The Loop That Doesn't Listen
The orchestrator was wired up — DebateAgentInvoker, DebatePromptAssembler, DebateResponseBuilder, the autonomous flag, everything constructed and stored on the session. But nothing started it. converse() sat there, fully assembled, waiting for a trigger that didn’t exist.
The trigger belongs in DebateChannelBackend.post(). Every debate message on the channel passes through this method via Qhorus gateway fan-out — it already pushes WebSocket events and dispatches sub-agent requests via CDI. Adding the autonomous trigger here means any message source can kick off the orchestrator, not just the raise_point MCP tool.
The interesting part is the isolation model. ConversationOrchestrator.converse() runs an internal ArrayDeque — it seeds the queue with the triggering message, invokes agents, builds responses, dispatches them to the channel, and enqueues the responses for the next iteration. It’s a closed loop. Messages that arrive on the channel from outside — a human clicking FLAG_HUMAN in the UI, someone calling raise_point manually — go through messageService.dispatch() and reach DebateChannelBackend.post(), but they never enter the orchestrator’s queue. The dispatcher is output-only: responses go out to the channel, nothing comes back in.
This means external interruption can’t work by posting to the channel. You need orchestrator.terminate(), which sets a volatile boolean the loop checks each iteration. I wired FLAG_HUMAN detection in post() — when a FLAG_HUMAN arrives on an autonomous session, it calls terminate() directly. Same for endDebate().
The exactly-once trigger uses an AtomicBoolean on DebateSession — markConverseStarted() is a single compareAndSet(false, true). Multiple messages can arrive in the window between the first dispatch and the virtual thread starting; only the first wins the CAS race. The virtual thread itself is straightforward: converse() returns Uni<ConversationOutcome> but the Uni wraps a synchronous loop, so .await().indefinitely() on a virtual thread is the right subscription model.
ContestedEscalation was the other discovery worth noting. The name suggests it handles FLAG_HUMAN escalation — the spec originally assumed this. It doesn’t. It checks for points in DISPUTED status that have been disputed more than N times. It’s about agents that can’t agree, not about human intervention. Two different termination paths, easily conflated.