openclaw - 💡(How to fix) Fix Feature Request: Expose user message in agent:bootstrap hook context for query-conditioned context compression [1 participants]

Summary

The agent:bootstrap hook event currently provides workspaceDir, cfg, sessionKey, and bootstrapFiles in its context, but does not include the incoming user message. Adding the user message (or at minimum a query/topic string) would enable powerful context compression techniques that can reduce injected workspace token counts by 90-97%.

Motivation

Workspace bootstrap files (AGENTS.md, SOUL.md, MEMORY.md, TOOLS.md, etc.) are injected into the context window on every inference call. For a workspace with substantial memory and tool notes, this can easily reach 35KB+ of static content per turn.

Tools like SMELT (Schema-Aware Markdown Compilation for Efficient Local Token Inference) can perform query-conditioned selective emission — scoring workspace content against the actual user query and returning only relevant sections. In testing on a production OpenClaw workspace:

This translates directly to:

• Lower API costs (fewer input tokens per call)
• Faster TTFT (less context to process)
• Better model attention (less irrelevant context to wade through)
• Reduced compaction frequency (smaller system prompts = more room for conversation)

Current Limitation

The agent:bootstrap event context includes:

context: {
  bootstrapFiles: WorkspaceBootstrapFile[],
  workspaceDir: string,
  cfg: OpenClawConfig,
  sessionKey: string,
}

Without the user message, hooks can only apply static compression (Layer 2 semantic compilation), which yields only ~3% savings. The big wins (90%+) come from Layer 4 query-conditioned retrieval, which requires knowing what the user is asking about.

Proposed Change

Add the user message text to the agent:bootstrap event context:

context: {
  bootstrapFiles: WorkspaceBootstrapFile[],
  workspaceDir: string,
  cfg: OpenClawConfig,
  sessionKey: string,
  userMessage?: string,  // The incoming user message that triggered this agent run
}

This would allow bootstrap hooks to:

1. Run query-conditioned retrieval on workspace files
2. Replace bootstrapFiles[].content with compressed versions
3. Dramatically reduce context size while preserving relevant information

Implementation Notes

• The field should be optional (undefined for heartbeats, gateway startup, etc.)
• For the first message in a session, this is straightforward
• For subsequent messages, the bootstrap files are already in context from the system prompt — but if OpenClaw re-injects them per turn, the same optimization applies
• Privacy consideration: the user message is already visible to the model; exposing it to a local hook running in the same Gateway process adds no new attack surface

Alternatives Considered

• External proxy between OpenClaw and the API provider: Works but fragile and adds latency
• Static compression only (no query conditioning): Only ~3% savings, not impactful
• Moving workspace files to tool-based retrieval: Would work but loses the "always available" property of bootstrap context

References

• SMELT paper
• SMELT repo
• OpenClaw hooks docs
• OpenClaw system prompt docs