Title: Bionic 1.0.2 - Reasoning mode causes ngPredictTokens HTTP 400
Environment:
- Bionic: 1.0.2 (build 3)
- OS: Windows 11 Pro 64-bit
- GPU: NVIDIA GeForce RTX 3090 Ti 24 GB
- Backend: LM Studio / LMS Core
Problem:
When Reasoning is enabled in Bionic, even the simplest prompt such as "Hello"
fails with the following error:
Engine protocol ngPredictTokens request returned 400:
{"error":{"code":400,"message":"Assistant message must contain either 'content' or 'tool_calls'!","type":"invalid_request_error"}}
Reproduction:
- Start Bionic.
- Select Qwen3.5 35B A3B.
- Enable Reasoning.
- Start a new conversation.
- Send: Hello
- Request fails with HTTP 400.
The same failure was reproduced with:
- Qwen3.5 35B A3B
- Mistral Nemo Instruct 2407
Control test:
Disable Reasoning and start a new conversation.
The same Qwen3.5 model works normally.
Voice transcription with the local Voxtral model also works correctly when
Reasoning is disabled.
Relevant Bionic log:
[LMSInternal][LMSAuthenticator][Client=LM Studio]
[Endpoint=appendUserDraftAndRunUntilWaiting]
Error in RPC handler:
Engine protocol ngPredictTokens request returned 400:
"Assistant message must contain either 'content' or 'tool_calls'!"
Observed:
Reasoning ON -> HTTP 400
Reasoning OFF -> works normally
Expected:
Reasoning mode should send a valid assistant message containing content
and/or tool_calls and should not cause the request to fail.
The issue appears independent of the selected LLM and reproducible specifically
when Bionic Reasoning mode is enabled.
Title: Bionic 1.0.2 - Reasoning mode causes ngPredictTokens HTTP 400
Environment:
Problem:
When Reasoning is enabled in Bionic, even the simplest prompt such as "Hello"
fails with the following error:
Engine protocol ngPredictTokens request returned 400:
{"error":{"code":400,"message":"Assistant message must contain either 'content' or 'tool_calls'!","type":"invalid_request_error"}}
Reproduction:
The same failure was reproduced with:
Control test:
Disable Reasoning and start a new conversation.
The same Qwen3.5 model works normally.
Voice transcription with the local Voxtral model also works correctly when
Reasoning is disabled.
Relevant Bionic log:
[LMSInternal][LMSAuthenticator][Client=LM Studio]
[Endpoint=appendUserDraftAndRunUntilWaiting]
Error in RPC handler:
Engine protocol ngPredictTokens request returned 400:
"Assistant message must contain either 'content' or 'tool_calls'!"
Observed:
Reasoning ON -> HTTP 400
Reasoning OFF -> works normally
Expected:
Reasoning mode should send a valid assistant message containing content
and/or tool_calls and should not cause the request to fail.
The issue appears independent of the selected LLM and reproducible specifically
when Bionic Reasoning mode is enabled.