/kit-tune

Auto-tune agent model tiers from measured performance. Thresholds live in .claude/metrics/tuning.json (min samples, promote-below-fit, demote toggle).

1. Refresh the scorecard first:

   "${CLAUDE_PLUGIN_ROOT}"/scripts/kit-stats.sh
   

2. Show the dry-run proposal (writes nothing):

   "${CLAUDE_PLUGIN_ROOT}"/scripts/kit-tune.sh --agents-dir "$CLAUDE_PROJECT_DIR/.claude/agents"
   

3. Summarize the proposed promotions/demotions and the evidence (runs + fit).
4. Only if the user passed --apply (or confirms), run:

   "${CLAUDE_PLUGIN_ROOT}"/scripts/kit-tune.sh --apply --agents-dir "$CLAUDE_PROJECT_DIR/.claude/agents"
   

   Then show git diff of the changed agent files so the human reviews before committing.

Never promote/demote without enough samples — the script enforces min_samples, but call it out if a proposal is based on thin data.