ci-parity

Catch the "passes locally, fails in CI" bugs before you push — statically, in about 50ms, with no Docker.

You edit, commit, push, wait 8 minutes, and CI fails on something your machine never sees: a Node version your workflow pins differently, a lockfile that's out of sync with npm ci, a secret the workflow injects that your shell doesn't have, a sed -i that only works on Linux. ci-parity reads your workflow files and your repo's config and predicts those failure classes locally.

$ npx ci-parity

ci-parity  scanned 1 workflow file(s) in .

  [fail-in-ci] Node major version disagrees across your config files
    - .nvmrc: 18 (Node 18.x)
    - package.json engines.node: >=20 (Node 20.x)
    - ci.yml setup-node node-version: 22 (Node 22.x)
    fix: Pick one source of truth. Put the version in .nvmrc and point the
         workflow at it with node-version-file, then match engines.node.

  [fail-in-ci] 1 dependency is in package.json but not in package-lock.json
    - out of sync: left-pad
    - workflow install: npm ci
    fix: Run `npm install` to regenerate package-lock.json, then commit it.

  [parity-risk] 1 env var your workflow injects is absent locally
    - DEPLOY_KEY <- secrets.DEPLOY_KEY (ci.yml); not in your shell or .env

  2 likely-to-fail-in-CI, 2 parity risk(s). Push will likely fail CI.

Exit code is 1 when something will likely fail CI, so it drops straight into a pre-push hook.

How it's different from what you already use

• actionlint checks that your workflow YAML is valid — expression types, action inputs, shellcheck on run: scripts. It does not compare your workflow to your local environment. ci-parity is exactly that comparison.
• act runs your workflow locally in Docker. That's heavier, its runner images are incomplete, secrets and services need manual wiring, and the OS still differs from GitHub's runners. ci-parity doesn't run anything — it predicts the mismatches statically. Use both: ci-parity for the fast pre-push gate, act when you need to actually execute a job.

What it checks

All static. No network, no Docker, no running your build.

• Multi-source Node/tool version drift. Compares the Node major across .nvmrc, .tool-versions, package.json engines.node, actions/setup-node in your workflows, and your locally-installed node. Single-source setups (node-version-file: .nvmrc) are treated as healthy and never flagged.
• packageManager vs lockfile. A packageManager: "pnpm@9" (Corepack) field with no pnpm-lock.yaml committed resolves installs differently in CI than locally.
• Lockfile staleness. Dependencies in package.json that aren't in package-lock.json — the exact drift that makes npm ci / frozen installs exit non-zero in CI while npm install passes on your machine.
• Env/secrets missing locally. Names a workflow injects from secrets.* / vars.* that aren't in your shell or a local .env. When a test reads one, local and CI can take different code paths.
• OS-only shell steps. run: steps using GNU-only forms (sed -i without a suffix, readlink -f, grep -P, date -d, stat -c) that behave differently on a macOS/Windows dev machine than the Linux runner.

Install

npx ci-parity          # one-off, no install
npm i -D ci-parity     # add to a project

GitHub Action

Run the same check as a CI job — catches the same failures on a teammate's PR who skipped the hook:

- uses: fernforge/[email protected]
  with:
    path: .        # optional, default "."
    strict: false  # optional, fail on warnings too

Pre-push hook

With simple-git-hooks:

{
  "simple-git-hooks": { "pre-push": "npx ci-parity" }
}

Or husky — drop npx ci-parity into .husky/pre-push. A non-zero exit blocks the push.

Options

ci-parity [path]     scan a repo (default: current directory)
--json               machine-readable output
--strict             exit 1 on parity risks (warnings) too, not just likely failures
--warn-only          always exit 0 (report without blocking)
-h, --help
-v, --version

Exit codes: 0 clean (or --warn-only), 1 a finding that will likely fail CI (or any finding under --strict), 2 internal error.

Scope

GitHub Actions + npm/pnpm/yarn today. It reasons about .github/workflows/*.yml and standard Node config files; other CI providers aren't covered yet. It predicts a specific, common set of failure classes — it's not a substitute for running your tests.

License

MIT. Built autonomously by an AI agent.