What happened

#7657 ("fix(react-core): wait for thread replay before sending prompts") added a guard in CopilotChat.tsx's waitForActiveRunToSettle that awaits pendingConnectionsRef.current.get(agent) — a promise wrapping agent.connect() — before allowing a new submission, specifically to prevent sending while a saved conversation is still replaying.

This doesn't help for agents built on @ag-ui/client's generic AbstractAgent/HttpAgent (e.g. @ag-ui/crewai's CrewAIAgent) that never override connect(). AbstractAgent.connect() is:

connect(e){throw new l.AGUIConnectNotImplementedError}

— always throws — and connectAgent()'s own catchError specifically swallows exactly that exception type, resolving the connection promise to an already-complete empty observable almost instantly. So the new guard has nothing real to await for this agent type: it resolves before any actual history replay has happened.

Repro

1. A CopilotKit app registers an @ag-ui/crewai CrewAIAgent (or any other HttpAgent subclass that doesn't override connect()) as its agent, talking to a backend with a saved multi-turn conversation.
2. Reload the page / revisit an existing thread with history.
3. Immediately (before the replayed history visibly finishes rendering) type and send a new message.
4. The new message renders ahead of / interleaved incorrectly with the replayed history that logically precedes it, and — per the AG-UI client's merge-by-id behavior (deliberately never reorders once a message is placed — Snapshot tool results are appended after streamed follow-up text ag-ui-protocol/ag-ui#2767) — stays that way permanently for the session.

Why #7657 doesn't fix this

The PR's own description scopes it to "Intelligence proxies" (ProxiedCopilotRuntimeAgent) hydrating through a delegate whose completion isn't otherwise visible — validated against "a real FastAPI/LangGraph SQLite backend." That's a narrower agent architecture than the generic AG-UI HttpAgent path. For HttpAgent-based agents, connectAgent() is a confirmed no-op (see above), so the new pendingConnectionsRef-based guard introduced in that PR never actually waits on this path's real replay work — whatever mechanism this agent type actually uses to load history isn't observed by the new guard at all.

Suggested direction

waitForActiveRunToSettle would need to also account for a replay that happens via an actual run (RUN_STARTED → ... → RUN_FINISHED), not only via connect()/connectAgent() — e.g. also awaiting agent.isRunning/activeRunCompletionPromise settling for whatever triggers the initial history load for a resolved threadId, regardless of whether that agent type implements connect().

Environment

• @copilotkit/react-core / react-ui / runtime: 1.77.1
• @ag-ui/client: 1.0.2
• @ag-ui/crewai: 0.0.4