MiniMax LLM Error (2013): Multiple system messages in requests

[DenXjun]

Pre-checks

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

Deployment Method

Source (setup.sh)

Steps to Reproduce

Issue Description

When using the MiniMax LLM, the following error occurs in Agent-to-Agent communication and trigger execution scenarios:

HTTP 400: {"type":"error","error":{"type":"bad_request_error","message":"invalid params, invalid chat setting (2013)","http_code":"400"}}

Affected Features

• Agent-to-Agent communication via Gateway
• Automatic trigger execution

Reproduction Steps

1. Configure an Agent using the MiniMax model
2. Have another Agent send a message to this Agent (via Gateway)
   OR
   Configure an automatic trigger execution
3. Observe the error message in backend.log

Expected vs Actual Behavior

Root Cause

The MiniMax LLM does not support multiple messages with the system role in the message list.

The issue exists in two files:

1. app/api/gateway.py - During Agent-to-Agent communication
2. app/services/trigger_daemon.py - During trigger execution

The code logic in both locations is as follows:

1. Construct a complete context-aware system_prompt
2. Add a system message to the messages list: {"role": "system", "content": system_prompt}
3. Call call_llm(messages)
4. The call_llm function internally adds another system message: api_messages = [LLMMessage(role="system", content=system_prompt)]

Result: The message list contains two system messages, which MiniMax rejects and returns error (2013).

Logs / Screenshots

2026-03-24 21:16:19 | ERROR | app.api.websocket:call_llm:262 - [LLM] LLMError provider=minimax model=MiniMax-M2.5 round=1: HTTP 400: {"type":"error","error":{"type":"bad_request_error","message":"invalid params, invalid chat setting (2013)","http_code":"400"}}

[Clawiee]

Hi @DenXjun,

Thank you for this detailed bug report! This is a very clear analysis of the MiniMax LLM error (2013).

I've reviewed the issue and can confirm:

• Problem: Multiple system messages being sent to MiniMax API, which doesn't support this
• Affected files: app/api/gateway.py and app/services/trigger_daemon.py
• Root cause: Double system message injection (once in the calling code, once inside call_llm())

This is a valid bug that needs to be fixed. I'm notifying our backend team (@姚劲) and product lead (@覃睿) to prioritize this fix.

Next steps:

1. Review the call_llm() function to understand the system message handling
2. Fix the duplicate system message issue in both affected files
3. Test with MiniMax model to ensure the fix works

Thanks again for the thorough investigation and clear reproduction steps! 🙏

/cc @姚劲 @覃睿 @湛雨桐

[wisdomqin]

Hi @DenXjun, thank you for the excellent root-cause analysis!

You are absolutely right — this was a duplicate system message bug. Both gateway.py and trigger_daemon.py were manually prepending a {"role": "system", ...} message before calling call_llm(), which then also adds its own system message internally. This caused MiniMax (and potentially other strict LLM providers) to reject the request with error 2013.

This has already been fixed in commit 1156e1e:

• Removed the manual system message construction from trigger_daemon.py and gateway.py
• call_llm() is now the single source of truth for system prompt injection via build_agent_context()
• The agent-to-agent alert message was preserved as a user-message prefix instead of a second system message

The fix is live on our production environment. If you are self-hosting, please update:

git pull origin main
docker cp backend/app clawith-backend-1:/app/
docker exec clawith-backend-1 find /app -name "__pycache__" -exec rm -rf {} + 2>/dev/null
docker compose restart backend

Thanks again for the detailed investigation — it directly contributed to a cleaner architecture! Closing this issue.