fix: resolve followup message failure after frontend tool calls#2
Open
fix: resolve followup message failure after frontend tool calls#2
Conversation
Adds end-to-end CopilotKit v2 chat support — streaming generative UI
(pie chart), frontend action tools, and persistent memory — running on
Bedrock AgentCore with Claude Sonnet via the Converse API.
## Backend (patterns/langgraph-single-agent)
### langgraph_agent.py — complete rewrite for AgentCore + CopilotKit
- Replaces `create_react_agent` + raw streaming with `create_agent`
(from `langchain.agents`) using `CopilotKitMiddleware`, connecting the
LangGraph agent to CopilotKit's frontend action / generative-UI protocol
- Adopts `BedrockAgentCoreApp` (`@app.entrypoint`) instead of FastAPI,
keeping AgentCore's native invocation model
- MCP tools fetched per-request from the AgentCore Gateway via
`MultiServerMCPClient` with a fresh OAuth2 token each time (avoids
token-expiry in long-running processes)
- Actor identity resolved from `forwardedProps` keys
(`actor_id`/`user_id`) or the `sub` claim in the Cognito Bearer JWT,
then threaded through `AgentCoreMemorySaver` so each user's history is
isolated
- `ActorAwareLangGraphAgent(LangGraphAGUIAgent)` adds three overrides
required by `AgentCoreMemorySaver`:
1. `_filter_orphan_tool_messages` — restores `tool_calls` stripped by
`clean_orphan_tool_calls` (frontend tools have no ToolMessage in the
checkpoint); ensures MESSAGES_SNAPSHOT carries `toolCalls` so the
rendered component (e.g. pie chart) is not removed when the snapshot
overwrites client state
2. `langgraph_default_merge_state` — prepends repaired AIMessages when
Run 2 (CopilotKit follow-up) adds a ToolMessage; without this,
`_fix_messages_for_bedrock` strips the `tool_use` content block
(because `tool_calls=[]`) → orphan ToolMessage → Bedrock API error
3. `get_checkpoint_before_message` — injects `actor_id` into the
LangGraph config for time-travel / edit history lookups
- `serialize_agui_event` / terminal-event guard in `invocations` ensure
the AG-UI stream always ends with `RUN_FINISHED` or `RUN_ERROR`
### requirements.txt
- Replaces `langgraph` + `langchain-aws` stubs with the full dependency
set: `ag-ui-protocol`, `partialjson`, `langgraph==1.0.10rc1`,
`langchain-aws==1.0.0`, `langchain-mcp-adapters`, `copilotkit` (local
vendor), `bedrock-agentcore`, and pinned versions for reproducibility
### Dockerfile
- Installs the local `copilotkit/` and `ag_ui_langgraph/` vendor packages
via `pip install -e` so no PyPI dependency is needed for these patched
libraries
## Vendored packages (copilotkit/ and ag_ui_langgraph/)
Local copies of CopilotKit SDK and ag-ui-langgraph with Bedrock
Converse API compatibility patches (sourced from PR mme:mme/local-copilotkit):
### copilotkit/copilotkit_lg_middleware.py — CopilotKitMiddleware
- `_fix_messages_for_bedrock`: strips unanswered `tool_calls` and syncs
`tool_use` content blocks before each Bedrock model call, preventing
the `toolUse` / `toolResult` interleaving errors the Converse API
requires
- `before_agent`: injects app context from `copilotkit.context` as a
`SystemMessage` at the start of each agent turn
- `after_model` / `after_agent`: intercepts frontend tool calls so they
are not forwarded to `ToolNode`, then restores them to the checkpoint
after the agent exits (enables CopilotKit's frontend-action loop)
- `awrap_model_call`: merges frontend tool definitions into the model
request so the LLM can call them
### ag_ui_langgraph/agent.py — LangGraphAGUIAgent
- `langgraph_default_merge_state`: fixes string `args` in checkpoint
`tool_calls`, replaces fake ToolMessages injected by
`patch_orphan_tool_calls` with the real AG-UI result, and deduplicates
tool definitions across runs
- `_filter_orphan_tool_messages` / `_ORPHAN_TOOL_MSG_RE`: removes fake
ToolMessages (pattern: "Tool call '…' with id '…' was interrupted
before completion.") that AgentCore's saver injects when a tool call
has no matching result in the checkpoint
## CopilotKit Lambda runtime (infra-cdk/lambdas/copilotkit-runtime/)
New TypeScript Lambda that sits between the frontend and the Python
AgentCore agent, acting as the CopilotKit server-side runtime:
- `CopilotRuntime` with `InMemoryAgentRunner` wraps the Python agent as
an `HttpAgent` (AG-UI over HTTP)
- `CopilotKitRunner` (extends `InMemoryAgentRunner`) overrides `connect`:
on reconnect it replays `TOOL_CALL_RESULT` events for every `toolCall`
in the `MESSAGES_SNAPSHOT`, preventing CopilotKit's `processAgentResult`
from re-triggering Run 2 when the user reloads the page
- Supports multiple named agents via env vars
(`LANGGRAPH_AGENTCORE_AG_UI_URL`, `STRANDS_AGENTCORE_AG_UI_URL`) or a
single `AGENTCORE_AG_UI_URL` fallback; agent selected by
`COPILOTKIT_AGENT_NAME`
## Frontend (frontend/)
### CopilotChatInterface.tsx (new)
- Loads runtime config (`copilotKitRuntimeUrl`) from `aws-exports.json`
- Wraps `<CopilotKitProvider>` + `<CopilotChat>` with Cognito Bearer
token forwarded as `Authorization` header
### Generative UI
- `PieChart.tsx`: Recharts-based pie chart component with legend,
dark-mode support, and a Zod schema (`PieChartPropsSchema`) so
CopilotKit can validate tool arguments before rendering
- `useGenerativeUi.ts`: registers `pieChart` as a controlled generative-UI
component via `useComponent` from `@copilotkit/react-core/v2`
- `useExampleSuggestions.ts`: registers suggested prompts in the chat UI
### ChatPage / ChatInterface routing
- `ChatPage.tsx` routes to the new `CopilotChatInterface` when
`copilotKitRuntimeUrl` is present in config; falls back to the existing
`ChatInterface` otherwise
- `main.tsx` / `auth.ts` updated for compatibility
### package.json
- Adds `@copilotkit/react-core`, `@copilotkit/react-ui`, `recharts`, and
`zod`; pins `@copilotkit/runtime-client-gql` for the v2 API surface
## Infrastructure
### CDK (infra-cdk/)
- `backend-stack.ts`: adds a `CopilotKitRuntimeFunction` (Node.js 22
Lambda, 512 MB, 5 min timeout) with `LANGGRAPH_AGENTCORE_AG_UI_URL` /
`STRANDS_AGENTCORE_AG_UI_URL` env vars pointing to the AgentCore
runtime URLs; exposes `/copilotkit/{proxy+}` on the existing API
Gateway with Cognito authorizer
- `fast-main-stack.ts`: threads the new Lambda construct through the
stack output
- `config.yaml`: adds `copilotKitRuntimeUrl` to the frontend runtime
config exported to `aws-exports.json`
### Terraform (infra-terraform/)
- `copilotkit_runtime.tf`: equivalent Lambda + IAM + API Gateway
resources for Terraform deployments
- `locals.tf`, `outputs.tf`, `ssm.tf`: expose `copilotkit_runtime_url`
as an SSM parameter and stack output
- `deploy-frontend.{py,sh}` / `scripts/deploy-frontend.py`: write
`copilotKitRuntimeUrl` into `aws-exports.json` during frontend deploy
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…oint-aws 1.0.5 Upgrade langgraph-checkpoint-aws from 1.0.1 to 1.0.5. Version 1.0.5 uses patch_orphan_tool_calls (injecting placeholder ToolMessages) instead of clean_orphan_tool_calls (stripping tool_calls). The vendored ag_ui_langgraph base class already handles patch_orphan_tool_calls correctly via _filter_orphan_tool_messages and langgraph_default_merge_state, and CopilotKitMiddleware handles Bedrock API errors via _fix_messages_for_bedrock. Remove from ActorAwareLangGraphAgent: - _reconstruct_tool_calls helper - _filter_orphan_tool_messages override - langgraph_default_merge_state override - get_checkpoint_before_message override Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…sage bugs Fix two bugs in the LangGraph AG-UI agent when used with CopilotKit frontend tools: **Fix 1: Infinite loop after frontend tool call (pie chart / createPieChart)** Root cause: `patch_orphan_tool_calls` (langgraph-checkpoint-aws v1.0.5) generates a new random ToolMessage ID on every checkpoint load. The previous approach tried to replace the placeholder by ID in `stream_input["messages"]`, but the ID from the first `aget_state()` call never matches the placeholder ID from the internal `astream_events` reload — so the real result was appended alongside the placeholder, causing Bedrock to reject duplicate `toolResult` IDs (ValidationException). Fix: added step 4 to `_fix_messages_for_bedrock` in `CopilotKitMiddleware` to deduplicate `ToolMessage`s by `tool_call_id` in-place before each Converse API call, keeping the real result over any "interrupted before completion" placeholder. Simplified `langgraph_default_merge_state` to pass incoming `ToolMessage`s through without attempting unreliable ID-based replacement. Also upgrades `langgraph-checkpoint-aws` from 1.0.1 to 1.0.5 which switches from `clean_orphan_tool_calls` (removes tool_calls, breaking continuation) to `patch_orphan_tool_calls` (adds placeholder ToolMessages, enabling continuation). **Fix 2: RUN_ERROR on multi-message follow-up turns** Root cause: when the checkpoint had more messages than the incoming request, the agent always triggered time-travel regeneration — including for legitimate follow-up turns (new user message after a prior exchange). Fix: checks whether all incoming non-ToolMessage IDs are already in the checkpoint before deciding to time-travel. If they are, it's a continuation; only time-travel when the last user message ID exists in the checkpoint but the incoming messages are a proper subset (indicating a genuine re-generation request). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…oList, TodoCanvas) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add explicit typed cast for agent.state in TodoCanvas to avoid untyped access - Import Todo type in TodoCanvas - Prevent saving empty values in TodoCard.saveEdit by keeping the editor open - Remove redundant `as "pending" | "completed"` cast in TodoList.toggleStatus - Remove dead null-check on non-nullable todos array in TodoList Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…Kit registrations Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…s into CopilotChatInterface Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… infrastructure Introduces a query_data LangChain tool backed by a 40-row financial CSV (db.csv), updates tools/__init__.py to export it, adds three passing unit tests, and adds conftest.py files that resolve the repo-root tools/ vs agent tools/ sys.path conflict. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…dos tools Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Remove manual terminal event tracking, exception wrapping, and serialize_agui_event helper from the invocations entrypoint — just forward events via model_dump, matching the FastAPI endpoint pattern. Also set admin_user_email and fix strands runtime ARN fallback. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wire @ag-ui/mcp-apps-middleware into the CopilotKit runtime Lambda so agents can use Excalidraw's MCP server for open generative UI. Add the corresponding "network diagram" suggestion to the frontend. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fix Bedrock ValidationException ("Expected toolResult blocks") that occurred on
follow-up messages after using shared state (Canvas/Todos). The root cause was
that _fix_messages_for_bedrock deduplicated ToolMessages by dropping the
placeholder (correct position, adjacent to AI message) and keeping the real
result (wrong position, appended at end). Bedrock requires toolResult blocks
immediately after their corresponding toolUse.
Three fixes in copilotkit_lg_middleware.py:
- Dedup now replaces placeholder in-place with real result, preserving position
- Unanswered tool_call detection uses adjacency check instead of global lookup
- Orphan ToolMessages are cleaned up after stripping tool_calls
Also adds app mode UI (Chat/App toggle), error handling in agent entrypoint,
and None-event filtering.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
ValidationException("Expected toolResult blocks") on follow-up messages after using shared state (Canvas/Todos) or any frontend tool callcopilotkit/copilotkit_lg_middleware.py:enableAppMode/enableChatModefrontend toolsRunErrorEventinstead of silent crash)printstatements withloggingTest plan