菜单
Read and report token usage stats from Hermes gateway logs. Use when asked "how many tokens", "token breakdown", "context size", "what's using my context", or any question about token/component breakdown in the current session or across recent sessions.
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 dispatching work through the Hermes gateway with /claw, especially when the task may resolve slash commands or hand off into AO worker orchestration.
12-hour autonomous evolution loop — observe, measure, diagnose, fix, repeat every 10min. Drives zero-touch rate up by finding friction and dispatching fixes.
Custom evolve loop for jleechanbrain orchestrator — drains dropped Slack thread backlog via /claw, fixes hermes issues, proposes new work items. Max 50 items, newest-first.
jleechanbrain / Hermes-specific harness — gateway, canary, deploy, lane backlog, staging vs prod. Use with user-scope harness-engineering.
Delegate coding change tasks to Agent-Orchestrator (AO). Triggered by the keyword "agento" anywhere in a message, or by default for any coding task. Never use mctrl unless explicitly requested.
Spawn an Agent-Orchestrator (AO) worker that uses the minimax CLI with the M3 model (via the minimax Anthropic-compatible API). Use when: the task should be dispatched to a worker that uses minimax M3 specifically, when the user says 'use minimax', 'use the M3 model', 'use the minimax worker', 'ao spawn minimax', or 'minimax CLI for AO'. This skill verifies the env (MINIMAX_MODEL, MINIMAX_API_KEY, ANTHROPIC_BASE_URL) end-to-end, then runs `ao spawn --agent minimax`. Verified 2026-06-13 against session ao-6355 / PR #678.
| name | context-token-audit |
| description | Read and report token usage stats from Hermes gateway logs. Use when asked "how many tokens", "token breakdown", "context size", "what's using my context", or any question about token/component breakdown in the current session or across recent sessions. |
Read actual token usage from Hermes gateway logs to report precise breakdown of context components.
~/.smartclaw_prod/logs/gateway.log
The gateway logs token stats with this format:
⏱️ Elapsed: 0.21s Context: 4 msgs, ~14,339 tokens
Also contains per-message headers when sending to the LLM API:
DEBUG anthropic._base_client: Request options: ...'content': None, 'json_data': {'max_tokens': 131072, 'messages': [...]
grep -E "^ ⏱" ~/.smartclaw_prod/logs/gateway.log | tail -10 | grep -oE "Context: [0-9]+ msgs, ~[0-9,]+ tokens"
Returns most recent sessions with msg count and token estimate.
grep -E "^ ⏱" ~/.smartclaw_prod/logs/gateway.log | awk '{print $NF}' | sed 's/[^0-9]//g' | sort -n | uniq -c | sort -rn | head -20
grep -E "^ ⏱" ~/.smartclaw_prod/logs/gateway.log | grep -oE "[0-9]+ msgs, ~[0-9,]+ tokens" | sort -t: -k3 -rn | head -20
When reporting token breakdown, use these known proportions from analysis:
| Component | Approx % | Notes |
|---|---|---|
| Tool definitions | ~45% | 35 tools with full JSON schemas |
| System prompt framework | ~18% | Fixed instruction structure |
| SOUL.md + AGENTS.md | ~25-27% | Combined policy files |
| Skills list (metadata only) | ~6% | 130+ entries, names + descriptions |
| Session context + profiles | ~4% | Slack thread, user profile, memory |
| TOOLS.md | <1% | Local setup notes |
Note: Skill content (SKILL.md files) loads on-demand via skill_view(), not pre-loaded. Only the skill metadata (name + description) is in the baseline context.
When asked "what's in my context" or "token breakdown":
~NN,NNN tokens figure is an estimate from the gateway, not the model's own count