fix-from-logs— agent skill

Output Contract

Produces per bug cluster: (1) one targeted code edit at the confirmed root cause location, (2) one named typed pytest function with mocks, (3) one-line root cause statement in the test comment.

Does NOT produce: architecture refactors, style fixes, unrelated test coverage, multiple competing fixes.

Hands off to: CI / pytest — the final action is confirming the full suite is green.

Process

Phase 1 — Ingest & Triage

1. Accept log input from paste, file path, or stdin reference.
2. Filter to FATAL/ERROR severity first; set INFO/DEBUG aside unless they precede the first error.
3. Cluster related errors: cascading failures share one root cause — group by error type, file, and timestamp proximity before analysing individual lines.
4. For each cluster, extract: error type, file path, line number, function name, request/correlation IDs, timestamp window.
5. Output a structured error inventory (1–3 clusters max; surface the highest-severity first).

Read references/rca-framework.md — Section 1 (triage sequence) and Section 2 (failure mode glossary).

Phase 2 — Localize (file → function → line)

1. Read the flagged file(s) in the codebase using the paths extracted in Phase 1.
2. Trace the call chain backward from the error site — the logged error is almost always a symptom, not the cause.
3. Apply 5 Whys: for each "why did this fail?" step, read the upstream caller or dependency until reaching an actionable root cause (something changeable in this codebase, not a framework or stdlib).
4. Read 20–30 lines of surrounding context in the affected file to understand: naming conventions, type usage, existing mock patterns, and error handling style. The fix must be native to this style.
5. Form 2–3 ranked hypotheses per cluster. Each hypothesis must have: root cause statement, location (file:line), confidence level (high/medium/low), and supporting evidence from the log + code.

Read references/rca-framework.md — Section 3 (5 Whys walkthrough).

Phase 3 — Plan Mode Gate (human approval required before any edits)

1. Call EnterPlanMode.
2. Present the following plan structure for each bug cluster:

   ## Bug [N]: [short description]
   Root cause: [one sentence, file:line]
   Evidence: [log excerpt + code snippet]
   Confidence: high / medium / low

   ### Proposed fix
   File: [path]
   Change: [one-sentence description of what changes and why]

   ### Test plan
   Test file: [path, e.g. tests/test_module.py]
   Test name: test_[exact_scenario_description]
   Mocks needed: [list of external dependencies to mock, e.g. payment gateway, DB call]
   Assertion: [what specific state the test will assert]

1. Wait for human to approve, edit, or redirect before proceeding.
2. Call ExitPlanMode only after the human confirms.

Phase 4 — Write Failing Test First

If you cannot write a test that fails on the bug, the root cause is wrong — return to Phase 2.

1. Create or append to the appropriate test file (e.g. tests/test_<module>.py).
2. Write a typed pytest function following references/test-writing-rules.md.
3. Use unittest.mock.patch or pytest-mock to isolate external dependencies (DB, HTTP, filesystem, time).
4. Assert the specific violated state — not just pytest.raises(Exception).
5. Run the test: pytest <test_file>::<test_name> -v.
6. Confirm it fails for the right reason (failure message matches the bug) before continuing.
7. If the test passes immediately: root cause was misidentified — return to Phase 2.

Read references/test-writing-rules.md before writing any test.

Phase 5 — Fix & Verify

1. Implement the minimum change that makes the failing test pass. No opportunistic cleanup, refactoring, or unrelated improvements.
2. Match the surrounding code style exactly: type annotations, naming conventions, error handling patterns.
3. Run: pytest <test_file>::<test_name> -v → must be green.
4. Run the full test suite: pytest → confirm no regressions.
5. If any existing test breaks: investigate before reverting. The fix may have exposed a pre-existing wrong test — surface this to the human rather than silently reverting.
6. After 3 failed fix attempts: stop. Present a clear blocker statement and ask the human for direction.

Rules

1. Never propose a fix before completing Phase 2 localization.
2. Never skip Phase 3 — EnterPlanMode before any file edit, every time.
3. Always write and run the failing test before writing the fix. Always.
4. If a freshly written test passes immediately: the root cause is wrong — back to Phase 2.
5. Fix scope = minimum change to make the failing test pass. No scope creep.
6. Test names must describe the exact failure scenario (test_checkout_raises_when_payment_returns_none, not test_edge_case).
7. All tests must use typed signatures: def test_foo(mock_bar: MagicMock) -> None:.
8. After 3 failed fix attempts: escalate to human with a clear blocker — never compound patches.
9. When the user flags something to never do again: update the relevant rule here or in the reference file immediately.
10. When a fix + test pair is approved: save it as an example in assets/approved-examples/.

<!-- Built with Agent Engineer Master — get your own production-ready skill: www.agentengineermaster.com/skill-engineer -->