ci-fix-tests

Auto-fix CodeceptJS Tests on CI

A non-interactive auto-repair loop. Reads the failed-tests baseline from output/ (where aiTrace left trace artifacts), attempts a small set of safe fixes, reruns only the failing scenarios, compares, and writes a report at output/ci-fix.md. Rolls back any edit that didn't help.

Conservative by design. CI is not the place to restructure tests, change abstractions, or fix anything that needs human judgement. The goal is "pass the run". Anything riskier is escalated through the report.

This skill assumes a working git checkout — rollback uses git checkout -- <file> to restore originals.

Workflow

1. Read the baseline failures

Use the codeceptjs-run-analysis skill against the existing output/trace_*/ artifacts to enumerate every failing scenario. Record:

• the set of failing tests (file path + scenario name)
• the failed step in each, and the locator / wait / assertion involved
• a cluster signature per failure (so unique vs systemic is clear)

Save this as the baseline — count, failing_set, clusters.

2. Pick safe fix candidates

Only attempt fixes from this allowed list:

• Locator drift — the failed locator no longer matches anything but a similar element exists. Use the codeceptjs-exploration skill (headless) to find candidates; pick a short locator scoped to a stable region — I.click('Save', '.toolbar') — in this order: visible text / accessible name → ARIA { role, name } → $name (customLocator) → composed CSS. Replace the locator at the failing step only.
• Missing wait for a spinner / loader / modal — the failed step's ARIA snapshot shows a spinner or skeleton present, or the target element appears later. Add a single matching I.waitFor* immediately before the failing step.
• I.wait(N) replacement — when a hardcoded sleep is the only thing between a failing assertion and a passing one and the gating element is identifiable, replace the sleep with a specific waitFor*.

Do not attempt:

• Auth / credential / env-var fixes — those are CI configuration, not test code.
• Refactoring (extracting page objects, custom helpers, splitting scenarios).
• Renaming Scenarios or tags — CI's --grep and pipelines reference them.
• Adding retries: N to a Scenario — masking flake is a human decision.
• Anything outside the failing test file (no codecept.conf.*, no helper, no PO edits).

If a failure doesn't match any allowed-fix pattern, skip it and record it as unresolved for the report.

3. Apply edits and rerun

Track every modification: file path, line number, before/after snippet, fix kind. Then rerun only the failing scenarios under the same config CI used:

npx codeceptjs run --grep '<scenario>|<scenario>|...'

Find the right config via package.json scripts or the CI workflow file. Match the worker count if run-workers was used originally.

4. Compare against the baseline

Use codeceptjs-run-analysis again to enumerate the failing set after the fix.

5. Decide: keep or roll back

• SUCCESS_FULL / SUCCESS_PARTIAL → keep the edits.
• NO_IMPROVEMENT / REGRESSED → roll back every edit with git checkout -- <files>. The on-disk state at the end of the skill must be identical to what it was at the start.

A regression is treated the same as no improvement — the attempt didn't work, the working state is what was on disk before. Don't try to "undo only the bad ones" — the whole batch goes.

6. Write the report (always)

Write output/ci-fix.md regardless of outcome — CI reads it to surface results.

Report format

The first line of Status: is the machine-parseable signal. The rest is for the human reading the failed CI build.

Things to avoid

• Editing files outside the failing test files. Config, helpers, page objects, and CI workflows are off-limits in this mode.
• Attempting more than one fix kind on the same scenario in one pass. Fix one thing, observe, then try another.
• retries: N to mask flakiness — human decision.
• Skipping rollback on NO_IMPROVEMENT. Even no-op edits can confuse the next run or the next reviewer.
• Writing a partial or missing output/ci-fix.md. CI depends on it; the absence of a report is itself a failure mode.
• Running the whole suite for the verification step — only the originally-failing scenarios, to keep CI time bounded.

Related skills

• codeceptjs-run-analysis — baseline, clustering, post-fix comparison
• codeceptjs-exploration — replacement locators for drift
• codeceptjs-fundamentals — config, helper, CI environment