infra-ci-cd-github-actions（infra CI CD GitHub actions 开发）

CI/CD Pipelines

Quick Guide: GitHub Actions for CI/CD. Affected detection for monorepo optimization (e.g., Turborepo --affected or --filter=...[origin/main]). Dependency and build output caching for fast CI. Quality gates: lint + type-check + test + build + coverage as required status checks. Multi-environment deployments with build promotion. OIDC authentication for cloud providers. Pin all action and runtime versions.

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST use affected/changed-package detection for PR builds - NEVER run full test suite on PRs)

(You MUST cache package manager dependencies and build outputs - CI without caching wastes 70% of runtime)

(You MUST pin action versions (actions/checkout@v6, oven-sh/setup-bun@v2, actions/cache@v5) - NEVER use @main or unversioned)

(You MUST implement quality gates (lint + type-check + test + build) as required status checks - block merge on failures)

(You MUST use OIDC for cloud provider auth where supported - NEVER use static long-lived credentials)

</critical_requirements>

Detailed Resources:

• examples/core.md - Pipeline config, jobs, caching, reusable workflows, composite actions, matrix builds
• examples/testing.md - Affected detection, quality gates
• examples/caching.md - Remote caching, Turborepo
• examples/security.md - OIDC auth, secrets rotation, artifact attestations
• examples/deployment.md - Multi-env, rollback
• examples/monitoring.md - CI metrics, GitHub Insights
• reference.md - Decision frameworks, anti-patterns, constants reference

Auto-detection: GitHub Actions, CI/CD pipelines, .github/workflows, Turborepo affected detection, remote cache, deployment automation, quality gates, OIDC authentication, secret rotation, artifact attestations, SLSA provenance, reusable workflows, composite actions, matrix builds, workflow_call

When to use:

• Setting up GitHub Actions workflows for monorepos
• Implementing affected detection for faster PR builds
• Configuring remote cache for shared build artifacts
• Setting up quality gates and branch protection rules
• Implementing OIDC authentication for cloud deployments
• Adding artifact attestations for supply chain security

When NOT to use:

• Projects not using GitHub (use your CI provider's native docs)
• No automated testing or build step needed

Key patterns covered:

• Pipeline configuration with parallel jobs and dependency caching
• Affected detection (Turborepo --affected flag or --filter=...[origin/main])
• Quality gates (lint, type-check, test, build as parallel jobs with dependencies)
• OIDC authentication (no static credentials for cloud providers)
• Reusable workflows (workflow_call, up to 10 levels total)
• Composite actions (using: composite, shared setup logic)
• Matrix builds (include/exclude, fail-fast, dynamic matrices)
• Artifact attestations (SLSA v1.0 Build Level 2 provenance)
• Multi-environment deployment with build promotion

Philosophy

CI/CD pipelines automate testing, building, and deployment. In a monorepo, intelligent caching and affected detection are critical for maintaining fast CI as the codebase grows.

Core principles:

• Fast feedback: PR builds should complete in < 5 minutes via affected detection and caching
• Build once, promote everywhere: Single build artifact deployed through preview/staging/production
• No static credentials: OIDC for cloud providers, secrets managers for rotating credentials
• Quality gates block merge: Lint, type-check, test, and build must all pass before merge

Core Patterns

Pattern 1: Pipeline Configuration

Separate install, parallel quality checks, then build.

# Recommended workflow structure:
# ci.yml      - lint, test, type-check, build (PR + main)
# deploy.yml  - production deployment from main
# preview.yml - preview deployments for PRs

Key decisions:

• Pin runtime and action versions (never use latest)
• Separate install job with cached dependencies, then fan out to parallel lint/test/type-check
• Build only after all quality gates pass
• Use concurrency with cancel-in-progress: true to avoid wasting resources

See examples/core.md for complete workflow examples.

Pattern 2: Affected Detection

Only test and build changed packages in monorepos.

Turborepo example (two approaches, choose one):

# Modern: --affected flag (auto-detects CI environment)
turbo run test --affected

# Manual: --filter with git comparison
turbo run test --filter=...[origin/main]

Key principle: PRs use affected detection for fast feedback (< 5 min). Main branch runs full suite.

Gotcha: New packages have no git history and get skipped by affected detection. Always check for new package.json files and fall back to full suite.

See examples/testing.md for PR vs main branch workflow examples.

Pattern 3: Quality Gates

Automated checks that must pass before merge.

Quality gate order:

1. Linting (code style and static analysis)
2. Type checking (TypeScript errors)
3. Tests with coverage (functionality validation)
4. Build verification (production build succeeds)
5. Bundle size check (performance regression prevention)
6. Security audit (dependency vulnerabilities)

Configure as required status checks in branch protection. Use strict: true to require branches be up-to-date before merge.

See examples/testing.md for comprehensive quality gate workflow.

Pattern 4: OIDC Authentication

Eliminate static credentials for cloud deployments.

# Key requirement for OIDC:
permissions:
  id-token: write # Required for OIDC token generation
  contents: read

OIDC eliminates: manual key rotation, permanent security risk from leaked keys, and untraceable deployments. Temporary credentials auto-expire (typically 1 hour).

See examples/security.md for AWS OIDC and token-based authentication examples.

Pattern 5: Reusable Workflows vs Composite Actions

Centralize CI/CD logic across repositories.

See examples/core.md for implementation examples.

Performance Optimization

Goal: CI runtime < 5 minutes for PR builds

Parallelization techniques:

• Separate install job, parallel lint/test/type-check jobs (saves 40% time)
• Matrix builds for multiple OS/versions (only on main, not PRs)
• Split test suites (unit, integration, e2e as parallel jobs)
• Use concurrency with cancel-in-progress: true to cancel outdated runs

Monitoring targets:

• CI runtime: < 5 min (PR), < 10 min (main)
• Cache hit rate: > 80% (remote cache)
• Failure rate: < 5% (excluding flaky tests)
• Time to deploy: < 10 min (commit to production)

<red_flags>

RED FLAGS

High Priority:

• Running full test suite on every PR - Use affected detection or CI takes 10+ minutes
• No caching configured - Reinstalling dependencies every run wastes 2-3 minutes
• Using latest for runtime versions - Non-deterministic builds break reproducibility
• Static cloud credentials in secrets - Use OIDC authentication, never store long-lived access keys
• Committing secrets to repository - Use GitHub Secrets, never hardcode credentials in YAML
• No quality gates on main branch - Missing lint/test/type-check allows broken code to merge

Medium Priority:

• Sequential jobs instead of parallel - Lint/test/type-check should fan out after install
• No concurrency limits - Multiple CI runs on same PR waste resources
• Rebuilding for each environment - Build once, promote artifact through environments
• No monitoring of CI performance - Cannot identify bottlenecks without tracking duration and cache hit rate
• Magic numbers in workflows - Hardcoded timeouts and thresholds with no documentation of intent

Common Mistakes:

• Not using fetch-depth: 0 for affected detection (git diff fails without history)
• Using needs: [all, previous, jobs] on every job (creates sequential execution)
• Not handling new packages in affected detection (they get skipped)

Gotchas & Edge Cases:

• fetch-depth: 0 required for affected detection (shallow clone breaks git diff)
• New packages have no git history so affected detection skips them
• actions/cache default limit is 10GB per repo (configurable up to 10TB, additional storage billed at $0.07/GiB/month)
• OIDC requires id-token: write permission or token generation fails silently
• Environment secrets override repository secrets with the same name
• Artifact attestations require attestations: write AND id-token: write AND contents: read
• Reusable workflows support 10 levels total (caller + 9 nested, increased from 4) and 50 unique per run
• actions/create-release is deprecated - use softprops/action-gh-release@v2 instead
• workflow_dispatch now supports 25 inputs (increased from 10)

</red_flags>

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md

(You MUST use affected/changed-package detection for PR builds - NEVER run full test suite on PRs)

(You MUST cache package manager dependencies and build outputs - CI without caching wastes 70% of runtime)

(You MUST pin action versions (actions/checkout@v6, oven-sh/setup-bun@v2, actions/cache@v5) - NEVER use @main or unversioned)

(You MUST implement quality gates (lint + type-check + test + build) as required status checks - block merge on failures)

(You MUST use OIDC for cloud provider auth where supported - NEVER use static long-lived credentials)

Failure to follow these rules will result in slow CI (10+ min), security vulnerabilities (leaked credentials), and broken builds (missing quality gates).

</critical_reminders>