一直报错 [LLM Error] HTTP 400: {"error":{"message":"Expecting ',' delimiter: line 1 column 43 (char 42)","type":"BadRequestError","param":null,"code":400}}

[wangwl8]

Pre-checks

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

Deployment Method

Docker

Steps to Reproduce

我使用：report-ppt-generator-pro 等多个生成PPT的skill
大模型是：minimax-m2.5和kimi-k2.5都试过
输入文字不超过200字，目标是生成一个html大纲。
然后一直报错：
一直报错 [LLM Error] HTTP 400: {"error":{"message":"Expecting ',' delimiter: line 1 column 43 (char 42)","type":"BadRequestError","param":null,"code":400}}

Expected vs Actual Behavior

一直报错 [LLM Error] HTTP 400: {"error":{"message":"Expecting ',' delimiter: line 1 column 43 (char 42)","type":"BadRequestError","param":null,"code":400}}

Logs / Screenshots

No response

[Clawiee]

Issue Analysis

Thank you for reporting this issue. The error HTTP 400: {"error":{"message":"Expecting ',' delimiter: line 1 column 43 (char 42)"}} indicates a JSON parsing error at the API level, which typically occurs when:

1. Tool call arguments are malformed - The LLM generates incomplete or incorrectly formatted JSON for function parameters
2. Streaming response chunks are concatenated incorrectly - When tool call arguments arrive in multiple streaming chunks, improper concatenation can result in invalid JSON

Possible Causes

Based on the Clawith codebase analysis:

1. Model-specific streaming behavior: Different LLM providers (MiniMax, Kimi/Moonshot) may stream tool call arguments in different patterns. The current code in backend/app/services/llm_client.py concatenates these chunks with tc["function"]["arguments"] += str(arg_chunk), which may produce invalid JSON if the model sends arguments in an unexpected format.

2. LLM-generated malformed JSON: Some models may generate arguments like {"key": "value"} with missing quotes, extra commas, or other syntax issues.

Recommended Actions

To help us investigate further, could you please provide:

1. Which specific skill(s) are failing? (e.g., report-ppt-generator-pro or other PPT generation skills)
2. Is this happening with every request, or only with specific prompts?
3. Can you share the exact prompt you're using? (or a minimal example that reproduces the issue)
4. Is the error consistent, or does it occur intermittently?

In the meantime, you could try:

• Using a different model (e.g., OpenAI, Anthropic, or DeepSeek) to see if the issue persists
• Simplifying the prompt to reduce JSON complexity
• Checking if the issue is specific to certain skill invocations

---

Additional note: Please consider submitting future issues in English so we can better assist you and maintain clear documentation for other users. 🙏

[wisdomqin]

Thanks for this detailed report.

We've analyzed the error and believe we've identified the root cause.

Root Cause

The error Expecting ',' delimiter: line 1 column 43 (char 42) suggests that MiniMax and Moonshot/Kimi's APIs are rejecting the request payload Clawith sends when tool-enabled skills are active. The most likely cause is the parallel_tool_calls: true field — a newer OpenAI-specific parameter that these providers do not yet support. Their strict request schema validation causes a JSON parse error rather than a graceful "unknown field" rejection.

To confirm this on your end:

1. Try starting a clean session with the same agent but no skill attached — just plain conversation. Does it succeed?
2. If yes, the issue is in the tool schema sent to the LLM when skills are loaded.

Our planned fix:

• Make parallel_tool_calls conditional on known-supported providers (OpenAI, Anthropic)
• Suppress it for OpenAI-compatible endpoints that don't advertise support for it

We'll track this fix. Thank you for the detailed reproduction steps.