Checkpoint

/skill:checkpoint — Quality Gate

Purpose: Fast quality gate that runs lint, build, and tests for changed packages. Does NOT run code review — use /skill:domain-review separately for that.

When to Use

• Before pushing a branch or creating a PR
• After a batch of edits to verify lint/typecheck/tests still pass
• As a sanity gate inside larger flows (invoked by /skill:pr, /skill:ship)

When NOT to Use

• You want review feedback, not just green tests — use /skill:domain-review
• No changes staged or tooling not yet configured — run /skill:configure first
• Full delivery flow needed — use /skill:pr or /skill:ship, which call checkpoint internally

Execute these steps in order. Report results at the end.

Step 1: Read the repo-delivery contract

Read the repo's CLAUDE.md ## Commands section for the required schema (see _internal/repo-delivery). Required fields:

• lint — lint command
• build — build/typecheck command
• test — test command
• package_manager — pnpm, npm, yarn, bun, pip, poetry, cargo, go, dotnet, ...
• runtime_probes — optional commands that prove native/runtime dependencies load

If the ## Commands section is missing, fail fast with:

Repo missing required CLAUDE.md ## Commands section.
See _internal/repo-delivery for the schema.

Do not infer commands from manifests. Do not fall back to package.json scripts. The contract is declared; if it's absent, the repo is not ready for this flow.

Step 2: Identify Changed Packages

Prefer the declared default_branch remote ref when available:

git diff origin/<default_branch> --name-only

If that ref does not exist, fall back in this order:

git diff <default_branch> --name-only
git diff HEAD~1 --name-only
git diff --name-only
git diff --cached --name-only

Do not fail the checkpoint only because origin/<default_branch> is absent; fresh clones, local-only branches, and test repos often lack that ref.

Extract unique package/service paths from the changed files. For a pnpm monorepo, map file paths to their workspace package names (check pnpm-workspace.yaml and each package's package.json).

If no changes vs origin/main, check for uncommitted changes:

git diff --name-only
git diff --cached --name-only

Report: "N files changed across M packages: [list]"

Step 3: Run Linter (if configured)

pnpm -r lint

Or the project-specific lint command from Step 1.

• If no lint script exists, skip and note "No linter configured."
• If lint fails, record the output but continue to build.

Step 4: Run Build / Typecheck

pnpm -r build

Or the project-specific build command from Step 1.

• If build fails, record the output but continue to tests.

Step 5: Run Runtime Probes

If runtime_probes are declared in CLAUDE.md ## Commands, run each command after build. Treat failures as checkpoint failures because they indicate code may compile while runtime dependencies, native bindings, generated clients, or drivers cannot load.

If no probes are declared, skip and note "No runtime probes configured."

Step 6: Run Tests for Changed Packages

For each changed package identified in Step 2:

pnpm --filter <package-name> test -- --run

• Use --run to prevent vitest from entering watch mode.
• If a package has no test script, skip it and note "No tests for ."
• If tests fail, record the output.

Step 6a: End-to-End Verification (conditional)

If the diff touches product source (not just tests/docs/config), invoke the built-in /verify to exercise the changed flow before reporting green. Tests prove the code compiles and passes assertions; /verify proves the change actually works when the affected flow is driven end-to-end. Record the outcome as an additional check in the report. If the built-in /verify is unavailable in the session, skip and note "Verify: SKIP (not available)."

Step 7: Report

Output a summary:

## Checkpoint Results

**Files changed:** N files across M packages
**Packages:** [list]

| Check      | Result | Details          |
|------------|--------|------------------|
| Lint       | PASS/FAIL/SKIP | [summary]  |
| Build      | PASS/FAIL | [summary]       |
| Probes     | PASS/FAIL/SKIP | [summary]  |
| Tests      | PASS/FAIL/SKIP | [summary]  |
| Verify     | PASS/FAIL/SKIP | [summary]  |

**Verdict: PASS / FAIL**

• PASS: All checks passed (or skipped with no failures).
• FAIL: Any check failed. List each failure with relevant output.

Notes

• This is the fast gate. It does NOT run /skill:domain-review.
• /skill:pr and /skill:ship call this skill as a prerequisite.
• Relationship to native skills: checkpoint is the concrete command-runner that satisfies superpowers:verification-before-completion's evidence requirement; the two compose rather than compete.
• Known pre-existing test failures (documented in CLAUDE.md or auto-memory) should be noted but not block the verdict unless they are new regressions.

Contract

• Inputs: working tree state. Reads ## Commands from CLAUDE.md for lint / build / test / package_manager (per _internal/repo-delivery).
• Preconditions: inside a git repo; toolchain available on PATH; not on the default branch unless explicitly invoked there.
• Outputs: PASS or FAIL verdict plus per-check breakdown (lint, build, test for changed packages).
• Postconditions: lint/build/test have actually run; the calling workflow may proceed only on PASS.
• Failure modes: any check fails → FAIL; calling workflow (/skill:pr, /skill:ship) must halt. No retries here.