Misc. bug: gpt-oss Jinja template error when assistant message has reasoning_content and tool_calls

[abhijitb11]

Summary

I had a working build of llama-cli and llama-server that was previously stably running gpt-oss-120b from ggml huggingface. After pulling and rebuilding many commits later, I found that during multi-turn chat requests (specifically while coding against this model using OpenCode) that 500 errors were starting to appear.

I narrowed down the issue to a specific PR that caused the regression. There is a potential one-line fix for this.

Details

• Version: 8069 (dc0732143)
• Module: llama-server / common / chat.cpp
• Problem: After common : move gpt-oss reasoning processing to init params #16937, the gpt-oss template throws 'content' is not defined when a multi-turn conversation has an assistant message with both reasoning_content and tool_calls
• Root cause: The reasoning_content -> thinking copy leaves content on the adjusted message, which the template's thinking branch doesn't expect
• Regression : from common : move gpt-oss reasoning processing to init params #16937

Error output:

Server error output

srv    operator(): got exception: {"error":{"code":500,"message":"...
  Error: Jinja Exception: Cannot pass both content and thinking in an
  assistant message with tool calls! Put the analysis message in one
  or the other, but not both.","type":"server_error"}}

[aldehir]

Pretty sure this bug has always been there, it's just that commentary messages are rare and only emitted by the 120b variant.