菜单
Use when the user invokes `/token-report` or asks for an analytical summary of which compression plugin (rtk / headroom / caveman) is effective for which workload. Aggregates token-meter JSONL, `rtk gain`, and the active Claude Code session transcript, then produces recommendations.
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
Use when the user invokes `/token-clear` or asks to delete old token-meter JSONL logs by retention period. Confirms before destructive deletion.
Use when the user invokes `/token-stats` or asks for a token volume / compression-effect report across recent days. Aggregates JSONL records by tool and plugin into a single page summary.
Use when the user invokes `/token-plugins` or asks to list, install, enable, or disable a compression plugin (rtk / caveman / headroom). Wraps each plugin's enable/disable mechanism behind a unified interface.
Use when the user invokes `/token-measure` or asks to enable/disable the token-meter measurement system, list per-tool measurement state, or override a single tool. Provides ON/OFF control and JSONL-backed activity listing.
Orchestrate parallel execution of multiple tasks via cmux workspaces. Each task gets its own git worktree + Claude Code session. The parent session acts as orchestrator, monitoring all child sessions. Dynamically discovers available agent types from .claude/agents/ and passes the list to child sessions, which select the appropriate agent themselves. Supports three layout modes: workspace (default, separate sidebar entries), split (panes within current workspace), and claude-teams (native Agent Teams via cmux claude-teams). Dispatches immediately without parent-side planning — each child handles its own brainstorming/planning in parallel. Use when: "parallel execution", "team dispatch", "run these at once", "run these in parallel", "dispatch tasks", "execute these simultaneously", or when 2+ independent tasks need concurrent execution.
Bring up an isolated development stack per git worktree. Reserves a port slot from a shared registry, writes .env.dispatch, then starts all services declared in .dev-up.yaml (docker-compose stacks and/or direct commands like `go run`, `pnpm dev`). Spawned commands are tracked with PID/PGID for clean teardown. Use when verifying the running app inside a worktree spawned by cmux-team-dispatch-task, or whenever you need parallel stacks without port collisions. Run the `setup` subcommand once per new project to generate .dev-up.yaml.
| name | token-report |
| description | Use when the user invokes `/token-report` or asks for an analytical summary of which compression plugin (rtk / headroom / caveman) is effective for which workload. Aggregates token-meter JSONL, `rtk gain`, and the active Claude Code session transcript, then produces recommendations. |
token-meter のログ + rtk gain + transcript を横断集計して、 「どんな処理にどのプラグインを使うと効くか」の質的レポートを生成する。
| 引数 | 動作 |
|---|---|
--since today|1d|7d|30d|all | 集計期間。デフォルトは 7d |
bun run "$CLAUDE_PROJECT_DIR/apps/token-meter/scripts/token-report.ts" --since 7d
tool_summary: tool 別 calls / input / output / ratio (output 降順)rtk.jsonl: token-meter 計測の rtk savingsrtk.gain: rtk gain --format json の global summary (Claude 外の rtk 使用も含む authoritative 値)headroom: post.compress 集計 (input_tokens - output_tokens = saved)caveman: prompt mode のため計測不可。transcript の assistant output_tokens × 0.65 を「常時 full モードだった場合の上限値」として推定top_wasteful_tool_calls: output_tokens シェア上位 5 件# Token 効果分析レポート (since=<期間>)
## サマリー
- 観測 tool 数: ...
- 観測 plugin: rtk (saved=X), headroom (saved=Y), caveman (推定 ≤Z)
## Plugin 別の効きと使いどころ
### rtk (Bash 出力圧縮)
- 計測: X tokens saved / Y calls
- 効いた典型: ... ← rtk.gain や tool_summary から具体的に
- 効かなかった典型: ...
- 推奨: 「<こういう Bash> のときに使う」
### headroom (大きな text の圧縮)
- 計測: ...
- 効いた典型: 入力 >5K で ratio が小さい call
- 効かなかった典型: router:noop で返るケース (短文 / 保護パターン)
- 推奨: ...
### caveman (応答テキスト短縮)
- 計測: 不可 (prompt mode)
- 推定上限: ~Z tokens (transcript 集計 × 0.65 / sonnet-4 ベンチマーク)
- 使いどころ: 長い説明応答が多い session、自由記述レポート系
- 注意: 推定値は「常時 full モード」前提の上限値。実値は `/caveman:caveman-stats` 参照
## 浪費が大きい上位 tool
| tool | output_tokens | share |
| ... | ... | ... |
コメント: rtk / headroom で削減余地があるものは具体的に提案する。
## 推奨アクション
- 「X tool の出力には rtk を当てる」「caveman を Y 系の応答で ON にする」など、
数字に裏打ちされた actionable な箇条書きを 3-5 個。
rtk.gain は Claude 外の rtk 使用も合算する。rtk.jsonl との差を「Claude セッション外」と説明できる。transcript_path が拾った session が今動かしている session と違うこともある。current_mode が null なら caveman は今使われていない。