[Epic] Hybrid ADK + AI SDK Migration

[aWN4Y25pa2EK]

Summary

Migrate the HCA-I2P MVP from the current multi-SDK approach to a hybrid architecture:

• Keep: Google ADK for agent orchestration (LlmAgent, Runner, SessionService, FunctionTool)
• Add: Vercel AI SDK with @openrouter/ai-sdk-provider for LLM calls
• Remove: Direct openai SDK usage, @opencode-ai/sdk dependency

Motivation

Current architecture uses 3 different approaches for LLM calls:

1. Custom OpenRouterLlm adapter with raw fetch (fragile format conversion)
2. openai SDK in llm-invoke.ts
3. @opencode-ai/sdk for extended thinking

Issues with current approach:

• Tool call ID instability (Date.now() collisions)
• No streaming support in ADK adapter
• Format conversion bugs (Gemini vs OpenAI message formats)
• Multiple codepaths for same functionality

Benefits of AI SDK

• Native OpenRouter provider with extended thinking support
• Unified API across providers
• Built-in streaming with proper backpressure
• Provider registry for multi-model routing
• Proper reasoning token tracking

Migration Phases

Phase 1: Add AI SDK Dependencies

• [Migration] Phase 1: Add AI SDK Dependencies #20
• Add ai and @openrouter/ai-sdk-provider packages
• Create provider registry with stage-specific presets

Phase 2: Create AI SDK LLM Tool

• [Migration] Phase 2: Create AI SDK LLM Tool #21
• New src/tools/ai-sdk-invoke.ts using AI SDK
• Update llmInvokeTool to use new implementation

Phase 3: Update OpenRouterLlm Internals

• [Migration] Phase 3: Update OpenRouterLlm Internals #22
• Replace raw fetch with AI SDK generateText()
• Fix tool call ID stability
• Add streaming support

Phase 4: Migrate Analysis Stages

• [Migration] Phase 4: Migrate Analysis Stages #23
• Update boundary, gap, navigation, decomposition stages
• Use provider registry for model routing

Phase 5: Remove OpenCode Dependency

• [Migration] Phase 5: Remove OpenCode Dependency #24
• Use OpenRouter :thinking variant for extended reasoning
• Remove @opencode-ai/sdk

Phase 6: Validation & Cleanup

• [Migration] Phase 6: Validation & Cleanup #25
• Remove openai package
• Clean up dead code
• E2E validation

Architecture After Migration

┌─────────────────────────────────────────────────┐
│                  Google ADK                      │
│  (Agent orchestration, Session, FunctionTool)   │
└─────────────────────┬───────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────┐
│              AI SDK Provider Registry            │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────┐  │
│  │  boundary   │  │    gap      │  │decomp:  │  │
│  │  (flash)    │  │  (opus)     │  │thinking │  │
│  └─────────────┘  └─────────────┘  └─────────┘  │
└─────────────────────┬───────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────┐
│         @openrouter/ai-sdk-provider              │
│            (OpenRouter API)                      │
└─────────────────────────────────────────────────┘

Related Files

• src/holons/adk/openrouter-llm.ts - Custom LLM adapter
• src/holons/adk/agents.ts - Agent definitions
• src/tools/llm-invoke.ts - OpenAI SDK wrapper
• src/tools/opencode-client.ts - Extended thinking
• src/config.ts - Configuration