[ENHANCEMENT] Stream model thoughts

[jamestmartin]

Edited in response to roomote's comments.

Problem (one or two sentences)

Users cannot audit or inspect the internal reasoning of models invoked through Roo Code with the Ollama provider, limiting their ability to diagnose failures, tune prompts, or audit model behavior.

Context (who is affected and when)

This affects users of the Ollama provider who would benefit from insight into model reasoning during API requests and subtasks.

Desired behavior (conceptual, not technical)

Roo Code should add support for viewing reasoning and managing reasoning effort to the Ollama provider.

Constraints / preferences (optional)

• Distinguish clearly between when an API request is queued and when it is streaming, enabling the user to distinguish between queued/loading states and active reasoning streams.
• Ollama supports streaming, so this information would preferably be real-time when available.

Request checklist

• I've searched existing Issues and Discussions for duplicates
• This describes a specific problem with clear context and impact

Roo Code Task Links (optional)

No response

Acceptance criteria (optional)

No response

Proposed approach (optional)

No response

Trade-offs / risks (optional)

No response

[roomote]

Thanks for this enhancement request. Id like to help, but I need some clarification to ensure I implement the right solution.

After exploring the codebase, I found that Roo Code already has built-in support for displaying model reasoning/thinking content:

• The ReasoningBlock component displays thinking content with a timer and collapsible UI
• Reasoning is streamed in real-time when models provide it (e.g., Anthropic extended thinking, DeepSeek reasoning_content, OpenAI reasoning_effort models)
• There's a UI setting "Collapse Thinking Blocks" to control the default display state

However, this only works when the model's API explicitly provides reasoning content (like Claude's extended thinking or o1's reasoning tokens). Many models (including most Ollama models) don't produce separate "thinking" output.

Could you help clarify what you're looking for?

1. Are you using a model that supports thinking/reasoning (like DeepSeek-R1, QwQ, or an Ollama model with thinking support like deepseek-r1:*)? If so and you're not seeing the reasoning block, that would be a bug.

2. Are you looking for a "raw streaming view" that shows the model's output as it streams, before tool parsing? This would be a new feature.

3. Are you looking for API request/response debugging? Something that shows the raw HTTP traffic or token-level details?

4. For the queued vs streaming distinction you mentioned - are you referring to showing when a request is waiting in queue (e.g., rate limited) vs actively receiving data? Roo Code does show different spinner states, but perhaps you want more detail?

Please mention @roomote in your response with more details about your use case, and I'd be happy to propose an implementation plan.

[jamestmartin]

@roomote All of my models support reasoning; in particular, right now, I'm using gpt-oss-20b and gpt-oss-120b with Ollama 0.15.0 and the latest published (non-nightly) version of Roo Code. I have never seen the reasoning block appear. It's always just been API Request with the loading spinner. Collapse Thinking Blocks is unset.

[roomote]

Thanks for the additional context! I've seen your comment about gpt-oss-20b and gpt-oss-120b with Ollama 0.15.0 not showing reasoning blocks despite having the Collapse Thinking Blocks setting unset.

This sounds like a bug where reasoning content from these specific models via Ollama isn't being detected. I'll investigate how Roo Code handles reasoning content from Ollama and what format these models use.

I'll open a PR to share progress once I identify the issue.

[roomote]

I've opened a draft PR with my attempt at addressing this issue: #10952

The PR adds support for Ollama's native thinking feature (available since Ollama 0.5.0+):

1. Native thinking field: The handler now checks for message.thinking field in Ollama responses and yields it as reasoning content
2. Think option: Passes the think option to Ollama chat requests when reasoning is enabled in settings
3. Effort mapping: Maps reasoning effort settings to Ollama's think option values (high/medium/low)

For this to work with your gpt-oss-20b and gpt-oss-120b models, you'll need:

• enableReasoningEffort: true in settings
• The model info to have supportsReasoningEffort: true (this may need to be detected/configured for Ollama models)

Feedback welcome!