CI/CD Audit Skill

Degree of freedom: MIXED — Steps 1–3 judgment [HIGH freedom]; Step 0 gh inventory and Step 4 deletes [LOW freedom — run exactly].

Systematic audit of GitHub Actions workflows to cut the Actions bill (minutes + storage) and speed up CI without losing test coverage or deploy safety. Uses the gh CLI for live billing, run history, and storage data.

How to reason

1. Observe — quote the trigger, runs-on, concurrency, or gh minute/storage number
2. Interpret — billed twice, doomed minutes, or a 10× runner on every push?
3. Classify — double-bill / no-concurrency / expensive-runner / doomed-job / storage / correct
4. Severity — by $ × safety; never score a cut that drops coverage

Worked example

Observe: build-mobile.yml runs-on: macos-15 on push: to every
branch; no if: gate; private repo.
Interpret: every push bills ~10× Ubuntu for an iOS job nobody requested.
Classify: expensive-runner on ordinary push.
Severity: Very high (~10×).
Finding: build-mobile.yml | macos on push | dispatch/tag-only | ~10×

Step 0: Inventory the account and pipelines [LOW freedom — run exactly]

Measure before optimizing. Only private repos consume the paid minute allowance; public repos get free minutes — don't spend effort there.

# Which repos actually cost money (private, active, recent pushes)?
gh repo list <owner> --limit 200 --json name,visibility,isArchived,pushedAt \
  --jq 'sort_by(.pushedAt)|reverse|.[]|select(.visibility=="PRIVATE" and .isArchived==false)|"\(.name)\t\(.pushedAt[0:10])"'

# Per repo: active workflows, run volume (last 14d), and runner types
gh api "repos/<owner>/<repo>/actions/workflows" --jq '.workflows[]|select(.state=="active")|.name'
gh run list --repo <owner>/<repo> --created ">=YYYY-MM-DD" --limit 200 --json databaseId --jq 'length'

# Storage (the other half of the bill)
gh api "repos/<owner>/<repo>/actions/artifacts" --paginate --jq '[.artifacts[]|select(.expired==false)|.size_in_bytes]|add'
gh api "repos/<owner>/<repo>/actions/cache/usage" --jq '.active_caches_size_in_bytes'

Rank repos by runs × runner-multiplier. A macos- job counts ~10x a ubuntu job; -large/bigger runners cost more than standard.

Step 1: Anti-pattern scan (per workflow) [HIGH freedom]

For each .github/workflows/*.yml, check for the recurring cost drivers:

• Double-billing triggers — the same jobs run on both push: [main]

and pullrequest for main (every merge bills twice). PR gates verification; a separate deploy.yml handles main. If both are needed, guard per-job with if: github.eventname == '...'.

• Missing concurrency — no cancel-in-progress, so rapid pushes/PR

syncs stack full runs instead of superseding.

• Expensive runners on ordinary pushes — macos- / -large firing

on every push instead of workflow_dispatch or version tags only.

• No path filters — docs/config-only changes run the full heavy suite

(add paths: / paths-ignore:).

• Doomed jobs — a costly job (visual/e2e/Lighthouse) runs in parallel

with a cheap gate and keeps burning minutes after the gate fails (gate it with needs: [...]).

• Long/default artifact retention — upload-artifact without

retention-days keeps outputs 90 days; heavy reports uploaded on success (if: always()) instead of if: failure().

• Self-uploading SARIF pile-ups — e.g. gitleaks-action uploads a tiny

SARIF artifact every run for 90 days.

• Over-frequent crons — daily schedule: where weekly suffices.
• Redundant rebuilds — two jobs each run the same build instead of

building once and sharing via artifact (only when their build env is identical — see Safety).

Step 2: Cost levers (highest impact first) [HIGH freedom]

Step 3: Fix patterns [HIGH freedom]

# Concurrency — every workflow. CI/verification cancels superseded runs:
concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: true

# Deploy/release: never cancel a live deploy. For combined verify-on-PR +
# deploy-on-push workflows, cancel PR churn only:
concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

# Expensive runner: dispatch/tag-only, not every push
build-ios:
  runs-on: macos-15
  if: ${{ inputs.platform == 'ios' || startsWith(github.ref, 'refs/tags/') }}
  timeout-minutes: 60   # cap runaway 10x spend

# Lean artifacts
- uses: actions/upload-artifact@v4
  if: failure()
  with: { name: report, path: report/, retention-days: 3 }

# Stop the gitleaks SARIF pile-up
env: { GITLEAKS_ENABLE_UPLOAD_ARTIFACT: "false" }

Step 4: Storage cleanup (destructive — confirm first) [LOW freedom — confirm first]

Artifacts and caches are ephemeral CI outputs; deleting them is safe but irreversible. Confirm with the user, then:

# Delete non-expired artifacts in a repo
gh api "repos/<owner>/<repo>/actions/artifacts" --paginate \
  --jq '.artifacts[]|select(.expired==false)|.id' \
| xargs -I{} gh api -X DELETE "repos/<owner>/<repo>/actions/artifacts/{}"

# Delete caches
gh api "repos/<owner>/<repo>/actions/caches" --paginate --jq '.actions_caches[].id' \
| xargs -I{} gh api -X DELETE "repos/<owner>/<repo>/actions/caches/{}"

Safety rules (do NOT trade coverage or deploys for cost)

• Never delete, skip, or weaken a test to save minutes. Make advisory

scans (npm audit, CVE scans) continue-on-error or move them to a schedule — keep the check.

• Keep the deploy gate. If deploy.yml triggers on the CI workflow_run

for push, do not remove the push trigger from CI.

• Only share a build across jobs when their build env is identical. Reusing

an artifact built with different secrets/flags (prod env, QA build stamps) ships or tests the wrong bundle.

• Update branch protection when renaming/merging jobs. If a required status

check's job name changes, PATCH requiredstatuschecks.contexts in the same change or merges hang. Verify: gh api repos/<owner>/<repo>/branches/<main>/protection/requiredstatuschecks --jq '.contexts'

• Verify with live runs. After pushing, confirm no startup_failure

(YAML parses) and that intended jobs skip/run: gh run view <id> --json jobs --jq '.jobs[]|"\(.conclusion // .status) \(.name)"'

Account backstops (one-time, GitHub UI — cannot be set via API)

• Default artifact/log retention → Settings → Actions → General → drop from

90 days to ~14 (applies to all repos, including future ones).

• Spending budget → Settings → Billing → Budgets → set an Actions budget

with 75/90/100% alerts; keep "Stop usage" off so production deploys never hard-break.

• Rotate any secret leaked in a workflow/remote (the user must revoke;

you can only strip it from configs).

Self-critique before reporting [LOW freedom — do not skip]

1. Evidenced — workflow YAML or gh count, not "CI feels expensive"
2. Reproducible — same trigger/runs-on still present
3. Severity justified — ranked by $ × runner multiplier
4. Right owner — silent bypass / ratchet → audit-gate-logic; hosting bill → audit-infra-cost
5. No-false-safety — no test deleted or skipped to save minutes; storage delete confirmed

Output: CI/CD Cost Audit Report

## CI/CD Audit: [owner]

### Spend snapshot
- Actions billable: ~$X/mo (private repos exhaust the included minutes)
- Top spenders: [repo — ~$Y, driver], ...
- Storage: [X GB artifacts / Y GB caches]

### Findings (prioritized)
| # | Repo | Workflow | Anti-pattern | Fix | Impact |
|---|------|----------|--------------|-----|--------|
| 1 | repo | build-mobile.yml | macOS on every push | dispatch/tag-only | ~10x |

### Already healthy
- [repos/workflows already using concurrency, gated runners, path filters]

### Manual actions (user-only)
- [ ] Default artifact/log retention → 14 days
- [ ] Actions spending budget + alerts
- [ ] Rotate leaked secret(s), if any

### Expected outcome
~$A → ~$B/mo, no loss of test coverage or deploy safety.

Related

• audit-gate-logic — does the gate actually stop what it claims (bypass / ratchet)
• audit-infra-cost — hosting / DB / egress (not Actions minutes)
• test-visual-regression — screenshot CI artifact retention
• audit-security / deploy-verify / workflow-pr / create-hook