| name | debug-backend |
| description | Systematic 5-step backend debugging flow for AI-coded FastAPI apps. Load this skill when a bug is reported, a test fails, or unexpected behavior appears in any backend layer. Forces layer isolation before touching code — prevents the "touch 8 files and make it worse" pattern. |
debug-backend
Stop. Do not edit any file yet. Work through these 5 steps in order.
Step 1 — Locate the layer
Run these greps to find where the bug lives. Check one layer at a time.
grep -n "HTTPException\|status_code\|Depends(" app/routers/<file>.py
grep -n "def \|raise \|return " app/services/<file>.py
grep -n "SELECT\|filter\|where\|FOR UPDATE" app/repositories/<file>.py
grep -n "httpx\|raise\|return\|except" app/providers/<file>.py
Decision: fix only in the layer where the bug lives. Never fix a service bug in the router.
Step 2 — Check the 3 most common AI antipatterns
These cause 80% of production bugs in AI-coded backends. Grep for each:
grep -rn "AsyncSession\|Session\|from sqlalchemy" app/services/
grep -rn "return {" app/providers/
grep -c "await.*service\|await.*repo\|if.*balance\|FOR UPDATE" app/routers/*.py
| Result | Root cause | Principle |
|---|
| Session in services | Hexagonal boundary broken | B1 |
return {} from provider | Missing ACL | B3 |
| Logic in router | Layer violation | A3 / code-standards |
No user_id filter in query | Multi-tenancy gap | security-rules |
Step 3 — Write a reproducing test first
Before changing any production code, write a failing test that captures the exact bug.
async def test_<bug_description>():
repo = FakeCreditsRepo()
repo.seed_balance("user-1", Decimal("5.00"))
service = CreditsUserService(repo=repo)
with pytest.raises(ValueError, match="Insufficient balance"):
await service.charge("user-1", Decimal("10.00"), idempotency_key="idem-1")
assert await repo.get_balance("user-1") == Decimal("5.00")
Run it: pytest tests/unit/test_<file>.py::test_<bug_description> -v
It must be RED before you fix anything.
Step 4 — Fix in the correct layer
Minimum change. No opportunistic refactoring. No unrelated cleanup.
Router bug → fix only app/routers/
Service bug → fix only app/services/
Repo bug → fix only app/repositories/
Provider bug → fix only app/providers/
After fixing: run the reproducing test. It must turn GREEN.
pytest tests/unit/test_<file>.py::test_<bug_description> -v
Step 5 — Run architecture lint
bash scripts/lint-architecture.sh
Must exit 0. If a check fails, the fix introduced a new architecture violation — revert and fix properly.
Common error → root cause table
| Error | Where to look first | Likely cause |
|---|
422 Unprocessable Entity | Router — Pydantic schema | Wrong field type or missing required field |
500 Internal Server Error | Service — exception not caught | Provider returned unexpected shape (dict not JobResult) |
KeyError: 'url' | Provider — ACL missing | Provider response changed, parser not updated |
InsufficientFunds on correct balance | Repo — concurrent hold | Two requests raced, no FOR UPDATE lock |
401 on authenticated route | Router — Depends(get_current_user_id) missing | Auth dependency not wired |
| Test passes, prod fails | Repo — fake vs real diverged | FakeCreditsRepo doesn't match CreditsRepoProtocol |
| Logs silent on error | Service — bare except: | Exception swallowed, add logger.exception(...) |
Verification
The skill was applied correctly when: