在沟通的时候，全部出的代码

[Apache012]

Pre-checks

• I have searched existing issues and this is not a duplicate.

Deployment Method

Source (setup.sh)

Steps to Reproduce

Expected vs Actual Behavior

期望正常回复

Logs / Screenshots

No response

[Clawiee]

Hi @Apache012,

Thank you for reporting this issue! 🙏

Friendly reminder: Could you please use English for issue titles and descriptions? This helps our international team members understand and resolve issues faster.

About your issue: I understand you're experiencing an problem where the system outputs code instead of normal responses during conversations. To help us investigate this bug, could you please provide:

1. Which tool or skill was being called when this happened?
2. Can you describe the conversation context? (What did you ask, what did you expect?)
3. Does this happen consistently or only occasionally?
4. Any error messages in the browser console or backend logs?

The screenshot you attached will definitely help our team debug this. Looking forward to your reply!

---

Translation of your issue title for reference: "When communicating, all output is code"

[Apache012]

请你给这些邮箱发送邮件吧，邮件的编码为 PAIN_POINT_001 这个，每一个邮箱发送一封
12:28
市
[LLM Error] HTTP 400: {"error":{"code":"InvalidParameter","message":"Invalid messages at 1:\n{\'role\': \'system\', \'content\': \'You are 市场运营部, an enterprise digital employee.\\n\\n## Role\\nFocuses on market research, industry analysis, competitive intelligence tracking, and trend insights\\n\\n\\n## Workspace \u0026 Tools\\n\\nYou have a dedicated workspace with this structure:\\n - focus.md → Your focus items — what you are currently tracking (ALWAYS read this first when waking u
很多时候还会出现这种错误.....

[yaojin3616]

Fixed in commit 4921623 (PR #416): Prevent tool call message pairs from being broken down.

Root cause: When the LLM was generating tool calls, the message history management could produce malformed conversation state where tool result messages appeared without their preceding tool call context, causing the LLM to output raw code or error responses instead of natural language.

Fix: Tool call/result message pairs are now kept atomic during history management.