---

name: agent-ops-cicd-github description: "Read, react to, and operate GitHub Actions workflow runs and status checks like a careful operator — never blindly rerun, never blindly approve."

Agent ops: CI/CD on GitHub

When you, as a developer-role agent, interact with CI/CD, you are operating production infrastructure. Behave accordingly.

When to use

• Anytime you open a PR or push a branch.
• When a CI run is red and you need to decide between fix-forward, rerun, and revert.
• When you are gating a merge on status checks.
• When you are creating or modifying a workflow YAML.

The loop

1. Identify the failing job. Use github_status_checks and the workflow run page. Note the job, step, and the first error line — not just the last.
2. Classify. Flaky? Real failure? Environment? Tool-version drift? Be specific. "Tests failed" is not a classification.
3. Fix-forward vs rerun. Rerun only if:
   • The failure is a known flake (record it as a planning issue if not already).
   • The failure is in a service the change does not touch (network, registry). In all other cases, fix-forward with a commit.
4. Cross-check. Before merging, confirm:
   • All required checks are green.
   • The PR description references the planning task / Jira ticket.
   • A reviewer (human or shadow-code-reviewer) has signed off.
5. Modify workflows carefully. When editing .github/workflows/*.yml:
   • Open a dedicated PR scoped to the workflow change only.
   • Run the workflow on the PR branch with workflow_dispatch or a draft PR first.
   • Never disable required checks without a release-readiness-checklist review.

Anti-patterns

• Blind reruns that hide real failures.
• Force-merging past required checks because "the test is wrong".
• Editing workflow YAML in the same PR as a feature change — the blast radius is too different.

Tools you will lean on

• github_status_checks, github_get_workflow_run, github_rerun_workflow (when wired) — to inspect and act on runs.
• agent_terminal_run — to reproduce the failing step locally before fix-forward.
• jira_create_task — to file a planning issue for known flakes.