Surface think blocks from reasoning models instead of stripping

[tjdoomer]

Summary

• Previously: <think>...</think> tags from reasoning models (DeepSeek-R1, QwQ, Qwen3-coder) were silently stripped
• Now: extracted and emitted as { thought: true, text } parts that the existing Turn system surfaces as DeltaEventType.Thought events
• New thinkBlockParser.ts — extractThinkBlocks() for non-streaming, StreamingThinkExtractor class for cross-chunk streaming
• OpenAI generator: replaced filterThinkTags stripping with extraction in both paths
• Anthropic generator: handles thinking and thinking_delta content blocks from Claude's extended thinking

Test plan

• 11 think block parser unit tests (single/multi chunk, partial tags, reset)
• Updated OpenAI generator tests to expect thought parts instead of stripped output
• All 1873 existing tests pass
• Full workspace build clean
• Manual: run a reasoning model (QwQ, DeepSeek-R1) via LM Studio and verify thinking is visible

Summary by CodeRabbit

• New Features

  • AI model internal reasoning (think blocks/thinking) is now preserved and emitted as separate components instead of being removed.
  • Internal reasoning from both OpenAI and Anthropic models can be selectively displayed in the UI with a toggle option.

• Documentation

  • Added roadmap documentation outlining the multi-phase implementation strategy for exposing internal reasoning blocks.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 342b0d03-479e-41cd-9df9-7a7f76673e82

📥 Commits

Reviewing files that changed from the base of the PR and between 55adf32 and adae23b.

📒 Files selected for processing (6)

• docs/roadmap/06-think-passthrough.md
• packages/core/src/core/anthropicContentGenerator.ts
• packages/core/src/core/openaiContentGenerator.test.ts
• packages/core/src/core/openaiContentGenerator.ts
• packages/core/src/utils/thinkBlockParser.test.ts
• packages/core/src/utils/thinkBlockParser.ts

---

📝 Walkthrough

Walkthrough

This change implements "Think Block Passthrough" functionality by introducing utilities to extract and preserve <think> blocks as separate GenAI parts instead of filtering them out. The OpenAI and Anthropic generators are updated to use these utilities for both streaming and non-streaming responses, and comprehensive tests validate the new behavior.

Changes

Cohort / File(s)	Summary
Think Block Parsing Utilities packages/core/src/utils/thinkBlockParser.ts, packages/core/src/utils/thinkBlockParser.test.ts	New utilities for extracting <think> blocks from content. extractThinkBlocks() handles non-streaming extraction via regex, while StreamingThinkExtractor manages chunk-aware buffering for streaming with state tracking. Tests validate both extraction strategies and boundary-spanning tag handling.
OpenAI Generator Updates packages/core/src/core/openaiContentGenerator.ts	Replaced stateful tag-filtering logic with StreamingThinkExtractor for streaming and extractThinkBlocks() for non-streaming. Think content now emits as separate { thought: true, text: ... } parts instead of being stripped.
OpenAI Generator Tests packages/core/src/core/openaiContentGenerator.test.ts	Renamed test suite from "think tag filtering" to "think tag handling" and updated expectations: think blocks now appear as separate thought parts alongside visible text parts in both streaming and non-streaming modes.
Anthropic Generator Support packages/core/src/core/anthropicContentGenerator.ts	Added handling for thinking_delta events in streaming and thinking blocks in non-streaming responses, yielding them as { thought: true, text: ... } parts.
Roadmap Documentation docs/roadmap/06-think-passthrough.md	New documentation outlining multi-phase implementation strategy for think block passthrough, current behavior across generators, and proposed UI rendering with configuration flags.

Sequence Diagram(s)

sequenceDiagram
    participant Client
    participant OpenAIGen as OpenAI Generator
    participant Extractor as StreamingThinkExtractor
    participant Output as GenAI Parts

    Client->>OpenAIGen: Stream chunk received
    OpenAIGen->>Extractor: process(chunk)
    
    alt Contains closing </think>
        Extractor->>Extractor: Detect </think>, finalize think content
        Extractor->>Output: { thought: true, text: thinkContent }
        Extractor->>Output: { text: visibleText }
    else No closing tag yet
        Extractor->>Extractor: Buffer incomplete think block
        Extractor->>Output: { text: visibleText }
    end
    
    Output->>Client: Emit parts with thought attribution

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~25 minutes

Possibly related PRs

• Fix Anthropic provider: SDK, streaming tool calls, tiktoken #2 — Concurrently modifies anthropicContentGenerator.ts streaming to add thinking_delta/thought parts handling alongside other streaming architecture changes (tool_use, @anthropic-ai/sdk migration).

Poem

🐰 Whisper, whisper, think blocks bound,
No more hidden underground!
Extracted clean, with streaming care,
Thoughts now visible, laid bare! ✨

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feature/think-passthrough

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}