.github

Imported from apache/airflow (.github/AGENTS.md). Install upstream with npx skills add apache/airflow --skill .github. Copyright stays with the author.

Table of Contents generated with DocToc

• The .github folder — Agent Instructions
  • Keep the CI-minutes estimates in sync with the workflows

The .github folder — Agent Instructions

Keep the CI-minutes estimates in sync with the workflows

breeze ci approve-workflow-runs (run by workflows/auto-approve-workflow-runs.yml) decides whether to approve an external contributor's CI runs from an estimate of the job-minutes they take. The numbers live in dev/breeze/src/airflow_breeze/utils/ci_minutes_estimate.py and are medians measured on recent Tests (AMD) PR runs on public runners.

After you change the CI workflows in a way that adds, removes, splits or resizes jobs (new test groups, a changed matrix, new composite workflows, a move between runner types), refresh the estimates once the change has run on enough PRs, and in any case about once a month:

1. Run uv run dev/calibrate_ci_minutes_estimates.py --since <date> from the repository root with an authenticated gh CLI. Pick a date that covers at least a week of runs after the workflow change.
2. Copy the medians and fits it prints into the constants of ci_minutes_estimate.py, and update the "Last calibrated" line next to them.
3. When you add a new job group to the workflows, add a matching group to both ci_minutes_estimate.py (with the selective-check output that turns it on) and the JOB_GROUPS patterns of the calibration script. Unmatched jobs fall into the base group.
4. Run uv run --project dev/breeze pytest dev/breeze/tests/test_ci_minutes_estimate.py dev/breeze/tests/test_workflow_run_approval.py -xvs.

The estimates cover public runners only. Self-hosted (RunsOn) runners need their own set once there is enough data from them.