CI/CD Generator

Lifecycle skill that scaffolds a complete GitHub Actions pipeline for a Go, Rust, TypeScript, or Python project. The workflow encodes opinionated defaults documented in the project owner's notes — coverage gates, N+1 query detection, race condition property-based testing (including Python 3.14t free-threaded), memory leak detection, and load testing — plus standard security gates (SAST, SCA, container scan, SBOM, secret scan).

When to Use

The skill has two operating modes — generation (the default, walks Phase 0 → 6 below) and audit (read-only review of an existing pipeline, walks Phase A0 → A5 in references/AUDIT.md).

Generation mode

• A new repository (Go, Rust, TypeScript, or Python) has no .github/workflows/ directory and the user wants CI/CD wired up
• An existing repository has an ad-hoc workflow that the user wants replaced with a complete, opinionated baseline
• The user explicitly asks for "criar CI", "gerar pipeline", "scaffold workflow", or invokes /valarmindskills:ci-cd-generator
• A polyglot monorepo needs per-language pipelines (run the skill once per project root)

Audit mode

• The repository already has workflows authored before this skill existed and the user wants them reviewed
• A security incident exposed a CI weakness (compromised secret, malicious action) and the user wants the full surface checked
• Pre-release gate: confirm the workflows match the security level the team thinks they are running at
• The user explicitly asks for "auditar pipeline", "audit CI pipeline", "review existing workflow"

This skill is language-aware and lifecycle-driven: it detects → asks the minimum needed → applies heuristics → emits YAML or report → validates. For multi-stack API security review (design + active testing + Go, Next.js, and Python stack-specific lifecycles) run from a CI pipeline, complement with @code-security-review (Go branch — references/golang/; Next branch — references/nextjs/; Python branch — references/python/). For Next.js performance audits, see @code-review references/NEXTJS.md; for Python performance/style sweeps, see @code-review references/PYTHON.md.

Do not use when

• The user wants to fix or refactor a single line in an existing workflow — open the YAML directly or use @github-pr-review. For a full pipeline audit, switch to audit mode instead.
• The CI platform is GitLab CI, CircleCI, Buildkite, or Jenkins — this skill is GitHub Actions only
• The project language is not Go, Rust, TypeScript (Node.js / Bun / Deno), or Python (3.13+/3.14+, FastAPI / Django / Flask)
• The user wants to deploy infrastructure (Terraform, Pulumi, Kubernetes manifests) — pipeline ≠ infra

Prerequisites

Required access:

• Read access to the repository root (for detection)
• Write access to .github/ and .github/workflows/
• Permission to commit on a branch (skill never commits without explicit user approval)
• If branch protection is part of the request: admin role on the repository

Phase 0 — Project Detection

Detect language, package manager, and runtime before generating anything. Run the steps in order; stop at the first conclusive match per axis.

# Step 1 — language
test -f go.mod                                                  && echo "language: go"
test -f Cargo.toml                                              && echo "language: rust"
test -f package.json                                            && echo "language: typescript"
test -f tsconfig.json                                           && echo "  ts-config: present"
{ test -f pyproject.toml || test -f requirements.txt || test -f setup.py; } && echo "language: python"

# Step 2 — TypeScript runtime (only if language=typescript)
test -f bun.lockb         && echo "runtime: bun, pm: bun"
test -f pnpm-lock.yaml    && echo "runtime: node, pm: pnpm"
test -f yarn.lock         && echo "runtime: node, pm: yarn"
test -f package-lock.json && echo "runtime: node, pm: npm"

# Step 3 — Rust workspace shape
grep -q '\[workspace\]' Cargo.toml 2>/dev/null && echo "rust: workspace"

# Step 4 — Go module shape
grep -E '^go [0-9]+\.[0-9]+' go.mod | head -1   # toolchain version
grep -q '^// +build' . -r 2>/dev/null && echo "go: legacy build tags present"

# Step 5 — Python PM + framework (only if language=python)
test -f uv.lock           && echo "pm: uv"
test -f poetry.lock       && echo "pm: poetry"
test -f Pipfile.lock      && echo "pm: pipenv"   # legacy
{ test -f requirements.txt && ! test -f uv.lock && ! test -f poetry.lock; } && echo "pm: pip"
grep -qE '(^|[[:space:]"])fastapi[>=<~!"[:space:]]' pyproject.toml requirements.txt 2>/dev/null && echo "framework: fastapi"
grep -qE '(^|[[:space:]"])django[>=<~!"[:space:]]'  pyproject.toml requirements.txt 2>/dev/null && echo "framework: django"
grep -qE '(^|[[:space:]"])flask[>=<~!"[:space:]]'   pyproject.toml requirements.txt 2>/dev/null && echo "framework: flask"
# Python version source of truth
test -f .python-version   && cat .python-version
grep -E '^requires-python' pyproject.toml 2>/dev/null

Persist as $LANG ∈ {go, rust, typescript, python}, plus per-language sub-fields ($RUNTIME, $PM, $WORKSPACE, $FRAMEWORK). The next phases branch on these values.

If no match is found, abort with a one-line notice. The skill does not invent a language.

Phase 1 — Collect Preferences

Ask the user only for inputs that cannot be inferred. Use one consolidated AskUserQuestion block, never one question per phase.

Three security levels:

• minimal — lint + test + build. No security gates. Use only for prototypes.
• standard (default) — adds CodeQL/Semgrep, dependency audit, secret scan
• strict — adds container scan (trivy), SBOM (syft), license check, signing

The full preferences-elicitation script lives in references/USER_HEURISTICS.md. It also documents the rationale for each default, citing the project owner's vault notes.

Phase 2 — Apply User Heuristics

The pipeline emits five opinionated checks regardless of language. Each is sourced from the project owner's notes and is mandatory unless the user opts out explicitly.

For each heuristic, references/USER_HEURISTICS.md contains:

• The exact rationale from the source notes
• The detection technique per language
• A copy-pasteable workflow snippet
• The expected failure mode if the heuristic catches something

Phase 3 — Wire Security Gates

Branch on the security level chosen in Phase 1. Each gate is a separate job that runs in parallel with the test job whenever possible.

Per-language tool selection (CodeQL languages, audit commands, container base scan strategy) is captured in references/SECURITY_GATES.md.

Anti-patterns the generator must refuse to emit:

• pull_request_target with checkout of the PR head (RCE vector)
• permissions: defaulting to write-all at the workflow level
• Third-party actions without a pinned commit SHA when security level = strict
• Secrets passed via env: at the workflow level (must be job- or step-scoped)
• actions/checkout@v4 followed by running untrusted scripts before any allowlist check

Phase 4 — Generate Workflow YAML

Emit the YAML in this skeleton, populated from the language-specific reference:

name: ci
on:
  push:
    branches: [main]
  pull_request:
    branches: [main]

permissions:
  contents: read

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

jobs:
  meta:
    name: actionlint
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: rhysd/actionlint@v1   # pinned via SHA in `strict`

  lint:
    needs: meta
    # populated from <language>.md

  test:
    needs: meta
    # populated from <language>.md
    # includes: race flag, coverage gate, N+1 assertion, leak detection

  security:
    needs: meta
    # populated from SECURITY_GATES.md per security-level

  build:
    needs: [lint, test, security]
    # populated from <language>.md

Per-language fully populated workflows (with matrix, cache, environment variables, and the canonical job graph) live in:

• references/GO.md — actions/setup-go@v5, go test -race -cover -covermode=atomic, staticcheck, golangci-lint, govulncheck, optional goreleaser
• references/RUST.md — actions-rust-lang/setup-rust-toolchain@v1, Swatinem/rust-cache@v2, cargo fmt --check, cargo clippy -D warnings, cargo nextest, cargo-llvm-cov, cargo-audit, cargo-deny
• references/TYPESCRIPT.md — setup-bun / setup-node + pnpm/action-setup, tsc --noEmit, eslint, vitest/jest with --detectOpenHandles --detectLeaks, N+1 query test template, size-limit
• references/PYTHON.md — actions/setup-python@v5 / astral-sh/setup-uv@v3 / snok/install-poetry, ruff check+ruff format --check, mypy --strict/pyright, pytest --cov-fail-under, bandit+pip-audit+safety, N+1 templates (Django assertNumQueries / SQLAlchemy event listener), free-threaded race (python3.14t + hypothesis), optional PyPI trusted publishing

Always pin third-party actions:

• standard level: @vN major-version pin
• strict level: full commit SHA pin (@<40-char-sha>) with a comment naming the version

Phase 5 — Validation & Dry-Run

Before reporting success, run the local checks:

# 1. Lint the YAML
actionlint .github/workflows/*.yml
yamllint -d relaxed .github/workflows/*.yml

# 2. Confirm secrets referenced in the YAML
grep -hoE '\$\{\{ secrets\.[A-Z_]+ \}\}' .github/workflows/*.yml | sort -u

# 3. Verify all third-party actions are pinned
grep -hE 'uses: ' .github/workflows/*.yml | grep -v 'actions/' | grep -v '@[a-f0-9]\{40\}\|@v[0-9]'

# 4. List required workflows for branch protection
grep -E '^  [a-z_-]+:$' .github/workflows/ci.yml | sed 's/[: ]//g'

Each finding is added to the post-generation report. If actionlint reports errors, abort and emit the diff for the user instead of writing the file.

The full validation matrix and the optional gh workflow run --ref <branch> ci.yml smoke test are in references/CHECKLIST.md.

Phase 6 — Deliver

1. Write the workflow file(s) to .github/workflows/. Default filename: ci.yml. Optional: release.yml, nightly-load.yml.
2. Write .github/dependabot.yml if Phase 1 enabled it.
3. Emit the generation report (see Output format).
4. Surface the suggested branch protection rules — do not apply them automatically. Provide the gh command the user can run.
5. Hand off to @github-commit to commit the new files; do not commit unless the user explicitly approves.

A worked end-to-end run of Phase 0 → Phase 6 for a Go API project is in EXAMPLE.md.

Audit mode (alternate entry)

When the user asks to audit an existing pipeline rather than generate one, branch at the very top of Phase 0 into the audit procedure documented in references/AUDIT.md. The audit walks Phase A0 → A5 (inventory → heuristic compliance → security gate compliance → anti-pattern sweep → action freshness → caching/concurrency hygiene) and emits a severity-ranked findings table with fix proposals.

The audit is read-only by default. The user can opt in per finding ID for the skill to apply unified-diff fixes; each fix is tagged SAFE / REVIEW / BREAKING and re-validated with actionlint before report-ok. The skill never commits.

Pick the entry point by user intent:

Constraints

• Never commit the generated YAML without explicit user approval — emit the diff first
• Never apply branch protection rules without an explicit gh api call confirmed by the user
• Never emit pull_request_target with PR-head checkout
• Never default permissions: to write-all at workflow level — start at contents: read and elevate per job
• Never pass secrets via the workflow-level env:; scope them to the job or step
• Never use third-party actions without a version pin (major tag minimum, SHA in strict)
• Never generate a pipeline for a language outside {go, rust, typescript, python} — abort with a one-line notice
• Never silently downgrade the security level when a tool is missing — surface the gap to the user
• Must load the matching language reference before emitting any YAML
• Must run actionlint on the generated file before reporting success
• Must include a header comment listing required secrets, the trigger graph, and the source skill
• Must keep the workflow under 500 lines per file; split into ci.yml + release.yml + nightly-load.yml when needed
• Must not invent secrets or repository variables — every ${{ secrets.X }} reference must be listed in the report

Output format

Generation report (printed verbatim after every successful run):

ci-cd-generator: <action>
  language:      <go | rust | typescript | python>
  runtime/pm:    <node+pnpm | bun | rust+nextest | go+toolchain x.y | python+uv | python+poetry | python+pip>
  framework:     <fastapi | django | flask | n/a>   # python only
  security:      <minimal | standard | strict>
  workflows:     <files written, paths under .github/>
  jobs:          <ordered list of job ids in ci.yml>
  required-secrets: <SECRET_A, SECRET_B, … or "none">

Suggested branch protection (run after committing):
  gh api -X PUT repos/<owner>/<repo>/branches/main/protection \
    -F required_status_checks.strict=true \
    -F 'required_status_checks.contexts[]=lint' \
    -F 'required_status_checks.contexts[]=test' \
    -F 'required_status_checks.contexts[]=security' \
    -F enforce_admins=true \
    -F required_pull_request_reviews.required_approving_review_count=1

Next: review the diff, then `/valarmindskills:github-commit`.

Related Skills

• @github-commit — commit the generated workflow files following Conventional Commits
• @github-release-note — generate release notes when the release workflow tags a version
• @github-pr-review — review the diff that introduces the workflow
• @clean-code — apply principles to the YAML (consistent naming, no duplication across jobs)
• @code-security-review — multi-stack security review job covering generic flows + Go (Gin/Fiber, references/golang/) + Next.js 16 App Router (references/nextjs/); active testing via references/TESTING_PHASES.md, run nightly via the load testing slot; 100-vuln class catalog via references/WEB_VULNERABILITIES.md when the pipeline must spell out which classes each gate addresses
• @code-debugger — diagnosing a failing generated workflow before reporting it as broken

References

• GO — Go pipeline template, tooling matrix, canonical workflow
• RUST — Rust pipeline template, tooling matrix, canonical workflow
• TYPESCRIPT — TypeScript pipeline template (Node + Bun), N+1 and leak templates
• PYTHON — Python 3.13/3.14 pipeline template (FastAPI / Django / Flask), uv/poetry/pip PM detection, N+1 templates, free-threaded race testing, PyPI trusted publishing
• USER_HEURISTICS — coverage gate, N+1, race PBT, leak detection, load testing — rationale and snippets
• SECURITY_GATES — SAST, SCA, container, SBOM, secret scan, license — per-language tool matrix
• CHECKLIST — post-generation validation, branch protection, smoke test
• AUDIT — audit mode procedure, detection rules, severity matrix, report format, opt-in fix application
• EXAMPLE — end-to-end worked example for a Go API project