ci-cd-integration

name	ci-cd-integration
description	Design CI/CD pipelines that run test suites. Covers GitHub Actions and GitLab CI templates, parallelism and sharding, artifact management, flaky-test quarantine, test-result publishing, coverage quality gates, OIDC keyless deploy, and copy-paste workflows for Playwright, Jest, and multi-stage pipelines. Use when: "CI/CD," "GitHub Actions," "pipeline," "test in CI," "GitLab CI," "continuous integration," "test automation pipeline," "shard tests in CI." Not for: per-test flaky healing at runtime — use test-reliability; go/no-go release decisions and smoke-test checklists — use release-readiness; test-result dashboards and trend reporting — use qa-metrics. Related: playwright-automation, qa-metrics, test-reliability, coverage-analysis, release-readiness.
license	MIT
metadata	{"author":"kindlmann","version":"2.0","category":"infrastructure"}

A 20-minute serial suite on every push destroys developer velocity; a green pipeline that retries flaky tests three times hides the race condition until it ships. This skill produces CI/CD pipelines that run the right tests at the right trigger, shard them across runners, store traces and reports as evidence, quarantine flaky tests instead of masking them, and gate merges on real coverage numbers. Use this skill when the question is about running tests in a pipeline, not writing them.

Discovery Questions

Check .agents/qa-project-context.md first — if it exists, use it and skip anything already answered there (especially team_maturity and existing CI conventions). Then:

1. Which CI platform? GitHub Actions, GitLab CI, CircleCI, Jenkins? This skill ships templates for GitHub Actions and GitLab CI.
2. What test types need to run? Unit, integration, E2E, visual, performance? Each has different resource and timing needs.
3. What is the current CI duration? Over 10 minutes means parallelism and sharding are mandatory, not optional.
4. How many developers push per day? High-frequency teams need aggressive concurrency cancellation and caching.
5. What triggers should run which tests? Not every push needs a full E2E suite — map triggers to suites before writing YAML.

Calibrate to team maturity

Set team_maturity in .agents/qa-project-context.md; pick the matching pipeline shape:

• startup — one job: lint + unit + one E2E smoke on PR. Fast feedback over completeness.
• growing — separate jobs for unit, integration, E2E. Parallelization, artifact uploads, result publishing, flaky quarantine.
• established — full matrix: sharded E2E, multi-environment promotion gates, perf and security scans, deploy-gated checks, SLA-backed pipelines.

---

Core Principles

1. Fast feedback: right tests at the right time. Unit tests on every push (under 2 min). E2E on PRs (under 10 min). Full suite on merge and nightly. The trigger-to-suite map below is the contract.
2. Parallel first: shard tests across workers. A 20-minute serial suite becomes 5 minutes across 4 shards. Always worth the runner cost.
3. Artifacts are evidence. Every run stores traces, screenshots, coverage, and HTML reports. Without artifacts, a CI failure is an undebuggable "reproduce locally" cycle.
4. Flaky tests need quarantine, not retries. Retrying hides the problem — the test passes on retry, the report is green, the race condition persists. Move flaky tests to a non-blocking job, track them, fix the root cause.
5. Quality gates get stricter toward production. Define what must pass at each stage; PR gate is fast and cheap, deploy gate is comprehensive.
6. Read thresholds from config, not from bash. Let the test runner enforce coverage via its own coverageThreshold/thresholds and exit non-zero. Scraping percentages out of stdout with regex is fragile across runner versions.

---

Pipeline Architecture

Push to branch:   lint+types (30s) → unit (1-2m)
PR opened:        + integration (2-3m) → E2E sharded (5-8m) → merge report
Merge to main:    full E2E ∥ visual ∥ perf budget → deploy (OIDC)
Nightly (cron):   full suite + npm audit + axe a11y + flaky quarantine

What runs when

Trigger	Tests	Max duration
Push to branch	lint, type-check, unit	2 min
PR opened/updated	+ integration, E2E smoke	10 min
Merge to main	+ full E2E, visual, perf budget	15 min
Nightly schedule	full suite, security, a11y, flaky quarantine	30 min
Release tag	full suite, smoke against staging	20 min

---

GitHub Actions

For complete copy-paste workflow files (unit, sharded Playwright E2E, full pipeline, nightly, PR gate), see references/github-actions-templates.md.

Action versions (June 2026)

Pin to the current major and let Dependabot bump them. The actions/* family runs on the Node 24 runner; Node 20 is deprecated on GH-hosted runners.

Action	Current major	Notes
actions/checkout	@v6
actions/setup-node	@v6	v5+ auto-caches only when packageManager is set; use cache: npm to be explicit
actions/cache	@v5	new cache service v2 backend
actions/upload-artifact	@v7	v7 can upload unzipped (archive: false)
actions/download-artifact	@v7	pair with upload-artifact major
dorny/test-reporter	@v3	v3 requires Node 24 runner; reporter keys unchanged
dorny/paths-filter	@v3
marocchino/sticky-pull-request-comment	@v3
slackapi/slack-github-action	@v2	floating major; see notification note before adopting v3

For supply-chain-sensitive pipelines, pin third-party actions (dorny, marocchino, slackapi, knapsack) to a full-length commit SHA with a version comment, and let Dependabot update the SHA: uses: dorny/test-reporter@<40-char-sha> # v3.0.0. First-party actions/* are lower risk; tags are acceptable there.

Key concepts

Concurrency groups cancel wasted runs when a branch gets multiple pushes:

concurrency:
  group: tests-${{ github.ref }}
  cancel-in-progress: true

Matrix sharding across runners:

strategy:
  fail-fast: false
  matrix:
    shard: [1, 2, 3, 4]
steps:
  - run: npx playwright test --shard=${{ matrix.shard }}/4

Caching browsers so they aren't re-downloaded every run:

- uses: actions/setup-node@v6
  with: { node-version: 22, cache: npm }

- name: Cache Playwright browsers
  id: playwright-cache
  uses: actions/cache@v5
  with:
    path: ~/.cache/ms-playwright
    key: playwright-${{ runner.os }}-${{ hashFiles('package-lock.json') }}

- name: Install Playwright browsers
  if: steps.playwright-cache.outputs.cache-hit != 'true'
  run: npx playwright install --with-deps chromium

Artifacts for reports and traces, and merging sharded reports into one HTML report — see references/github-actions-templates.md (E2E workflow). The merge job uses actions/download-artifact@v7 with pattern: test-results-* then npx playwright merge-reports --reporter=html.

Smarter sharding at scale

Past 10–15 shards, naïve hash-based splitting wastes runner time on uneven shards. Use a timing-aware balancer:

• knapsack-pro — timing-data based, supports Playwright/Jest/Cypress/RSpec; distributes by historical duration.
• CloudBees Smart Tests (formerly Launchable) — ML prioritization + Test Impact Analysis; runs only the tests likely to fail for the diff.
• Datadog Test Optimization — TIA + flake management; shard-balancing by historical time.
• Trunk Flaky Tests — flake-aware quarantine + retry budgeting.

Before reaching for a paid balancer: Playwright's --shard already distributes by file and balances on duration from prior runs. To inspect or feed custom timing data, dump it yourself — npx playwright test --reporter=json | jq '[.suites[].specs[] | {file: .file, duration: .tests[].results[].duration}]'. For Jest, jest-slow-test-reporter surfaces the slowest specs so you can split or fix them.

For self-hosted runners on Kubernetes, use Actions Runner Controller (arc-runner-set / gha-runner-scale-set) — Helm-installed, auto-scales runner pods per workflow. Replaces the deprecated runner-deployment CRD.

Required status checks

Protect main in Settings → Branches → Branch protection rules: enable "Require status checks to pass before merging," add lint, unit-tests, and e2e (all shards) as required checks, and enable "Require branches to be up to date."

---

GitLab CI

For the full pipeline, see references/gitlab-ci-template.md. Key points:

• Stages [validate, test, e2e, deploy]; node:22-alpine for lint/unit, mcr.microsoft.com/playwright:v1.60.0-noble for E2E (keep this pinned to your installed @playwright/test minor).
• Parallel sharding: parallel: 4 exposes CI_NODE_INDEX/CI_NODE_TOTAL; run npx playwright test --shard=$CI_NODE_INDEX/$CI_NODE_TOTAL.
• Coverage: emit a cobertura coverage_report artifact and a junit report; GitLab reads the percentage and test results from those. The legacy coverage: stdout regex is a fragile fallback across Jest versions — prefer the cobertura report.

---

Advanced Patterns

Test result publishing to PR comments

- name: Publish test results
  uses: dorny/test-reporter@v3
  if: ${{ !cancelled() }}
  with:
    name: Test Results
    path: test-results/junit.xml
    reporter: jest-junit  # use java-junit for a Playwright JUnit report

For the sticky coverage PR comment (marocchino/sticky-pull-request-comment@v3), see references/github-actions-templates.md (PR Quality Gate).

Conditional test execution

Only test what changed. Use dorny/paths-filter@v3 to set outputs, then gate steps on them — see references/github-actions-templates.md (Conditional execution).

Flaky test quarantine

Separate flaky tests into a non-blocking job so they run in CI but don't block merges:

e2e-stable:        # required for merge
  steps:
    - run: npx playwright test --grep-invert @flaky

e2e-quarantine:    # non-blocking
  continue-on-error: true
  steps:
    - run: npx playwright test --grep @flaky
    - if: failure()
      run: echo "::warning::Quarantined tests failed. Review and fix or remove."

Tag flaky tests at the source so the grep splits them:

test('sometimes fails due to race condition @flaky', async ({ page }) => {
  // runs in CI but doesn't block merges
});

If a quarantined test passes 10 consecutive runs, remove the @flaky tag. For runtime self-healing of a single flaky test (selector recovery, auto-retry policy), use test-reliability.

Cache strategies

Layer	Path	Cache key
Node modules	(handled by setup-node cache: npm)	automatic
Playwright browsers	~/.cache/ms-playwright	pw-{os}-{hash(package-lock.json)}
Build cache (Next.js)	.next/cache	nextjs-{os}-{hash(lockfile)}-{hash(src)}
Test fixtures	e2e/fixtures/.cache	test-data-{hash(seed.sql)}

Use actions/cache@v5 for layers 2–4; add restore-keys on build caches for partial matches.

OIDC keyless deploy

Don't store a long-lived DEPLOY_TOKEN. Use GitHub Actions OIDC to assume a cloud role for short-lived credentials — nothing static to leak or rotate:

deploy:
  permissions:
    id-token: write   # request the OIDC JWT
    contents: read
  steps:
    - uses: aws-actions/configure-aws-credentials@v6
      with:
        role-to-assume: arn:aws:iam::123456789012:role/gha-deploy
        aws-region: eu-central-1
    - run: ./deploy.sh production   # uses short-lived STS creds, no static secret

The IAM role's trust policy pins the sub claim to your repo and branch. GCP (google-github-actions/auth) and Azure (azure/login) have equivalent OIDC flows.

Slack/Teams notification on failure

Use slackapi/slack-github-action@v2 with webhook-type: incoming-webhook, gated on if: failure() && github.ref == 'refs/heads/main' so only main-branch failures notify. Before moving to @v3, note v3 changed payload handling for workflow-trigger webhooks (no longer flattened/stringified) — verify your payload against the v3 docs first. Full example in references/github-actions-templates.md (Nightly Full Suite).

---

Quality Gates

Gate	When	Required checks	Blocking?
PR Gate	PR opened/updated	lint, type-check, unit, coverage threshold	Yes
Merge Gate	Before merge to main	+ E2E smoke suite	Yes
Deploy Gate	Before production deploy	+ full E2E, visual, perf budget	Yes
Nightly Gate	Scheduled 2am daily	full suite, npm audit, axe a11y	Alert only

PR Gate (under 3 minutes)

Enforce the coverage floor in the test runner's config, not in bash. In jest.config.js (or vitest.config.ts coverage.thresholds):

coverageThreshold: { global: { lines: 80, statements: 80, branches: 70 } }

Then jest --coverage exits non-zero when coverage drops, so the job fails with no extra script. If you must read the number in CI (e.g. to print it), have Jest emit json-summary and read the file — there is no coverage-summary CLI:

- run: npm test -- --ci --coverage   # exits 1 if below coverageThreshold
- name: Print coverage (optional)
  run: |
    PCT=$(jq '.total.lines.pct' coverage/coverage-summary.json)
    echo "Line coverage: ${PCT}%"

(json-summary reporter writes coverage/coverage-summary.json. For nyc/c8 projects, nyc report --reporter=text-summary. The standalone istanbul CLI is deprecated — don't use istanbul report.)

Merge Gate (under 10 minutes)

PR Gate + E2E smoke. Configure as required status checks in branch protection.

Deploy Gate (under 15 minutes)

Needs [unit-tests, e2e-tests, visual-tests], then a perf budget check (npx lhci autorun / lhci assert --config=lighthouserc.json) before the OIDC deploy step above.

Nightly Gate (up to 30 minutes)

Full E2E across all browsers, security scan, a11y audit, flaky quarantine. Wire the security and a11y steps as real jobs, not just prose:

- run: npm audit --audit-level=high   # fails on high/critical advisories
- run: npx playwright test --grep @a11y   # specs that call @axe-core/playwright

Where the @a11y-tagged specs use @axe-core/playwright:

import AxeBuilder from '@axe-core/playwright';
test('home page has no a11y violations @a11y', async ({ page }) => {
  await page.goto('/');
  const results = await new AxeBuilder({ page }).analyze();
  expect(results.violations).toEqual([]);
});

Results go to Slack, not as blocking checks.

---

Anti-Patterns

1. Running all tests on every commit

A 20-minute full suite on every push destroys velocity. Use the trigger-to-suite map: fast tests on push, comprehensive on PR and merge.

2. No artifact storage

Without traces, screenshots, and logs, every CI failure becomes a "reproduce locally" cycle that wastes hours. Upload artifacts on if: ${{ !cancelled() }}.

3. Retrying flaky tests without tracking them

retries: 3 hides flakiness — the report is green but the race condition persists. Quarantine, track, fix the root cause.

4. CI-only failures without local reproduction

If a test only fails in CI, document why (timezone, missing env var, screen resolution) and add a script that replicates CI locally with the same Playwright image you run in CI — don't pin a stale image. See references/github-actions-templates.md (Local repro).

5. Shared state between CI jobs

Jobs that read files from sibling jobs without artifacts or needs. Each job starts fresh; pass data via upload-artifact/download-artifact.

6. No concurrency controls

Multiple runs for the same branch waste runners. Always use a concurrency group with cancel-in-progress: true.

7. Hardcoded secrets in workflow files

Never put tokens, passwords, or keys in YAML. Use repo secrets (${{ secrets.X }}) or GitLab CI/CD variables — and prefer OIDC keyless auth over any long-lived deploy token.

8. Ignoring job timeouts

A stuck test can hold a runner for hours. Set timeout-minutes on every job and actionTimeout/navigationTimeout in the Playwright config.

---

Verification

Prove the pipeline before relying on it. Smallest check first:

1. Lint the workflow syntax — actionlint .github/workflows/*.yml catches expression, needs, and shell-quoting errors before they fail at runtime. Add a yamllint .github/workflows/ pass for indentation. Run actionlint as a job too.
2. Dry-run a job locally — act -j unit-tests runs the job in a container so you can iterate without pushing.
3. Confirm required checks appear — push to a throwaway branch, open a draft PR, and verify the expected check runs (lint, unit-tests, e2e) show up and that the coverage gate fails when you drop coverage below the threshold.
4. Verify artifacts — download the run's artifacts from the Actions UI (or gh run download <id>) and confirm playwright-report/ and traces are present.

---

Done When

• A trigger map exists: push runs lint+unit only; the PR workflow gates E2E behind if: github.event_name == 'pull_request' (verify the YAML, not "integration runs somewhere").
• actionlint .github/workflows/*.yml exits 0.
• A non-blocking quarantine job runs --grep @flaky with continue-on-error: true; the stable job runs --grep-invert @flaky and is in the required checks list.
• Test artifacts (reports, screenshots, traces) upload on if: ${{ !cancelled() }} with an explicit retention-days/expire_in.
• Concurrency groups with cancel-in-progress: true are set on the PR/test workflows.
• Coverage is enforced by the runner's coverageThreshold/thresholds (job exits non-zero below the floor) — no coverage-summary CLI scrape.
• Branch protection lists lint, unit-tests, and e2e as required status checks.
• No long-lived deploy token in YAML — deploy uses OIDC (id-token: write + cloud role) or, at minimum, a secrets-store reference.

---

Related Skills

• playwright-automation — writing the E2E tests, Page Object Model, and the playwright.config.ts whose sharding/timeouts this pipeline drives.
• test-reliability — runtime self-healing of one flaky test (selector recovery, retry policy); go there to fix a flaky test, come here to quarantine it in CI.
• qa-metrics — turning the JUnit/coverage artifacts this pipeline produces into dashboards and flakiness trends.
• release-readiness — the human go/no-go decision and release checklist that consumes these gate results; this skill builds the gates, that one decides on them.
• coverage-analysis — finding the coverage gaps and setting the threshold this skill's PR gate enforces.

Reference Files (in references/)

• github-actions-templates.md — copy-paste unit, sharded Playwright E2E (+ report merge), full pipeline, nightly (Slack + audit + axe), and PR quality-gate workflows, plus conditional execution and local-repro snippets.
• gitlab-ci-template.md — full .gitlab-ci.yml with parallel sharding, cobertura coverage, and JUnit MR reporting.