cadence: Skill for Claude Code

PR land (protected main)

main requires three status checks — ci-success (ci.yml), security-success (security.yml) and codeql-success (codeql.yml) — and enforce_admins is on, so there is no direct-push path, for anyone, and a red security or CodeQL run blocks the merge as surely as a red test leg. Everything lands branch → PR → green → squash merge.

1 — Preflight locally (cheaper than a CI round-trip)

• Run pnpm turbo run lint typecheck test build before pushing.
• Windows caveat: Linux CI is canonical. A few Windows-local test failures are known environment issues (pnpm/tempRepo/spawn races) — if something fails only locally and the failure smells environmental, push and let the CI matrix decide rather than debugging the dev box.
• Before any git operation that touches shared tree state — stash, reset, checkout onto main — re-check git status immediately beforehand, not from a check done minutes earlier in this same session. Another process (a subagent, a background build, a second terminal) can have changed the tree in between.

2 — Commit hygiene

• Stage explicitly, never git add -A. Off-limits: live .cadence/state.json
  • STATE.md dirt mid-loop (the settle commit owns those), .agents/, launch/ (local-only by explicit decision), .claude/scheduled_tasks.lock, stray uncommitted SESSION-*.md handoffs.
• Single-commit settle convention when a phase is closing: source + tests + docs + phase artifacts land together in one commit, no separate chore: settle commit after it.
• Feature PRs carry their .changeset/*.md. Conventional-commit subject with the phase id: feat: <what> (phase NNN).

3 — Push + PR

git push -u origin <branch>
gh pr create --title "<subject>" --body "<what/why, invariants checked>"

One logical change per PR. The PR body ends with the standard 🤖 Generated with [Claude Code](https://claude.com/claude-code) footer.

4 — Babysit CI (the flake protocol)

• Watch with gh pr checks <n> --watch (or poll gh pr checks <n>).
• On red, get the failing leg + test before reacting: gh run view <run-id> --log-failed | head -100.
• Re-run once, without investigation, only when all three hold: a single OS/Node leg is red, the diff can't plausibly touch the failing area, and the failure matches a known flake (reference: macOS/Node22 timeout in settle-codereview-convergence.test.ts; new node --test subprocess spawns under parallel load have flaked before).
• Anything else — multiple legs, plausible relation to the diff, an unfamiliar failure, or a second red after the re-run — is real until proven otherwise: investigate, don't loop re-runs.