majiayu000 Always
- Keep PRs small and auditable; avoid drive-by refactors.
- If invariants/plumbing are touched: run the canonical proof command listed in EVALSETS.yaml (currently:
npx nx run evalsets:migration-smoke). - Never hand-edit vendored contracts; use sync scripts + drift gates.
Claude lane (enforced triggers)
If any of the following are true, you MUST propose a Claude Code run before proceeding:
- Repo-wide enumeration / dependency mapping across many files.
- Independent audit/review requested (migration correctness, invariants, compliance).
- Large-context synthesis from long documents.
When proposing Claude, provide all of:
- Claude prompt (ready to copy/paste)
- Artifact path to save output into (default:
docs/reviews/YYYYMMDD-HHMM_<slug>.md) - Acceptance criteria (what the artifact must contain)
Do not proceed until the artifact exists, unless the user explicitly declines Claude. If declined, proceed in Codex but label increased risk.
Claude Code (optional specialist)
If the user asks for large mechanical edits/refactors, or if the change would touch many files, propose a Claude Code handoff.
When proposing Claude, output a paste-ready “Claude Work Order” with:
- Goal
- Allowed paths
- Hard constraints (no churn, no invariant changes, etc.)
- Verification commands
- Expected deliverable summary Do not proceed with Claude unless the user explicitly says “yes”.
Claude Code (optional specialist)
Follow the AGENTS.md handoff protocol. Codex must provide a paste-ready work order (goal, allowed paths, constraints, verification, expected output), then re-verify and update the worklog after Claude returns results.