CI Fix

A PR's checks are red. Find the actual failure (not the noise around it), figure out whether the bug is in the code, the test, or the environment, and fix it at the right layer.

Use this skill when

• The user pastes a CI failure log and asks what is wrong.
• The user says "my PR's checks are red" or "the build broke after rebase" and wants to know why.
• The user has a flaky test that fails in CI but not locally.
• The user wants to fix a recurring CI issue once and for all.

Do NOT use this skill when

• The user wants to design a new CI pipeline from scratch. Out of scope for the MVP.
• The CI is red because the application code is genuinely broken and the failure mode is interesting. Use debugging; come back here once the root cause is in hand.
• The user wants to "skip the failing test to unblock the merge". Push back — that is technical debt with a CVSS score.

Procedure

1. Read the actual failure line. CI logs are mostly noise — banners, setup output, retries. Search for Error, FAIL, exit code, the test runner's failure prefix, or the build tool's failure prefix. The first failure is usually the one that caused all the others.
2. Classify the failure:
   • Build — compilation, type, lint, packaging.
   • Test — unit, integration, e2e, snapshot, smoke.
   • Lint / format — style enforcement.
   • Deploy — auth, infra, environment promotion.
   • Infra — runner OOM, disk full, network flake, container pull rate limit. The category determines who is responsible for the fix.
3. Reproduce locally. Run the exact failing command from the workflow file with the same arguments and (where possible) the same Node/Bun/Python version. A reproducible local failure is 90% of the fix; "works on my machine" without checking the versions is a smell.
4. Check for environment differences when local doesn't reproduce:
   • Runner OS (Ubuntu version, macOS arch).
   • Toolchain version pin (Node 20 vs 22, Bun 1.1 vs 1.3).
   • Environment variables present in CI but not locally (or vice versa).
   • File system case sensitivity (Linux yes, macOS often no, Windows no).
   • Locale (LANG, LC_ALL) — date formats, sort order.
   • Parallelism / concurrency settings (CI often runs serially or with fewer cores; tests built for parallel runs may pass at low concurrency).
   • Time zone (CI is usually UTC).
5. Fix the root cause, not the symptom. Anti-patterns to avoid:
   • Pinning a dependency to dodge a real bug.
   • Retrying a failing test until it passes (if: failure()-style reruns).
   • Adding continue-on-error: true to make the job green.
   • Marking a real failure as a "known flake" without an issue filed.
   • Disabling a check on main because it is annoying on PRs. Each of these makes the next failure harder to find.
6. Verify the fix in a fresh CI run. Don't ship the fix and walk away — a green checkmark on the same commit that was failing is the proof.
7. If the failure was caused by a flaky test, write a comment near the test (and ideally an issue) describing the symptom, the intermittent root cause, and what would make the test deterministic.
8. When the CI config itself is at fault (wrong runner, missing secret, broken matrix), edit the workflow file in the same PR as the fix and call it out in the PR description. A "fix CI" commit on main without explanation is a tax on every future debugger.
9. Watch out for caches. A green CI run on a cached build can hide a real failure that a fresh checkout would surface. If a fix seems to "just work", bust the cache and re-run before trusting it.
10. Distinguish new failures from pre-existing red main. If main is already red, the PR may be inheriting the failure. Fix main first, then rebase the PR; mixing the two makes review impossible.

Examples

In scope: "Tests pass locally but fail in CI with Cannot find module './foo'."

→ Case-sensitivity mismatch. The file is Foo.ts, the import is ./foo. Linux is case-sensitive, macOS usually isn't. Fix the import (or rename the file) to match exactly, and add a lint to catch the next one.

In scope: "Snapshot test fails in CI on every PR."

→ Read the diff. If the new snapshot is correct, update it and commit the snapshot file alongside the change. If the snapshot varies between machines (locale, time zone, asset hash), either strip the variance from the snapshot or replace the snapshot test with a behavioral assertion.

In scope: "Build OOMs on the GitHub runner."

→ A standard runner has limited RAM. Reduce parallelism, split the build, or move to a larger runner. Don't --max-old-space-size your way out of an actual memory leak in the build.

In scope: "Lockfile drift — bun install --frozen-lockfile fails."

→ Someone updated a dependency without updating the lockfile, or the lockfile was edited by a different tool. Reinstall locally, commit the regenerated lockfile, and consider adding a CI step that fails on lockfile drift before running anything else.

Out of scope: "The deploy is failing because the prod database is down."

→ This is an incident, not a CI fix. Page someone.

Out of scope: "Make the build twice as fast."

→ Speed work without a failure is a refactor — use refactor-plan or measure first.

Self-check before responding

• Did I find the actual failure line, not just the noisiest red output?
• Did I classify the failure (build / test / lint / deploy / infra)?
• Did I reproduce locally with the same toolchain versions?
• Did I check environment differences when local couldn't reproduce?
• Did I propose a root-cause fix instead of a retry, skip, or continue-on-error?
• Did I verify the fix in a fresh CI run before declaring done?
• For flakes, did I leave breadcrumbs (comment + issue) so the next person isn't starting from zero?