Problem

Every agent request incurs ~6-7 seconds of latency before the model even starts generating. The bottleneck is the bundle-tools phase which loads all MCP servers and skills upfront.

Trace evidence

[trace:embedded-run] prep stages: totalMs=5889 stages=bundle-tools:5792ms@5809ms
[trace:embedded-run] prep stages: totalMs=6020 stages=bundle-tools:5898ms@5915ms
[trace:embedded-run] prep stages: totalMs=6527 stages=bundle-tools:6435ms@6449ms
[trace:embedded-run] prep stages: totalMs=7207 stages=bundle-tools:7132ms@7207ms

In all cases, bundle-tools accounts for 96-99% of total prep time.

Environment

• OpenClaw v2026.6.8
• Node v26.0.0
• macOS (arm64)
• MCP servers: Notion, Z.ai, Web Reader, Web Search Prime
• ~30+ skills loaded

Expected behavior

Tools should be lazily loaded on-demand, or at least cached/warm-pooled across requests. Simple text replies should not require 6+ seconds of tool initialization.

Impact

• Every user message — including simple chat — has 6-7s added latency
• Chatty sessions feel noticeably slow compared to similar AI assistants
• Streaming partially masks this but the delay is still present before first token

Suggested fix

1. Lazy tool loading: only initialize tools when the model decides to call them
2. Tool cache/warm pool: keep initialized tool connections alive across requests in the same session
3. Fast path: if the model has no tool calls (simple text reply), skip tool initialization entirely