🐛 Bug: copilotkit_emit_state partial state not merged with current_graph_state in STATE_SNAPSHOT

[Valen-C12]

♻️ Reproduction Steps

1. Create a LangGraph agent with custom state fields (e.g., thinking_steps, run_titles, current_run_id)
2. Use copilotkit_emit_state to emit partial state updates during node execution:

# In node, emit only specific fields
await copilotkit_emit_state(config, {"current_run_id": run_id, "current_turn_index": turn_index})
# Later, emit different fields
await copilotkit_emit_state(config, {"thinking_steps": new_steps})
# Even later, emit another field
await copilotkit_emit_state(config, {"run_titles": current_titles})

3. Observe frontend useAgent hook - each STATE_SNAPSHOT event replaces previous state instead of merging

🔎 Observed Behavior

When copilotkit_emit_state emits partial state (e.g., {"run_titles": {...}}):

• The STATE_SNAPSHOT event contains only {"run_titles": {...}}
• Frontend loses previously emitted fields like thinking_steps
• State "flickers" - fields appear, disappear, then reappear

Console logs showing the issue:

[useAgent] state: {current_run_id: '...', current_turn_index: 6}
[useAgent] state: {thinking_steps: Array(43)}  // ✅ has steps
[useAgent] state: {run_titles: {...}}          // ❌ thinking_steps GONE!
[useAgent] state: {thinking_steps: Array(43)}  // ✅ back again

✅ Expected Behavior

According to documentation:

"the underlying mechanism in _emit_state_sync_event will merge this with the current_graph_state before emitting a complete state object to the frontend"

Each STATE_SNAPSHOT should contain the merged state, preserving all fields.

📃 Root Cause Analysis

In langgraph_agent.py (v0.1.76 and v0.1.77), lines 437-447:

if manually_emit_intermediate_state:
    manually_emitted_state = cast(Any, event["data"])
    yield self._emit_state_sync_event(
        thread_id=thread_id,
        run_id=run_id,
        node_name=node_name,
        state=manually_emitted_state,  # ❌ Only partial state!
        running=True,
        active=True
    ) + "\n"
    continue  # ❌ Skips merge with current_graph_state

The continue statement bypasses lines 479-482 where normal state sync merges with current_graph_state.

💡 Suggested Fix

Merge the manually emitted state with current_graph_state before emitting:

if manually_emit_intermediate_state:
    manually_emitted_state = cast(Any, event["data"])
    # Merge with current graph state
    merged_state = {**current_graph_state, **manually_emitted_state}
    current_graph_state.update(manually_emitted_state)
    yield self._emit_state_sync_event(
        thread_id=thread_id,
        run_id=run_id,
        node_name=node_name,
        state=merged_state,  # ✅ Full merged state
        running=True,
        active=True
    ) + "\n"
    continue

📦 Versions

• copilotkit (Python SDK): 0.1.76 / 0.1.77
• @copilotkit/react-core: 1.51.2
• langgraph: 1.0.7+

🔗 Related Issues

• 🐛 Bug: State emitted using copilotkitEmitState gets overwritten when calling .invoke() in langgraph node #2179 - State emitted using copilotkitEmitState gets overwritten
• 🐛 Bug: The context passed from useCopilotReadable to LangGraph is getting lost at the graph node. #2769 - Context passed from useCopilotReadable getting lost (similar overwrite-not-merge pattern)
• 🐛 Bug: How to emit state correctly in CoAgents? #934 - State overridden when streaming

[copilotkit-support-bot]

This is a confirmed bug in the LangGraphAgent implementation within the Python SDK. When copilotkit_emit_state is used, the partial state update is emitted directly to the frontend without being merged with the existing agent state.

Details

This is a confirmed bug in the LangGraphAgent implementation within the Python SDK. When copilotkit_emit_state is used, the partial state update is emitted directly to the frontend without being merged with the existing agent state. Furthermore, this partial state is stored and continues to override the full state for subsequent events in the same execution cycle, leading to the observed "flickering" effect.

🔎 Analysis

The issue resides in sdk-python/copilotkit/langgraph_agent.py within the _stream_events method:

1. Partial Emission: In the manually_emit_intermediate_state handler (lines 458-472), the code captures event["data"] and passes it directly to _emit_state_sync_event. Since this data only contains the fields passed to copilotkit_emit_state, the resulting STATE_SNAPSHOT is incomplete.
2. State Overwriting: The partial state is saved in the manually_emitted_state variable. At line 491, updated_state = manually_emitted_state or current_graph_state causes all subsequent streaming events (like tool call chunks) to use this partial dict as the baseline state, losing all other state fields and messages until the node exits.

✅ Solution

The manually emitted state must be merged into the current_graph_state accumulator, and the full merged state should be

---

Was this helpful?

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

✅ This was helpful | ❌ Need more help