Selig
4c966a3ad2
6 custom skills (assign-task, dispatch-webhook, daily-briefing, task-capture, qmd-brain, tts-voice) with technical documentation. Compatible with Claude Code, OpenClaw, Codex CLI, and OpenCode.
49 lines
4.9 KiB
JSON
Executable File
49 lines
4.9 KiB
JSON
Executable File
{
|
||
"title": "Token use & costs",
|
||
"content": "OpenClaw tracks **tokens**, not characters. Tokens are model-specific, but most\nOpenAI-style models average \\~4 characters per token for English text.\n\n## How the system prompt is built\n\nOpenClaw assembles its own system prompt on every run. It includes:\n\n* Tool list + short descriptions\n* Skills list (only metadata; instructions are loaded on demand with `read`)\n* Self-update instructions\n* Workspace + bootstrap files (`AGENTS.md`, `SOUL.md`, `TOOLS.md`, `IDENTITY.md`, `USER.md`, `HEARTBEAT.md`, `BOOTSTRAP.md` when new). Large files are truncated by `agents.defaults.bootstrapMaxChars` (default: 20000).\n* Time (UTC + user timezone)\n* Reply tags + heartbeat behavior\n* Runtime metadata (host/OS/model/thinking)\n\nSee the full breakdown in [System Prompt](/concepts/system-prompt).\n\n## What counts in the context window\n\nEverything the model receives counts toward the context limit:\n\n* System prompt (all sections listed above)\n* Conversation history (user + assistant messages)\n* Tool calls and tool results\n* Attachments/transcripts (images, audio, files)\n* Compaction summaries and pruning artifacts\n* Provider wrappers or safety headers (not visible, but still counted)\n\nFor a practical breakdown (per injected file, tools, skills, and system prompt size), use `/context list` or `/context detail`. See [Context](/concepts/context).\n\n## How to see current token usage\n\n* `/status` → **emoji‑rich status card** with the session model, context usage,\n last response input/output tokens, and **estimated cost** (API key only).\n* `/usage off|tokens|full` → appends a **per-response usage footer** to every reply.\n * Persists per session (stored as `responseUsage`).\n * OAuth auth **hides cost** (tokens only).\n* `/usage cost` → shows a local cost summary from OpenClaw session logs.\n\n* **TUI/Web TUI:** `/status` + `/usage` are supported.\n* **CLI:** `openclaw status --usage` and `openclaw channels list` show\n provider quota windows (not per-response costs).\n\n## Cost estimation (when shown)\n\nCosts are estimated from your model pricing config:\n\nThese are **USD per 1M tokens** for `input`, `output`, `cacheRead`, and\n`cacheWrite`. If pricing is missing, OpenClaw shows tokens only. OAuth tokens\nnever show dollar cost.\n\n## Cache TTL and pruning impact\n\nProvider prompt caching only applies within the cache TTL window. OpenClaw can\noptionally run **cache-ttl pruning**: it prunes the session once the cache TTL\nhas expired, then resets the cache window so subsequent requests can re-use the\nfreshly cached context instead of re-caching the full history. This keeps cache\nwrite costs lower when a session goes idle past the TTL.\n\nConfigure it in [Gateway configuration](/gateway/configuration) and see the\nbehavior details in [Session pruning](/concepts/session-pruning).\n\nHeartbeat can keep the cache **warm** across idle gaps. If your model cache TTL\nis `1h`, setting the heartbeat interval just under that (e.g., `55m`) can avoid\nre-caching the full prompt, reducing cache write costs.\n\nFor Anthropic API pricing, cache reads are significantly cheaper than input\ntokens, while cache writes are billed at a higher multiplier. See Anthropic’s\nprompt caching pricing for the latest rates and TTL multipliers:\n[https://docs.anthropic.com/docs/build-with-claude/prompt-caching](https://docs.anthropic.com/docs/build-with-claude/prompt-caching)\n\n### Example: keep 1h cache warm with heartbeat\n\n## Tips for reducing token pressure\n\n* Use `/compact` to summarize long sessions.\n* Trim large tool outputs in your workflows.\n* Keep skill descriptions short (skill list is injected into the prompt).\n* Prefer smaller models for verbose, exploratory work.\n\nSee [Skills](/tools/skills) for the exact skill list overhead formula.",
|
||
"code_samples": [
|
||
{
|
||
"code": "models.providers.<provider>.models[].cost",
|
||
"language": "unknown"
|
||
}
|
||
],
|
||
"headings": [
|
||
{
|
||
"level": "h2",
|
||
"text": "How the system prompt is built",
|
||
"id": "how-the-system-prompt-is-built"
|
||
},
|
||
{
|
||
"level": "h2",
|
||
"text": "What counts in the context window",
|
||
"id": "what-counts-in-the-context-window"
|
||
},
|
||
{
|
||
"level": "h2",
|
||
"text": "How to see current token usage",
|
||
"id": "how-to-see-current-token-usage"
|
||
},
|
||
{
|
||
"level": "h2",
|
||
"text": "Cost estimation (when shown)",
|
||
"id": "cost-estimation-(when-shown)"
|
||
},
|
||
{
|
||
"level": "h2",
|
||
"text": "Cache TTL and pruning impact",
|
||
"id": "cache-ttl-and-pruning-impact"
|
||
},
|
||
{
|
||
"level": "h3",
|
||
"text": "Example: keep 1h cache warm with heartbeat",
|
||
"id": "example:-keep-1h-cache-warm-with-heartbeat"
|
||
},
|
||
{
|
||
"level": "h2",
|
||
"text": "Tips for reducing token pressure",
|
||
"id": "tips-for-reducing-token-pressure"
|
||
}
|
||
],
|
||
"url": "llms-txt#token-use-&-costs",
|
||
"links": []
|
||
} |