fix(orchestrator): LiveStateTurnGuard ramène la LED live-state à Done au drop du futur ask cross-project — #281 (RAII armé après publication Working, désarmé sur sorties explicites mark_target_done_best_effort, Drop spawn best-effort UpdateLiveState Done avec last_delegation=ticket; test régression dropped_ask_future_marks_live_state_done; QA verte: cargo fmt, orchestrator_service 78/78, npm build frontend, AppImage rebuildée 2026-09-08 13:52; réserve documentée: échec préexistant launch_conventionfile_injects_assigned_skills_in_order hors périmètre)

crates/application/src/orchestrator/service.rs

@@ -328,6 +328,79 @@ impl Drop for BusyTurnGuard {
     }
 }
 
+/// Best-effort live-state cleanup paired with [`BusyTurnGuard`].
+///
+/// `BusyTurnGuard` owns synchronous FIFO/busy cleanup only. Live-state writes are
+/// async, so this guard captures the already-materialized updater and spawns a
+/// best-effort `Done` transition when the ask future is dropped before an
+/// explicit success/error branch can clean the human-facing status row. This is
+/// the cross-project cancel/timeout case: the target turn is gone, but the
+/// persisted LED must not remain `working` until restart/TTL.
+struct LiveStateTurnGuard {
+    update: Option<Arc<UpdateLiveState>>,
+    target: AgentId,
+    ticket: TicketId,
+    armed: bool,
+}
+
+impl LiveStateTurnGuard {
+    fn new(update: Option<Arc<UpdateLiveState>>, target: AgentId, ticket: TicketId) -> Self {
+        Self {
+            update,
+            target,
+            ticket,
+            armed: true,
+        }
+    }
+
+    fn disarm(&mut self) {
+        self.armed = false;
+    }
+}
+
+impl Drop for LiveStateTurnGuard {
+    fn drop(&mut self) {
+        if !self.armed {
+            return;
+        }
+        let Some(update) = self.update.clone() else {
+            return;
+        };
+        let target = self.target;
+        let ticket = self.ticket;
+        match tokio::runtime::Handle::try_current() {
+            Ok(handle) => {
+                handle.spawn(async move {
+                    let outcome = update
+                        .execute(UpdateLiveStateInput {
+                            agent_id: target,
+                            ticket: Some(ticket),
+                            intent: String::new(),
+                            status: domain::live_state::WorkStatus::Done,
+                            progress: None,
+                            last_delegation: Some(ticket),
+                        })
+                        .await;
+                    if let Err(e) = outcome {
+                        crate::diag!(
+                            "[live-state] drop cleanup failed: target={target} ticket={ticket} err={e}"
+                        );
+                    } else {
+                        crate::diag!(
+                            "[live-state] drop cleanup marked target done: target={target} ticket={ticket}"
+                        );
+                    }
+                });
+            }
+            Err(_) => {
+                crate::diag!(
+                    "[live-state] drop cleanup skipped without tokio runtime: target={target} ticket={ticket}"
+                );
+            }
+        }
+    }
+}
+
 /// Fournit les faits OS/runtime (exe + endpoint projet) pour écrire la déclaration MCP
 /// réelle quand l'orchestrateur (re)lance une cible sur le chemin `ask`. Implémenté dans
 /// app-tauri (seul détenteur de current_exe/$APPIMAGE/mcp_endpoint).
@@ -2122,6 +2195,11 @@ impl OrchestratorService {
             Self::distill_intent(&task),
         )
         .await;
+        let live_state_update = self
+            .live_state
+            .as_ref()
+            .and_then(|provider| provider.live_state_for(&project.root));
+        let mut live_state_guard = LiveStateTurnGuard::new(live_state_update, agent_id, ticket_id);
         // Garde RAII de fin de tour, armé JUSTE après l'enqueue (cible `Busy`). Couvre
         // toutes les sorties — `return Err` des bras du select, OU **futur abandonné
         // (drop)** — en ramenant la cible `Idle` au Drop (cf. [`BusyTurnGuard`]). C'est
@@ -2253,6 +2331,7 @@ impl OrchestratorService {
                 }
                 self.mark_target_done_best_effort(&project.root, agent_id, ticket_id)
                     .await;
+                live_state_guard.disarm();
                 crate::diag!(
                     "[rendezvous] ask error (structured): target={target} (agent {agent_id}) \
                      ticket={ticket_id} after_ms={} err={err}",
@@ -2271,6 +2350,7 @@ impl OrchestratorService {
                 .await?;
                 self.mark_target_done_best_effort(&project.root, agent_id, ticket_id)
                     .await;
+                live_state_guard.disarm();
                 crate::diag!(
                     "[rendezvous] ask CEILING (structured, still active): target={target} \
                      (agent {agent_id}) ticket={ticket_id} after_ms={}",
@@ -2290,6 +2370,7 @@ impl OrchestratorService {
                 .await?;
                 self.mark_target_done_best_effort(&project.root, agent_id, ticket_id)
                     .await;
+                live_state_guard.disarm();
                 crate::diag!(
                     "[rendezvous] ask TIMEOUT (structured): target={target} (agent {agent_id}) \
                      ticket={ticket_id} after_ms={} (aucun progrès observé sur la fenêtre)",
@@ -2325,6 +2406,7 @@ impl OrchestratorService {
         // résultat (chemin structuré) ⇒ `Done` + `last_delegation` = ticket résolu.
         self.mark_target_done_best_effort(&project.root, agent_id, ticket_id)
             .await;
+        live_state_guard.disarm();
         Ok(self.reply_outcome(agent_id, target, result))
     }
 

crates/application/tests/orchestrator_service.rs

@@ -2865,6 +2865,45 @@ async fn dropped_ask_future_frees_busy_target() {
     );
 }
 
+/// Régression cross-project/workstate — si le client MCP appelant abandonne un
+/// `idea_project_ask_agent`, le futur `ask` est droppé. Le garde Busy libère déjà
+/// la FIFO ; il faut aussi nettoyer le live-state, sinon l'UI garde une LED
+/// `working`/jaune jusqu'au restart ou à la TTL.
+#[tokio::test]
+async fn dropped_ask_future_marks_live_state_done() {
+    let agent = scratch_agent(aid(1), "architect", "agents/architect.md");
+    let fx = ask_fixture(FakeContexts::with_agent(&agent, "# persona"));
+    seed_live_pty(&fx.sessions, aid(1), sid(800));
+
+    let svc = Arc::clone(&fx.service);
+    let ask = tokio::spawn(async move { svc.dispatch(&project(), cmd(ASK_JSON)).await });
+    await_until(|| fx.mailbox.pending(&aid(1)) == 1).await;
+    await_until(|| {
+        fx.live
+            .entry_for(aid(1))
+            .is_some_and(|entry| entry.status == WorkStatus::Working)
+    })
+    .await;
+
+    ask.abort();
+    let _ = ask.await;
+
+    await_until(|| {
+        fx.live
+            .entry_for(aid(1))
+            .is_some_and(|entry| entry.status == WorkStatus::Done)
+    })
+    .await;
+    let entry = fx
+        .live
+        .entry_for(aid(1))
+        .expect("target live entry present");
+    assert_eq!(
+        entry.last_delegation, entry.ticket,
+        "drop cleanup preserves the ticket as last_delegation"
+    );
+}
+
 /// Régression (garde RAII) — une **2e délégation** vers une cible dont un 1er `ask` a
 /// été abandonné est bien livrée : l'agent n'est pas coincé `Busy`, son tour suivant
 /// démarre et peut être résolu par `idea_reply`.