fix: preserve terminal turn failures

crates/llm-worker/src/worker.rs

@@ -262,13 +262,17 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     }
 
     fn drain_cancel_queue(&mut self) {
-        use tokio::sync::mpsc::error::TryRecvError;
-        loop {
-            match self.cancel_rx.try_recv() {
-                Ok(()) => continue,
-                Err(TryRecvError::Empty) | Err(TryRecvError::Disconnected) => break,
-            }
-        }
+        while self.cancel_rx.try_recv().is_ok() {}
+    }
+
+    /// Discard pending cancellation notifications while the worker is idle.
+    ///
+    /// Cancellation is a running-turn control signal. Callers that own a higher
+    /// level run state can use this before starting a new turn so an old idle
+    /// signal does not poison the next request, while cancellation queued after
+    /// the run has been accepted remains observable by the turn loop.
+    pub fn clear_pending_cancel(&mut self) {
+        self.drain_cancel_queue();
     }
 
     fn try_cancelled(&mut self) -> bool {
@@ -1058,7 +1062,6 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
     /// Internal turn execution logic
     async fn run_turn_loop(&mut self) -> Result<WorkerResult, WorkerError> {
         self.reset_interruption_state();
-        self.drain_cancel_queue();
         let tool_definitions = self.build_tool_definitions();
 
         info!(
@@ -1363,9 +1366,27 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
                             "elapsed_ms": stream_started.elapsed().as_millis() as u64,
                         }),
                     );
-                    match wait_for_first_stream_event(stream, DEFAULT_FIRST_STREAM_EVENT_TIMEOUT)
-                        .await
-                    {
+                    let first_event_result = tokio::select! {
+                        first_event = wait_for_first_stream_event(stream, DEFAULT_FIRST_STREAM_EVENT_TIMEOUT) => first_event,
+                        cancel = self.cancel_rx.recv() => {
+                            if cancel.is_some() {
+                                info!("Cancelled before first stream event");
+                            }
+                            self.emit_lifecycle_trace(
+                                turn,
+                                llm_call,
+                                "stream_first_event_cancelled",
+                                json!({
+                                    "attempt": attempt,
+                                    "elapsed_ms": stream_started.elapsed().as_millis() as u64,
+                                }),
+                            );
+                            self.timeline.abort_current_block();
+                            self.last_run_interrupted = true;
+                            return Err(WorkerError::Cancelled);
+                        }
+                    };
+                    match first_event_result {
                         Ok(FirstStreamEvent::Ready(stream)) => return Ok(stream),
                         Ok(FirstStreamEvent::Empty(stream)) => return Ok(stream),
                         Err(err) => {
@@ -1502,6 +1523,17 @@ impl<C: LlmClient, S: WorkerState> Worker<C, S> {
                             }
                             self.emit_stream_event(turn, llm_call, &event);
                             self.timeline.dispatch(&event);
+                            if let Event::Error(err) = &event {
+                                self.timeline.abort_current_block();
+                                self.timeline.flush_usage();
+                                self.last_run_interrupted = true;
+                                return Err(WorkerError::Client(ClientError::Api {
+                                    status: None,
+                                    code: err.code.clone(),
+                                    message: err.message.clone(),
+                                    retry_after: None,
+                                }));
+                            }
                         }
                         None => break,
                     }

crates/pod/src/controller.rs

@@ -617,6 +617,11 @@ async fn controller_loop<C, St>(
         // here so the status flip → drive_turn → finish sequence lives
         // in one place, regardless of which Method caused it.
         if let Some(run) = pending.take() {
+            // Cancellation is meaningful only for an accepted running turn. Clear
+            // idle/stale signals before the status flip; any Cancel/Pause received
+            // after this point is delivered to the turn and must not be discarded by
+            // the Worker at run start.
+            pod.worker_mut().clear_pending_cancel();
             set_controller_status(&shared_state, &runtime_dir, &event_tx, PodStatus::Running).await;
             let parent_originated = run.is_parent_originated();
             let (new_status, shutdown) = match run {

crates/pod/tests/controller_test.rs

@@ -5,7 +5,7 @@ use std::sync::{Arc, Mutex};
 use async_trait::async_trait;
 use futures::{Stream, StreamExt};
 use llm_worker::Worker;
-use llm_worker::llm_client::event::{Event as LlmEvent, ResponseStatus, StatusEvent};
+use llm_worker::llm_client::event::{ErrorEvent, Event as LlmEvent, ResponseStatus, StatusEvent};
 use llm_worker::llm_client::types::Item;
 use llm_worker::llm_client::{ClientError, LlmClient, Request};
 use llm_worker::tool::{Tool, ToolDefinition, ToolError, ToolMeta, ToolOutput};
@@ -254,6 +254,52 @@ async fn run_end_returns_to_idle_without_busy_status() {
     assert_eq!(handle.shared_state.get_status(), PodStatus::Idle);
 }
 
+#[tokio::test]
+async fn provider_stream_error_records_run_errored() {
+    let client = MockClient::new(vec![LlmEvent::Error(ErrorEvent {
+        code: Some("context_length_exceeded".into()),
+        message: "request too large".into(),
+    })]);
+    let pod = make_pod(client).await;
+    let handle = spawn_controller(pod).await;
+    let mut rx = handle.subscribe();
+
+    handle.send(Method::run_text("ping")).await.unwrap();
+
+    assert!(
+        drain_until(&mut rx, std::time::Duration::from_secs(2), |e| matches!(
+            e,
+            Event::Error {
+                code: protocol::ErrorCode::ProviderError,
+                message,
+            } if message.contains("context_length_exceeded")
+        ))
+        .await,
+        "provider stream error should be surfaced as a live provider error"
+    );
+    wait_for_status(&handle, PodStatus::Idle).await;
+
+    let (entries, _rx) = handle.sink.subscribe_with_snapshot();
+    assert!(
+        entries.iter().any(|entry| matches!(
+            entry,
+            LogEntry::RunErrored { message, .. }
+                if message.contains("context_length_exceeded")
+        )),
+        "provider stream error should be persisted as RunErrored"
+    );
+    assert!(
+        !entries.iter().any(|entry| matches!(
+            entry,
+            LogEntry::RunCompleted {
+                result: llm_worker::WorkerResult::Finished,
+                ..
+            }
+        )),
+        "provider stream error must not be recorded as a finished run"
+    );
+}
+
 /// Mid-turn re-attach: a client connecting while the worker is still
 /// running observes the in-flight `UserInput` entry in the connect-time
 /// `Event::Snapshot`. This is the load-bearing property of the new

work-items/open/20260529-205844-session-pod-state-boundary/thread.md

@@ -17,3 +17,15 @@ Artifacts:
 - `artifacts/review-r2.md`
 
 ---
+
+<!-- event: fix author: insomnia at: 2026-05-30T00:08:00Z -->
+
+## Parent-side validation fix
+
+After merging the approved implementation, post-merge validation failed on `cargo test -p pod --test controller_test empty_turn_pause_rolls_back_and_snapshot_does_not_restore_input`.
+
+The parent took over the stopped/failed handoff and fixed the adjacent turn-control regression directly on main: cancellation received immediately after the controller accepts a run was being lost before the worker reached its first stream event wait, so empty turns could hang instead of rolling back. The fix preserves idle stale-cancel cleanup at the controller boundary and makes first-event waiting cancellation-aware.
+
+While investigating the child Pod's `context_length_exceeded` ping failure, the parent also fixed provider terminal stream errors so `Event::Error` is not only a live TUI event: terminal provider errors now fail the worker turn and persist `RunErrored` instead of allowing an empty `RunCompleted::Finished`.
+
+---