write-pipeline

Write a CI/CD pipeline for $ARGUMENTS.

Process (sequential — do not skip steps)

Step 1: Reconnaissance

Before writing any pipeline configuration:

1. Identify the stack — language, framework, package manager, test runner, linter
2. Check existing pipelines — look for .github/workflows/, .gitlab-ci.yml, Jenkinsfile, azure-pipelines.yml
3. Check existing scripts — package.json scripts, Makefile, Taskfile, scripts/ directory
4. Identify deployment targets — cloud provider, container registry, CDN, serverless
5. Check for monorepo structure — multiple projects/services in one repo?

Step 2: Pipeline Architecture

Every pipeline follows this ordering principle: fail fast — cheapest checks first.

Lint/Format → Build → Unit Tests → Integration Tests → Security Scan → Deploy

If any stage fails, subsequent stages do not run. Total pipeline time budget: under 10 minutes for the fast path (lint + build + unit tests).

Step 3: Stage Design

Stage 1: Lint and Format Check (fastest — run first)

# Purpose: catch style and type errors in <30 seconds
- name: Lint
  run: |
    npm run lint
    npm run typecheck
    npm run format:check  # --check flag, never auto-fix in CI

Rules:

• Format check, not format fix — CI detects, developers fix locally
• Type checking is a lint stage, not a build stage
• If lint takes >60 seconds, something is wrong

Stage 2: Build

# Purpose: compile/bundle and verify the artifact is producible
- name: Build
  run: npm run build

Rules:

• Build output is cached or uploaded as an artifact for later stages
• Build once, deploy the same artifact everywhere (dev, staging, prod)
• Never rebuild for each environment — use environment variables at runtime

Stage 3: Unit Tests (fast feedback)

- name: Unit Tests
  run: CI=true npm test -- --coverage

Rules:

• Run in run mode, never watch mode (CI=true or explicit --run flag)
• Coverage report uploaded as artifact
• Fail if coverage drops below threshold (configure per-project)
• Timeout: 5 minutes max. If unit tests take longer, they aren't unit tests

Stage 4: Integration Tests

- name: Integration Tests
  run: CI=true npm run test:integration
  services:
    postgres:
      image: postgres:16-alpine
      env:
        POSTGRES_PASSWORD: test

Rules:

• Run after unit tests pass (don't waste time on integration if units fail)
• Use service containers for databases, message brokers, etc.
• Timeout: 10 minutes max
• Separate from unit tests — different command, different config

Stage 5: Security Scan

- name: Security Scan
  run: |
    npm audit --audit-level=high
    # or: trivy fs . --severity HIGH,CRITICAL

Rules:

• Fail on HIGH and CRITICAL vulnerabilities
• Allow suppression file for acknowledged false positives (with expiry dates)
• Run SAST (static analysis) in addition to dependency audit
• Container image scanning if building Docker images

Stage 6: Deploy

- name: Deploy
  if: github.ref == 'refs/heads/main' && success()
  run: ./scripts/deploy.sh

Rules:

• Deploy only on main branch, only after ALL checks pass
• Use environment protection rules (manual approval for production)
• Deploy the artifact built in Stage 2 — never rebuild
• Include a smoke test after deployment
• Rollback plan documented or automated

Caching Strategy

Cache aggressively to reduce pipeline time:

# Node.js
- uses: actions/cache@v4
  with:
    path: node_modules
    key: node-${{ hashFiles('package-lock.json') }}

# .NET
- uses: actions/cache@v4
  with:
    path: ~/.nuget/packages
    key: nuget-${{ hashFiles('**/*.csproj') }}

# Python
- uses: actions/cache@v4
  with:
    path: ~/.cache/pip
    key: pip-${{ hashFiles('requirements*.txt') }}

# Docker layers
- uses: docker/build-push-action@v5
  with:
    cache-from: type=gha
    cache-to: type=gha,mode=max

Rules:

• Cache key MUST include a hash of the lockfile/dependency manifest
• Restore keys for partial cache hits (same prefix, different hash)
• Cache the dependency directory, not the build output (build output changes more frequently)
• Monitor cache hit rates — below 80% means the key strategy is wrong

Matrix Strategy

Use matrix builds for multi-version or multi-project testing:

# Multi-version testing
strategy:
  matrix:
    node-version: [20, 22]
  fail-fast: true  # Stop all jobs if one fails

# Monorepo auto-discovery
strategy:
  matrix:
    project: ${{ fromJson(needs.detect-changes.outputs.projects) }}

Rules:

• fail-fast: true — no point running other versions if one fails
• For monorepos, detect changed projects first and only build/test affected ones
• Matrix dimension should be small (2-3 values). Combinatorial explosion wastes CI minutes

Monorepo CI (if applicable)

For monorepo projects:

1. Change detection — determine which projects changed using path filters or git diff
2. Selective execution — only build/test projects that changed or depend on changed code
3. Dependency graph — if project A depends on project B and B changed, test A too
4. Shared workflows — extract common steps into reusable workflows/templates

# GitHub Actions path filter
on:
  push:
    paths:
      - 'services/api/**'
      - 'packages/shared/**'  # shared dependency

Secrets Management

• Never hardcode secrets — use repository/environment secrets
• Minimal scope — each secret available only to the jobs that need it
• Rotation — document which secrets exist and when they were last rotated
• Masking — pipeline platform should auto-mask secrets in logs. Verify this
• Environment-specific — production secrets only available in production deployment jobs
• No secrets in build args — use multi-stage builds or runtime injection

Version Pinning

# Pin action versions to full SHA (not tags)
- uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11  # v4.1.1

# Pin tool versions
- uses: actions/setup-node@v4
  with:
    node-version-file: '.nvmrc'  # or package.json engines

Rules:

• Pin GitHub Actions to commit SHA, not tag (tags can be moved)
• Pin language/tool versions to the project's version file (.nvmrc, global.json, .python-version)
• Pin Docker base images to digest, not just tag
• Dependabot or Renovate for automated version updates

Anti-Patterns (NEVER do these)

• Watch mode in CI — tests never exit. Always use CI=true or --run
• Rebuilding for each environment — build once, configure at runtime
• No caching — every run downloads the internet. Cache dependencies
• Deploy from feature branches — only main/release branches deploy
• Secrets in pipeline files — use secret management, never inline
• No timeout — set job-level and step-level timeouts. Runaway processes cost money
• Success-only notifications — alert on failure, not on success. Nobody reads 200 green emails
• Manual artifact passing — use pipeline artifacts or cache, not manual copy between stages

Output

Pipeline design affects all four DORA metrics: deployment frequency (how often the pipeline runs), lead time for changes (pipeline duration), change failure rate (test/gate effectiveness), and time to restore service (rollback speed).

Deliver:

1. Complete pipeline configuration file (.github/workflows/*.yml or equivalent)
2. Any supporting scripts referenced by the pipeline
3. .dockerignore or equivalent if building containers
4. Brief explanation of stage ordering and caching decisions

Related Skills

• /devops:write-dockerfile — pipelines that build containers need a Dockerfile. Ensure the pipeline's build stage matches the Dockerfile's target.