gpt-oss tool calling broken after recent update - parser error during response generation

[liampierc3]

Bug Description

Tool calling with gpt-oss-20b fails during response generation with parsing error. The model successfully executes tool calls, but crashes when trying to generate a response with the results.

Error Message

Failed to regenerate message
Failed to parse tool call: Expected "<|message|>", but got "to=function" at index 89.

Environment

• LM Studio Version: [Updated on Dec 12, 2024 - latest version]
• Model: mlx-community/gpt-oss-20b-MXFP4-Q8
• Platform: macOS (Apple Silicon)
• MCP Servers Enabled:
  • @modelcontextprotocol/server-filesystem
  • @modelcontextprotocol/server-brave-search

Steps to Reproduce

1. Load gpt-oss-20b model with MCP tools enabled
2. Ask a question that triggers search tools (e.g., "Search my memory for X")
3. Observe: Tools execute successfully multiple times
4. Observe: Model attempts to generate response
5. Error occurs: Parser crashes with the error above
6. No output is returned to user

Expected Behavior

The model should successfully generate a response using the tool results, as it did before the recent LM Studio update.

Actual Behavior

• Tools execute successfully ✓
• Model begins response generation ✓
• Parser encounters gpt-oss's native tool call format ✗
• Parser expects <|message|> at position 89 but finds to=function instead ✗
• Generation crashes, no output returned ✗

Technical Details

gpt-oss uses this chat template format for tool calls:
<|start|>assistant to=functions.TOOL_NAME<|channel|>commentary json<|message|>{"args": "..."}<|call|>

The new LM Studio parser appears to expect <|message|> earlier in the sequence, causing it to fail when it encounters to=function at index 89.

Evidence of the issue

• This worked correctly before the December 12, 2024 update
• The same setup works perfectly with Qwen3-VL-8B-Instruct-MLX-8bit model
• The gpt-oss chat template hasn't changed - it's defined in the model's tokenizer_config.json
• Only LM Studio's parser behavior has changed

Workaround

Using Qwen3-VL-8B-Instruct-MLX-8bit instead of gpt-oss works correctly with identical MCP server configuration.

Additional Context

gpt-oss has a unique chat template that uses special tokens like <|channel|>, <|message|>, and to= syntax for tool routing. The model expects its native format, but the new LM Studio parser appears to have broken compatibility with this format.

[ryan-the-crayon]

Hi. When LM Studio fails to parse a tool call, it will also provide the raw output if you "expand" the tool call. See the screenshot attached:

Can attach what the LLM actually generated? My suspicion is that gpt-oss-20b produced an invalid tool call.

(I was able to get gpt-oss-20b to produce the above invalid tool call by messing up its parameters)

[liampierc3]

Does it on every model for me. Heres the screenshots.

[liampierc3]

This issue is gone now after the last update.

[iamdroppy]

Still happening 0.4.2