cm-harness

Applies human-approved self-improvement findings to agent harness files as atomic commits.

3|Updated Jan 29, 2026
One-click install
npx skills add https://github.com/PurdueRCAC/rcac-mcp --skill cm-harness-purduercac
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: cm-harness
Source: https://github.com/PurdueRCAC/rcac-mcp/tree/main/.agents/skills/cm-harness
Command: npx skills add https://github.com/PurdueRCAC/rcac-mcp --skill cm-harness-purduercac

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Feature work generates friction findings logged in META.md files, but turning those observations into actual improvements to agent skills, templates, and scripts requires a careful, human-gated process that avoids weakening safety gates or thrashing prior decisions. ## Core Features & Use Cases - Finding triage: Reads open findings from spec/{slug}/META.md via meta_status.py, filtered by id, severity, or across all features with --all. - Human-gated application: Previews a concrete diff per finding, confirms with the human, then applies each fix as its own atomic [harness] commit on main (or a harness/{slug} branch with PR mode). - Anti-thrash ledger: Records every applied, rejected, and deferred decision in harness-log.md so future runs detect reverts of recent changes or repeats of rejected fixes. - Use Case: After a feature merges, run /cm-harness my-feature --dry-run to preview fixes for its open findings, then apply the approved ones as individual commits with status flips and ledger entries. ## Quick Start Ask the agent to run /cm-harness with a feature slug to review and apply that feature's open harness findings as atomic commits on main.

Frequently Asked Questions about cm-harness

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I apply harness improvement findings from a feature's META.md?

Run /cm-harness with the feature slug to enumerate open findings via meta_status.py, preview a diff per finding, confirm with the human, and apply each fix as its own atomic [harness] commit on main. Use --dry-run first to preview without making changes.

How do I preview harness changes without committing them?

Pass --dry-run to /cm-harness to execute all steps up to the per-finding diff preview, then stop with no edits, commits, or branches. This is the recommended first pass and is automatic when reading a META.md from an unmerged branch.

Can cm-harness modify product source code or lifecycle state?

No. cm-harness is meta/maintenance only and edits only files under .agents/ such as skills, templates, scripts, and docs. It never touches product code, GOAL/PLAN/TECH/REVIEW artifacts, the FSM, and never writes new META.md findings.

What prevents cm-harness from weakening safety gates?

Any fix that would loosen tests, the sandboxed CLI verify, blind-review integrity, auth posture, or invariants.md items requires an explicit typed human override and is treated as suspect by default. Findings proposing to allow a real SSH host in sandbox.sh are default rejects.

Why does cm-harness read harness-log.md before applying fixes?

The ledger is anti-thrash memory recording every prior applied, rejected, and deferred decision with commit SHAs. A proposed fix that reverts a recent change or repeats a rejected one is flagged to the human instead of being silently re-applied.