rumil-clean

rumil-clean

Turns a review into applied fixes, safely. There's a fast lane and a careful lane.

Two modes

Mode A — pipeline (rumil-internal)

When you already have a completed evaluate call on a question and want rumil's own clean pipelines to run against its evaluation text, this mode wraps them:

PYTHONPATH=.claude/lib uv run python -m rumil_skills.run_clean_pipeline \
    grounding <eval_call_id>

PYTHONPATH=.claude/lib uv run python -m rumil_skills.run_clean_pipeline \
    feedback <eval_call_id>

grounding runs clean.grounding.run_grounding_feedback — creates source pages and grounds claims in real evidence.

feedback runs clean.feedback.run_feedback_update — applies proposed changes from a feedback-style evaluation (new claims, new investigations, link fixes).

Both are rumil-mediated: the actual mutations are chosen by a rumil-internal LLM pipeline, not by you. You're just the trigger. The run is tagged origin=claude-code skill=rumil-clean pipeline=... and gets its own trace URL. Use this when you trust the rumil clean pipeline to do the right thing and just want it to run.

Prereq: the user must have already run /rumil-dispatch evaluate <question_id> to produce an EVALUATE call with evaluation text in its review_json.

Mode B — interactive (cc-mediated, accreting-only)

This is the conversational lane. The user typically lands here after running /rumil-review <question_id> and getting back a punch list. Your job is to walk the punch list one item at a time, proposing each mutation, talking it through, and applying it after explicit user consent. Strict rules:

1. Accreting-only. Every applied move goes through apply_move.py --accreting-only. The allowlist is: CREATE_* moves, non-destructive LINK_* moves, FLAG_FUNNINESS, REPORT_DUPLICATE, PROPOSE_CONCEPT, LOAD_PAGE. Never propose or apply REMOVE_LINK, UPDATE_EPISTEMIC, CHANGE_LINK_ROLE, or PROMOTE_CONCEPT from this skill — they modify or migrate existing state. If a punch-list item requires one of those, say so and suggest the user use the destructive tool manually or run the rumil-mediated clean pipeline instead.

2. Talk through, then act. Never apply a move without stating what you're about to do and giving the user a chance to object. Format:

   "Proposing: FLAG_FUNNINESS on be6d1a1d (the AI-governance-determines-space-allocation claim) — the headline frames it as a claim but the body reads as a meta-reframe. OK to apply?"

3. Dry-run first for anything non-trivial. For moves with complex payloads (links with reasoning, claims with long content), run apply_move.py --dry-run first so you can show the user what the validated payload looks like before committing.

4. One move per turn. Don't batch five mutations in a single response. Each move should be its own little propose-apply cycle so the user can stop or redirect anytime.

5. Always gloss IDs. When you cite a page or call, include a 3-8 word summary in parens: be6d1a1d (the claim about power-driven allocation). Never drop bare hex.

Invocation

There's no single ! block here because the skill dispatches based on args. Parse $ARGUMENTS yourself:

• If the first arg is --pipeline grounding|feedback followed by an eval call id → run Mode A via run_clean_pipeline.
• If the first arg is a question ID → run Mode B: first load the question context via gather_review_context, then walk whatever punch list is already in the conversation (or ask the user for one).

If no punch list exists yet, suggest the user run /rumil-review <qid> first, unless they explicitly want you to produce one here.

What to show the user after each applied move

Just the one-line skill output:

⚙ cc-mediated move: <type>
• created page <short_id>
<one-line headline of result.message>

Plus (once per session) the envelope trace URL, so they can open the rumil frontend to see the whole envelope Call and the moves flowing into it.

Escape hatches

• If the user changes their mind mid-session, stop applying and summarize what's been done + what's still pending from the punch list.
• If a proposed mutation would require destructive capability (edit an existing claim, remove a link, change a link role), say so and skip it — do not escalate to non-accreting moves. The user can handle those manually.
• If the user wants to switch to Mode A, just exit the conversational loop and fire run_clean_pipeline directly.