[v2] Message view freezes mid-run in long multi-tool runs (LangGraphAgent); state updates flow but UI stops re-rendering

Environment

• @copilotkit/react-core / react-ui / runtime: 1.64.2 (all v2 imports)
• Agent: LangGraphAgent (LangGraph dev server, stream_mode=['events','values','updates','messages-tuple'])
• LLM: DeepSeek (reasoning model) via OpenAI-compatible API
• Next.js 16 + React 19, production build

Symptom

In long multi-tool runs (e.g. query tool → long reasoning → file-generation tool → final answer), the chat view stops re-rendering mid-run while the run is still executing:

• The last visible content is typically the assistant text right before a tool call.
• The run completes successfully server-side (verified in traces and server logs).
• The client keeps receiving and processing events — we instrumented agent.subscribe({ onEvent }) and events (STATE_SNAPSHOT, RAW, …) keep flowing after the UI freezes.
• No console errors.
• After the freeze, nothing brings the content back — waiting past run completion does not help. Only starting a new run flushes the accumulated content into view.

What we verified (elimination)

• Server SSE stream is complete and well-formed (replayed the run: all ~2000 events valid JSON, RUN_FINISHED present).
• Component renderers are fine in isolation (SSR smoke tests with the exact tool payloads).
• Not throttling-related: we don't set throttleMs/defaultThrottleMs, so the unthrottled path is used.
• A "reconcile watchdog" subscriber that returns { messages: [...messages] } (fresh array reference) on onRunFinishedEvent (outcome success) / onRunFailed / onRunErrorEvent does not unfreeze the view — suggesting the blockage is deeper than the MemoizedAssistantMessage reference comparison.
• Two empirical red lines we hit while experimenting (both reproduce reliably):
  1. Returning a messages mutation from a subscriber while the run is paused on an interrupt breaks interrupt UI (the choice card from useInterrupt never renders).
  2. Returning periodic messages mutations during an active run corrupts in-flight streaming (the view freezes almost immediately).

Repro characteristics

• Intermittent for short runs; near-deterministic for long runs (multiple tool calls + 20–60s reasoning + a tool whose arguments are a large JSON payload, ~10–20 KB streamed as TOOL_CALL_ARGS deltas).
• Freeze point in our logs consistently lands right after a generate_file-like tool call's result, or right after an assistant message that carries both visible text and tool calls.
• Our integration details that may matter: assistant messages from DeepSeek carry content as an array of parts (reasoning_content + text); heavy STATE_SNAPSHOT traffic (500+ per run) from LangGraph values stream mode.

Happy to provide

• Full client-side event log (window.__evts probe) vs. server SSE comparison for a frozen run
• Langfuse trace showing the run completing successfully
• A minimal repro if you can point us at a reference LangGraph + v2 setup

Ask

1. Is this a known class of issue in v2's message reconciliation (memoized message view vs. in-place state mutation)?
2. Is there a supported way to force the message view to re-sync from agent.messages after a run? (Our subscriber mutation does not achieve this.)
3. Guidance on the two red lines above — are subscriber messages mutations during interrupts / mid-run supposed to be safe?

Start by reproducing the long multi-tool run and trace agent.subscribe({ onEvent }), agent.messages, and the MemoizedAssistantMessage and useInterrupt paths. Compare the event flow after TOOL_CALL_ARGS, STATE_SNAPSHOT, and RUN_FINISHED with the rendered message state. Done means the message view continues updating through long runs without breaking interrupt UI or streaming reconciliation.