floom-skill-evals

name	floom-skill-evals
description	Use when the user asks to create, run, or review skill evaluations with baseline-vs-with-skill A/B tests, Docker-isolated task runs, deterministic verifiers, receipt generation, or claims like pass-rate lift for a skill or pack.
metadata	{"version":"0.1.0"}

Floom Skill Evals

Run simple, deterministic skill receipts:

1. Create eval tasks for a skill.
2. Run baseline and with-skill conditions in isolated Docker workspaces.
3. Grade with deterministic verifier.
4. Save result.json and a markdown receipt.

When to use

Use this skill for requests like:

• "evaluate this skill"
• "baseline vs with skill"
• "run skill receipts"
• "prove skill lift"
• "Floom Verified style run"

Commands

1) Scaffold an eval task

python3 scripts/scaffold_eval_task.py \
  --skill seo-audit \
  --task technical-homepage-audit \
  --out ./skills/seo-audit/evals

This creates:

• <out>/<task>/task.md
• <out>/<task>/input/
• <out>/<task>/expected.json
• <out>/<task>/verifier.py
• <out>/<task>/eval.json

2) Run A/B eval

python3 scripts/run_skill_eval.py \
  --eval-dir ./skills/seo-audit/evals/technical-homepage-audit \
  --skill-path ~/.codex/skills/seo-audit \
  --agent codex

Outputs:

• runs/<timestamp>/result.json
• runs/<timestamp>/report.md
• condition workspaces and logs

Default policy rejects scripted output commands (echo, printf, heredoc writes) in agent_command. Use --allow-scripted only for runner plumbing tests.

Execution model

• baseline: run agent command without skill mount.
• with_skill: run same command with skill mounted at /skill and env FLOOM_SKILL_PATH=/skill.
• verifier: runs deterministic checker inside each condition workspace.
• audit: simple deterministic summary from pass/fail/runtime deltas.

Required eval contract

eval.json must include:

• id
• skill
• timeout_minutes
• trials
• agent
• agent_command.baseline
• agent_command.with_skill
• verifier
• expected_output

Command placeholders supported:

• {task_dir} mounted as /task
• {output_path} path in workspace where agent writes result

Notes

• This v0 runner is Docker-first and local-first.
• Host keeps credentials; container only gets mounted task workspace.
• If API keys are needed, pass scoped env explicitly with --pass-env OPENAI_API_KEY etc.
• Receipts include provenance hashes for verifier and both condition commands.