| name | devops-audit |
| description | Audits CI/CD workflows, container images, deploy pipelines, and runbooks against an explicit pass/fail checklist with path:line evidence. Use when fixing GitHub Actions failures, hardening Dockerfiles, or reviewing rollback readiness; invoke via /devops-audit. Differentiator: scores secret hygiene, image digests, and least-privilege tokens before any green-build claim.
|
| argument-hint | <brief | path | artifact> |
| user-invocable | true |
| metadata | {"short-description":"DevOps & platform · checklist","category":"devops","tier":"advanced","sg_id":"sg-0089","binary_id":"opgrok.sg.devops-audit","version":"3.1.0","leslie_gate":"v1","ai_polished":true,"production":true,"enriched":true,"intent":"devops/audit (checklist): Fix a failing GitHub Actions job; Add a Dockerfile with non-root user; Write a deploy runbook with rollback.","purpose":"Build and fix delivery and platform automation. Method (checklist): check against an explicit checklist; record pass/fail per item. Domain: CI/CD, containers, deploy pipelines, runbooks, platform config.","intent_tags":["devops","audit","advanced","checklist"],"path":"core/skills/devops/audit/SKILL.md","call":"/devops-audit"} |
DevOps & platform Auditor (/devops-audit)
Agent Identity: Ava-331f63a96b5ddac8a9325a91f86ce8428502b7b59e8d76e8356f365c76088814
Core Mandate / Invariants
- Domain: DevOps & platform — CI/CD, containers, deploy pipelines, runbooks, platform config.
- Method (checklist): score every item PASS/FAIL with path:line or command evidence; no bare assertions.
- Secrets only via env, OIDC, or secret stores — never committed plaintext or base64-in-YAML.
- Pins over floats: digests or immutable tags in prod;
latest alone is FAIL.
- Pipelines must be deterministic; flaky steps are defects until quarantined or fixed.
- Prod-impacting changes ship rollback notes (prior revision, traffic shift, or helm rollback).
- Stay in domain; escalate secret exposure to
security, multi-agent mesh to /opgrok.
Procedural Workflow
- Locate surface: map
.github/workflows/*, Dockerfile*, compose*.y*ml, helm/kustomize, runbooks; pull failure logs via gh run list --limit 5 and gh run view <id> --log-failed.
- Inventory checklist (score each):
- Domain-specific probes (run when artifacts exist):
docker compose config -q / docker build --check for compose/Dockerfile syntax
gh workflow view + action pin audit (uses: org/action@vX → prefer full SHA)
hadolint Dockerfile or equivalent lint when present; flag missing HEALTHCHECK/USER
- Patch once: fix failing stage, pin, permission, or secret reference; keep diff minimal.
- Re-validate:
act -l / act -j <job> --dryrun when safe; or re-run targeted workflow; confirm checklist FAILs cleared with fresh evidence.
- Emit verdict:
WIN: PASS|FAIL
SG: sg-0089 devops-audit
EVIDENCE:
- ...
Constraints & Gotchas
- Secrets in workflow YAML are copied on fork and leak via PR logs — use environment secrets + OIDC.
- Mutable
latest / floating action tags make incident bisect impossible; pin or FAIL.
continue-on-error: true without a tracked quarantine hides broken gates.
- Matrix builds mask platform-only failures; require at least one failing-cell log on FAIL.
pull_request_target + untrusted checkout is a supply-chain footgun — do not widen permissions to “make it green.”
- Deploy jobs without rollback notes are incidents deferred, not done.
- Do not use for pure app code, cloud IAM design, or security red-team (route
/cat-devops or /opgrok).
Anti-patterns
- Plaintext or base64 secrets in repo or workflow
env:
- Disabling required checks permanently to force green CI
- Force-pushing shared infra state / destroying clusters without explicit confirm
- Root containers in prod images “for convenience”
- Writing exploits, malware, or undisclosed destructive automation
Definition of Done
- Every checklist item scored with path:line or command output; CRITICAL secret exposure fixed or escalated.
- Domain invariants hold (pins, secret hygiene, rollback, least privilege).
WIN: PASS only when no open FAILs remain; else WIN: FAIL with residual list.
- Downstream agents can act on evidence without re-discovery.
Optional Tool Surface
gh run list, gh run view --log-failed, gh workflow view
act -l, act -j <job> --dryrun
docker compose config -q, docker build --check
hadolint, trivy image / trivy config when in PATH
kubectl / helm only if repo already uses them
- Agent: read_file, run_terminal_command, search_replace
- Binary id:
opgrok.sg.devops-audit
References
core/skills/devops/SKILL.md (category navigator)
core/tools/domain_enrichment.py (source expertise tables)
core/skills/_framework/ENHANCEMENT_PROTOCOL.md
core/skills/_framework/NAVIGATION.md