[Migration] Phase 3: Update OpenRouterLlm Internals

[aWN4Y25pa2EK]

Parent Epic

Part of #19 - Hybrid ADK + AI SDK Migration

Objective

Refactor the ADK OpenRouter LLM adapter to use AI SDK internally instead of raw fetch calls, while maintaining the ADK BaseLlm interface.

Tasks

• Refactor openrouter-llm.ts to use AI SDK generateText()
• Update message format conversion for AI SDK
• Add reasoning token tracking to usageMetadata
• Update runner.ts token extraction for new format
• Test with all 5 agents (root, architect, reviewer, cursor, deep)
• Remove raw fetch and retry logic (handled by AI SDK)

Current State

// Current: Raw fetch with manual format conversion
async *generateContentAsync(llmRequest: LlmRequest): AsyncGenerator<LlmResponse> {
  const messages = geminiToOpenAI(llmRequest.contents, systemInstruction);
  const response = await fetch(`${this.baseUrl}/chat/completions`, {
    method: "POST",
    headers: { ... },
    body: JSON.stringify(requestBody),
  });
  // Manual retry logic, format conversion...
}

Target State

// Target: AI SDK with native features
async *generateContentAsync(llmRequest: LlmRequest): AsyncGenerator<LlmResponse> {
  const { text, toolCalls, usage, reasoningText } = await generateText({
    model: this.openrouter(modelId),
    messages: this.convertToAISDKMessages(llmRequest.contents),
    tools: this.convertToAISDKTools(llmRequest.config?.tools),
    providerOptions: isThinkingModel ? {
      openrouter: { reasoning: { effort: 'high' } }
    } : undefined,
  });
  
  yield this.convertToGeminiResponse(text, toolCalls, usage, reasoningText);
}

Key Changes

1. Replace fetch with AI SDK

• Remove geminiToOpenAI() manual conversion
• Use AI SDK's native message format
• Remove manual retry logic (AI SDK handles it)

2. Add Reasoning Token Support

usageMetadata: {
  promptTokenCount: usage.promptTokens,
  candidatesTokenCount: usage.completionTokens,
  totalTokenCount: usage.totalTokens,
  reasoningTokenCount: usage.reasoningTokens,  // NEW
}

3. Support :thinking Variant

const isThinkingModel = modelId.includes(':thinking');
providerOptions: isThinkingModel ? {
  openrouter: { reasoning: { effort: 'high' } }
} : undefined,

Files to Modify

• src/holons/adk/openrouter-llm.ts - Main refactor
• src/holons/adk/runner.ts - Update token extraction
• src/holons/adk/agents.ts - Test model configurations

Bugs to Fix During Migration

Bug	Location	Fix
Unstable tool_call_id	Line 101	Use crypto.randomUUID()
Token format mismatch	Lines 264-276	Support both Gemini and OpenAI formats
No streaming	Line 268	Document as future enhancement

Acceptance Criteria

• All 5 ADK agents work with refactored adapter
• Tool calling functions correctly
• Token usage properly extracted from AI SDK response
• Reasoning tokens tracked for :thinking models
• No regression in existing functionality

Estimated Effort

2 days