gh-actions-diagnostics

name	gh-actions-diagnostics
description	Diagnose GitHub Actions failures with gh by tracing failing runs, jobs, and step logs. Trigger on CI red/failing checks, flaky workflows, rerun decisions, or release pipeline breakages.
compatibility	Requires GitHub CLI and workflow log visibility; collector can auto-download gh into ~/.cache/prism/bin (set GH_VERSION / GH_BIN / PRISM_BIN_DIR).
metadata	{"prism-agents":"github-cli"}

GH Actions diagnostics

Inputs expected

• Workflow name, run URL/ID, branch/SHA, or PR number.
• Desired scope (latest failure vs historical flake pattern).

Command workflow (gh CLI)

1. Locate candidate runs.
   • gh run list --limit 20
   • gh run list --branch <branch> --limit 20
   • gh run list --workflow "<workflow>" --limit 20
2. Inspect failing run details and jobs.
   • gh run view <run-id>
   • gh run view <run-id> --json jobs,conclusion,createdAt,updatedAt,headSha,event,workflowName
3. Pull failing logs.
   • gh run view <run-id> --log-failed
   • For full context: gh run view <run-id> --log
4. Pattern-match root causes.
   • Infra/transient: timeouts, network, rate limits.
   • Deterministic: test assertion, migration mismatch, unknown flags.
   • Config drift: missing secrets/env/permissions.
5. Recommend next action.
   • rerun safe? (transient)
   • patch required? (deterministic/config)
   • handoff target (k8s/argo/docs) if failure is downstream.

Output requirements

• summary: short root-cause hypothesis with confidence.
• findings: failing job, failing step, first concrete error, recurrence signal.
• artifacts: include run/job IDs and key log snippets.
• confidence: high when error is explicit and reproducible; medium/low otherwise.

Guardrails

• Do not trigger reruns or workflow dispatch without explicit approval.
• Clearly separate observed evidence from hypotheses.
• If repo/run identifiers are missing, return insufficient_evidence with a callback request to parent for exact run URL/ID.
• Do not output fabricated file paths, run IDs, or generic placeholders as observed evidence.