fix(conversation): recover dead ACP turns after agent process loss (#514)

## Summary
- Add ACP turn recovery policy support to automatically replay safe
retryable turns once after upstream agent/process failures.
- Evict unhealthy ACP tasks after recoverable terminal errors so later
stop/cancel/send flows bind to a fresh runtime instead of a dead task.
- Keep team child-turn status handling transparent across recovered
replays and cover recovery safety boundaries with focused tests.

## Test Plan
- [x] `just push -u origin feat/agent-turn-recoverable-error` (migration
immutability, cargo fix/fmt, clippy, and `cargo nextest run
--workspace`: 6379 passed, 18 skipped)
- [x] Local AionUi Dev solo smoke for Claude, Codex, and OpenCode: kill
ACP child process immediately after sending; subsequent stop/cancel/send
remains usable
- [x] Local AionUi Dev team smoke: kill ACP child process during a team
turn; child turn auto-recovers and the team run completes

---------

Co-authored-by: zynx <>

Author: piorpua

crates/aionui-ai-agent/src/protocol/send_error.rs

@@ -217,6 +217,10 @@ impl AgentSendError {
         self.stream_error
     }
 
+    pub fn from_stream_error_data(stream_error: AgentStreamErrorData) -> Self {
+        Self { stream_error }
+    }
+
     pub fn code(&self) -> Option<AgentErrorCode> {
         self.stream_error.code
     }

crates/aionui-conversation/src/agent_health_policy.rs

@@ -22,7 +22,10 @@ impl AgentHealthPolicy {
         outcome: &RelayOutcome,
         lifecycle: RuntimeLifecycleState,
     ) -> AgentHealthAction {
-        if lifecycle != RuntimeLifecycleState::Active {
+        if matches!(
+            lifecycle,
+            RuntimeLifecycleState::Deleting | RuntimeLifecycleState::ShuttingDown
+        ) {
             return AgentHealthAction::Keep;
         }
         if agent_type != AgentType::Acp || !outcome.terminal.is_error() {
@@ -45,11 +48,11 @@ mod tests {
 
     fn error_outcome(code: Option<AgentErrorCode>) -> RelayOutcome {
         RelayOutcome {
-            system_responses: Vec::new(),
             terminal: RelayTerminal::Error {
                 code,
                 retryable: Some(true),
             },
+            ..RelayOutcome::default()
         }
     }
 
@@ -98,15 +101,27 @@ mod tests {
     #[test]
     fn channel_closed_does_not_evict_by_default() {
         let outcome = RelayOutcome {
-            system_responses: Vec::new(),
             terminal: RelayTerminal::ChannelClosed,
+            ..RelayOutcome::default()
         };
         assert_eq!(
             AgentHealthPolicy::decide(AgentType::Acp, &outcome, RuntimeLifecycleState::Active),
             AgentHealthAction::Keep
         );
     }
 
+    #[test]
+    fn cancelling_acp_terminal_error_evicts_task() {
+        let outcome = error_outcome(Some(AgentErrorCode::UnknownUpstreamError));
+        assert!(matches!(
+            AgentHealthPolicy::decide(AgentType::Acp, &outcome, RuntimeLifecycleState::Cancelling),
+            AgentHealthAction::EvictAcpTask {
+                clear_model_seed: false,
+                ..
+            }
+        ));
+    }
+
     #[test]
     fn non_active_lifecycle_keeps_task() {
         let outcome = error_outcome(Some(AgentErrorCode::UserLlmProviderModelNotFound));

crates/aionui-conversation/src/lib.rs

@@ -24,6 +24,7 @@ pub mod stream_relay;
 pub mod task_options;
 mod turn_continuation_policy;
 mod turn_orchestrator;
+mod turn_recovery_policy;
 
 pub use error::ConversationError;
 pub use response_middleware::{