[Bug] 客户端是Openclaw时报错

[deepsourceonline]

版本 / Version

v1.5.0

部署方式 / Deployment

自建服务器 / Self-hosted

问题描述 / What happened?

19:34:30 WARN Cascade error step 19:34:30 WARN Account XXXXXXXX@lanbojini.online failed (model_error) on gpt-5.4-high, trying next 19:34:30 ERROR Stream error after retries: Your request was blocked by our content policy. Please remove sensitive or unsafe content from your prompt, memories, and other settings and try again.

期望响应正常

复现步骤 / Steps to reproduce

使用Openclaw 发起请求，claude-opus-4-6, gpt-5.4-high 都会出现

日志 / Logs

19:34:30
WARN
Cascade error step
19:34:30
WARN
Account XXXXXXXX@lanbojini.online failed (model_error) on gpt-5.4-high, trying next
19:34:30
ERROR
Stream error after retries: Your request was blocked by our content policy. Please remove sensitive or unsafe content from your prompt, memories, and other settings and try again.

模型 / Model

claude-opus-4-6/gpt-5.4-high

环境 / Environment

Ubuntu 服务器部署的Windsurf API

[dwgx]

能贴一下具体报错日志吗 光标题看不出来是什么问题

[deepsourceonline]

19:34:30
WARN
Cascade error step
19:34:30
WARN
Account XXXXXXXX@lanbojini.online failed (model_error) on gpt-5.4-high, trying next
19:34:30
ERROR
Stream error after retries: Your request was blocked by our content policy. Please remove sensitive or unsafe content from your prompt, memories, and other settings and try again.

大佬，上面这个就是我在服务器部署咱们的服务后，客户端用openclaw 访问 claude-opus-4-6, gpt-5.4-high 模型的时候会的报错信息

[deepsourceonline]

这个是咱们平台的截图

[dwgx]

应该是注入太激进了你们可以自己改prompt试试

[dwgx]

查了 content policy blocked 的原因

之前我们注入的工具指令用了 IMPORTANT/CRITICAL/MUST/Do NOT 这种命令式语气 加上负面引用("I don't have access to tools") Windsurf 的内容安全层可能把这当成 prompt injection 攻击

79057dc 把注入语气全部柔化了 去掉了所有激进指令词 改成自然陈述

实测工具调用仍然正常 Read/Bash 都 PASS

不过 content policy 也可能是用户自己的 prompt 触发的 这个不是我们能控制的 如果还出现 看看是不是 prompt 里有敏感内容

update.sh 拉最新试试

[dwgx]

支持了 前端改prompt 你可以自己试试

[deepsourceonline]

好的，感谢大佬，我现在去测试

[deepsourceonline]

大佬，还是有，提示词很简单，就是一个你好，如图

[dwgx]

好的 等你测试结果