fix(websockets): /api/ws broadcast is a single global callback — multi-client delivery and chat-history persistence both break #14814

Summary

The main event WebSocket (/api/ws) cannot serve more than one client. The broadcast handler is
held as a single scalar callback, so each new connection overwrites the previous one, and the
first disconnect clears it for everyone. A third-order effect is that chat history stops being
persisted.

Found during a research audit comparing AutoBot's event/state layer against the Agent Host Protocol
(see Sources). This is a pre-existing defect, independent of any protocol adoption.

Root cause

autobot-backend/event_manager.py:31 declares the callback as a scalar, not a collection:

self._websocket_broadcast_callback: Callable[[Dict[str, Any]], Awaitable[None]] | None = None

register_websocket_broadcast (event_manager.py:53) assigns straight over it, and
autobot-backend/api/websockets.py:681 calls that once per accepted connection via
get_event_bus().register_ws_broadcast(broadcast_event).

Impact

1. Last-writer-wins. Client B connecting silently stops Client A receiving anything on this
   path. No error, no warning, no log.
2. First disconnect mutes every remaining client. api/websockets.py:694 clears the callback to
   None in the endpoint's finally — including for a connection that was already superseded.
   Clients that are still connected go permanently silent until they reconnect.
3. Chat history stops persisting. _add_to_chat_history is invoked from inside
   broadcast_event (api/websockets.py:510, helper at :459). No registered callback means no
   broadcast, which means no history write. This is a data-loss path, not only a UI path.

Blast radius

PersistStrategy.NONE routes to this single-callback path (events/bus.py:85). Current callers:

• api/workflow.py — 7 call sites (:562, :658, :693, :716, :843, :870, :986)
• chat_workflow/cot_events.py:201 — chain-of-thought streaming
• api/agent.py:194 — agent events
• orchestrator.py:968, orchestration/workflow_runner.py:695
• services/approval_gate_service.py:362 — approval gates
• services/npu_worker_manager.py:902, services/load_balancer.py:649
• worker_node.py:298,399,415, diagnostics.py:173,204,265

The frontend's primary socket connects to exactly this endpoint:
autobot-frontend/src/services/GlobalWebSocketService.ts:146,153.

Reproduction

1. Open the GUI in two browser tabs (both establish /api/ws).
2. Trigger a workflow or a chat turn.
3. Only the most recently connected tab receives events.
4. Close that tab. The remaining tab now receives nothing, and new events are no longer written to
   chat history.

Proposed fix

• autobot-backend/event_manager.py — hold a set[Callable] instead of a scalar; turn
  register_websocket_broadcast into add/discard semantics; fan out to all registered callbacks in
  publish, removing any that raise a connection error.
• autobot-backend/events/bus.py — widen register_ws_broadcast to match (register/unregister
  rather than set/clear).
• autobot-backend/api/websockets.py:671-697 — register and unregister this connection's
  handler rather than overwriting a global.
• autobot-backend/api/websockets.py:459,510 — move _add_to_chat_history out of the broadcast
  handler so persistence no longer depends on a client being attached.

Acceptance criteria

• Two concurrent /api/ws clients both receive every published event.
• Disconnecting one client does not affect delivery to the others.
• Events are written to chat history when zero clients are connected.
• Regression test asserting all three, driving the production endpoint (not the helper) and
  asserting at the far boundary — a mutation to the fan-out must fail it.

Sources

Audit context and the full comparison: docs/research/agent-host-and-agent-client-protocols.md
(PR pending on branch research-ahp-acp).

• Agent Host Protocol — https://microsoft.github.io/agent-host-protocol/ ·
  https://github.com/microsoft/agent-host-protocol (MIT)
• AHP docs/guide/reconciliation.md and docs/specification/subscriptions.md describe the
  per-subscription fan-out model this endpoint is missing.