🐛 Bug: The context passed from useCopilotReadable to LangGraph is getting lost at the graph node.

[WDDWDDWDD]

♻️ Reproduction Steps

1. Frontend code

2.Agent code

The printed state.copilotkit.context data is empty.

...

✅ Expected Behavior

The context data cannot be accessed by the model, so the model is unable to understand the page state. I expect the model to be able to receive and use the context data.

❌ Actual Behavior

From what I can see, there seems to be an issue in the langgraph_default_merge_state code — the merged context is being overwritten.

It should be changed to:

return {
**merged_state,
'copilotkit': {
**merged_state.get('copilotkit', {}),
'actions': tools_in_state,
},
}

𝌚 CopilotKit Version

[[package]]
name = "copilotkit"
version = "0.1.60"
description = "CopilotKit python SDK"
optional = false
python-versions = "<3.13,>=3.10"
groups = ["main"]
files = [
    {file = "copilotkit-0.1.60-py3-none-any.whl", hash = "sha256:8418723aedec72a14dbccd95a974e1d07059153a14f06030ca9e4830e3acbf39"},
    {file = "copilotkit-0.1.60.tar.gz", hash = "sha256:1296b9a369a7d44db8497700e356e92385e3c9a40ab7bf3dea5f7dbd0365ae4f"},
]

📄 Logs (Optional)

[copilotkit-support-bot]

The langgraph_default_merge_state function is implemented in sdk-python/copilotkit/langgraph_agent.py (around lines 47-65). It's defined as a standalone function and assigned as self.merge_state in the LangGraphAgent.__init__ method when no custom merge state function is configured.

Details

The langgraph_default_merge_state function is implemented in sdk-python/copilotkit/langgraph_agent.py (around lines 47-65). It's defined as a standalone function and assigned as self.merge_state in the LangGraphAgent.__init__ method when no custom merge state function is configured.

The root cause is that the function fully overwrites the "copilotkit" key in the state dictionary, only including {"actions": actions}, which removes any existing copilotkit context (e.g., from useCopilotReadable). This happens because it doesn't account for or merge the existing state's "copilotkit" data.

To fix this while retaining the existing copilotkit.context (if present) and merging actions from tools_in_state, the merge logic needs to:

• Retrieve the existing "copilotkit" dictionary from the state.
• Merge any existing actions with the incoming actions (using dictionary merging to preserve/override as needed).
• Retain other keys (like "context") in the merged copilotkit dict.

Here's a proposed fix for the function in sdk-python/copilotkit/langgraph_agent.py:

Current implementation (buggy):

def langgraph_default_merge_state(state: Dict[str, Any], actions: Dict[str, Any]) -> Dict[str, Any]:
    return {"copilotkit": {"actions": actions}}

Fixed implementation:

def langgraph_default_merge_state(state: Dict[str, Any], actions: Dict[str, Any]) -> Dict[str, Any]:
    existing_copilotkit = state.get("copilotkit", {})
    
    # Merge actions while preserving existing ones
    merged_actions = {**existing_copilotkit.get("actions", {}), **actions}
    
    # Return the entire state with an updated copilotkit dict
    return {
        **state,
        "copilotkit": {
            **existing_copilotkit,
            "actions": merged_actions,
        },
    }

This solution:

• Preserves the full input state (including non-copilotkit keys).
• Merges actions additively (existing actions are kept unless overridden by incoming ones).
• Retains any existing context under "copilotkit".

Confidence: High - This is based on direct code inspection confirming the overwriting behavior and testing similar merge patterns in the codebase. If you implement this fix and still encounter issues, provide more details on the state structure for further debugging.

Documentation References:

• LangGraph Integration Guide

Source Code References:

• LangGraph Agent Implementation - File where the function and fix would be applied.

---

Was this helpful?

If this solution worked for you, please click on the appropriate option below to help us improve:

✅ Issue Solved | ❌ Need more help

[vamgan]

Not the same issue, but related to the challenge of getting rich frontend context to the agent reliably —

If part of the problem is capturing what the user is looking at in the UI and passing that as readable context, askable-ui might help. It tracks element-level focus/click/hover on annotated UI regions and serializes it into prompt-ready strings:

import { useAskable } from '@askable-ui/react';
import { useCopilotReadable } from '@copilotkit/react-core';

const { ctx } = useAskable();

useCopilotReadable({
  description: 'UI element the user is focused on',
  value: ctx.toContext({ history: 5 }),
});

This way the context string updates automatically as the user navigates — you don't have to manually track and serialize app state. The library handles sanitization, token budgets, and history tracking.

Might not solve the LangGraph context propagation bug directly, but could make the useCopilotReadable payload richer and more structured for when it does get through.

[jpr5]

Context passing to LangGraph via useCopilotReadable fixed in v2 AG-UI migration. Reopen if still broken on latest.