Test Fix Loop

Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.

Test Fix Loop

Smallest cooking-routines skill. Hand the agent a failing test (or a set of them) and a marker file; the cooking Stop hook will keep the agent looping through fix-and-rerun cycles until either the tests pass or the iteration counter hits 10. Marker file is the trust boundary; remove it to end the loop.

When to Use

• A unit/integration test is failing and the fix is mechanical (assertion update, missing import, simple refactor)
• You want the agent to iterate without supervision until green
• You explicitly do not want the agent to ship the fix anywhere — cross-repo-refactor handles branch+push; this skill only edits the working tree

When NOT to Use

• Tests are failing because of a real design bug — fix the design first, don't loop
• Tests touch multi-repo state — use cross-repo-refactor instead
• You don't have a clean working tree — the loop edits files; commit or stash first

How to Run

1. Confirm a clean working tree: git status --short returns no output, OR you accept that the loop will edit working-tree files.
2. Identify the failing test command (e.g., npm test -- failing-test.spec.ts, bun test path/to/test, pytest tests/foo.py::test_bar).
3. Create the marker file with the agent's PID:

   mkdir -p ~/.claude/cooking
   PID=$$
   START_EPOCH=$(ps -p $PID -o lstart= | sed 's/^ *//;s/ *$//')
   printf '{"start_epoch": "%s"}' "$START_EPOCH" > ~/.claude/cooking/active-$PID
   

4. Tell Claude: "Run <failing test command> in cook mode. Fix the test until it passes."
5. The agent runs the test, edits the failing implementation/assertion, reruns, repeats. The Stop hook returns exit 2 after each turn, so the agent keeps iterating until termination.
6. The skill must remove the marker file in a finally-equivalent path (whether the loop terminates cleanly or aborts).

TERMINATION

Stop when ANY of these are true:

• The failing test command exits with status 0 (test is green)
• Iteration count reaches 10 (the cooking Stop hook hard-caps regardless of marker)
• The agent determines the failure is a real design bug, not a mechanical fix — in this case write a brief note describing the bug and exit
• The marker file is removed (manually or by the agent)

When termination triggers:

rm -f ~/.claude/cooking/active-$PID ~/.claude/cooking/iter-$PID

Anti-patterns

• Don't suppress the failing assertion to "make it green" — that's removing tests, which the spec forbids.
• Don't silently delete the test if it can't be made to pass — write a note and exit.
• Don't run this on a dirty working tree without warning the user; the loop edits files.
• Don't use this for end-to-end / browser tests where each iteration costs minutes — use investigation-loop instead and reason about the failure first.

Verification

/cook-status should show this loop's marker as active while running, and as gone after termination. The iteration counter file (~/.claude/cooking/iter-<pid>) should be ≤ 10 at any observation point.

Related

• Stop hook: hooks/cooking/stop-hook.sh
• Status command: /cook-status
• Sibling skills: investigation-loop, cross-repo-refactor, dreaming-consolidator