Durable assistant turns and realtime events

POST /api/chatbot/sessions/{sessionId}/messages persists a turn and returns 202 immediately. The response carries turnId, deadline, model ID, revision, and event cursor. Use GET /api/chatbot/turns/{turnId} for durable status, POST /cancel, /continue, /confirm, /undo, /rebase, /checkpoints/{checkpointId}/rollback, and /feedback for turn actions.

Realtime uses authenticated fetch-based SSE only:

GET /api/chatbot/turns/{turnId}/events
X-Auth-Token: <session token>
Last-Event-ID: <optional cursor>
Accept: text/event-stream

Every event contains an event ID, turn ID, sequence, type, timestamp, and safe payload. Common event types include turn.accepted, turn.stage, tool.started, tool.completed, turn.workflow.routed, turn.plan.ready, model.slice, model.checkpoint, model.checkpoint.committed, turn.validation.completed, turn.needs_input, turn.needs_confirmation, turn.completed, turn.failed, turn.cancelled, turn.feedback, and source-coverage progress payloads. Clients must poll turn status when SSE disconnects and must tolerate replayed events.

Status payloads may include coveragePercent, unresolved source units or obligations, provider-call/token counters, repair attempts, workflow phase, work items, and remainingWork. Reload the model after each model.checkpoint event. Run explicit model validation separately when EVL feedback is required; assistant checkpoints use structural validation only.

Provider token text streaming and WebSocket access are intentionally not supported. The UI exposes only factual stage/checkpoint/validation/coverage progress, never private reasoning.

Internal workflow events can identify routing and durable work-item progress, but they are observability data. Clients cannot select or override the assistant's internal conceptual/agent strategy through SSE or REST.

Conceptual work items and obligation-review evidence remain private until a complete atomic checkpoint is committed. SSE never exposes provider reasoning content.