What problem does it solve? It prevents undisciplined coding in the mj-agent project by enforcing HITL Stage 8 methodology: test-first behavior changes, root-cause bug fixes instead of symptom patches, fresh evidence before claiming completion, and strict handling of the three change flavors (pure code, in-source canonical prompts/skills, infrastructure). ## Core Features & Use Cases - Red-green-refactor enforcement: Requires writing a failing test, observing RED, implementing minimal code to GREEN, then refactoring for any behavior change. - Root-cause-first debugging: Mandates a stable reproduction path and root-level fixes, explicitly rejecting try-except error swallowing or disabling failing tests. - 3-flavor classification: Routes changes by path — flavor A (pure code with TDD), flavor B (in-source canonical SKILL.md/prompt edits that always trigger a mandatory HITL stop), and flavor C (infra changes requiring healthchecks, docker compose rehearsal, and uv lock). - Use Case: After a Plan and SPEC are confirmed, a developer says "开始编码" and the skill produces a step-by-step coding plan with an implementation trace (red-green log, root-cause log, infra log, fresh evidence) ready for handoff to Stage 10 verification. ## Quick Start Ask the agent to start implementing the confirmed plan using the mj-agent-flow-implement skill so it classifies the change flavor and applies the correct coding methodology.
