a5c-ai - For medium and high stakes changes (low stakes can proceed inline)
- When structured execution with verification is needed
Process
- Load plan - Validate plan exists and is approved
- Stakes enforcement - High: halt without plan. Medium: ask. Low: proceed.
- Worktree isolation - Offer git worktree based on stakes level
- Execute steps - For each: mark in-progress, locate files, read, modify, verify, mark complete
- Phase checkpoints - Summarize and ask: continue, review, or pause
- Failure handling - Stop immediately, investigate, propose fix, get approval for deviations
- Code review - Run code-reviewer agent (APPROVE / APPROVE_WITH_NITS / REQUEST_CHANGES)
- Security review - Run security-reviewer agent (halt if failed)
- Completion summary - Steps completed, files changed, test status, plan location
Key Rules
- Follow the plan strictly; deviations require explicit approval
- Verify before declaring done; run verification after each step
- Track progress visibly via plan document updates
- Read files before modifying them
- Complete code and security reviews before finishing
Tool Use
Invoke via babysitter process: methodologies/rpikit/rpikit-implement