name: ci-failure-triage description: Classify a CI failure (logs + commit + repo state) as flake, infra, real-break, or dep and emit a typed runx.ci.triage.v1 routing packet for a downstream issue-intake run. Read-only: opens no item, reruns nothing, pages no one. source: type: cli-tool command: node args: - run.mjs timeout_seconds: 30 sandbox: profile: readonly cwd_policy: skill-directory inputs: logs: type: string required: true description: Raw CI failure output (build/test logs) to classify. commit: type: string required: false description: Commit SHA under test, for provenance in the packet. repo_state: type: string required: false description: Short repo state hint (clean/dirty/unknown). min_confidence: type: string required: false description: escalation_policy.min_confidence; below this the run refuses (needs_agent). runx: category: ops input_resolution: required: - logs

ci-failure-triage

CI failure triage is the first decision in incident response: a fast wrong call creates noise, and a slow right call burns the production window. This skill reads a CI failure and makes exactly one call, with cited evidence, so the next governed step starts from a typed decision instead of a raw log dump.

What it does

Given the failure logs (plus optional commit and repo_state) and a min_confidence threshold, it classifies the failure as one of:

| verdict | meaning | consequence emitted | | --- | --- | --- | | real-break | a code defect visible in the logs | routing decision -> issue-to-pr | | dep | a dependency / resolution failure | routing decision -> issue-to-pr | | infra | runner / network / disk / daemon failure | read-only operator page note | | flake | known-flaky / intermittent, dominant signal | read-only rerun verdict |

The output is a typed runx.ci.triage.v1 packet:

{
  "schema": "runx.ci.triage.v1",
  "classification": { "verdict": "real-break", "confidence": 0.9, "evidence_refs": ["log:L2", "log:L3"] },
  "routing_decision": { "recommended_lane": "issue-to-pr", "rationale": "..." },
  "handoff": { "seam": "dispatch-by-naming", "downstream": ["issue-intake", "issue-to-pr", "pr-review-note"] }
}

Boundaries (why an operator can trust it)

• Read-only. It opens no tracking item, reruns no CI, and pages no operator. It produces a classification and a routing decision; a separate issue-intake / issue-to-pr / pr-review-note run is the governed commencement gate.
• Evidence-bound. Every verdict cites the exact log lines (log:L<n>) it is based on. It refuses to assert a root cause that is not visible in the supplied logs.
• It refuses when it should. Empty logs, truncated logs, no decisive signal, or a computed confidence below min_confidence all block the run for a human lane (needs_agent) and emit no routing.

Inputs

• logs (required): the raw build/test output of the failed CI run.
• commit (optional): commit SHA under test, recorded for provenance.
• repo_state (optional): short hint such as clean, dirty, unknown.
• min_confidence (optional, default 0.6): refuse below this confidence.

Output

A single runx.ci.triage.v1 packet on stdout with classification and exactly one of routing_decision, rerun_verdict, or page_note.

Harness

Two locked cases (runx harness .):

• real_break_clear_logs -> sealed, verdict: real-break, lane issue-to-pr.
• ambiguous_truncated_logs -> refused (needs_agent), no routing.