gpt-oss interactions seem incomplete

[jfriv]

I am trying to use the new openai/gpt-oss:20b model in Android Studio in agent mode via ollama. I tried a few different configurations and seem to have it working as an openai model with a custom apiBase. The problem I am seeing is that the interactions I'm having with the model with continue in agent mode seem incomplete. Tool calling appears to work but there are only ever commentary on a few of the responses, and sometimes the interaction just stops without a status update or prompt to continue. I see there is an issue with "agent" mode being greyed out for gpt-oss as reported in this issue #6984 and resolved by this PR #7006 but my custom model configuration with ollama is not affected by this issue because the agent mode is not greyed out.

Here is my model config:

name: Local Assistant
version: 1.0.0
schema: v1
models:
  - name: gpt-oss:20b
    provider: openai
    model: gpt-oss:20b 
    apiBase: http://192.168.86.111:11434/v1
    apiKey: xxx
    roles:
      - chat
      - edit
      - apply
    capabilities:
      - tool_use

Seems to me continue is not properly parsing OpenAI's Harmony Response Format. Anyone else seeing this and have a fix?

[RomneyDa]

@jfriv I'm curious, does this specifically happen during file edits when it tries to apply code?

[chezsmithy]

@RomneyDa see my similar comment on discord when I activated it for Bedrock.

[jfriv]

I'm not really seeing any consistency, just seems like its dropping almost all text/conversation content, and continue is only parsing code blocks and tool calling. Could be openai/gpt-oss:20b just doesn't talk much 😆 although, I recall seeing some markup mixed in the text response that was visible in the continue view. Sorry, I haven't really had any time to mess around with it any further in the last week. I will follow up with more detail if I'm able to find more time.