skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
onsager-ai/dev-skills63 installs

ci-triage

Investigate failed GitHub Actions runs and PR checks, classify evidence and reproduce using repository gates. Use for failed checks, red main, red PRs or uncertain CI failures.

How do I install this agent skill?

npx skills add https://github.com/onsager-ai/dev-skills --skill ci-triage
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill provides a structured process for CI triage and appears safe for use. It includes security measures like log redaction and requires execution in isolated environments. The identified risks are inherent to the skill's primary function of investigating repository failures.

  • Socketpass

    No alerts

  • Snykwarn

    Risk: MEDIUM · 1 issue

What does this agent skill do?

CI triage

Scope

This skill owns evidence-based classification. The repo contract/overlay owns reproduction commands, expected failures, issue conventions and warning severity. Triage does not authorize code publication, reverts, issue/comments or mentions. Load pr-lifecycle when repair changes an authorized PR.

Prerequisites

Identify the run, workflow, exact tested SHA and event. For PRs distinguish the head from the tested merge preview. Use harness-operations when retrieving checks/logs requires a capability mapping. Authentication and log availability must be observed; do not assume all connectors expose the same fields or limitations.

If expected PR runs are missing or a "no checks reported" alert arrives, follow PR CI trigger recovery before failure classification. Passing manual dispatches do not establish PR checks.

Procedure

  1. Read run/job status, first failed step and bounded logs through the available GitHub tools. Redact credentials and personal data. If logs are unavailable, retain the status/configuration evidence and report the access limit.
  2. Compare the tested diff with the previous comparable run, including workflow, inputs, dependencies and environment. An earlier green commit alone does not establish a flake or prove causation. Identify a suspect change with evidence; do not blame an author from the merge-message shape alone.
  3. Reproduce in an authorized isolated checkout at the exact tested ref or merge result using its own gate commands and fixtures. Do not switch a parked primary checkout to the suspect branch. Browser failures use the repo's browser workflow; another product's routes or fixtures are not interchangeable.
  4. Classify with the following taxonomy. Missing logs, credentials, tooling or a reproducible environment limit confidence; report them rather than guessing.
  5. Follow the repo's issue convention only when authorized. If it uses a rolling main-red issue, search for an existing open tracker before creating one. Append concise evidence instead of opening duplicates. Do not notify people merely because a template contains a mention. Close a tracker only with authority and a cited green run for the affected workflow.
  6. For a repair, hand off to the repo workflow and pre-push/PR lifecycle. For a retry, retain the failed run and compare the same tested inputs. A green retry alone does not justify weakening a required assertion.

Classification

BucketEvidenceNext action
regressionDeterministic failure tied to the tested change or its merged interactionReproduce and repair the demonstrated defect
flakeSame comparable inputs alternate outcomes, with evidence of a nondeterministic causeRecord/retry and address the cause
infraRunner, service, network or dependency provisioning failed independently of product assertionsRepair/retry the environment
needs-humanEvidence or authority is insufficient for a reliable diagnosisGather available evidence; request any pending human decision through the structured question capability in harness-operations

If a human must choose the next action or supply missing context, use the structured question capability in harness-operations, then wait before dependent work. A plain-text needs-human report alone does not request that decision when the tool is available. If unavailable, state the limitation and use the established handoff channel. Missing infrastructure alone does not require a human choice.

Completion

Report workflow/run URL, tested SHA, first failing step, bucket, confidence, reproduction result, bounded evidence and next action. A blocked reproduction is not a successful local check. Keep check status tied to the exact SHA.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/onsager-ai/dev-skills/ci-triage">View ci-triage on skillZs</a>