让 before_agent_start 的 systemPrompt 真正生效，减少上下文外泄

适用于“插件已返回 `systemPrompt` 但模型仍回显内部指令”场景：把策略指令放进系统提示而非用户前缀，降低小模型把路由信息输出给用户的概率。

GITHUBDiscovered 2026-02-12Author PauloLuan

Prerequisites

You are using plugin hook `before_agent_start` and can inspect behavior with test prompts.
Your stack can deploy a patch or upgrade once fix is merged upstream.

Steps

Reproduce the issue by returning both `prependContext` and `systemPrompt`, then check which one affects final model behavior.
Move sensitive routing directives from user-visible prepend text into system-level prompt path.
Regression-test with smaller models (e.g., Haiku-class) for echo/leakage cases.
Ship patch/upgrade and keep a fallback rule to strip accidental prompt echoes in output filters.

Commands

openclaw gateway status

openclaw gateway restart

openclaw status

Verify

User-facing replies stop echoing internal intent-routing text while hook-driven behavior still takes effect.

Caveats

Behavior differs across model families; every target model should be revalidated（需验证）.
Avoid storing secrets directly in system prompts; use scoped references where possible.

Source attribution

This tip is aggregated from community/public sources and preserved with attribution.