ls-self-improve

Improve skill trigger accuracy by scanning installed skill catalogs to evaluate existing evals, measuring false positives and false negatives, then retraining underperforming skills through the /train mutation loop to update trigger phrases and compiled help text.

False negatives and false positives in skill descriptions occur when metadata drifts or trigger phrases fail to match user intent reliably, causing the agent to skip useful tools or fire the wrong tool during tool-routing orchestration across sessions.

Automate agent self-improvement by locating each skill's SKILL.md and eval set, running the skill eval runner to identify low-accuracy skills, then spawning trainer agents via LiteHarness prompts to mutate descriptions and optionally recompile with LiteCLI.

LiteCLI is optional for recompiling updated skill description fields, ensuring compiled help text and matching patterns stay current after the /train mutation loop modifies trigger phrases in the SKILL.md metadata.

Run a self-improvement cycle after adding new skills to your catalog or noticing tool misfires, ensuring the agent consistently selects the right tools by refreshing trigger accuracy and compiled help text across repeated sessions.