ratchetlisted

# /ratchet — Autonomous Improvement Loop An autoresearch-style ratchet that continuously improves your agent. Each iteration: improve → run agent → eval → keep (if better) or revert (if worse) → repeat. --- ## Step 1: Configure (MANDATORY — do not skip) You MUST ask the user to confirm the configuration before starting the loop. Never proceed without explicit confirmation. ### If `program.md` exists Read it. If any values look like placeholders (e.g., "Describe what you want to improve", "your_agent.py"), treat them as empty. Present the current configuration to the user and ask them to confirm or modify: > **Here's the current ratchet configuration:** > > 1. **Objective:** {objective from file} > 2. **Agent run command:** `{command from file}` > 3. **Traces directory:** {traces_dir from file} > 4. **Metrics to optimize:** > {list each metric with direction and weight} > 5. **Stopping conditions:** {max_iterations} iterations, {max_duration_hours}h, plateau after {plateau_patience} > > **Does this look right, or would you like to change anything?** Wait for the user to confirm or provide modifications. Update `program.md` with any changes. ### If `program.md` does not exist Ask the user each question: 1. **Objective** — what do you want to improve about your agent? 2. **Agent run command** — the shell command that runs your agent and generates traces (e.g., `uv run python examples/technova_agent.py`) 3. **Traces directory** — where traces are written (default: