cmdher

A deliberately small harness layer for Command Code.

The goal is simple: let a cheap, capable implementation model do the work, then require one independent review before the result is treated as complete — without growing another full agent framework.

v0 scope

• AGENTS.md defines the project-wide workflow and completion contract.
• .commandcode/agents/code-reviewer.md provides one independent, read-only reviewer pinned to GLM-5.3.
• The intended default implementer is MiMo V2.5 Pro, selected at the Command Code session level.
• Review is policy-enforced in v0. There is intentionally no planner, deep agent, task graph, memory subsystem, retry orchestrator, or custom Mod yet.

Workflow

1. Implement the requested change.
2. Run the relevant validation commands and record the observed results.
3. Delegate one independent review to code-reviewer.
4. If the reviewer returns PASS, hand off the result with evidence.
5. If it returns CHANGES_REQUIRED, fix the findings and run one re-review. If material findings remain after that, stop and report them instead of claiming completion.

Non-goals

cmdher is not intended to compete with full orchestration harnesses. New machinery should only be added after repeated real-world failures show that the small setup is insufficient.