Revises the research harness workflow rules, hooks, and skills through an approval-gated process.
npx skills add https://github.com/coollx/stable-harness --skill meta-coollx
Please help me install this Agent Skill. Skill: meta Source: https://github.com/coollx/stable-harness/tree/main/.claude/skills/meta Command: npx skills add https://github.com/coollx/stable-harness --skill meta-coollx
harness: lines in docs/tasks/BACKLOG.md plus friction observed in the current conversation. - Diagnosis and Minimal Revision: Identifies the rule or gap causing friction, reads the original rationale in AGENTS.md, and prefers deleting or simplifying rules over adding new ones. - Adversarial Risk Review: Dispatches adversarial reviewer agents to attack substantial proposals (routing-table rows, hook behaviors, new rules) before they are applied. - Dual Application: Applies approved changes both to the current project and to the research-harness template repository so future projects inherit them. - Use Case: A researcher notices a hook repeatedly blocking legitimate commits. They invoke /meta, which surfaces the friction, traces the hook's original rationale, proposes a simplification, gets adversarial review, and applies the change only after explicit approval. ## Quick Start Invoke /meta explicitly and ask it to review the harness friction items in the backlog and propose minimal revisions to AGENTS.md.High-intent search queries and answers about installing and using this skill.
Invoke /meta explicitly, which collects friction candidates, diagnoses the rule's original rationale, and proposes a minimal revision. The change is applied only after your explicit approval and is committed with a [meta] spine line.
No. The skill has a hard rule that it may only run when the researcher explicitly invokes /meta. The agent may recommend invoking it, but must stop and report if triggered any other way.
Approved changes are applied in two places: the current project and the research-harness template repository, so future projects inherit them. The template edit requires separate explicit approval before being applied.
Substantial changes such as routing-table rows, hook behaviors, or new rules are attacked by adversarial reviewer agents before application. Proposals are also checked against red-flag words like sha, gate, stamp, and tuning constant.
Do not use it for ordinary research tasks or content work; it exists solely to modify the harness itself. If you were not explicitly invoked by the researcher, the correct behavior is to stop and say so.
lepahcAutomate end-of-session reflections and export insights to the planning backlog.
rlesniakGenerate conventional commit messages using structured XML templates.
SpencerGossReview proposals with four mandatory counterarguments and a proposer response workflow.
rkumar261Restrict Edit and Write operations to a configured directory for a session.