Token Counter

解析本地 OpenClaw 会话转录、会话元数据与 cron 定义，按类别、客户端、模型和工具归因 token 用量，支持日报/周报、单会话下钻、JSON 导出与每日快照保存。工具归因与客户端检测为启发式/规则式。

mkhaytman87

热度 85.59

最近更新 2026-05-11 16:28:05

分类：数据依赖：代码执行 / 读写文件应用场景：数据处理 / 报告撰写 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add token-counter

Metadata

name: token-counter
description: Track and analyze OpenClaw token usage across main, cron, and sub-agent sessions with category, client, model, and tool attribution. Use when the user asks where tokens are being spent, wants daily/weekly token reports, needs per-session drilldowns, or is planning token-cost optimizations and needs evidence from transcript data.

Token Counter

Overview

Use this skill to produce token usage reports from local OpenClaw data. It parses session transcripts (.jsonl), session metadata, and cron definitions, then reports usage by category, client, tool, model, and top token consumers.

Quick Start

Run:

$OPENCLAW_SKILLS_DIR/token-counter/scripts/token-counter --period 7d

Common Commands

Basic report:

$OPENCLAW_SKILLS_DIR/token-counter/scripts/token-counter --period 7d

Focus on selected breakdowns:

$OPENCLAW_SKILLS_DIR/token-counter/scripts/token-counter \
  --period 1d \
  --breakdown tools,category,client

Analyze one session:

$OPENCLAW_SKILLS_DIR/token-counter/scripts/token-counter \
  --session agent:main:cron:d3d76f7a-7090-41c3-bb19-e2324093f9b1

Export JSON:

$OPENCLAW_SKILLS_DIR/token-counter/scripts/token-counter \
  --period 30d \
  --format json \
  --output $OPENCLAW_WORKSPACE/token-usage/token-usage-30d.json

Persist daily snapshot:

$OPENCLAW_SKILLS_DIR/token-counter/scripts/token-counter \
  --period 1d \
  --save

This writes JSON to: $OPENCLAW_WORKSPACE/token-usage/daily/YYYY-MM-DD.json

Defaults and Data Sources

Sessions index: $OPENCLAW_DATA_DIR/agents/main/sessions/sessions.json
Session transcripts: $OPENCLAW_DATA_DIR/agents/main/sessions/*.jsonl
Cron definitions: $OPENCLAW_DATA_DIR/cron/jobs.json

The parser reads assistant usage fields for token counts and uses tool-call records for attribution.

Notes on Attribution

Tool token attribution is heuristic: assistant-message tokens are split across tool calls in that message.
Session totalTokens may come from either session index metadata or transcript usage sums (max is used).
Client detection is rules-based (personal, bonsai, mixed, unknown) using path/domain/email markers.

Validation

Run:

python3 $OPENCLAW_SKILLS_DIR/skill-creator/scripts/quick_validate.py \
  $OPENCLAW_SKILLS_DIR/token-counter

References

See:

references/classification-rules.md for category/client detection logic and keyword mapping.