eval-triggers

IMPORTANT: This skill is expensive (makes many LLM API calls). Only run when the user explicitly asks for it. Never run proactively.

Before running, ask the user which eval source to run unless they already specified it in $ARGUMENTS:

• Claude Code only — uses run_trigger_eval.py --track claude
• Codex only — uses run_trigger_eval.py --track codex
• Both — runs both tracks sequentially

Step 1 — Classify:

The runner writes to the canonical layout at evals/results/trigger/<skill>/<track>/<source>/results.json. You do not pass an output directory.

For Claude Code:

python3 evals/scripts/run_trigger_eval.py --track claude --source combined

For Codex:

python3 evals/scripts/run_trigger_eval.py --track codex

Step 2 — Analyze:

python3 evals/scripts/analyze_trigger_results.py \
  --track claude --source frontmatter

This writes the cross-skill summary to evals/results/trigger/_summary/<track>/<source>/summary.{json,md}. Report:

• The gate verdict (pass/fail at F1 ≥ 0.90, with the list of skills below threshold).
• Per-skill precision, recall, F1.
• Any noteworthy false negatives / positives.

Step 3 — Compare (optional):

If the user provides a baseline summary or you have one from a previous run:

python3 evals/scripts/compare_trigger_runs.py \
  --baseline <baseline-summary>.json \
  --candidate evals/results/trigger/_summary/<track>/<source>/summary.json \
  --output-markdown local/comparison.md

Summarize regressions and improvements. For ad-hoc baselines that should not be committed, store them under local/ (gitignored).