What problem does it solve? AI agent skills and plugins change constantly, but there is no systematic way to know whether a recent edit degraded real-world behavior. This Skill closes that loop by collecting actual agent run traces, labeling failures, attributing regressions to recently changed skills or plugins, and proposing concrete fixes. ## Core Features & Use Cases - Trace Dataset Collection: Harvests real agent runs from Loki events and Postgres into a durable JSONL dataset with transcripts, tool calls, and ok/fumbled/failed labels. - Regression Mining & Attribution: Clusters failed and fumbled runs by skill and plugin contribution, then confirms root causes via agent self-reflection (debrief) and deep-dive workers. - Evaluation Loop: Re-runs replay-safe tasks with fresh agents, scores them with a completion/efficiency rubric, and keeps skill text edits only when measured scores improve. - Use Case: After a daily batch of ~100 runs, detect that a recently edited skill correlates with five fumbled runs, identify the exact instruction causing failures, apply a fix, verify the rubric score improves, and open a PR for human review. ## Quick Start Run the self-evolution batch flow to collect yesterday's agent runs and report any regressions tied to recently changed skills.
