fix(orchestrator): la délégation simple vers UX renvoie son Final structuré (#207, QA verte)

Une demande idea_ask_agent vers UX échouait avec « finished without a structured Final response » alors qu'Architect répondait. Deux causes corrigées:

- session/process: un tour CLI terminé en code non nul SANS stdout (ex. erreur 401/auth) renvoyait une sortie vide au lieu du stderr réel; l'orchestrateur concluait alors à un « no Final ». On remonte désormais le stderr effectif (drain sync + async + helper process_exit_message + test dédié).

- orchestrator/service: la branche « ask waiting on background » masquait le premier Final structuré; la réponse synchrone redevient le premier Final (suppression de active_rendezvous_has_background_tasks / mark_rendezvous_task_waiting).

Nettoyage runtime corollaire (#204): arrêt du live runtime stale (PTY + structuré) d'un agent recréé afin que la délégation route vers l'agent courant (LiveSessions, stop_live_agent_runtime_in_project, wiring backend via DeleteAgent::with_runtime_cleanup, lifecycle, workstate cascade).

crates/application/src/agent/lifecycle.rs

@@ -43,7 +43,10 @@ use crate::model_server::{
 };
 use crate::project::project_context_path;
 use crate::skill::{AgentCapability, AssignedSkillResolver, ResolveAgentCapabilities};
-use crate::terminal::{StructuredSessions, TerminalSessions};
+use crate::terminal::{
+    stop_live_agent_runtime_in_project, CloseTerminal, LiveSessions, StructuredSessions,
+    TerminalSessions,
+};
 use crate::workstate::GetLiveStateLean;
 
 /// Directory (relative to `.ideai/`) under which agent contexts are written.
@@ -1033,13 +1036,29 @@ pub struct DeleteAgentInput {
 pub struct DeleteAgent {
     contexts: Arc<dyn AgentContextStore>,
     events: Arc<dyn EventBus>,
+    runtime_cleanup: Option<(Arc<LiveSessions>, Arc<CloseTerminal>)>,
 }
 
 impl DeleteAgent {
     /// Builds the use case.
     #[must_use]
     pub fn new(contexts: Arc<dyn AgentContextStore>, events: Arc<dyn EventBus>) -> Self {
-        Self { contexts, events }
+        Self {
+            contexts,
+            events,
+            runtime_cleanup: None,
+        }
+    }
+
+    /// Wires live-runtime cleanup performed before the manifest entry is deleted.
+    #[must_use]
+    pub fn with_runtime_cleanup(
+        mut self,
+        live: Arc<LiveSessions>,
+        close: Arc<CloseTerminal>,
+    ) -> Self {
+        self.runtime_cleanup = Some((live, close));
+        self
     }
 
     /// Drops the manifest entry for the agent.
@@ -1058,6 +1077,11 @@ impl DeleteAgent {
         if entries.len() == before {
             return Err(AppError::NotFound(format!("agent {}", input.agent_id)));
         }
+        if let Some((live, close)) = &self.runtime_cleanup {
+            let _stopped =
+                stop_live_agent_runtime_in_project(&input.project, input.agent_id, live, close)
+                    .await?;
+        }
         let manifest = AgentManifest::new(manifest.version, entries)
             .map_err(|e| AppError::Invalid(e.to_string()))?;
         self.contexts

crates/application/src/error.rs

@@ -73,15 +73,14 @@ pub enum AppError {
         node_id: NodeId,
     },
 
-    /// The delegation target finished its turn and **returned to its prompt without
-    /// calling `idea_reply`**, so the rendezvous could produce no answer. Surfaced
-    /// promptly (after a short grace window) instead of blocking the caller until the
-    /// long safety timeout. **Retryable**: the caller may re-ask, ideally reminding the
-    /// target to render its result via `idea_reply` (never plain text). Carries the
-    /// target's display name.
+    /// The delegation target finished its structured/headless turn without producing
+    /// a [`ReplyEvent::Final`](domain::ports::ReplyEvent::Final), so the rendezvous
+    /// could not capture an inline answer. Surfaced promptly instead of blocking the
+    /// caller until the long safety timeout. **Retryable**: the caller may re-ask for a
+    /// concise final response. Carries the target's display name.
     #[error(
-        "agent {0} returned to its prompt without calling idea_reply (no answer was \
-         rendered); retry the request and ensure the agent replies via idea_reply"
+        "agent {0} finished without a structured Final response (no inline answer was \
+         captured); retry the request and ask for a concise final response"
     )]
     TargetReturnedNoReply(String),
 

crates/application/src/orchestrator/service.rs

@@ -6,7 +6,7 @@
 //! that turns a request into the *same* calls the UI makes:
 //!
 //! - `spawn_agent`     → [`CreateAgentFromScratch`] (if unknown) then [`LaunchAgent`],
-//! - `stop_agent`      → resolve the agent's live session, then [`CloseTerminal`],
+//! - `stop_agent`      → stop the agent's live PTY or structured runtime,
 //! - `update_agent_context` → [`UpdateAgentContext`].
 //!
 //! It talks **only** to use cases and ports ([`ProfileStore`], [`TerminalSessions`]):
@@ -53,7 +53,10 @@ use crate::orchestrator::{
     WriteMemoryInput,
 };
 use crate::skill::{CreateSkill, CreateSkillInput, ReadSkill, ReadSkillInput};
-use crate::terminal::{CloseTerminal, CloseTerminalInput, StructuredSessions, TerminalSessions};
+use crate::terminal::{
+    stop_live_agent_runtime_in_project, CloseTerminal, LiveSessions, StructuredSessions,
+    TerminalSessions,
+};
 use crate::workstate::{
     GetLiveStateLean, StopAgentExecutionCascade, UpdateLiveState, UpdateLiveStateInput,
 };
@@ -431,9 +434,9 @@ pub struct OrchestratorService {
     /// Séparé de [`WaitForGraph`] qui reste un objet domaine minimal de détection de
     /// cycle, sans API de traversal.
     active_waits: StdMutex<Vec<(AgentId, AgentId)>>,
-    /// Composite business rendezvous currently driven by a target agent. Used to
-    /// structurally attach `idea_run_in_background` tasks launched during
-    /// `idea_ask_agent` without parsing the model's textual `Final`.
+    /// Headless rendezvous currently driven by a target agent. Used to structurally
+    /// attach `idea_run_in_background` tasks for observability/cancellation while the
+    /// synchronous answer remains the first structured `Final`.
     active_rendezvous: StdMutex<HashMap<RuntimeAgentKey, ActiveRendezvous>>,
     /// Bus d'événements pour publier [`DomainEvent::AgentReplied`] à l'issue d'un
     /// `ask` réussi (§17.4). Injecté via [`Self::with_events`] ; `None` ⇒ pas de
@@ -751,19 +754,6 @@ impl OrchestratorService {
         }
     }
 
-    fn active_rendezvous_has_background_tasks(
-        &self,
-        agent: RuntimeAgentKey,
-        rendezvous_id: RendezvousId,
-    ) -> bool {
-        self.active_rendezvous
-            .lock()
-            .unwrap_or_else(std::sync::PoisonError::into_inner)
-            .get(&agent)
-            .filter(|active| active.link.rendezvous_id == rendezvous_id)
-            .is_some_and(|active| !active.background_tasks.is_empty())
-    }
-
     /// Returns the transitive set of agents currently waited on by `agent`.
     ///
     /// Used by user-driven cancellation: stopping A while A waits on B should also
@@ -893,27 +883,6 @@ impl OrchestratorService {
             .await
     }
 
-    async fn mark_rendezvous_task_waiting(&self, task_id: Option<TaskId>) -> Result<(), AppError> {
-        let (Some(store), Some(task_id)) = (&self.background_tasks, task_id) else {
-            return Ok(());
-        };
-        let task = store
-            .get(task_id)
-            .await
-            .map_err(|err| AppError::Store(err.to_string()))?
-            .ok_or_else(|| AppError::Store(format!("background task {task_id} not found")))?;
-        if task.is_terminal() || task.state == BackgroundTaskState::Waiting {
-            return Ok(());
-        }
-        let waiting = task
-            .transition(BackgroundTaskState::Waiting, self.now_ms())
-            .map_err(|err| AppError::Internal(err.to_string()))?;
-        store
-            .save(&waiting)
-            .await
-            .map_err(|err| AppError::Store(err.to_string()))
-    }
-
     async fn complete_rendezvous_task_failure(
         &self,
         project: &Project,
@@ -1963,7 +1932,7 @@ impl OrchestratorService {
         // Timeout de tour piloté par profil (lot 2) + armement du seuil de stall, AVANT
         // l'enqueue qui démarre le tour (le médiateur arme alors sa fenêtre de vivacité).
         let turn_timeout = self.turn_timeout_for(project, agent_id).await;
-        let pending = input.enqueue_silent(agent_key, ticket);
+        let _pending = input.enqueue_silent(agent_key, ticket);
         let rendezvous_link = BackgroundTaskRendezvousLink {
             rendezvous_id: RendezvousId::new_random(),
             ticket_id,
@@ -2071,7 +2040,7 @@ impl OrchestratorService {
         };
 
         // Borne par la fenêtre d'inactivité (réarmée sur signe de vie) sous plafond absolu.
-        let initial_result = match self
+        let result = match self
             .run_ask_with_watchdog(wait, turn_timeout, &project.root, agent_id, target, started)
             .await
         {
@@ -2163,64 +2132,6 @@ impl OrchestratorService {
             }
         };
 
-        let result = if self
-            .active_rendezvous_has_background_tasks(agent_key, rendezvous_link.rendezvous_id)
-        {
-            self.mark_rendezvous_task_waiting(rendezvous_task).await?;
-            crate::diag!(
-                "[rendezvous] ask waiting on background: target={target} \
-                 (agent {agent_id}) ticket={ticket_id} rendezvous={} first_final_len={}",
-                rendezvous_link.rendezvous_id,
-                initial_result.len(),
-            );
-            let remaining = self.ask_ceiling.saturating_sub(started.elapsed());
-            match tokio::time::timeout(remaining, pending).await {
-                Ok(Ok(domain::mailbox::TurnResolution::Replied(content))) => content,
-                Ok(Ok(domain::mailbox::TurnResolution::ReturnedToPromptNoReply)) => {
-                    let err = AppError::TargetReturnedNoReply(target.to_owned());
-                    self.complete_rendezvous_task_failure(
-                        project,
-                        agent_id,
-                        rendezvous_task,
-                        format!("NoReply: {err}"),
-                    )
-                    .await?;
-                    self.mark_target_done_best_effort(&project.root, agent_id, ticket_id)
-                        .await;
-                    return Err(err);
-                }
-                Ok(Err(err)) => {
-                    let err = AppError::Process(err.to_string());
-                    self.complete_rendezvous_task_failure(
-                        project,
-                        agent_id,
-                        rendezvous_task,
-                        format!("NoReply: {err}"),
-                    )
-                    .await?;
-                    self.mark_target_done_best_effort(&project.root, agent_id, ticket_id)
-                        .await;
-                    return Err(err);
-                }
-                Err(_elapsed) => {
-                    self.complete_rendezvous_task_failure(
-                        project,
-                        agent_id,
-                        rendezvous_task,
-                        format!(
-                            "Timeout: composite rendezvous ceiling reached for target {target}"
-                        ),
-                    )
-                    .await?;
-                    self.mark_target_done_best_effort(&project.root, agent_id, ticket_id)
-                        .await;
-                    return Err(AppError::TargetCeilingActive(target.to_owned()));
-                }
-            }
-        } else {
-            initial_result
-        };
-
         self.complete_rendezvous_task_success(project, agent_id, rendezvous_task, &result)
             .await?;
 
@@ -2665,7 +2576,7 @@ impl OrchestratorService {
                     .as_ref()
                     .and_then(|p| p.runtime_for(project, agent_id)),
                 allow_structured_alongside_pty: true,
-                require_structured: false,
+                require_structured: true,
             })
             .await?;
 
@@ -2774,7 +2685,7 @@ impl OrchestratorService {
         })
     }
 
-    /// `stop_agent`: translate the agent name → its live session → `CloseTerminal`.
+    /// `stop_agent`: translate the agent name → its live runtime → stop it.
     async fn stop_agent(
         &self,
         project: &Project,
@@ -2785,14 +2696,19 @@ impl OrchestratorService {
             .await?
             .ok_or_else(|| AppError::NotFound(format!("agent {name}")))?;
 
-        let session_id = self
-            .sessions
-            .session_for_agent_in_project(project.id, &agent_id)
-            .ok_or_else(|| AppError::NotFound(format!("running session for agent {name}")))?;
-
-        self.close_terminal
-            .execute(CloseTerminalInput { session_id })
-            .await?;
+        let structured = self
+            .structured
+            .clone()
+            .unwrap_or_else(|| Arc::new(StructuredSessions::new()));
+        let live = LiveSessions::new(Arc::clone(&self.sessions), structured);
+        if stop_live_agent_runtime_in_project(project, agent_id, &live, &self.close_terminal)
+            .await?
+            .is_empty()
+        {
+            return Err(AppError::NotFound(format!(
+                "running session for agent {name}"
+            )));
+        }
 
         Ok(OrchestratorOutcome {
             detail: format!("stopped agent {name}"),

crates/application/src/terminal/mod.rs

@@ -33,6 +33,7 @@ pub use registry::{
     TerminalSessions,
 };
 pub use usecases::{
-    CloseTerminal, CloseTerminalInput, CloseTerminalOutput, OpenTerminal, OpenTerminalInput,
-    OpenTerminalOutput, ResizeTerminal, ResizeTerminalInput, WriteToTerminal, WriteToTerminalInput,
+    stop_live_agent_runtime_in_project, CloseTerminal, CloseTerminalInput, CloseTerminalOutput,
+    OpenTerminal, OpenTerminalInput, OpenTerminalOutput, ResizeTerminal, ResizeTerminalInput,
+    StoppedAgentRuntime, WriteToTerminal, WriteToTerminalInput,
 };

crates/application/src/terminal/usecases.rs

@@ -4,13 +4,13 @@ use std::sync::Arc;
 
 use domain::ports::{EventBus, PtyPort, SpawnSpec};
 use domain::{
-    DomainEvent, NodeId, ProjectPath, PtySize, SessionId, SessionKind, SessionStatus,
-    TerminalSession,
+    AgentId, DomainEvent, NodeId, Project, ProjectPath, PtySize, SessionId, SessionKind,
+    SessionStatus, TerminalSession,
 };
 
 use crate::error::AppError;
 
-use super::registry::TerminalSessions;
+use super::registry::{LiveSessionKind, LiveSessions, TerminalSessions};
 
 /// Input for [`OpenTerminal::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -230,6 +230,59 @@ impl CloseTerminal {
     }
 }
 
+/// Runtime session stopped for one agent.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct StoppedAgentRuntime {
+    /// Agent whose live runtime was stopped.
+    pub agent_id: AgentId,
+    /// Live session id removed from its registry.
+    pub session_id: SessionId,
+    /// Runtime registry that owned the session.
+    pub kind: LiveSessionKind,
+}
+
+/// Stops one agent's live runtime across both registries.
+///
+/// PTY sessions are closed through [`CloseTerminal`]. Structured sessions are
+/// removed from their registry before `shutdown()` so relaunch guards stop seeing
+/// them as live even while shutdown runs.
+pub async fn stop_live_agent_runtime_in_project(
+    project: &Project,
+    agent_id: AgentId,
+    live: &LiveSessions,
+    close: &CloseTerminal,
+) -> Result<Vec<StoppedAgentRuntime>, AppError> {
+    let mut stopped = Vec::new();
+
+    if let Some(session_id) = live
+        .structured
+        .session_id_for_agent_in_project(project.id, &agent_id)
+    {
+        if let Some(session) = live.structured.remove(&session_id) {
+            session
+                .shutdown()
+                .await
+                .map_err(|e| AppError::Process(e.to_string()))?;
+        }
+        stopped.push(StoppedAgentRuntime {
+            agent_id,
+            session_id,
+            kind: LiveSessionKind::Structured,
+        });
+    }
+
+    if let Some(session_id) = live.pty.session_for_agent_in_project(project.id, &agent_id) {
+        close.execute(CloseTerminalInput { session_id }).await?;
+        stopped.push(StoppedAgentRuntime {
+            agent_id,
+            session_id,
+            kind: LiveSessionKind::Pty,
+        });
+    }
+
+    Ok(stopped)
+}
+
 /// The platform default interactive shell.
 ///
 /// Resolution policy lives in the application layer (a metier default), not the

crates/application/src/workstate/actions.rs

@@ -15,7 +15,9 @@ use domain::{AgentId, NodeId, Project, RuntimeAgentKey, SessionId};
 
 use crate::error::AppError;
 use crate::orchestrator::OrchestratorService;
-use crate::terminal::{CloseTerminal, CloseTerminalInput, LiveSessionKind, LiveSessions};
+use crate::terminal::{
+    stop_live_agent_runtime_in_project, CloseTerminal, LiveSessionKind, LiveSessions,
+};
 
 /// Input for [`AttachLiveAgent::execute`].
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -223,42 +225,15 @@ impl StopLiveAgent {
         project: &Project,
         agent_id: AgentId,
     ) -> Result<StopLiveAgentOutput, AppError> {
-        // PTY first: delegate to the existing close primitive (removes + kills).
-        if let Some(session_id) = self
-            .live
-            .pty
-            .session_for_agent_in_project(project.id, &agent_id)
-        {
-            self.close
-                .execute(CloseTerminalInput { session_id })
-                .await?;
-            return Ok(StopLiveAgentOutput {
-                agent_id,
-                session_id,
-                kind: LiveSessionKind::Pty,
-            });
-        }
-        // Structured: remove from the registry first (so the uniqueness guard no
-        // longer sees a live session), then shut the session down out of the lock.
-        if let Some(session_id) = self
-            .live
-            .structured
-            .session_id_for_agent_in_project(project.id, &agent_id)
-        {
-            if let Some(session) = self.live.structured.remove(&session_id) {
-                session
-                    .shutdown()
-                    .await
-                    .map_err(|e| AppError::Process(e.to_string()))?;
-            }
-            return Ok(StopLiveAgentOutput {
-                agent_id,
-                session_id,
-                kind: LiveSessionKind::Structured,
-            });
-        }
-        Err(AppError::NotFound(format!(
-            "running session for agent {agent_id}"
-        )))
+        stop_live_agent_runtime_in_project(project, agent_id, &self.live, &self.close)
+            .await?
+            .into_iter()
+            .next()
+            .map(|stopped| StopLiveAgentOutput {
+                agent_id: stopped.agent_id,
+                session_id: stopped.session_id,
+                kind: stopped.kind,
+            })
+            .ok_or_else(|| AppError::NotFound(format!("running session for agent {agent_id}")))
     }
 }

crates/application/tests/agent_lifecycle.rs

@@ -47,10 +47,11 @@ use domain::{PtySize, SessionId, SkillId, SkillRef};
 use uuid::Uuid;
 
 use application::{
-    CreateAgentFromScratch, CreateAgentInput, DeleteAgent, DeleteAgentInput, LaunchAgent,
-    LaunchAgentInput, ListAgents, ListAgentsInput, PermissionProjectorRegistry, ReadAgentContext,
-    ReadAgentContextInput, StructuredRoutingMode, StructuredSessions, TerminalSessions,
-    UpdateAgentContext, UpdateAgentContextInput, UpdateAgentEffort, UpdateAgentEffortInput,
+    CloseTerminal, CreateAgentFromScratch, CreateAgentInput, DeleteAgent, DeleteAgentInput,
+    LaunchAgent, LaunchAgentInput, ListAgents, ListAgentsInput, LiveSessions,
+    PermissionProjectorRegistry, ReadAgentContext, ReadAgentContextInput, StructuredRoutingMode,
+    StructuredSessions, TerminalSessions, UpdateAgentContext, UpdateAgentContextInput,
+    UpdateAgentEffort, UpdateAgentEffortInput,
 };
 
 // ---------------------------------------------------------------------------
@@ -585,6 +586,7 @@ impl PtyPort for FakePty {
 struct FakeSession {
     id: SessionId,
     conversation_id: Option<String>,
+    shutdowns: Arc<Mutex<Vec<SessionId>>>,
 }
 
 #[async_trait]
@@ -602,6 +604,7 @@ impl AgentSession for FakeSession {
     }
 
     async fn shutdown(&self) -> Result<(), AgentSessionError> {
+        self.shutdowns.lock().unwrap().push(self.id);
         Ok(())
     }
 }
@@ -613,6 +616,7 @@ struct FakeStructuredFactory {
     efforts: Arc<Mutex<Vec<Option<String>>>>,
     envs: Arc<Mutex<Vec<Vec<(String, String)>>>>,
     policies: Arc<Mutex<Vec<Option<domain::ports::StructuredProviderLaunchPolicy>>>>,
+    shutdowns: Arc<Mutex<Vec<SessionId>>>,
     next_session: SessionId,
 }
 
@@ -624,6 +628,7 @@ impl FakeStructuredFactory {
             efforts: Arc::new(Mutex::new(Vec::new())),
             envs: Arc::new(Mutex::new(Vec::new())),
             policies: Arc::new(Mutex::new(Vec::new())),
+            shutdowns: Arc::new(Mutex::new(Vec::new())),
             next_session,
         }
     }
@@ -679,6 +684,7 @@ impl AgentSessionFactory for FakeStructuredFactory {
         Ok(Arc::new(FakeSession {
             id: self.next_session,
             conversation_id: Some("engine-session-1".to_owned()),
+            shutdowns: Arc::clone(&self.shutdowns),
         }))
     }
 }
@@ -1048,6 +1054,56 @@ async fn delete_removes_entry_then_unknown_is_not_found() {
     assert_eq!(err.code(), "NOT_FOUND", "got {err:?}");
 }
 
+#[tokio::test]
+async fn delete_agent_cleans_structured_then_pty_runtime_before_manifest_removal() {
+    let a = scratch_agent(aid(1), "Backend", "agents/backend.md", pid(9));
+    let contexts = FakeContexts::with_agent(&a, "ctx");
+    let sessions = Arc::new(TerminalSessions::new());
+    let structured = Arc::new(StructuredSessions::new());
+    let live = Arc::new(LiveSessions::new(
+        Arc::clone(&sessions),
+        Arc::clone(&structured),
+    ));
+    let pty = FakePty::new(trace(), sid(42));
+    let close = Arc::new(CloseTerminal::new(
+        Arc::new(pty.clone()),
+        Arc::clone(&sessions),
+    ));
+    let shutdowns = Arc::new(Mutex::new(Vec::new()));
+    structured.insert_in_project(
+        project().id,
+        Arc::new(FakeSession {
+            id: sid(555),
+            conversation_id: Some("engine-session-stale".to_owned()),
+            shutdowns: Arc::clone(&shutdowns),
+        }),
+        a.id,
+        nid(7),
+    );
+    seed_live_agent_session(&sessions, a.id, nid(8), sid(42));
+
+    let delete = DeleteAgent::new(Arc::new(contexts.clone()), Arc::new(SpyBus::default()))
+        .with_runtime_cleanup(live, close);
+
+    delete
+        .execute(DeleteAgentInput {
+            project: project(),
+            agent_id: a.id,
+        })
+        .await
+        .unwrap();
+
+    assert_eq!(*shutdowns.lock().unwrap(), vec![sid(555)]);
+    assert_eq!(pty.kills(), vec![sid(42)]);
+    assert!(structured
+        .session_for_agent_in_project(project().id, &a.id)
+        .is_none());
+    assert!(sessions
+        .session_for_agent_in_project(project().id, &a.id)
+        .is_none());
+    assert!(contexts.manifest().entries.is_empty());
+}
+
 // ---------------------------------------------------------------------------
 // LaunchAgent
 // ---------------------------------------------------------------------------
@@ -1369,6 +1425,7 @@ async fn chat_intent_reattaches_existing_structured_session_after_stale_pty_tran
         Arc::new(FakeSession {
             id: sid(555),
             conversation_id: Some("engine-session-stale".to_owned()),
+            shutdowns: Arc::new(Mutex::new(Vec::new())),
         }),
         agent.id,
         stale_custom_node,

crates/application/tests/orchestrator_service.rs

@@ -390,6 +390,10 @@ impl RecordingBackgroundStore {
     fn task(&self, id: TaskId) -> Option<BackgroundTask> {
         self.tasks.lock().unwrap().get(&id).cloned()
     }
+
+    fn tasks(&self) -> Vec<BackgroundTask> {
+        self.tasks.lock().unwrap().values().cloned().collect()
+    }
 }
 #[async_trait]
 impl BackgroundTaskStore for RecordingBackgroundStore {
@@ -1009,6 +1013,41 @@ async fn stop_agent_kills_the_right_session() {
     assert!(fx.sessions.session(&sid(777)).is_none());
 }
 
+#[tokio::test]
+async fn stop_agent_stops_structured_session() {
+    let agent = scratch_agent(aid(1), "dev-backend", "agents/dev-backend.md");
+    let contexts = FakeContexts::with_agent(&agent, "# persona");
+    let fx = ask_fixture(contexts);
+    let session = Arc::new(CompletionSession {
+        id: sid(9010),
+        agent: agent.id,
+        completions: fx.mailbox.completions(),
+    }) as Arc<dyn AgentSession>;
+    fx.structured
+        .insert_in_project(project_id(), session, agent.id, nid(9010));
+    assert!(fx
+        .structured
+        .session_for_agent_in_project(project_id(), &agent.id)
+        .is_some());
+
+    fx.service
+        .dispatch(
+            &project(),
+            cmd(r#"{ "action":"stop_agent", "name":"dev-backend" }"#),
+        )
+        .await
+        .expect("stop ok");
+
+    assert!(
+        fx.pty.kills().is_empty(),
+        "structured stop must not kill PTY"
+    );
+    assert!(fx
+        .structured
+        .session_for_agent_in_project(project_id(), &agent.id)
+        .is_none());
+}
+
 #[tokio::test]
 async fn stop_agent_without_live_session_is_not_found() {
     let agent = scratch_agent(aid(1), "dev-backend", "agents/dev-backend.md");
@@ -1025,6 +1064,47 @@ async fn stop_agent_without_live_session_is_not_found() {
     assert_eq!(err.code(), "NOT_FOUND", "got {err:?}");
 }
 
+#[tokio::test]
+async fn ask_agent_after_delete_recreate_same_name_uses_new_agent_id() {
+    let old = scratch_agent(aid(1), "dev-backend", "agents/dev-backend.md");
+    let new = scratch_agent(aid(2), "dev-backend", "agents/dev-backend-2.md");
+    let contexts = FakeContexts::with_agent(&new, "# recreated persona");
+    let fx = ask_fixture(contexts);
+
+    let old_session = Arc::new(CompletionSession {
+        id: sid(9001),
+        agent: old.id,
+        completions: fx.mailbox.completions(),
+    }) as Arc<dyn AgentSession>;
+    fx.structured
+        .insert_in_project(project_id(), old_session, old.id, nid(9001));
+    fx.mailbox
+        .completions()
+        .push(old.id, TestCompletion::Replied("old-agent".to_owned()));
+    fx.mailbox
+        .completions()
+        .push(new.id, TestCompletion::Replied("new-agent".to_owned()));
+
+    let out = fx
+        .service
+        .dispatch(
+            &project(),
+            cmd(r#"{ "type":"agent.message", "targetAgent":"dev-backend", "task":"ping" }"#),
+        )
+        .await
+        .unwrap();
+
+    assert_eq!(out.reply.as_deref(), Some("new-agent"));
+    assert!(fx
+        .structured
+        .session_for_agent_in_project(project_id(), &old.id)
+        .is_some());
+    assert!(fx
+        .structured
+        .session_for_agent_in_project(project_id(), &new.id)
+        .is_some());
+}
+
 #[tokio::test]
 async fn update_agent_context_overwrites_md() {
     let agent = scratch_agent(aid(1), "dev-backend", "agents/dev-backend.md");
@@ -4198,6 +4278,96 @@ impl AgentSession for RateLimitedSession {
     }
 }
 
+struct EmptyStructuredSession {
+    id: SessionId,
+}
+
+#[async_trait]
+impl AgentSession for EmptyStructuredSession {
+    fn id(&self) -> SessionId {
+        self.id
+    }
+
+    fn conversation_id(&self) -> Option<String> {
+        Some(format!("engine-{}", self.id))
+    }
+
+    async fn send(&self, _prompt: &str) -> Result<ReplyStream, AgentSessionError> {
+        Ok(Box::new(Vec::<ReplyEvent>::new().into_iter()))
+    }
+
+    async fn shutdown(&self) -> Result<(), AgentSessionError> {
+        Ok(())
+    }
+}
+
+struct FailingStructuredSession {
+    id: SessionId,
+    message: String,
+}
+
+#[async_trait]
+impl AgentSession for FailingStructuredSession {
+    fn id(&self) -> SessionId {
+        self.id
+    }
+
+    fn conversation_id(&self) -> Option<String> {
+        Some(format!("engine-{}", self.id))
+    }
+
+    async fn send(&self, _prompt: &str) -> Result<ReplyStream, AgentSessionError> {
+        Err(AgentSessionError::Io(self.message.clone()))
+    }
+
+    async fn shutdown(&self) -> Result<(), AgentSessionError> {
+        Ok(())
+    }
+}
+
+struct BackgroundThenFinalSession {
+    id: SessionId,
+    service: Arc<OrchestratorService>,
+}
+
+#[async_trait]
+impl AgentSession for BackgroundThenFinalSession {
+    fn id(&self) -> SessionId {
+        self.id
+    }
+
+    fn conversation_id(&self) -> Option<String> {
+        Some(format!("engine-{}", self.id))
+    }
+
+    async fn send(&self, _prompt: &str) -> Result<ReplyStream, AgentSessionError> {
+        self.service
+            .dispatch(
+                &project(),
+                OrchestratorCommand::RunInBackground {
+                    owner: aid(1),
+                    label: "correlated child".to_owned(),
+                    command: "bash".to_owned(),
+                    args: vec!["-lc".to_owned(), "sleep 60".to_owned()],
+                    cwd: None,
+                    deadline_ms: None,
+                },
+            )
+            .await
+            .map_err(|err| AgentSessionError::Io(err.to_string()))?;
+        Ok(Box::new(
+            vec![ReplyEvent::Final {
+                content: "business final".to_owned(),
+            }]
+            .into_iter(),
+        ))
+    }
+
+    async fn shutdown(&self) -> Result<(), AgentSessionError> {
+        Ok(())
+    }
+}
+
 #[derive(Default)]
 struct NoopScheduler {
     next: Mutex<u128>,
@@ -4241,9 +4411,24 @@ struct StructuredAskFixture {
     pty: FakePty,
     bus: SpyBus,
     session_limits: Arc<SessionLimitService>,
+    background_store: Arc<RecordingBackgroundStore>,
+    background_runner: Arc<RecordingBackgroundRunner>,
 }
 
 fn structured_ask_fixture(contexts: FakeContexts) -> StructuredAskFixture {
+    structured_ask_fixture_with_background(contexts, false)
+}
+
+fn structured_ask_fixture_with_observable_background(
+    contexts: FakeContexts,
+) -> StructuredAskFixture {
+    structured_ask_fixture_with_background(contexts, true)
+}
+
+fn structured_ask_fixture_with_background(
+    contexts: FakeContexts,
+    enable_background: bool,
+) -> StructuredAskFixture {
     let profiles = Arc::new(FakeProfiles::new(vec![claude_profile()]));
     let sessions = Arc::new(TerminalSessions::new());
     let pty = FakePty::new(sid(777));
@@ -4251,6 +4436,15 @@ fn structured_ask_fixture(contexts: FakeContexts) -> StructuredAskFixture {
     let mailbox = Arc::new(TestMailbox::new());
     let structured = Arc::new(StructuredSessions::new());
     let factory = CountingFactory::new();
+    let background_store = Arc::new(RecordingBackgroundStore::default());
+    let background_runner = Arc::new(RecordingBackgroundRunner::default());
+    let background_clock = Arc::new(FixedMillisClock(42)) as Arc<dyn domain::ports::Clock>;
+    let spawn_background = Arc::new(SpawnBackgroundCommand::new(
+        Arc::clone(&background_store) as Arc<dyn BackgroundTaskStore>,
+        Arc::clone(&background_runner) as Arc<dyn BackgroundTaskRunner>,
+        Arc::clone(&background_clock),
+        Arc::new(SeqIds::new()) as Arc<dyn IdGenerator>,
+    ));
 
     let create = Arc::new(CreateAgentFromScratch::new(
         Arc::new(contexts.clone()),
@@ -4301,8 +4495,7 @@ fn structured_ask_fixture(contexts: FakeContexts) -> StructuredAskFixture {
         Arc::new(NoopResumer),
     ));
 
-    let service = Arc::new(
-        OrchestratorService::new(
+    let mut service = OrchestratorService::new(
         create,
         launch,
         list,
@@ -4321,8 +4514,20 @@ fn structured_ask_fixture(contexts: FakeContexts) -> StructuredAskFixture {
     .with_session_limits(Arc::clone(&session_limits))
     // Même registre que le launcher : `ask_agent` y relit la session fraîchement
     // insérée par `launch_structured`.
-        .with_structured(Arc::clone(&structured)),
+    .with_structured(Arc::clone(&structured));
+
+    if enable_background {
+        service = service
+            .with_background_tasks(
+                Arc::clone(&background_store) as Arc<dyn BackgroundTaskStore>,
+                Arc::clone(&background_clock),
+            )
+            .with_spawn_background_command(spawn_background)
+            .with_background_task_runner(
+                Arc::clone(&background_runner) as Arc<dyn BackgroundTaskRunner>
             );
+    }
+    let service = Arc::new(service);
 
     StructuredAskFixture {
         service,
@@ -4332,6 +4537,8 @@ fn structured_ask_fixture(contexts: FakeContexts) -> StructuredAskFixture {
         pty,
         bus,
         session_limits,
+        background_store,
+        background_runner,
     }
 }
 
@@ -4446,6 +4653,138 @@ async fn ask_structured_target_live_as_pty_keeps_visible_terminal() {
     );
 }
 
+/// Régression #201 : une tâche de fond lancée pendant le tour headless reste corrélée
+/// au rendez-vous pour l'observabilité, mais elle ne prolonge plus la réponse
+/// synchrone. Le premier `Final` structured est le succès métier.
+#[tokio::test]
+async fn ask_structured_final_wins_even_with_attached_background_task() {
+    let agent = scratch_agent(aid(1), "architect", "agents/architect.md");
+    let fx = structured_ask_fixture_with_observable_background(FakeContexts::with_agent(
+        &agent,
+        "# persona",
+    ));
+    fx.structured.insert_in_project(
+        project_id(),
+        Arc::new(BackgroundThenFinalSession {
+            id: sid(9800),
+            service: Arc::clone(&fx.service),
+        }),
+        aid(1),
+        nid(700),
+    );
+
+    let out = fx
+        .service
+        .dispatch(&project(), cmd(ASK_JSON))
+        .await
+        .expect("Final must win even when a background task is attached");
+
+    assert_eq!(out.reply.as_deref(), Some("business final"));
+    assert_eq!(
+        fx.background_runner.specs().len(),
+        1,
+        "the child background command is still spawned for observability"
+    );
+
+    let tasks = fx.background_store.tasks();
+    let rendezvous_task = tasks
+        .iter()
+        .find(|task| {
+            matches!(
+                task.kind,
+                domain::BackgroundTaskKind::HeadlessRendezvous { .. }
+            )
+        })
+        .expect("headless rendezvous task persisted");
+    assert_eq!(
+        rendezvous_task.state,
+        BackgroundTaskState::Completed,
+        "the rendezvous itself is marked successful inline"
+    );
+
+    let child_task = tasks
+        .iter()
+        .find(|task| matches!(task.kind, domain::BackgroundTaskKind::Command { .. }))
+        .expect("attached command task persisted");
+    assert_eq!(
+        child_task.state,
+        BackgroundTaskState::Running,
+        "the attached background task remains observable without gating the reply"
+    );
+    assert!(
+        child_task.rendezvous.is_some(),
+        "the child task keeps its rendezvous correlation"
+    );
+}
+
+/// Régression #201 : si le flux structuré se clôt sans `Final`, l'erreur publique
+/// décrit l'absence de réponse inline exploitable, sans mentionner l'ancien protocole
+/// `idea_reply`.
+#[tokio::test]
+async fn ask_structured_closed_stream_without_final_returns_inline_no_reply_error() {
+    let agent = scratch_agent(aid(1), "architect", "agents/architect.md");
+    let fx = structured_ask_fixture(FakeContexts::with_agent(&agent, "# persona"));
+    fx.structured.insert_in_project(
+        project_id(),
+        Arc::new(EmptyStructuredSession { id: sid(9801) }),
+        aid(1),
+        nid(700),
+    );
+
+    let err = fx
+        .service
+        .dispatch(&project(), cmd(ASK_JSON))
+        .await
+        .expect_err("closed stream without Final must be a typed no-reply");
+
+    assert_eq!(err.code(), "TARGET_RETURNED_NO_REPLY");
+    let message = err.to_string();
+    assert!(
+        message.contains("structured Final response"),
+        "message must explain the missing inline Final, got: {message}"
+    );
+    assert!(
+        !message.contains("idea_reply"),
+        "message must not mention the legacy reply tool, got: {message}"
+    );
+}
+
+/// Régression #207 : si la CLI structurée échoue avant tout stdout exploitable
+/// (ex. `401 Unauthorized`), l'orchestrateur doit remonter l'erreur réelle en
+/// `PROCESS` au lieu de maquiller l'échec en faux no-reply.
+#[tokio::test]
+async fn ask_structured_session_error_surfaces_process_message_not_no_reply() {
+    let agent = scratch_agent(aid(1), "architect", "agents/architect.md");
+    let fx = structured_ask_fixture(FakeContexts::with_agent(&agent, "# persona"));
+    fx.structured.insert_in_project(
+        project_id(),
+        Arc::new(FailingStructuredSession {
+            id: sid(9802),
+            message: "codex exec exited with code 1: 401 Unauthorized: Missing bearer"
+                .to_owned(),
+        }),
+        aid(1),
+        nid(700),
+    );
+
+    let err = fx
+        .service
+        .dispatch(&project(), cmd(ASK_JSON))
+        .await
+        .expect_err("session startup failure must surface as a process error");
+
+    assert_eq!(err.code(), "PROCESS");
+    let message = err.to_string();
+    assert!(
+        message.contains("401 Unauthorized: Missing bearer"),
+        "public error must preserve stderr detail, got: {message}"
+    );
+    assert!(
+        !message.contains("structured Final response"),
+        "public error must not fall back to no-reply wording, got: {message}"
+    );
+}
+
 /// Sous-lot adjacent #7/F2 : une limite détectée pendant un rendez-vous structuré
 /// doit aussi passer par `SessionLimitService::on_rate_limited` pour la cible, afin
 /// de publier le flux mono-agent `AgentRateLimited` puis `AgentResumeScheduled`.

crates/backend/src/lib.rs

@@ -1514,6 +1514,10 @@ impl BackendCore {
         // `profile.structured_adapter`. Injectés dans LaunchAgent (routage §17.4) et
         // ChangeAgentProfile (shutdown polymorphe au hot-swap).
         let structured_sessions = Arc::new(StructuredSessions::new());
+        let live_sessions = Arc::new(LiveSessions::new(
+            Arc::clone(&terminal_sessions),
+            Arc::clone(&structured_sessions),
+        ));
         let openai_tool_invoker = Arc::new(LateBoundOpenAiToolInvoker::new());
         let openai_tool_invoker_port = Arc::clone(&openai_tool_invoker) as Arc<dyn ToolInvoker>;
         let session_factory = Arc::new(
@@ -2065,10 +2069,10 @@ impl BackendCore {
         ));
         let read_agent_context = Arc::new(ReadAgentContext::new(Arc::clone(&contexts_port)));
         let update_agent_context = Arc::new(UpdateAgentContext::new(Arc::clone(&contexts_port)));
-        let delete_agent = Arc::new(DeleteAgent::new(
-            Arc::clone(&contexts_port),
-            Arc::clone(&events_port),
-        ));
+        let delete_agent = Arc::new(
+            DeleteAgent::new(Arc::clone(&contexts_port), Arc::clone(&events_port))
+                .with_runtime_cleanup(Arc::clone(&live_sessions), Arc::clone(&close_terminal)),
+        );
         // LaunchAgent shares the SAME pty_port and terminal_sessions as the terminal
         // use cases — indispensable for transport bridges to work correctly.
         //
@@ -2661,10 +2665,6 @@ impl BackendCore {
                 }
             });
         }
-        let live_sessions = Arc::new(LiveSessions::new(
-            Arc::clone(&terminal_sessions),
-            Arc::clone(&structured_sessions),
-        ));
         let delete_model_artifact = Arc::new(DeleteModelArtifact::new(
             Arc::clone(&model_server_registry_port),
             Arc::clone(&model_server_probe_port),

crates/infrastructure/src/session/mod.rs

@@ -186,6 +186,58 @@ mod tests {
         assert!(matches!(err, AgentSessionError::Start(_)), "vu: {err:?}");
     }
 
+    #[tokio::test]
+    async fn run_turn_nonzero_without_stdout_surfaces_stderr() {
+        use std::io::Write as _;
+        use std::sync::atomic::{AtomicU64, Ordering};
+
+        static C: AtomicU64 = AtomicU64::new(0);
+        let mut path = std::env::temp_dir();
+        path.push(format!(
+            "idea-empty-failing-cli-{}-{}",
+            std::process::id(),
+            C.fetch_add(1, Ordering::Relaxed)
+        ));
+        {
+            let mut file = std::fs::File::create(&path).expect("write fake cli");
+            file.write_all(
+                b"#!/bin/sh\nprintf '%s\\n' '401 Unauthorized: Missing bearer' >&2\nexit 1\n",
+            )
+            .expect("write fake cli");
+            file.sync_all().expect("sync fake cli");
+        }
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            let mut perms = std::fs::metadata(&path).unwrap().permissions();
+            perms.set_mode(0o755);
+            std::fs::set_permissions(&path, perms).unwrap();
+        }
+
+        let spec = super::process::SpawnLine {
+            command: path.to_string_lossy().into_owned(),
+            args: Vec::new(),
+            cwd: "/".to_owned(),
+            env: Vec::new(),
+            stdin: None,
+            sandbox: None,
+        };
+
+        let err = run_turn(&spec, None, None, None)
+            .await
+            .expect_err("empty nonzero process must be a session error");
+        match err {
+            AgentSessionError::Io(message) => {
+                assert!(
+                    message.contains("401 Unauthorized: Missing bearer"),
+                    "stderr must be preserved, got: {message}"
+                );
+            }
+            other => panic!("expected Io, got {other:?}"),
+        }
+        let _ = std::fs::remove_file(&path);
+    }
+
     // -- parse_event Claude (format RÉEL vérifié 2026-06-09) --------------
 
     #[test]

crates/infrastructure/src/session/process.rs

@@ -32,7 +32,7 @@ use std::sync::mpsc::Sender;
 use std::sync::Arc;
 use std::time::Duration;
 
-use tokio::io::{AsyncBufReadExt, AsyncWriteExt, BufReader};
+use tokio::io::{AsyncBufReadExt, AsyncReadExt, AsyncWriteExt, BufReader};
 use tokio::process::Command;
 
 use domain::ports::AgentSessionError;
@@ -219,7 +219,7 @@ fn drain_sandboxed(
     killer_tx: tokio::sync::oneshot::Sender<Arc<std::sync::Mutex<std::process::Child>>>,
     line_tap: Option<Sender<String>>,
 ) -> Result<Vec<String>, AgentSessionError> {
-    use std::io::{BufRead, BufReader as StdBufReader, Write as _};
+    use std::io::{BufRead, BufReader as StdBufReader, Read as _, Write as _};
     use std::process::{Command as StdCommand, Stdio};
     use std::sync::Mutex as StdMutex;
 
@@ -266,6 +266,17 @@ fn drain_sandboxed(
         .stdout
         .take()
         .ok_or_else(|| AgentSessionError::Io("stdout pipe indisponible".to_owned()))?;
+    let stderr = child
+        .stderr
+        .take()
+        .ok_or_else(|| AgentSessionError::Io("stderr pipe indisponible".to_owned()))?;
+    let stderr_handle = std::thread::spawn(move || {
+        let mut bytes = Vec::new();
+        StdBufReader::new(stderr)
+            .read_to_end(&mut bytes)
+            .map(|_| bytes)
+            .map_err(|e| AgentSessionError::Io(e.to_string()))
+    });
     let child = Arc::new(StdMutex::new(child));
     // Donne au côté async de quoi tuer l'enfant en cas de timeout.
     let _ = killer_tx.send(Arc::clone(&child));
@@ -285,8 +296,20 @@ fn drain_sandboxed(
     }
 
     // 6. Reap (pas de zombie). Lock tenu brièvement.
-    if let Ok(mut c) = child.lock() {
-        let _ = c.wait();
+    let status = child
+        .lock()
+        .map_err(|_| AgentSessionError::Io("child lock empoisonné".to_owned()))?
+        .wait()
+        .map_err(|e| AgentSessionError::Io(e.to_string()))?;
+    let stderr_bytes = stderr_handle
+        .join()
+        .map_err(|_| AgentSessionError::Io("stderr thread terminé sans résultat".to_owned()))??;
+    if !status.success() && collected.is_empty() {
+        return Err(AgentSessionError::Io(process_exit_message(
+            &command,
+            status.code(),
+            &stderr_bytes,
+        )));
     }
     Ok(collected)
 }
@@ -333,6 +356,14 @@ async fn drain(
         .stdout
         .take()
         .ok_or_else(|| AgentSessionError::Io("stdout pipe indisponible".to_owned()))?;
+    let mut stderr = child
+        .stderr
+        .take()
+        .ok_or_else(|| AgentSessionError::Io("stderr pipe indisponible".to_owned()))?;
+    let stderr_task = tokio::spawn(async move {
+        let mut bytes = Vec::new();
+        stderr.read_to_end(&mut bytes).await.map(|_| bytes)
+    });
     let mut lines = BufReader::new(stdout).lines();
 
     let mut collected = Vec::new();
@@ -350,9 +381,21 @@ async fn drain(
     }
 
     // Réifie le statut du process pour ne pas laisser de zombie ; un statut non nul
-    // n'est pas en soi une erreur de session (la CLI peut signaler une erreur
-    // *dans* sa sortie structurée, captée par le parser).
-    let _ = child.wait().await.map_err(map_io)?;
+    // n'est pas en soi une erreur de session si la CLI a imprimé une sortie
+    // structurée, captée ensuite par le parser. Sans stdout, il n'y a rien à parser :
+    // on remonte le stderr réel au lieu de laisser l'appelant conclure à un no-Final.
+    let status = child.wait().await.map_err(map_io)?;
+    let stderr_bytes = stderr_task
+        .await
+        .map_err(|e| AgentSessionError::Io(format!("lecture stderr: {e}")))?
+        .map_err(map_io)?;
+    if !status.success() && collected.is_empty() {
+        return Err(AgentSessionError::Io(process_exit_message(
+            &spec.command,
+            status.code(),
+            &stderr_bytes,
+        )));
+    }
     Ok(collected)
 }
 
@@ -360,3 +403,14 @@ async fn drain(
 fn map_io(e: io::Error) -> AgentSessionError {
     AgentSessionError::Io(e.to_string())
 }
+
+fn process_exit_message(command: &str, code: Option<i32>, stderr: &[u8]) -> String {
+    let status = code.map_or_else(|| "signal".to_owned(), |code| format!("code {code}"));
+    let stderr = String::from_utf8_lossy(stderr);
+    let stderr = stderr.trim();
+    if stderr.is_empty() {
+        format!("{command} exited with {status} without structured output")
+    } else {
+        format!("{command} exited with {status}: {stderr}")
+    }
+}