Early termination
This is the first time using this, so I'm not sure on the debugging process. I added the template to LM Studio prompt template, and after using VSCode for about 2 minutes in planning mode, it spat out this. Any thoughts? Thanks 😃 The version is qwen3.6-froggeric-v21.3.
<function=read_file> <parameter=filePath> /Users/me/Documents/Git/example/shared/src/example.ts <parameter=startLine> 1 <parameter=endLine> 300
There seems to be an issue with this breaking tool calls, I am also using LM Studio and with Ornith and Qwen3.6 in OpenCode and Pi.dev it would break tool_calls and just print the tool call XML into chat, I reset the template to default and they started working again.
Same for me, I had to revert to the standard template on Qwen 3.6 27b - this ruins toolcalling so quickly. With the standard template it randomly stops, but at least I can just tell it to continue and then it keeps going. Once you get into the broken toolcall situation, it stays broken.
I second the issue, I also had to revert to the original chat template
I use OpenCode on llama.cpp
When you see raw XML printed as text in your chat box, it means the client or inference server is not intercepting the output as a function call. This happens in two scenarios:
- The client app (or harness) did not send the tools schema array in the API request payload. Without tools passed in the payload, the server treats the request as a plain chat turn.
- On llama-server, native tool parsing was not enabled. Ensure your server is launched with --reasoning-format deepseek and that your client is connected to an endpoint configured for function calling.
If you are using an older tool harness that strictly expects JSON instead of XML, you can also pass --chat-template-kwargs '{"tool_call_format": "json"}'. Closing this topic.