|
|
a88323716d
|
feat(stage-1-2): hierarchical sub-agent team — sister-agent + SubTask tracking
Stage 1 + 2 통합 구현 — 사용자 피드백 반영:
manager / principal / lead / junior 계층 구조 추가.
부장이 복잡도를 판단해서 팀을 동적으로 꾸린다.
Design:
- .plans/design/hierarchy.md — 전체 설계 문서 (점수/tier/plan/escalation)
Prisma schema:
- SubTask (계층 트리 + 역할/모델/state/complexity)
- SubTaskEvent (JSONL 스타일 이벤트 로그)
rails (core):
- src/hierarchy/roles.ts — 4역할 기본 config + 모델 매핑
manager/principal: gpt-5.4 / glm-5.1
lead: gpt-codex-5.3 / glm-5
junior: glm-5-turbo / gpt-5
- src/hierarchy/complexity.ts — 규칙 기반 스코어러 (7 factors, 0-100)
- src/hierarchy/planner.ts — tier → DecompositionPlan (trivial/simple/moderate/complex/massive)
+ concurrency budget 강제
- src/hierarchy/store.ts — Prisma CRUD + tree builder
- src/handoff/http-transport.ts — 실제 HTTP transport (MockTransport 대체)
- src/handoff/build.ts — config + env 기반 transport 빌더
(env RAILS_TRANSPORT_MODE + RAILS_AGENT_{STAGE}_HOST 오버라이드)
- src/server/http.ts — sub-task 엔드포인트 4개 추가
POST /api/sub-tasks
PATCH /api/sub-tasks/:id
POST /api/sub-tasks/:id/events
GET /api/pipelines/:id/sub-tasks (tree view)
sister-agent (new sub-project):
- sister-agent/ — 각 LXC 에 배포될 Node.js 데몬
- src/types.ts, roles.ts, complexity.ts, planner.ts
- src/spawn.ts — executeInvocation: 복잡도 점수 → plan → spawn 트리
simulation mode (현재는 work 를 delay 로 시뮬레이트, LLM 연동은 후속)
- src/rails-client.ts — sub-task / event push
- src/server.ts — POST /invoke 엔드포인트 (port 18801)
LXC 리소스 실측 기반 기본 concurrency:
104/106/107: 8 concurrent sub-agents
105 (narang, build 중): 6
검증: tsc --noEmit ✓ | vitest 105/105 ✓ | rails build ✓ | sister-agent build ✓
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-10 17:11:16 +09:00 |
|
|
|
30e782a32d
|
feat(sprint-005): Resilience — retry + backoff + escalation + xhigh 금지
Sprint 005 전체 구현 — F5 (중간 끊김/타임아웃 무한대기) 해결:
Resilience core:
- src/resilience/backoff.ts — exponential backoff + jitter (base 1s, cap 30s)
+ cancellable sleep
- src/resilience/classifier.ts — error → {retryable, reason}
retryable: timeout, network, rate_limit, transient
non-retryable: permission, config, invariant(ZodError)
휴리스틱: ETIMEDOUT/ECONNREFUSED/429 등 메시지 패턴 감지
+ assertAllowedThinkingTier('xhigh' 금지)
- src/resilience/retry.ts — withRetry 래퍼
non-retryable은 즉시 중단, max retries 초과시 classification 반환
- src/resilience/kill.ts — child process SIGTERM→SIGKILL grace 처리
+ 전역 cleanup handler (SIGINT/SIGTERM)
- src/resilience/escalate.ts — recordEscalation + EscalationNotifier
+ listEscalations / resolveEscalation
Prisma:
- Escalation 모델 추가 (pipelineId, reason, errorCategory, attempts, contextSnapshot)
- Pipeline.escalations 역참조
Runner 통합:
- transport.invoke 를 withRetry 로 래핑
- 실패시 classification 기반으로 자동 escalation 기록 (non-retryable만)
- RunOptions 에 maxRetries / notifier 추가
CLI:
- rails resume <pipeline-id> — escalated → idle 전이
- rails abort <pipeline-id> [-r reason] — 강제 종료
Tests (25 신규, 82 total pass):
- backoff: 기본값/지수/캡/jitter 범위/abort
- classifier: 6 error 클래스 + 3 휴리스틱 + xhigh 금지
- withRetry: 성공/재시도 후 성공/non-retryable 즉시 중단/max 초과/abort
검증: tsc --noEmit ✓ | vitest 82/82 ✓ | build ✓ | CLI ✓
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-10 15:40:50 +09:00 |
|
|
|
0af4bbc685
|
feat(sprint-001): XState FSM + Prisma + CLI 뼈대 — 결정론적 파이프라인 코어
Sprint 001 전체 구현:
Foundation:
- package.json (pnpm + Node 22 + TypeScript strict)
- tsconfig.json (strict + noUncheckedIndexedAccess)
- .env.example (DATABASE_URL, DISCORD_TOKEN, etc.)
- vitest.config.ts
Core:
- src/env.ts — Zod 환경변수 검증
- src/logger.ts — pino 구조화 로거
- src/orchestrator/events.ts — Zod discriminated union 이벤트 스키마
- src/orchestrator/context.ts — PipelineContext 타입 + 팩토리
- src/orchestrator/machine.ts — XState v5 결정론적 FSM
States: idle → planning → implementing → reviewing → deploying → done
+ retrying (exponential backoff 준비) + escalated + aborted
- src/orchestrator/persist.ts — Prisma 기반 상태 영속화
- prisma/schema.prisma — MariaDB 스키마 (pipelines, state_transitions, actor_spawns, contracts)
CLI (citty):
- rails start <project> — 파이프라인 생성
- rails status [id] — 상태 조회 + 타임라인
- rails serve — 오케스트레이터 서버 (Sprint 004 에서 완성)
Tests (9/9 pass):
- happy path (idle → done)
- REQUEST_CHANGES 재작업 루프 + max review round escalation
- retryable/non-retryable 에러 분기
- RESUME / ABORT
- context 추적
검증: pnpm tsc --noEmit ✓ | pnpm vitest run 9/9 ✓ | pnpm build ✓ | rails --help ✓
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-10 13:46:13 +09:00 |
|
