🏃 Sprint — every TODO item gets a full mini-sprint: autonomous plan→implement→review for Claude Code via git-hook agent chain, isolated worktree, deterministic gate & fail-closed review
Agent
Implements ONE planned item in the isolated worktree, exactly one commit per item. Use after planner + navigator, and again on each gate/review loopback to fix only the reported findings.
Agent
Finds existing patterns, conventions and the blast radius via CodeGraph BEFORE code is written. Use proactively before implementation and before conventions review.
Agent
Turns ONE TODO item into a bounded plan with verifiable acceptance criteria and a risk tier. Use proactively as the first step of each item, before any code change.
Agent
Checks adherence to the codebase's existing conventions and that the solution is the best-practice fit for THIS project's stack. Use after the correctness reviewer. Read-only + writes its findings JSON.
Agent
Checks the implementation against the acceptance criteria for correctness, edge cases and missing tests. Use AFTER the deterministic gate is green. Read-only + writes its findings JSON.
Agent
STAGE 2 (scaffold — not yet wired into the loop). Verifies UI implementation in the browser via screenshots + a11y tree. Use only for items with a frontend/UI part; Figma optional.