What problem does it solve?
Agents repeat the same mistakes across sessions because their prompts, skills, and memory never get fixed. This Skill gives an agent a structured protocol to detect failure signatures in its own trajectory and repair its harness mid-session, with changes promoted to main only through a reviewed change request.
Core Features & Use Cases
- Failure signature detection: Scans recent turns for repeated tool failures, rediscovery loops, stalled objectives, repeated multi-step patterns, and exception-raising code.
- Four-pass refinement: Runs CRUD passes over prompts, sub-agents, skills/tools, and memory, with deletion treated as a first-class outcome.
- Two operating modes: In-session self-refinement that lands edits on the session branch immediately, and project-level reflection where the harness-reflector agent fans out session-reviewer sub-agents and aggregates findings.
- Use Case: An agent notices the same shell command has failed twice in a session, so it pauses, records the working alternative in a skill, commits the harness change, opens a change request, and resumes its task with the fix active on the next turn.
Quick Start
Load the kortix-harness-refinement skill and review my recent turns for failure signatures, then apply the four-pass refinement to the harness.
