fix: Add AWS Bedrock Extended Thinking support for Claude models (#4343, #4201)

[hannesrudolph]

Description

Fixes #4343
Fixes #4201

This PR adds support for AWS Bedrock Extended Thinking (reasoning mode) for Claude 3.7/4 models, with improved error handling and retry logic.

Changes Made

1. Added Extended Thinking Support

• Updated model definitions to include supportsReasoningBudget: true for:
  • Claude 3.7 Sonnet (anthropic.claude-3-7-sonnet-20250219-v1:0)
  • Claude 4 Sonnet (anthropic.claude-sonnet-4-20250514-v1:0)
  • Claude 4 Opus (anthropic.claude-opus-4-20250514-v1:0)

2. Implemented Reasoning Configuration

• Extended thinking is disabled by default (as per AWS requirements)
• Only enabled when user explicitly sets enableReasoningEffort: true
• Added additionalModelRequestFields with thinking configuration when enabled
• Uses correct anthropic_version: "bedrock-20250514"
• Removes incompatible parameters (temperature, topP) when thinking is enabled

3. Enhanced Stream Processing

• Added support for new event types:
  • thinking content blocks
  • thinking_delta events
  • signature_delta events
• Properly yields reasoning chunks with type: "reasoning"

4. Improved Error Handling

• Added retry logic with exponential backoff (up to 3 retries)
• Added specific error type for thinking-related errors
• Improved generic error message from "Unknown Error" to more descriptive text
• Errors now properly parsed and displayed to users

5. Updated UI

• Added ThinkingBudget component to Bedrock settings
• Shows "Enable Reasoning Mode" toggle for supported models
• Allows configuration of max thinking tokens (1024-80% of max output tokens)

Testing

• All existing tests pass
• Added comprehensive test suite for reasoning functionality
• Manual testing completed:
  • Verified reasoning is disabled by default
  • Verified reasoning only enabled when explicitly requested
  • Tested error handling and retry logic
  • Confirmed UI shows reasoning toggle only for supported models

Verification of Acceptance Criteria

• "Enable reasoning" option now available for Bedrock Claude models
• Better error handling with retry logic (addresses @hannesrudolph's concern)
• Improved error parsing and messages
• Support for Claude 3.7/4 extended thinking via AWS Bedrock
• Reasoning is opt-in only (disabled by default)

Checklist

• Code follows project style guidelines
• Self-review completed
• Comments added for complex logic
• Documentation updated (if needed)
• No breaking changes
• Tests added for new functionality

Additional Context

This implementation follows the technical requirements where extended thinking/reasoning is disabled by default and only enabled when explicitly requested by the user. This maintains backward compatibility while adding the new functionality requested in the issues.

---

Important

Adds AWS Bedrock Extended Thinking support for Claude models with new configurations, error handling, and tests.

• Behavior:
  • Adds support for AWS Bedrock Extended Thinking for Claude 3.7/4 models in bedrock.ts.
  • Extended thinking is disabled by default; enabled with enableReasoningEffort: true.
  • Removes incompatible parameters (temperature, topP) when thinking is enabled.
• Stream Processing:
  • Handles new event types: thinking, thinking_delta, signature_delta in bedrock.ts.
  • Yields reasoning chunks with type: "reasoning".
• Error Handling:
  • Adds retry logic with exponential backoff in createMessage() in bedrock.ts.
  • Introduces specific error type for thinking-related errors.
  • Improves error messages and parsing.
• Testing:
  • Adds bedrock-reasoning.spec.ts for testing reasoning functionality.
  • Tests include default behavior, enabled reasoning, unsupported models, and event handling.

^{This description was created by for 13e272c. You can customize this summary. It will automatically update as commits are pushed.}

[daniel-lxs]

I can't actually test if this actually works but, I left some suggestions based on how the Anthropic provider works.

[hannesrudolph]

Review Issues Addressed ✅

Fixed the remaining type safety issue:

• Added StreamChunk interface with proper type: 'reasoning' | 'text' typing
• Updated all generator methods to use Generator<StreamChunk, void, unknown> instead of any

All other review feedback was already implemented. Tests pass, no lint errors.

[hannesrudolph]

Type Safety Fix Applied ✅

Fixed the CI test failures by resolving type compatibility issues:

• Replaced custom StreamChunk interface with proper ApiStreamChunk import
• Updated all generator method signatures to use ApiStreamChunk type
• Resolved yield* delegation type mismatches between generator methods

All local tests pass and TypeScript compilation is clean. The CI failures should now be resolved.

[hannesrudolph]

this implementation is no good.