Skip to content

SKILL.md

Latest commit

 

History

History
100 lines (77 loc) · 3.89 KB

SKILL.md

File metadata and controls

100 lines (77 loc) · 3.89 KB
Raw
name fix-ci
description Read every failing check in one pass, fix in dependency order, push once. Use when CI is failing, tests are broken, GitHub Actions report errors, or PR checks show red.
disable-model-invocation true
allowed-tools
Bash(gh *)
Bash(git *)
Bash(pnpm *)
Bash(./mvnw *)
Read
Grep
Glob
metadata
source version
internal
2.0.0

Fix CI

Diagnose every failing check in one pass, then push once. Fixing 2 of 5 failures burns a push cycle and a full CI run.

1. Wait for the run to finish

PAGER=cat gh pr view --json number,statusCheckRollup --jq '{
  pr: .number,
  failed: [.statusCheckRollup[] | select(.conclusion == "FAILURE") | .name],
  pending: [.statusCheckRollup[] | select(.status == "IN_PROGRESS" or .status == "QUEUED") | .name]
}'

Anything pending means you do not yet know the failure set.

2. Read the job summaries before the logs

The quality and test workflows write a markdown table to $GITHUB_STEP_SUMMARY naming, per failed check, the exact command that fixes it, and emit the same text as ::error:: annotations. That table is the prescription. Logs are only for what it cannot express — a test assertion, a type error.

RUN_ID=$(PAGER=cat gh run list --branch "$(git branch --show-current)" --limit 1 \
  --json databaseId,conclusion --jq '[.[] | select(.conclusion == "failure")][0].databaseId')
PAGER=cat gh api "repos/{owner}/{repo}/actions/runs/$RUN_ID/jobs" \
  --jq '[.jobs[] | select(.conclusion == "failure" and (.name | test("CI Status|all-ci") | not)) | {id, name}]'

Then per failed job id:

PAGER=cat gh api "repos/{owner}/{repo}/actions/jobs/$JOB_ID/logs" 2>&1 | tail -80

Read every failure before changing anything.

3. Fix in dependency order

Formatting, then lint, then types, then behaviour — an earlier fix routinely erases a later failure. Each leg's annotation names its own command; this table is only what the annotation cannot tell you.

Failure What it actually means
routeTree.gen.ts is stale Only a Vite build writes it. cd webapp && pnpm run build, then commit the file.
README images are stale The storybook job runs export:readme-assets after test:storybook, so the job goes red having printed a clean pass line. Run pnpm --filter webapp run export:readme-assets and commit docs/images/readme.
Biome version skew check:biome-pin compares package.json, node_modules and the $schema URLs in both biome.jsonc files. Fix the pin; do not reformat.
Migrations gate A changelog that reached main was edited, renamed or deleted, or a master.xml <include> was not appended at the end. Fix forward with a new changeset; never edit the released file.
verify-changesets The PR touches shipped code with no .changeset/*.md. /land-pr step 9 has the rules.
App Server leg red on a docs-only PR Expected, not a misconfiguration: docs/** is inside the application-server paths filter, because docs:lint and check:diagrams run on that leg.

4. Reproduce locally before pushing

pnpm run format
pnpm run check

check runs every leg CI runs except those needing Docker or a live credential — docs:lint included. If check is green and CI is not, the difference is one of those.

The server reactor builds generated clients before the application. Run the unit tier with:

cd server && ./mvnw -pl application -am test -Dsurefire.includedGroups=unit -T 2C --batch-mode -q

-Dgroups is ignored: the POM binds ${surefire.includedGroups}, and a POM element beats the -Dgroups user property. server/AGENTS.md § Build traps has the other three tiers.

5. Commit and push once

git add -A
git commit -m "fix(<scope>): resolve ci failures"
git push
PAGER=cat gh pr checks --watch

If the same check fails twice with the same diagnosis, re-read the log rather than retrying the fix.