10 KiB
Tech spec: RTC task invalidation — Phase 1 (client-only)
Context
RTC invalidations for cloud agent tasks cause excessive GET /api/v1/agent/runs requests. During a bug bash with multiple concurrent agents on a team, this triggered 429 rate limiting that blocked spawning new agents.
Current flow
The server sends AmbientTaskUpdated { TaskId, Timestamp } over the websocket on every task state transition, session link update, conversation ID update, and task creation. The client receives this in the Listener → UpdateManager → AgentConversationsModel chain:
listener.rs:113-116—ObjectUpdateMessage::AmbientTaskUpdated { task_id, timestamp }arrives with both fields.update_manager.rs:1119-1125—handle_ambient_task_changeddiscardstask_id(param named_task_id) and emitsUpdateManagerEvent::AmbientTaskUpdated { timestamp }with only the timestamp.agent_conversations_model.rs:673-696—handle_update_manager_eventthrottles at 5s, then callsfetch_tasks_updated_after(timestamp)which hitsGET /api/v1/agent/runs?limit=100&updated_after={ts}— a list fetch of all recently-updated tasks.
Three consumer surfaces
- Conversation details panel (
terminal/view/ambient_agent/view_impl.rs:938-977): pane-level sidebar showing one task. Usesget_or_async_fetch_task_data(task_id)which hitsGET /agent/runs/{task_id}with per-task dedup. Not connected to RTC directly — free-rides on the list-fetch populatingself.tasks. Today this works because the list-fetch fires unconditionally on every RTC event. But once we gate the list-fetch on whether views are open (change 3), the details panel loses its data source and needs its own RTC path. - Agent management view (
workspace/view.rs:8023-8048): full-page dashboard. Shows all tasks (personal + team). Registers withregister_view_open/register_view_closed. - Conversation list view (
workspace/view/left_panel.rs:1030-1047): left panel sidebar. Shows personal tasks only (OwnerFilter::PersonalOnlyatconversation_list/view_model.rs:73). Also registers withregister_view_open/register_view_closed.
Problems
task_iddiscarded → forces broad list-fetch on every RTC event- Details panel has no direct RTC path → relying indirectly on
AgentConversationsModel. - Every RTC event triggers a list-fetch even if no list view is open
- No recovery if websocket misses a message (polling fully disabled when RTC is on,
agent_conversations_model.rs:981-983)
Out of scope: spawn.rs session polling
ambient_agents/spawn.rs has a separate polling loop (poll_run_until_joinable_session, spawn.rs:165-308) that polls GET /agent/runs/{task_id} every 3s (TASK_STATUS_POLL_INTERVAL, spawn.rs:23) to detect when a session becomes joinable. Not affected by these changes.
The tab IS registered in ActiveAgentViewsModel on TaskSpawned (model.rs:1261-1262), so the RTC handler (change 2a) will see has_open_tab = true and trigger redundant re-fetches during spawn. This is a minor inefficiency (~4-5 extra single-task requests per spawn) deferred for now — the big win is eliminating list-fetches.
RTC cannot replace spawn.rs because:
- spawn.rs drives the session state machine (
WaitingForSession→AgentRunning) by emittingAmbientAgentEvent::SessionStarted(spawn.rs:292-295), which triggers the shared session join (model.rs:1311-1346). RTC only refreshes cached task data. - spawn.rs handles timeouts, error/terminal states, followup stale-state skipping, and cancellation.
- spawn.rs extracts
SessionJoinInfo::from_task(spawn.rs:278) each poll; RTC events only carrytask_id+timestamp.
Relevant files
app/src/server/cloud_objects/listener.rs:113-116— websocket message type withtask_idapp/src/server/cloud_objects/update_manager.rs:1119-1126— discardstask_idapp/src/server/cloud_objects/update_manager.rs:137-142—UpdateManagerEventenumapp/src/ai/agent_conversations_model.rs:56—RTC_TASK_REFRESH_THROTTLE(5s)app/src/ai/agent_conversations_model.rs:673-696—handle_update_manager_eventapp/src/ai/agent_conversations_model.rs:735-768—fetch_tasks_updated_afterapp/src/ai/agent_conversations_model.rs:932-961—register_view_open/register_view_closedapp/src/ai/agent_conversations_model.rs:975-1001—should_be_pollingapp/src/ai/agent_conversations_model.rs:1519-1601—get_or_async_fetch_task_dataapp/src/ai/active_agent_views_model.rs:83-93— tracks focused conversations and ambient sessionsapp/src/terminal/view/ambient_agent/view_impl.rs:938-977— details panel data fetchapp/src/workspace/view/conversation_list/view_model.rs:68-91— personal-only filter
Proposed changes
1. Pass task_id through the event chain
In update_manager.rs, add task_id to the event:
// update_manager.rs:137-142
enum UpdateManagerEvent {
// ...
AmbientTaskUpdated { task_id: AmbientAgentTaskId, timestamp: DateTime<Utc> },
}
Rename _task_id → task_id in handle_ambient_task_changed and include it in the emitted event. This requires importing AmbientAgentTaskId in update_manager.rs.
2. Per-surface RTC dispatch in AgentConversationsModel
Replace the current handle_update_manager_event (which unconditionally list-fetches) with a dispatch that routes based on what's open.
New handle_update_manager_event logic:
fn handle_update_manager_event(&mut self, event: &UpdateManagerEvent, ctx: &mut ModelContext<Self>) {
let UpdateManagerEvent::AmbientTaskUpdated { task_id, timestamp } = event else {
return;
};
let has_list_consumers = self
.active_data_consumers_per_window
.values()
.any(|views| !views.is_empty());
if has_list_consumers {
// (a) List views: if management view or conversation list is open, do a throttled list-fetch.
self.handle_rtc_for_list_views(*timestamp, ctx);
} else {
let has_open_tab = ActiveAgentViewsModel::as_ref(ctx)
.get_terminal_view_id_for_ambient_task(*task_id)
.is_some();
if has_open_tab {
// (b) Details panel: if any window has this task focused, do a targeted single-task fetch.
// This still respects per-task dedup and failure cooldowns.
self.async_fetch_task(task_id, ctx);
} else {
// (c) No list surface or open tab: mark dirty for a later list refresh.
record_earliest_rtc_task_refresh_timestamp(&mut self.dirty_since, *timestamp);
}
}
}
2a. Open-tab check
Check ActiveAgentViewsModel for whether the task_id has an open ambient session tab:
let has_open_tab = ActiveAgentViewsModel::as_ref(ctx)
.get_terminal_view_id_for_ambient_task(*task_id)
.is_some();
This covers any window where the task is open in a tab (not just the focused window). It only runs when no list surface is open, so one RTC event does not trigger both a single-task fetch and a list-fetch.
2b. async_fetch_task
Call the shared task-fetch path which already has per-task dedup (TaskFetchState::InFlight), backoff for failures, and emits TasksUpdated on completion.
2c. handle_rtc_for_list_views
Extract the existing throttle logic from today's handle_update_manager_event into this method. Identical behavior to today — throttled list-fetch.
We keep the list-fetch (rather than batching single-task fetches) because: (1) the management view shows team tasks, so it needs to discover new tasks created by teammates — single-task fetches can only refresh known task_ids; (2) during bursts (20 tasks changing in a 5s window), 1 list-fetch is cheaper than 20 individual requests; (3) the big win is gating — not doing the list-fetch at all when the view isn't open.
3. Dirty-on-open flush
Add a dirty_since: Option<DateTime<Utc>> field to AgentConversationsModel.
When an RTC event arrives while no list surface is open and the task does not have an open tab, keep the earliest timestamp in dirty_since.
In register_view_open, after the existing logic, flush dirty state with one list refresh:
if let Some(dirty_since) = self.dirty_since.take() {
self.fetch_tasks_updated_after(dirty_since, ctx);
}
This keeps the closed-view path at zero requests, then performs one bounded list refresh when a list view becomes visible again.
Summary of request reduction
Before: every RTC event → 1 list-fetch (GET /agent/runs?updated_after=...), throttled to 1 per 5s.
After:
- Details panel open, no list surface open → 1 single-task fetch per event (deduped by
TaskFetchState) - Management/convo list open → same list-fetch as today (throttled)
- No list surface or open tab → 0 requests, dirty timestamp only
For a team of 10 running 5 agents with 4-5 state changes each: before = ~250 list-fetches across team in 5 min. After = ~0 list-fetches for users not looking at views, plus ~1 single-task fetch per user per agent they have open.
Testing and validation
Unit tests in agent_conversations_model_tests.rs:
- RTC event with
task_idwhen details panel has that task open and no list surface is open → targeted task refresh path used - RTC event when no list surface or open tab is present → earliest dirty timestamp recorded
register_view_openwithdirty_sinceset → onefetch_tasks_updated_aftercall
Manual verification:
- Add
log::info!("[lili] ...")inhandle_update_manager_eventto count fetches before/after - Run multiple cloud agents on a team, verify no list-fetches when management view is closed
- Open details panel for an agent, verify state changes appear within ~3s
- Close all views, let agents run, reopen management view → verify dirty tasks load
Parallelization
Not beneficial — all changes are in agent_conversations_model.rs and update_manager.rs with tight coupling between them. Single-agent serial implementation is the right approach.