Lots of changes... not done yet.

AGENTS.md

@@ -40,6 +40,7 @@ Environment variables:
 - Set `GALAXY_BEDROCK_DIAGNOSTICS=1` to enable Bedrock diagnostic output, including:
   - `Error_<timestamp>.txt` snapshot files written to the repository root on request/stream failures (includes the serialized Bedrock context window, tool definitions, protobuf request debug payload, captured Bedrock diagnostic lines, and log tails)
   - Per-event Bedrock diagnostic logs written to `bedrock-diagnostics.log` in the active Warp log directory
+- Set `GALAXY_TOOL_DIAGNOSTICS=1` to enable verbose local tool queue/execution debug logs and cancellation backtraces. These diagnostics are disabled during routine operation.
 
 ### AI Provider Architecture
 
@@ -139,7 +140,10 @@ Key invariants:
 - Direct-provider `RequestFileEdits` views must register from streaming output before provider-run completion; preprocessing results must survive delayed view registration, and `NotReady` retries must remain automatic rather than emitting a synthetic user permission decision
 - A clean direct-provider `ProviderRunOutcome::Completed` explicitly finalizes the conversation as `Success` after terminal output projection, even if earlier turns added tool actions; child-completion waits rely on that status
 - Provider actions and results must correlate by `(conversation_id, run_id, epoch, call_id)`; stale or duplicate callbacks must not advance a run
-- Active provider runs must checkpoint before external work, persist without credentials, normalize unsafe restored states, and reconcile command state before continuing
+- Action status/result lookups and archived results are keyed by `(conversation_id, action_id)`; callers must supply the owning conversation and must not fall back to a global action-ID search
+- Action blocked/executing/finished events carry `conversation_id`; UI subscribers must match it, and CLI shell-control mutations must also match the active block's requested-command action ID
+- Active provider runs must checkpoint before external work, persist without credentials, normalize unsafe restored states, and reconcile command state before continuing; cancellation intent stays checkpointed until the terminal outcome is projected and `finish_active_provider_run` performs cleanup
+- Same-conversation direct-provider follow-ups queue behind the cancelling generation; the old run keeps the active slot until terminal projection and cleanup, then the next generation starts, and stale callbacks are ignored by stream identity
 - Known tools are in `KNOWN_TOOLS` in `response_translator.rs`; definitions are built by `tool_definition_for_name()` in `convert_request.rs`
 - Direct-provider normal and plan turns must advertise `read_plan`, `create_plan`, and `edit_plan` when the matching document capabilities are enabled; plan-creation requests should call `create_plan` after research rather than only returning prose
 - Unknown or invalid tool calls receive one correlated synthetic error result and a visible `AgentOutput` message; the durable run owns any continuation
@@ -151,10 +155,13 @@ Key invariants:
 - Progressive summaries are prepended to provider requests as a user/assistant pair; background summarization remains independent of the active provider run
 - Loop prevention in `controller.rs` detects repeated tool failures (3+ identical) and injects a corrective instruction
 - Direct-provider long-running shell follow-ups retain CLI tasks as UI projections while command monitoring and completion stay owned by the same provider run
+- A direct-provider command completion is only queued when the terminal reports it; the CLI task remains active until the provider run applies that completion at a safe boundary and deactivates it
 - Direct-provider completed-command assessments are hidden, tool-free root-task turns whose output and hidden input survive CLI-task deactivation and restoration
 - ACP remains a separate session-owned runtime; direct-provider cleanup must not move ACP lifecycle into `ProviderRun`
 - Orchestrated child conversations are leaf workers: nested `RunAgents` and legacy `StartAgent` calls are rejected before autonomy or permission bypasses, and child requests do not advertise delegation tools
 - Direct-provider `RunAgents` remains pending until every local child reaches `Success`, `Error`, or `Cancelled`, or is removed/deleted; recoverable `Blocked`, `TransientError`, and `WaitingForEvents` states remain pending, and the hosted 30-second startup timeout must not apply to these completion waits
+- `StartAgentWaitPolicy` is selected from child execution mode, not parent `run_id`: local children wait for completion and only remote/hosted children use startup acknowledgement
+- Hosted `RunAgents` startup timeouts detach the exact `StartAgentRequestId`; late launch callbacks must not register the child after the timeout result, while children linked before cancellation remain independently running
 
 ### Platform Setup
 - `./script/bootstrap` - Platform-specific setup plus common agent skill installation from `skills-lock.json`; prompts for project/global when an install or update is needed unless a target flag or environment override is provided.

app/src/ai/agent/conversation.rs

@@ -1930,7 +1930,8 @@ impl AIConversation {
     ) -> String {
         let mut result = Vec::new();
         for exchange in self.all_exchanges() {
-            let formatted_exchange = exchange.format_for_copy(action_model);
+            let formatted_exchange =
+                exchange.format_for_copy_for_conversation(action_model, Some(self.id()));
             if !formatted_exchange.is_empty() {
                 result.push(formatted_exchange);
             }

app/src/ai/agent/mod.rs

@@ -581,6 +581,14 @@ impl AIAgentOutput {
     pub fn format_for_copy(
         &self,
         action_model: Option<&crate::ai::blocklist::BlocklistAIActionModel>,
+    ) -> String {
+        self.format_for_copy_for_conversation(action_model, None)
+    }
+
+    pub fn format_for_copy_for_conversation(
+        &self,
+        action_model: Option<&crate::ai::blocklist::BlocklistAIActionModel>,
+        conversation_id: Option<conversation::AIConversationId>,
     ) -> String {
         let mut result = Vec::new();
         let mut last_was_action = false;
@@ -612,8 +620,12 @@ impl AIAgentOutput {
                 }
                 AIAgentOutputMessageType::Action(action) => {
                     // Include action results from the action model if available
-                    if let Some(action_model) = action_model {
+                    if let (Some(action_model), Some(conversation_id)) =
-                        if let Some(action_result) = action_model.get_action_result(&action.id) {
+                        (action_model, conversation_id)
+                    {
+                        if let Some(action_result) =
+                            action_model.get_action_result(conversation_id, &action.id)
+                        {
                             result.push(format!("{}", MarkdownActionResult(&action_result.result)));
                             // Add an extra newline after tool call results for readability
                             result.push(String::new());
@@ -1222,6 +1234,9 @@ impl<'a> std::fmt::Display for MarkdownActionResult<'a> {
                 RequestCommandOutputResult::CancelledBeforeExecution => {
                     write!(f, "\n_Command cancelled_")
                 }
+                RequestCommandOutputResult::ExecutionError { command, message } => {
+                    write!(f, "\n_Command `{command}` was not executed: {message}_")
+                }
                 RequestCommandOutputResult::Denylisted { command } => {
                     write!(
                         f,
@@ -3225,9 +3240,19 @@ impl AIAgentExchange {
     pub fn format_output_for_copy(
         &self,
         action_model: Option<&crate::ai::blocklist::BlocklistAIActionModel>,
+    ) -> String {
+        self.format_output_for_copy_for_conversation(action_model, None)
+    }
+
+    pub fn format_output_for_copy_for_conversation(
+        &self,
+        action_model: Option<&crate::ai::blocklist::BlocklistAIActionModel>,
+        conversation_id: Option<conversation::AIConversationId>,
     ) -> String {
         match self.output_status.output() {
-            Some(output) => output.get().format_for_copy(action_model),
+            Some(output) => output
+                .get()
+                .format_for_copy_for_conversation(action_model, conversation_id),
             None => String::new(),
         }
     }
@@ -3238,9 +3263,18 @@ impl AIAgentExchange {
     pub fn format_for_copy(
         &self,
         action_model: Option<&crate::ai::blocklist::BlocklistAIActionModel>,
+    ) -> String {
+        self.format_for_copy_for_conversation(action_model, None)
+    }
+
+    pub fn format_for_copy_for_conversation(
+        &self,
+        action_model: Option<&crate::ai::blocklist::BlocklistAIActionModel>,
+        conversation_id: Option<conversation::AIConversationId>,
     ) -> String {
         let input_text = self.format_input_for_copy();
-        let output_text = self.format_output_for_copy(action_model);
+        let output_text =
+            self.format_output_for_copy_for_conversation(action_model, conversation_id);
         let has_user_input = !input_text.is_empty();
         let has_agent_output = !output_text.is_empty();
 

app/src/ai/agent_sdk/driver/output.rs

@@ -61,6 +61,9 @@ pub mod text {
                     RequestCommandOutputResult::CancelledBeforeExecution => {
                         writeln!(w, "{CANCELLED_MESSAGE}")
                     }
+                    RequestCommandOutputResult::ExecutionError { command, message } => {
+                        writeln!(w, "Command `{command}` was not executed: {message}")
+                    }
                     RequestCommandOutputResult::Denylisted { .. } => {
                         writeln!(
                             w,
@@ -829,6 +832,11 @@ pub mod json {
                     RequestCommandOutputResult::CancelledBeforeExecution => {
                         Some(JsonMessage::ToolCanceled)
                     }
+                    RequestCommandOutputResult::ExecutionError { message, .. } => {
+                        Some(JsonMessage::ToolError {
+                            error: Cow::Borrowed(message),
+                        })
+                    }
                     RequestCommandOutputResult::Denylisted { .. } => Some(JsonMessage::ToolError {
                         error: Cow::Borrowed(
                             "Command was not allowed to run due to presence on denylist",

app/src/ai/blocklist/action_model.rs

@@ -31,7 +31,7 @@ pub use execute::{
     ReadFileContextResult, RequestFileEditsExecutor, RequestFileEditsFormatKind,
     RequestFileEditsTelemetryEvent, RunAgentsExecutor, RunAgentsExecutorEvent,
     RunAgentsSpawningSnapshot, ShellCommandExecutor, ShellCommandExecutorEvent, StartAgentExecutor,
-    StartAgentExecutorEvent, StartAgentRequest, StartAgentRequestId,
+    StartAgentExecutorEvent, StartAgentRequest, StartAgentRequestId, StartAgentWaitPolicy,
 };
 use futures::future::{join_all, BoxFuture};
 use galaxy_agent_core::{
@@ -355,6 +355,45 @@ fn sort_action_results_by_order(
     results.sort_by_key(|result| action_order.get(&result.id).copied().unwrap_or(usize::MAX));
 }
 
+fn action_result_for_conversation<'a>(
+    finished_action_results: &'a HashMap<AIConversationId, Vec<Arc<AIAgentActionResult>>>,
+    provider_finished_action_results: &'a HashMap<
+        (AIConversationId, ExternalWorkId),
+        Vec<Arc<AIAgentActionResult>>,
+    >,
+    past_action_results: &'a HashMap<(AIConversationId, AIAgentActionId), Arc<AIAgentActionResult>>,
+    conversation_id: AIConversationId,
+    action_id: &AIAgentActionId,
+) -> Option<&'a Arc<AIAgentActionResult>> {
+    finished_action_results
+        .get(&conversation_id)
+        .into_iter()
+        .chain(provider_finished_action_results.iter().filter_map(
+            |((result_conversation_id, _), results)| {
+                (*result_conversation_id == conversation_id).then_some(results)
+            },
+        ))
+        .flat_map(|results| results.iter())
+        .find(|result| &result.id == action_id)
+        .or_else(|| past_action_results.get(&(conversation_id, action_id.clone())))
+}
+
+fn pending_action_status(
+    pending_actions: &HashMap<AIConversationId, VecDeque<AIAgentAction>>,
+    running_actions: &HashMap<AIConversationId, RunningActions>,
+    conversation_id: AIConversationId,
+    action_id: &AIAgentActionId,
+    is_view_only: bool,
+) -> Option<AIActionStatus> {
+    let actions = pending_actions.get(&conversation_id)?;
+    let index = actions.iter().position(|action| &action.id == action_id)?;
+    if index == 0 && !is_view_only && !running_actions.contains_key(&conversation_id) {
+        Some(AIActionStatus::Blocked)
+    } else {
+        Some(AIActionStatus::Queued)
+    }
+}
+
 fn domain_tool_result(action_result: &AIAgentActionResult, permission_denied: bool) -> ToolResult {
     let status = if permission_denied {
         ToolResultStatus::Denied
@@ -587,6 +626,9 @@ fn action_result_failure_summary(result: &AIAgentActionResultType) -> Option<Str
         AIAgentActionResultType::RunAgents(RunAgentsResult::Failure { error }) => {
             Some(error.clone())
         }
+        AIAgentActionResultType::RequestCommandOutput(
+            RequestCommandOutputResult::ExecutionError { message, .. },
+        ) => Some(message.clone()),
         AIAgentActionResultType::RequestCommandOutput(
             RequestCommandOutputResult::Completed { .. }
             | RequestCommandOutputResult::CancelledBeforeExecution
@@ -734,7 +776,7 @@ pub struct BlocklistAIActionModel {
         HashMap<(AIConversationId, AIAgentActionId), ProviderToolExecutionRef>,
 
     /// Past actions and their corresponding statuses from previous AI exchanges.
-    past_action_results: HashMap<AIAgentActionId, Arc<AIAgentActionResult>>,
+    past_action_results: HashMap<(AIConversationId, AIAgentActionId), Arc<AIAgentActionResult>>,
 
     /// The ID of the terminal view this controller is associated with.
     terminal_view_id: EntityId,
@@ -774,6 +816,7 @@ impl BlocklistAIActionModel {
                 let execution_ref = me.provider_tool_execution_ref(*conversation_id, action_id);
                 ctx.emit(BlocklistAIActionEvent::ExecutingAction {
                     action_id: action_id.clone(),
+                    conversation_id: *conversation_id,
                     execution_ref: execution_ref.clone(),
                 });
                 ctx.emit(BlocklistAIActionEvent::ToolLifecycle {
@@ -863,6 +906,7 @@ impl BlocklistAIActionModel {
         );
         ctx.emit(BlocklistAIActionEvent::ExecutingAction {
             action_id: action_id.clone(),
+            conversation_id,
             execution_ref: self.provider_tool_execution_ref(conversation_id, action_id),
         });
     }
@@ -988,8 +1032,8 @@ impl BlocklistAIActionModel {
             .get(&conversation_id)
             .map(|q| q.len())
             .unwrap_or(0);
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] try_to_execute_available_actions: conversation={:?}, pending_count={}",
+            "try_to_execute_available_actions: conversation={:?}, pending_count={}",
             conversation_id,
             pending_count
         );
@@ -1000,14 +1044,14 @@ impl BlocklistAIActionModel {
                 .and_then(|queue| queue.front())
                 .cloned()
             else {
-                log::info!(
+                crate::ai::tool_diagnostics::tool_debug!(
-                    "[tool-debug] try_to_execute_available_actions: no more pending actions"
+                    "try_to_execute_available_actions: no more pending actions"
                 );
                 return;
             };
 
-            log::info!(
+            crate::ai::tool_diagnostics::tool_debug!(
-                "[tool-debug] try_to_execute_available_actions: trying action id={:?}, type={:?}",
+                "try_to_execute_available_actions: trying action id={:?}, type={:?}",
                 front_action.id,
                 std::mem::discriminant(&front_action.action)
             );
@@ -1019,8 +1063,8 @@ impl BlocklistAIActionModel {
                     current_phase,
                     ctx,
                 ) {
-                    log::info!(
+                    crate::ai::tool_diagnostics::tool_debug!(
-                        "[tool-debug] try_to_execute_available_actions: cannot start in current phase {:?}",
+                        "try_to_execute_available_actions: cannot start in current phase {:?}",
                         current_phase
                     );
                     return;
@@ -1033,12 +1077,12 @@ impl BlocklistAIActionModel {
                 ActionExecutionInitiator::Automatic,
                 ctx,
             ) else {
-                log::info!("[tool-debug] try_to_execute_available_actions: start_pending_action_by_id returned None (blocked)");
+                crate::ai::tool_diagnostics::tool_debug!("try_to_execute_available_actions: start_pending_action_by_id returned None (blocked)");
                 return;
             };
 
-            log::info!(
+            crate::ai::tool_diagnostics::tool_debug!(
-                "[tool-debug] try_to_execute_available_actions: action started, result={:?}",
+                "try_to_execute_available_actions: action started, result={:?}",
                 std::mem::discriminant(&result)
             );
 
@@ -1048,7 +1092,9 @@ impl BlocklistAIActionModel {
                     phase: RunningActionPhase::Serial
                 }
             ) {
-                log::info!("[tool-debug] try_to_execute_available_actions: serial async action, stopping loop");
+                crate::ai::tool_diagnostics::tool_debug!(
+                    "try_to_execute_available_actions: serial async action, stopping loop"
+                );
                 return;
             }
         }
@@ -1081,14 +1127,6 @@ impl BlocklistAIActionModel {
         }
     }
 
-    /// Returns all pending actions for all conversations.
-    pub fn get_pending_actions(&self) -> Vec<&AIAgentAction> {
-        self.pending_actions
-            .values()
-            .flat_map(|queue| queue.iter())
-            .collect()
-    }
-
     /// Returns all pending actions for a specific conversation.
     pub fn get_pending_actions_for_conversation(
         &self,
@@ -1106,11 +1144,15 @@ impl BlocklistAIActionModel {
         self.blocked_action_for_conversation(&conversation_id)
     }
 
-    /// Returns a pending action by its ID, searching across all conversations.
+    /// Returns a pending action by its ID within the given conversation.
-    pub fn get_pending_action_by_id(&self, action_id: &AIAgentActionId) -> Option<&AIAgentAction> {
+    pub fn get_pending_action_by_id(
+        &self,
+        conversation_id: AIConversationId,
+        action_id: &AIAgentActionId,
+    ) -> Option<&AIAgentAction> {
         self.pending_actions
-            .values()
+            .get(&conversation_id)?
-            .flat_map(|queue| queue.iter())
+            .iter()
             .find(|action| &action.id == action_id)
     }
 
@@ -1142,7 +1184,11 @@ impl BlocklistAIActionModel {
         self.running_actions
             .get(&conversation_id)
             .and_then(RunningActions::first_action_id)
-            .and_then(|action_id| self.executor.as_ref(app).async_executing_action(action_id))
+            .and_then(|action_id| {
+                self.executor
+                    .as_ref(app)
+                    .async_executing_action(conversation_id, action_id)
+            })
     }
 
     /// Returns whether there is a pending or running action for the active conversation.
@@ -1194,53 +1240,58 @@ impl BlocklistAIActionModel {
         self.finished_action_results.get(&conversation_id)
     }
 
-    /// Returns the `AIActionStatus` for the action corresponding to the given `id`, if any.
+    /// Returns the status for an action within the given conversation.
-    pub fn get_action_status(&self, id: &AIAgentActionId) -> Option<AIActionStatus> {
+    pub fn get_action_status(
-        for (conversation_id, pending_actions_for_conversation) in &self.pending_actions {
+        &self,
-            for (index, action) in pending_actions_for_conversation.iter().enumerate() {
+        conversation_id: AIConversationId,
-                if &action.id != id {
+        id: &AIAgentActionId,
-                    continue;
+    ) -> Option<AIActionStatus> {
-                }
+        if let Some(status) = pending_action_status(
-
+            &self.pending_actions,
-                if index == 0
+            &self.running_actions,
-                    && !self.is_view_only
+            conversation_id,
-                    && !self.running_actions.contains_key(conversation_id)
+            id,
-                {
+            self.is_view_only,
-                    return Some(AIActionStatus::Blocked);
+        ) {
-                }
+            return Some(status);
-
-                return Some(AIActionStatus::Queued);
-            }
         }
 
         self.running_actions
-            .values()
+            .get(&conversation_id)
-            .find(|running| running.contains(id))
+            .filter(|running| running.contains(id))
             .map(|_| AIActionStatus::RunningAsync)
             .or_else(|| {
-                self.get_action_result(id)
+                self.get_action_result(conversation_id, id)
                     .map(|result| AIActionStatus::Finished(result.clone()))
             })
             .or_else(|| {
                 self.pending_preprocessed_actions
-                    .values()
+                    .get(&conversation_id)
-                    .any(|preprocessing| preprocessing.contains(id))
+                    .is_some_and(|preprocessing| preprocessing.contains(id))
                     .then_some(AIActionStatus::Preprocessing)
             })
     }
 
-    pub fn get_action_result(&self, id: &AIAgentActionId) -> Option<&Arc<AIAgentActionResult>> {
+    pub fn get_action_result(
-        // Search through all conversations' finished action results
+        &self,
-        self.finished_action_results
+        conversation_id: AIConversationId,
-            .values()
+        id: &AIAgentActionId,
-            .chain(self.provider_finished_action_results.values())
+    ) -> Option<&Arc<AIAgentActionResult>> {
-            .flat_map(|results| results.iter())
+        action_result_for_conversation(
-            .find(|result| &result.id == id)
+            &self.finished_action_results,
-            .or_else(|| self.past_action_results.get(id))
+            &self.provider_finished_action_results,
+            &self.past_action_results,
+            conversation_id,
+            id,
+        )
     }
 
     /// Bulk restore action results from a list of exchanges (used when loading conversations from tasks)
-    pub fn restore_action_results_from_exchanges(&mut self, exchanges: Vec<&AIAgentExchange>) {
+    pub fn restore_action_results_from_exchanges(
+        &mut self,
+        conversation_id: AIConversationId,
+        exchanges: Vec<&AIAgentExchange>,
+    ) {
         for exchange in exchanges.iter() {
             for input in &exchange.input {
                 if let AIAgentInput::ActionResult { result, .. } = input {
@@ -1257,7 +1308,7 @@ impl BlocklistAIActionModel {
                         );
                     }
                     self.past_action_results
-                        .insert(result_id, Arc::new(result_to_insert));
+                        .insert((conversation_id, result_id), Arc::new(result_to_insert));
                 }
             }
         }
@@ -1267,18 +1318,16 @@ impl BlocklistAIActionModel {
     /// from the confirmation card.
     pub fn execute_run_agents(
         &mut self,
+        conversation_id: AIConversationId,
         action_id: &AIAgentActionId,
         request: ai::agent::action::RunAgentsRequest,
         ctx: &mut ModelContext<Self>,
     ) {
-        let mut found = None;
+        let Some(action) = self
-        for (conv_id, queue) in self.pending_actions.iter_mut() {
+            .pending_actions
-            if let Some(action) = queue.iter_mut().find(|action| &action.id == action_id) {
+            .get_mut(&conversation_id)
-                found = Some((*conv_id, action));
+            .and_then(|queue| queue.iter_mut().find(|action| &action.id == action_id))
-                break;
+        else {
-            }
-        }
-        let Some((conversation_id, action)) = found else {
             log::warn!(
                 "BlocklistAIActionModel::execute_run_agents: no pending action for {action_id:?}"
             );
@@ -1299,20 +1348,19 @@ impl BlocklistAIActionModel {
     /// the time the action becomes blocked on user confirmation.
     pub fn deny_run_agents(
         &mut self,
+        conversation_id: AIConversationId,
         action_id: &AIAgentActionId,
         reason: String,
         ctx: &mut ModelContext<Self>,
     ) {
-        let mut found: Option<(AIConversationId, AIAgentAction)> = None;
+        let Some(action) = self
-        for (conv_id, queue) in self.pending_actions.iter_mut() {
+            .pending_actions
-            if let Some(idx) = queue.iter().position(|a| &a.id == action_id) {
+            .get_mut(&conversation_id)
-                if let Some(action) = queue.remove(idx) {
+            .and_then(|queue| {
-                    found = Some((*conv_id, action));
+                let index = queue.iter().position(|action| &action.id == action_id)?;
-                }
+                queue.remove(index)
-                break;
+            })
-            }
+        else {
-        }
-        let Some((conversation_id, action)) = found else {
             log::warn!(
                 "BlocklistAIActionModel::deny_run_agents: no pending action for {action_id:?}"
             );
@@ -1451,6 +1499,7 @@ impl BlocklistAIActionModel {
                 let execution_ref = self.provider_tool_execution_ref(conversation_id, &action.id);
                 ctx.emit(BlocklistAIActionEvent::ActionBlockedOnUserConfirmation {
                     action_id: action.id.clone(),
+                    conversation_id,
                     execution_ref: execution_ref.clone(),
                 });
                 ctx.emit(BlocklistAIActionEvent::ToolLifecycle {
@@ -1698,14 +1747,14 @@ impl BlocklistAIActionModel {
         conversation_id: AIConversationId,
         ctx: &mut ModelContext<Self>,
     ) {
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] queue_actions: queuing {} actions for conversation {:?}",
+            "queue_actions: queuing {} actions for conversation {:?}",
             actions.len(),
             conversation_id
         );
         for (i, action) in actions.iter().enumerate() {
-            log::info!(
+            crate::ai::tool_diagnostics::tool_debug!(
-                "[tool-debug] queue_actions:   [{}] id={:?}, type={:?}",
+                "queue_actions:   [{}] id={:?}, type={:?}",
                 i,
                 action.id,
                 std::mem::discriminant(&action.action)
@@ -1856,7 +1905,7 @@ impl BlocklistAIActionModel {
         reason: CancellationReason,
         ctx: &mut ModelContext<Self>,
     ) {
-        let status = self.get_action_status(action_id);
+        let status = self.get_action_status(conversation_id, action_id);
         let permission_denied = is_permission_denial(reason, status.as_ref());
         if self
             .running_actions
@@ -1864,7 +1913,7 @@ impl BlocklistAIActionModel {
             .is_some_and(|running| running.contains(action_id))
         {
             self.executor.update(ctx, |executor, ctx| {
-                executor.cancel_running_async_action(action_id, Some(reason), ctx)
+                executor.cancel_running_async_action(conversation_id, action_id, Some(reason), ctx)
             });
         } else {
             let Some(pending_actions_for_conversation) =
@@ -1933,12 +1982,14 @@ impl BlocklistAIActionModel {
         };
         for action in actions_to_cancel.drain(..).collect_vec() {
             log::info!(
-                "Canceling pending action of type {:?} conversation_id={conversation_id:?} action_id={:?}, reason={:?}, backtrace=\n{}",
+                "Canceling pending action of type {:?} conversation_id={conversation_id:?} action_id={:?}, reason={:?}",
                 AIAgentActionTypeDiscriminants::from(&action.action),
                 action.id,
-                reason,
+                reason
-                std::backtrace::Backtrace::force_capture()
             );
+            if let Some(backtrace) = crate::ai::tool_diagnostics::capture_backtrace() {
+                log::debug!("Pending action cancellation backtrace:\n{backtrace}");
+            }
             self.cancel_pending_action(conversation_id, action, reason, false, ctx);
         }
     }
@@ -2053,7 +2104,7 @@ impl BlocklistAIActionModel {
 
         for result in finished_action_results.iter() {
             self.past_action_results
-                .insert(result.id.clone(), result.clone());
+                .insert((conversation_id, result.id.clone()), result.clone());
         }
         finished_action_results
             .into_iter()
@@ -2091,7 +2142,8 @@ impl BlocklistAIActionModel {
             .remove(&(conversation_id, work_id.clone()))
             .unwrap_or_default();
         for result in results {
-            self.past_action_results.insert(result.id.clone(), result);
+            self.past_action_results
+                .insert((conversation_id, result.id.clone()), result);
         }
     }
 
@@ -2123,33 +2175,28 @@ impl BlocklistAIActionModel {
     /// respective functions.
     pub fn handle_requested_command_accepted(
         &mut self,
+        conversation_id: AIConversationId,
         action_id: &AIAgentActionId,
         command: String,
         ctx: &mut ModelContext<Self>,
     ) {
-        // Search through all pending conversations to find the action and conversation ID
+        let Some(action) = self
-        let mut found_conversation_id = None;
+            .pending_actions
-        for (conversation_id, pending_actions_for_conversation) in self.pending_actions.iter_mut() {
+            .get_mut(&conversation_id)
-            if let Some(action) = pending_actions_for_conversation
+            .and_then(|actions| actions.iter_mut().find(|action| action.id == *action_id))
-                .iter_mut()
+        else {
-                .find(|action| action.id == *action_id)
-            {
-                if let AIAgentActionType::RequestCommandOutput {
-                    command: original_command,
-                    ..
-                } = &mut action.action
-                {
-                    *original_command = command;
-                    found_conversation_id = Some(*conversation_id);
-                    break;
-                }
-            }
-        }
-
-        let Some(conversation_id) = found_conversation_id else {
             log::warn!("Ignoring acceptance for non-pending requested command: {action_id:?}");
             return;
         };
+        let AIAgentActionType::RequestCommandOutput {
+            command: original_command,
+            ..
+        } = &mut action.action
+        else {
+            log::warn!("Ignoring acceptance for non-command action: {action_id:?}");
+            return;
+        };
+        *original_command = command;
 
         self.execute_action(action_id, conversation_id, ctx);
     }
@@ -2161,8 +2208,8 @@ impl BlocklistAIActionModel {
         cancellation_reason: Option<CancellationReason>,
         ctx: &mut ModelContext<Self>,
     ) {
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] handle_action_result: action_id={:?}, result_type={:?}, cancellation={:?}",
+            "handle_action_result: action_id={:?}, result_type={:?}, cancellation={:?}",
             action_result.id,
             std::mem::discriminant(&action_result.result),
             cancellation_reason
@@ -2453,11 +2500,13 @@ pub enum BlocklistAIActionEvent {
     /// Emitted when the action with the given ID requires user confirmation to execute.
     ActionBlockedOnUserConfirmation {
         action_id: AIAgentActionId,
+        conversation_id: AIConversationId,
         execution_ref: Option<ProviderToolExecutionRef>,
     },
     /// Emitted when the action with the given ID begins execution.
     ExecutingAction {
         action_id: AIAgentActionId,
+        conversation_id: AIConversationId,
         execution_ref: Option<ProviderToolExecutionRef>,
     },
     /// Emitted when the action with the given ID has finished.
@@ -2496,6 +2545,25 @@ impl BlocklistAIActionEvent {
             BlocklistAIActionEvent::InsertCodeReviewComments { action_id, .. } => action_id,
         }
     }
+
+    pub fn conversation_id(&self) -> Option<AIConversationId> {
+        match self {
+            BlocklistAIActionEvent::ActionBlockedOnUserConfirmation {
+                conversation_id, ..
+            }
+            | BlocklistAIActionEvent::ExecutingAction {
+                conversation_id, ..
+            }
+            | BlocklistAIActionEvent::FinishedAction {
+                conversation_id, ..
+            } => Some(*conversation_id),
+            BlocklistAIActionEvent::QueuedAction { .. }
+            | BlocklistAIActionEvent::ToolLifecycle { .. }
+            | BlocklistAIActionEvent::InitProject(_)
+            | BlocklistAIActionEvent::ToggleCodeReview(_)
+            | BlocklistAIActionEvent::InsertCodeReviewComments { .. } => None,
+        }
+    }
 }
 
 impl Entity for BlocklistAIActionModel {

app/src/ai/blocklist/action_model/execute.rs

@@ -74,6 +74,7 @@ use serde::{Deserialize, Serialize};
 pub use shell_command::{ShellCommandExecutor, ShellCommandExecutorEvent};
 pub use start_agent::{
     StartAgentExecutor, StartAgentExecutorEvent, StartAgentRequest, StartAgentRequestId,
+    StartAgentWaitPolicy,
 };
 pub use suggest_new_conversation::NewConversationDecision;
 use suggest_new_conversation::SuggestNewConversationExecutor;
@@ -245,9 +246,36 @@ pub(super) enum TryExecuteResult {
 #[derive(Clone)]
 struct AsyncExecutingAction {
     action: AIAgentAction,
-    /// The conversation this action belongs to so cancellation and follow-up scheduling remain
+}
-    /// scoped even when several conversations have async actions in flight.
+
-    conversation_id: AIConversationId,
+type AsyncExecutingActionKey = (AIConversationId, AIAgentActionId);
+
+#[derive(Default)]
+struct AsyncExecutingActions(
+    std::collections::HashMap<AsyncExecutingActionKey, AsyncExecutingAction>,
+);
+
+impl AsyncExecutingActions {
+    fn insert(&mut self, conversation_id: AIConversationId, running: AsyncExecutingAction) {
+        self.0
+            .insert((conversation_id, running.action.id.clone()), running);
+    }
+
+    fn get(
+        &self,
+        conversation_id: AIConversationId,
+        action_id: &AIAgentActionId,
+    ) -> Option<&AsyncExecutingAction> {
+        self.0.get(&(conversation_id, action_id.clone()))
+    }
+
+    fn remove(
+        &mut self,
+        conversation_id: AIConversationId,
+        action_id: &AIAgentActionId,
+    ) -> Option<AsyncExecutingAction> {
+        self.0.remove(&(conversation_id, action_id.clone()))
+    }
 }
 
 impl AsyncExecutingAction {
@@ -286,10 +314,8 @@ pub struct BlocklistAIActionExecutor {
     send_message_executor: ModelHandle<SendMessageToAgentExecutor>,
     ask_user_question_executor: ModelHandle<AskUserQuestionExecutor>,
     wait_for_events_executor: ModelHandle<WaitForEventsExecutor>,
-    /// The actions currently executing asynchronously, keyed by action ID.
+    /// The actions currently executing asynchronously, scoped by conversation and action ID.
-    /// We track them per action rather than as a single slot so multiple actions from the same
+    async_executing_actions: AsyncExecutingActions,
-    /// parallel phase can complete independently.
-    async_executing_actions: std::collections::HashMap<AIAgentActionId, AsyncExecutingAction>,
     restored_action_ids: HashSet<AIAgentActionId>,
 
     /// Reference to the terminal model for checking session sharing state.
@@ -390,9 +416,13 @@ impl BlocklistAIActionExecutor {
         }
     }
 
-    pub fn async_executing_action(&self, action_id: &AIAgentActionId) -> Option<&AIAgentAction> {
+    pub fn async_executing_action(
+        &self,
+        conversation_id: AIConversationId,
+        action_id: &AIAgentActionId,
+    ) -> Option<&AIAgentAction> {
         self.async_executing_actions
-            .get(action_id)
+            .get(conversation_id, action_id)
             .map(|running| &running.action)
     }
 
@@ -408,13 +438,16 @@ impl BlocklistAIActionExecutor {
     }
 
     pub(super) fn has_running_ask_user_question(&self, conversation_id: AIConversationId) -> bool {
-        self.async_executing_actions.values().any(|running| {
+        self.async_executing_actions
-            running.conversation_id == conversation_id
+            .0
-                && matches!(
+            .iter()
-                    running.action.action,
+            .any(|((running_conversation_id, _), running)| {
-                    AIAgentActionType::AskUserQuestion { .. }
+                *running_conversation_id == conversation_id
-                )
+                    && matches!(
-        })
+                        running.action.action,
+                        AIAgentActionType::AskUserQuestion { .. }
+                    )
+            })
     }
 
     /// Returns the action_id of any running WaitForEvents action for the
@@ -424,10 +457,9 @@ impl BlocklistAIActionExecutor {
         &self,
         conversation_id: AIConversationId,
     ) -> Option<AIAgentActionId> {
-        self.async_executing_actions
+        self.async_executing_actions.0.iter().find_map(
-            .iter()
+            |((running_conversation_id, action_id), running)| {
-            .find_map(|(action_id, running)| {
+                if *running_conversation_id == conversation_id
-                if running.conversation_id == conversation_id
                     && matches!(
                         running.action.action,
                         AIAgentActionType::WaitForEvents { .. }
@@ -437,7 +469,8 @@ impl BlocklistAIActionExecutor {
                 } else {
                     None
                 }
-            })
+            },
+        )
     }
 
     pub fn shell_command_executor(&self) -> &ModelHandle<ShellCommandExecutor> {
@@ -642,8 +675,8 @@ impl BlocklistAIActionExecutor {
         is_user_initiated: bool,
         ctx: &mut ModelContext<Self>,
     ) -> TryExecuteResult {
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] try_to_execute_action: action_id={:?}, type={:?}, is_user_initiated={}",
+            "try_to_execute_action: action_id={:?}, type={:?}, is_user_initiated={}",
             action.id,
             std::mem::discriminant(&action.action),
             is_user_initiated
@@ -651,7 +684,9 @@ impl BlocklistAIActionExecutor {
 
         // We should never actually execute actions in view-only mode.
         if self.is_shared_session_viewer() {
-            log::info!("[tool-debug] try_to_execute_action: BLOCKED - shared session viewer mode");
+            crate::ai::tool_diagnostics::tool_debug!(
+                "try_to_execute_action: BLOCKED - shared session viewer mode"
+            );
             return TryExecuteResult::NotExecuted {
                 reason: NotExecutedReason::WaitingOnSharer,
                 action: Box::new(action),
@@ -664,8 +699,8 @@ impl BlocklistAIActionExecutor {
         };
         let can_auto_execute = self.should_autoexecute(input, ctx);
         let is_agent_autonomous = AppExecutionMode::as_ref(ctx).is_autonomous();
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] try_to_execute_action: can_auto_execute={}, is_agent_autonomous={}",
+            "try_to_execute_action: can_auto_execute={}, is_agent_autonomous={}",
             can_auto_execute,
             is_agent_autonomous
         );
@@ -677,8 +712,8 @@ impl BlocklistAIActionExecutor {
             || can_auto_execute
             || (is_agent_autonomous && action.action.is_request_command_output()));
         if needs_confirmation {
-            log::info!(
+            crate::ai::tool_diagnostics::tool_debug!(
-                "[tool-debug] try_to_execute_action: NEEDS CONFIRMATION - action_id={:?}",
+                "try_to_execute_action: NEEDS CONFIRMATION - action_id={:?}",
                 action.id
             );
             return TryExecuteResult::NotExecuted {
@@ -713,8 +748,8 @@ impl BlocklistAIActionExecutor {
             }
         }
 
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] try_to_execute_action: EXECUTING action_id={:?}, type={:?}",
+            "try_to_execute_action: EXECUTING action_id={:?}, type={:?}",
             action.id,
             std::mem::discriminant(&action.action)
         );
@@ -870,8 +905,8 @@ impl BlocklistAIActionExecutor {
         };
 
         let action_id = action_clone.id.clone();
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] try_to_execute_action: execution result type={:?} for action_id={:?}",
+            "try_to_execute_action: execution result type={:?} for action_id={:?}",
             match &execution {
                 AnyActionExecution::NotReady => "NotReady",
                 AnyActionExecution::InvalidAction => "InvalidAction",
@@ -882,8 +917,8 @@ impl BlocklistAIActionExecutor {
         );
         match execution {
             AnyActionExecution::NotReady => {
-                log::info!(
+                crate::ai::tool_diagnostics::tool_debug!(
-                    "[tool-debug] try_to_execute_action: NOT READY - action_id={:?}",
+                    "try_to_execute_action: NOT READY - action_id={:?}",
                     action_id
                 );
                 TryExecuteResult::NotExecuted {
@@ -893,7 +928,7 @@ impl BlocklistAIActionExecutor {
             }
             AnyActionExecution::InvalidAction => {
                 log::error!(
-                    "[tool-debug] try_to_execute_action: INVALID ACTION - action_id={:?}",
+                    "try_to_execute_action: invalid action, action_id={:?}",
                     action_id
                 );
                 debug_assert!(false, "Tried to execute AIAgentAction with wrong executor.");
@@ -907,10 +942,9 @@ impl BlocklistAIActionExecutor {
                 on_complete,
             } => {
                 self.async_executing_actions.insert(
-                    action_id.clone(),
+                    conversation_id,
                     AsyncExecutingAction {
                         action: action_clone,
-                        conversation_id,
                     },
                 );
                 if !is_restored {
@@ -919,15 +953,21 @@ impl BlocklistAIActionExecutor {
                         conversation_id,
                     });
                 }
-                log::info!("[tool-debug] try_to_execute_action: spawning ASYNC execution for action_id={:?}", action_id);
+                crate::ai::tool_diagnostics::tool_debug!(
+                    "try_to_execute_action: spawning ASYNC execution for action_id={:?}",
+                    action_id
+                );
                 ctx.spawn(execute_future, move |me, result, ctx| {
-                    let Some(running) = me.async_executing_actions.remove(&action_id) else {
+                    let Some(running) = me
-                        log::warn!("[tool-debug] try_to_execute_action: async action completed but not found in executing map, action_id={:?}", action_id);
+                        .async_executing_actions
+                        .remove(conversation_id, &action_id)
+                    else {
+                        log::warn!("try_to_execute_action: async action completed but not found in executing map, conversation_id={conversation_id}, action_id={action_id:?}");
                         return;
                     };
                     let result = on_complete(result, ctx);
-                    log::info!(
+                    crate::ai::tool_diagnostics::tool_debug!(
-                        "[tool-debug] try_to_execute_action: ASYNC action COMPLETED action_id={:?}, result_type={:?}",
+                        "try_to_execute_action: ASYNC action COMPLETED action_id={:?}, result_type={:?}",
                         action_id,
                         std::mem::discriminant(&result)
                     );
@@ -937,7 +977,7 @@ impl BlocklistAIActionExecutor {
                             task_id: running.action.task_id,
                             result,
                         }),
-                        conversation_id: running.conversation_id,
+                        conversation_id,
                         cancellation_reason: None,
                     });
                 });
@@ -981,6 +1021,7 @@ impl BlocklistAIActionExecutor {
 
     pub fn cancel_running_async_action(
         &mut self,
+        conversation_id: AIConversationId,
         action_id: &AIAgentActionId,
         reason: Option<CancellationReason>,
         ctx: &mut ModelContext<Self>,
@@ -989,13 +1030,42 @@ impl BlocklistAIActionExecutor {
         if self.is_shared_session_viewer() {
             return;
         }
-        if let Some(running) = self.async_executing_actions.remove(action_id) {
+        if self
+            .async_executing_actions
+            .get(conversation_id, action_id)
+            .is_some_and(|running| {
+                matches!(
+                    running.action.action,
+                    AIAgentActionType::RequestCommandOutput { .. }
+                )
+            })
+        {
+            let termination_requested = self.shell_command_executor.update(ctx, |executor, ctx| {
+                executor.cancel_execution(action_id, ctx)
+            });
+            if termination_requested {
+                // Keep the action in flight until block completion proves the process stopped.
+                // Its normal async completion will report the actual terminal exit status.
+                return;
+            }
+        }
+        if let Some(running) = self
+            .async_executing_actions
+            .remove(conversation_id, action_id)
+        {
             let action_kind = AIAgentActionTypeDiscriminants::from(&running.action.action);
             log::info!(
-                "Canceling running async action of type {action_kind:?} action_id={action_id:?}, reason={reason:?}, backtrace=\n{}",
+                "Canceling running async action of type {action_kind:?} action_id={action_id:?}, reason={reason:?}"
-                std::backtrace::Backtrace::force_capture()
             );
-            if running.is_shell_command_action() {
+            if let Some(backtrace) = crate::ai::tool_diagnostics::capture_backtrace() {
+                log::debug!("Running action cancellation backtrace:\n{backtrace}");
+            }
+            if running.is_shell_command_action()
+                && !matches!(
+                    running.action.action,
+                    AIAgentActionType::RequestCommandOutput { .. }
+                )
+            {
                 self.shell_command_executor.update(ctx, |executor, ctx| {
                     executor.cancel_execution(&running.action.id, ctx);
                 });
@@ -1007,6 +1077,10 @@ impl BlocklistAIActionExecutor {
                 self.run_agents_executor.update(ctx, |executor, ctx| {
                     executor.cancel_execution(&running.action.id, ctx);
                 });
+            } else if matches!(running.action.action, AIAgentActionType::StartAgent { .. }) {
+                self.start_agent_executor.update(ctx, |executor, _| {
+                    executor.cancel_execution(&running.action.id);
+                });
             } else if let AIAgentActionType::WaitForEvents { tool_call_id, .. } =
                 &running.action.action
             {
@@ -1023,7 +1097,7 @@ impl BlocklistAIActionExecutor {
                     task_id: running.action.task_id,
                     result: running.action.action.cancelled_result(),
                 }),
-                conversation_id: running.conversation_id,
+                conversation_id,
                 cancellation_reason: reason,
             });
         }
@@ -1037,13 +1111,14 @@ impl BlocklistAIActionExecutor {
     ) {
         let action_ids = self
             .async_executing_actions
+            .0
             .iter()
-            .filter_map(|(action_id, running)| {
+            .filter_map(|((running_conversation_id, action_id), _)| {
-                (running.conversation_id == conversation_id).then_some(action_id.clone())
+                (*running_conversation_id == conversation_id).then_some(action_id.clone())
             })
             .collect::<Vec<_>>();
         for action_id in action_ids {
-            self.cancel_running_async_action(&action_id, reason, ctx);
+            self.cancel_running_async_action(conversation_id, &action_id, reason, ctx);
         }
     }
 
@@ -1493,6 +1568,95 @@ async fn read_file_as_binary(file_path: &std::path::Path) -> Result<Vec<u8>, Fil
     async_fs::read(file_path).await.map_err(FileLoadError::from)
 }
 
+#[cfg(test)]
+mod async_executing_action_tests {
+    use super::*;
+    use crate::ai::agent::task::TaskId;
+
+    fn action(id: &str, task_id: &str) -> AIAgentAction {
+        AIAgentAction {
+            id: AIAgentActionId::from(id.to_owned()),
+            action: AIAgentActionType::InitProject,
+            task_id: TaskId::new(task_id.to_owned()),
+            requires_result: true,
+            tool_name: Some("init_project".to_owned()),
+        }
+    }
+
+    #[test]
+    fn duplicate_action_ids_can_execute_concurrently_in_different_conversations() {
+        let first_conversation = AIConversationId::new();
+        let second_conversation = AIConversationId::new();
+        let duplicate_id = AIAgentActionId::from("duplicate".to_owned());
+        let mut running = AsyncExecutingActions::default();
+
+        running.insert(
+            first_conversation,
+            AsyncExecutingAction {
+                action: action("duplicate", "first-task"),
+            },
+        );
+        running.insert(
+            second_conversation,
+            AsyncExecutingAction {
+                action: action("duplicate", "second-task"),
+            },
+        );
+
+        assert_eq!(running.0.len(), 2);
+        assert_eq!(
+            running
+                .get(first_conversation, &duplicate_id)
+                .unwrap()
+                .action
+                .task_id,
+            TaskId::new("first-task".to_owned())
+        );
+        assert_eq!(
+            running
+                .get(second_conversation, &duplicate_id)
+                .unwrap()
+                .action
+                .task_id,
+            TaskId::new("second-task".to_owned())
+        );
+    }
+
+    #[test]
+    fn duplicate_action_completion_and_cancellation_remove_only_the_matching_conversation() {
+        let first_conversation = AIConversationId::new();
+        let second_conversation = AIConversationId::new();
+        let duplicate_id = AIAgentActionId::from("duplicate".to_owned());
+        let mut running = AsyncExecutingActions::default();
+        running.insert(
+            first_conversation,
+            AsyncExecutingAction {
+                action: action("duplicate", "first-task"),
+            },
+        );
+        running.insert(
+            second_conversation,
+            AsyncExecutingAction {
+                action: action("duplicate", "second-task"),
+            },
+        );
+
+        let completed = running.remove(first_conversation, &duplicate_id).unwrap();
+        assert_eq!(
+            completed.action.task_id,
+            TaskId::new("first-task".to_owned())
+        );
+        assert!(running.get(second_conversation, &duplicate_id).is_some());
+
+        let cancelled = running.remove(second_conversation, &duplicate_id).unwrap();
+        assert_eq!(
+            cancelled.action.task_id,
+            TaskId::new("second-task".to_owned())
+        );
+        assert!(running.0.is_empty());
+    }
+}
+
 #[cfg(all(test, feature = "local_fs"))]
 #[path = "execute_tests.rs"]
 mod tests;

app/src/ai/blocklist/action_model/execute/call_mcp_tool.rs

@@ -85,7 +85,7 @@ impl CallMCPToolExecutor {
 
         #[cfg(not(target_family = "wasm"))]
         {
-            log::info!("[tool-debug] CallMCPToolExecutor::execute called");
+            crate::ai::tool_diagnostics::tool_debug!("CallMCPToolExecutor::execute called");
             let server_output_id = get_server_output_id(input.conversation_id, ctx);
             let AIAgentAction {
                 action:
@@ -97,21 +97,21 @@ impl CallMCPToolExecutor {
                 ..
             } = input.action
             else {
-                log::error!("[tool-debug] CallMCPToolExecutor::execute: action type mismatch!");
+                log::error!("CallMCPToolExecutor::execute: action type mismatch");
                 return ActionExecution::InvalidAction;
             };
 
             let name_owned = name.to_owned();
             let name_clone = name_owned.clone();
-            log::info!(
+            crate::ai::tool_diagnostics::tool_debug!(
-                "[tool-debug] CallMCPToolExecutor: tool_name={}, server_id={:?}, input={}",
+                "CallMCPToolExecutor: tool_name={}, server_id={:?}, input={}",
                 name,
                 server_id,
                 serde_json::to_string(input).unwrap_or_else(|_| "<serialize error>".to_string())
             );
 
             let serde_json::Value::Object(mut arguments) = input.clone() else {
-                log::error!("[tool-debug] CallMCPToolExecutor: input is not an object!");
+                log::error!("CallMCPToolExecutor: input is not an object");
                 return ActionExecution::Sync(AIAgentActionResultType::CallMCPTool(
                     CallMCPToolResult::Error("MCP server tool input not an object".to_owned()),
                 ));
@@ -143,15 +143,15 @@ impl CallMCPToolExecutor {
 
             let Some(reconnecting_peer) = templatable_peer else {
                 log::error!(
-                    "[tool-debug] CallMCPToolExecutor: MCP server for tool '{}' NOT FOUND",
+                    "CallMCPToolExecutor: MCP server for tool '{}' not found",
                     name_owned
                 );
                 return ActionExecution::Sync(AIAgentActionResultType::CallMCPTool(
                     CallMCPToolResult::Error("MCP server for tool not found".to_owned()),
                 ));
             };
-            log::info!(
+            crate::ai::tool_diagnostics::tool_debug!(
-                "[tool-debug] CallMCPToolExecutor: found MCP server peer for tool '{}'",
+                "CallMCPToolExecutor: found MCP server peer for tool '{}'",
                 name_owned
             );
 
@@ -314,8 +314,8 @@ fn handle_call_tool_result(
     tool_name: String,
     ctx: &galaxyui::AppContext,
 ) -> AIAgentActionResultType {
-    log::info!(
+    crate::ai::tool_diagnostics::tool_debug!(
-        "[tool-debug] handle_call_tool_result: tool_name={}, is_ok={}",
+        "handle_call_tool_result: tool_name={}, is_ok={}",
         tool_name,
         res.is_ok()
     );

app/src/ai/blocklist/action_model/execute/file_glob.rs

@@ -108,8 +108,8 @@ impl FileGlobExecutor {
         else {
             return ActionExecution::InvalidAction;
         };
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] FileGlobExecutor::execute: patterns={:?}, path={:?}",
+            "FileGlobExecutor::execute: patterns={:?}, path={:?}",
             patterns,
             path
         );

app/src/ai/blocklist/action_model/execute/grep.rs

@@ -237,8 +237,8 @@ impl GrepExecutor {
         else {
             return ActionExecution::InvalidAction;
         };
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] GrepExecutor::execute: queries={:?}, path={:?}",
+            "GrepExecutor::execute: queries={:?}, path={:?}",
             queries,
             path
         );

app/src/ai/blocklist/action_model/execute/read_files.rs

@@ -91,8 +91,8 @@ impl ReadFilesExecutor {
         else {
             return ActionExecution::InvalidAction;
         };
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] ReadFilesExecutor::execute: {} files requested",
+            "ReadFilesExecutor::execute: {} files requested",
             locations.len()
         );
 

app/src/ai/blocklist/action_model/execute/request_file_edits.rs

@@ -173,14 +173,14 @@ impl RequestFileEditsExecutor {
         else {
             return ActionExecution::InvalidAction;
         };
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] RequestFileEditsExecutor::execute: action_id={:?}",
+            "RequestFileEditsExecutor::execute: action_id={:?}",
             id
         );
 
         let Some(diff_view) = self.diff_views.get(id) else {
-            log::warn!(
+            crate::ai::tool_diagnostics::tool_debug!(
-                "[tool-debug] RequestFileEditsExecutor: no diff view found for action_id={:?}",
+                "RequestFileEditsExecutor: no diff view found for action_id={:?}",
                 id
             );
             return ActionExecution::NotReady;

app/src/ai/blocklist/action_model/execute/run_agents.rs

@@ -21,7 +21,7 @@ use warpui::{Entity, EntityId, ModelContext, ModelHandle, SingletonEntity};
 
 use super::start_agent::{
     StartAgentDispatch, StartAgentExecutor, StartAgentExecutorEvent, StartAgentOutcome,
-    StartAgentWaitPolicy,
+    StartAgentRequestId, StartAgentWaitPolicy,
 };
 use super::{
     child_agent_delegation_denial_reason, ActionExecution, AnyActionExecution, ExecuteActionInput,
@@ -146,9 +146,12 @@ impl RunAgentsExecutor {
         ctx: &mut ModelContext<Self>,
     ) {
         self.recovery_action_ids.remove(action_id);
-        self.start_agent_executor.update(ctx, |executor, _| {
+        let detached_dispatches = self.start_agent_executor.update(ctx, |executor, _| {
-            executor.cancel_dispatches_for_action(action_id);
+            executor.cancel_dispatches_for_action(action_id)
         });
+        log::info!(
+            "RunAgents cancellation detached {detached_dispatches} pending child dispatch(es) for action {action_id}"
+        );
         if self.pending.remove(action_id).is_some() {
             ctx.emit(RunAgentsExecutorEvent::SpawningFinished {
                 action_id: action_id.clone(),
@@ -163,6 +166,22 @@ impl RunAgentsExecutor {
     ) {
         for agent in agents {
             let RunAgentsAgentOutcomeKind::Launched { agent_id } = &agent.kind else {
+                let RunAgentsAgentOutcomeKind::Completed { agent_id, .. } = &agent.kind else {
+                    continue;
+                };
+                let Some(normalized_name) = normalize_agent_name(&agent.name) else {
+                    continue;
+                };
+                self.launched_agents
+                    .entry(conversation_id)
+                    .or_default()
+                    .insert(
+                        normalized_name,
+                        ExistingLaunchedAgent {
+                            name: agent.name.clone(),
+                            agent_id: agent_id.clone(),
+                        },
+                    );
                 continue;
             };
             let Some(normalized_name) = normalize_agent_name(&agent.name) else {
@@ -372,6 +391,10 @@ impl RunAgentsExecutor {
         );
 
         let mut slots: Vec<ChildSlot> = Vec::with_capacity(agent_run_configs.len());
+        let wait_policy = match &run_execution_mode {
+            RunAgentsExecutionMode::Local => StartAgentWaitPolicy::Completion,
+            RunAgentsExecutionMode::Remote { .. } => StartAgentWaitPolicy::Startup,
+        };
         for cfg in &agent_run_configs {
             let normalized_name = normalize_agent_name(&cfg.name)
                 .expect("validated RunAgents requests have non-empty agent names");
@@ -382,7 +405,7 @@ impl RunAgentsExecutor {
                         cfg.name.clone(),
                         parent_conversation_id,
                         child_conversation_id,
-                        parent_run_id.clone(),
+                        wait_policy,
                         exec_ctx,
                     )
                 });
@@ -485,9 +508,9 @@ impl RunAgentsExecutor {
 
         ctx.spawn(
             async move {
-                let outcomes = join_all(slots.into_iter().map(resolve_child_slot)).await;
+                let resolved_slots = join_all(slots.into_iter().map(resolve_child_slot)).await;
                 #[cfg(not(target_family = "wasm"))]
-                for (slot_index, kind) in outcomes.iter().enumerate() {
+                for (slot_index, resolved) in resolved_slots.iter().enumerate() {
                     log::info!(
                         "RunAgents child launch outcome action_id={} parent_conversation_id={} \
                          agent_name={} slot_index={} outcome={}",
@@ -498,21 +521,32 @@ impl RunAgentsExecutor {
                             .map(String::as_str)
                             .unwrap_or("<unknown>"),
                         slot_index,
-                        run_agents_agent_outcome_kind_label(kind)
+                        run_agents_agent_outcome_kind_label(&resolved.outcome)
                     );
                 }
-                outcomes
+                resolved_slots
             },
-            move |me, outcomes, ctx| {
+            move |me, resolved_slots, ctx| {
                 if !me.is_pending(&action_id_for_aggr) {
                     return;
                 }
+                let timed_out_request_ids = resolved_slots
+                    .iter()
+                    .filter_map(|resolved| resolved.timed_out_request_id)
+                    .collect::<Vec<_>>();
+                if !timed_out_request_ids.is_empty() {
+                    me.start_agent_executor.update(ctx, |executor, _| {
+                        for request_id in timed_out_request_ids {
+                            executor.detach_dispatch(request_id);
+                        }
+                    });
+                }
                 let agents: Vec<RunAgentsAgentOutcome> = agent_run_configs_for_result
                     .iter()
-                    .zip(outcomes)
+                    .zip(resolved_slots)
-                    .map(|(cfg, kind)| RunAgentsAgentOutcome {
+                    .map(|(cfg, resolved)| RunAgentsAgentOutcome {
                         name: cfg.name.clone(),
-                        kind,
+                        kind: resolved.outcome,
                     })
                     .collect();
                 me.record_launched_agents(parent_conversation_id_for_result, &agents);
@@ -526,7 +560,7 @@ impl RunAgentsExecutor {
                         "action_id": action_id_for_aggr.to_string(),
                         "parent_conversation_id": parent_conversation_id_for_result.to_string(),
                         "agent_count": agents.len(),
-                        "launched_count": agents.iter().filter(|agent| matches!(agent.kind, RunAgentsAgentOutcomeKind::Launched { .. })).count(),
+                        "launched_count": agents.iter().filter(|agent| matches!(agent.kind, RunAgentsAgentOutcomeKind::Launched { .. } | RunAgentsAgentOutcomeKind::Completed { .. })).count(),
                         "failed_count": agents.iter().filter(|agent| matches!(agent.kind, RunAgentsAgentOutcomeKind::Failed { .. })).count(),
                         "agents": agents
                             .iter()
@@ -536,6 +570,12 @@ impl RunAgentsExecutor {
                                     "status": "launched",
                                     "agent_id": agent_id.as_str(),
                                 }),
+                                RunAgentsAgentOutcomeKind::Completed { agent_id, output } => serde_json::json!({
+                                    "name": agent.name.as_str(),
+                                    "status": "completed",
+                                    "agent_id": agent_id.as_str(),
+                                    "output": output,
+                                }),
                                 RunAgentsAgentOutcomeKind::Failed { error } => serde_json::json!({
                                     "name": agent.name.as_str(),
                                     "status": "failed",
@@ -723,6 +763,7 @@ fn start_agent_execution_mode_label(mode: &StartAgentExecutionMode) -> &'static
 fn run_agents_agent_outcome_kind_label(kind: &RunAgentsAgentOutcomeKind) -> &'static str {
     match kind {
         RunAgentsAgentOutcomeKind::Launched { .. } => "launched",
+        RunAgentsAgentOutcomeKind::Completed { .. } => "completed",
         RunAgentsAgentOutcomeKind::Failed { .. } => "failed",
     }
 }
@@ -732,18 +773,30 @@ enum ChildSlot {
     Pending(StartAgentDispatch),
 }
 
-async fn resolve_child_slot(slot: ChildSlot) -> RunAgentsAgentOutcomeKind {
+#[derive(Debug)]
+struct ResolvedChildSlot {
+    outcome: RunAgentsAgentOutcomeKind,
+    timed_out_request_id: Option<StartAgentRequestId>,
+}
+
+async fn resolve_child_slot(slot: ChildSlot) -> ResolvedChildSlot {
     resolve_child_slot_with_timeout(slot, SPAWN_TIMEOUT).await
 }
 
 async fn resolve_child_slot_with_timeout(
     slot: ChildSlot,
     spawn_timeout: Duration,
-) -> RunAgentsAgentOutcomeKind {
+) -> ResolvedChildSlot {
     let dispatch = match slot {
-        ChildSlot::Failed(error) => return RunAgentsAgentOutcomeKind::Failed { error },
+        ChildSlot::Failed(error) => {
+            return ResolvedChildSlot {
+                outcome: RunAgentsAgentOutcomeKind::Failed { error },
+                timed_out_request_id: None,
+            };
+        }
         ChildSlot::Pending(dispatch) => dispatch,
     };
+    let request_id = dispatch.request_id;
 
     let outcome = match dispatch.wait_policy {
         StartAgentWaitPolicy::Completion => dispatch.receiver.recv().await.ok(),
@@ -754,31 +807,41 @@ async fn resolve_child_slot_with_timeout(
             {
                 futures::future::Either::Left((outcome, _)) => outcome.ok(),
                 futures::future::Either::Right((_, _)) => {
+                    dispatch.mark_detached();
                     log::warn!(
                         "Agent spawn timed out after {} seconds",
                         spawn_timeout.as_secs()
                     );
-                    return RunAgentsAgentOutcomeKind::Failed {
+                    return ResolvedChildSlot {
-                        error: format!(
+                        outcome: RunAgentsAgentOutcomeKind::Failed {
-                            "Agent failed to start within {} seconds. \
+                            error: format!(
-                             The harness binary may not be installed.",
+                                "Agent failed to start within {} seconds. \
-                            spawn_timeout.as_secs()
+                                 The harness binary may not be installed.",
-                        ),
+                                spawn_timeout.as_secs()
+                            ),
+                        },
+                        timed_out_request_id: Some(request_id),
                     };
                 }
             }
         }
     };
 
-    match outcome {
+    let outcome = match outcome {
-        Some(StartAgentOutcome::Started { agent_id })
+        Some(StartAgentOutcome::Started { agent_id }) => {
-        | Some(StartAgentOutcome::Completed { agent_id, .. }) => {
             RunAgentsAgentOutcomeKind::Launched { agent_id }
         }
+        Some(StartAgentOutcome::Completed { agent_id, output }) => {
+            RunAgentsAgentOutcomeKind::Completed { agent_id, output }
+        }
         Some(StartAgentOutcome::Error(error)) => RunAgentsAgentOutcomeKind::Failed { error },
         None => RunAgentsAgentOutcomeKind::Failed {
             error: "Child agent was cancelled before completion".to_string(),
         },
+    };
+    ResolvedChildSlot {
+        outcome,
+        timed_out_request_id: None,
     }
 }
 
@@ -967,6 +1030,19 @@ fn existing_launched_agents_for_conversation(
                 };
                 for agent in agents {
                     let RunAgentsAgentOutcomeKind::Launched { agent_id } = &agent.kind else {
+                        let RunAgentsAgentOutcomeKind::Completed { agent_id, .. } = &agent.kind
+                        else {
+                            continue;
+                        };
+                        let Some(normalized_name) = normalize_agent_name(&agent.name) else {
+                            continue;
+                        };
+                        existing_agents.entry(normalized_name).or_insert_with(|| {
+                            ExistingLaunchedAgent {
+                                name: agent.name.clone(),
+                                agent_id: agent_id.clone(),
+                            }
+                        });
                         continue;
                     };
                     let Some(normalized_name) = normalize_agent_name(&agent.name) else {

app/src/ai/blocklist/action_model/execute/run_agents_tests.rs

@@ -1,4 +1,6 @@
 use std::collections::HashMap;
+use std::sync::atomic::AtomicBool;
+use std::sync::Arc;
 
 use ai::agent::action::{RunAgentsAgentRunConfig, RunAgentsExecutionMode, RunAgentsRequest};
 use ai::agent::orchestration_config::{
@@ -525,12 +527,16 @@ fn completion_slots_are_polled_concurrently_and_preserve_request_order() {
         let (second_sender, second_receiver) = async_channel::bounded(1);
         let slots = vec![
             ChildSlot::Pending(StartAgentDispatch {
+                request_id: StartAgentRequestId::from_raw_for_test(1),
                 receiver: first_receiver,
                 wait_policy: StartAgentWaitPolicy::Completion,
+                detached: Arc::new(AtomicBool::new(false)),
             }),
             ChildSlot::Pending(StartAgentDispatch {
+                request_id: StartAgentRequestId::from_raw_for_test(2),
                 receiver: second_receiver,
                 wait_policy: StartAgentWaitPolicy::Completion,
+                detached: Arc::new(AtomicBool::new(false)),
             }),
             ChildSlot::Failed("prelaunch failure".to_string()),
         ];
@@ -557,15 +563,16 @@ fn completion_slots_are_polled_concurrently_and_preserve_request_order() {
         let outcomes = outcomes.await;
 
         assert!(matches!(
-            &outcomes[0],
+            &outcomes[0].outcome,
             RunAgentsAgentOutcomeKind::Failed { error } if error == "first failed"
         ));
         assert!(matches!(
-            &outcomes[1],
+            &outcomes[1].outcome,
-            RunAgentsAgentOutcomeKind::Launched { agent_id } if agent_id == "second-agent"
+            RunAgentsAgentOutcomeKind::Completed { agent_id, output }
+                if agent_id == "second-agent" && output == "done"
         ));
         assert!(matches!(
-            &outcomes[2],
+            &outcomes[2].outcome,
             RunAgentsAgentOutcomeKind::Failed { error } if error == "prelaunch failure"
         ));
     });
@@ -577,8 +584,10 @@ fn completion_wait_ignores_spawn_timeout() {
         let (sender, receiver) = async_channel::bounded(1);
         let completion = Box::pin(resolve_child_slot_with_timeout(
             ChildSlot::Pending(StartAgentDispatch {
+                request_id: StartAgentRequestId::from_raw_for_test(1),
                 receiver,
                 wait_policy: StartAgentWaitPolicy::Completion,
+                detached: Arc::new(AtomicBool::new(false)),
             }),
             Duration::from_millis(1),
         ));
@@ -598,8 +607,9 @@ fn completion_wait_ignores_spawn_timeout() {
             .unwrap();
 
         assert!(matches!(
-            completion.await,
+            completion.await.outcome,
-            RunAgentsAgentOutcomeKind::Launched { agent_id } if agent_id == "child-agent"
+            RunAgentsAgentOutcomeKind::Completed { agent_id, output }
+                if agent_id == "child-agent" && output == "done"
         ));
     });
 }
@@ -610,21 +620,73 @@ fn startup_wait_retains_spawn_timeout() {
         let (_sender, receiver) = async_channel::bounded(1);
         let outcome = resolve_child_slot_with_timeout(
             ChildSlot::Pending(StartAgentDispatch {
+                request_id: StartAgentRequestId::from_raw_for_test(1),
                 receiver,
                 wait_policy: StartAgentWaitPolicy::Startup,
+                detached: Arc::new(AtomicBool::new(false)),
             }),
             Duration::from_millis(1),
         )
         .await;
 
+        assert!(outcome.timed_out_request_id.is_some());
+        assert_eq!(
+            outcome.timed_out_request_id,
+            Some(StartAgentRequestId::from_raw_for_test(1))
+        );
         assert!(matches!(
-            outcome,
+            outcome.outcome,
             RunAgentsAgentOutcomeKind::Failed { error }
                 if error.contains("Agent failed to start within")
         ));
     });
 }
 
+#[test]
+fn startup_timeout_detaches_exact_pending_request() {
+    App::test((), |mut app| async move {
+        let state = initialize_run_agents_test(&mut app, ExecutionMode::App);
+        let start_agent_executor = state.start_agent_executor;
+        let parent_conversation_id = state.conversation_id;
+        let dispatch = start_agent_executor.update(&mut app, |executor, ctx| {
+            executor.dispatch(
+                AIAgentActionId::from("run-agents-timeout".to_string()),
+                "child".to_string(),
+                "work".to_string(),
+                StartAgentExecutionMode::Remote {
+                    environment_id: "environment".to_string(),
+                    skill_references: Vec::new(),
+                    model_id: "model".to_string(),
+                    computer_use_enabled: false,
+                    worker_host: String::new(),
+                    harness_type: "oz".to_string(),
+                    title: String::new(),
+                    auth_secret_name: None,
+                },
+                None,
+                parent_conversation_id,
+                Some("parent-run".to_string()),
+                ctx,
+            )
+        });
+        let request_id = dispatch.request_id;
+
+        let resolved =
+            resolve_child_slot_with_timeout(ChildSlot::Pending(dispatch), Duration::from_millis(1))
+                .await;
+        let timed_out_request_id = resolved
+            .timed_out_request_id
+            .expect("startup timeout should expose request identity");
+        start_agent_executor.update(&mut app, |executor, _| {
+            assert!(executor.detach_dispatch(timed_out_request_id));
+        });
+
+        start_agent_executor.read(&app, |executor, _| {
+            assert!(!executor.has_pending_dispatch_for_test(request_id));
+        });
+    });
+}
+
 fn initialize_run_agents_test(app: &mut App, mode: ExecutionMode) -> RunAgentsTestState {
     initialize_settings_for_tests_with_mode(app, mode, false);
     let global_resource_handles = GlobalResourceHandles::mock(app);

app/src/ai/blocklist/action_model/execute/shell_command.rs

@@ -13,7 +13,6 @@ use galaxy_core::execution_mode::AppExecutionMode;
 use galaxy_util::path::ShellFamily;
 use galaxyui::r#async::{Spawnable, Timer};
 use galaxyui::{Entity, EntityId, ModelContext, ModelHandle, SingletonEntity};
-use itertools::Itertools;
 use parking_lot::FairMutex;
 
 use super::{ActionExecution, AnyActionExecution, ExecuteActionInput, PreprocessActionInput};
@@ -37,11 +36,11 @@ use crate::{send_telemetry_from_ctx, TelemetryEvent};
 
 pub struct ShellCommandExecutor {
     active_session: ModelHandle<ActiveSession>,
-    block_finished_senders: HashMap<BlockSelector, oneshot::Sender<()>>,
+    block_finished_senders: HashMap<BlockSelector, Vec<oneshot::Sender<()>>>,
     /// Senders used by `Check now` and the automatic monitor watchdog to force a long-running
     /// shell command's pending poll future to resolve immediately with a fresh snapshot,
     /// bypassing the agent-set timeout.
-    force_refresh_senders: HashMap<BlockSelector, oneshot::Sender<()>>,
+    force_refresh_senders: HashMap<BlockSelector, Vec<oneshot::Sender<()>>>,
     terminal_model: Arc<FairMutex<TerminalModel>>,
     terminal_view_id: EntityId,
     /// Sender to notify when user hands control back to agent after TransferShellCommandControlToUser.
@@ -80,24 +79,39 @@ impl ShellCommandExecutor {
         event: &ModelEvent,
         _ctx: &mut ModelContext<Self>,
     ) {
-        // We wait for precmd for the block _after_ the requested command's block so that
+        // Precmd provides fresh CWD metadata, while BlockCompleted is definitive completion
-        // downstream checks for current working directory are fresh. The precmd hook is when
+        // evidence for shells that never deliver a subsequent precmd.
-        // the shell relays current working directory to warp.
+        if matches!(
-        if let ModelEvent::BlockMetadataReceived(BlockMetadataReceivedEvent { .. }) = event {
+            event,
+            ModelEvent::BlockMetadataReceived(BlockMetadataReceivedEvent { .. })
+                | ModelEvent::BlockCompleted(_)
+        ) {
             let model = self.terminal_model.lock();
-            let block_finished_senders = self.block_finished_senders.drain().collect_vec();
+            let block_finished_senders = self.block_finished_senders.drain().collect::<Vec<_>>();
-            for (block_selector, block_finished_tx) in block_finished_senders.into_iter() {
+            for (block_selector, block_finished_txs) in block_finished_senders {
-                if let Some(block) = block_selector.get_block(&model) {
+                let completed_block = block_selector.get_block(&model).filter(|block| {
-                    if block.is_command_finished() {
+                    block.is_command_finished()
+                        && match event {
+                            ModelEvent::BlockCompleted(completed) => {
+                                block.id() == &completed.block_id
+                            }
+                            ModelEvent::BlockMetadataReceived(_) => true,
+                            _ => false,
+                        }
+                });
+                if completed_block.is_some() {
+                    for block_finished_tx in block_finished_txs {
                         if let Err(e) = block_finished_tx.send(()) {
                             log::warn!(
                                 "Failed to notify block completion for running requested command: {e:?}"
                             )
                         }
-                    } else {
-                        self.block_finished_senders
-                            .insert(block_selector, block_finished_tx);
                     }
+                } else {
+                    // The requested-command association may not exist yet. Keep all waiters until
+                    // this selector resolves and its block actually completes, or it is cancelled.
+                    self.block_finished_senders
+                        .insert(block_selector, block_finished_txs);
                 }
             }
         }
@@ -195,8 +209,8 @@ impl ShellCommandExecutor {
         input: ExecuteActionInput,
         ctx: &mut ModelContext<Self>,
     ) -> impl Into<AnyActionExecution> {
-        log::info!(
+        crate::ai::tool_diagnostics::tool_debug!(
-            "[tool-debug] ShellCommandExecutor::execute: action_type={:?}",
+            "ShellCommandExecutor::execute: action_type={:?}",
             std::mem::discriminant(&input.action.action)
         );
         let model = self.terminal_model.lock();
@@ -204,12 +218,6 @@ impl ShellCommandExecutor {
         // Determine the action we want to take based on the input.
         let action_id = input.action.id.clone();
 
-        let command = model
-            .block_list()
-            .active_block()
-            .command_with_secrets_unobfuscated(false)
-            .clone();
-
         let handle = ctx.handle();
         match &input.action.action {
             AIAgentActionType::RequestCommandOutput {
@@ -222,18 +230,13 @@ impl ShellCommandExecutor {
                     .active_block()
                     .is_active_and_long_running()
                 {
-                    // Another command is still running (e.g. stuck in a pager). Return an error
+                    let running_command = model
-                    // result so the model receives feedback and can adapt. Using Completed with a
+                        .block_list()
-                    // non-zero exit code ensures a follow-up request is triggered.
+                        .active_block()
-                    return ActionExecution::Sync(AIAgentActionResultType::RequestCommandOutput(
+                        .command_with_secrets_unobfuscated(false);
-                        RequestCommandOutputResult::Completed {
+                    return ActionExecution::Sync(terminal_busy_execution_error(
-                            command: command.clone(),
+                        command,
-                            block_id: model.block_list().active_block().id().clone(),
+                        &running_command,
-                            output: "Error: Cannot execute command because another command is still running in the terminal.".to_string(),
-                            exit_code: ExitCode::from(1),
-                            start_ts: None,
-                            completed_ts: None,
-                        },
                     ));
                 }
                 // If another conversation has taken over the agent view since this command
@@ -275,8 +278,7 @@ impl ShellCommandExecutor {
                         // Remove the senders from the maps.
                         if let Some(handle) = handle.upgrade(ctx) {
                             handle.update(ctx, |me, _| {
-                                me.block_finished_senders.remove(&block_selector);
+                                me.prune_closed_senders(&block_selector);
-                                me.force_refresh_senders.remove(&block_selector);
                             });
                         }
 
@@ -339,8 +341,7 @@ impl ShellCommandExecutor {
                         // Remove the senders from the maps.
                         if let Some(handle) = handle.upgrade(ctx) {
                             handle.update(ctx, |me, _| {
-                                me.block_finished_senders.remove(&block_selector);
+                                me.prune_closed_senders(&block_selector);
-                                me.force_refresh_senders.remove(&block_selector);
                             });
                         }
 
@@ -371,6 +372,7 @@ impl ShellCommandExecutor {
                         },
                     ));
                 }
+                let command = block.command_with_secrets_unobfuscated(false);
                 drop(model);
 
                 let block_selector = BlockSelector::Id(block_id.clone());
@@ -380,8 +382,7 @@ impl ShellCommandExecutor {
                         // Remove the senders from the maps.
                         if let Some(handle) = handle.upgrade(ctx) {
                             handle.update(ctx, |me, _| {
-                                me.block_finished_senders.remove(&block_selector);
+                                me.prune_closed_senders(&block_selector);
-                                me.force_refresh_senders.remove(&block_selector);
                             });
                         }
 
@@ -419,7 +420,9 @@ impl ShellCommandExecutor {
                 // Set up a future to also wait for block completion.
                 let (block_finished_tx, block_finished_rx) = oneshot::channel();
                 self.block_finished_senders
-                    .insert(block_selector.clone(), block_finished_tx);
+                    .entry(block_selector.clone())
+                    .or_default()
+                    .push(block_finished_tx);
 
                 // Build the future that captures terminal model and block data.
                 let transfer_future = {
@@ -491,7 +494,7 @@ impl ShellCommandExecutor {
                     // Clean up.
                     if let Some(handle) = handle.upgrade(ctx) {
                         handle.update(ctx, |me, _| {
-                            me.block_finished_senders.remove(&block_selector);
+                            me.prune_closed_senders(&block_selector);
                             me.control_handback_sender = None;
                         });
                     }
@@ -520,13 +523,17 @@ impl ShellCommandExecutor {
         // Create a channel to notify us when we receive block metadata.
         let (block_metadata_received_tx, block_metadata_received_rx) = oneshot::channel();
         self.block_finished_senders
-            .insert(block_selector.clone(), block_metadata_received_tx);
+            .entry(block_selector.clone())
+            .or_default()
+            .push(block_metadata_received_tx);
 
         // Create a channel so `Check now` or the automatic monitor watchdog can short-circuit
         // the timeout and deliver the agent a fresh snapshot immediately.
         let (force_refresh_tx, force_refresh_rx) = oneshot::channel();
         self.force_refresh_senders
-            .insert(block_selector.clone(), force_refresh_tx);
+            .entry(block_selector.clone())
+            .or_default()
+            .push(force_refresh_tx);
 
         // Create a future that resolves when we should send a result to the agent.
         let terminal_model = self.terminal_model.clone();
@@ -600,7 +607,12 @@ impl ShellCommandExecutor {
                             completed_ts: block.completed_ts().cloned(),
                         }
                     } else {
-                        let grid_contents = if model.is_alt_screen_active() {
+                        let selected_block_owns_alt_screen = selected_block_owns_alt_screen(
+                            model.is_alt_screen_active(),
+                            model.active_block_id(),
+                            block.id(),
+                        );
+                        let grid_contents = if selected_block_owns_alt_screen {
                             formatted_terminal_contents_for_input(
                                 model.alt_screen().grid_handler(),
                                 None,
@@ -618,7 +630,7 @@ impl ShellCommandExecutor {
                             block_id: block.id().clone(),
                             grid_contents,
                             cursor: CURSOR_MARKER,
-                            is_alt_screen_active: model.is_alt_screen_active(),
+                            is_alt_screen_active: selected_block_owns_alt_screen,
                             is_preempted,
                         }
                     }
@@ -630,23 +642,50 @@ impl ShellCommandExecutor {
         }
     }
 
-    pub(super) fn cancel_execution(&mut self, id: &AIAgentActionId, _ctx: &mut ModelContext<Self>) {
+    pub(super) fn cancel_execution(
+        &mut self,
+        id: &AIAgentActionId,
+        ctx: &mut ModelContext<Self>,
+    ) -> bool {
         let terminal_model = self.terminal_model.lock();
-        let active_block = terminal_model.block_list().active_block();
+        let requested_selector = BlockSelector::RequestedCommandId(id.clone());
-        if !active_block.is_active_and_long_running() {
+        let requested_block_is_running = requested_selector
-            return;
+            .get_block(&terminal_model)
-        }
+            .is_some_and(|block| block.is_active_and_long_running() && !block.finished());
-
+        let selector = if requested_block_is_running {
-        let selector = if active_block
+            requested_selector
-            .requested_command_action_id()
-            .is_some_and(|requested_command_id| requested_command_id == id)
-        {
-            BlockSelector::RequestedCommandId(id.clone())
         } else {
-            BlockSelector::Id(active_block.id().clone())
+            BlockSelector::Id(terminal_model.active_block_id().clone())
         };
-        self.block_finished_senders.remove(&selector);
+        // Cancelling the wait future alone would report cancellation while the process keeps
-        self.force_refresh_senders.remove(&selector);
+        // running. Terminate the exact requested command before resolving the action as cancelled.
+        if requested_block_is_running {
+            ctx.emit(ShellCommandExecutorEvent::CancelExecution {
+                action_id: id.clone(),
+            });
+        }
+        if !requested_block_is_running {
+            self.block_finished_senders.remove(&selector);
+            self.force_refresh_senders.remove(&selector);
+        }
+        requested_block_is_running
+    }
+
+    fn prune_closed_senders(&mut self, selector: &BlockSelector) {
+        Self::prune_closed_sender_group(&mut self.block_finished_senders, selector);
+        Self::prune_closed_sender_group(&mut self.force_refresh_senders, selector);
+    }
+
+    fn prune_closed_sender_group(
+        senders: &mut HashMap<BlockSelector, Vec<oneshot::Sender<()>>>,
+        selector: &BlockSelector,
+    ) {
+        if let Some(selector_senders) = senders.get_mut(selector) {
+            selector_senders.retain(|sender| !sender.is_canceled());
+            if selector_senders.is_empty() {
+                senders.remove(selector);
+            }
+        }
     }
 
     /// Force any in-flight poll for the given long-running command block to resolve
@@ -657,9 +696,8 @@ impl ShellCommandExecutor {
     /// control to the user). Returns whether a matching poll was successfully refreshed.
     pub fn force_refresh_block(&mut self, block_id: &BlockId) -> bool {
         let terminal_model = self.terminal_model.lock();
-        // Find a sender whose selector resolves to this block. In practice there is at
+        // Find every pending poll whose selector resolves to this block. Multiple provider polls
-        // most one: a given block can have at most one in-flight `action_result_future`
+        // may legitimately wait on the same command and must be refreshed together.
-        // at a time.
         let matching_selector = self
             .force_refresh_senders
             .keys()
@@ -674,8 +712,12 @@ impl ShellCommandExecutor {
         drop(terminal_model);
 
         if let Some(selector) = matching_selector {
-            if let Some(sender) = self.force_refresh_senders.remove(&selector) {
+            if let Some(senders) = self.force_refresh_senders.remove(&selector) {
-                return sender.send(()).is_ok();
+                let mut refreshed = false;
+                for sender in senders {
+                    refreshed |= sender.send(()).is_ok();
+                }
+                return refreshed;
             }
         }
         false
@@ -714,6 +756,21 @@ fn command_for_execution(
     }
 }
 
+fn terminal_busy_execution_error(command: &str, running_command: &str) -> AIAgentActionResultType {
+    AIAgentActionResultType::RequestCommandOutput(RequestCommandOutputResult::ExecutionError {
+        command: command.to_string(),
+        message: format!("terminal is busy running command '{running_command}'"),
+    })
+}
+
+fn selected_block_owns_alt_screen(
+    is_alt_screen_active: bool,
+    active_block_id: &BlockId,
+    selected_block_id: &BlockId,
+) -> bool {
+    is_alt_screen_active && active_block_id == selected_block_id
+}
+
 #[derive(Debug, Clone, Hash, PartialEq, Eq)]
 enum BlockSelector {
     Id(BlockId),
@@ -919,7 +976,9 @@ pub enum ShellCommandExecutorEvent {
         input: Bytes,
         mode: AIAgentPtyWriteMode,
     },
-    CancelExecution,
+    CancelExecution {
+        action_id: AIAgentActionId,
+    },
     /// Emitted when the agent requests to transfer control of a long-running command to the user.
     TransferControlToUser {
         action_id: AIAgentActionId,

app/src/ai/blocklist/action_model/execute/shell_command_tests.rs

@@ -1,19 +1,29 @@
 use std::sync::Arc;
+use std::task::Poll;
 
 use async_channel::unbounded;
 use futures::channel::oneshot;
+use futures::{pin_mut, poll};
 use parking_lot::FairMutex;
 use warpui::{App, EntityId};
 
-use super::{command_for_execution, ActionResult, BlockSelector, ShellCommandExecutor};
+use super::{
-use crate::ai::agent::ShellCommandDelay;
+    command_for_execution, selected_block_owns_alt_screen, terminal_busy_execution_error,
-use crate::terminal::event::{BlockMetadataReceivedEvent, BlockWorkingDirectoryUpdatedEvent};
+    ActionResult, BlockSelector, ShellCommandExecutor,
+};
+use crate::ai::agent::{
+    AIAgentActionId, AIAgentActionResultType, RequestCommandOutputResult, ShellCommandDelay,
+};
+use crate::terminal::event::{
+    BlockCompletedEvent, BlockMetadataReceivedEvent, BlockType, BlockWorkingDirectoryUpdatedEvent,
+};
 use crate::terminal::model::block::{BlockId, BlockMetadata};
 use crate::terminal::model::session::active_session::ActiveSession;
 use crate::terminal::model::session::Sessions;
 use crate::terminal::model::terminal_model::{BlockIndex, TerminalModel};
 use crate::terminal::model_events::{ModelEvent, ModelEventDispatcher};
 use crate::terminal::shell::ShellType;
+use crate::AIConversationId;
 
 #[test]
 fn wait_commands_disable_implicit_posix_pagers_without_masking_exit_status() {
@@ -31,6 +41,41 @@ fn wait_commands_disable_implicit_posix_pagers_without_masking_exit_status() {
     );
 }
 
+#[test]
+fn terminal_busy_is_an_execution_error_for_the_unstarted_command() {
+    let result = terminal_busy_execution_error("cargo test", "sleep 120");
+
+    assert!(matches!(
+        result,
+        AIAgentActionResultType::RequestCommandOutput(
+            RequestCommandOutputResult::ExecutionError { command, message }
+        ) if command == "cargo test"
+            && message == "terminal is busy running command 'sleep 120'"
+    ));
+}
+
+#[test]
+fn targeted_poll_uses_alt_screen_only_for_its_owning_block() {
+    let active_block_id = BlockId::new();
+    let selected_block_id = BlockId::new();
+
+    assert!(!selected_block_owns_alt_screen(
+        true,
+        &active_block_id,
+        &selected_block_id
+    ));
+    assert!(selected_block_owns_alt_screen(
+        true,
+        &active_block_id,
+        &active_block_id
+    ));
+    assert!(!selected_block_owns_alt_screen(
+        false,
+        &active_block_id,
+        &active_block_id
+    ));
+}
+
 /// Locks in the contract that `ShellCommandExecutor`'s requested-command finish
 /// detector reacts only to `BlockMetadataReceived` (precmd) and not to
 /// `BlockWorkingDirectoryUpdated` (OSC 7). The detector relies on
@@ -63,7 +108,7 @@ fn block_working_directory_updated_does_not_drain_finish_senders() {
         let selector = BlockSelector::Id(block_id);
         let (tx, _rx) = oneshot::channel::<()>();
         executor.update(&mut app, |executor, _ctx| {
-            executor.block_finished_senders.insert(selector, tx);
+            executor.block_finished_senders.insert(selector, vec![tx]);
         });
         assert_eq!(
             app.read(|ctx| executor.as_ref(ctx).block_finished_senders.len()),
@@ -88,8 +133,7 @@ fn block_working_directory_updated_does_not_drain_finish_senders() {
              that map is reserved for precmd (BlockMetadataReceived)"
         );
 
-        // Precmd event — the senders map should be drained (and since the
+        // An unrelated precmd cannot resolve this selector, so its waiter must survive.
-        // block isn't in the terminal model, the sender is dropped).
         model_event_dispatcher.update(&mut app, |_dispatcher, ctx| {
             ctx.emit(ModelEvent::BlockMetadataReceived(
                 BlockMetadataReceivedEvent {
@@ -102,8 +146,8 @@ fn block_working_directory_updated_does_not_drain_finish_senders() {
         });
         assert_eq!(
             app.read(|ctx| executor.as_ref(ctx).block_finished_senders.len()),
-            0,
+            1,
-            "BlockMetadataReceived should drain the finish senders"
+            "BlockMetadataReceived must retain unresolved finish senders"
         );
     });
 }
@@ -138,7 +182,7 @@ fn force_refresh_block_reports_and_resolves_matching_poll() {
         executor.update(&mut app, |executor, _| {
             executor
                 .force_refresh_senders
-                .insert(BlockSelector::Id(block_id.clone()), tx);
+                .insert(BlockSelector::Id(block_id.clone()), vec![tx]);
             assert!(executor.force_refresh_block(&block_id));
             assert!(!executor.force_refresh_block(&block_id));
         });
@@ -149,7 +193,7 @@ fn force_refresh_block_reports_and_resolves_matching_poll() {
         executor.update(&mut app, |executor, _| {
             executor
                 .force_refresh_senders
-                .insert(BlockSelector::Id(block_id.clone()), tx);
+                .insert(BlockSelector::Id(block_id.clone()), vec![tx]);
         });
         terminal_model.lock().finish_block();
         assert!(executor.update(&mut app, |executor, _| {
@@ -158,6 +202,150 @@ fn force_refresh_block_reports_and_resolves_matching_poll() {
     });
 }
 
+#[test]
+fn requested_command_waiter_survives_early_metadata_and_resolves_after_association() {
+    App::test((), |mut app| async move {
+        let terminal_view_id = EntityId::new();
+        let sessions = app.add_model(|_| Sessions::new_for_test());
+        let (_model_events_tx, model_events_rx) = unbounded();
+        let model_event_dispatcher =
+            app.add_model(|ctx| ModelEventDispatcher::new(model_events_rx, sessions.clone(), ctx));
+        let active_session = app.add_model(|ctx| {
+            ActiveSession::new(sessions.clone(), model_event_dispatcher.clone(), ctx)
+        });
+        let terminal_model = Arc::new(FairMutex::new(TerminalModel::mock(None, None)));
+        let executor = app.add_model(|ctx| {
+            ShellCommandExecutor::new(
+                active_session,
+                terminal_model.clone(),
+                &model_event_dispatcher,
+                terminal_view_id,
+                ctx,
+            )
+        });
+        let action_id = AIAgentActionId::from("requested-command".to_string());
+        let result_future = executor.update(&mut app, |executor, _| {
+            executor.action_result_future(
+                BlockSelector::RequestedCommandId(action_id.clone()),
+                Some(ShellCommandDelay::OnCompletion),
+            )
+        });
+        pin_mut!(result_future);
+
+        model_event_dispatcher.update(&mut app, |_dispatcher, ctx| {
+            ctx.emit(ModelEvent::BlockMetadataReceived(
+                BlockMetadataReceivedEvent {
+                    block_metadata: BlockMetadata::new(None, Some("/tmp/early".to_string())),
+                    block_index: BlockIndex::zero(),
+                    is_after_in_band_command: false,
+                    is_done_bootstrapping: true,
+                },
+            ));
+        });
+        assert!(matches!(poll!(&mut result_future), Poll::Pending));
+
+        terminal_model
+            .lock()
+            .simulate_long_running_block("printf done", "done");
+        let block_id = terminal_model.lock().active_block_id().clone();
+        terminal_model
+            .lock()
+            .block_list_mut()
+            .active_block_mut()
+            .set_agent_interaction_mode_for_requested_command(
+                action_id,
+                None,
+                AIConversationId::new(),
+            );
+        terminal_model.lock().finish_block();
+        model_event_dispatcher.update(&mut app, |_dispatcher, ctx| {
+            ctx.emit(ModelEvent::BlockCompleted(block_completed_event(
+                block_id.clone(),
+            )));
+        });
+
+        assert!(matches!(
+            result_future.await,
+            ActionResult::CommandFinished {
+                block_id: result_block_id,
+                ..
+            } if result_block_id == block_id
+        ));
+    });
+}
+
+#[test]
+fn duplicate_completion_polls_for_same_block_both_resolve_on_block_completed() {
+    App::test((), |mut app| async move {
+        let terminal_view_id = EntityId::new();
+        let sessions = app.add_model(|_| Sessions::new_for_test());
+        let (_model_events_tx, model_events_rx) = unbounded();
+        let model_event_dispatcher =
+            app.add_model(|ctx| ModelEventDispatcher::new(model_events_rx, sessions.clone(), ctx));
+        let active_session = app.add_model(|ctx| {
+            ActiveSession::new(sessions.clone(), model_event_dispatcher.clone(), ctx)
+        });
+        let terminal_model = Arc::new(FairMutex::new(TerminalModel::mock(None, None)));
+        terminal_model
+            .lock()
+            .simulate_long_running_block("sleep 1", "finished");
+        let block_id = terminal_model.lock().active_block_id().clone();
+        let executor = app.add_model(|ctx| {
+            ShellCommandExecutor::new(
+                active_session,
+                terminal_model.clone(),
+                &model_event_dispatcher,
+                terminal_view_id,
+                ctx,
+            )
+        });
+
+        let first = executor.update(&mut app, |executor, _| {
+            executor.action_result_future(
+                BlockSelector::Id(block_id.clone()),
+                Some(ShellCommandDelay::OnCompletion),
+            )
+        });
+        let second = executor.update(&mut app, |executor, _| {
+            executor.action_result_future(
+                BlockSelector::Id(block_id.clone()),
+                Some(ShellCommandDelay::OnCompletion),
+            )
+        });
+        pin_mut!(first);
+        pin_mut!(second);
+        assert!(matches!(poll!(&mut first), Poll::Pending));
+        assert!(matches!(poll!(&mut second), Poll::Pending));
+
+        terminal_model.lock().finish_block();
+        model_event_dispatcher.update(&mut app, |_dispatcher, ctx| {
+            ctx.emit(ModelEvent::BlockCompleted(block_completed_event(
+                block_id.clone(),
+            )));
+        });
+
+        let first_result = first.await;
+        let second_result = second.await;
+        assert!(matches!(first_result, ActionResult::CommandFinished { .. }));
+        assert!(matches!(
+            second_result,
+            ActionResult::CommandFinished { .. }
+        ));
+    });
+}
+
+fn block_completed_event(block_id: BlockId) -> BlockCompletedEvent {
+    BlockCompletedEvent {
+        block_latency_data: None,
+        block_type: BlockType::Restored,
+        num_secrets_obfuscated: 0,
+        block_index: BlockIndex::zero(),
+        block_id,
+        session_id: None,
+        restored_block_was_local: None,
+    }
+}
+
 #[test]
 fn force_refresh_wakes_on_completion_poll_with_preempted_snapshot() {
     App::test((), |mut app| async move {

app/src/ai/blocklist/action_model/execute/start_agent.rs

@@ -1,4 +1,6 @@
 use std::collections::HashMap;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
 
 use futures::future::BoxFuture;
 use futures::FutureExt;
@@ -42,9 +44,24 @@ pub enum StartAgentWaitPolicy {
     Completion,
 }
 
+fn wait_policy_for_execution_mode(mode: &StartAgentExecutionMode) -> StartAgentWaitPolicy {
+    match mode {
+        StartAgentExecutionMode::Local { .. } => StartAgentWaitPolicy::Completion,
+        StartAgentExecutionMode::Remote { .. } => StartAgentWaitPolicy::Startup,
+    }
+}
+
 pub struct StartAgentDispatch {
+    pub request_id: StartAgentRequestId,
     pub receiver: async_channel::Receiver<StartAgentOutcome>,
     pub wait_policy: StartAgentWaitPolicy,
+    pub(super) detached: Arc<AtomicBool>,
+}
+
+impl StartAgentDispatch {
+    pub(super) fn mark_detached(&self) {
+        self.detached.store(true, Ordering::Release);
+    }
 }
 
 fn invalid_local_child_harness_error(harness_type: &str) -> String {
@@ -139,6 +156,7 @@ struct PendingStartAgent {
     /// Set once the child conversation is synchronously created.
     child_conversation_id: Option<AIConversationId>,
     sender: async_channel::Sender<StartAgentOutcome>,
+    detached: Arc<AtomicBool>,
     /// Direct Bedrock/OpenAI parents do not have a server run id or an
     /// orchestration event stream. Keep the tool call open until their local
     /// child finishes, then return the child's output inline.
@@ -176,9 +194,17 @@ impl StartAgentExecutor {
         ctx: &mut ModelContext<Self>,
     ) {
         let child_link_event = {
-            let Some(pending) = self.pending.get_mut(&request_id) else {
+            let Some(pending) = self.pending.get(&request_id) else {
                 return;
             };
+            if pending.detached.load(Ordering::Acquire) {
+                self.pending.remove(&request_id);
+                return;
+            }
+            let pending = self
+                .pending
+                .get_mut(&request_id)
+                .expect("pending request was checked above");
             pending.child_conversation_id = Some(child_conversation_id);
             if let Some(agent_name) = pending.run_agents_child_name.clone() {
                 Some(StartAgentExecutorEvent::RunAgentsChildConversationCreated {
@@ -595,16 +621,13 @@ impl StartAgentExecutor {
             }
         };
 
-        // In local mode (no parent_run_id), block until the child finishes
+        // Local children return their completed work; remote children acknowledge startup and
-        // so the parent model receives the child's output as the tool result.
+        // continue through the hosted orchestration lifecycle.
-        let wait_policy = if parent_run_id.is_none() {
+        let wait_policy = wait_policy_for_execution_mode(&execution_mode);
-            StartAgentWaitPolicy::Completion
-        } else {
-            StartAgentWaitPolicy::Startup
-        };
 
         let (sender, receiver) = async_channel::bounded(1);
         let request_id = self.next_request_id();
+        let detached = Arc::new(AtomicBool::new(false));
         self.pending.insert(
             request_id,
             PendingStartAgent {
@@ -613,6 +636,7 @@ impl StartAgentExecutor {
                 parent_conversation_id,
                 child_conversation_id: None,
                 sender,
+                detached,
                 wait_policy,
             },
         );
@@ -667,24 +691,23 @@ impl StartAgentExecutor {
         parent_run_id: Option<String>,
         ctx: &mut ModelContext<Self>,
     ) -> StartAgentDispatch {
-        let wait_policy = if parent_run_id.is_none() {
+        let wait_policy = wait_policy_for_execution_mode(&execution_mode);
-            StartAgentWaitPolicy::Completion
-        } else {
-            StartAgentWaitPolicy::Startup
-        };
         let (sender, receiver) = async_channel::bounded(1);
+        let request_id = self.next_request_id();
+        let detached = Arc::new(AtomicBool::new(false));
         if let Some(error) = child_agent_delegation_denial_reason(parent_conversation_id, ctx) {
             let _ = sender.try_send(StartAgentOutcome::Error(error));
             return StartAgentDispatch {
+                request_id,
                 receiver,
                 wait_policy,
+                detached,
             };
         }
 
         let (prompt, execution_mode) =
             normalize_legacy_local_child_harness_command(prompt, execution_mode);
         let prompt = compose_leaf_agent_prompt(&prompt);
-        let request_id = self.next_request_id();
         self.pending.insert(
             request_id,
             PendingStartAgent {
@@ -693,6 +716,7 @@ impl StartAgentExecutor {
                 parent_conversation_id,
                 child_conversation_id: None,
                 sender,
+                detached: detached.clone(),
                 wait_policy,
             },
         );
@@ -708,8 +732,10 @@ impl StartAgentExecutor {
             },
         )));
         StartAgentDispatch {
+            request_id,
             receiver,
             wait_policy,
+            detached,
         }
     }
 
@@ -719,16 +745,12 @@ impl StartAgentExecutor {
         name: String,
         parent_conversation_id: AIConversationId,
         child_conversation_id: AIConversationId,
-        parent_run_id: Option<String>,
+        wait_policy: StartAgentWaitPolicy,
         ctx: &mut ModelContext<Self>,
     ) -> StartAgentDispatch {
-        let wait_policy = if parent_run_id.is_none() {
-            StartAgentWaitPolicy::Completion
-        } else {
-            StartAgentWaitPolicy::Startup
-        };
         let (sender, receiver) = async_channel::bounded(1);
         let request_id = self.next_request_id();
+        let detached = Arc::new(AtomicBool::new(false));
         self.pending.insert(
             request_id,
             PendingStartAgent {
@@ -737,19 +759,54 @@ impl StartAgentExecutor {
                 parent_conversation_id,
                 child_conversation_id: Some(child_conversation_id),
                 sender,
+                detached: detached.clone(),
                 wait_policy,
             },
         );
         self.record_child_conversation(request_id, child_conversation_id, ctx);
         StartAgentDispatch {
+            request_id,
             receiver,
             wait_policy,
+            detached,
         }
     }
 
-    pub fn cancel_dispatches_for_action(&mut self, action_id: &AIAgentActionId) {
+    /// Detaches one exact dispatch. If its launch callback is already queued,
-        self.pending
+    /// the shared marker prevents that callback from linking a late child.
-            .retain(|_, pending| &pending.action_id != action_id);
+    pub fn detach_dispatch(&mut self, request_id: StartAgentRequestId) -> bool {
+        let Some(pending) = self.pending.remove(&request_id) else {
+            return false;
+        };
+        pending.detached.store(true, Ordering::Release);
+        true
+    }
+
+    /// Test-only lookup for request ownership without exposing executor internals.
+    #[cfg(test)]
+    pub fn has_pending_dispatch_for_test(&self, request_id: StartAgentRequestId) -> bool {
+        self.pending.contains_key(&request_id)
+    }
+
+    pub fn cancel_dispatches_for_action(&mut self, action_id: &AIAgentActionId) -> usize {
+        let request_ids = self
+            .pending
+            .iter()
+            .filter_map(|(request_id, pending)| {
+                (&pending.action_id == action_id).then_some(*request_id)
+            })
+            .collect::<Vec<_>>();
+        let detached_count = request_ids.len();
+        for request_id in request_ids {
+            self.detach_dispatch(request_id);
+        }
+        detached_count
+    }
+
+    /// Cancels only the caller's pending tool wait. A child that was already created keeps
+    /// running independently and remains available in conversation history.
+    pub(super) fn cancel_execution(&mut self, action_id: &AIAgentActionId) {
+        self.cancel_dispatches_for_action(action_id);
     }
 
     pub(super) fn preprocess_action(

app/src/ai/blocklist/action_model/execute/start_agent_tests.rs

@@ -257,6 +257,71 @@ fn dispatch_denies_child_conversation_defense_in_depth() {
     });
 }
 
+#[test]
+fn local_execution_waits_for_completion() {
+    assert_eq!(
+        wait_policy_for_execution_mode(&StartAgentExecutionMode::local_with_defaults()),
+        StartAgentWaitPolicy::Completion
+    );
+}
+
+#[test]
+fn detach_dispatch_rejects_late_child_callback() {
+    App::test((), |mut app| async move {
+        initialize_history_persistence_for_tests(&mut app);
+        let terminal_view_id = EntityId::new();
+        let history_model = app.add_singleton_model(|_| BlocklistAIHistoryModel::new_for_test());
+        let executor = app.add_model(StartAgentExecutor::new);
+        let parent_conversation_id = history_model.update(&mut app, |history, ctx| {
+            history.start_new_conversation(terminal_view_id, false, false, false, ctx)
+        });
+        let dispatch = executor.update(&mut app, |executor, ctx| {
+            executor.dispatch(
+                AIAgentActionId::from("run-agents".to_string()),
+                "child".to_string(),
+                "work".to_string(),
+                StartAgentExecutionMode::Remote {
+                    environment_id: "environment".to_string(),
+                    skill_references: Vec::new(),
+                    model_id: "model".to_string(),
+                    computer_use_enabled: false,
+                    worker_host: String::new(),
+                    harness_type: "oz".to_string(),
+                    title: String::new(),
+                    auth_secret_name: None,
+                },
+                None,
+                parent_conversation_id,
+                Some(PARENT_RUN_ID.to_string()),
+                ctx,
+            )
+        });
+        assert!(executor.update(&mut app, |executor, _| {
+            executor.detach_dispatch(dispatch.request_id)
+        }));
+
+        let child_conversation_id = history_model.update(&mut app, |history, ctx| {
+            history.start_new_child_conversation(
+                terminal_view_id,
+                "child".to_string(),
+                parent_conversation_id,
+                None,
+                ctx,
+            )
+        });
+        history_model.update(&mut app, |history, ctx| {
+            history.record_new_conversation_request_complete(
+                dispatch.request_id,
+                child_conversation_id,
+                ctx,
+            );
+        });
+
+        executor.read(&app, |executor, _| assert!(executor.pending.is_empty()));
+        assert!(dispatch.receiver.try_recv().is_err());
+    });
+}
+
 #[test]
 fn legacy_local_codex_command_prompt_normalizes_to_local_harness() {
     let (prompt, execution_mode) = normalize_legacy_local_child_harness_command(
@@ -878,6 +943,69 @@ fn direct_provider_error_preserves_child_for_inspection() {
     });
 }
 
+#[test]
+fn cancelling_standalone_start_agent_drops_dispatch_but_keeps_child_running() {
+    App::test((), |mut app| async move {
+        initialize_history_persistence_for_tests(&mut app);
+        let terminal_view_id = EntityId::new();
+        let history_model = app.add_singleton_model(|_| BlocklistAIHistoryModel::new_for_test());
+        let executor = app.add_model(StartAgentExecutor::new);
+        let parent_conversation_id = history_model.update(&mut app, |history, ctx| {
+            history.start_new_conversation(terminal_view_id, false, false, false, ctx)
+        });
+        let action = build_start_agent_action(
+            StartAgentVersion::V1,
+            StartAgentExecutionMode::local_with_defaults(),
+        );
+        let execution = executor.update(&mut app, |executor, ctx| {
+            executor
+                .execute(
+                    ExecuteActionInput {
+                        action: &action,
+                        conversation_id: parent_conversation_id,
+                    },
+                    ctx,
+                )
+                .into()
+        });
+        let child_conversation_id = history_model.update(&mut app, |history, ctx| {
+            history.start_new_child_conversation(
+                terminal_view_id,
+                "child".to_string(),
+                parent_conversation_id,
+                None,
+                ctx,
+            )
+        });
+        history_model.update(&mut app, |history, ctx| {
+            history.record_new_conversation_request_complete(
+                FIRST_REQUEST_ID,
+                child_conversation_id,
+                ctx,
+            );
+        });
+
+        executor.update(&mut app, |executor, _| {
+            executor.cancel_execution(&action.id);
+        });
+
+        executor.read(&app, |executor, _| assert!(executor.pending.is_empty()));
+        history_model.read(&app, |history, _| {
+            assert_eq!(
+                history
+                    .conversation(&child_conversation_id)
+                    .expect("child should remain in history")
+                    .status(),
+                &ConversationStatus::InProgress
+            );
+        });
+        let AnyActionExecution::Async { execute_future, .. } = execution else {
+            panic!("expected async StartAgent execution");
+        };
+        let _ = execute_future.await;
+    });
+}
+
 #[test]
 fn removing_direct_provider_child_resolves_pending_wait() {
     App::test((), |mut app| async move {
@@ -975,7 +1103,7 @@ fn reattach_reuses_persisted_child_without_launching_another_agent() {
                 "child".to_string(),
                 parent_conversation_id,
                 child_conversation_id,
-                None,
+                StartAgentWaitPolicy::Completion,
                 ctx,
             )
         });

app/src/ai/blocklist/action_model_tests.rs

@@ -346,3 +346,128 @@ fn only_rejecting_a_blocked_action_is_a_permission_denial() {
         Some(&AIActionStatus::Blocked),
     ));
 }
+
+#[test]
+fn duplicate_action_ids_resolve_only_within_the_requested_conversation() {
+    let first_conversation = AIConversationId::new();
+    let second_conversation = AIConversationId::new();
+    let duplicate_id = AIAgentActionId::from("duplicate".to_string());
+    let first_result = make_action_result("duplicate");
+    let mut second_result = action_result("duplicate", AIAgentActionResultType::InitProject);
+    second_result.task_id = TaskId::new("second-task".to_string());
+    let second_result = Arc::new(second_result);
+    let finished_results = HashMap::from([(first_conversation, vec![first_result.clone()])]);
+    let provider_results = HashMap::new();
+    let archive = HashMap::from([
+        (
+            (first_conversation, duplicate_id.clone()),
+            first_result.clone(),
+        ),
+        (
+            (second_conversation, duplicate_id.clone()),
+            second_result.clone(),
+        ),
+    ]);
+
+    assert!(Arc::ptr_eq(
+        action_result_for_conversation(
+            &finished_results,
+            &provider_results,
+            &archive,
+            first_conversation,
+            &duplicate_id,
+        )
+        .unwrap(),
+        &first_result,
+    ));
+    assert!(Arc::ptr_eq(
+        action_result_for_conversation(
+            &finished_results,
+            &provider_results,
+            &archive,
+            second_conversation,
+            &duplicate_id,
+        )
+        .unwrap(),
+        &second_result,
+    ));
+    assert!(action_result_for_conversation(
+        &finished_results,
+        &provider_results,
+        &archive,
+        AIConversationId::new(),
+        &duplicate_id,
+    )
+    .is_none());
+}
+
+#[test]
+fn cancellation_permission_inference_uses_the_matching_conversation_status() {
+    let blocked_conversation = AIConversationId::new();
+    let queued_conversation = AIConversationId::new();
+    let duplicate_id = AIAgentActionId::from("duplicate".to_string());
+    let pending_actions = HashMap::from([
+        (blocked_conversation, VecDeque::from([action("duplicate")])),
+        (
+            queued_conversation,
+            VecDeque::from([action("first"), action("duplicate")]),
+        ),
+    ]);
+    let running_actions = HashMap::new();
+
+    let blocked_status = pending_action_status(
+        &pending_actions,
+        &running_actions,
+        blocked_conversation,
+        &duplicate_id,
+        false,
+    );
+    let queued_status = pending_action_status(
+        &pending_actions,
+        &running_actions,
+        queued_conversation,
+        &duplicate_id,
+        false,
+    );
+
+    assert!(is_permission_denial(
+        CancellationReason::ManuallyCancelled,
+        blocked_status.as_ref(),
+    ));
+    assert!(!is_permission_denial(
+        CancellationReason::ManuallyCancelled,
+        queued_status.as_ref(),
+    ));
+}
+
+#[test]
+fn action_lifecycle_events_disambiguate_duplicate_ids_by_conversation() {
+    let first_conversation = AIConversationId::new();
+    let second_conversation = AIConversationId::new();
+    let duplicate_id = AIAgentActionId::from("duplicate".to_owned());
+    let events = [
+        BlocklistAIActionEvent::ActionBlockedOnUserConfirmation {
+            action_id: duplicate_id.clone(),
+            conversation_id: first_conversation,
+            execution_ref: None,
+        },
+        BlocklistAIActionEvent::ExecutingAction {
+            action_id: duplicate_id.clone(),
+            conversation_id: second_conversation,
+            execution_ref: None,
+        },
+        BlocklistAIActionEvent::FinishedAction {
+            action_id: duplicate_id.clone(),
+            conversation_id: first_conversation,
+            cancellation_reason: None,
+            execution_ref: None,
+        },
+    ];
+
+    assert_eq!(events[0].conversation_id(), Some(first_conversation));
+    assert_eq!(events[1].conversation_id(), Some(second_conversation));
+    assert_eq!(events[2].conversation_id(), Some(first_conversation));
+    assert!(events
+        .iter()
+        .all(|event| event.action_id() == &duplicate_id));
+}

app/src/ai/blocklist/block.rs

@@ -3578,7 +3578,10 @@ impl AIBlock {
                 }
 
                 // Set the state based on the action status from the action model
-                let action_status = self.action_model.as_ref(ctx).get_action_status(action_id);
+                let action_status = self
+                    .action_model
+                    .as_ref(ctx)
+                    .get_action_status(self.client_ids.conversation_id, action_id);
 
                 let is_reverted = BlocklistAIHistoryModel::as_ref(ctx)
                     .conversation(&self.client_ids.conversation_id)
@@ -3673,6 +3676,7 @@ impl AIBlock {
             RequestedCommandViewEvent::Accepted => {
                 self.action_model.update(ctx, |action_model, ctx| {
                     action_model.handle_requested_command_accepted(
+                        self.client_ids.conversation_id,
                         action_id,
                         view.as_ref(ctx).command_text().to_string(),
                         ctx,
@@ -3691,7 +3695,10 @@ impl AIBlock {
             RequestedCommandViewEvent::UpdatedExpansionState { is_expanded } => {
                 // We only care about expansion state updates when the command
                 // is running or finished (i.e. when it has a block).
-                let action_status = self.action_model.as_ref(ctx).get_action_status(action_id);
+                let action_status = self
+                    .action_model
+                    .as_ref(ctx)
+                    .get_action_status(self.client_ids.conversation_id, action_id);
                 let has_finished_command_block = {
                     let terminal_model = self.terminal_model.lock();
                     terminal_model
@@ -3890,7 +3897,7 @@ impl AIBlock {
         if self
             .action_model
             .as_ref(ctx)
-            .get_action_status(action_id)
+            .get_action_status(self.client_ids.conversation_id, action_id)
             .is_some_and(|status| status.is_blocked())
         {
             ctx.focus(&view);
@@ -4274,7 +4281,10 @@ impl AIBlock {
         // but it's not incorrect to populate if it is, and we rely on this for
         // for restored conversations because action model events don't re-fire
         // after the view is created.
-        let action_status = self.action_model.as_ref(ctx).get_action_status(action_id);
+        let action_status = self
+            .action_model
+            .as_ref(ctx)
+            .get_action_status(self.client_ids.conversation_id, action_id);
         if let Some(view) = self.search_codebase_view.get(action_id) {
             let files = if let Some(AIActionStatus::Finished(ref result)) = action_status {
                 if let AIAgentActionResultType::SearchCodebase(SearchCodebaseResult::Success {
@@ -4708,7 +4718,11 @@ impl AIBlock {
     pub fn is_blocked_on_user_confirmation(&self, app: &AppContext) -> bool {
         self.requested_action_ids
             .iter()
-            .filter_map(|id| self.action_model.as_ref(app).get_action_status(id))
+            .filter_map(|id| {
+                self.action_model
+                    .as_ref(app)
+                    .get_action_status(self.client_ids.conversation_id, id)
+            })
             .any(|status| status.is_blocked())
     }
 
@@ -4734,7 +4748,12 @@ impl AIBlock {
         ctx.subscribe_to_model(action_model, |me, action_model, event, ctx| {
             let action_id = event.action_id();
 
-            if me.is_finished() || !me.requested_action_ids.contains(action_id) {
+            if event
+                .conversation_id()
+                .is_some_and(|conversation_id| conversation_id != me.client_ids.conversation_id)
+                || me.is_finished()
+                || !me.requested_action_ids.contains(action_id)
+            {
                 // Technically, this subscription should be unregistered after `is_finished` is
                 // set to true, but it seems that the callback is called once more after the `unsubscribe_to_model`
                 // call, so early return here if this is errantly being called.
@@ -4828,7 +4847,7 @@ impl AIBlock {
                     {
                         let should_collapse = action_model
                             .as_ref(ctx)
-                            .get_action_result(action_id)
+                            .get_action_result(me.client_ids.conversation_id, action_id)
                             .is_none_or(|result| match &result.result {
                                 AIAgentActionResultType::RequestCommandOutput(
                                     RequestCommandOutputResult::Completed { exit_code, .. },
@@ -4843,7 +4862,9 @@ impl AIBlock {
                     }
 
                     if let Some(view) = me.search_codebase_view.get(action_id) {
-                        let new_status = action_model.as_ref(ctx).get_action_status(action_id);
+                        let new_status = action_model
+                            .as_ref(ctx)
+                            .get_action_status(me.client_ids.conversation_id, action_id);
                         view.update(ctx, |view, ctx| {
                             view.update_status(new_status);
                             ctx.notify();
@@ -4852,7 +4873,9 @@ impl AIBlock {
 
                     // Create subagent panel state for finished StartAgent actions
                     if let Some(AIActionStatus::Finished(result)) =
-                        action_model.as_ref(ctx).get_action_status(action_id)
+                        action_model
+                            .as_ref(ctx)
+                            .get_action_status(me.client_ids.conversation_id, action_id)
                     {
                         if let AIAgentActionResultType::StartAgent(
                             crate::ai::agent::StartAgentResult::Success { agent_id, .. },
@@ -4874,7 +4897,11 @@ impl AIBlock {
                     let action_statuses = me
                         .requested_action_ids
                         .iter()
-                        .filter_map(|id| action_model.as_ref(ctx).get_action_status(id))
+                        .filter_map(|id| {
+                            action_model
+                                .as_ref(ctx)
+                                .get_action_status(me.client_ids.conversation_id, id)
+                        })
                         .collect_vec();
 
                     // Detecting links on SearchCodebase tool call outputs
@@ -4907,7 +4934,9 @@ impl AIBlock {
                                     view.update_render_read_file_args(
                                         &me.find_state,
                                         files.clone(),
-                                        action_model.as_ref(ctx).get_action_status(action_id),
+                                        action_model
+                                            .as_ref(ctx)
+                                            .get_action_status(me.client_ids.conversation_id, action_id),
                                     );
                                     ctx.notify();
                                 })
@@ -4917,7 +4946,9 @@ impl AIBlock {
 
                     // Open the AI document pane when documents are created or edited
                     if let Some(action_result) =
-                        action_model.as_ref(ctx).get_action_result(action_id)
+                        action_model
+                            .as_ref(ctx)
+                            .get_action_result(me.client_ids.conversation_id, action_id)
                     {
                         match &action_result.result {
                             AIAgentActionResultType::CreateDocuments(
@@ -5677,7 +5708,9 @@ impl AIBlock {
     /// This hides their keybindings in the UI and makes them less interactive.
     pub fn ignore_passive_actions(&mut self, ctx: &mut ViewContext<Self>) {
         self.action_model.update(ctx, |action_model, ctx| {
-            for action in action_model.get_pending_actions() {
+            for action in
+                action_model.get_pending_actions_for_conversation(&self.client_ids.conversation_id)
+            {
                 if let Some(edit) = self.requested_edits.get(&action.id) {
                     edit.view.update(ctx, |view, ctx| view.dismiss(ctx));
                 } else if let Some(suggested_prompt) = self.unit_tests_suggestions.get(&action.id) {
@@ -5730,7 +5763,12 @@ impl AIBlock {
                     .view
                     .update(ctx, |view, ctx| view.commit_and_get_command_text(ctx));
                 self.action_model.update(ctx, |action_model, ctx| {
-                    action_model.handle_requested_command_accepted(&action_id, command_text, ctx);
+                    action_model.handle_requested_command_accepted(
+                        self.client_ids.conversation_id,
+                        &action_id,
+                        command_text,
+                        ctx,
+                    );
                 });
                 ctx.notify();
             }
@@ -5758,12 +5796,11 @@ impl AIBlock {
     /// Finds the undismissed passive code diff across all pending actions.
     /// This is needed because passive code diffs are NOT added to the active conversation by default, when they first appear.
     pub(crate) fn find_undismissed_code_diff(&self, app: &AppContext) -> Option<&RequestedEdit> {
-        let all_pending_actions = self.action_model.as_ref(app).get_pending_actions();
-
         // Find any RequestFileEdits action that has a corresponding passive code diff view.
         // Note that we only expect a maximum of 1 passive code diff to be undismissed at any given time.
-        all_pending_actions
+        self.action_model
-            .iter()
+            .as_ref(app)
+            .get_pending_actions_for_conversation(&self.client_ids.conversation_id)
             .find_map(|action| match &action.action {
                 AIAgentActionType::RequestFileEdits {
                     file_edits: _,
@@ -5803,7 +5840,10 @@ impl AIBlock {
             .is_none_or(|output| {
                 output.get().actions().last().is_none_or(|action| {
                     let is_streaming = self.model.status(app).is_streaming();
-                    let status = self.action_model.as_ref(app).get_action_status(&action.id);
+                    let status = self
+                        .action_model
+                        .as_ref(app)
+                        .get_action_status(self.client_ids.conversation_id, &action.id);
                     is_streaming || status.is_some_and(|status| status.is_running())
                 })
             })
@@ -5830,7 +5870,7 @@ impl AIBlock {
             .any(|(action_id, requested_command)| {
                 self.action_model
                     .as_ref(app)
-                    .get_action_status(action_id)
+                    .get_action_status(self.client_ids.conversation_id, action_id)
                     .is_some_and(|status| status.is_running())
                     && requested_command.view.as_ref(app).is_header_expanded()
             })
@@ -5930,7 +5970,10 @@ impl AIBlock {
             return String::new();
         };
         let output = output.get();
-        output.format_for_copy(Some(self.action_model.as_ref(app)))
+        output.format_for_copy_for_conversation(
+            Some(self.action_model.as_ref(app)),
+            Some(self.client_ids.conversation_id),
+        )
     }
 
     /// Gets AI output text for copying from the preceding user query until the next user query
@@ -5985,8 +6028,10 @@ impl AIBlock {
         // Collect all AI outputs from start_idx to end_idx (exclusive)
         let mut combined_result = Vec::new();
         for exchange in exchanges.iter().take(end_idx).skip(start_idx) {
-            let formatted_output =
+            let formatted_output = exchange.format_output_for_copy_for_conversation(
-                exchange.format_output_for_copy(Some(self.action_model.as_ref(app)));
+                Some(self.action_model.as_ref(app)),
+                Some(self.client_ids.conversation_id),
+            );
             if !formatted_output.is_empty() {
                 combined_result.push(formatted_output);
             }
@@ -7158,7 +7203,7 @@ impl TypedActionView for AIBlock {
                     let Some(result) = self
                         .action_model
                         .as_ref(ctx)
-                        .get_action_result(action_id)
+                        .get_action_result(self.client_ids.conversation_id, action_id)
                         .map(Arc::clone)
                     else {
                         continue;

app/src/ai/blocklist/block/cli.rs

@@ -1170,7 +1170,7 @@ impl View for CLISubagentView {
                         let is_cancelled = self
                             .action_model
                             .as_ref(app)
-                            .get_action_status(&action.id)
+                            .get_action_status(self.conversation_id, &action.id)
                             .is_some_and(|status| status.is_cancelled());
                         if blocked_action.is_none() && !is_cancelled && !should_hide_responses {
                             if let Some(rendered_action) = render_action(action.action.clone(), app)

app/src/ai/blocklist/block/cli_controller.rs

@@ -41,6 +41,7 @@ pub enum UserTakeOverReason {
 
 #[derive(Debug, Clone, Default)]
 struct ActiveCLISubagentState {
+    initial_requested_command_conversation_id: Option<AIConversationId>,
     initial_requested_command_action_id: Option<AIAgentActionId>,
     task_id: Option<TaskId>,
     last_snapshot_at: Option<Instant>,
@@ -171,9 +172,21 @@ impl CLISubagentController {
         });
 
         ctx.subscribe_to_model(action_model, |me, _, event, ctx| match event {
-            BlocklistAIActionEvent::ActionBlockedOnUserConfirmation { .. } => {
+            BlocklistAIActionEvent::ActionBlockedOnUserConfirmation {
+                action_id,
+                conversation_id,
+                ..
+            } => {
                 let mut terminal_model = me.terminal_model.lock();
                 let active_block = terminal_model.block_list_mut().active_block_mut();
+                if !matches_active_requested_command(
+                    *conversation_id,
+                    action_id,
+                    active_block.ai_conversation_id(),
+                    active_block.requested_command_action_id(),
+                ) {
+                    return;
+                }
                 active_block.update_is_agent_blocked(true);
 
                 let action_id = active_block.requested_command_action_id().cloned();
@@ -183,9 +196,21 @@ impl CLISubagentController {
                     agent_has_control: active_block.is_agent_in_control(),
                 });
             }
-            BlocklistAIActionEvent::ExecutingAction { .. } => {
+            BlocklistAIActionEvent::ExecutingAction {
+                action_id,
+                conversation_id,
+                ..
+            } => {
                 let mut terminal_model = me.terminal_model.lock();
                 let active_block = terminal_model.block_list_mut().active_block_mut();
+                if !matches_active_requested_command(
+                    *conversation_id,
+                    action_id,
+                    active_block.ai_conversation_id(),
+                    active_block.requested_command_action_id(),
+                ) {
+                    return;
+                }
                 active_block.update_is_agent_blocked(false);
 
                 let action_id = active_block.requested_command_action_id().cloned();
@@ -197,12 +222,13 @@ impl CLISubagentController {
             }
             BlocklistAIActionEvent::FinishedAction {
                 action_id: finished_action_id,
+                conversation_id,
                 ..
             } => {
                 let action_result = me
                     .action_model
                     .as_ref(ctx)
-                    .get_action_result(finished_action_id);
+                    .get_action_result(*conversation_id, finished_action_id);
                 let initial_command_finished_without_snapshot =
                     action_result.is_some_and(|result| {
                         matches!(
@@ -222,14 +248,22 @@ impl CLISubagentController {
                     .cloned();
                 let mut terminal_model = me.terminal_model.lock();
                 let active_block = terminal_model.block_list_mut().active_block_mut();
-                active_block.update_is_agent_blocked(false);
+                if matches_active_requested_command(
+                    *conversation_id,
+                    finished_action_id,
+                    active_block.ai_conversation_id(),
+                    active_block.requested_command_action_id(),
+                ) {
+                    active_block.update_is_agent_blocked(false);
 
-                let active_command_action_id = active_block.requested_command_action_id().cloned();
+                    let active_command_action_id =
-                ctx.emit(CLISubagentEvent::UpdatedControl {
+                        active_block.requested_command_action_id().cloned();
-                    block_id: active_block.id().clone(),
+                    ctx.emit(CLISubagentEvent::UpdatedControl {
-                    requested_command_action_id: active_command_action_id,
+                        block_id: active_block.id().clone(),
-                    agent_has_control: active_block.is_agent_in_control(),
+                        requested_command_action_id: active_command_action_id,
-                });
+                        agent_has_control: active_block.is_agent_in_control(),
+                    });
+                }
 
                 // Updates the last snapshot timestamp for the active block after the agent has read the block output.
                 if let Some(snapshot_block_id) = snapshot_block_id {
@@ -244,18 +278,17 @@ impl CLISubagentController {
                 if initial_command_finished_without_snapshot {
                     me.active_subagents_by_block.retain(|_, state| {
                         state.task_id.is_some()
-                            || state.initial_requested_command_action_id.as_ref()
+                            || !matches_requested_command_identity(
-                                != Some(finished_action_id)
+                                *conversation_id,
+                                finished_action_id,
+                                state.initial_requested_command_conversation_id,
+                                state.initial_requested_command_action_id.as_ref(),
+                            )
                     });
                 }
+                drop(terminal_model);
                 if let Some(block_id) = command_finished_block_id {
-                    if let Some(completion) = me
+                    me.advance_completed_subagent(&block_id, ctx);
-                        .active_subagents_by_block
-                        .get_mut(&block_id)
-                        .and_then(|state| state.completion.as_mut())
-                    {
-                        completion.final_turn_started = true;
-                    }
                 }
             }
             _ => (),
@@ -322,7 +355,7 @@ impl CLISubagentController {
                 };
                 drop(terminal_model);
 
-                let provider_consumed_completion = completion.as_ref().is_some_and(|completion| {
+                let provider_accepted_completion = completion.as_ref().is_some_and(|completion| {
                     me.controller.update(ctx, |controller, ctx| {
                         controller.accept_provider_command_completion(
                             completion.conversation_id,
@@ -344,13 +377,9 @@ impl CLISubagentController {
                 if has_last_snapshot {
                     ctx.emit(CLISubagentEvent::UpdatedLastSnapshot);
                 }
-                if provider_consumed_completion {
+                if provider_accepted_completion {
-                    me.finish_subagent(
+                    // The provider controller owns deactivation after it applies the queued
-                        &block_id,
+                    // completion at a safe run boundary.
-                        conversation_id,
-                        requested_command_action_id,
-                        ctx,
-                    );
                     return;
                 }
                 if !me.active_subagents_by_block.contains_key(&block_id) {
@@ -628,11 +657,18 @@ impl CLISubagentController {
     ///
     /// The placeholder lets command completion and action-result events arrive in either order
     /// without losing the completion that a subsequently-created CLI monitor needs.
-    pub fn track_requested_command(&mut self, block_id: &BlockId, action_id: &AIAgentActionId) {
+    pub fn track_requested_command(
-        self.active_subagents_by_block
+        &mut self,
+        block_id: &BlockId,
+        conversation_id: AIConversationId,
+        action_id: &AIAgentActionId,
+    ) {
+        let state = self
+            .active_subagents_by_block
             .entry(block_id.clone())
-            .or_default()
+            .or_default();
-            .initial_requested_command_action_id = Some(action_id.clone());
+        state.initial_requested_command_conversation_id = Some(conversation_id);
+        state.initial_requested_command_action_id = Some(action_id.clone());
     }
 
     /// Force the currently in-flight poll for the given long-running command block to
@@ -879,6 +915,10 @@ impl CLISubagentController {
             requested_command_action_id: action_id.clone(),
             agent_has_control,
         });
+        self.active_subagents_by_block
+            .entry(block_id.clone())
+            .or_default()
+            .initial_requested_command_conversation_id = Some(conversation_id);
         self.active_subagents_by_block
             .entry(block_id.clone())
             .or_default()
@@ -1027,6 +1067,7 @@ fn command_finished_block_id(result: &AIAgentActionResultType) -> Option<&BlockI
         AIAgentActionResultType::RequestCommandOutput(
             RequestCommandOutputResult::LongRunningCommandSnapshot { .. }
             | RequestCommandOutputResult::CancelledBeforeExecution
+            | RequestCommandOutputResult::ExecutionError { .. }
             | RequestCommandOutputResult::Denylisted { .. },
         )
         | AIAgentActionResultType::WriteToLongRunningShellCommand(
@@ -1085,6 +1126,26 @@ fn should_nudge_monitor_turn(last_exchange_has_action: bool, monitor_nudge_sent:
     !last_exchange_has_action && !monitor_nudge_sent
 }
 
+fn matches_active_requested_command(
+    event_conversation_id: AIConversationId,
+    event_action_id: &AIAgentActionId,
+    active_conversation_id: Option<AIConversationId>,
+    active_requested_command_id: Option<&AIAgentActionId>,
+) -> bool {
+    active_conversation_id == Some(event_conversation_id)
+        && active_requested_command_id == Some(event_action_id)
+}
+
+fn matches_requested_command_identity(
+    event_conversation_id: AIConversationId,
+    event_action_id: &AIAgentActionId,
+    requested_command_conversation_id: Option<AIConversationId>,
+    requested_command_action_id: Option<&AIAgentActionId>,
+) -> bool {
+    requested_command_conversation_id == Some(event_conversation_id)
+        && requested_command_action_id == Some(event_action_id)
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -1121,4 +1182,51 @@ mod tests {
         assert!(!should_nudge_monitor_turn(false, true));
         assert!(!should_nudge_monitor_turn(true, false));
     }
+
+    #[test]
+    fn shell_control_event_must_match_conversation_and_requested_command() {
+        let active_conversation_id = AIConversationId::new();
+        let other_conversation_id = AIConversationId::new();
+        let active_action_id = AIAgentActionId::from("same-action".to_owned());
+        let other_action_id = AIAgentActionId::from("other-action".to_owned());
+
+        assert!(matches_active_requested_command(
+            active_conversation_id,
+            &active_action_id,
+            Some(active_conversation_id),
+            Some(&active_action_id),
+        ));
+        assert!(!matches_active_requested_command(
+            other_conversation_id,
+            &active_action_id,
+            Some(active_conversation_id),
+            Some(&active_action_id),
+        ));
+        assert!(!matches_active_requested_command(
+            active_conversation_id,
+            &other_action_id,
+            Some(active_conversation_id),
+            Some(&active_action_id),
+        ));
+    }
+
+    #[test]
+    fn requested_command_identity_rejects_duplicate_id_from_another_conversation() {
+        let active_conversation_id = AIConversationId::new();
+        let other_conversation_id = AIConversationId::new();
+        let duplicate_action_id = AIAgentActionId::from("duplicate-action".to_owned());
+
+        assert!(matches_requested_command_identity(
+            active_conversation_id,
+            &duplicate_action_id,
+            Some(active_conversation_id),
+            Some(&duplicate_action_id),
+        ));
+        assert!(!matches_requested_command_identity(
+            other_conversation_id,
+            &duplicate_action_id,
+            Some(active_conversation_id),
+            Some(&duplicate_action_id),
+        ));
+    }
 }

app/src/ai/blocklist/block/model/helper.rs

@@ -149,7 +149,11 @@ impl<T: ?Sized + AIBlockModel> AIBlockModelHelper for T {
         let output = output.get();
         output.messages.iter().find_map(|message| {
             if let AIAgentOutputMessageType::Action(action) = &message.message {
-                if let Some(status) = action_model.as_ref(app).get_action_status(&action.id) {
+                if let Some(status) = self.conversation_id(app).and_then(|conversation_id| {
+                    action_model
+                        .as_ref(app)
+                        .get_action_status(conversation_id, &action.id)
+                }) {
                     return status.is_blocked().then_some(action.clone());
                 }
             }

app/src/ai/blocklist/block/status_bar.rs

@@ -328,10 +328,27 @@ impl BlocklistAIStatusBar {
             },
         );
 
-        ctx.subscribe_to_model(&action_model, |_, _, event, ctx| match event {
+        ctx.subscribe_to_model(&action_model, |me, _, event, ctx| match event {
-            BlocklistAIActionEvent::ExecutingAction { .. }
+            BlocklistAIActionEvent::ExecutingAction {
-            | BlocklistAIActionEvent::FinishedAction { .. } => ctx.notify(),
+                conversation_id, ..
-            _ => (),
+            }
+            | BlocklistAIActionEvent::FinishedAction {
+                conversation_id, ..
+            } if me
+                .active_exchange_model
+                .as_ref()
+                .is_some_and(|model| model.conversation_id(ctx) == Some(*conversation_id)) =>
+            {
+                ctx.notify();
+            }
+            BlocklistAIActionEvent::QueuedAction { .. }
+            | BlocklistAIActionEvent::ActionBlockedOnUserConfirmation { .. }
+            | BlocklistAIActionEvent::ExecutingAction { .. }
+            | BlocklistAIActionEvent::FinishedAction { .. }
+            | BlocklistAIActionEvent::ToolLifecycle { .. }
+            | BlocklistAIActionEvent::InitProject(_)
+            | BlocklistAIActionEvent::ToggleCodeReview(_)
+            | BlocklistAIActionEvent::InsertCodeReviewComments { .. } => {}
         });
         ctx.subscribe_to_model(model_event_dispatcher, |me, _, event, ctx| match event {
             ModelEvent::AfterBlockStarted { block_id, .. } => {

app/src/ai/blocklist/block/view_impl.rs

@@ -1079,6 +1079,7 @@ impl View for AIBlock {
 
         contents.add_child(output::render(
             output::Props {
+                conversation_id: self.client_ids.conversation_id,
                 model: self.model.as_ref(),
                 state_handles: &self.state_handles,
                 action_buttons: &self.action_buttons,

app/src/ai/blocklist/block/view_impl/orchestration.rs

@@ -420,7 +420,10 @@ pub(super) fn render_send_message(
 ) -> Box<dyn Element> {
     let appearance = Appearance::as_ref(app);
     let theme = appearance.theme();
-    let status = props.action_model.as_ref(app).get_action_status(action_id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, action_id);
     let orchestrator_agent_id = props
         .model
         .conversation(app)
@@ -564,7 +567,10 @@ pub(super) fn render_start_agent(
 ) -> Box<dyn Element> {
     let appearance = Appearance::as_ref(app);
     let theme = appearance.theme();
-    let status = props.action_model.as_ref(app).get_action_status(action_id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, action_id);
 
     if let Some(AIActionStatus::Finished(result)) = &status {
         let AIAgentActionResultType::StartAgent(result) = &result.result else {

app/src/ai/blocklist/block/view_impl/output.rs

@@ -56,6 +56,7 @@ use super::{
 };
 use crate::ai::agent::api::ServerConversationToken;
 use crate::ai::agent::comment::ReviewComment;
+use crate::ai::agent::conversation::AIConversationId;
 use crate::ai::agent::icons::{self, gray_stop_icon, yellow_stop_icon};
 use crate::ai::agent::task::TaskId;
 use crate::ai::agent::{
@@ -143,6 +144,7 @@ fn should_render_requested_edit(action_status: Option<&AIActionStatus>) -> bool
 /// Data required to render the AI block output component.
 #[derive(Copy, Clone)]
 pub(crate) struct Props<'a> {
+    pub(crate) conversation_id: AIConversationId,
     pub(crate) model: &'a dyn AIBlockModel<View = AIBlock>,
     pub(super) state_handles: &'a AIBlockStateHandles,
     pub(super) action_buttons: &'a HashMap<AIAgentActionId, ActionButtons>,
@@ -436,7 +438,7 @@ pub(super) fn render(props: Props, app: &AppContext) -> Box<dyn Element> {
                             let is_action_done = props
                                 .action_model
                                 .as_ref(app)
-                                .get_action_status(id)
+                                .get_action_status(props.conversation_id, id)
                                 .as_ref()
                                 .is_some_and(|status| status.is_done());
                             if !is_action_done {
@@ -476,7 +478,7 @@ pub(super) fn render(props: Props, app: &AppContext) -> Box<dyn Element> {
                                 let agent_action_results = props
                                     .action_model
                                     .as_ref(app)
-                                    .get_action_result(id)
+                                    .get_action_result(props.conversation_id, id)
                                     .map(|action_result| action_result.as_ref());
 
                                 // checks if the read file action result is completed and successful.
@@ -565,8 +567,10 @@ pub(super) fn render(props: Props, app: &AppContext) -> Box<dyn Element> {
                             id,
                             ..
                         }) => {
-                            let action_status =
+                            let action_status = props
-                                props.action_model.as_ref(app).get_action_status(id);
+                                .action_model
+                                .as_ref(app)
+                                .get_action_status(props.conversation_id, id);
 
                             if should_render_requested_edit(action_status.as_ref()) {
                                 if let Some(requested_edit) = props.requested_edits.get(id) {
@@ -656,7 +660,7 @@ pub(super) fn render(props: Props, app: &AppContext) -> Box<dyn Element> {
                             let is_action_done = props
                                 .action_model
                                 .as_ref(app)
-                                .get_action_status(id)
+                                .get_action_status(props.conversation_id, id)
                                 .as_ref()
                                 .is_some_and(|status| status.is_done());
                             if !is_action_done {
@@ -1374,7 +1378,13 @@ fn render_runtime_activity(
         }
     }
 
-    Some(render_tool_pane_shell(content.finish(), false, false, app))
+    Some(render_tool_pane_shell(
+        content.finish(),
+        false,
+        is_expanded,
+        false,
+        app,
+    ))
 }
 
 fn should_render_stopped_output(props: Props, app: &AppContext) -> bool {
@@ -1473,7 +1483,10 @@ fn render_search_codebase(
     id: &AIAgentActionId,
     app: &AppContext,
 ) -> Option<Box<dyn Element>> {
-    let status = props.action_model.as_ref(app).get_action_status(id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, id);
     let appearance = Appearance::as_ref(app);
     let theme = appearance.theme();
 
@@ -1974,7 +1987,10 @@ fn render_read_files(
     parsed_skill: Option<&ai::skills::ParsedSkill>,
     action_index: usize,
 ) -> Box<dyn Element> {
-    let status = props.action_model.as_ref(app).get_action_status(id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, id);
     let appearance = Appearance::as_ref(app);
     let formatted_files =
         render_read_files_text(props.into(), file_names, app, appearance, action_index);
@@ -2091,7 +2107,10 @@ fn maybe_render_edit_document(
     id: &AIAgentActionId,
     app: &AppContext,
 ) -> Option<Box<dyn Element>> {
-    let status = props.action_model.as_ref(app).get_action_status(id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, id);
 
     // Document operations are always auto-executed for now
     if status.as_ref().is_some_and(|status| status.is_blocked()) {
@@ -2101,7 +2120,7 @@ fn maybe_render_edit_document(
     let agent_action_results = props
         .action_model
         .as_ref(app)
-        .get_action_result(id)
+        .get_action_result(props.conversation_id, id)
         .map(|action_result| action_result.as_ref());
 
     let Some(AIAgentActionResult {
@@ -2128,7 +2147,10 @@ fn maybe_render_create_document(
     id: &AIAgentActionId,
     app: &AppContext,
 ) -> Option<Box<dyn Element>> {
-    let status = props.action_model.as_ref(app).get_action_status(id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, id);
 
     // Document operations are always auto-executed for now
     if status.as_ref().is_some_and(|status| status.is_blocked()) {
@@ -2138,7 +2160,7 @@ fn maybe_render_create_document(
     let agent_action_results = props
         .action_model
         .as_ref(app)
-        .get_action_result(id)
+        .get_action_result(props.conversation_id, id)
         .map(|action_result| action_result.as_ref());
 
     let Some(AIAgentActionResult {
@@ -2441,7 +2463,7 @@ fn render_suggest_new_conversation(
     let status = props
         .action_model
         .as_ref(app)
-        .get_action_status(action_id)
+        .get_action_status(props.conversation_id, action_id)
         .unwrap_or(AIActionStatus::Finished(Arc::new(AIAgentActionResult {
             result: AIAgentActionResultType::SuggestNewConversation(
                 SuggestNewConversationResult::Cancelled,
@@ -2549,7 +2571,10 @@ fn create_formatted_text_for_grep(
     let appearance = Appearance::as_ref(app);
     let theme = appearance.theme();
 
-    let action_status = props.action_model.as_ref(app).get_action_status(id);
+    let action_status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, id);
     let is_cancelled = action_status
         .as_ref()
         .is_some_and(|status| status.is_cancelled());
@@ -2653,7 +2678,10 @@ fn create_formatted_text_for_file_glob(
     let appearance = Appearance::as_ref(app);
     let theme = appearance.theme();
 
-    let action_status = props.action_model.as_ref(app).get_action_status(id);
+    let action_status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, id);
     let is_cancelled = action_status
         .as_ref()
         .is_some_and(|status| status.is_cancelled());
@@ -2754,7 +2782,10 @@ fn render_file_retrieval_tool(
     app: &AppContext,
 ) -> Box<dyn Element> {
     let appearance = Appearance::as_ref(app);
-    let status = props.action_model.as_ref(app).get_action_status(action_id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, action_id);
 
     let mut config = RenderableAction::new_with_formatted_text(tool_formatted_text, app);
 
@@ -2871,7 +2902,10 @@ fn render_read_mcp_resource(
     app: &AppContext,
 ) -> Box<dyn Element> {
     let appearance = Appearance::as_ref(app);
-    let status = props.action_model.as_ref(app).get_action_status(action_id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, action_id);
 
     let mut renderable_action = RenderableAction::new(name, app);
 
@@ -2948,11 +2982,14 @@ fn render_upload_artifact(
     app: &AppContext,
 ) -> Box<dyn Element> {
     let appearance = Appearance::as_ref(app);
-    let status = props.action_model.as_ref(app).get_action_status(action_id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, action_id);
     let result = props
         .action_model
         .as_ref(app)
-        .get_action_result(action_id)
+        .get_action_result(props.conversation_id, action_id)
         .and_then(|result| match &result.result {
             AIAgentActionResultType::UploadArtifact(upload_result) => Some(upload_result),
             _ => None,
@@ -3011,7 +3048,7 @@ fn render_use_computer(
     let has_screenshot = props
         .action_model
         .as_ref(app)
-        .get_action_result(action_id)
+        .get_action_result(props.conversation_id, action_id)
         .is_some_and(|result| {
             matches!(
                 &result.result,
@@ -3057,7 +3094,10 @@ fn render_request_computer_use(
     app: &AppContext,
 ) -> Box<dyn Element> {
     let appearance = Appearance::as_ref(app);
-    let status = props.action_model.as_ref(app).get_action_status(action_id);
+    let status = props
+        .action_model
+        .as_ref(app)
+        .get_action_status(props.conversation_id, action_id);
 
     let mut renderable_action = RenderableAction::new(&request.task_summary, app);
 
@@ -3638,7 +3678,13 @@ pub fn action_icon<V: View>(
     app: &AppContext,
 ) -> galaxyui::elements::Icon {
     let appearance = Appearance::as_ref(app);
-    let status = action_model.as_ref(app).get_action_status(action_id);
+    let status = ai_block_model
+        .conversation_id(app)
+        .and_then(|conversation_id| {
+            action_model
+                .as_ref(app)
+                .get_action_status(conversation_id, action_id)
+        });
     match status {
         Some(status) => match status {
             AIActionStatus::Preprocessing => icons::gray_circle_icon(appearance),

app/src/ai/blocklist/controller.rs

@@ -10,7 +10,7 @@ mod pending_response_streams;
 pub mod response_stream;
 pub(super) mod shared_session;
 mod slash_command;
-use std::collections::{BTreeMap, HashMap, HashSet};
+use std::collections::{BTreeMap, HashMap, HashSet, VecDeque};
 #[cfg(not(target_family = "wasm"))]
 use std::path::PathBuf;
 use std::sync::Arc;
@@ -23,7 +23,8 @@ use futures::channel::oneshot;
 use galaxy_agent_core::{
     turn_control, ExternalWorkId, PendingToolBatch, PendingToolCallState, ProviderRun,
     ProviderRunFailureKind, ProviderRunId, ProviderRunLimits, ProviderRunOutcome, ProviderRunState,
-    ToolLoopGuard, TurnCommand, TurnCommandSender, TurnRequest,
+    StopReason, ToolLoopGuard, ToolResult, ToolResultStatus, TurnCommand, TurnCommandSender,
+    TurnRequest,
 };
 use galaxy_core::assertions::safe_assert;
 use input_context::{input_context_for_request, parse_context_attachments};
@@ -630,6 +631,32 @@ enum ProviderCommandResult {
     },
 }
 
+fn convert_provider_tool_batch(
+    action_context: &ProviderActionContext,
+    batch: &PendingToolBatch,
+) -> (Vec<(AIAgentAction, bool)>, Vec<ToolResult>) {
+    let mut actions = Vec::new();
+    let mut invalid_results = Vec::new();
+    for pending in batch
+        .calls
+        .iter()
+        .filter(|pending| pending.state.result().is_none())
+    {
+        match action_context.action_from_tool_call(&pending.call) {
+            Ok(action) => actions.push((
+                action,
+                matches!(pending.state, PendingToolCallState::RecoveryPending),
+            )),
+            Err(message) => invalid_results.push(ToolResult {
+                call_id: pending.call.id.clone(),
+                content: format!("Invalid {} tool input: {message}", pending.call.name),
+                status: ToolResultStatus::Error,
+            }),
+        }
+    }
+    (actions, invalid_results)
+}
+
 struct ActiveProviderRunSlot {
     stream_id: ResponseStreamId,
     response_stream: ModelHandle<ResponseStream>,
@@ -650,6 +677,13 @@ struct ActiveProviderRunSlot {
     monitor_prose_continuations: usize,
 }
 
+struct QueuedProviderRun {
+    slot: ActiveProviderRunSlot,
+    base_provider_config: crate::ai::provider::ProviderConfig,
+    cli_provider_config: crate::ai::provider::ProviderConfig,
+    request_params: api::RequestParams,
+}
+
 #[derive(Clone)]
 struct ActiveProviderRunCheckpoint {
     run: ProviderRun,
@@ -663,6 +697,7 @@ struct ActiveProviderRunCheckpoint {
 struct PreparedRestoredProviderRun {
     snapshot: ActiveProviderRunSnapshot,
     profiles: BTreeMap<String, ProviderRunProfile>,
+    projection_was_initialized: bool,
 }
 
 impl ActiveProviderRunCheckpoint {
@@ -706,6 +741,8 @@ struct ActiveProviderRunSnapshot {
     root_task_id: TaskId,
     did_input_contain_user_query: bool,
     persistence_offset: usize,
+    #[serde(default)]
+    cancellation_reason: Option<CancellationReason>,
     committed_provider_batch: Option<ExternalWorkId>,
     #[serde(default)]
     finished_provider_batch: Option<ExternalWorkId>,
@@ -746,6 +783,7 @@ impl ActiveProviderRunSnapshot {
             root_task_id: slot.root_task_id.clone(),
             did_input_contain_user_query: slot.did_input_contain_user_query,
             persistence_offset: checkpoint.persistence_offset,
+            cancellation_reason: slot.cancellation_reason,
             committed_provider_batch: slot.committed_provider_batch.clone(),
             finished_provider_batch: slot.finished_provider_batch.clone(),
             command_action_refs: slot.command_action_refs.clone(),
@@ -996,11 +1034,25 @@ fn normalize_restored_provider_snapshot(
 fn apply_restored_provider_command_evidence(
     conversation_id: AIConversationId,
     snapshot: &mut ActiveProviderRunSnapshot,
-    evidence: RestoredProviderCommandEvidence,
+    evidence: Option<RestoredProviderCommandEvidence>,
 ) -> Result<(), String> {
     let Some(monitor) = snapshot.command_monitor.as_ref() else {
         return Ok(());
     };
+    let Some(evidence) = evidence else {
+        snapshot.pending_monitor_observation = None;
+        snapshot.pending_command_completion = Some(PendingProviderCommandCompletion {
+            block_id: monitor.block_id.clone(),
+            initial_requested_command_action_id: Some(
+                monitor.initial_requested_command_action_id.clone(),
+            ),
+            command: monitor.command.clone(),
+            output: "The monitored command was interrupted while Galaxy was offline; its terminal block is no longer available."
+                .to_owned(),
+            exit_code: 130,
+        });
+        return Ok(());
+    };
     if evidence.conversation_id != Some(conversation_id)
         || evidence.requested_command_action_id.as_ref()
             != Some(&monitor.initial_requested_command_action_id)
@@ -1038,6 +1090,20 @@ fn apply_restored_provider_command_evidence(
     Ok(())
 }
 
+fn restored_projection_was_initialized(
+    has_output: bool,
+    has_server_output_id: bool,
+    has_added_messages: bool,
+) -> Result<bool, String> {
+    match (has_output, has_server_output_id, has_added_messages) {
+        (false, false, false) => Ok(false),
+        (true, true, _) => Ok(true),
+        (false, true, _) | (false, false, true) | (true, false, _) => {
+            Err("restored provider projection exchange is partially initialized".to_owned())
+        }
+    }
+}
+
 fn provider_execution_matches_active_work(
     run_id: &ProviderRunId,
     active_work_id: Option<&ExternalWorkId>,
@@ -1147,6 +1213,7 @@ fn classify_provider_command_result(
                 command: Some(command.clone()),
             }),
             RequestCommandOutputResult::CancelledBeforeExecution
+            | RequestCommandOutputResult::ExecutionError { .. }
             | RequestCommandOutputResult::Denylisted { .. } => None,
         },
         AIAgentActionResultType::WriteToLongRunningShellCommand(result) => match result {
@@ -1430,7 +1497,7 @@ fn provider_llm_lifecycle(projection: &ProviderRunProjection) -> Option<Provider
             error: Some(error.message.clone()),
         },
         ProviderRunProjection::ModelEvent { .. } | ProviderRunProjection::ToolBatchReady { .. } => {
-            return None
+            return None;
         }
     };
     Some(lifecycle)
@@ -1535,8 +1602,11 @@ fn provider_run_terminal_remote_log_record(
 }
 
 enum ProviderDriveMessage {
-    Response(warp_multi_agent_api::ResponseEvent),
+    Projection {
-    Lifecycle(ProviderLlmLifecycle),
+        lifecycle: Option<ProviderLlmLifecycle>,
+        events: Vec<warp_multi_agent_api::ResponseEvent>,
+        acknowledgement: oneshot::Sender<Result<(), String>>,
+    },
     Checkpoint {
         checkpoint: ActiveProviderRunCheckpoint,
         acknowledgement: oneshot::Sender<Result<(), String>>,
@@ -1559,6 +1629,7 @@ pub struct BlocklistAIController {
 
     in_flight_response_streams: PendingResponseStreams,
     active_provider_runs: HashMap<AIConversationId, ActiveProviderRunSlot>,
+    queued_provider_runs: HashMap<AIConversationId, VecDeque<QueuedProviderRun>>,
     restoring_provider_runs: HashSet<AIConversationId>,
 
     /// The ID of the terminal surface this controller is associated with.
@@ -2048,6 +2119,7 @@ impl BlocklistAIController {
             terminal_model,
             in_flight_response_streams: PendingResponseStreams::new(),
             active_provider_runs: HashMap::new(),
+            queued_provider_runs: HashMap::new(),
             restoring_provider_runs: HashSet::new(),
             terminal_surface_id,
             should_refresh_available_llms_on_stream_finish: false,
@@ -2606,6 +2678,10 @@ impl BlocklistAIController {
         if self
             .in_flight_response_streams
             .has_active_stream_for_conversation(conversation_id, ctx)
+            && !self
+                .active_provider_runs
+                .get(&conversation_id)
+                .is_some_and(|slot| slot.cancellation_reason.is_some())
             || self
                 .action_model
                 .as_ref(ctx)
@@ -4563,7 +4639,7 @@ impl BlocklistAIController {
             .all_inputs()
             .any(|input| input.is_user_query());
         ctx.subscribe_to_model(&response_stream, move |me, _, event, ctx| {
-            me.handle_response_stream_event(
+            let _ = me.handle_response_stream_event(
                 input_contains_user_query,
                 event,
                 &response_stream_clone,
@@ -4625,15 +4701,24 @@ impl BlocklistAIController {
         } else {
             None
         };
-        self.in_flight_response_streams.register_new_stream(
+        if provider_configs.is_some()
-            response_stream_id.clone(),
+            && self
-            conversation_data.id,
+                .active_provider_runs
-            response_stream.clone(),
+                .contains_key(&conversation_data.id)
-            CancellationReason::FollowUpSubmitted {
+        {
-                is_for_same_conversation: true,
+            self.in_flight_response_streams
-            },
+                .register_additional_stream(response_stream_id.clone(), response_stream.clone());
-            ctx,
+        } else {
-        );
+            self.in_flight_response_streams.register_new_stream(
+                response_stream_id.clone(),
+                conversation_data.id,
+                response_stream.clone(),
+                CancellationReason::FollowUpSubmitted {
+                    is_for_same_conversation: true,
+                },
+                ctx,
+            );
+        }
         if let Some((base_provider_config, cli_provider_config)) = provider_configs {
             let provider_run_id = ProviderRunId::new(format!(
                 "{}:{}",
@@ -4646,37 +4731,50 @@ impl BlocklistAIController {
                 .expect("conversation exists while starting provider run")
                 .get_root_task_id()
                 .clone();
-            self.active_provider_runs.insert(
+            let slot = ActiveProviderRunSlot {
-                conversation_data.id,
+                stream_id: response_stream_id.clone(),
-                ActiveProviderRunSlot {
+                response_stream,
-                    stream_id: response_stream_id.clone(),
+                did_input_contain_user_query: input_contains_user_query,
-                    response_stream,
+                run_id: provider_run_id,
-                    did_input_contain_user_query: input_contains_user_query,
+                root_task_id,
-                    run_id: provider_run_id,
+                projection_target: provider_projection_target
-                    root_task_id,
+                    .expect("provider projection target was validated"),
-                    projection_target: provider_projection_target
+                run: None,
-                        .expect("provider projection target was validated"),
+                checkpoint: None,
-                    run: None,
+                turn_control: None,
-                    checkpoint: None,
+                cancellation_reason: None,
-                    turn_control: None,
+                committed_provider_batch: None,
-                    cancellation_reason: None,
+                finished_provider_batch: None,
-                    committed_provider_batch: None,
+                command_action_refs: HashMap::new(),
-                    finished_provider_batch: None,
+                command_monitor: None,
-                    command_action_refs: HashMap::new(),
+                pending_monitor_observation: None,
-                    command_monitor: None,
+                pending_command_completion: None,
-                    pending_monitor_observation: None,
+                monitor_prose_continuations: 0,
-                    pending_command_completion: None,
+            };
-                    monitor_prose_continuations: 0,
+            if self
-                },
+                .active_provider_runs
-            );
+                .contains_key(&conversation_data.id)
-            self.prepare_active_provider_run(
+            {
-                conversation_data.id,
+                self.queued_provider_runs
-                response_stream_id.clone(),
+                    .entry(conversation_data.id)
-                base_provider_config,
+                    .or_default()
-                cli_provider_config,
+                    .push_back(QueuedProviderRun {
-                request_params.clone(),
+                        slot,
-                ctx,
+                        base_provider_config,
-            );
+                        cli_provider_config,
+                        request_params: request_params.clone(),
+                    });
+            } else {
+                self.active_provider_runs.insert(conversation_data.id, slot);
+                self.prepare_active_provider_run(
+                    conversation_data.id,
+                    response_stream_id.clone(),
+                    base_provider_config,
+                    cli_provider_config,
+                    request_params.clone(),
+                    ctx,
+                );
+            }
         }
 
         // Skip the context reset for a fired queued-prompt row (`is_queued_prompt`): its
@@ -4806,21 +4904,29 @@ impl BlocklistAIController {
                 let Some(task) = conversation.get_task(&snapshot.projection_target.task_id) else {
                     return Err("restored provider projection task is missing".to_string());
                 };
-                if !task
+                let Some(exchange) = task
                     .exchanges()
-                    .any(|exchange| exchange.id == snapshot.projection_target.exchange_id)
+                    .find(|exchange| exchange.id == snapshot.projection_target.exchange_id)
-                {
+                else {
                     return Err(
                         "restored provider projection exchange is missing from its task"
                             .to_string(),
                     );
-                }
+                };
-                Ok(())
+                let output = exchange.output_status.output();
+                restored_projection_was_initialized(
+                    output.is_some(),
+                    output.is_some_and(|output| output.get().server_output_id.is_some()),
+                    !exchange.added_message_ids.is_empty(),
+                )
             });
-        if let Err(error) = history_validation {
+        let projection_was_initialized = match history_validation {
-            self.fail_restored_provider_run(conversation_id, error, ctx);
+            Ok(initialized) => initialized,
-            return;
+            Err(error) => {
-        }
+                self.fail_restored_provider_run(conversation_id, error, ctx);
+                return;
+            }
+        };
 
         if let Err(error) = normalize_restored_provider_snapshot(&mut snapshot) {
             self.fail_restored_provider_run(conversation_id, error, ctx);
@@ -4868,7 +4974,11 @@ impl BlocklistAIController {
                         ProviderRunProfile::new(cli_runtime, cli_monitor_request.clone()),
                     );
                 }
-                Ok::<_, anyhow::Error>(PreparedRestoredProviderRun { snapshot, profiles })
+                Ok::<_, anyhow::Error>(PreparedRestoredProviderRun {
+                    snapshot,
+                    profiles,
+                    projection_was_initialized,
+                })
             },
             move |me, result, ctx| {
                 me.handle_prepared_restored_provider_run(conversation_id, result, ctx);
@@ -4886,19 +4996,18 @@ impl BlocklistAIController {
         };
         let evidence = {
             let terminal_model = self.terminal_model.lock();
-            let block = terminal_model
+            terminal_model
                 .block_list()
                 .block_with_id(&monitor.block_id)
-                .ok_or_else(|| "restored provider command block is missing".to_string())?;
+                .map(|block| RestoredProviderCommandEvidence {
-            RestoredProviderCommandEvidence {
+                    conversation_id: block.ai_conversation_id(),
-                conversation_id: block.ai_conversation_id(),
+                    requested_command_action_id: block.requested_command_action_id().cloned(),
-                requested_command_action_id: block.requested_command_action_id().cloned(),
+                    cli_task_id: block.cli_subagent_task_id().cloned(),
-                cli_task_id: block.cli_subagent_task_id().cloned(),
+                    command: block.command_to_string(),
-                command: block.command_to_string(),
+                    state: block.state(),
-                state: block.state(),
+                    output: block.output_to_string(),
-                output: block.output_to_string(),
+                    exit_code: block.exit_code().value(),
-                exit_code: block.exit_code().value(),
+                })
-            }
         };
         apply_restored_provider_command_evidence(conversation_id, snapshot, evidence)
     }
@@ -4930,7 +5039,11 @@ impl BlocklistAIController {
             self.restoring_provider_runs.remove(&conversation_id);
             return;
         }
-        let PreparedRestoredProviderRun { snapshot, profiles } = match result {
+        let PreparedRestoredProviderRun {
+            snapshot,
+            profiles,
+            projection_was_initialized,
+        } = match result {
             Ok(prepared) => prepared,
             Err(error) => {
                 self.fail_restored_provider_run(conversation_id, error.to_string(), ctx);
@@ -4948,6 +5061,7 @@ impl BlocklistAIController {
             root_task_id,
             did_input_contain_user_query,
             persistence_offset,
+            cancellation_reason,
             committed_provider_batch,
             finished_provider_batch,
             command_action_refs,
@@ -4960,13 +5074,21 @@ impl BlocklistAIController {
         let transcript = provider_run.transcript();
         let offset = persistence_offset.min(transcript.len());
         let messages_sent = Arc::new(std::sync::Mutex::new(transcript[offset..].to_vec()));
-        let coordinator = match ProviderRunCoordinator::new(provider_run, profiles) {
+        let mut coordinator = match ProviderRunCoordinator::new(provider_run, profiles) {
             Ok(coordinator) => coordinator,
             Err(error) => {
                 self.fail_restored_provider_run(conversation_id, error.to_string(), ctx);
                 return;
             }
         };
+        if let Some(reason) = cancellation_reason {
+            if !coordinator.run().is_terminal() {
+                if let Err(error) = coordinator.run_mut().cancel(reason.to_string()) {
+                    self.fail_restored_provider_run(conversation_id, error.to_string(), ctx);
+                    return;
+                }
+            }
+        }
         let model = LLMId::from(response_config.model_id.as_str());
         let ai_identifiers = AIIdentifiers {
             client_conversation_id: Some(conversation_id),
@@ -4984,7 +5106,7 @@ impl BlocklistAIController {
         let stream_id = response_stream.as_ref(ctx).id().clone();
         let response_stream_clone = response_stream.clone();
         ctx.subscribe_to_model(&response_stream, move |me, _, event, ctx| {
-            me.handle_response_stream_event(
+            let _ = me.handle_response_stream_event(
                 did_input_contain_user_query,
                 event,
                 &response_stream_clone,
@@ -5032,7 +5154,10 @@ impl BlocklistAIController {
                 projection_target,
                 run: Some(ActiveProviderRun {
                     coordinator,
-                    projector: ProviderRunResponseProjector::restored(response_config.clone()),
+                    projector: ProviderRunResponseProjector::restored(
+                        response_config.clone(),
+                        projection_was_initialized,
+                    ),
                     response_config,
                     action_context,
                     messages_sent,
@@ -5040,7 +5165,7 @@ impl BlocklistAIController {
                 }),
                 checkpoint: None,
                 turn_control: None,
-                cancellation_reason: None,
+                cancellation_reason,
                 committed_provider_batch,
                 finished_provider_batch,
                 command_action_refs,
@@ -5320,26 +5445,29 @@ impl BlocklistAIController {
                 let checkpoint_sender = sender.clone();
                 let result = run
                     .coordinator
-                    .drive_until_blocked_with_checkpoint(
+                    .drive_until_blocked_with_acknowledgements(
                         turn_control,
                         |projection| {
-                            if let Some(lifecycle) = provider_llm_lifecycle(&projection) {
+                            let lifecycle = provider_llm_lifecycle(&projection);
+                            let events = run.projector.project(projection);
+                            let projection_sender = projection_sender.clone();
+                            Box::pin(async move {
+                                let events = events?;
+                                let (acknowledgement, receiver) = oneshot::channel();
                                 projection_sender
-                                    .try_send(ProviderDriveMessage::Lifecycle(lifecycle))
+                                    .send(ProviderDriveMessage::Projection {
+                                        lifecycle,
+                                        events,
+                                        acknowledgement,
+                                    })
+                                    .await
                                     .map_err(|_| {
-                                        "provider lifecycle projection receiver was closed"
+                                        "provider projection receiver was closed".to_string()
-                                            .to_string()
                                     })?;
-                            }
+                                receiver.await.map_err(|_| {
-                            for event in run.projector.project(projection)? {
+                                    "provider projection acknowledgement was dropped".to_string()
-                                projection_sender
+                                })?
-                                    .try_send(ProviderDriveMessage::Response(event))
+                            })
-                                    .map_err(|_| {
-                                        "provider response projection receiver was closed"
-                                            .to_string()
-                                    })?;
-                            }
-                            Ok(())
                         },
                         move |provider_run| {
                             let checkpoint_sender = checkpoint_sender.clone();
@@ -5385,29 +5513,40 @@ impl BlocklistAIController {
             return;
         }
         match message {
-            ProviderDriveMessage::Response(event) => {
+            ProviderDriveMessage::Projection {
+                lifecycle,
+                events,
+                acknowledgement,
+            } => {
                 let response_stream = slot.response_stream.clone();
                 let did_input_contain_user_query = slot.did_input_contain_user_query;
-                let event = ResponseStream::projected_event(event);
+                let mut result = Ok(());
-                self.handle_response_stream_event(
+                for event in events {
-                    did_input_contain_user_query,
+                    let event = ResponseStream::projected_event(event);
-                    &event,
+                    if let Err(error) = self.handle_response_stream_event(
-                    &response_stream,
+                        did_input_contain_user_query,
-                    ctx,
+                        &event,
-                );
+                        &response_stream,
-            }
+                        ctx,
-            ProviderDriveMessage::Lifecycle(lifecycle) => {
+                    ) {
+                        result = Err(error);
+                        break;
+                    }
+                }
                 #[cfg(not(target_family = "wasm"))]
-                remote_logging::log_model_event(
+                if let Some(lifecycle) = lifecycle.as_ref() {
-                    ctx,
+                    remote_logging::log_model_event(
-                    provider_llm_lifecycle_remote_log_record(
+                        ctx,
-                        conversation_id,
+                        provider_llm_lifecycle_remote_log_record(
-                        stream_id,
+                            conversation_id,
-                        &lifecycle,
+                            stream_id,
-                    ),
+                            lifecycle,
-                );
+                        ),
+                    );
+                }
                 #[cfg(target_family = "wasm")]
                 let _ = lifecycle;
+                let _ = acknowledgement.send(result);
             }
             ProviderDriveMessage::Checkpoint {
                 checkpoint,
@@ -5691,35 +5830,60 @@ impl BlocklistAIController {
         batch: PendingToolBatch,
         ctx: &mut ModelContext<Self>,
     ) {
-        let conversion = self
+        let Some(run) = self
             .active_provider_runs
-            .get(&conversation_id)
+            .get_mut(&conversation_id)
-            .and_then(|slot| slot.run.as_ref())
+            .and_then(|slot| slot.run.as_mut())
-            .map(|run| {
+        else {
-                batch
+            return;
-                    .calls
+        };
-                    .iter()
+        let (converted_actions, invalid_results) =
-                    .filter(|pending| pending.state.result().is_none())
+            convert_provider_tool_batch(&run.action_context, &batch);
-                    .map(|pending| {
+        for result in &invalid_results {
-                        run.action_context
+            if let Err(error) = run
-                            .action_from_tool_call(&pending.call)
+                .coordinator
-                            .map(|action| {
+                .run_mut()
-                                (
+                .complete_tool(&batch.work_id, result.clone())
-                                    action,
+            {
-                                    matches!(pending.state, PendingToolCallState::RecoveryPending),
+                self.fail_active_provider_run(
-                                )
+                    conversation_id,
-                            })
+                    format!("failed to record invalid provider tool input: {error}"),
-                    })
+                    ctx,
-                    .collect::<Result<Vec<_>, _>>()
+                );
-            });
-        let converted_actions = match conversion {
-            Some(Ok(actions)) => actions,
-            Some(Err(message)) => {
-                self.fail_active_provider_run(conversation_id, message, ctx);
                 return;
             }
-            None => return,
+        }
-        };
+        if converted_actions.is_empty() {
+            if let Err(error) = run.coordinator.run_mut().commit_tool_batch(&batch.work_id) {
+                self.fail_active_provider_run(
+                    conversation_id,
+                    format!("failed to commit invalid provider tool batch: {error}"),
+                    ctx,
+                );
+                return;
+            }
+            if let Err(error) = self.persist_active_provider_run(conversation_id, ctx) {
+                self.fail_active_provider_run(
+                    conversation_id,
+                    format!("failed to persist invalid provider tool results: {error}"),
+                    ctx,
+                );
+                return;
+            }
+            self.drive_active_provider_run(conversation_id, ctx);
+            return;
+        }
+        let mut executable_batch = batch.clone();
+        for pending in &mut executable_batch.calls {
+            if let Some(result) = invalid_results
+                .iter()
+                .find(|result| result.call_id == pending.call.id)
+            {
+                pending.state = PendingToolCallState::Resolved {
+                    result: result.clone(),
+                };
+            }
+        }
         let stream_id = self.active_provider_runs[&conversation_id]
             .stream_id
             .clone();
@@ -5806,7 +5970,7 @@ impl BlocklistAIController {
                 actions,
                 recovery_action_ids,
                 conversation_id,
-                &batch,
+                &executable_batch,
                 ctx,
             )
         });
@@ -6038,6 +6202,33 @@ impl BlocklistAIController {
         }
     }
 
+    fn detach_cancelled_provider_command(
+        &mut self,
+        conversation_id: AIConversationId,
+        block_id: &BlockId,
+        ctx: &mut ModelContext<Self>,
+    ) {
+        let detached = {
+            let mut terminal_model = self.terminal_model.lock();
+            let active_block = terminal_model.block_list_mut().active_block_mut();
+            if active_block.id() == block_id
+                && active_block.ai_conversation_id() == Some(conversation_id)
+                && active_block.is_active_and_long_running()
+            {
+                active_block.set_user_control_with_stop_reason();
+                true
+            } else {
+                false
+            }
+        };
+        self.deactivate_provider_cli_task(conversation_id, block_id, ctx);
+        if !detached {
+            log::warn!(
+                "Could not detach cancelled provider command for conversation {conversation_id:?} block {block_id:?}"
+            );
+        }
+    }
+
     fn handle_provider_actions_finished(
         &mut self,
         conversation_id: AIConversationId,
@@ -6219,7 +6410,7 @@ impl BlocklistAIController {
         let did_input_contain_user_query = slot.did_input_contain_user_query;
         for event in events {
             let event = ResponseStream::projected_event(event);
-            self.handle_response_stream_event(
+            let _ = self.handle_response_stream_event(
                 did_input_contain_user_query,
                 &event,
                 &response_stream,
@@ -6237,9 +6428,35 @@ impl BlocklistAIController {
             ),
         );
         match outcome {
-            ProviderRunOutcome::Completed(_) => {
+            ProviderRunOutcome::Completed(completion) => match completion.stop_reason {
-                self.finalize_completed_provider_conversation(conversation_id, ctx);
+                StopReason::Completed => {
-            }
+                    self.finalize_completed_provider_conversation(conversation_id, ctx);
+                }
+                StopReason::Cancelled => {
+                    BlocklistAIHistoryModel::handle(ctx).update(ctx, |history_model, ctx| {
+                        history_model.update_conversation_status(
+                            self.terminal_surface_id,
+                            conversation_id,
+                            ConversationStatus::Cancelled,
+                            ctx,
+                        );
+                    });
+                }
+                StopReason::MaxTokens
+                | StopReason::ContextWindowExceeded
+                | StopReason::Refusal
+                | StopReason::ToolLoopLimit
+                | StopReason::Other(_) => {
+                    BlocklistAIHistoryModel::handle(ctx).update(ctx, |history_model, ctx| {
+                        history_model.update_conversation_status(
+                            self.terminal_surface_id,
+                            conversation_id,
+                            ConversationStatus::Error,
+                            ctx,
+                        );
+                    });
+                }
+            },
             // Failed outcomes are finalized by the projected InternalError event.
             ProviderRunOutcome::Failed(_) => {}
             ProviderRunOutcome::Cancelled { .. } => {
@@ -6314,6 +6531,13 @@ impl BlocklistAIController {
         response_stream: &ModelHandle<ResponseStream>,
         ctx: &mut ModelContext<Self>,
     ) {
+        if !self
+            .active_provider_runs
+            .get(&conversation_id)
+            .is_some_and(|slot| &slot.stream_id == stream_id)
+        {
+            return;
+        }
         if let Err(error) = self.clear_persisted_active_provider_run(conversation_id, ctx) {
             log::error!("Failed to clear persisted provider run during cleanup: {error}");
         }
@@ -6334,6 +6558,38 @@ impl BlocklistAIController {
             request_usage_model.refresh_request_usage_async(ctx);
         });
         self.maybe_refresh_ai_overages(ctx);
+        self.start_next_queued_provider_run(conversation_id, ctx);
+    }
+
+    fn start_next_queued_provider_run(
+        &mut self,
+        conversation_id: AIConversationId,
+        ctx: &mut ModelContext<Self>,
+    ) {
+        let next = self
+            .queued_provider_runs
+            .get_mut(&conversation_id)
+            .and_then(VecDeque::pop_front);
+        if self
+            .queued_provider_runs
+            .get(&conversation_id)
+            .is_some_and(VecDeque::is_empty)
+        {
+            self.queued_provider_runs.remove(&conversation_id);
+        }
+        let Some(next) = next else {
+            return;
+        };
+        let stream_id = next.slot.stream_id.clone();
+        self.active_provider_runs.insert(conversation_id, next.slot);
+        self.prepare_active_provider_run(
+            conversation_id,
+            stream_id,
+            next.base_provider_config,
+            next.cli_provider_config,
+            next.request_params,
+            ctx,
+        );
     }
 
     fn cancel_active_provider_run(
@@ -6342,53 +6598,59 @@ impl BlocklistAIController {
         reason: CancellationReason,
         ctx: &mut ModelContext<Self>,
     ) -> bool {
-        let Some(mut slot) = self.active_provider_runs.remove(&conversation_id) else {
+        let cancellation_outcome = reason.conversation_outcome();
+        let Some(slot) = self.active_provider_runs.get_mut(&conversation_id) else {
             return false;
         };
         slot.cancellation_reason = Some(reason);
+        let command_block_id = if matches!(cancellation_outcome, CancellationOutcome::Cancelled) {
+            let monitor = slot.command_monitor.take();
+            if let Some(monitor) = &monitor {
+                slot.command_action_refs
+                    .remove(&monitor.initial_requested_command_action_id);
+            }
+            slot.pending_monitor_observation = None;
+            slot.pending_command_completion = None;
+            monitor.map(|monitor| monitor.block_id)
+        } else {
+            None
+        };
         if let Some(turn_control) = &slot.turn_control {
             let _ = turn_control.try_send(TurnCommand::Cancel);
         }
-        if let Some(mut run) = slot.run.take() {
+        let should_drive = if let Some(run) = slot.run.as_mut() {
             if !run.coordinator.run().is_terminal() {
                 let _ = run.coordinator.run_mut().cancel(reason.to_string());
             }
-            if let Ok(mut messages_sent) = run.messages_sent.lock() {
+            true
-                let transcript = run.coordinator.run().transcript();
+        } else {
-                let offset = run.persistence_offset.min(transcript.len());
+            false
-                *messages_sent = transcript[offset..].to_vec();
+        };
-            }
+
+        // Keep the terminal run and its slot durable until the normal driver path projects the
+        // cancellation and finalizes it through `finish_active_provider_run`.
+        if let Err(error) = self.persist_active_provider_run(conversation_id, ctx) {
+            log::error!("Failed to persist provider cancellation: {error}");
         }
 
         self.action_model.update(ctx, |action_model, ctx| {
             action_model.cancel_all_pending_actions(conversation_id, Some(reason), ctx);
         });
 
-        let cancellation_outcome = reason.conversation_outcome();
         if FeatureFlag::AgentSharedSessions.is_enabled()
             && !matches!(cancellation_outcome, CancellationOutcome::KeepInProgress)
         {
             self.send_cancellation_to_viewers(ctx);
         }
-        BlocklistAIHistoryModel::handle(ctx).update(ctx, |history_model, ctx| {
-            history_model.mark_response_stream_cancelled(
-                &slot.stream_id,
-                conversation_id,
-                self.terminal_surface_id,
-                reason,
-                ctx,
-            );
-        });
         if matches!(cancellation_outcome, CancellationOutcome::Cancelled) {
             self.set_input_mode_for_cancellation(ctx);
+            if let Some(block_id) = command_block_id {
+                self.detach_cancelled_provider_command(conversation_id, &block_id, ctx);
+            }
+        }
+        if should_drive {
+            self.drive_active_provider_run(conversation_id, ctx);
         }
-
-        self.cleanup_active_provider_run(
-            conversation_id,
-            &slot.stream_id,
-            &slot.response_stream,
-            ctx,
-        );
         true
     }
 
@@ -6535,7 +6797,7 @@ impl BlocklistAIController {
     ) {
         let stream_clone = stream.clone();
         ctx.subscribe_to_model(&stream, move |me, _, event, ctx| {
-            me.handle_response_stream_event(false, event, &stream_clone, ctx);
+            let _ = me.handle_response_stream_event(false, event, &stream_clone, ctx);
         });
         self.in_flight_response_streams.register_new_stream(
             stream_id,
@@ -6739,7 +7001,7 @@ impl BlocklistAIController {
         event: &ResponseStreamEvent,
         response_stream: &ModelHandle<ResponseStream>,
         ctx: &mut ModelContext<Self>,
-    ) {
+    ) -> Result<(), String> {
         let stream_id = response_stream.as_ref(ctx).id().clone();
 
         match event {
@@ -6749,14 +7011,16 @@ impl BlocklistAIController {
                     .conversation_for_response_stream(&stream_id)
                 else {
                     log::warn!("Could not find conversation for response stream: {stream_id:?}");
-                    return;
+                    return Err(format!(
+                        "could not find conversation for response stream {stream_id:?}"
+                    ));
                 };
                 let Some(event) = event.consume() else {
                     debug_assert!(
                         false,
                         "This model should only have a single subscriber that takes ownership over the event."
                     );
-                    return;
+                    return Err("response stream event was already consumed".to_string());
                 };
                 let history_model = BlocklistAIHistoryModel::handle(ctx);
                 match event {
@@ -6794,7 +7058,7 @@ impl BlocklistAIController {
                             }
                         }
                         let Some(event) = event.r#type else {
-                            return;
+                            return Err("response event did not contain a type".to_string());
                         };
                         match event {
                             warp_multi_agent_api::response_event::Type::Init(init_event) => {
@@ -6909,6 +7173,9 @@ impl BlocklistAIController {
                                     log::error!(
                                         "Failed to apply client actions to conversation: {e:?}"
                                     );
+                                    return Err(format!(
+                                        "failed to apply provider client actions: {e:?}"
+                                    ));
                                 }
                             }
                         }
@@ -6954,7 +7221,9 @@ impl BlocklistAIController {
                             log::warn!(
                                 "Could not find conversation for response stream: {stream_id:?}"
                             );
-                            return;
+                            return Err(format!(
+                                "could not find conversation for response stream {stream_id:?}"
+                            ));
                         };
                         id
                     }
@@ -6980,7 +7249,7 @@ impl BlocklistAIController {
                     })
                 else {
                     log::warn!("Conversation not found.");
-                    return;
+                    return Err("conversation not found for completed response stream".to_string());
                 };
                 #[cfg(not(target_family = "wasm"))]
                 if let Some(metadata) = response_stream.as_ref(ctx).acp_session_metadata() {
@@ -7036,7 +7305,7 @@ impl BlocklistAIController {
                 for new_exchange_id in new_exchange_ids {
                     let Some(exchange) = exchanges.exchange_with_id(new_exchange_id) else {
                         log::warn!("Exchange not found.");
-                        return;
+                        return Err("exchange not found for completed response stream".to_string());
                     };
                     was_passive_request |= exchange.has_passive_request();
                     is_any_exchange_unfinished |= !exchange.output_status.is_finished();
@@ -7328,6 +7597,7 @@ impl BlocklistAIController {
                 self.maybe_refresh_ai_overages(ctx);
             }
         }
+        Ok(())
     }
 
     /// Sets the terminal input state after an AI request is cancelled.

app/src/ai/blocklist/controller/pending_response_streams.rs

@@ -91,6 +91,14 @@ impl PendingResponseStreams {
         self.streams.insert(stream_id, stream);
     }
 
+    pub fn register_additional_stream(
+        &mut self,
+        stream_id: ResponseStreamId,
+        stream: ModelHandle<ResponseStream>,
+    ) {
+        self.streams.insert(stream_id, stream);
+    }
+
     pub fn cleanup_stream(&mut self, stream_id: &ResponseStreamId) {
         self.streams.remove(stream_id);
     }
@@ -142,9 +150,11 @@ impl PendingResponseStreams {
             for response_stream in streams_to_cancel.into_iter() {
                 log::info!(
                     "Canceling active stream for conversation_id={conversation_id:?}, \
-                     reason={reason}, backtrace=\n{}",
+                     reason={reason}"
-                    std::backtrace::Backtrace::force_capture()
                 );
+                if let Some(backtrace) = crate::ai::tool_diagnostics::capture_backtrace() {
+                    log::debug!("Active stream cancellation backtrace:\n{backtrace}");
+                }
                 response_stream.update(ctx, |stream, ctx| {
                     stream.cancel(reason, conversation_id, ctx)
                 });

app/src/ai/blocklist/controller/shared_session.rs

@@ -354,7 +354,7 @@ impl BlocklistAIController {
             if self
                 .action_model
                 .as_ref(ctx)
-                .get_action_result(&result.id)
+                .get_action_result(conversation_id, &result.id)
                 .is_none()
             {
                 self.action_model.update(ctx, |action_model, ctx| {

app/src/ai/blocklist/controller_tests.rs

@@ -25,6 +25,7 @@ use crate::ai::agent::{
     WriteToLongRunningShellCommandResult,
 };
 use crate::ai::ambient_agents::AmbientAgentTaskId;
+use crate::ai::blocklist::action_model::StartAgentWaitPolicy;
 use crate::ai::blocklist::{
     BlocklistAIHistoryEvent, BlocklistAIHistoryModel, PendingAttachment, PendingFile, RequestInput,
     ResponseStream, ResponseStreamId, StartAgentExecutor,
@@ -265,6 +266,7 @@ fn provider_snapshot(conversation_id: AIConversationId) -> super::ActiveProvider
         root_task_id: task_id,
         did_input_contain_user_query: true,
         persistence_offset: 0,
+        cancellation_reason: None,
         committed_provider_batch: None,
         finished_provider_batch: None,
         command_action_refs: HashMap::new(),
@@ -275,6 +277,278 @@ fn provider_snapshot(conversation_id: AIConversationId) -> super::ActiveProvider
     }
 }
 
+#[test]
+fn provider_snapshot_persists_cancellation_reason() {
+    let conversation_id = AIConversationId::new();
+    let mut snapshot = provider_snapshot(conversation_id);
+    snapshot.cancellation_reason = Some(CancellationReason::ManuallyCancelled);
+
+    let restored = super::ActiveProviderRunSnapshot::parse(
+        &serde_json::to_string(&snapshot).expect("cancellation snapshot should serialize"),
+    )
+    .expect("cancellation snapshot should parse");
+
+    assert_eq!(
+        restored.cancellation_reason,
+        Some(CancellationReason::ManuallyCancelled)
+    );
+    assert!(!restored.run.is_terminal());
+}
+
+#[test]
+fn cancelling_provider_startup_keeps_slot_and_durable_cancellation() {
+    App::test((), |mut app| async move {
+        initialize_app_for_terminal_view(&mut app);
+        let terminal = add_window_with_terminal(&mut app, None);
+
+        terminal.update(&mut app, |terminal, ctx| {
+            let terminal_surface_id = terminal.id();
+            let conversation_id =
+                BlocklistAIHistoryModel::handle(ctx).update(ctx, |history_model, ctx| {
+                    history_model.start_new_conversation(
+                        terminal_surface_id,
+                        false,
+                        false,
+                        false,
+                        ctx,
+                    )
+                });
+            let snapshot = provider_snapshot(conversation_id);
+            let stream_id = ResponseStreamId::new_for_test();
+            let response_stream =
+                ctx.add_model(|_| ResponseStream::new_for_test(stream_id.clone()));
+            let checkpoint = super::ActiveProviderRunCheckpoint {
+                run: snapshot.run.clone(),
+                base_request: snapshot.base_request.clone(),
+                cli_monitor_request: snapshot.cli_monitor_request.clone(),
+                response_config: snapshot.response_config.clone(),
+                action_context: snapshot.action_context.clone(),
+                persistence_offset: snapshot.persistence_offset,
+            };
+
+            terminal.ai_controller().update(ctx, |controller, ctx| {
+                controller.active_provider_runs.insert(
+                    conversation_id,
+                    super::ActiveProviderRunSlot {
+                        stream_id,
+                        response_stream,
+                        did_input_contain_user_query: snapshot.did_input_contain_user_query,
+                        run_id: snapshot.run.id().clone(),
+                        root_task_id: snapshot.root_task_id,
+                        projection_target: snapshot.projection_target,
+                        run: None,
+                        checkpoint: Some(checkpoint),
+                        turn_control: None,
+                        cancellation_reason: None,
+                        committed_provider_batch: None,
+                        finished_provider_batch: None,
+                        command_action_refs: HashMap::new(),
+                        command_monitor: None,
+                        pending_monitor_observation: None,
+                        pending_command_completion: None,
+                        monitor_prose_continuations: 0,
+                    },
+                );
+
+                assert!(controller.cancel_active_provider_run(
+                    conversation_id,
+                    CancellationReason::ManuallyCancelled,
+                    ctx,
+                ));
+                assert_eq!(
+                    controller
+                        .active_provider_runs
+                        .get(&conversation_id)
+                        .and_then(|slot| slot.cancellation_reason),
+                    Some(CancellationReason::ManuallyCancelled)
+                );
+            });
+
+            let conversation = BlocklistAIHistoryModel::as_ref(ctx)
+                .conversation(&conversation_id)
+                .expect("cancelled provider conversation should remain durable");
+            let persisted = super::ActiveProviderRunSnapshot::parse(
+                conversation
+                    .active_provider_run_json()
+                    .expect("cancelled provider run should remain checkpointed"),
+            )
+            .expect("persisted cancellation should parse");
+            assert_eq!(
+                persisted.cancellation_reason,
+                Some(CancellationReason::ManuallyCancelled)
+            );
+            assert_eq!(conversation.status(), &ConversationStatus::InProgress);
+        });
+    });
+}
+
+#[test]
+fn same_conversation_follow_up_waits_for_cancelled_provider_generation_cleanup() {
+    App::test((), |mut app| async move {
+        initialize_app_for_terminal_view(&mut app);
+        let terminal = add_window_with_terminal(&mut app, None);
+
+        terminal.update(&mut app, |terminal, ctx| {
+            let terminal_surface_id = terminal.id();
+            let conversation_id =
+                BlocklistAIHistoryModel::handle(ctx).update(ctx, |history_model, ctx| {
+                    history_model.start_new_conversation(
+                        terminal_surface_id,
+                        false,
+                        false,
+                        false,
+                        ctx,
+                    )
+                });
+            let mut old_snapshot = provider_snapshot(conversation_id);
+            start_snapshot_tool(&mut old_snapshot, "old-tool");
+            assert!(matches!(
+                old_snapshot.run.state(),
+                ProviderRunState::AwaitingTools { .. }
+            ));
+
+            let old_stream_id = ResponseStreamId::new_for_test();
+            let old_response_stream =
+                ctx.add_model(|_| ResponseStream::new_for_test(old_stream_id.clone()));
+            let new_stream_id = ResponseStreamId::new_for_test();
+            let new_response_stream =
+                ctx.add_model(|_| ResponseStream::new_for_test(new_stream_id.clone()));
+            let new_snapshot = provider_snapshot(conversation_id);
+
+            terminal.ai_controller().update(ctx, |controller, ctx| {
+                controller.active_provider_runs.insert(
+                    conversation_id,
+                    super::ActiveProviderRunSlot {
+                        stream_id: old_stream_id.clone(),
+                        response_stream: old_response_stream.clone(),
+                        did_input_contain_user_query: true,
+                        run_id: old_snapshot.run.id().clone(),
+                        root_task_id: old_snapshot.root_task_id.clone(),
+                        projection_target: old_snapshot.projection_target.clone(),
+                        run: None,
+                        checkpoint: Some(super::ActiveProviderRunCheckpoint {
+                            run: old_snapshot.run.clone(),
+                            base_request: old_snapshot.base_request.clone(),
+                            cli_monitor_request: old_snapshot.cli_monitor_request.clone(),
+                            response_config: old_snapshot.response_config.clone(),
+                            action_context: old_snapshot.action_context.clone(),
+                            persistence_offset: old_snapshot.persistence_offset,
+                        }),
+                        turn_control: None,
+                        cancellation_reason: Some(CancellationReason::FollowUpSubmitted {
+                            is_for_same_conversation: true,
+                        }),
+                        committed_provider_batch: None,
+                        finished_provider_batch: None,
+                        command_action_refs: HashMap::new(),
+                        command_monitor: None,
+                        pending_monitor_observation: None,
+                        pending_command_completion: None,
+                        monitor_prose_continuations: 0,
+                    },
+                );
+                controller
+                    .queued_provider_runs
+                    .entry(conversation_id)
+                    .or_default()
+                    .push_back(super::QueuedProviderRun {
+                        slot: super::ActiveProviderRunSlot {
+                            stream_id: new_stream_id.clone(),
+                            response_stream: new_response_stream,
+                            did_input_contain_user_query: true,
+                            run_id: new_snapshot.run.id().clone(),
+                            root_task_id: new_snapshot.root_task_id,
+                            projection_target: new_snapshot.projection_target,
+                            run: None,
+                            checkpoint: None,
+                            turn_control: None,
+                            cancellation_reason: None,
+                            committed_provider_batch: None,
+                            finished_provider_batch: None,
+                            command_action_refs: HashMap::new(),
+                            command_monitor: None,
+                            pending_monitor_observation: None,
+                            pending_command_completion: None,
+                            monitor_prose_continuations: 0,
+                        },
+                        base_provider_config: crate::ai::provider::ProviderConfig::None,
+                        cli_provider_config: crate::ai::provider::ProviderConfig::None,
+                        request_params: crate::ai::agent::api::RequestParams::new_for_test(),
+                    });
+
+                assert_eq!(
+                    controller.active_provider_runs[&conversation_id].stream_id,
+                    old_stream_id
+                );
+                controller.cleanup_active_provider_run(
+                    conversation_id,
+                    &old_stream_id,
+                    &old_response_stream,
+                    ctx,
+                );
+                assert_eq!(
+                    controller.active_provider_runs[&conversation_id].stream_id,
+                    new_stream_id
+                );
+                assert!(!controller
+                    .queued_provider_runs
+                    .contains_key(&conversation_id));
+
+                // A delayed callback from the old generation cannot remove its replacement.
+                controller.cleanup_active_provider_run(
+                    conversation_id,
+                    &old_stream_id,
+                    &old_response_stream,
+                    ctx,
+                );
+                assert_eq!(
+                    controller.active_provider_runs[&conversation_id].stream_id,
+                    new_stream_id
+                );
+            });
+        });
+    });
+}
+
+#[test]
+fn cancelled_provider_command_detaches_running_process_to_user() {
+    App::test((), |mut app| async move {
+        initialize_app_for_terminal_view(&mut app);
+        let terminal = add_window_with_terminal(&mut app, None);
+
+        terminal.update(&mut app, |terminal, ctx| {
+            let conversation_id = AIConversationId::new();
+            let task_id = TaskId::new("provider-command-monitor".to_owned());
+            let block_id = {
+                let mut terminal_model = terminal.model.lock();
+                terminal_model.simulate_long_running_block("sleep 100", "running");
+                let active_block = terminal_model.block_list_mut().active_block_mut();
+                active_block.set_is_agent_tagged_in(true);
+                active_block
+                    .set_agent_interaction_mode_for_agent_monitored_command(
+                        &task_id,
+                        conversation_id,
+                    )
+                    .expect("command should become agent monitored");
+                active_block.id().clone()
+            };
+
+            terminal.ai_controller().update(ctx, |controller, ctx| {
+                controller.detach_cancelled_provider_command(conversation_id, &block_id, ctx);
+            });
+
+            let terminal_model = terminal.model.lock();
+            let active_block = terminal_model.block_list().active_block();
+            assert!(!active_block.is_agent_in_control());
+            assert!(active_block
+                .long_running_control_state()
+                .and_then(|state| state.user_take_over_reason())
+                .is_some_and(|reason| reason.is_stop()));
+            assert!(active_block.is_active_and_long_running());
+        });
+    });
+}
+
 fn start_snapshot_tool(
     snapshot: &mut super::ActiveProviderRunSnapshot,
     call_id: &str,
@@ -305,6 +579,105 @@ fn start_snapshot_tool(
     batch.work_id
 }
 
+#[test]
+fn malformed_provider_tool_inputs_become_correlated_errors_without_dropping_valid_calls() {
+    let conversation_id = AIConversationId::new();
+    let mut snapshot = provider_snapshot(conversation_id);
+    let Some(ProviderRunStep::CallModel(call)) = snapshot.run.next_step().unwrap() else {
+        panic!("expected provider model call");
+    };
+    snapshot
+        .run
+        .accept_model_turn(
+            &call.work_id,
+            CompletedModelTurn {
+                assistant_content: vec![],
+                tool_calls: vec![
+                    ToolCall {
+                        id: "bad-read".to_owned(),
+                        name: "read_files".to_owned(),
+                        arguments: serde_json::json!({"files": "not-an-array"}),
+                    },
+                    ToolCall {
+                        id: "good-grep".to_owned(),
+                        name: "grep".to_owned(),
+                        arguments: serde_json::json!({"queries": ["ProviderRun"]}),
+                    },
+                ],
+                usage: Usage::default(),
+                stop_reason: StopReason::Completed,
+                advertised_tools: BTreeSet::from(["grep".to_owned(), "read_files".to_owned()]),
+            },
+        )
+        .unwrap();
+    let Some(ProviderRunStep::DispatchTools(batch)) = snapshot.run.next_step().unwrap() else {
+        panic!("expected provider tool batch");
+    };
+
+    let (actions, errors) = super::convert_provider_tool_batch(&snapshot.action_context, &batch);
+
+    assert_eq!(actions.len(), 1);
+    assert_eq!(actions[0].0.id.to_string(), "good-grep");
+    assert_eq!(errors.len(), 1);
+    assert_eq!(errors[0].call_id, "bad-read");
+    assert_eq!(errors[0].status, galaxy_agent_core::ToolResultStatus::Error);
+    assert!(errors[0].content.contains("expected an array"));
+}
+
+#[test]
+fn malformed_provider_tool_error_can_be_committed_and_run_continues() {
+    let conversation_id = AIConversationId::new();
+    let mut snapshot = provider_snapshot(conversation_id);
+    let Some(ProviderRunStep::CallModel(call)) = snapshot.run.next_step().unwrap() else {
+        panic!("expected provider model call");
+    };
+    snapshot
+        .run
+        .accept_model_turn(
+            &call.work_id,
+            CompletedModelTurn {
+                assistant_content: vec![],
+                tool_calls: vec![ToolCall {
+                    id: "bad-read".to_owned(),
+                    name: "read_files".to_owned(),
+                    arguments: serde_json::json!({}),
+                }],
+                usage: Usage::default(),
+                stop_reason: StopReason::Completed,
+                advertised_tools: BTreeSet::from(["read_files".to_owned()]),
+            },
+        )
+        .unwrap();
+    let Some(ProviderRunStep::DispatchTools(batch)) = snapshot.run.next_step().unwrap() else {
+        panic!("expected provider tool batch");
+    };
+    let (actions, errors) = super::convert_provider_tool_batch(&snapshot.action_context, &batch);
+    assert!(actions.is_empty());
+
+    snapshot
+        .run
+        .complete_tool(&batch.work_id, errors[0].clone())
+        .unwrap();
+    snapshot.run.commit_tool_batch(&batch.work_id).unwrap();
+
+    assert!(matches!(
+        snapshot.run.state(),
+        ProviderRunState::ReadyToCallModel
+    ));
+    let MessageContent::MultiPart(parts) = &snapshot.run.transcript().last().unwrap().content
+    else {
+        panic!("expected correlated tool result");
+    };
+    assert!(matches!(
+        &parts[0],
+        ContentPart::ToolResult {
+            tool_use_id,
+            is_error: true,
+            ..
+        } if tool_use_id == "bad-read"
+    ));
+}
+
 fn attach_snapshot_command_monitor(
     snapshot: &mut super::ActiveProviderRunSnapshot,
     conversation_id: AIConversationId,
@@ -548,7 +921,7 @@ fn restored_active_command_rebuilds_monitor_observation() {
     super::apply_restored_provider_command_evidence(
         conversation_id,
         &mut snapshot,
-        super::RestoredProviderCommandEvidence {
+        Some(super::RestoredProviderCommandEvidence {
             conversation_id: Some(conversation_id),
             requested_command_action_id: Some(action_id),
             cli_task_id: Some(cli_task_id.clone()),
@@ -556,7 +929,7 @@ fn restored_active_command_rebuilds_monitor_observation() {
             state: BlockState::Executing,
             output: "running".to_owned(),
             exit_code: 0,
-        },
+        }),
     )
     .unwrap();
 
@@ -582,7 +955,7 @@ fn restored_completed_command_rebuilds_exact_completion_evidence() {
     super::apply_restored_provider_command_evidence(
         conversation_id,
         &mut snapshot,
-        super::RestoredProviderCommandEvidence {
+        Some(super::RestoredProviderCommandEvidence {
             conversation_id: Some(conversation_id),
             requested_command_action_id: Some(action_id.clone()),
             cli_task_id: Some(cli_task_id),
@@ -590,7 +963,7 @@ fn restored_completed_command_rebuilds_exact_completion_evidence() {
             state: BlockState::DoneWithExecution,
             output: "done".to_owned(),
             exit_code: 17,
-        },
+        }),
     )
     .unwrap();
 
@@ -608,6 +981,78 @@ fn restored_completed_command_rebuilds_exact_completion_evidence() {
     assert_eq!(completion.exit_code, 17);
 }
 
+#[test]
+fn restored_missing_command_block_becomes_interrupted_completion_evidence() {
+    let conversation_id = AIConversationId::new();
+    let mut snapshot = provider_snapshot(conversation_id);
+    let (action_id, block_id, _) = attach_snapshot_command_monitor(&mut snapshot, conversation_id);
+    snapshot.pending_monitor_observation = Some(super::PendingProviderMonitorObservation {
+        block_id: block_id.clone(),
+        cli_task_id: TaskId::new("stale-cli-task".to_owned()),
+    });
+
+    super::apply_restored_provider_command_evidence(conversation_id, &mut snapshot, None).unwrap();
+
+    assert!(snapshot.pending_monitor_observation.is_none());
+    let completion = snapshot
+        .pending_command_completion
+        .expect("missing terminal block should become interrupted-command evidence");
+    assert_eq!(completion.block_id, block_id);
+    assert_eq!(
+        completion.initial_requested_command_action_id,
+        Some(action_id)
+    );
+    assert_eq!(completion.command, "sleep 10");
+    assert_eq!(completion.exit_code, 130);
+    assert!(completion.output.contains("interrupted"));
+}
+
+#[test]
+fn restored_evidence_is_ignored_without_a_command_monitor() {
+    let conversation_id = AIConversationId::new();
+    let mut snapshot = provider_snapshot(conversation_id);
+    let evidence = super::RestoredProviderCommandEvidence {
+        conversation_id: Some(conversation_id),
+        requested_command_action_id: None,
+        cli_task_id: None,
+        command: "sleep 10".to_owned(),
+        state: BlockState::Executing,
+        output: "running".to_owned(),
+        exit_code: 0,
+    };
+
+    super::apply_restored_provider_command_evidence(conversation_id, &mut snapshot, Some(evidence))
+        .unwrap();
+    assert!(snapshot.pending_monitor_observation.is_none());
+    assert!(snapshot.pending_command_completion.is_none());
+}
+
+#[test]
+fn restored_projection_accepts_empty_or_complete_and_rejects_partial_state() {
+    assert_eq!(
+        super::restored_projection_was_initialized(false, false, false).unwrap(),
+        false
+    );
+    assert_eq!(
+        super::restored_projection_was_initialized(true, true, false).unwrap(),
+        true
+    );
+    assert_eq!(
+        super::restored_projection_was_initialized(true, true, true).unwrap(),
+        true
+    );
+    for state in [
+        (false, false, true),
+        (false, true, false),
+        (true, false, false),
+    ] {
+        assert_eq!(
+            super::restored_projection_was_initialized(state.0, state.1, state.2).unwrap_err(),
+            "restored provider projection exchange is partially initialized"
+        );
+    }
+}
+
 #[test]
 fn restored_command_rejects_mismatched_or_invalid_terminal_evidence() {
     let conversation_id = AIConversationId::new();
@@ -619,7 +1064,7 @@ fn restored_command_rejects_mismatched_or_invalid_terminal_evidence() {
         super::apply_restored_provider_command_evidence(
             conversation_id,
             &mut snapshot,
-            super::RestoredProviderCommandEvidence {
+            Some(super::RestoredProviderCommandEvidence {
                 conversation_id: Some(AIConversationId::new()),
                 requested_command_action_id: Some(action_id.clone()),
                 cli_task_id: Some(cli_task_id.clone()),
@@ -627,7 +1072,7 @@ fn restored_command_rejects_mismatched_or_invalid_terminal_evidence() {
                 state: BlockState::Executing,
                 output: String::new(),
                 exit_code: 0,
-            },
+            }),
         )
         .unwrap_err(),
         "restored provider command block identity does not match"
@@ -636,7 +1081,7 @@ fn restored_command_rejects_mismatched_or_invalid_terminal_evidence() {
         super::apply_restored_provider_command_evidence(
             conversation_id,
             &mut snapshot,
-            super::RestoredProviderCommandEvidence {
+            Some(super::RestoredProviderCommandEvidence {
                 conversation_id: Some(conversation_id),
                 requested_command_action_id: Some(action_id),
                 cli_task_id: Some(cli_task_id),
@@ -644,7 +1089,7 @@ fn restored_command_rejects_mismatched_or_invalid_terminal_evidence() {
                 state: BlockState::Background,
                 output: String::new(),
                 exit_code: 0,
-            },
+            }),
         )
         .unwrap_err(),
         "restored provider command block has an invalid state"
@@ -2042,7 +2487,7 @@ fn completed_provider_run_with_prior_action_resolves_child_completion_wait() {
                 "child".to_owned(),
                 parent_conversation_id,
                 child_conversation_id,
-                None,
+                StartAgentWaitPolicy::Completion,
                 ctx,
             )
         });

app/src/ai/blocklist/inline_action/ask_user_question_view.rs

@@ -782,7 +782,11 @@ impl AskUserQuestionView {
         };
 
         ctx.subscribe_to_model(&action_model, |me, _, event, ctx| {
-            if event.action_id() != me.action_id() {
+            if event.action_id() != me.action_id()
+                || event
+                    .conversation_id()
+                    .is_some_and(|conversation_id| conversation_id != me.conversation_id)
+            {
                 return;
             }
 
@@ -879,7 +883,8 @@ impl AskUserQuestionView {
     /// conversations still render deterministically.
     fn action_status(&self, app: &AppContext) -> Option<AIActionStatus> {
         let action_model = self.action_model.as_ref(app);
-        if let Some(status) = action_model.get_action_status(self.action_id()) {
+        if let Some(status) = action_model.get_action_status(self.conversation_id, self.action_id())
+        {
             return Some(status);
         }
 

app/src/ai/blocklist/inline_action/code_diff_view.rs

@@ -695,12 +695,26 @@ impl CodeDiffView {
             session_platform,
             ctx,
         );
+        let action_id = (*action_id).clone();
 
         ctx.subscribe_to_model(
             &action_model,
             move |me, action_model, event, ctx| match event {
-                BlocklistAIActionEvent::FinishedAction { action_id, .. } if !me.is_complete() => {
+                BlocklistAIActionEvent::FinishedAction {
-                    match action_model.as_ref(ctx).get_action_status(&me.action_id) {
+                    action_id: event_action_id,
+                    conversation_id: event_conversation_id,
+                    ..
+                } if !me.is_complete()
+                    && *event_action_id == me.action_id
+                    && me.identifiers.client_conversation_id == Some(*event_conversation_id) =>
+                {
+                    let Some(conversation_id) = me.identifiers.client_conversation_id else {
+                        return;
+                    };
+                    match action_model
+                        .as_ref(ctx)
+                        .get_action_status(conversation_id, &me.action_id)
+                    {
                         Some(AIActionStatus::Blocked) => {
                             me.state = CodeDiffState::WaitingForUser;
                             ctx.notify();

app/src/ai/blocklist/inline_action/requested_command.rs

@@ -412,7 +412,7 @@ impl RequestedCommandView {
 
         let is_finished = action_model
             .as_ref(ctx)
-            .get_action_result(&action_id)
+            .get_action_result(client_ids.conversation_id, &action_id)
             .is_some();
 
         if !is_finished {
@@ -424,16 +424,24 @@ impl RequestedCommandView {
                         ctx.notify();
                     }
                     BlocklistAIActionEvent::ActionBlockedOnUserConfirmation {
-                        action_id, ..
+                        action_id,
-                    } if *action_id == me.action_id => {
+                        conversation_id,
+                        ..
+                    } if *conversation_id == me.client_ids.conversation_id
+                        && *action_id == me.action_id =>
+                    {
                         if me.action_type.is_requested_command() {
                             me.ensure_editor(ctx);
                         }
                         me.set_is_header_expanded(true, ctx);
                         ctx.notify();
                     }
-                    BlocklistAIActionEvent::ExecutingAction { action_id, .. }
+                    BlocklistAIActionEvent::ExecutingAction {
-                        if *action_id == me.action_id =>
+                        action_id,
+                        conversation_id,
+                        ..
+                    } if *conversation_id == me.client_ids.conversation_id
+                        && *action_id == me.action_id =>
                     {
                         // For shared-session viewers, sync the command text from the action when it starts executing.
                         if me.action_model.as_ref(ctx).is_view_only() {
@@ -467,11 +475,15 @@ impl RequestedCommandView {
                         }
                         ctx.notify();
                     }
-                    BlocklistAIActionEvent::FinishedAction { action_id, .. } => {
+                    BlocklistAIActionEvent::FinishedAction {
+                        action_id,
+                        conversation_id,
+                        ..
+                    } if *conversation_id == me.client_ids.conversation_id => {
                         let Some(action_result) = me
                             .action_model
                             .as_ref(ctx)
-                            .get_action_result(action_id)
+                            .get_action_result(me.client_ids.conversation_id, action_id)
                             .cloned()
                         else {
                             log::info!("Got finished action event without result: {action_id}.");
@@ -724,7 +736,7 @@ impl RequestedCommandView {
     fn is_waiting_for_user_confirmation(&self, app: &AppContext) -> bool {
         self.action_model
             .as_ref(app)
-            .get_action_status(&self.action_id)
+            .get_action_status(self.client_ids.conversation_id, &self.action_id)
             .is_some_and(|status| status.is_blocked())
     }
 
@@ -750,7 +762,9 @@ impl RequestedCommandView {
             let Some(mouse_state_handle) =
                 self.citation_state_handles.get(copied_citation).cloned()
             else {
-                log::warn!("Tried to retrieve mouse state handle for citation, but no mouse state handle exists.");
+                log::warn!(
+                    "Tried to retrieve mouse state handle for citation, but no mouse state handle exists."
+                );
                 return None;
             };
             render_citation(
@@ -1108,7 +1122,7 @@ impl RequestedCommandView {
         let action_status = self
             .action_model
             .as_ref(app)
-            .get_action_status(&self.action_id);
+            .get_action_status(self.client_ids.conversation_id, &self.action_id);
 
         let mut title: Cow<'static, str>;
         let mut font_override = None;
@@ -1457,7 +1471,7 @@ impl View for RequestedCommandView {
         let action_status = self
             .action_model
             .as_ref(app)
-            .get_action_status(&self.action_id);
+            .get_action_status(self.client_ids.conversation_id, &self.action_id);
 
         let is_last_output_message_in_output = self
             .block_model
@@ -1635,6 +1649,7 @@ impl View for RequestedCommandView {
         let container = render_tool_pane_shell(
             content.finish(),
             has_highlighted_border,
+            self.is_header_expanded,
             should_remove_bottom_margin,
             app,
         );

app/src/ai/blocklist/inline_action/run_agents_card_view.rs

@@ -489,29 +489,36 @@ impl RunAgentsCardView {
 
         // Re-render when this action finishes or becomes blocked.
         let action_id_for_action_events = action_id.clone();
-        ctx.subscribe_to_model(&action_model, move |me, _, event, ctx| match event {
+        ctx.subscribe_to_model(&action_model, move |me, _, event, ctx| {
-            BlocklistAIActionEvent::FinishedAction { action_id, .. }
+            if event.conversation_id().is_some_and(|conversation_id| {
-                if action_id == &action_id_for_action_events =>
+                me.block_model.conversation_id(ctx) != Some(conversation_id)
-            {
+            }) {
-                ctx.notify();
+                return;
             }
-            BlocklistAIActionEvent::ActionBlockedOnUserConfirmation { action_id, .. }
+            match event {
-                if action_id == &action_id_for_action_events =>
+                BlocklistAIActionEvent::FinishedAction { action_id, .. }
-            {
+                    if action_id == &action_id_for_action_events =>
-                // Normal case: streaming is complete and the action is
+                {
-                // ready for user confirmation. Re-render so the card
+                    ctx.notify();
-                // transitions from the "Configuring agents..." placeholder
-                // to the full confirmation UI.
-                resolve_interactive_defaults(&mut me.state, &*me.block_model, ctx);
-                oc::repopulate_all_pickers(&mut me.state.orch, &me.handles.pickers, ctx);
-                me.refresh_accept_button_state(ctx);
-                me.maybe_auto_open_create_modal(ctx);
-                if let Some(conversation_id) = me.block_model.conversation_id(ctx) {
-                    me.emit_orchestration_entered_once(conversation_id, ctx);
                 }
-                ctx.notify();
+                BlocklistAIActionEvent::ActionBlockedOnUserConfirmation { action_id, .. }
+                    if action_id == &action_id_for_action_events =>
+                {
+                    // Normal case: streaming is complete and the action is
+                    // ready for user confirmation. Re-render so the card
+                    // transitions from the "Configuring agents..." placeholder
+                    // to the full confirmation UI.
+                    resolve_interactive_defaults(&mut me.state, &*me.block_model, ctx);
+                    oc::repopulate_all_pickers(&mut me.state.orch, &me.handles.pickers, ctx);
+                    me.refresh_accept_button_state(ctx);
+                    me.maybe_auto_open_create_modal(ctx);
+                    if let Some(conversation_id) = me.block_model.conversation_id(ctx) {
+                        me.emit_orchestration_entered_once(conversation_id, ctx);
+                    }
+                    ctx.notify();
+                }
+                _ => {}
             }
-            _ => {}
         });
 
         // Repopulate the model picker when available Warp LLMs change.
@@ -713,8 +720,11 @@ impl RunAgentsCardView {
         let request = self.state.to_request();
         self.emit_decision(RunAgentsCardDecision::Accept, ctx);
         let action_id = self.action_id.clone();
+        let Some(conversation_id) = self.block_model.conversation_id(ctx) else {
+            return;
+        };
         self.action_model.update(ctx, |action_model, action_ctx| {
-            action_model.execute_run_agents(&action_id, request, action_ctx);
+            action_model.execute_run_agents(conversation_id, &action_id, request, action_ctx);
         });
     }
 
@@ -806,10 +816,13 @@ impl RunAgentsCardView {
         if self.block_model.is_restored() {
             return;
         }
+        let Some(conversation_id) = self.block_model.conversation_id(ctx) else {
+            return;
+        };
         if matches!(
             self.action_model
                 .as_ref(ctx)
-                .get_action_status(&self.action_id),
+                .get_action_status(conversation_id, &self.action_id),
             Some(AIActionStatus::Finished(_)) | Some(AIActionStatus::RunningAsync)
         ) {
             return;
@@ -1093,9 +1106,13 @@ impl View for RunAgentsCardView {
     fn render(&self, app: &AppContext) -> Box<dyn Element> {
         let appearance = Appearance::as_ref(app);
         let status = self
-            .action_model
+            .block_model
-            .as_ref(app)
+            .conversation_id(app)
-            .get_action_status(&self.action_id);
+            .and_then(|conversation_id| {
+                self.action_model
+                    .as_ref(app)
+                    .get_action_status(conversation_id, &self.action_id)
+            });
 
         if let Some(AIActionStatus::Finished(result)) = &status {
             if let AIAgentActionResultType::RunAgents(orchestrate_result) = &result.result {
@@ -1208,8 +1225,16 @@ impl TypedActionView for RunAgentsCardView {
             RunAgentsCardViewAction::AcceptWithoutOrchestration => {
                 self.emit_decision(RunAgentsCardDecision::AcceptWithoutOrchestration, ctx);
                 let action_id = self.action_id.clone();
+                let Some(conversation_id) = self.block_model.conversation_id(ctx) else {
+                    return;
+                };
                 self.action_model.update(ctx, |action_model, action_ctx| {
-                    action_model.deny_run_agents(&action_id, String::new(), action_ctx);
+                    action_model.deny_run_agents(
+                        conversation_id,
+                        &action_id,
+                        String::new(),
+                        action_ctx,
+                    );
                 });
             }
             RunAgentsCardViewAction::ToggleAcceptMenu => {
@@ -1537,14 +1562,31 @@ pub(crate) fn format_terminal_state(result: &RunAgentsResult) -> (String, Status
                 .iter()
                 .filter(|a| matches!(a.kind, RunAgentsAgentOutcomeKind::Launched { .. }))
                 .count();
-            if launched == total {
+            let completed = agents
+                .iter()
+                .filter(|a| matches!(a.kind, RunAgentsAgentOutcomeKind::Completed { .. }))
+                .count();
+            let successful = launched + completed;
+            if completed > 0 && completed == total {
+                let label = if total == 1 {
+                    "Completed 1 agent".to_string()
+                } else {
+                    format!("Completed {total} agents")
+                };
+                (label, StatusKind::Success)
+            } else if launched == 0 && completed > 0 {
+                (
+                    format!("Completed {completed} of {total} agents"),
+                    StatusKind::Mixed,
+                )
+            } else if successful == total {
                 let label = if total == 1 {
                     "Spawned 1 agent".to_string()
                 } else {
                     format!("Spawned {total} agents")
                 };
                 (label, StatusKind::Success)
-            } else if launched == 0 {
+            } else if successful == 0 {
                 // Every child failed to launch: surface a terminal failure
                 // rather than the in-progress-looking mixed state.
                 let label = if total == 1 {
@@ -1555,7 +1597,7 @@ pub(crate) fn format_terminal_state(result: &RunAgentsResult) -> (String, Status
                 (label, StatusKind::Failure)
             } else {
                 (
-                    format!("Spawned {launched} of {total} agents"),
+                    format!("Spawned {successful} of {total} agents"),
                     StatusKind::Mixed,
                 )
             }
@@ -1713,7 +1755,8 @@ fn render_run_agents_child_row(
     app: &AppContext,
 ) -> Box<dyn Element> {
     let outcome_conversation_id = outcome.and_then(|outcome| match &outcome.kind {
-        RunAgentsAgentOutcomeKind::Launched { agent_id } => {
+        RunAgentsAgentOutcomeKind::Launched { agent_id }
+        | RunAgentsAgentOutcomeKind::Completed { agent_id, .. } => {
             conversation_id_for_agent_id(agent_id, app)
         }
         RunAgentsAgentOutcomeKind::Failed { .. } => None,
@@ -1757,6 +1800,9 @@ fn render_run_agents_child_row(
             RunAgentsAgentOutcomeKind::Launched { .. } => {
                 (ConversationStatus::Success, "Started".to_string())
             }
+            RunAgentsAgentOutcomeKind::Completed { .. } => {
+                (ConversationStatus::Success, "Completed".to_string())
+            }
             RunAgentsAgentOutcomeKind::Failed { error } => (
                 ConversationStatus::Error,
                 if error.trim().is_empty() {

app/src/ai/blocklist/inline_action/run_agents_card_view_tests.rs

@@ -327,6 +327,16 @@ mod format_terminal_state_tests {
         }
     }
 
+    fn completed(name: &str, agent_id: &str) -> RunAgentsAgentOutcome {
+        RunAgentsAgentOutcome {
+            name: name.to_string(),
+            kind: RunAgentsAgentOutcomeKind::Completed {
+                agent_id: agent_id.to_string(),
+                output: format!("{name} output"),
+            },
+        }
+    }
+
     fn launched_result(agents: Vec<RunAgentsAgentOutcome>) -> RunAgentsResult {
         RunAgentsResult::Launched {
             model_id: "auto".to_string(),
@@ -368,6 +378,30 @@ mod format_terminal_state_tests {
         assert!(matches!(kind, StatusKind::Mixed));
     }
 
+    #[test]
+    fn all_completed_uses_completed_label_and_success_status() {
+        let result = launched_result(vec![
+            completed("a", "a-1"),
+            completed("b", "a-2"),
+            completed("c", "a-3"),
+        ]);
+        let (label, kind) = format_terminal_state(&result);
+        assert_eq!(label, "Completed 3 agents");
+        assert!(matches!(kind, StatusKind::Success));
+    }
+
+    #[test]
+    fn mixed_completed_and_failed_uses_completed_label_and_mixed_status() {
+        let result = launched_result(vec![
+            completed("a", "a-1"),
+            failed("b", "boom"),
+            completed("c", "a-3"),
+        ]);
+        let (label, kind) = format_terminal_state(&result);
+        assert_eq!(label, "Completed 2 of 3 agents");
+        assert!(matches!(kind, StatusKind::Mixed));
+    }
+
     #[test]
     fn all_failed_uses_failure_status_not_mixed() {
         let result = launched_result(vec![

app/src/ai/blocklist/inline_action/tool_pane.rs

@@ -14,6 +14,7 @@ use crate::ai::blocklist::block::view_impl::{
 pub(crate) fn render_tool_pane_shell(
     content: Box<dyn Element>,
     has_highlighted_border: bool,
+    spans_conversation_width: bool,
     should_remove_bottom_margin: bool,
     app: &AppContext,
 ) -> Box<dyn Element> {
@@ -25,7 +26,7 @@ pub(crate) fn render_tool_pane_shell(
     };
 
     Container::new(content)
-        .with_margin_left(if has_highlighted_border {
+        .with_margin_left(if has_highlighted_border || spans_conversation_width {
             CONTENT_HORIZONTAL_PADDING
         } else {
             CONTENT_HORIZONTAL_PADDING + icon_size(app) + 16.

app/src/ai/mod.rs

@@ -61,6 +61,7 @@ pub mod request_usage_model;
 pub(crate) mod restored_conversations;
 pub(crate) mod runtime;
 pub(crate) mod skills;
+pub(crate) mod tool_diagnostics;
 pub(crate) mod voice;
 pub use agent_tips::*;
 use galaxyui::AppContext;

app/src/ai/runtime/event_translator.rs

@@ -62,9 +62,12 @@ impl ProviderRunResponseProjector {
         }
     }
 
-    pub(crate) fn restored(config: RuntimeResponseConfig) -> Self {
+    pub(crate) fn restored(
+        config: RuntimeResponseConfig,
+        projection_was_initialized: bool,
+    ) -> Self {
         Self {
-            translator: RuntimeResponseTranslator::restored(config),
+            translator: RuntimeResponseTranslator::restored(config, projection_was_initialized),
             has_started_model_turn: false,
             finished: false,
         }
@@ -88,9 +91,11 @@ impl ProviderRunResponseProjector {
                 })
             }
             ProviderRunProjection::ModelEvent { event, .. } => self.translator.translate(event),
+            ProviderRunProjection::ModelRetry { .. } => {
+                Ok(self.translator.discard_failed_turn_output())
+            }
             ProviderRunProjection::ModelTurnRequested { .. }
             | ProviderRunProjection::ModelTurnFinished { .. }
-            | ProviderRunProjection::ModelRetry { .. }
             | ProviderRunProjection::ToolBatchReady { .. } => Ok(Vec::new()),
         }
     }
@@ -130,8 +135,15 @@ impl RuntimeResponseTranslator {
         Self::with_initialization(config, false)
     }
 
-    pub(crate) fn restored(config: RuntimeResponseConfig) -> Self {
+    pub(crate) fn restored(
-        Self::with_initialization(config, true)
+        mut config: RuntimeResponseConfig,
+        projection_was_initialized: bool,
+    ) -> Self {
+        // The task and exchange already exist in restored history. If its output was never
+        // initialized, replay only the stream Init rather than duplicating task/input messages.
+        config.needs_create_task = false;
+        config.user_query = None;
+        Self::with_initialization(config, projection_was_initialized)
     }
 
     fn with_initialization(config: RuntimeResponseConfig, initialized: bool) -> Self {
@@ -143,7 +155,7 @@ impl RuntimeResponseTranslator {
             reasoning_message_id: None,
             activity_message_ids: HashMap::new(),
             activities: HashMap::new(),
-            has_visible_output: initialized,
+            has_visible_output: false,
             usage: Usage::default(),
             context_usage: None,
         }
@@ -163,9 +175,7 @@ impl RuntimeResponseTranslator {
             }
             AgentEvent::ReasoningCompleted { text, .. } => {
                 self.initialize(&mut events);
-                if self.reasoning_message_id.is_none() && !text.is_empty() {
+                self.complete_reasoning(&text, &mut events);
-                    self.add_or_append_reasoning(&text, &mut events);
-                }
             }
             AgentEvent::RuntimeActivityUpdated { activity } => {
                 if self.config.capabilities.host_tool_execution {
@@ -244,6 +254,25 @@ impl RuntimeResponseTranslator {
         self.reasoning_message_id = None;
     }
 
+    fn discard_failed_turn_output(&mut self) -> Vec<ResponseEvent> {
+        let mut events = Vec::new();
+        if let Some(message_id) = self.text_message_id.take() {
+            events.push(build_replace_text_message(
+                &self.config.task_id,
+                &message_id,
+                "",
+            ));
+        }
+        if let Some(message_id) = self.reasoning_message_id.take() {
+            events.push(build_replace_reasoning_message(
+                &self.config.task_id,
+                &message_id,
+                "",
+            ));
+        }
+        events
+    }
+
     pub(crate) fn set_task_id(&mut self, task_id: impl Into<String>) {
         let task_id = task_id.into();
         if self.config.task_id == task_id {
@@ -315,6 +344,18 @@ impl RuntimeResponseTranslator {
         }
     }
 
+    fn complete_reasoning(&mut self, text: &str, events: &mut Vec<ResponseEvent>) {
+        if let Some(message_id) = &self.reasoning_message_id {
+            events.push(build_replace_reasoning_message(
+                &self.config.task_id,
+                message_id,
+                text,
+            ));
+        } else if !text.is_empty() {
+            self.add_or_append_reasoning(text, events);
+        }
+    }
+
     fn upsert_runtime_activity(
         &mut self,
         activity: RuntimeActivity,
@@ -441,6 +482,58 @@ fn build_reasoning_message(
     runtime_client_action(action)
 }
 
+fn build_replace_reasoning_message(task_id: &str, message_id: &str, text: &str) -> ResponseEvent {
+    let message = api::Message {
+        id: message_id.to_owned(),
+        task_id: task_id.to_owned(),
+        request_id: String::new(),
+        timestamp: None,
+        server_message_data: String::new(),
+        citations: Vec::new(),
+        fetched_memories: Vec::new(),
+        message: Some(api::message::Message::AgentReasoning(
+            api::message::AgentReasoning {
+                reasoning: text.to_owned(),
+                finished_duration: None,
+            },
+        )),
+    };
+    runtime_client_action(api::client_action::Action::UpdateTaskMessage(
+        api::client_action::UpdateTaskMessage {
+            task_id: task_id.to_owned(),
+            message: Some(message),
+            mask: Some(prost_types::FieldMask {
+                paths: vec!["agent_reasoning.reasoning".to_owned()],
+            }),
+        },
+    ))
+}
+
+fn build_replace_text_message(task_id: &str, message_id: &str, text: &str) -> ResponseEvent {
+    runtime_client_action(api::client_action::Action::UpdateTaskMessage(
+        api::client_action::UpdateTaskMessage {
+            task_id: task_id.to_owned(),
+            message: Some(api::Message {
+                id: message_id.to_owned(),
+                task_id: task_id.to_owned(),
+                request_id: String::new(),
+                timestamp: None,
+                server_message_data: String::new(),
+                citations: Vec::new(),
+                fetched_memories: Vec::new(),
+                message: Some(api::message::Message::AgentOutput(
+                    api::message::AgentOutput {
+                        text: text.to_owned(),
+                    },
+                )),
+            }),
+            mask: Some(prost_types::FieldMask {
+                paths: vec!["agent_output.text".to_owned()],
+            }),
+        },
+    ))
+}
+
 fn runtime_activity_fallback_text(activity: &RuntimeActivity) -> String {
     let title = &activity.title;
     let status = activity.status.as_ref().map(|status| match status {

app/src/ai/runtime/event_translator_tests.rs

@@ -45,7 +45,7 @@ fn restored_provider_projection_skips_stream_initialization() {
         capabilities: RuntimeCapabilities::provider(),
         empty_output_message: None,
     };
-    let mut projector = ProviderRunResponseProjector::restored(config);
+    let mut projector = ProviderRunResponseProjector::restored(config, true);
     let work_id = galaxy_agent_core::ExternalWorkId {
         run_id: galaxy_agent_core::ProviderRunId::new("run"),
         epoch: galaxy_agent_core::RunEpoch::new(2),
@@ -82,6 +82,56 @@ fn restored_provider_projection_skips_stream_initialization() {
     ));
 }
 
+#[test]
+fn restored_uninitialized_projection_replays_init_before_live_delta() {
+    let config = RuntimeResponseConfig {
+        task_id: "task".to_owned(),
+        conversation_id: "conversation".to_owned(),
+        needs_create_task: true,
+        user_query: Some("already persisted".to_owned()),
+        model_id: "model".to_owned(),
+        max_context_tokens: Some(1_000),
+        capabilities: RuntimeCapabilities::provider(),
+        empty_output_message: None,
+    };
+    let mut projector = ProviderRunResponseProjector::restored(config, false);
+    let work_id = galaxy_agent_core::ExternalWorkId {
+        run_id: galaxy_agent_core::ProviderRunId::new("run"),
+        epoch: galaxy_agent_core::RunEpoch::new(2),
+    };
+
+    let started = projector
+        .project(ProviderRunProjection::ModelTurnStarted {
+            work_id: work_id.clone(),
+            profile: galaxy_agent_core::ProviderRequestProfile::new("base"),
+            runtime_id: "runtime".to_owned(),
+            model_id: "model".to_owned(),
+            runtime_request_id: "request".to_owned(),
+            retry_attempt: 0,
+            elapsed_ms: 1,
+        })
+        .unwrap();
+    let delta = projector
+        .project(ProviderRunProjection::ModelEvent {
+            work_id,
+            event: AgentEvent::TextDelta {
+                text: "continued".to_owned(),
+            },
+        })
+        .unwrap();
+
+    assert_eq!(started.len(), 1);
+    assert!(matches!(
+        started[0].r#type,
+        Some(response_event::Type::Init(_))
+    ));
+    assert_eq!(delta.len(), 1);
+    assert!(matches!(
+        delta[0].r#type,
+        Some(response_event::Type::ClientActions(_))
+    ));
+}
+
 #[test]
 fn provider_followup_turn_starts_a_distinct_text_message() {
     let mut projector = ProviderRunResponseProjector::new(RuntimeResponseConfig {
@@ -261,6 +311,127 @@ fn reasoning_uses_the_native_reasoning_message_contract() {
     ));
 }
 
+#[test]
+fn reasoning_completed_authoritatively_replaces_streamed_reasoning() {
+    let mut translator = provider_translator();
+    let streamed = translator
+        .translate(AgentEvent::ReasoningDelta {
+            text: "draft reasoning".to_owned(),
+        })
+        .expect("reasoning delta");
+    let completed = translator
+        .translate(AgentEvent::ReasoningCompleted {
+            text: "authoritative reasoning".to_owned(),
+            signature: Some("signature".to_owned()),
+        })
+        .expect("reasoning completion");
+
+    let Some(response_event::Type::ClientActions(streamed_actions)) = &streamed[1].r#type else {
+        panic!("expected streamed reasoning action");
+    };
+    let Some(client_action::Action::AddMessagesToTask(add)) = &streamed_actions.actions[0].action
+    else {
+        panic!("expected streamed reasoning message");
+    };
+    let message_id = add.messages[0].id.clone();
+    let Some(response_event::Type::ClientActions(completed_actions)) = &completed[0].r#type else {
+        panic!("expected completed reasoning action");
+    };
+    let Some(client_action::Action::UpdateTaskMessage(update)) =
+        &completed_actions.actions[0].action
+    else {
+        panic!("completed reasoning must replace the streamed value");
+    };
+    let message = update.message.as_ref().expect("replacement message");
+    assert_eq!(message.id, message_id);
+    assert!(matches!(
+        &message.message,
+        Some(message::Message::AgentReasoning(reasoning))
+            if reasoning.reasoning == "authoritative reasoning"
+    ));
+    assert_eq!(
+        update.mask.as_ref().expect("replacement mask").paths,
+        ["agent_reasoning.reasoning"]
+    );
+}
+
+#[test]
+fn provider_retry_clears_failed_attempt_output_before_new_messages() {
+    let mut projector = ProviderRunResponseProjector::new(RuntimeResponseConfig {
+        task_id: "task".to_owned(),
+        conversation_id: "conversation".to_owned(),
+        needs_create_task: false,
+        user_query: None,
+        model_id: "model".to_owned(),
+        max_context_tokens: Some(1_000),
+        capabilities: RuntimeCapabilities::provider(),
+        empty_output_message: None,
+    });
+    let work_id = galaxy_agent_core::ExternalWorkId {
+        run_id: galaxy_agent_core::ProviderRunId::new("run"),
+        epoch: galaxy_agent_core::RunEpoch::new(1),
+    };
+    projector
+        .project(ProviderRunProjection::ModelTurnStarted {
+            work_id: work_id.clone(),
+            profile: galaxy_agent_core::ProviderRequestProfile::new("base"),
+            runtime_id: "runtime".to_owned(),
+            model_id: "model".to_owned(),
+            runtime_request_id: "request".to_owned(),
+            retry_attempt: 0,
+            elapsed_ms: 1,
+        })
+        .unwrap();
+    for event in [
+        AgentEvent::TextDelta {
+            text: "failed text".to_owned(),
+        },
+        AgentEvent::ReasoningDelta {
+            text: "failed reasoning".to_owned(),
+        },
+    ] {
+        projector
+            .project(ProviderRunProjection::ModelEvent {
+                work_id: work_id.clone(),
+                event,
+            })
+            .unwrap();
+    }
+    let retry = projector
+        .project(ProviderRunProjection::ModelRetry {
+            work_id,
+            profile: galaxy_agent_core::ProviderRequestProfile::new("base"),
+            runtime_id: "runtime".to_owned(),
+            model_id: "model".to_owned(),
+            retry_attempt: 1,
+            elapsed_ms: 2,
+            error: galaxy_agent_core::AgentError::new(
+                galaxy_agent_core::AgentErrorKind::Transport,
+                "retry",
+            ),
+        })
+        .unwrap();
+
+    assert_eq!(retry.len(), 2);
+    for event in retry {
+        let Some(response_event::Type::ClientActions(actions)) = event.r#type else {
+            panic!("cleanup must use a client action");
+        };
+        let Some(client_action::Action::UpdateTaskMessage(update)) = &actions.actions[0].action
+        else {
+            panic!("cleanup must replace failed output");
+        };
+        let message = update.message.as_ref().expect("cleanup message");
+        match &message.message {
+            Some(message::Message::AgentOutput(output)) => assert!(output.text.is_empty()),
+            Some(message::Message::AgentReasoning(reasoning)) => {
+                assert!(reasoning.reasoning.is_empty())
+            }
+            _ => panic!("cleanup must target visible text or reasoning"),
+        }
+    }
+}
+
 #[test]
 fn session_activity_updates_the_same_structured_message() {
     let mut translator = session_translator();

app/src/ai/runtime/provider_run_coordinator.rs

@@ -5,16 +5,17 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use futures::future::BoxFuture;
-use futures::StreamExt;
+use futures::{FutureExt as _, StreamExt};
 use galaxy_agent_core::{
-    AgentError, AgentErrorKind, AgentEvent, AgentRuntime, CompletedModelTurn, ContentPart,
+    turn_control, AgentError, AgentErrorKind, AgentEvent, AgentRuntime, CompletedModelTurn,
-    ExternalWorkId, ModelFailureDisposition, PendingToolBatch, ProviderModelCall,
+    ContentPart, ExternalWorkId, ModelFailureDisposition, PendingToolBatch, ProviderModelCall,
     ProviderRequestProfile, ProviderRun, ProviderRunFailureKind, ProviderRunId, ProviderRunLimits,
     ProviderRunOutcome, ProviderRunPhase, ProviderRunProtocolError, ProviderRunState,
-    ProviderRunStep, RunEpoch, RuntimeKind, StopReason, ToolEvent, TurnControl, TurnRequest, Usage,
+    ProviderRunStep, RunEpoch, RuntimeKind, StopReason, ToolEvent, TurnCommand, TurnControl,
+    TurnRequest, Usage,
 };
 use instant::Instant;
-use warpui::r#async::FutureExt as _;
+use warpui::r#async::Timer;
 
 use crate::ai::agent::conversation::AIConversationId;
 
@@ -321,31 +322,55 @@ impl ProviderRunCoordinator {
         &mut self,
         control: TurnControl,
         mut project: F,
-        mut checkpoint: C,
+        checkpoint: C,
     ) -> Result<ProviderRunBlock, ProviderRunCoordinatorError>
     where
         F: FnMut(ProviderRunProjection) -> Result<(), String>,
         C: FnMut(ProviderRun) -> BoxFuture<'static, Result<(), String>>,
+    {
+        self.drive_until_blocked_with_acknowledgements(
+            control,
+            move |projection| {
+                let result = project(projection);
+                Box::pin(async move { result })
+            },
+            checkpoint,
+        )
+        .await
+    }
+
+    pub(crate) async fn drive_until_blocked_with_acknowledgements<F, C>(
+        &mut self,
+        control: TurnControl,
+        mut project: F,
+        mut checkpoint: C,
+    ) -> Result<ProviderRunBlock, ProviderRunCoordinatorError>
+    where
+        F: FnMut(ProviderRunProjection) -> BoxFuture<'static, Result<(), String>>,
+        C: FnMut(ProviderRun) -> BoxFuture<'static, Result<(), String>>,
     {
         loop {
             match self.run.next_step()? {
                 Some(ProviderRunStep::CallModel(call)) => {
                     if !self.checkpoint_or_fail(&mut checkpoint).await? {
-                        continue;
+                        return self.terminal_block();
                     }
-                    self.drive_model_call(call, control.clone(), &mut project)
+                    self.drive_model_call_acknowledged(call, control.clone(), &mut project)
                         .await?;
                 }
                 Some(ProviderRunStep::DispatchTools(batch)) => {
                     if !self.checkpoint_or_fail(&mut checkpoint).await? {
-                        continue;
+                        return self.terminal_block();
                     }
-                    if !self.project_or_fail(
+                    if !self
-                        ProviderRunProjection::ToolBatchReady {
+                        .project_or_fail_acknowledged(
-                            batch: batch.clone(),
+                            ProviderRunProjection::ToolBatchReady {
-                        },
+                                batch: batch.clone(),
-                        &mut project,
+                            },
-                    )? {
+                            &mut project,
+                        )
+                        .await?
+                    {
                         continue;
                     }
                     if batch.is_complete() {
@@ -356,7 +381,7 @@ impl ProviderRunCoordinator {
                 }
                 Some(ProviderRunStep::Done(outcome)) => {
                     if !self.checkpoint_or_fail(&mut checkpoint).await? {
-                        continue;
+                        return self.terminal_block();
                     }
                     return Ok(ProviderRunBlock::Done(outcome));
                 }
@@ -368,7 +393,7 @@ impl ProviderRunCoordinator {
                         let work_id = work_id.clone();
                         let stop_reason = stop_reason.clone();
                         if !self.checkpoint_or_fail(&mut checkpoint).await? {
-                            continue;
+                            return self.terminal_block();
                         }
                         return Ok(ProviderRunBlock::AwaitingDriver {
                             work_id,
@@ -413,14 +438,26 @@ impl ProviderRunCoordinator {
         }
     }
 
-    async fn drive_model_call<F>(
+    fn terminal_block(&mut self) -> Result<ProviderRunBlock, ProviderRunCoordinatorError> {
+        match self.run.next_step()? {
+            Some(ProviderRunStep::Done(outcome)) => Ok(ProviderRunBlock::Done(outcome)),
+            Some(ProviderRunStep::CallModel(_) | ProviderRunStep::DispatchTools(_)) | None => Err(
+                ProviderRunCoordinatorError::Core(ProviderRunProtocolError::UnexpectedState {
+                    expected: ProviderRunPhase::Failed,
+                    actual: self.run.state().phase(),
+                }),
+            ),
+        }
+    }
+
+    async fn drive_model_call_acknowledged<F>(
         &mut self,
         call: ProviderModelCall,
         control: TurnControl,
         project: &mut F,
     ) -> Result<(), ProviderRunCoordinatorError>
     where
-        F: FnMut(ProviderRunProjection) -> Result<(), String>,
+        F: FnMut(ProviderRunProjection) -> BoxFuture<'static, Result<(), String>>,
     {
         let Some(profile) = self.profiles.get(call.profile.as_str()).cloned() else {
             self.run.fail(
@@ -440,39 +477,66 @@ impl ProviderRunCoordinator {
             .collect::<BTreeSet<_>>();
         let runtime_id = profile.runtime.descriptor().id.clone();
         let model_id = profile.request.model.as_str().to_string();
-        if !self.project_or_fail(
+        if !self
-            ProviderRunProjection::ModelTurnRequested {
+            .project_or_fail_acknowledged(
-                work_id: call.work_id.clone(),
+                ProviderRunProjection::ModelTurnRequested {
-                profile: call.profile.clone(),
+                    work_id: call.work_id.clone(),
-                runtime_id: runtime_id.clone(),
+                    profile: call.profile.clone(),
-                model_id: model_id.clone(),
+                    runtime_id: runtime_id.clone(),
-                retry_attempt: call.retry_attempt,
+                    model_id: model_id.clone(),
-            },
+                    retry_attempt: call.retry_attempt,
-            project,
+                },
-        )? {
+                project,
+            )
+            .await?
+        {
             return Ok(());
         }
         let request = request_for_model_call(profile.request.clone(), &call);
         let started_at = Instant::now();
-        let stream = match profile
+        let (attempt_sender, attempt_control) = turn_control();
-            .runtime
+        let start_future = profile.runtime.start_turn(request, attempt_control).fuse();
-            .start_turn(request, control)
+        let timeout = futures::FutureExt::fuse(Timer::after(self.model_start_timeout));
-            .with_timeout(self.model_start_timeout)
+        futures::pin_mut!(start_future, timeout);
-            .await
+        let mut control_open = true;
-        {
+        let start_result = loop {
-            Ok(Ok(stream)) => stream,
+            let command = if control_open {
-            Ok(Err(error)) => {
+                futures::future::Either::Left(control.receive())
-                self.handle_model_failure(&call, &profile, started_at, error, project)?;
+            } else {
+                futures::future::Either::Right(futures::future::pending())
+            }
+            .fuse();
+            futures::pin_mut!(command);
+            futures::select_biased! {
+                result = start_future => break Some(result),
+                _ = timeout => {
+                    let _ = attempt_sender.send(TurnCommand::Cancel).await;
+                    break None;
+                }
+                command = command => match command {
+                    Ok(command) => {
+                        let _ = attempt_sender.send(command).await;
+                    }
+                    Err(_) => control_open = false,
+                },
+            }
+        };
+        let stream = match start_result {
+            Some(Ok(stream)) => stream,
+            Some(Err(error)) => {
+                self.handle_model_failure_acknowledged(&call, &profile, started_at, error, project)
+                    .await?;
                 return Ok(());
             }
-            Err(_) => {
+            None => {
-                self.handle_model_failure(
+                self.handle_model_failure_acknowledged(
                     &call,
                     &profile,
                     started_at,
                     provider_timeout_error("start", self.model_start_timeout),
                     project,
-                )?;
+                )
+                .await?;
                 return Ok(());
             }
         };
@@ -480,157 +544,230 @@ impl ProviderRunCoordinator {
         let mut buffer = ModelTurnBuffer::default();
 
         loop {
-            let event = match stream
+            let next_event = stream.next().fuse();
-                .next()
+            let timeout = futures::FutureExt::fuse(Timer::after(self.model_event_idle_timeout));
-                .with_timeout(self.model_event_idle_timeout)
+            futures::pin_mut!(next_event, timeout);
-                .await
+            let event_result = loop {
-            {
+                let command = if control_open {
-                Ok(Some(Ok(event))) => event,
+                    futures::future::Either::Left(control.receive())
-                Ok(Some(Err(error))) => {
+                } else {
-                    self.handle_model_failure(&call, &profile, started_at, error, project)?;
+                    futures::future::Either::Right(futures::future::pending())
+                }
+                .fuse();
+                futures::pin_mut!(command);
+                futures::select_biased! {
+                    event = next_event => break Some(event),
+                    _ = timeout => {
+                        let _ = attempt_sender.send(TurnCommand::Cancel).await;
+                        break None;
+                    }
+                    command = command => match command {
+                        Ok(command) => {
+                            let _ = attempt_sender.send(command).await;
+                        }
+                        Err(_) => control_open = false,
+                    },
+                }
+            };
+            let event = match event_result {
+                Some(Some(Ok(event))) => event,
+                Some(Some(Err(error))) => {
+                    self.handle_model_failure_acknowledged(
+                        &call, &profile, started_at, error, project,
+                    )
+                    .await?;
                     return Ok(());
                 }
-                Ok(None) => break,
+                Some(None) => break,
-                Err(_) => {
+                None => {
-                    self.handle_model_failure(
+                    self.handle_model_failure_acknowledged(
                         &call,
                         &profile,
                         started_at,
                         provider_timeout_error("event", self.model_event_idle_timeout),
                         project,
-                    )?;
+                    )
+                    .await?;
                     return Ok(());
                 }
             };
             match event {
                 AgentEvent::TurnStarted { runtime_request_id } => {
                     if buffer.started {
-                        self.handle_model_failure(
+                        self.handle_model_failure_acknowledged(
                             &call,
                             &profile,
                             started_at,
                             protocol_error("provider emitted more than one TurnStarted event"),
                             project,
-                        )?;
+                        )
+                        .await?;
                         return Ok(());
                     }
                     if runtime_request_id.is_empty() {
-                        self.handle_model_failure(
+                        self.handle_model_failure_acknowledged(
                             &call,
                             &profile,
                             started_at,
                             protocol_error("provider emitted an empty runtime request ID"),
                             project,
-                        )?;
+                        )
+                        .await?;
                         return Ok(());
                     }
                     buffer.started = true;
-                    if !self.project_or_fail(
+                    if !self
-                        ProviderRunProjection::ModelTurnStarted {
+                        .project_or_fail_acknowledged(
-                            work_id: call.work_id.clone(),
+                            ProviderRunProjection::ModelTurnStarted {
-                            profile: call.profile.clone(),
+                                work_id: call.work_id.clone(),
-                            runtime_id: runtime_id.clone(),
+                                profile: call.profile.clone(),
-                            model_id: model_id.clone(),
+                                runtime_id: runtime_id.clone(),
-                            runtime_request_id,
+                                model_id: model_id.clone(),
-                            retry_attempt: call.retry_attempt,
+                                runtime_request_id,
-                            elapsed_ms: elapsed_millis(started_at),
+                                retry_attempt: call.retry_attempt,
-                        },
+                                elapsed_ms: elapsed_millis(started_at),
-                        project,
+                            },
-                    )? {
+                            project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                 }
                 AgentEvent::TextDelta { text } => {
-                    if !self.ensure_model_started(&call, &profile, started_at, &buffer, project)? {
+                    if !self
+                        .ensure_model_started_acknowledged(
+                            &call, &profile, started_at, &buffer, project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                     buffer.text.push_str(&text);
-                    if !self.project_or_fail(
+                    if !self
-                        ProviderRunProjection::ModelEvent {
+                        .project_or_fail_acknowledged(
-                            work_id: call.work_id.clone(),
+                            ProviderRunProjection::ModelEvent {
-                            event: AgentEvent::TextDelta { text },
+                                work_id: call.work_id.clone(),
-                        },
+                                event: AgentEvent::TextDelta { text },
-                        project,
+                            },
-                    )? {
+                            project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                 }
                 AgentEvent::ReasoningDelta { text } => {
-                    if !self.ensure_model_started(&call, &profile, started_at, &buffer, project)? {
+                    if !self
+                        .ensure_model_started_acknowledged(
+                            &call, &profile, started_at, &buffer, project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                     buffer.reasoning.push_str(&text);
-                    if !self.project_or_fail(
+                    if !self
-                        ProviderRunProjection::ModelEvent {
+                        .project_or_fail_acknowledged(
-                            work_id: call.work_id.clone(),
+                            ProviderRunProjection::ModelEvent {
-                            event: AgentEvent::ReasoningDelta { text },
+                                work_id: call.work_id.clone(),
-                        },
+                                event: AgentEvent::ReasoningDelta { text },
-                        project,
+                            },
-                    )? {
+                            project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                 }
                 AgentEvent::ReasoningCompleted { text, signature } => {
-                    if !self.ensure_model_started(&call, &profile, started_at, &buffer, project)? {
+                    if !self
+                        .ensure_model_started_acknowledged(
+                            &call, &profile, started_at, &buffer, project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
-                    if !text.is_empty() {
+                    buffer.reasoning.clone_from(&text);
-                        buffer.reasoning.clone_from(&text);
-                    }
                     buffer.reasoning_signature.clone_from(&signature);
-                    if !self.project_or_fail(
+                    if !self
-                        ProviderRunProjection::ModelEvent {
+                        .project_or_fail_acknowledged(
-                            work_id: call.work_id.clone(),
+                            ProviderRunProjection::ModelEvent {
-                            event: AgentEvent::ReasoningCompleted { text, signature },
+                                work_id: call.work_id.clone(),
-                        },
+                                event: AgentEvent::ReasoningCompleted { text, signature },
-                        project,
+                            },
-                    )? {
+                            project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                 }
                 AgentEvent::Tool {
                     event: ToolEvent::Proposed { call: tool_call },
                 } => {
-                    if !self.ensure_model_started(&call, &profile, started_at, &buffer, project)? {
+                    if !self
+                        .ensure_model_started_acknowledged(
+                            &call, &profile, started_at, &buffer, project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                     buffer.tool_calls.push(tool_call);
                 }
                 AgentEvent::UsageUpdated { usage } => {
-                    if !self.ensure_model_started(&call, &profile, started_at, &buffer, project)? {
+                    if !self
+                        .ensure_model_started_acknowledged(
+                            &call, &profile, started_at, &buffer, project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                     buffer.usage.clone_from(&usage);
                     let cumulative_usage = combined_usage(self.run.usage(), &usage);
-                    if !self.project_or_fail(
+                    if !self
-                        ProviderRunProjection::ModelEvent {
+                        .project_or_fail_acknowledged(
-                            work_id: call.work_id.clone(),
+                            ProviderRunProjection::ModelEvent {
-                            event: AgentEvent::UsageUpdated {
+                                work_id: call.work_id.clone(),
-                                usage: cumulative_usage,
+                                event: AgentEvent::UsageUpdated {
+                                    usage: cumulative_usage,
+                                },
                             },
-                        },
+                            project,
-                        project,
+                        )
-                    )? {
+                        .await?
+                    {
                         return Ok(());
                     }
                 }
                 AgentEvent::TurnStopped { reason } => {
-                    if !self.ensure_model_started(&call, &profile, started_at, &buffer, project)? {
+                    if !self
+                        .ensure_model_started_acknowledged(
+                            &call, &profile, started_at, &buffer, project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
-                    if !self.project_or_fail(
+                    if !self
-                        ProviderRunProjection::ModelTurnFinished {
+                        .project_or_fail_acknowledged(
-                            work_id: call.work_id.clone(),
+                            ProviderRunProjection::ModelTurnFinished {
-                            profile: call.profile.clone(),
+                                work_id: call.work_id.clone(),
-                            runtime_id: runtime_id.clone(),
+                                profile: call.profile.clone(),
-                            model_id: model_id.clone(),
+                                runtime_id: runtime_id.clone(),
-                            stop_reason: reason.clone(),
+                                model_id: model_id.clone(),
-                            retry_attempt: call.retry_attempt,
+                                stop_reason: reason.clone(),
-                            elapsed_ms: elapsed_millis(started_at),
+                                retry_attempt: call.retry_attempt,
-                            tool_call_count: buffer.tool_calls.len(),
+                                elapsed_ms: elapsed_millis(started_at),
-                        },
+                                tool_call_count: buffer.tool_calls.len(),
-                        project,
+                            },
-                    )? {
+                            project,
+                        )
+                        .await?
+                    {
                         return Ok(());
                     }
                     if reason == StopReason::Cancelled {
@@ -657,7 +794,7 @@ impl ProviderRunCoordinator {
                 | AgentEvent::ContextUsageUpdated { .. }
                 | AgentEvent::UserInputAccepted { .. }
                 | AgentEvent::RuntimeNotice { .. } => {
-                    self.handle_model_failure(
+                    self.handle_model_failure_acknowledged(
                         &call,
                         &profile,
                         started_at,
@@ -665,23 +802,25 @@ impl ProviderRunCoordinator {
                             "direct-provider transport emitted a non-model lifecycle event",
                         ),
                         project,
-                    )?;
+                    )
+                    .await?;
                     return Ok(());
                 }
             }
         }
 
-        self.handle_model_failure(
+        self.handle_model_failure_acknowledged(
             &call,
             &profile,
             started_at,
             protocol_error("provider stream ended before TurnStopped"),
             project,
-        )?;
+        )
+        .await?;
         Ok(())
     }
 
-    fn ensure_model_started<F>(
+    async fn ensure_model_started_acknowledged<F>(
         &mut self,
         call: &ProviderModelCall,
         profile: &ProviderRunProfile,
@@ -690,22 +829,23 @@ impl ProviderRunCoordinator {
         project: &mut F,
     ) -> Result<bool, ProviderRunCoordinatorError>
     where
-        F: FnMut(ProviderRunProjection) -> Result<(), String>,
+        F: FnMut(ProviderRunProjection) -> BoxFuture<'static, Result<(), String>>,
     {
         if buffer.started {
             return Ok(true);
         }
-        self.handle_model_failure(
+        self.handle_model_failure_acknowledged(
             call,
             profile,
             started_at,
             protocol_error("provider emitted model output before TurnStarted"),
             project,
-        )?;
+        )
+        .await?;
         Ok(false)
     }
 
-    fn handle_model_failure<F>(
+    async fn handle_model_failure_acknowledged<F>(
         &mut self,
         call: &ProviderModelCall,
         profile: &ProviderRunProfile,
@@ -714,7 +854,7 @@ impl ProviderRunCoordinator {
         project: &mut F,
     ) -> Result<(), ProviderRunCoordinatorError>
     where
-        F: FnMut(ProviderRunProjection) -> Result<(), String>,
+        F: FnMut(ProviderRunProjection) -> BoxFuture<'static, Result<(), String>>,
     {
         let disposition = self
             .run
@@ -737,7 +877,7 @@ impl ProviderRunCoordinator {
                     ));
                 }
             };
-            self.project_or_fail(
+            self.project_or_fail_acknowledged(
                 ProviderRunProjection::ModelRetry {
                     work_id: call.work_id.clone(),
                     profile: call.profile.clone(),
@@ -748,20 +888,21 @@ impl ProviderRunCoordinator {
                     error,
                 },
                 project,
-            )?;
+            )
+            .await?;
         }
         Ok(())
     }
 
-    fn project_or_fail<F>(
+    async fn project_or_fail_acknowledged<F>(
         &mut self,
         event: ProviderRunProjection,
         project: &mut F,
     ) -> Result<bool, ProviderRunCoordinatorError>
     where
-        F: FnMut(ProviderRunProjection) -> Result<(), String>,
+        F: FnMut(ProviderRunProjection) -> BoxFuture<'static, Result<(), String>>,
     {
-        match project(event) {
+        match project(event).await {
             Ok(()) => Ok(true),
             Err(message) => {
                 self.run.fail(

app/src/ai/runtime/provider_run_coordinator_tests.rs

@@ -79,6 +79,7 @@ struct StallingRuntime {
     first_attempt_stall: FirstAttemptStall,
     attempts: AtomicUsize,
     requests: Mutex<Vec<TurnRequest>>,
+    controls: Mutex<Vec<TurnControl>>,
 }
 
 impl StallingRuntime {
@@ -93,12 +94,22 @@ impl StallingRuntime {
             first_attempt_stall,
             attempts: AtomicUsize::new(0),
             requests: Mutex::new(Vec::new()),
+            controls: Mutex::new(Vec::new()),
         }
     }
 
     fn requests(&self) -> Vec<TurnRequest> {
         self.requests.lock().unwrap().clone()
     }
+
+    fn attempt_commands(&self) -> Vec<Option<TurnCommand>> {
+        self.controls
+            .lock()
+            .unwrap()
+            .iter()
+            .map(|control| control.try_receive().ok())
+            .collect()
+    }
 }
 
 #[async_trait]
@@ -110,9 +121,10 @@ impl AgentRuntime for StallingRuntime {
     async fn start_turn(
         &self,
         request: TurnRequest,
-        _control: TurnControl,
+        control: TurnControl,
     ) -> Result<AgentEventStream, AgentError> {
         self.requests.lock().unwrap().push(request);
+        self.controls.lock().unwrap().push(control);
         let attempt = self.attempts.fetch_add(1, Ordering::Relaxed);
         if attempt == 0 {
             match self.first_attempt_stall {
@@ -778,6 +790,64 @@ async fn recoverable_start_failure_retries_the_same_work_identity() {
     assert_eq!(retry.1, 1);
 }
 
+#[tokio::test]
+async fn model_progress_waits_for_each_projection_acknowledgement() {
+    let runtime = Arc::new(ScriptedRuntime::new(vec![answer_turn()]));
+    let mut coordinator = coordinator(runtime.clone());
+    let (acknowledgement_sender, acknowledgement_receiver) = async_channel::unbounded();
+    let (observed_sender, observed_receiver) = async_channel::unbounded();
+    let (_sender, control) = turn_control();
+
+    let drive = async {
+        coordinator
+            .drive_until_blocked_with_acknowledgements(
+                control,
+                move |projection| {
+                    let acknowledgement_receiver = acknowledgement_receiver.clone();
+                    let observed_sender = observed_sender.clone();
+                    Box::pin(async move {
+                        observed_sender.send(projection).await.unwrap();
+                        acknowledgement_receiver.recv().await.unwrap()
+                    })
+                },
+                |_| Box::pin(async { Ok(()) }),
+            )
+            .await
+            .unwrap()
+    };
+    let driver = async {
+        let first = observed_receiver.recv().await.unwrap();
+        assert!(matches!(
+            first,
+            ProviderRunProjection::ModelTurnRequested { .. }
+        ));
+        assert_eq!(runtime.requests().len(), 0);
+        assert!(observed_receiver.try_recv().is_err());
+        acknowledgement_sender.send(Ok(())).await.unwrap();
+
+        let second = observed_receiver.recv().await.unwrap();
+        assert!(matches!(
+            second,
+            ProviderRunProjection::ModelTurnStarted { .. }
+        ));
+        assert_eq!(runtime.requests().len(), 1);
+        assert!(observed_receiver.try_recv().is_err());
+        acknowledgement_sender.send(Ok(())).await.unwrap();
+
+        loop {
+            let projection = observed_receiver.recv().await.unwrap();
+            let finished = matches!(projection, ProviderRunProjection::ModelTurnFinished { .. });
+            acknowledgement_sender.send(Ok(())).await.unwrap();
+            if finished {
+                break;
+            }
+        }
+    };
+
+    let (block, ()) = futures::join!(drive, driver);
+    assert!(matches!(block, ProviderRunBlock::AwaitingDriver { .. }));
+}
+
 fn assert_single_retry_lifecycle(
     projections: &[ProviderRunProjection],
     expected_timeout_stage: &str,
@@ -868,6 +938,10 @@ async fn model_start_timeout_retries_the_same_work_identity() {
 
     assert!(matches!(block, ProviderRunBlock::AwaitingDriver { .. }));
     assert_eq!(runtime.requests().len(), 2);
+    assert_eq!(
+        runtime.attempt_commands(),
+        vec![Some(TurnCommand::Cancel), None]
+    );
     assert_eq!(coordinator.run().model_retries(), 1);
     assert_single_retry_lifecycle(&projections, "start timed out", false);
 }
@@ -887,10 +961,43 @@ async fn model_event_idle_timeout_retries_the_same_work_identity() {
 
     assert!(matches!(block, ProviderRunBlock::AwaitingDriver { .. }));
     assert_eq!(runtime.requests().len(), 2);
+    assert_eq!(
+        runtime.attempt_commands(),
+        vec![Some(TurnCommand::Cancel), None]
+    );
     assert_eq!(coordinator.run().model_retries(), 1);
     assert_single_retry_lifecycle(&projections, "event timed out", true);
 }
 
+#[tokio::test]
+async fn persistent_checkpoint_failure_terminates_without_redrive() {
+    let runtime = Arc::new(ScriptedRuntime::new(vec![answer_turn()]));
+    let mut coordinator = coordinator(runtime.clone());
+    let checkpoint_attempts = Arc::new(AtomicUsize::new(0));
+    let attempts = checkpoint_attempts.clone();
+    let (_sender, control) = turn_control();
+
+    let block = coordinator
+        .drive_until_blocked_with_checkpoint(
+            control,
+            |_| Ok(()),
+            move |_| {
+                attempts.fetch_add(1, Ordering::Relaxed);
+                Box::pin(async { Err("database unavailable".to_string()) })
+            },
+        )
+        .await
+        .unwrap();
+
+    let ProviderRunBlock::Done(ProviderRunOutcome::Failed(failure)) = block else {
+        panic!("checkpoint failure must terminate the drive");
+    };
+    assert_eq!(failure.kind, ProviderRunFailureKind::ExternalWork);
+    assert!(failure.message.contains("database unavailable"));
+    assert_eq!(checkpoint_attempts.load(Ordering::Relaxed), 1);
+    assert!(runtime.requests().is_empty());
+}
+
 #[tokio::test]
 async fn tool_projection_failure_synthesizes_a_result_and_fails_the_run() {
     let runtime = Arc::new(ScriptedRuntime::new(vec![tool_turn()]));
@@ -953,6 +1060,38 @@ async fn stream_without_terminal_event_fails_instead_of_committing_partial_outpu
     assert_eq!(coordinator.run().transcript().len(), 1);
 }
 
+#[tokio::test]
+async fn reasoning_completed_authoritatively_replaces_canonical_reasoning() {
+    let runtime = Arc::new(ScriptedRuntime::new(vec![Ok(vec![
+        started("request-reasoning"),
+        Ok(AgentEvent::ReasoningDelta {
+            text: "draft".to_string(),
+        }),
+        Ok(AgentEvent::ReasoningCompleted {
+            text: "final".to_string(),
+            signature: Some("signature".to_string()),
+        }),
+        stopped(StopReason::Completed),
+    ])]));
+    let mut coordinator = coordinator(runtime);
+    let (_sender, control) = turn_control();
+
+    let block = coordinator
+        .drive_until_blocked(control, |_| Ok(()))
+        .await
+        .unwrap();
+
+    assert!(matches!(block, ProviderRunBlock::AwaitingDriver { .. }));
+    let MessageContent::MultiPart(parts) = &coordinator.run().transcript()[1].content else {
+        panic!("expected canonical reasoning content");
+    };
+    assert!(matches!(
+        parts.as_slice(),
+        [ContentPart::Reasoning { text, signature }]
+            if text == "final" && signature.as_deref() == Some("signature")
+    ));
+}
+
 #[tokio::test]
 async fn provider_cancellation_does_not_commit_partial_assistant_content() {
     let expected_transcript = request().messages;

app/src/ai/runtime/rig_request.rs

@@ -835,6 +835,16 @@ fn build_system_prompt(
                 .join(", "),
         );
         prompt.push_str(".\nNever invent tool names or parameters. Check command exit codes and tool error results before claiming success.\n");
+        if tools.iter().any(|tool| tool.name == "run_shell_command") {
+            let has_file_tools = tools
+                .iter()
+                .any(|tool| matches!(tool.name.as_str(), "file_glob" | "grep" | "read_files"));
+            if has_file_tools {
+                prompt.push_str(
+                    "Prefer `file_glob`, `grep`, and `read_files` for file discovery, content search, and file reading when they are available. Reserve `run_shell_command` for operations those specialized tools cannot perform; do not use shell `find`, `grep`, `rg`, `cat`, `head`, or `tail` as substitutes.\n",
+                );
+            }
+        }
         if tools.iter().any(|tool| tool.name == "create_plan") {
             prompt.push_str(
                 "Plan document creation is available through `create_plan`. When the user asks you to create a plan for review, research first as needed, then call `create_plan`; do not merely return the plan as prose or claim that no plan-creation tool is available. If the user asks to review the plan before implementation, creating the document and presenting it for review is the requested outcome; do not implement it until they approve.\n",

app/src/ai/runtime/rig_request_tests.rs

@@ -127,6 +127,28 @@ fn builds_a_rig_turn_directly_from_galaxy_request_state() {
     ));
 }
 
+#[test]
+fn system_prompt_prefers_specialized_file_tools_over_shell_substitutes() {
+    let mut params = RequestParams::new_for_test();
+    params.input = vec![user_query("Find every Rust file containing ProviderRun")];
+
+    let prepared = prepare_rig_turn(
+        &config(),
+        params,
+        vec![
+            ToolType::RunShellCommand,
+            ToolType::FileGlob,
+            ToolType::Grep,
+            ToolType::ReadFiles,
+        ],
+        Vec::new(),
+    );
+    let prompt = prepared.request.system_prompt.expect("system prompt");
+
+    assert!(prompt.contains("Prefer `file_glob`, `grep`, and `read_files`"));
+    assert!(prompt.contains("do not use shell `find`, `grep`, `rg`, `cat`, `head`, or `tail`"));
+}
+
 #[test]
 fn normal_turn_advertises_plan_creation_and_corrects_false_unavailability_claims() {
     let mut params = RequestParams::new_for_test();

app/src/ai/runtime/rig_tool.rs

@@ -25,6 +25,9 @@ pub(crate) fn action_from_tool_call(
     mcp_tool_aliases: &HashMap<String, MCPToolTarget>,
 ) -> Result<AIAgentAction, String> {
     let input = &call.arguments;
+    if !input.is_object() {
+        return Err(format!("invalid {} input: expected an object", call.name));
+    }
     let action = if let Some(target) = mcp_tool_aliases.get(&call.name) {
         AIAgentActionType::CallMCPTool {
             server_id: target.server_id,
@@ -33,197 +36,220 @@ pub(crate) fn action_from_tool_call(
         }
     } else {
         match call.name.as_str() {
-        "run_shell_command" => AIAgentActionType::RequestCommandOutput {
+            "run_shell_command" => AIAgentActionType::RequestCommandOutput {
-            command: string(input, "command"),
+                command: required_nonempty_string(input, "command")?,
-            is_read_only: Some(boolean(input, "is_read_only")),
+                is_read_only: Some(optional_boolean(input, "is_read_only")?.unwrap_or(false)),
-            is_risky: Some(boolean(input, "is_risky")),
+                is_risky: Some(optional_boolean(input, "is_risky")?.unwrap_or(false)),
-            wait_until_completion: boolean(input, "wait_until_complete"),
+                wait_until_completion: optional_boolean(input, "wait_until_complete")?
-            uses_pager: Some(boolean(input, "uses_pager")),
+                    .unwrap_or(false),
-            rationale: None,
+                uses_pager: Some(optional_boolean(input, "uses_pager")?.unwrap_or(false)),
-            citations: Vec::new(),
+                rationale: None,
-        },
+                citations: Vec::new(),
-        "read_files" => AIAgentActionType::ReadFiles(ReadFilesRequest {
+            },
-            locations: input
+            "read_files" => AIAgentActionType::ReadFiles(ReadFilesRequest {
-                .get("files")
+                locations: required_array(input, "files")?
-                .and_then(serde_json::Value::as_array)
+                    .iter()
-                .into_iter()
+                    .enumerate()
-                .flatten()
+                    .map(|(index, file)| file_location(file, index))
-                .filter_map(file_location)
+                    .collect::<Result<_, _>>()?,
-                .collect(),
+            }),
-        }),
+            "apply_file_diffs" => AIAgentActionType::RequestFileEdits {
-        "apply_file_diffs" => AIAgentActionType::RequestFileEdits {
+                file_edits: file_edits(input)?,
-            file_edits: file_edits(input),
+                title: Some(required_string(input, "summary")?),
-            title: nonempty_string(input, "summary"),
+            },
-        },
+            "grep" => AIAgentActionType::Grep {
-        "grep" => AIAgentActionType::Grep {
+                queries: required_strings(input, "queries")?,
-            queries: strings(input, "queries"),
+                path: optional_string(input, "path")?.unwrap_or_default(),
-            path: string(input, "path"),
+            },
-        },
+            "file_glob" => AIAgentActionType::FileGlob {
-        "file_glob" => AIAgentActionType::FileGlob {
+                patterns: required_strings(input, "patterns")?,
-            patterns: strings(input, "patterns"),
+                path: optional_string(input, "path")?.filter(|path| !path.is_empty()),
-            path: nonempty_string(input, "path"),
+            },
-        },
+            "search_codebase" => AIAgentActionType::SearchCodebase(SearchCodebaseRequest {
-        "search_codebase" => AIAgentActionType::SearchCodebase(SearchCodebaseRequest {
+                query: required_string(input, "query")?,
-            query: string(input, "query"),
+                partial_paths: optional_strings(input, "path_filters")?
-            partial_paths: nonempty_strings(input, "path_filters"),
+                    .filter(|paths| !paths.is_empty()),
-            codebase_path: nonempty_string(input, "path"),
+                codebase_path: optional_string(input, "path")?.filter(|path| !path.is_empty()),
-        }),
+            }),
-        "write_to_long_running_shell_command" => {
+            "write_to_long_running_shell_command" => {
-            AIAgentActionType::WriteToLongRunningShellCommand {
+                AIAgentActionType::WriteToLongRunningShellCommand {
-                block_id: string(input, "command_id").into(),
+                    block_id: required_nonempty_string(input, "command_id")?.into(),
-                input: string(input, "input").into_bytes().into(),
+                    input: required_string(input, "input")?.into_bytes().into(),
-                mode: match input.get("mode").and_then(serde_json::Value::as_str) {
+                    mode: match optional_string(input, "mode")?.as_deref() {
-                    Some("line") => AIAgentPtyWriteMode::Line,
+                        Some("line") => AIAgentPtyWriteMode::Line,
-                    Some("block") => AIAgentPtyWriteMode::Block,
+                        Some("block") => AIAgentPtyWriteMode::Block,
-                    Some("raw") | Some(_) | None => AIAgentPtyWriteMode::Raw,
+                        Some("raw") | None => AIAgentPtyWriteMode::Raw,
-                },
+                        Some(mode) => {
+                            return Err(format!(
+                            "invalid field \"mode\": expected \"raw\", \"line\", or \"block\", got {mode:?}"
+                        ));
+                        }
+                    },
+                }
             }
-        }
+            "interrupt_shell_command" => AIAgentActionType::WriteToLongRunningShellCommand {
-        "interrupt_shell_command" => AIAgentActionType::WriteToLongRunningShellCommand {
+                block_id: required_nonempty_string(input, "command_id")?.into(),
-            block_id: string(input, "command_id").into(),
+                input: vec![galaxy_terminal::model::escape_sequences::C0::ETX].into(),
-            input: vec![galaxy_terminal::model::escape_sequences::C0::ETX].into(),
+                mode: AIAgentPtyWriteMode::Raw,
-            mode: AIAgentPtyWriteMode::Raw,
+            },
-        },
+            "read_shell_command_output" => AIAgentActionType::ReadShellCommandOutput {
-        "read_shell_command_output" => AIAgentActionType::ReadShellCommandOutput {
+                block_id: required_nonempty_string(input, "command_id")?.into(),
-            block_id: string(input, "command_id").into(),
+                delay: Some(ShellCommandDelay::Duration(Duration::from_secs(
-            delay: Some(ShellCommandDelay::Duration(Duration::from_secs(
+                    optional_bounded_u64(
-                input
+                        input,
-                    .get("wait_seconds")
+                        "wait_seconds",
-                    .and_then(serde_json::Value::as_u64)
+                        crate::ai::bedrock::request_translator::COMMAND_MONITOR_MAX_POLL_SECONDS,
-                    .unwrap_or(2)
+                    )?
-                    .min(crate::ai::bedrock::request_translator::COMMAND_MONITOR_MAX_POLL_SECONDS),
+                    .unwrap_or(2),
-            ))),
+                ))),
-        },
+            },
-        "read_mcp_resource" => AIAgentActionType::ReadMCPResource {
+            "read_mcp_resource" => AIAgentActionType::ReadMCPResource {
-            server_id: uuid(input, "server_id"),
+                server_id: Some(required_uuid(input, "server_id")?),
-            name: String::new(),
+                name: String::new(),
-            uri: nonempty_string(input, "uri"),
+                uri: Some(required_string(input, "uri")?),
-        },
+            },
-        "read_plan" | "read_documents" | "read_notebook" => {
+            "read_plan" | "read_documents" | "read_notebook" => {
-            AIAgentActionType::ReadDocuments(ReadDocumentsRequest {
+                AIAgentActionType::ReadDocuments(ReadDocumentsRequest {
-                document_ids: strings(input, "document_ids")
+                    document_ids: required_strings(input, "document_ids")?
-                    .into_iter()
-                    .filter_map(|id| AIDocumentId::try_from(id).ok())
-                    .collect(),
-            })
-        }
-        "create_plan" | "create_documents" | "create_notebook" => {
-            AIAgentActionType::CreateDocuments(CreateDocumentsRequest {
-                documents: input
-                    .get("documents")
-                    .and_then(serde_json::Value::as_array)
-                    .into_iter()
-                    .flatten()
-                    .filter_map(|document| {
-                        Some(DocumentToCreate {
-                            title: document.get("title")?.as_str()?.to_string(),
-                            content: document.get("content")?.as_str()?.to_string(),
-                        })
-                    })
-                    .collect(),
-            })
-        }
-        "edit_plan" | "edit_documents" | "edit_notebook" => {
-            AIAgentActionType::EditDocuments(EditDocumentsRequest {
-                diffs: input
-                    .get("diffs")
-                    .and_then(serde_json::Value::as_array)
-                    .into_iter()
-                    .flatten()
-                    .filter_map(|diff| {
-                        Some(DocumentDiff {
-                            document_id: AIDocumentId::try_from(diff.get("document_id")?.as_str()?)
-                                .ok()?,
-                            search: string(diff, "search"),
-                            replace: string(diff, "replace"),
-                        })
-                    })
-                    .collect(),
-            })
-        }
-        "run_agents" => AIAgentActionType::RunAgents(RunAgentsRequest {
-            summary: string(input, "summary"),
-            base_prompt: string(input, "base_prompt"),
-            skills: skill_references(input, skill_path_origin),
-            model_id: string(input, "model_id"),
-            harness_type: string(input, "harness_type"),
-            execution_mode: run_agents_execution_mode(input),
-            agent_run_configs: input
-                .get("agent_run_configs")
-                .and_then(serde_json::Value::as_array)
-                .into_iter()
-                .flatten()
-                .map(|config| RunAgentsAgentRunConfig {
-                    name: string(config, "name"),
-                    prompt: string(config, "prompt"),
-                    title: string(config, "title"),
-                })
-                .collect(),
-            plan_id: string(input, "plan_id"),
-            harness_auth_secret_name: None,
-        }),
-        "start_agent" => AIAgentActionType::StartAgent {
-            version: StartAgentVersion::V1,
-            name: string(input, "name"),
-            prompt: string(input, "prompt"),
-            execution_mode: StartAgentExecutionMode::local_with_defaults(),
-            lifecycle_subscription: None,
-        },
-        "send_message_to_agent" => AIAgentActionType::SendMessageToAgent {
-            addresses: vec![string(input, "agent_id")],
-            subject: String::new(),
-            message: string(input, "message"),
-        },
-        "ask_user_question" => AIAgentActionType::AskUserQuestion {
-            questions: vec![AskUserQuestionItem {
-                question_id: Uuid::new_v4().to_string(),
-                question: string(input, "question"),
-                question_type: AskUserQuestionType::MultipleChoice {
-                    is_multiselect: false,
-                    options: strings(input, "options")
                         .into_iter()
-                        .enumerate()
+                        .map(|id| {
-                        .map(|(index, label)| AskUserQuestionOption {
+                            AIDocumentId::try_from(id.clone()).map_err(|_| {
-                            label,
+                                format!("invalid document_ids entry: {id:?} is not a document ID")
-                            recommended: index == 0,
+                            })
                         })
-                        .collect(),
+                        .collect::<Result<_, _>>()?,
-                    supports_other: true,
+                })
-                },
+            }
-            }],
+            "create_plan" | "create_documents" | "create_notebook" => {
-        },
+                AIAgentActionType::CreateDocuments(CreateDocumentsRequest {
-        "read_skill" => {
+                    documents: required_array(input, "documents")?
-            let skill = string(input, "skill");
+                        .iter()
-            let skill = match input
+                        .enumerate()
-                .get("reference_type")
+                        .map(|(index, document)| {
-                .and_then(serde_json::Value::as_str)
+                            require_object(document, &format!("documents[{index}]"))?;
-            {
+                            Ok(DocumentToCreate {
-                Some("bundled") => SkillReference::BundledSkillId(skill),
+                                title: required_string(document, "title")?,
-                Some("path") | Some(_) | None => SkillReference::Path(
+                                content: required_string(document, "content")?,
-                    skill_path_origin
+                            })
-                        .location_for_path(skill)
+                        })
-                        .map_err(|error| error.to_string())?,
+                        .collect::<Result<_, String>>()?,
-                ),
+                })
-            };
+            }
-            AIAgentActionType::ReadSkill(ReadSkillRequest { skill })
+            "edit_plan" | "edit_documents" | "edit_notebook" => {
-        }
+                AIAgentActionType::EditDocuments(EditDocumentsRequest {
-        "fetch_conversation" => AIAgentActionType::FetchConversation {
+                    diffs: required_array(input, "diffs")?
-            conversation_id: string(input, "conversation_id"),
+                        .iter()
-        },
+                        .enumerate()
-        name if name.starts_with("mcp__") => {
+                        .map(|(index, diff)| {
-            let mut parts = name.splitn(3, "__");
+                            require_object(diff, &format!("diffs[{index}]"))?;
-            let _prefix = parts.next();
+                            let document_id = required_string(diff, "document_id")?;
-            let server_id = parts.next().and_then(|value| Uuid::parse_str(value).ok());
+                            Ok(DocumentDiff {
-            let name = parts
+                                document_id: AIDocumentId::try_from(document_id.clone())
-                .next()
+                                    .map_err(|_| format!("invalid document_id: {document_id:?}"))?,
-                .unwrap_or_else(|| name.strip_prefix("mcp__").unwrap_or(name))
+                                search: required_string(diff, "search")?,
-                .to_string();
+                                replace: required_string(diff, "replace")?,
-            AIAgentActionType::CallMCPTool {
+                            })
-                server_id,
+                        })
-                name,
+                        .collect::<Result<_, String>>()?,
-                input: input.clone(),
+                })
+            }
+            "run_agents" => AIAgentActionType::RunAgents(RunAgentsRequest {
+                summary: required_nonempty_string(input, "summary")?,
+                base_prompt: optional_string(input, "base_prompt")?.unwrap_or_default(),
+                skills: skill_references(input, skill_path_origin)?,
+                model_id: optional_string(input, "model_id")?.unwrap_or_default(),
+                harness_type: optional_string(input, "harness_type")?.unwrap_or_default(),
+                execution_mode: run_agents_execution_mode(input)?,
+                agent_run_configs: nonempty_required_array(input, "agent_run_configs")?
+                    .iter()
+                    .enumerate()
+                    .map(|(index, config)| {
+                        require_object(config, &format!("agent_run_configs[{index}]"))?;
+                        Ok(RunAgentsAgentRunConfig {
+                            name: required_nonempty_string(config, "name")?,
+                            prompt: required_nonempty_string(config, "prompt")?,
+                            title: optional_string(config, "title")?.unwrap_or_default(),
+                        })
+                    })
+                    .collect::<Result<_, String>>()?,
+                plan_id: optional_string(input, "plan_id")?.unwrap_or_default(),
+                harness_auth_secret_name: None,
+            }),
+            "start_agent" => AIAgentActionType::StartAgent {
+                version: StartAgentVersion::V1,
+                name: required_nonempty_string(input, "name")?,
+                prompt: required_nonempty_string(input, "prompt")?,
+                execution_mode: StartAgentExecutionMode::local_with_defaults(),
+                lifecycle_subscription: None,
+            },
+            "send_message_to_agent" => AIAgentActionType::SendMessageToAgent {
+                addresses: vec![required_string(input, "agent_id")?],
+                subject: String::new(),
+                message: required_string(input, "message")?,
+            },
+            "transfer_shell_command_control_to_user" => {
+                AIAgentActionType::TransferShellCommandControlToUser {
+                    reason: required_nonempty_string(input, "reason")?,
+                }
+            }
+            "wait_for_events" => AIAgentActionType::WaitForEvents {
+                tool_call_id: call.id.clone(),
+                idle_timeout_seconds: optional_nonnegative_i32(input, "idle_timeout_seconds")?
+                    .unwrap_or(0),
+            },
+            "ask_user_question" => AIAgentActionType::AskUserQuestion {
+                questions: vec![AskUserQuestionItem {
+                    question_id: Uuid::new_v4().to_string(),
+                    question: required_string(input, "question")?,
+                    question_type: AskUserQuestionType::MultipleChoice {
+                        is_multiselect: false,
+                        options: optional_strings(input, "options")?
+                            .unwrap_or_default()
+                            .into_iter()
+                            .enumerate()
+                            .map(|(index, label)| AskUserQuestionOption {
+                                label,
+                                recommended: index == 0,
+                            })
+                            .collect(),
+                        supports_other: true,
+                    },
+                }],
+            },
+            "read_skill" => {
+                let skill = required_string(input, "skill")?;
+                let skill = match required_string(input, "reference_type")?.as_str() {
+                    "bundled" => SkillReference::BundledSkillId(skill),
+                    "path" => SkillReference::Path(
+                        skill_path_origin
+                            .location_for_path(skill)
+                            .map_err(|error| error.to_string())?,
+                    ),
+                    reference_type => {
+                        return Err(format!(
+                        "invalid reference_type: expected \"path\" or \"bundled\", got {reference_type:?}"
+                    ));
+                    }
+                };
+                AIAgentActionType::ReadSkill(ReadSkillRequest { skill })
+            }
+            "fetch_conversation" => AIAgentActionType::FetchConversation {
+                conversation_id: required_string(input, "conversation_id")?,
+            },
+            name if name.starts_with("mcp__") => {
+                let mut parts = name.splitn(3, "__");
+                let _prefix = parts.next();
+                let server_id = parts.next().and_then(|value| Uuid::parse_str(value).ok());
+                let name = parts
+                    .next()
+                    .unwrap_or_else(|| name.strip_prefix("mcp__").unwrap_or(name))
+                    .to_string();
+                AIAgentActionType::CallMCPTool {
+                    server_id,
+                    name,
+                    input: input.clone(),
+                }
             }
-        }
             name => return Err(format!("unsupported Rig tool proposal: {name}")),
         }
     };
@@ -242,155 +268,309 @@ pub(crate) fn action_from_tool_call(
     })
 }
 
-fn string(input: &serde_json::Value, key: &str) -> String {
+fn require_object(input: &serde_json::Value, field: &str) -> Result<(), String> {
     input
-        .get(key)
+        .is_object()
-        .and_then(serde_json::Value::as_str)
+        .then_some(())
-        .unwrap_or_default()
+        .ok_or_else(|| format!("invalid {field}: expected an object"))
-        .to_string()
 }
 
-fn nonempty_string(input: &serde_json::Value, key: &str) -> Option<String> {
+fn required_string(input: &serde_json::Value, key: &str) -> Result<String, String> {
-    let value = string(input, key);
-    (!value.is_empty()).then_some(value)
-}
-
-fn boolean(input: &serde_json::Value, key: &str) -> bool {
     input
         .get(key)
-        .and_then(serde_json::Value::as_bool)
+        .ok_or_else(|| format!("missing required field {key:?}"))?
-        .unwrap_or(false)
+        .as_str()
-}
-
-fn strings(input: &serde_json::Value, key: &str) -> Vec<String> {
-    input
-        .get(key)
-        .and_then(serde_json::Value::as_array)
-        .into_iter()
-        .flatten()
-        .filter_map(serde_json::Value::as_str)
         .map(ToOwned::to_owned)
+        .ok_or_else(|| format!("invalid field {key:?}: expected a string"))
+}
+
+fn required_nonempty_string(input: &serde_json::Value, key: &str) -> Result<String, String> {
+    let value = required_string(input, key)?;
+    if value.trim().is_empty() {
+        return Err(format!(
+            "invalid field {key:?}: expected a non-empty string"
+        ));
+    }
+    Ok(value)
+}
+
+fn optional_string(input: &serde_json::Value, key: &str) -> Result<Option<String>, String> {
+    input
+        .get(key)
+        .map(|value| {
+            value
+                .as_str()
+                .map(ToOwned::to_owned)
+                .ok_or_else(|| format!("invalid field {key:?}: expected a string"))
+        })
+        .transpose()
+}
+
+fn required_array<'a>(
+    input: &'a serde_json::Value,
+    key: &str,
+) -> Result<&'a Vec<serde_json::Value>, String> {
+    input
+        .get(key)
+        .ok_or_else(|| format!("missing required field {key:?}"))?
+        .as_array()
+        .ok_or_else(|| format!("invalid field {key:?}: expected an array"))
+}
+
+fn nonempty_required_array<'a>(
+    input: &'a serde_json::Value,
+    key: &str,
+) -> Result<&'a Vec<serde_json::Value>, String> {
+    let values = required_array(input, key)?;
+    if values.is_empty() {
+        return Err(format!("invalid field {key:?}: expected at least one item"));
+    }
+    Ok(values)
+}
+
+fn required_strings(input: &serde_json::Value, key: &str) -> Result<Vec<String>, String> {
+    strings_from_array(required_array(input, key)?, key)
+}
+
+fn optional_strings(input: &serde_json::Value, key: &str) -> Result<Option<Vec<String>>, String> {
+    input
+        .get(key)
+        .map(|value| {
+            let values = value
+                .as_array()
+                .ok_or_else(|| format!("invalid field {key:?}: expected an array"))?;
+            strings_from_array(values, key)
+        })
+        .transpose()
+}
+
+fn strings_from_array(values: &[serde_json::Value], key: &str) -> Result<Vec<String>, String> {
+    values
+        .iter()
+        .enumerate()
+        .map(|(index, value)| {
+            value
+                .as_str()
+                .map(ToOwned::to_owned)
+                .ok_or_else(|| format!("invalid {key}[{index}]: expected a string"))
+        })
         .collect()
 }
 
-fn nonempty_strings(input: &serde_json::Value, key: &str) -> Option<Vec<String>> {
+fn required_uuid(input: &serde_json::Value, key: &str) -> Result<Uuid, String> {
-    let values = strings(input, key);
+    let value = required_string(input, key)?;
-    (!values.is_empty()).then_some(values)
+    Uuid::parse_str(&value).map_err(|_| format!("invalid field {key:?}: expected a UUID"))
 }
 
-fn uuid(input: &serde_json::Value, key: &str) -> Option<Uuid> {
+fn optional_boolean(input: &serde_json::Value, key: &str) -> Result<Option<bool>, String> {
     input
         .get(key)
-        .and_then(serde_json::Value::as_str)
+        .map(|value| {
-        .and_then(|value| Uuid::parse_str(value).ok())
+            value
+                .as_bool()
+                .ok_or_else(|| format!("invalid field {key:?}: expected a boolean"))
+        })
+        .transpose()
+}
+
+fn optional_bounded_u64(
+    input: &serde_json::Value,
+    key: &str,
+    maximum: u64,
+) -> Result<Option<u64>, String> {
+    input
+        .get(key)
+        .map(|value| {
+            let value = value
+                .as_u64()
+                .ok_or_else(|| format!("invalid field {key:?}: expected a non-negative integer"))?;
+            if value > maximum {
+                return Err(format!(
+                    "invalid field {key:?}: expected an integer no greater than {maximum}"
+                ));
+            }
+            Ok(value)
+        })
+        .transpose()
+}
+
+fn optional_nonnegative_i32(input: &serde_json::Value, key: &str) -> Result<Option<i32>, String> {
+    input
+        .get(key)
+        .map(|value| {
+            value
+                .as_i64()
+                .and_then(|value| i32::try_from(value).ok())
+                .filter(|value| *value >= 0)
+                .ok_or_else(|| {
+                    format!("invalid field {key:?}: expected a non-negative 32-bit integer")
+                })
+        })
+        .transpose()
 }
 
 fn skill_references(
     input: &serde_json::Value,
     skill_path_origin: &SkillPathOrigin,
-) -> Vec<SkillReference> {
+) -> Result<Vec<SkillReference>, String> {
-    input
+    let Some(skills) = optional_array(input, "skills")? else {
-        .get("skills")
+        return Ok(Vec::new());
-        .and_then(serde_json::Value::as_array)
+    };
-        .into_iter()
+    skills
-        .flatten()
+        .iter()
-        .filter_map(|skill| {
+        .enumerate()
-            let reference = string(skill, "skill");
+        .map(|(index, skill)| {
-            if reference.is_empty() {
+            require_object(skill, &format!("skills[{index}]"))?;
-                return None;
+            let reference = required_string(skill, "skill")?;
-            }
+            match required_string(skill, "reference_type")?.as_str() {
-            match skill
+                "bundled" => Ok(SkillReference::BundledSkillId(reference)),
-                .get("reference_type")
+                "path" => skill_path_origin
-                .and_then(serde_json::Value::as_str)
-            {
-                Some("bundled") => Some(SkillReference::BundledSkillId(reference)),
-                Some("path") | Some(_) | None => skill_path_origin
                     .location_for_path(reference)
-                    .ok()
+                    .map(SkillReference::Path)
-                    .map(SkillReference::Path),
+                    .map_err(|error| error.to_string()),
+                reference_type => Err(format!(
+                    "invalid skills[{index}].reference_type: expected \"path\" or \"bundled\", got {reference_type:?}"
+                )),
             }
         })
         .collect()
 }
 
-fn run_agents_execution_mode(input: &serde_json::Value) -> RunAgentsExecutionMode {
+fn run_agents_execution_mode(input: &serde_json::Value) -> Result<RunAgentsExecutionMode, String> {
     let Some(execution_mode) = input.get("execution_mode") else {
-        return RunAgentsExecutionMode::Local;
+        return Ok(RunAgentsExecutionMode::Local);
     };
-    let mode_type = execution_mode
+    require_object(execution_mode, "execution_mode")?;
-        .get("type")
+    match optional_string(execution_mode, "type")?.as_deref() {
-        .and_then(serde_json::Value::as_str)
+        Some("remote") => Ok(RunAgentsExecutionMode::Remote {
-        .or_else(|| execution_mode.as_str());
+            environment_id: optional_string(execution_mode, "environment_id")?.unwrap_or_default(),
-    match mode_type {
+            worker_host: optional_string(execution_mode, "worker_host")?.unwrap_or_default(),
-        Some("remote") => RunAgentsExecutionMode::Remote {
+            computer_use_enabled: optional_boolean(execution_mode, "computer_use_enabled")?
-            environment_id: string(execution_mode, "environment_id"),
+                .unwrap_or(false),
-            worker_host: string(execution_mode, "worker_host"),
+        }),
-            computer_use_enabled: boolean(execution_mode, "computer_use_enabled"),
+        Some("local") | None => {
-        },
+            optional_string(execution_mode, "environment_id")?;
-        Some("local") | Some(_) | None => RunAgentsExecutionMode::Local,
+            optional_string(execution_mode, "worker_host")?;
+            optional_boolean(execution_mode, "computer_use_enabled")?;
+            Ok(RunAgentsExecutionMode::Local)
+        }
+        Some(mode_type) => Err(format!(
+            "invalid execution_mode.type: expected \"local\" or \"remote\", got {mode_type:?}"
+        )),
     }
 }
 
-fn file_location(file: &serde_json::Value) -> Option<FileLocations> {
+fn file_location(file: &serde_json::Value, file_index: usize) -> Result<FileLocations, String> {
     if let Some(name) = file.as_str() {
-        return Some(FileLocations {
+        return Ok(FileLocations {
             name: name.to_string(),
             lines: Vec::new(),
         });
     }
-    let name = file
+    require_object(file, &format!("files[{file_index}]"))?;
-        .get("path")
+    let name = required_string(file, "path")?;
-        .or_else(|| file.get("name"))?
+    let lines = match file.get("line_ranges") {
-        .as_str()?
+        None => Vec::new(),
-        .to_string();
+        Some(value) => value
-    let lines = file
+            .as_array()
-        .get("line_ranges")
+            .ok_or_else(|| format!("invalid files[{file_index}].line_ranges: expected an array"))?
-        .and_then(serde_json::Value::as_array)
+            .iter()
-        .into_iter()
+            .enumerate()
-        .flatten()
+            .map(|(range_index, range)| {
-        .filter_map(|range| {
+                require_object(
-            let start = usize::try_from(range.get("start")?.as_u64()?).ok()?;
+                    range,
-            let end = usize::try_from(range.get("end")?.as_u64()?).ok()?;
+                    &format!("files[{file_index}].line_ranges[{range_index}]"),
-            (start > 0 && end >= start).then_some(start..end)
+                )?;
-        })
+                let start = required_line_number(range, "start", file_index, range_index)?;
-        .collect();
+                let inclusive_end = required_line_number(range, "end", file_index, range_index)?;
-    Some(FileLocations { name, lines })
+                if inclusive_end < start {
+                    return Err(format!(
+                        "invalid files[{file_index}].line_ranges[{range_index}]: end must be greater than or equal to start"
+                    ));
+                }
+                let exclusive_end = inclusive_end.checked_add(1).ok_or_else(|| {
+                    format!(
+                        "invalid files[{file_index}].line_ranges[{range_index}].end: inclusive end is too large"
+                    )
+                })?;
+                Ok(start..exclusive_end)
+            })
+            .collect::<Result<_, String>>()?,
+    };
+    Ok(FileLocations { name, lines })
 }
 
-fn file_edits(input: &serde_json::Value) -> Vec<FileEdit> {
+fn required_line_number(
-    let diffs = input
+    range: &serde_json::Value,
-        .get("diffs")
+    key: &str,
-        .and_then(serde_json::Value::as_array)
+    file_index: usize,
-        .into_iter()
+    range_index: usize,
-        .flatten()
+) -> Result<usize, String> {
-        .map(|diff| {
+    let value = range
-            FileEdit::Edit(ParsedDiff::StrReplaceEdit {
+        .get(key)
-                file: nonempty_string(diff, "file_path"),
+        .ok_or_else(|| format!("missing required field {key:?}"))?
-                search: nonempty_string(diff, "search"),
+        .as_u64()
-                replace: nonempty_string(diff, "replace"),
+        .and_then(|value| usize::try_from(value).ok())
-            })
+        .filter(|value| *value > 0)
-        });
+        .ok_or_else(|| {
-    let creates = input
+            format!(
-        .get("new_files")
+                "invalid files[{file_index}].line_ranges[{range_index}].{key}: expected a positive integer"
-        .and_then(serde_json::Value::as_array)
+            )
-        .into_iter()
+        })?;
-        .flatten()
+    Ok(value)
-        .map(|file| FileEdit::Create {
+}
-            file: nonempty_string(file, "file_path"),
+
-            content: nonempty_string(file, "content"),
+fn file_edits(input: &serde_json::Value) -> Result<Vec<FileEdit>, String> {
-        });
+    let mut edits = Vec::new();
-    let deletes = input
+    if let Some(diffs) = optional_array(input, "diffs")? {
-        .get("deleted_files")
+        for (index, diff) in diffs.iter().enumerate() {
-        .and_then(serde_json::Value::as_array)
+            require_object(diff, &format!("diffs[{index}]"))?;
-        .into_iter()
+            edits.push(FileEdit::Edit(ParsedDiff::StrReplaceEdit {
-        .flatten()
+                file: Some(required_string(diff, "file_path")?),
-        .map(|file| FileEdit::Delete {
+                search: Some(required_string(diff, "search")?),
-            file: file
+                replace: Some(required_string(diff, "replace")?),
+            }));
+        }
+    }
+    if let Some(files) = optional_array(input, "new_files")? {
+        for (index, file) in files.iter().enumerate() {
+            require_object(file, &format!("new_files[{index}]"))?;
+            edits.push(FileEdit::Create {
+                file: Some(required_string(file, "file_path")?),
+                content: Some(required_string(file, "content")?),
+            });
+        }
+    }
+    if let Some(files) = optional_array(input, "deleted_files")? {
+        for (index, file) in files.iter().enumerate() {
+            let path = file
                 .as_str()
-                .map(ToOwned::to_owned)
+                .ok_or_else(|| format!("invalid deleted_files[{index}]: expected a string"))?;
-                .or_else(|| nonempty_string(file, "file_path")),
+            edits.push(FileEdit::Delete {
-        });
+                file: Some(path.to_owned()),
-    diffs.chain(creates).chain(deletes).collect()
+            });
+        }
+    }
+    if edits.is_empty() {
+        return Err(
+            "invalid file edits: expected at least one diff, new file, or deleted file".to_string(),
+        );
+    }
+    Ok(edits)
+}
+
+fn optional_array<'a>(
+    input: &'a serde_json::Value,
+    key: &str,
+) -> Result<Option<&'a Vec<serde_json::Value>>, String> {
+    input
+        .get(key)
+        .map(|value| {
+            value
+                .as_array()
+                .ok_or_else(|| format!("invalid field {key:?}: expected an array"))
+        })
+        .transpose()
 }
 
 #[cfg(test)]

app/src/ai/runtime/rig_tool_tests.rs

@@ -46,6 +46,26 @@ fn shell_calls_become_domain_actions_without_a_proto_round_trip() {
     ));
 }
 
+#[test]
+fn transfer_control_calls_become_domain_actions() {
+    let action = action_from_tool_call(
+        "task-1",
+        &call(
+            "transfer_shell_command_control_to_user",
+            serde_json::json!({"reason": "The command needs interactive input"}),
+        ),
+        &SkillPathOrigin::Local,
+        &HashMap::new(),
+    )
+    .unwrap();
+
+    assert!(matches!(
+        action.action,
+        AIAgentActionType::TransferShellCommandControlToUser { reason }
+            if reason == "The command needs interactive input"
+    ));
+}
+
 #[test]
 fn create_plan_calls_become_document_actions() {
     let action = action_from_tool_call(
@@ -72,6 +92,227 @@ fn create_plan_calls_become_document_actions() {
     assert_eq!(request.documents[0].content, "# Implementation plan");
 }
 
+#[test]
+fn read_files_converts_advertised_inclusive_ranges_to_half_open_ranges() {
+    let action = action_from_tool_call(
+        "task-1",
+        &call(
+            "read_files",
+            serde_json::json!({
+                "files": [{
+                    "path": "/tmp/example.rs",
+                    "line_ranges": [
+                        {"start": 1, "end": 1},
+                        {"start": 10, "end": 25}
+                    ]
+                }]
+            }),
+        ),
+        &SkillPathOrigin::Local,
+        &HashMap::new(),
+    )
+    .unwrap();
+
+    let AIAgentActionType::ReadFiles(request) = action.action else {
+        panic!("expected read-files action");
+    };
+    assert_eq!(request.locations[0].lines, vec![1..2, 10..26]);
+}
+
+#[test]
+fn known_tools_reject_malformed_required_inputs() {
+    let cases = [
+        ("read_files", serde_json::json!({}), "files"),
+        (
+            "read_files",
+            serde_json::json!({"files": "not-an-array"}),
+            "expected an array",
+        ),
+        (
+            "read_files",
+            serde_json::json!({"files": [{"path": "/tmp/a", "line_ranges": [{"start": 0, "end": 1}]}]}),
+            "positive integer",
+        ),
+        (
+            "read_files",
+            serde_json::json!({"files": [{"path": "/tmp/a", "line_ranges": [{"start": 3, "end": 2}]}]}),
+            "greater than or equal",
+        ),
+        (
+            "read_files",
+            serde_json::json!({"files": [{"path": "/tmp/a", "line_ranges": [{"start": 1, "end": u64::MAX}]}]}),
+            "inclusive end is too large",
+        ),
+        (
+            "grep",
+            serde_json::json!({"queries": ["ok", 7]}),
+            "queries[1]",
+        ),
+        (
+            "file_glob",
+            serde_json::json!({"patterns": false}),
+            "expected an array",
+        ),
+        (
+            "search_codebase",
+            serde_json::json!({"query": 42}),
+            "expected a string",
+        ),
+        (
+            "apply_file_diffs",
+            serde_json::json!({"summary": "edit", "diffs": [{"file_path": "/tmp/a", "search": "x"}]}),
+            "replace",
+        ),
+        (
+            "apply_file_diffs",
+            serde_json::json!({"summary": "Nothing to do"}),
+            "at least one diff",
+        ),
+        (
+            "run_shell_command",
+            serde_json::json!({"command": 42}),
+            "expected a string",
+        ),
+        (
+            "run_shell_command",
+            serde_json::json!({"command": "  "}),
+            "non-empty string",
+        ),
+        (
+            "run_shell_command",
+            serde_json::json!({"command": "pwd", "is_read_only": "yes"}),
+            "expected a boolean",
+        ),
+        (
+            "write_to_long_running_shell_command",
+            serde_json::json!({"command_id": "command-1", "input": "yes", "mode": "words"}),
+            "mode",
+        ),
+        (
+            "interrupt_shell_command",
+            serde_json::json!({}),
+            "command_id",
+        ),
+        (
+            "read_shell_command_output",
+            serde_json::json!({"command_id": 12}),
+            "expected a string",
+        ),
+        (
+            "read_shell_command_output",
+            serde_json::json!({"command_id": "command-1", "wait_seconds": 11}),
+            "no greater than",
+        ),
+        (
+            "run_agents",
+            serde_json::json!({"summary": "Investigate", "agent_run_configs": []}),
+            "at least one item",
+        ),
+        (
+            "run_agents",
+            serde_json::json!({"summary": "Investigate", "agent_run_configs": [{"name": "one"}]}),
+            "prompt",
+        ),
+        (
+            "run_agents",
+            serde_json::json!({"summary": "Investigate", "agent_run_configs": [{"name": "one", "prompt": "Inspect"}], "execution_mode": {"type": "other"}}),
+            "execution_mode.type",
+        ),
+        (
+            "run_agents",
+            serde_json::json!({"summary": "Investigate", "agent_run_configs": [{"name": "one", "prompt": "Inspect"}], "skills": [{"skill": "test", "reference_type": "other"}]}),
+            "skills[0].reference_type",
+        ),
+        (
+            "start_agent",
+            serde_json::json!({"name": "worker"}),
+            "prompt",
+        ),
+        (
+            "transfer_shell_command_control_to_user",
+            serde_json::json!({"reason": false}),
+            "expected a string",
+        ),
+        (
+            "wait_for_events",
+            serde_json::json!({"idle_timeout_seconds": -1}),
+            "non-negative",
+        ),
+        (
+            "create_plan",
+            serde_json::json!({"documents": [{"title": "Plan"}]}),
+            "content",
+        ),
+        (
+            "read_skill",
+            serde_json::json!({"skill": "/tmp/SKILL.md", "reference_type": "other"}),
+            "reference_type",
+        ),
+        (
+            "fetch_conversation",
+            serde_json::json!({"conversation_id": null}),
+            "expected a string",
+        ),
+    ];
+
+    for (name, arguments, expected_error) in cases {
+        let error = action_from_tool_call(
+            "task-1",
+            &call(name, arguments),
+            &SkillPathOrigin::Local,
+            &HashMap::new(),
+        )
+        .unwrap_err();
+        assert!(
+            error.contains(expected_error),
+            "{name} error {error:?} did not contain {expected_error:?}"
+        );
+    }
+}
+
+#[test]
+fn known_tools_preserve_legitimate_optional_defaults() {
+    let cases = [
+        ("grep", serde_json::json!({"queries": ["needle"]})),
+        ("file_glob", serde_json::json!({"patterns": ["**/*.rs"]})),
+        (
+            "ask_user_question",
+            serde_json::json!({"question": "Continue?"}),
+        ),
+        (
+            "apply_file_diffs",
+            serde_json::json!({"summary": "Create file", "new_files": [{"file_path": "/tmp/new", "content": ""}]}),
+        ),
+        ("run_shell_command", serde_json::json!({"command": "pwd"})),
+        (
+            "write_to_long_running_shell_command",
+            serde_json::json!({"command_id": "command-1", "input": ""}),
+        ),
+        (
+            "read_shell_command_output",
+            serde_json::json!({"command_id": "command-1"}),
+        ),
+        (
+            "run_agents",
+            serde_json::json!({
+                "summary": "Investigate",
+                "agent_run_configs": [{"name": "worker", "prompt": "Inspect"}]
+            }),
+        ),
+        ("wait_for_events", serde_json::json!({})),
+    ];
+
+    for (name, arguments) in cases {
+        action_from_tool_call(
+            "task-1",
+            &call(name, arguments),
+            &SkillPathOrigin::Local,
+            &HashMap::new(),
+        )
+        .unwrap_or_else(|error| panic!("{name} rejected optional defaults: {error}"));
+    }
+}
+
 #[test]
 fn edit_calls_preserve_file_edits_in_the_domain_model() {
     let action = action_from_tool_call(

app/src/ai/tool_diagnostics.rs

@@ -0,0 +1,46 @@
+use std::backtrace::Backtrace;
+use std::ffi::OsStr;
+use std::sync::OnceLock;
+
+const ENV_VAR: &str = "GALAXY_TOOL_DIAGNOSTICS";
+
+fn env_value_is_enabled(value: Option<&OsStr>) -> bool {
+    value
+        .and_then(OsStr::to_str)
+        .is_some_and(|value| value == "1" || value.eq_ignore_ascii_case("true"))
+}
+
+pub(crate) fn is_enabled() -> bool {
+    static ENABLED: OnceLock<bool> = OnceLock::new();
+    *ENABLED.get_or_init(|| env_value_is_enabled(std::env::var_os(ENV_VAR).as_deref()))
+}
+
+pub(crate) fn capture_backtrace() -> Option<Backtrace> {
+    is_enabled().then(Backtrace::force_capture)
+}
+
+macro_rules! tool_debug {
+    ($($arg:tt)*) => {
+        if $crate::ai::tool_diagnostics::is_enabled() {
+            log::debug!("[tool-debug] {}", format_args!($($arg)*));
+        }
+    };
+}
+
+pub(crate) use tool_debug;
+
+#[cfg(test)]
+mod tests {
+    use std::ffi::OsStr;
+
+    use super::env_value_is_enabled;
+
+    #[test]
+    fn diagnostic_env_accepts_only_explicit_true_values() {
+        assert!(env_value_is_enabled(Some(OsStr::new("1"))));
+        assert!(env_value_is_enabled(Some(OsStr::new("TRUE"))));
+        assert!(!env_value_is_enabled(Some(OsStr::new("0"))));
+        assert!(!env_value_is_enabled(Some(OsStr::new("yes"))));
+        assert!(!env_value_is_enabled(None));
+    }
+}

app/src/terminal/view.rs

@@ -7254,8 +7254,22 @@ impl TerminalView {
         event: &BlocklistAIActionEvent,
         ctx: &mut ViewContext<Self>,
     ) {
+        let event_matches_active_conversation = || {
+            let Some(event_conversation_id) = event.conversation_id() else {
+                return true;
+            };
+            self.model
+                .lock()
+                .block_list()
+                .active_block()
+                .ai_conversation_id()
+                == Some(event_conversation_id)
+        };
         match event {
             BlocklistAIActionEvent::ActionBlockedOnUserConfirmation { .. } => {
+                if !event_matches_active_conversation() {
+                    return;
+                }
                 let is_agent_in_control = self
                     .model
                     .lock()
@@ -7267,14 +7281,20 @@ impl TerminalView {
                 }
             }
             BlocklistAIActionEvent::ExecutingAction { .. } => {
-                self.redetermine_terminal_focus(ctx);
+                if event_matches_active_conversation() {
+                    self.redetermine_terminal_focus(ctx);
+                }
                 ctx.notify();
             }
-            BlocklistAIActionEvent::FinishedAction { action_id, .. } => {
+            BlocklistAIActionEvent::FinishedAction {
+                action_id,
+                conversation_id,
+                ..
+            } => {
                 // Refresh git line changes when files are potentially updated by an action
                 let action_result = action_model
                     .as_ref(ctx)
-                    .get_action_result(action_id)
+                    .get_action_result(*conversation_id, action_id)
                     .cloned();
 
                 let maybe_modified_files = action_result
@@ -7644,7 +7664,11 @@ impl TerminalView {
                 drop(model);
 
                 self.cli_subagent_controller.update(ctx, |controller, _| {
-                    controller.track_requested_command(&block_id, action_id);
+                    controller.track_requested_command(
+                        &block_id,
+                        parent_conversation_id,
+                        action_id,
+                    );
                 });
 
                 ctx.emit(Event::ExecuteCommand(ExecuteCommandEvent {
@@ -7672,10 +7696,23 @@ impl TerminalView {
             ShellCommandExecutorEvent::WriteToPty { input, mode } => {
                 self.write_agent_bytes_to_pty(input.to_vec(), mode, ctx);
             }
-            ShellCommandExecutorEvent::CancelExecution => {
+            ShellCommandExecutorEvent::CancelExecution { action_id } => {
                 // We need to manually invoke ctrl-c to terminate the running command because the
                 // user's ctrl-c was directed to the AIBlock instead of the command's shell block.
-                self.ctrl_c(ctx);
+                let is_exact_active_command = self
+                    .model
+                    .lock()
+                    .block_list()
+                    .active_block()
+                    .requested_command_action_id()
+                    .is_some_and(|requested_id| requested_id == action_id);
+                if is_exact_active_command {
+                    self.ctrl_c(ctx);
+                } else {
+                    log::warn!(
+                        "Refusing to interrupt active terminal command for stale requested action {action_id}"
+                    );
+                }
             }
             ShellCommandExecutorEvent::TransferControlToUser { reason, .. } => {
                 // Transfer control of the long-running command to the user.

app/src/terminal/view/context_menu.rs

@@ -178,8 +178,10 @@ impl TerminalView {
 
         let mut result = Vec::new();
         for exchange in conversation.root_task_exchanges() {
-            let formatted_exchange =
+            let formatted_exchange = exchange.format_for_copy_for_conversation(
-                exchange.format_for_copy(Some(self.ai_action_model.as_ref(ctx)));
+                Some(self.ai_action_model.as_ref(ctx)),
+                Some(conversation_id),
+            );
             if !formatted_exchange.is_empty() {
                 result.push(formatted_exchange);
             }

app/src/terminal/view/load_ai_conversation.rs

@@ -364,7 +364,9 @@ impl TerminalView {
                             }) => {
                                 if let Some(result) =
                                     self.ai_action_model.read(ctx, |action_model, _| {
-                                        action_model.get_action_result(&action.id).cloned()
+                                        action_model
+                                            .get_action_result(conversation_id, &action.id)
+                                            .cloned()
                                     })
                                 {
                                     if let AIAgentActionResultType::CreateDocuments(
@@ -402,7 +404,9 @@ impl TerminalView {
                             AIAgentActionType::EditDocuments { .. } => {
                                 if let Some(result) =
                                     self.ai_action_model.read(ctx, |action_model, _| {
-                                        action_model.get_action_result(&action.id).cloned()
+                                        action_model
+                                            .get_action_result(conversation_id, &action.id)
+                                            .cloned()
                                     })
                                 {
                                     if let AIAgentActionResultType::EditDocuments(
@@ -449,8 +453,10 @@ impl TerminalView {
 
         for conversation in &conversations {
             self.ai_action_model.update(ctx, |action_model, _ctx| {
-                action_model
+                action_model.restore_action_results_from_exchanges(
-                    .restore_action_results_from_exchanges(exchanges_for_blocklist(conversation));
+                    conversation.id(),
+                    exchanges_for_blocklist(conversation),
+                );
             });
         }
 

crates/ai/src/agent/action_result/convert.rs

@@ -94,6 +94,25 @@ impl TryFrom<RequestCommandOutputResult> for api::request::input::tool_call_resu
                     },
                 ),
             ),
+            RequestCommandOutputResult::ExecutionError { command, message } => Ok(
+                api::request::input::tool_call_result::Result::RunShellCommand(
+                    #[allow(deprecated)]
+                    api::RunShellCommandResult {
+                        command,
+                        output: Default::default(),
+                        exit_code: Default::default(),
+                        result: Some(api::run_shell_command_result::Result::CommandFinished(
+                            api::ShellCommandFinished {
+                                command_id: String::new(),
+                                output: format!("Command was not executed: {message}"),
+                                exit_code: 1,
+                                start_ts: None,
+                                finish_ts: None,
+                            },
+                        )),
+                    },
+                ),
+            ),
             RequestCommandOutputResult::Denylisted { command } =>
             {
                 #[allow(deprecated)]
@@ -1551,6 +1570,14 @@ impl From<RunAgentsAgentOutcome> for api::run_agents_result::AgentOutcome {
                     api::run_agents_result::LaunchedAgent { agent_id },
                 )
             }
+            // The legacy wire schema has no completed-child shape. Preserve the child identity;
+            // direct-provider history retains the richer local result and output.
+            RunAgentsAgentOutcomeKind::Completed {
+                agent_id,
+                output: _,
+            } => api::run_agents_result::agent_outcome::Result::Launched(
+                api::run_agents_result::LaunchedAgent { agent_id },
+            ),
             RunAgentsAgentOutcomeKind::Failed { error } => {
                 api::run_agents_result::agent_outcome::Result::Failed(
                     api::run_agents_result::FailedAgent { error },

crates/ai/src/agent/action_result/convert_tests.rs

@@ -28,3 +28,22 @@ fn ask_user_question_skipped_by_auto_approve_converts_to_skipped_answers() {
         Some(AskUserQuestionAnswer::Skipped(()))
     ));
 }
+
+#[test]
+fn completed_run_agents_child_converts_to_legacy_launched_wire_outcome() {
+    let outcome = RunAgentsAgentOutcome {
+        name: "research".to_string(),
+        kind: RunAgentsAgentOutcomeKind::Completed {
+            agent_id: "child-1".to_string(),
+            output: "local output".to_string(),
+        },
+    };
+
+    let converted = api::run_agents_result::AgentOutcome::from(outcome);
+    assert!(matches!(
+        converted.result,
+        Some(api::run_agents_result::agent_outcome::Result::Launched(
+            api::run_agents_result::LaunchedAgent { agent_id }
+        )) if agent_id == "child-1"
+    ));
+}

crates/ai/src/agent/action_result/mod.rs

@@ -165,6 +165,7 @@ impl AIAgentActionResultType {
                     None,
                 ),
                 RequestCommandOutputResult::CancelledBeforeExecution
+                | RequestCommandOutputResult::ExecutionError { .. }
                 | RequestCommandOutputResult::Denylisted { .. } => result.to_string(),
             },
             Self::WriteToLongRunningShellCommand(result) => match result {
@@ -418,6 +419,8 @@ pub enum RequestCommandOutputResult {
     /// A running command canceled via ctrl-c
     /// would have Completed result with exit code 130.
     CancelledBeforeExecution,
+    /// The command could not start because the terminal was unavailable for execution.
+    ExecutionError { command: String, message: String },
     /// The command was denied because it was present on the denylist.
     Denylisted { command: String },
 }
@@ -427,14 +430,16 @@ impl RequestCommandOutputResult {
         match self {
             Self::Completed { exit_code, .. } => exit_code.was_successful(),
             Self::LongRunningCommandSnapshot { .. } => true,
-            Self::CancelledBeforeExecution | Self::Denylisted { .. } => false,
+            Self::CancelledBeforeExecution
+            | Self::ExecutionError { .. }
+            | Self::Denylisted { .. } => false,
         }
     }
 
     pub fn failed(&self) -> bool {
         match self {
             Self::Completed { exit_code, .. } => !exit_code.was_successful(),
-            Self::Denylisted { .. } => true,
+            Self::ExecutionError { .. } | Self::Denylisted { .. } => true,
             Self::CancelledBeforeExecution | Self::LongRunningCommandSnapshot { .. } => false,
         }
     }
@@ -444,6 +449,7 @@ impl RequestCommandOutputResult {
         match self {
             Self::Completed { command, .. }
             | Self::LongRunningCommandSnapshot { command, .. }
+            | Self::ExecutionError { command, .. }
             | Self::Denylisted { command } => command.clone(),
             Self::CancelledBeforeExecution => "cancelled".to_string(),
         }
@@ -473,6 +479,9 @@ impl Display for RequestCommandOutputResult {
             RequestCommandOutputResult::CancelledBeforeExecution => {
                 write!(f, "Command output cancelled")
             }
+            RequestCommandOutputResult::ExecutionError { command, message } => {
+                write!(f, "Command '{command}' could not be executed: {message}")
+            }
             RequestCommandOutputResult::Denylisted { .. } => {
                 write!(f, "Command output was on denylist")
             }
@@ -1042,7 +1051,9 @@ impl AIAgentActionResultType {
                 | TransferShellCommandControlToUserResult::CommandFinished { .. },
             ) => true,
             Self::AskUserQuestion(AskUserQuestionResult::Success { .. }) => true,
-            Self::RunAgents(RunAgentsResult::Launched { .. }) => true,
+            Self::RunAgents(RunAgentsResult::Launched { agents, .. }) => agents
+                .iter()
+                .any(|agent| !matches!(agent.kind, RunAgentsAgentOutcomeKind::Failed { .. })),
             Self::WaitForEvents(WaitForEventsResult::Completed) => true,
             _ => false,
         }
@@ -1076,6 +1087,12 @@ impl AIAgentActionResultType {
             | Self::RunAgents(RunAgentsResult::Failure { .. } | RunAgentsResult::Denied { .. }) => {
                 true
             }
+            Self::RunAgents(RunAgentsResult::Launched { agents, .. }) => {
+                !agents.is_empty()
+                    && agents
+                        .iter()
+                        .all(|agent| matches!(agent.kind, RunAgentsAgentOutcomeKind::Failed { .. }))
+            }
             _ => false,
         }
     }
@@ -1627,6 +1644,7 @@ pub struct RunAgentsAgentOutcome {
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
 pub enum RunAgentsAgentOutcomeKind {
     Launched { agent_id: String },
+    Completed { agent_id: String, output: String },
     Failed { error: String },
 }
 
@@ -1654,6 +1672,17 @@ impl RunAgentsResult {
                         "computer_use_enabled": computer_use_enabled,
                     }),
                 };
+                let children_completed = agents.iter().all(|agent| {
+                    matches!(
+                        agent.kind,
+                        RunAgentsAgentOutcomeKind::Completed { .. }
+                            | RunAgentsAgentOutcomeKind::Failed { .. }
+                    )
+                });
+                let all_failed = !agents.is_empty()
+                    && agents.iter().all(|agent| {
+                        matches!(agent.kind, RunAgentsAgentOutcomeKind::Failed { .. })
+                    });
                 let agents = agents
                     .iter()
                     .map(|agent| match &agent.kind {
@@ -1662,6 +1691,14 @@ impl RunAgentsResult {
                             "status": "launched",
                             "agent_id": agent_id,
                         }),
+                        RunAgentsAgentOutcomeKind::Completed { agent_id, output } => {
+                            serde_json::json!({
+                                "name": agent.name,
+                                "status": "completed",
+                                "agent_id": agent_id,
+                                "output": output,
+                            })
+                        }
                         RunAgentsAgentOutcomeKind::Failed { error } => serde_json::json!({
                             "name": agent.name,
                             "status": "failed",
@@ -1670,9 +1707,13 @@ impl RunAgentsResult {
                     })
                     .collect::<Vec<_>>();
                 serde_json::json!({
-                    "status": "launched",
+                    "status": if all_failed { "failure" } else { "launched" },
-                    "completion_state": "children_running",
+                    "completion_state": if children_completed { "children_completed" } else { "children_running" },
-                    "instruction": "Child agents have only been launched, not completed. Wait for child-agent updates before repeating this work or reporting final results.",
+                    "instruction": if children_completed {
+                        "Child agents reached terminal states. Use their structured outputs and errors to complete the task."
+                    } else {
+                        "Child agents have only been launched, not completed. Wait for child-agent updates before repeating this work or reporting final results."
+                    },
                     "model_id": model_id,
                     "harness_type": harness_type,
                     "execution_mode": execution_mode,
@@ -1701,7 +1742,13 @@ impl Display for RunAgentsResult {
             RunAgentsResult::Launched { agents, .. } => {
                 let launched = agents
                     .iter()
-                    .filter(|a| matches!(a.kind, RunAgentsAgentOutcomeKind::Launched { .. }))
+                    .filter(|a| {
+                        matches!(
+                            a.kind,
+                            RunAgentsAgentOutcomeKind::Launched { .. }
+                                | RunAgentsAgentOutcomeKind::Completed { .. }
+                        )
+                    })
                     .count();
                 write!(
                     f,

crates/ai/src/agent/action_result/mod_tests.rs

@@ -1,8 +1,22 @@
 use super::{
-    AIAgentActionResultType, RunAgentsAgentOutcome, RunAgentsAgentOutcomeKind,
+    AIAgentActionResultType, RequestCommandOutputResult, RunAgentsAgentOutcome,
-    RunAgentsLaunchedExecutionMode, RunAgentsResult, StartAgentResult, StartAgentVersion,
+    RunAgentsAgentOutcomeKind, RunAgentsLaunchedExecutionMode, RunAgentsResult, StartAgentResult,
+    StartAgentVersion,
 };
 
+#[test]
+fn shell_execution_error_is_failed_but_not_cancelled() {
+    let result =
+        AIAgentActionResultType::RequestCommandOutput(RequestCommandOutputResult::ExecutionError {
+            command: "cargo test".to_string(),
+            message: "terminal is busy".to_string(),
+        });
+
+    assert!(result.is_failed());
+    assert!(!result.is_cancelled());
+    assert!(result.model_content().contains("terminal is busy"));
+}
+
 #[test]
 fn deserializes_legacy_start_agent_success_without_version_as_v1() {
     let result: StartAgentResult =
@@ -134,3 +148,72 @@ fn run_agents_model_content_serializes_terminal_non_launch_outcomes() {
         assert_eq!(content, expected);
     }
 }
+
+#[test]
+fn completed_local_run_agents_preserves_outputs_and_terminal_state() {
+    let result = AIAgentActionResultType::RunAgents(RunAgentsResult::Launched {
+        model_id: "model".to_string(),
+        harness_type: "codex".to_string(),
+        execution_mode: RunAgentsLaunchedExecutionMode::Local,
+        agents: vec![RunAgentsAgentOutcome {
+            name: "research".to_string(),
+            kind: RunAgentsAgentOutcomeKind::Completed {
+                agent_id: "child-1".to_string(),
+                output: "Found the root cause".to_string(),
+            },
+        }],
+    });
+
+    let content: serde_json::Value = serde_json::from_str(&result.model_content()).unwrap();
+    assert_eq!(content["status"], "launched");
+    assert_eq!(content["completion_state"], "children_completed");
+    assert_eq!(content["agents"][0]["status"], "completed");
+    assert_eq!(content["agents"][0]["output"], "Found the root cause");
+    assert!(result.is_successful());
+    assert!(!result.is_failed());
+}
+
+#[test]
+fn all_failed_run_agents_is_failure_but_mixed_batch_is_successful() {
+    let failed = |name: &str, error: &str| RunAgentsAgentOutcome {
+        name: name.to_string(),
+        kind: RunAgentsAgentOutcomeKind::Failed {
+            error: error.to_string(),
+        },
+    };
+    let result = AIAgentActionResultType::RunAgents(RunAgentsResult::Launched {
+        model_id: "model".to_string(),
+        harness_type: "codex".to_string(),
+        execution_mode: RunAgentsLaunchedExecutionMode::Local,
+        agents: vec![failed("one", "first error"), failed("two", "second error")],
+    });
+
+    let content: serde_json::Value = serde_json::from_str(&result.model_content()).unwrap();
+    assert_eq!(content["status"], "failure");
+    assert_eq!(content["agents"][0]["error"], "first error");
+    assert_eq!(content["agents"][1]["error"], "second error");
+    assert!(result.is_failed());
+    assert!(!result.is_successful());
+
+    let mixed = AIAgentActionResultType::RunAgents(RunAgentsResult::Launched {
+        model_id: "model".to_string(),
+        harness_type: "codex".to_string(),
+        execution_mode: RunAgentsLaunchedExecutionMode::Local,
+        agents: vec![
+            failed("one", "first error"),
+            RunAgentsAgentOutcome {
+                name: "two".to_string(),
+                kind: RunAgentsAgentOutcomeKind::Completed {
+                    agent_id: "child-2".to_string(),
+                    output: "useful output".to_string(),
+                },
+            },
+        ],
+    });
+    let mixed_content: serde_json::Value = serde_json::from_str(&mixed.model_content()).unwrap();
+    assert_eq!(mixed_content["status"], "launched");
+    assert_eq!(mixed_content["agents"][0]["error"], "first error");
+    assert_eq!(mixed_content["agents"][1]["output"], "useful output");
+    assert!(mixed.is_successful());
+    assert!(!mixed.is_failed());
+}

crates/galaxy_agent_core/src/provider_run.rs

@@ -820,7 +820,8 @@ impl ProviderRun {
                     call_id: call.call.id.clone(),
                 })
             }
-            PendingToolCallState::PermissionPending { .. } | PendingToolCallState::Executing => {
+            PendingToolCallState::Executing => Ok(()),
+            PendingToolCallState::PermissionPending { .. } => {
                 Err(invalid_tool_transition(call, "tool start"))
             }
         }
@@ -840,6 +841,9 @@ impl ProviderRun {
                 call.state = PendingToolCallState::Resolved { result };
                 Ok(())
             }
+            PendingToolCallState::Resolved {
+                result: completed_result,
+            } if completed_result == &result => Ok(()),
             PendingToolCallState::Resolved { .. } => {
                 Err(ProviderRunProtocolError::DuplicateToolUpdate {
                     call_id: call.call.id.clone(),

crates/galaxy_agent_core/src/provider_run_tests.rs

@@ -358,6 +358,89 @@ fn parallel_tool_results_commit_atomically_in_original_call_order() {
     assert_eq!(ids, vec!["first", "second"]);
 }
 
+#[test]
+fn duplicate_tool_start_is_idempotent() {
+    let mut run = run();
+    let batch = accept_tool_turn(
+        &mut run,
+        tool_turn(vec![tool_call("read", "read_files")], &["read_files"]),
+    );
+
+    run.start_tool(&batch.work_id, "read").unwrap();
+    let started = run.clone();
+    run.start_tool(&batch.work_id, "read").unwrap();
+
+    assert_eq!(run, started);
+}
+
+#[test]
+fn identical_tool_completion_is_idempotent() {
+    let mut run = run();
+    let batch = accept_tool_turn(
+        &mut run,
+        tool_turn(vec![tool_call("read", "read_files")], &["read_files"]),
+    );
+    let result = successful_result("read", "contents");
+
+    run.complete_tool(&batch.work_id, result.clone()).unwrap();
+    let completed = run.clone();
+    run.complete_tool(&batch.work_id, result).unwrap();
+
+    assert_eq!(run, completed);
+}
+
+#[test]
+fn conflicting_tool_completion_is_rejected_without_mutation() {
+    let mut run = run();
+    let batch = accept_tool_turn(
+        &mut run,
+        tool_turn(vec![tool_call("read", "read_files")], &["read_files"]),
+    );
+    run.complete_tool(&batch.work_id, successful_result("read", "contents"))
+        .unwrap();
+    let completed = run.clone();
+
+    assert_eq!(
+        run.complete_tool(&batch.work_id, successful_result("read", "different"))
+            .unwrap_err(),
+        ProviderRunProtocolError::DuplicateToolUpdate {
+            call_id: "read".to_string(),
+        }
+    );
+    assert_eq!(run, completed);
+}
+
+#[test]
+fn duplicate_tool_callbacks_do_not_prevent_eventual_batch_completion() {
+    let mut run = run();
+    let batch = accept_tool_turn(
+        &mut run,
+        tool_turn(
+            vec![
+                tool_call("first", "read_files"),
+                tool_call("second", "grep"),
+            ],
+            &["read_files", "grep"],
+        ),
+    );
+    let first_result = successful_result("first", "one");
+
+    run.start_tool(&batch.work_id, "first").unwrap();
+    run.start_tool(&batch.work_id, "first").unwrap();
+    run.complete_tool(&batch.work_id, first_result.clone())
+        .unwrap();
+    run.complete_tool(&batch.work_id, first_result).unwrap();
+    run.complete_tool(&batch.work_id, successful_result("second", "two"))
+        .unwrap();
+
+    let ProviderRunState::AwaitingTools { batch: completed } = run.state() else {
+        panic!("expected completed tool batch");
+    };
+    assert!(completed.is_complete());
+    run.commit_tool_batch(&batch.work_id).unwrap();
+    assert_eq!(run.state().phase(), ProviderRunPhase::ReadyToCallModel);
+}
+
 #[test]
 fn exact_batch_submission_rejects_missing_duplicate_and_unknown_results_without_mutation() {
     let mut run = run();