feat(copilotkit): integrate CopilotKit v2 with Bedrock AgentCore

[blove]

Summary

Full integration of CopilotKit v2 with AWS Bedrock AgentCore, enabling a production-ready fullstack AI chat application where frontend React components can participate in the agent's tool loop (generative UI). The PR covers backend runtime, infrastructure (CDK + Terraform), frontend, and two runtime bug fixes verified against a live deployment.

---

What changed

Backend: CopilotKit runtime proxy

• New Lambda (infra-cdk/lambdas/copilotkit-runtime/) — TypeScript Lambda that acts as a CopilotKit Runtime, translating CopilotKit Cloud protocol to AG-UI invocations against the AgentCore runtime endpoint
• API Gateway — new /copilotkit route added to the existing backend stack; URL exported as CopilotKitRuntimeUrl
• CDK and Terraform modules updated in parallel (infra-cdk/lib/backend-stack.ts, infra-terraform/modules/backend/copilotkit_runtime.tf)

Backend: LangGraph AG-UI agent

• ag_ui_langgraph/ — vendored AG-UI agent base class implementing the AG-UI streaming protocol on top of LangGraph; handles checkpoint loading, time-travel regeneration, state merging, and Bedrock-compatible message fixups
• copilotkit/ — CopilotKit middleware package implementing AgentMiddleware for LangGraph: injects frontend tools into model calls, intercepts frontend tool calls to prevent backend tool execution, restores intercepted tool calls to checkpoint on agent exit, and manages app context injection via system messages
• patterns/langgraph-single-agent/langgraph_agent.py — updated to use LangGraphAGUIAgent + CopilotKitMiddleware; exposes the agent via AG-UI protocol with Cognito JWT auth and M2M gateway token fetch
• langgraph-checkpoint-aws upgraded 1.0.1 → 1.0.5

Frontend

• @copilotkit/react-core and @copilotkit/react-ui added
• CopilotChatInterface — wraps CopilotKit provider + CopilotChat; reads CopilotKitRuntimeUrl from runtime config
• PieChart generative UI component (recharts) — registered as a frontend tool so the agent can render charts directly in the chat
• useGenerativeUi hook — exports useMakePieChart action for the CopilotKit tool loop
• Auth updated to surface accessToken (required by AgentCore JWT authorizer)
• Runtime config extended with copilotKitRuntimeUrl

Bug fixes

Fix 1 — Infinite loop after frontend tool call

When the agent called a frontend tool (e.g. createPieChart) and the frontend returned the result, the agent re-called the same tool on every continuation instead of responding with text.

Root cause: patch_orphan_tool_calls (langgraph-checkpoint-aws v1.0.5) generates a new random ToolMessage ID on every checkpoint load. ID-based replacement of the "interrupted" placeholder always failed — the real result was appended alongside the placeholder, and Bedrock rejected the duplicate toolResult IDs (ValidationException).

Fix: _fix_messages_for_bedrock in CopilotKitMiddleware now deduplicates ToolMessages by tool_call_id in-place before each Converse API call, keeping the real result over any "interrupted before completion" placeholder. langgraph_default_merge_state simplified to pass incoming ToolMessages through without the unreliable ID-replacement logic.

Fix 2 — RUN_ERROR on multi-turn follow-up messages

New follow-up messages (brand-new message ID, not yet in any checkpoint) were incorrectly treated as time-travel regeneration requests and returned RUN_ERROR.

Fix: prepare_stream now checks is_continuation — whether all incoming non-ToolMessage IDs are already in the checkpoint — before triggering time-travel. Genuine follow-up turns fall through to normal continuation; time-travel only fires when the last user message ID is already in the checkpoint.

---

Architecture

Browser (React + CopilotKit)
  └─ CopilotKit Cloud
       └─ API Gateway /copilotkit  (CopilotKit Runtime Lambda)
            └─ AgentCore Runtime  (LangGraph agent + CopilotKitMiddleware)
                 ├─ Bedrock (claude-sonnet-4-5)
                 ├─ AgentCore Memory (langgraph-checkpoint-aws)
                 └─ MCP Gateway (backend tools)

Test plan

• Deploy: cd infra-cdk && npx cdk deploy
• Open Amplify URL; sign in via Cognito
• Send "Create a pie chart" — verify createPieChart tool fires and chart renders in chat
• Confirm agent responds with text after chart renders (no re-call loop)
• Send a follow-up question in the same thread — verify correct contextual answer, no RUN_ERROR
• Start a new thread, send two sequential messages — both return text responses
• Verify Terraform path: cd infra-terraform && terraform apply deploys equivalent infrastructure

🤖 Generated with Claude Code