OpenAI reasoning

[AnthonyRonning]

OpenAI-compatible reasoning models like GPT‑OSS do separate out reasoning text from the final “delta content” answer in the API response.

Here’s how it works:
• GPT‑OSS (20B and 120B) produces explicit chain-of-thought (CoT) output that’s sent in a dedicated reasoning channel. The actual answer content is separately delivered in the content property .
• In streaming mode, these models emit reasoning prose incrementally via events such as response.reasoning_text.delta, and signal completion with response.reasoning_text.done. This ensures the reasoning log is traceable and separable from the final output .
• The final channel only shows the final answer. Developers can configure whether to exclude the reasoning entirely (e.g. setting reasoning: { exclude: true }) if they want a clean, concise response .

This architecture is part of OpenAI’s Harmony response format, supported when using the OpenAI-compatible Responses API. It is specifically designed to keep the reasoning trace and output message as separate structured items — making it easier to interpret, audit, or hide the reasoning path if needed .

⸻

🧠 Summary table

Feature GPT‑OSS Behavior
Reasoning Text / CoT Provided in a separate reasoning channel
Final Answer Content Delivered purely in the content property
Streaming Support Yes — via response.reasoning_text.delta and …done events
Exclude Reasoning Option Yes — can disable reasoning output explicitly

[AnthonyRonning]

@claude i think we have to handle OpenAI reasoning content a little differently. Looks like another value to parse. We may need to handle it like the deepseek "thinking" text does. In our markdown formatting/parsing for chat messages.

The main difference is that the deepseek one just type in the normal delta content

I would like reasoning content to also render that way too. Please work on this based on the guide about reasoning content above.

[claude]

Claude finished @AnthonyRonning's task —— View job • claude/issue-182-20250806-0715 • Create PR ➔

---

Implementation Complete! ✅

I've successfully implemented OpenAI reasoning content support for GPT-OSS and similar models.

Tasks Completed

• Analyze current DeepSeek thinking implementation
• Find OpenAI streaming/parsing code
• Implement reasoning content handling
• Test and validate changes
• Run linting and build commands
• Commit and push changes

What Was Implemented

Key Changes in useChatSession.ts:

• Added reasoning content capture from chunk.reasoning_text.delta events
• Wrap reasoning in <think> tags for consistency with existing DeepSeek implementation
• Combine reasoning and regular content during streaming
• Maintain backward compatibility with existing thinking UI

Technical Details:

• Reasoning content is captured via (chunk as { reasoning_text?: { delta?: string } })?.reasoning_text?.delta
• Content is wrapped in <think> tags so existing parseThinkingTags() function can process it
• During streaming, reasoning appears first, followed by the final answer
• Uses the same collapsible thinking block UI as DeepSeek models

Testing:

• Code passes all linting checks
• Build completes successfully
• Pre-commit hooks pass formatting and build validation

The implementation reuses all existing thinking UI components and parsing logic, so OpenAI reasoning models will render identically to DeepSeek thinking - with collapsible sections, timing information, and the same visual styling.

---