| name | agf-running-sit-tests |
| description | Use when an execution-layer dev (frontend-dev / backend-dev / ai-agent-dev / ml-engineer / miniapp-dev) has finished feature code + Unit tests and is about to enter code-review. Provides the SIT scope, environment, AC-driven integration walk, and evidence sink (progress/<role>.md). SIT is now a dev-owned step, not a separate QA stage. |
Running System Integration Tests (SIT)
Use this skill when:
- A dev finished feature code + Unit tests and is about to enter code-review
- A dev needs to verify a fix doesn't regress integration (API ↔ DB ↔ external)
What SIT is — and is not
SIT verifies that independently-developed components compose correctly — frontend ↔ backend ↔ DB ↔ external services. It is NOT:
- Unit tests (already covered by
pytest / vitest on the branch before SIT)
- E2E tests (real browser + real user journey — run downstream after code-review)
- UAT (business owner final sign-off — product-lead drives, downstream of E2E)
You just wrote the Unit tests, so you have the clearest picture of the unit-vs-integration boundary. If a failure reproduces by running just the backend unit tests with mocks, it's a unit-level miss, not a SIT finding; fold it back into the unit suite rather than writing it up as a SIT defect.
Pre-conditions
If any precondition fails: SendMessage product-lead, do not proceed.
Environment
Default SIT environment is local docker-compose, brought up via the root Makefile（本地开发一键 SSOT，依赖管理走 uv——见 ADR-000；不要手写 pip/alembic/uvicorn 命令绕开 uv.lock）:
make dev
make migrate
For LLM-dependent features, set provider env vars per agf-wiring-multi-llm-sdk skill. Use a dedicated SIT API key with a hard daily spend cap so a runaway test doesn't drain the budget.
Execution sequence
Walk every AC from docs/changes/<change>/tasks.md（旧 feature fallback docs/prd/[feature].md）. For each AC at the integration layer:
- Setup — record exact starting state (DB rows / fixture / user logged in)
- Action — step-by-step what triggers the integration (frontend button click → API call → DB write → external service callback)
- Expected — copy the AC verbatim
- Actual — capture HTTP status, response body excerpt, DB row diff, log lines
- Verdict — Pass / Fail / Blocked (with reason)
Verify, don't assume. Don't write "Passed" because the code looks right. Run the action, capture the actual response, compare. Per .claude/standards/coding.md "Verify before assert" — paste the actual command output into the progress entry.
Evidence collection
For each AC verdict:
- HTTP responses:
curl -i output or browser DevTools Network export
- DB state changes:
psql -c "SELECT ..." before/after dumps
- Logs: relevant lines from backend console + frontend console
- Screenshots: only when UI behavior is the subject of the AC
Keep evidence inline in the **SIT 证据** section of progress/<role>.md (small payloads). Large/binary artifacts → store under progress/evidence/[feature]/ and reference by path.
Evidence Output
SIT no longer produces a standalone report under docs/qa/. All evidence lives in progress/<role>.md under the **SIT 证据** section of the task entry — pass = single AC-tagged line (✅ AC-N (integration): <一句话>), fail/blocked expands命令 + 输出 + 偏差.
Format authority: .claude/standards/ac-lifecycle.md → 完整条目格式 (5-section format 状态 / Skills / SIT 证据 / 质量门 / 下一步 + the **SIT 证据** block rules). The progress file is archived into docs/qa/[feature]-process-log.md after UAT sign-off (product-lead), so SIT evidence survives without a separate report artifact.
Hand-off
完成 SIT 自跑后，先自检再报告——跑 bash .claude/scripts/agf-advisory.sh progress/<role>.md（advisory 机筛统一入口，ADR-026 D2）机筛 placeholder / 漏证据 / pass 含失败 token / 质量门矛盾（advisory，不阻断），把 flag 的修掉再 SendMessage，省一轮 code-review 打回。这是全链路唯一一次机筛：reviewer 的 SIT Audit 不重跑 advisory（ADR-011 决策 2 + ADR-026 D2）。然后：
- All AC pass: SendMessage product-lead — "Implementation + SIT done, ready for code-review"，引用
progress/<role>.md 条目路径 + 时间戳
- Some AC fail / blocked: 仍写完
progress/<role>.md 的 SIT 段（如实记录 fail / blocked），但 SendMessage 写明阻塞原因与影响范围，由 product-lead 决定本轮是否就修
- 不再直接 SendMessage 下游 QA 角色（已不参与 SIT）；E2E 由 product-lead 在 code-review (含 SIT Audit) 通过后单独启动
Anti-patterns
- ❌ Marking AC as Pass without actually triggering the action
- ❌ Skipping ACs because "they look obviously fine"
- ❌ Lumping multiple ACs into one "Passed all" line — every AC needs its own verdict + evidence
- ❌ Running SIT against a stale environment (forgot to apply migrations / restart server)
- ❌ Using production API keys for SIT
- ❌ Skipping SIT to ship faster — code-reviewer will reject the SIT Audit (see
.claude/agents/code-reviewer.md "SIT Audit" section) and you'll be sent back
- ❌ Treating a missing unit test as a SIT defect — fold it back into the unit suite, don't smuggle it through integration
