openclaw - 💡(How to fix) Fix Context budget: framework + tool schemas add ~10K tokens to every prompt [1 participants]

Summary

Framework instructions and tool JSON schemas add ~10K+ tokens of fixed overhead to every single API call, regardless of conversation complexity. For a typical 31K token prompt, this is one-third of total spend — and it's the same cost whether the user sends "hello" or a complex multi-step request.

Breakdown

Full token accounting from a production Slack conversation (Claude Sonnet 4.6, 200K context):

The non-controllable overhead is ~9,573 tokens (30%) — tool schemas + framework text + envelope. This is baked into every request.

Cost Impact

At Anthropic's Sonnet 4.6 pricing ($3/M input tokens):

• ~10K overhead × every message × every agent adds up fast
• For an agent handling 200 messages/day across channels: ~$6/day just in framework overhead
• The overhead is amplified by prompt caching misses (cache write is 1.25x input rate)

Requests

1. promptMode as a user-facing config option

The code already supports PromptMode: "full" | "minimal" | "none" in buildAgentSystemPrompt(), and resolvePromptModeForSession() already handles all three modes. But it's hardcoded:

// pi-embedded-BaSvmUpW.js:173561
function resolvePromptModeForSession(sessionKey) {
    if (!sessionKey) return "full";
    return isSubagentSessionKey(sessionKey) || isCronSessionKey(sessionKey) ? "minimal" : "full";
}

Request: expose promptMode as a per-agent config option:

{
  "agents": {
    "defaults": {
      "promptMode": "none"
    }
  }
}

This lets advanced users who have comprehensive SOUL.md files opt out of the framework sections entirely. The "none" path already works — it returns "You are a personal assistant running inside OpenClaw." and lets the user's own files handle everything.

2. Tool schema optimization

29 tools × ~250 tokens each = ~7,250 tokens of JSON Schema sent on every request. Options:

• Lazy tool loading: Only include tool schemas for tools the agent has actually used in the current session, plus a "discover more tools" meta-tool
• Schema compression: Tool descriptions are verbose. Many optional parameters could use shorter descriptions or omit descriptions for self-evident fields
• Tool profiles with schema impact visibility: Show token cost per tool in openclaw tools list so users can make informed decisions about which tools to enable

3. Framework text audit

The 2,323 tokens of framework text include sections that could be trimmed or made conditional:

• ## Reply Tags (76 tokens) — only needed for channels that support reply tags
• ## Silent Replies (detailed formatting rules) — could be a single line
• ## OpenClaw CLI Quick Reference — rarely needed in conversation
• ## OpenClaw Self-Update — only needed when user asks for updates
• ## Heartbeats (detailed rules) — only needed for heartbeat runs
• ## Documentation (links) — static, could be in SOUL.md instead

A "lean" prompt mode that keeps only Safety + Tooling + Workspace + Runtime would save ~800-1,000 tokens.

Environment

• OpenClaw version: 2026.2.22-2
• Platform: macOS (darwin arm64)
• Primary model: Claude Sonnet 4.6 (200K context)
• Channels: Slack (socket mode), OpenWebUI, direct API
• Context engine: custom plugin (nightshift-context) with reranker-based compression
• Memory: memory-lancedb with auto-recall (30 memories) + conversation memory (9 turns)