезаются без ретрая (D-23-промпты ≤ 4 для greeting/farewell), клэмп всех реплик — знакомство/обсуждение/прощание;
• D-112: баланс ролей — дон входит в лимит чёрной команды, MAFIA = лимит − 1 при игре с доном;
• example.env — полный набор runtime-переменных.
• Документация: чек-лист пилота и запросы метрик (NFR-010), обновления спеки (RULE-133, D-111, D-112), contracts.

Проверка

• ruff check, ruff format --check, ty check, uv run pytest -q — 547 passed;
• frontend npm test (vitest) — 165 passed;
• docker compose config валиден; все 4 сервиса поднимаются и отвечают: api health 200, прокси /api/v1 через frontend 200, страницы /, /sessions, /players, /stats, /test — 200;
• 3 пилотные сессии на реальном GigaChat (журналы в volume journals).

# Контекст Эпик E7 (пилот на реальных LLM): цель — прогнать live-сессии на GigaChat, собрать метрики NFR-010 и подготовить продакшен-развёртывание через Docker Compose. # Что входит - Live-процесс game-core (TASK-53): session runner, python -m game_core, активация сессии, wakeup/start-now, восстановление запланированных сессий из журнала (D-81), управление жизненным циклом соединения с БД. - LLM-агент (TASK-53): живой игрок на GigaChat/Kimi через LangChain, промпты знакомства и голосований; KIMI_BASE_URL из env; GIGACHAT_VERIFY_SSL_CERTS для проверки TLS. - Контент: каталог темпераментов — 20 записей (D-94, D-95). - Docker (TASK-56): Dockerfile и wiring в compose для game-core, api и frontend (замена заглушек); frontend — standalone-сборка NextJS, api — uv-образ; healthcheck'и и depends_on у всех сервисов. - Фиксы по пилотам: - RULE-094/D-111: реплики длиннее 6 предложений обрезаются без ретрая (D-23-промпты ≤ 4 для greeting/farewell), клэмп всех реплик — знакомство/обсуждение/прощание; - D-112: баланс ролей — дон входит в лимит чёрной команды, MAFIA = лимит − 1 при игре с доном; - example.env — полный набор runtime-переменных. - Документация: чек-лист пилота и запросы метрик (NFR-010), обновления спеки (RULE-133, D-111, D-112), contracts. # Проверка - ruff check, ruff format --check, ty check, uv run pytest -q — 547 passed; - frontend npm test (vitest) — 165 passed; - docker compose config валиден; все 4 сервиса поднимаются и отвечают: api health 200, прокси /api/v1 через frontend 200, страницы /, /sessions, /players, /stats, /test — 200; - 3 пилотные сессии на реальном GigaChat (журналы в volume journals).

geekiot added 16 commits 2026-08-11 11:27:48 +05:00

docs(contracts): GigaChat TLS check via env GIGACHAT_VERIFY_SSL_CERTS (TASK-53) fb9e68ec82

feat(config): GIGACHAT_VERIFY_SSL_CERTS env for GigaChat TLS check (TASK-53) 19652a2f4c

feat(data): temperaments catalog content — 20 entries (D-94, D-95, TASK-53) 2ddff07b4f

docs(plan): E7 pilot dashboard queries; jq in devShell (TASK-53) 0281283c79

docs(contracts): Kimi base URL via env KIMI_BASE_URL, default Kimi Code (TASK-53) 9155d8a944

fix(game-core): Kimi base URL from env, default Kimi Code endpoint (TASK-53) be8c14dfe9

feat(game-core): live LLM player agent + greeting/continue_vote prompts (TASK-53) 40ade3df1d

feat(game-core): live process — session runner, python -m game_core, activation fix (TASK-53) ... d737fd8d7d

runner.py composes phase automata around PlayerAgent (live providers, LangGraph
Postgres memory, stats materialization); __main__.py hosts the admin API and
the scheduler loop with NFR-003 startup recovery. Activation moves from
tick/start-now to open_session: tick/start-now mark_active broke open_session
(active->active) and the scheduled->skipped path (FR-003); per the state
machine the session becomes active after a successful selection.

feat(game-core): Dockerfile and compose wiring — TASK-56 increment (TASK-53) ... bf554fb2fe

- Dockerfile: uv workspace sync --no-editable (editable .pth pointed to
  builder-stage paths, import game_core failed in the runtime image);
  prompts/*.md packaged via hatchling package data; migrations, alembic.ini,
  data copied; healthcheck on /admin/settings.
- docker-compose.yml: game-core service replaced the stub with the built
  image (DATABASE_URL to postgres, JOURNAL_DIR named volume, loopback-only
  port 8001 per D-77).
- admin_api: ADMIN_HOST overridable via env for containers (default stays
  127.0.0.1 per D-77); test reloads the module for both cases.

docs(plan): TASK-53 checklist passed — readiness for the pilot 648a6388eb

docs(env): complete example.env — ADMIN_HOST, GAME_CORE_ADMIN_URL, DATA_DIR, runtime settings overrides 54e26feb04

fix(game-core): wakeup/start-now execution, scheduled journal recovery, memory connection lifecycle (D-81) ... ee4ba47fc4

- WakeupScheduler moved into scheduler.py with consume_forced(); start_now
  creates a planned_at=now session; the main loop drains forced wakeups first.
- SessionStore.status_of() added; __main__ recovers scheduled journals on
  restart so their seq/pos reload correctly (NFR-001) instead of crashing.
- _postgres_memory uses an explicit psycopg connection (autocommit) held for
  the session lifetime and closed after the run; PlayerMemoryStore.close().

Tests: TestWakeupScheduler (forced consumed before tick, skipped when already
opened, none by default, rejected while active), scheduled journal reopen and
corrupt-tail failure in test_sessions.py. Test isolation: LangGraph store
tables are dropped after the postgres roundtrip and admin settings are
restored, so the shared session-scoped DB stays migration-clean.

fix(game-core): speech over 6 sentences is truncated, not re-picked (RULE-094, D-111) ... dee7c6b46a

GigaChat ignores the prompt limit and produced 4-5+ sentence speeches, which
counted as invalid answers and ran the D-99 backoff until technical_dropout
(observed in pilot E7). The prompt still asks for at most 4 sentences
(PROMPT_SPEECH_SENTENCES), the game-core hard limit is 6 and longer speeches
are truncated via truncate_sentences() instead of raising.

Spec: RULE-094/D-23 updated, new D-111 in registry, NFR-010, llm-cost,
PlayerSpeech schema description, ASSUMPTIONS A-5, CHANGELOG; historical plan
note in e3-game-core.md. Tests: truncation instead of re-pick, unit tests for
truncate_sentences (incl. trailing fragment without a terminator).

fix(domain): count don inside the black-team limit in role composition ... 972e96560f

The "Мафия" column in game-rules §7 means the whole black team including
the don (RULE-133, D-112): at 8–9 players the black team stays at 2
(1 mafia + don), and the 10-player fork is 2–3 black including the don.
Pilot 2 showed 4 black at 10 players gave black an instant win.

Updates role_composition, the game-core fixture (don now used for the
reflection failure), tests (incl. fixture event counts 249 → 315), and the
spec docs (RULE-133, D-112).

fix(game-core): clamp all spoken replies to RULE-094 sentence limit ... 4cce2903da

Greeting, discussion, and farewell speeches now enforce the D-111
hard limit of 6 sentences via _clamp_speech. Previously only the
discussion phase clamped in DiscussionPhase, so GigaChat introduction
and farewell replies could reach up to 21 sentences (pilot 3).

Prompts for greeting/farewell also carry the prompt-side <= 4
sentences instruction (D-23) via max_sentences template var
(PROMPT_VERSION 2 -> 3); spec RULE-094 updated accordingly.

build(docker): real api and frontend images in compose (TASK-56) dc4ccdef63

geekiot changed title from feat/pilot-e7 to feat(e7): live-сессии пилота — реальные LLM, run-процесс, Docker-образы, фиксы лимитов речи и баланса (TASK-53…TASK-56) 2026-08-12 22:49:32 +05:00

geekiot merged commit 0bd057cc0f into main 2026-08-12 22:52:26 +05:00

geekiot deleted branch feat/pilot-e7 2026-08-12 22:52:26 +05:00

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

geekiot/mafia-agents!8

No description provided.