agentleFS
andymaiandymai/brepjs/.claude/skills/ci-triage/SKILL.md

ci-triage

Skill142 starsChanged 8 days agoApache-2.0View on GitHub

This skill should be used when a brepjs GitHub Actions job is red or behaving oddly on github.com (a remote CI run, not a local pre-commit/pre-push hook) — "CI failed", "ci-pass is failing", "npm ci failed with EUSAGE/ETARGET in CI", "test shard timed out", "function-lookup.md diff failed in the build job", "check:patterns fails in CI on code I didn't touch", "npm publish workflow failed", "docs deploy didn't run", "playground smoke is red", "OSV scan is red", "release PR won't auto-merge", or npm ci fails in a CI run after a Dependabot bump on a workspace "*" dependency. Maps each CI job's known failure modes to causes and fixes. For a hook blocking a local commit, use quality-gates or git-pr-workflow instead.

17 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

1 rule. 0 are common, 0 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 13 edited versions we read keep these.

  • 0

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    apps/playground is outside root lint + vitest: only tsc -b (via playground-build) gates it in CI.

    line 133

“Common” is measured: kept by at least 6 in 10 of the 13 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

CI failure triage

Diagnose a red or stuck brepjs GitHub Actions run. Triage in two moves: first classify which workflow and job failed and on what event (PR vs push-to-main vs manual dispatch vs deployment); then match the symptom to the table below.

Most failure modes are documented in inline comments in the workflow files themselves — cite line refs (e.g. .github/workflows/ci.yml:244-250) rather than re-deriving. This skill is the index and the recovery recipes.

Job map — what runs, when

The single required PR check is ci-pass (ci.yml:431-459). It needs every gate job and fails if any result is failure or cancelled (a skipped result passes). Its needs-list: changes, typecheck, lint, quality, build, playground-build, packages-viewer, packages-verify, packages-sheetmetal, packages-bim, voxel-wasm-rust, test, test-brepkit, size, benchmark. Deliberately not in the list: coverage.

The changes job (ci.yml:18-67) path-filters into three outputs — code, playground, voxel — that gate the rest. Branches named release-please--* skip the filter entirely (ci.yml:29-36): release PRs run no code jobs (they touch only CHANGELOG/version/manifest).

JobCommand(s)Event / gate
typechecknpm run typecheckPR+push, code
lintnpm run lint + format:checkPR+push, code
qualitycheck:boundaries + check:patterns + knipPR+push, code
buildnpm run build + docs:api + function-lookup staleness gatePR+push, code
playground-buildbuild root + brepjs-bim + brepjs-sheetmetal + apps/playground (mirrors Vercel)PR+push, playground
packages-viewer / -sheetmetal / -bimtypecheck+lint+test+build per workspace (root build first for -sheetmetal/-bim; viewer builds no root)PR+push, code
packages-verify (brepjs-cad)build root+viewer+cad, then eval, smoke, smoke:standalonePR+push, code
voxel-wasm-rustcargo test + cargo clippy -D warningsPR+push, voxel only
testtest:ci --shard=N/4, matrix [1,2,3,4]PR+push, code
test-brepkitvitest run --project brepkit, then scripts/checkKernelBaseline.ts against the known-failure listPR+push, code
coveragetest:full (maxWorkers=2, timeout 180s)push-only, continue-on-error
sizesize-limit-actionPR-only
benchmarkcompare vs main, 25% regression thresholdPR-only

Consequence: a PR failure can never come from coverage (push-only); a main-only failure can never come from size/benchmark (PR-only).

Concurrency: ci-${{ github.ref }}, cancel-in-progress on non-main refs only (ci.yml:10-12).

Symptom → cause → fix

Ordered roughly by frequency.

npm ci fails with EUSAGE in every job's setup step

Lock file out of sync with package.json ("lock file does not satisfy"). Classic trigger: a Dependabot group bump silently drops the two top-level @emnapi/* peer-dep entries (package-lock.json node_modules/@emnapi/core + node_modules/@emnapi/runtime, ~lines 1180-1201) that satisfy @napi-rs/wasm-runtime's peerDependencies (~lines 2285-2286). Fix: surgically restore the dropped entries — never regenerate the lockfile from scratch. Run --package-lock-only from main to reproduce the correct shape. See the git-pr-workflow and companion-packages skills for lockfile-surgery discipline.

npm ci fails with ETARGET / 404 on an internal package (in a CI run)

An internal workspace package is pinned to a version not yet on npm. Two causes:

  1. release-please re-pin: the node-workspace plugin repins a leaf's brepjs dep to a pending (unmerged) root release version. This is why leaf release PRs are held until root merges+publishes (see auto-merge below).
  2. 0.x caret exclusion: a ^0.<older> range excludes the next 0.(x+1) minor, so an internal bump 404s any consumer still on the caret.

Fix: use an open floor range, not a pin. Main already does this — packages/brepjs-cad/package.json:60 ("brepjs": ">=18.117.1") and packages/brepjs-voxel/package.json:24 ("brepjs-voxel-wasm": ">=0.2.0"). Full mechanics in the release-publishing skill.

Note: for internal workspace devDeps the "*" wildcard is intentional (e.g. brepjs-cad's "brepjs-viewer": "*", package.json:81) — pinning it to a concrete version is what repeatedly broke npm ci when the pin outran what was published. A Dependabot alert on such a "*" spec is a false positive: raise the floor (>=x.y.z) if it must be cleared, don't pin it or bump the lockfile.

build job fails on git diff --exit-code docs/function-lookup.md

A *Fns.ts change added/renamed an exported function but docs/function-lookup.md wasn't regenerated (the gate at ci.yml:112-117 runs docs:generate-lookup → prettier --write → diff; pre-commit only reminds, so it's easy to miss locally). Fix: regenerate + prettier-align + commit the file — full recipe in the adding-operations skill.

quality job fails check:patterns on code the PR didn't touch

An unbaselined pattern violation reached main and now fails the quality job on every open PR at once. Fix: land a tiny baseline-bump PR first, then rebase the others. Mechanism and baseline mechanics live in the quality-gates skill.

test shard times out or stalls

OCCT WASM linear memory grows monotonically across tests; a long run over-commits the 16 GB runner into swap and a fork stalls past the 90s per-test timeout (#1102, rationale at ci.yml:244-250). The 4-way shard exists precisely so each fork exits before accumulation hits that threshold. Do not "fix" by widening the timeout. Check whether one shard is genuinely slower (real leak / new heavy test) or just an infra blip — re-run the single shard. Coverage can't shard (vitest --merge-reports can't recombine v8 coverage here), which is why the coverage job runs single-runner with a raised timeout. Test-authoring details in the writing-tests skill.

test-brepkit fails against the known-failure list

The job ignores vitest's exit code and compares the failures with tests/knownFailures/brepkit.json, where each file names its kernel issue. Each ::error:: line names the cause:

  • new brepkit failure: the change broke a test on brepkit only. Fix the code when brepkit is right. When brepkit differs by design, register a divergence in tests/helpers/kernelDivergences.ts. Only a real kernel fault goes on the list, after it is filed in andymai/brepkit.
  • listed test no longer fails: brepkit (usually a brepkit-wasm bump) or the change fixed it. Remove it by regenerating the list.
  • listed file did not run or no tracking issue: a renamed test file, or a new file on the list with TODO tracking.

Regenerate from a full run: npx vitest run --project brepkit --reporter=json --outputFile=brepkit-report.json, then npx tsx scripts/checkKernelBaseline.ts brepkit --update brepkit-report.json. The update keeps each file's tracking. A hang (one example, bench-vise, never returns on brepkit) cannot be stopped by a test timeout, because the kernel call is synchronous. Skip it through the registry before it starts.

packages-verify fails on smoke or smoke:standalone

smoke runs the built brepjs-cad CLI under plain node (catches bin/TS-load breaks that tsx/vitest hide in-repo). smoke:standalone packs the tarball into a clean project with no brepjs present (catches bundling / fallback-resolve breaks invisible in-repo). A green in-repo test with a red standalone smoke means a packaging or resolution regression, not a logic bug. See the companion-packages skill.

playground-build fails but the library builds fine

Type-resolution or bundler regression scoped to apps/playground (this job mirrors what Vercel runs). Historical example: a duplicate @types/three install broke tsc -b only inside the playground (#963, ci.yml:119-137). See the playground-examples skill.

npm publish failed

Publishes are OIDC-bound to the exact workflow filename — inlining npm publish elsewhere fails auth. Never re-run the failed push-event job; re-dispatch the same workflow. Recovery recipes:

  • Root brepjs: gh workflow run release-please.yml -f republish=true (input at release-please.yml:6-12; publish gate at :146-149).
  • A satellite (cad/bim/sheetmetal/viewer/opencascade): gh workflow run publish-<pkg>.yml -f dry_run=false --ref <release-tag> — dry_run defaults to true, and dispatch against the immutable tag, not main.

Full recipe set and the OIDC-binding rationale: the release-publishing skill.

prepack fails: "Too many files" or ".d.ts.map"

scripts/validate-pack.sh (root prepack hook, package.json:216) caps the packed tarball at MAX_FILES=600 and rejects any .d.ts.map sidecars (count must be 0). It fires during pack/publish, not PR CI, so an overflow first shows up as a red publish-brepjs after the release PR merges: the version is tagged and GitHub-released but missing from npm. rollupTypes: false emits one .d.ts per source module, so every new src/ file adds one packed file. Count with npm run build && npm pack --dry-run --ignore-scripts. If the growth is just new modules, raise MAX_FILES (ci: commit, no release) and then gh workflow run release-please.yml -f republish=true. Fix .d.ts.map by keeping declarationMap off in vite.config.ts.

"Deploy API Docs" didn't run after a main merge

docs.yml triggers on workflow_run of CI completing on main with conclusion == 'success'. A red CI on main blocks the docs deploy. This is exactly why coverage is continue-on-error (ci.yml:260-266): an informational threshold dip must not flip CI's conclusion to failure and starve the docs gate. Emergency redeploy: gh workflow run docs.yml (workflow_dispatch bypasses the CI gate, main-only).

"Playground Smoke" is red

playground-smoke.yml runs on deployment_status from Vercel, loads the deployed playground headless, and requires the WASM kernel to reach Ready within 60s. A red here is a CSP / COEP / COOP / asset-path class regression that blocks engine init — not a build-class failure (the static build already passed).

"OSV Scan" is red

osv-scan.yml scans package-lock.json for known-vulnerable versions. It splits by event: scan-pr ("OSV Scan (report-only)") on pull requests never fails the merge (findings surface in the Security tab only), while scan-main ("OSV Scan (blocking)") on push-to-main and the weekly cron fails closed. So a red OSV check on a PR is informational; a red one on main is a real gate. Fix a genuine finding by bumping the vulnerable transitive dep (lockfile surgery, see git-pr-workflow).

Dependabot flags a security alert on a workspace "*" dependency

False positive. An internal devDep pinned to "*" (e.g. brepjs-cad's "brepjs-viewer": "*") is intentional — pinning it to a concrete version is what repeatedly broke npm ci. Fix: raise the spec floor to >=x.y.z if the alert must be cleared — do not pin the spec or bump the lockfile. Mechanism in the companion-packages skill.

A release PR won't auto-merge

Expected during a release train: while the root brepjs release PR (branch release-please--branches--main--components--brepjs) is open, every leaf PR (cad/bim/sheetmetal) is held; leaves auto-merge only after root merges+publishes. brepjs-opencascade is held permanently (manual, expensive WASM build). Mechanism at release-please.yml:57-133; deep dive in the release-publishing skill.

Looks like a failure but isn't

  • coverage red on main: informational only (continue-on-error, push-only, out of ci-pass). Not a gate.
  • benchmark "main install failed" warning: continue-on-error on the main-baseline install so a broken main can't deadlock the PR that fixes it (ci.yml:323-330). The PR's own benchmark still runs; only the comparison is skipped.
  • benchmark shows ±9-15% noise: shared runners vary run-to-run; the threshold is 25% for that reason (ci.yml:366).
  • held release PRs: opencascade always held; leaves held while root is open. See above.
  • skipped jobs: a skipped needs result passes ci-pass — only failure/cancelled fail it.

Blind spots — things CI never runs

  • alwaysExclude tests never execute in any CI project (vitest.config.ts): tests/brepkit-adapter.test.ts, tests/brepkit-validation.test.ts, tests/kernel-agreement.test.ts, tests/io-stress.test.ts, plus benchmarks/, the cad/viewer packages (own jobs), apps/, and worktree dirs. They rot silently — run them locally when touching that code.
  • The occt and manifold projects never run in CI: the gates are occt-wasm (test) and brepkit (test-brepkit). Run npm run test:occt or npx vitest run --project manifold locally when touching those adapters.
  • test-brepkit skips the docs snippets: it does not run docs:extract-tests, so tests/docs/extracted.test.ts runs on occt-wasm only.
  • apps/playground is outside root lint + vitest: only tsc -b (via playground-build) gates it in CI. See playground-examples.

Stale docs — do not inherit

CONTRIBUTING.md (~lines 222-235) lists coverage thresholds as 83/73/64/73 and claims "CI runs the full test suite with coverage enforcement." Both are wrong. Actual thresholds (vitest.config.ts): statements 85, branches 71, functions 91, lines 88. Coverage is main-only, informational, and continue-on-error — never a PR gate. Trust vitest.config.ts, not CONTRIBUTING.md.

Additional resources

Sibling skills (each fact lives in exactly one — link, don't duplicate):

  • release-publishing — release-please serialization (root-before-leaves), OIDC filename binding, per-package dispatch recipes, republish recovery, concurrency/label races.
  • quality-gates — check:patterns / knip / boundary rule details and baseline mechanics.
  • git-pr-workflow — hooks, conventional commits, worktrees, merge / auto-merge, lockfile surgery.
  • companion-packages — workspace deps, ETARGET / build order, workspace "*" specs.
  • writing-tests — test skeleton, shard/coverage authoring, kernel divergence skips.
  • adding-operations — the function-lookup gate and export surfaces.
  • playground-examples — the three playground gates (types, geometry, thumbnail).

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.