菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
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.
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.
| name | rumil-trace |
| description | 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. |
| allowed-tools | Bash |
| argument-hint | <call_id> [--brief] [--only llm_exchange] [--last-n N] [--system-once] [--user-only N] [--response-only N] |
Loads a call's full trace into the Claude Code conversation. By default every LLM exchange is printed verbatim so you (and Claude) can see the model's actual words — summaries lose the signal needed to spot confusion.
Use filters for large traces:
--brief shortens bodies and hides the system prompt--only <event> shows only events of one type (e.g. llm_exchange, error)--last-n N trims to the final N exchanges--no-exchanges shows just the event timeline--system-once prints the system prompt only on its first occurrence; subsequent
identical prompts are replaced with (system prompt unchanged from exchange N).
Cuts ~60-70% off multi-round-call output where the system prompt repeats verbatim.--user-only N for each exchange, render only the last N chars of the user
message (prefixed with ... (truncated, M chars total) if trimmed).--response-only N same shape, applied to the assistant response.These flags combine — e.g. --system-once --user-only 2000 --response-only 2000
gives a compact read of a long multi-round call with the high-signal tail of
each exchange and no system-prompt repetition.
setopt no_glob 2>/dev/null; set -f; PYTHONPATH=.claude/lib uv run python -m rumil_skills.trace $ARGUMENTS