rumil

Iterate on versus's pipeline — review recent (or freshly-fired) completion / judging runs, identify concrete improvement opportunities, test them via /rumil-forks, and consolidate findings into a ranked punch list. Spawns parallel agents for trace investigation and fork experiments so wall-clock stays low. Use when the user wants to "review the recent versus runs and find improvements," "play with these traces," "see what we could do better in d&e / two_phase," or anytime after a versus generation/judging batch when it's worth turning the runs into actionable code/prompt changes. Default scope is one forethought essay × Sonnet × budget 4, but every input is overridable.

Background knowledge for working with the rumil research workspace from Claude Code. Explains the two-lane provenance model (rumil-mediated vs cc-mediated), when to use which rumil-* skill, how workspace/session state works, and the visibility/attribution story. Auto-load this whenever the user asks about rumil, wants to inspect research, dispatch a call, discuss a question, apply moves, or review a trace.

Run orchestrator-driven essay completions for the versus pairwise-eval pipeline. Per essay × prefix × model × workflow, fires a rumil orchestrator (TwoPhase, DraftAndEdit) and writes the resulting continuation as a `versus_texts` row tagged `kind="completion"`, `source_id="orch:<workflow>:<model>:c<hash8>"` — pickable as a contestant by `rumil-versus-judge` afterwards. Use when the user wants to A/B orchestrator-produced continuations against single-shot model continuations or against the human baseline. Both `--orch two_phase` and `--orch draft_and_edit` are usable. For single-shot completions (no orch), use `rumil-versus-generate`.

Run pairwise judgments on versus essay-continuation pairs. Default mode is the unified blind path (single-turn LLM call, no tools, no DB) — claude-* models route direct to Anthropic, others through OpenRouter. --variant orch fires a full TwoPhaseOrchestrator run per pair. Use when the user wants to measure how rumil discriminates on pairs with known ground truth (human continuation vs. model continuations), compare blind judges against workspace-aware ones, or top up pending judgments after adding new dimensions or models.

Load a rumil run into the conversation by run ID. Shows a tree of every call in the run (id, type, status, cost, event summary) grouped under the run. Use when you have a trace URL like /traces/<run_id> and want to see the whole run, not a single call. For one call's verbatim trace use rumil-trace instead. Pass --compare <other_run_id> to A/B two runs side-by-side.

Dump a rumil call's full execution trace — trace events plus every LLM exchange verbatim (system prompt, user message, response, tool calls). Use to inspect what happened inside a specific call, debug model confusion, or review a run for quality. Takes a short (8-char) or full call ID.

Run versus's generation scripts — completions and paraphrases — via versus/scripts/run_{completions,paraphrases}.py. Each model in the config routes to its provider (OpenRouter for non-claude ids today; claude-* may route direct to Anthropic in future iterations). Handles the uv --with dance and --config / path anchoring. Default scope is the canonical active essay set; pass --include-stale to widen. Use when the user wants to produce or top up versus completion/paraphrase rows. **Judgments are not in this skill** — the unified judge entry point lives in rumil-versus-judge.

Edit and re-fire a captured LLM exchange to see how the model would respond under different conditions — tweaked system prompt, edited message stack, different tool list, different model, different temperature, adaptive thinking on/off (Opus 4.7/4.6, Sonnet 4.6). Side-effect-free (no tool execution, no workspace mutation, no trace recording). Use when the user wants to probe model behavior on an existing exchange — "what if the system prompt said X", "would the model still call that tool if Y was different", "is this exchange stable across samples", "try sampling at higher temp", "what if thinking was off". Admin-only feature; results persist to exchange_forks for side-by-side comparison in the UI. Subcommands: show / fire / list. Pair with /rumil-trace to find exchange_ids.

Fire a Google Deep Research run and (on resume) save the synthesis as a Source page in the active rumil workspace. Default mode is fire-and-forget — it starts the run in background and returns the interaction id immediately, so you can keep workshopping in chat while it runs. Resume the same id later to poll to completion and create the Source page. Use whenever the user wants deep web research on a topic or question and is fine with a several-minute wait. Pair with /rumil-ingest afterwards if you want the synthesis extracted into considerations on a question.

Fire the rumil orchestrator against an existing question — a full multi-call research loop with a budget. This is the CC-initiated equivalent of `main.py --continue <qid> --budget N`. Use when the user wants real research done on a question, not just a single call. For one targeted call, use /rumil-dispatch instead. Budget defaults to 10; since that's not cheap, confirm with the user before firing if they didn't specify. Trigger when the user says things like "investigate this more", "run some research on this", "give Q# N calls of budget", or right after /rumil-ask when they want to immediately start investigating.

Scan recent rumil calls for signs of model confusion. Default mode uses fast, free heuristics (error events, non-complete status, exchange errors, cost outliers, thin output). Pass --deep for an LLM-based structured verdict on the top heuristic candidates, cached in scan log. Pass --structural <question_id> for graph health checks on a question's subtree (barren questions, orphans, load-bearing fragile claims). Use when the user wants to review recent runs for quality, triage a batch of calls, or find specific traces worth inspecting.

Add a new evaluation agent that assesses a specific quality dimension of research runs. Use this skill when the user describes a type of issue, failure mode, or quality dimension they want an evaluation agent to check for — e.g. "I want to evaluate whether runs are citing sources properly" or "add an eval agent for calibration quality". Also trigger when the user wants to modify the set of evaluation dimensions or asks how to add a new eval agent.

Fire one rumil call (find_considerations, assess, scout-*, web_research, create_view) as Claude Code. This is the "rumil-mediated" lane — a normal rumil call with its usual context-building, prompts, and tools. Use when the user wants to investigate a question deeper, assess a judgement, run a specific scout, or produce a distilled view summary. The run is tagged with origin=claude-code in the trace.

Add a new research question to the active rumil workspace. Creates the question page via the cc-mediated envelope lane but does NOT run any research calls — chain with /rumil-orchestrate or /rumil-dispatch afterward to investigate. Use whenever the user wants to pose a new root question or add a sub-question (either explicitly via /rumil-ask or mid-conversation when they say things like "add a question about X" or "let's track a new question for X"). Pass the headline as the argument; optionally --parent <qid> for a subquestion, --abstract/--content for more detail.

Commit a source to the workspace and run extraction calls against a question. Creates a Source page (from a file/URL) or reuses an existing one, then runs ingest rounds that turn the source into considerations on the target question. This is the mutating counterpart to rumil-read — use it when the user is ready to spend budget extracting considerations from a source. Tagged origin=claude-code in the run's config. If the user only wants to look at a source or stash it without extraction, reach for rumil-read or rumil-read --save instead.

List root questions in the active rumil workspace. Shows short ID, date, and headline for each. Use this as a lightweight overview before diving into a specific question with rumil-show.

Inspect a single rumil page by short ID — full content, provenance (which call created it), epistemic scores, superseded state, and all incoming/outgoing links with target headlines resolved. Use whenever you see a non-question page's short ID (a claim, judgement, concept, view, view_item, view_meta, wiki page, source) in a trace, punch list, or conversation and want to read its actual content and how it connects. Complements rumil-show which is question-specific.

Fetch a file or URL into the conversation — the view-only lane for sources. Default mode is pure fetch-and-print with no DB and no LLM, using the same scraping/extraction pipeline that rumil-ingest would use so you can preview exactly what the ingest step would see. Flags layer more behavior — --summary for an LLM headline, --save to persist as a Source page in the workspace. Use whenever the user shares a URL, PDF, or file they want Claude to look at, especially when it's not yet clear whether the source should become a workspace artifact.

Structured review of all research on a rumil question. Loads the subtree, all recent calls targeting the question (with brief traces), and any existing confusion-scan verdicts, then asks Claude to produce a ranked punch list of problems worth fixing — with per-item attribution and a suggested action (dispatch, apply_move, edit_prompt, or ignore). Use when the user wants to audit a question's research, not just look at individual calls.

Embedding-based search over the active rumil workspace. Takes a free-text query and returns the most semantically similar pages (questions, claims, judgements, concepts). Default mode is a fast top-N lookup with similarity scores — use proactively mid-conversation to check what the workspace already knows about a topic, find related research, or check for duplicates before creating a new question. Pass --full for the multi-tier context builder when you want a deeper rendered read.

Show a rumil question — its research subtree, embedding-based neighbors across the workspace, and recent calls that have run against it. Takes a short (8-char) or full question ID. Use whenever the user asks to view, inspect, or look at a question.

Show, list, or set the active rumil workspace for this Claude Code session. Workspace is rumil's scoping primitive — every other rumil-* skill reads it. Use with no args to list; pass a name to switch.

Guided cleanup of research on a rumil question. Two modes — (a) pipeline mode wraps rumil's existing grounding/feedback clean pipelines given an evaluate call id; (b) interactive mode (default) walks a punch list conversationally, proposes each mutation, talks it through with the user, and applies accreting-only moves via the chat envelope after explicit consent. Use when rumil-review has produced a punch list and the user wants to act on it safely.