| name | babysit-ci |
| description | Monitor CI, auto-fix small issues, pause on bigger problems, retrigger flaky runs |
Babysit CI
Monitor CI for the current branch/PR. Auto-fix mechanical issues, retrigger flaky jobs,
pause and describe anything needing judgment.
CI Surfaces
| System | What runs |
|---|
GitHub Actions UnitTests.yml | CPU tests, Julia 1.10/1.11/1 × Linux/Windows/macOS |
GitHub Actions Format.yml | Runic formatting diff, posted as a PR comment |
GitHub Actions Documenter.yml | docs build & deploy |
GitHub Actions Downstream.yml | downstream-package compatibility (e.g. JustRelax) |
GitHub Actions SpellCheck.yml | typos check |
| Buildkite | CUDA + AMDGPU GPU tests (pipeline generated by forerunner from .buildkite/run_tests.yml) |
Step 1: Find the Runs
gh run list --branch $(git branch --show-current) --limit 10
gh pr checks <PR_NUMBER>
Buildkite logs are linked from the PR check; gh run view <RUN_ID> --log-failed covers
the GitHub side.
Step 2: Triage Each Failure
Auto-fix (commit and push without asking)
| Failure | Fix |
|---|
| Runic formatting comment | git runic main, commit the result |
| SpellCheck typo | fix the typo |
Aqua ambiguity in test_Aqua.jl | add the disambiguating method |
| Doctest output mismatch | update expected output in the docstring |
| Missing export / undefined export | fix the export in src/common.jl |
Docs build: missing docstring in API.md | add the reference |
Commit with a descriptive message, push, and keep monitoring.
Retrigger (likely flaky)
- Passed locally, failed in CI with no code-relevant error
- Timeout, network/download error,
Pkg.instantiate/registry failure
- Buildkite agent lost / infrastructure error
gh run rerun <RUN_ID> --failed
git commit --allow-empty -m "Retrigger CI" && git push
Pause and describe (needs judgment)
- A test assertion fails (tolerance, wrong values) — possible real regression
- GPU-only failure ("dynamic function invocation", CPU/GPU result mismatch) — usually type
instability, a real bug
- Precompile/load failure of the package or an extension
- Downstream workflow failure — may mean a breaking API change
- Several jobs failing with a common suspicious root cause
When pausing, report: which job(s), the key error lines, your assessment, suggested fixes.
Step 3: Confirm All Green
gh pr checks <PR_NUMBER>
Notes
- Always read the actual log before acting — never diagnose from the job name
- GPU CI needs the cluster queues (
cuda, rocm) to have available agents; a pending
Buildkite job is not a failure
- Never force-push or rewrite history to fix CI — add new commits