Start

This blocks while monitoring the researcher. The stop hook fires every ~120s with analysis. Act on the guidance: CONTINUE / INVESTIGATE / PIVOT.

3. Active supervision

On each stop hook cycle:

• Read the metric trend — is it improving, stalled, or regressing?
• Read deviation reports — is the researcher following its own protocol?
• Make a decision and ACT (see CLAUDE.md for the mandatory thinking protocol)

If stalled:

1. Read the researcher's current prompts: pixi run researcher-dot-claude-read skill, pixi run researcher-dot-claude-read evaluator, pixi run researcher-dot-claude-read improver
2. Analyze what the researcher is doing wrong (bad iteration discipline, not pivoting, repeating failed approaches, etc.)
3. Edit prompts: echo "new content" | pixi run researcher-dot-claude-edit <name>
4. Restart: pixi run researcher-stop && pixi run researcher-loop --no-clean

4. When the loop returns

Analyze results:

• pixi run researcher-history — metric progression
• pixi run researcher-dot-claude-history — what prompt changes were made

Decide: restart with same prompts, edit prompts and restart, or report success.

PARALLEL VARIANT MODE

Use when testing multiple researcher SKILL.md variants simultaneously.

Full Workflow

1. Create researcher variants in researcher_variants/
2. For each:

   cat researcher_variants/X.md | pixi run researcher-dot-claude-edit skill
   pixi run researcher-variant start --id rv-X
   

3. Monitor: pixi run researcher-variant list
4. Compare: pixi run researcher-variant compare
5. Park all: pixi run researcher-variant park --id rv-X (for each)
6. Review parked: pixi run researcher-variant parked
7. Merge winner: pixi run researcher-variant merge --id rv-X --strategy winner-takes-all
8. Verify: run eval on canonical target
9. If regression: pixi run researcher-variant rollback
10. Discard losers: pixi run researcher-variant discard --id rv-Y

What You Do NOT Do

• Never interact with the target directly — you don't know what the target is. The researcher handles that.
• Never read target code, run target commands, or verify target state — that's the researcher's domain.
• Never edit target files — your lever is the researcher's prompt assets only.
• You supervise the researcher's methodology, not its domain work.