fix-tests

Make the suite green without making it weaker.

The failure mode this skill exists to prevent is the one where a suite goes green by shrinking — a deleted assertion, a .skip, a narrowed glob, a loosened matcher. That is not a fix; it is the bug plus a lie. Every step below is built around detecting it.

1. Resolve the test profile

Read .claude/testing/profile.json. If it is missing or --reconfigure was passed, run the discovery ladder in ${CLAUDE_PLUGIN_ROOT}/shared/TEST-PROFILE.md: manifests, then runner configs, then CI workflows, then at most four questions for gaps that genuinely block a correct run.

Resolve typecheck.command and lint.command on the same pass, the way coverage.command is resolved: the profile's own value first, then package.json scripts.typecheck / scripts["type-check"] and scripts.lint, then the step in the CI job that runs it. Never ask for either and never invent one — nothing found is command: null, which a ship run skips with a stated reason. A ship run's inter-wave checkpoint has no typecheck to run unless this field is filled, and a typecheck failure is a halting kind there.

This is the only skill permitted to interview the user about tests. Every later run is zero-question.

Persist the profile and mirror it into the ## Testing block of the first instruction file that exists — CLAUDE.md, then .claude/CLAUDE.md, then AGENTS.md — between the managed markers, idempotently.

2. Run and baseline

Run the unit command under the permissions already granted to this project. This skill does not pre-approve a shell. Record total, passed, failed, skipped — this is the baseline that step 6 asserts against.

Split failures into:

• Newly broken — not failing at --baseline <ref> (or all failures when no baseline was given)
• Pre-existing — already failing at the baseline. Report only. Do not absorb them into this fix unless the user explicitly widens scope.

3. Flake check

Re-run only the failing files or test names, once, using single_file / filter_syntax.

• Still failing → a real failure; continue.
• Passed on re-run → record it in profile.known_flaky and do not "fix" it. Mention it in the report. A test that passes on retry is a flaky test, and "fixing" it usually means deleting the thing that caught a race.

4. Cluster by error signature

With --from-log, parse the log into the same clusters without an initial run — but still record a live baseline before applying any fix.

5. Fix by root cause

iterations = 0, MAX = 5

WHILE failures remain AND iterations < MAX AND NOT --dry-run:
  1. Pick the cluster covering the MOST failures
  2. Choose ONE fix — source first
  3. Apply it
  4. Re-run ONLY the affected tests
  5. Drop resolved failures; iterations++

Source first. When a test and the code disagree, the default assumption is that the code is wrong. Change the test only when you can state why the test's expectation was incorrect — and say that out loud in the report.

Never: delete a test, add .skip/.only, loosen an assertion to match observed output, narrow a glob, raise a timeout to hide a race, or catch-and-ignore an error to make a run pass.

--dry-run stops after clustering: print the diagnoses and proposed fixes, edit nothing.

After 5 iterations, stop and report the unresolved clusters and what was tried. Do not thrash.

6. Full re-run and no-regression assert

Run the full suite again, then assert against the baseline:

• total >= baseline.total
• skipped <= baseline.skipped

If either assert fails, revert the weakening change — something deleted tests or added skips. A green suite that runs fewer tests than it did before is a regression wearing a green badge.

7. Persist

• Refresh .claude/testing/profile.json (updated, known_flaky, notes) with anything learned about running tests. Preserve every key you did not resolve yourself.
• After a green full unit run, record unit.timeout_ms as that run's wall time in whole milliseconds, rounded up — a JSON integer, never seconds, a decimal or a string, any of which either aborts verification planning or reads as a far smaller allowance. Never write a value below a wall time you measured: a ship run derives its inter-wave verify budget from this field, and a value under the real runtime makes every checkpoint skip the suite. Do not apply a multiplier and do not copy a number from elsewhere — the measured time is the input.
• Confirm the ## Testing block is still a single idempotent pair of markers.
• Write at most 3 .claude/memory/failure-modes/<slug>.md entries for genuinely recurring gotchas, with index lines in .claude/memory/MEMORY.md. One-off typos do not belong in memory.
• Write .claude/metrics/fix-tests-<YYYYMMDD-HHMMSS>.json with the baseline, final counts, clusters, fixes applied, and anything unresolved.

8. Report

Baseline → final counts. Root causes fixed, one line each. Failures reverted as pre-existing. Flaky tests recorded. Anything still red, and what was tried. If you changed a test rather than source, say which and why.