Retry on burn mode (#34669)

Now we only auto-retry if burn mode is enabled. We also show a "Retry"
button (so you don't have to type "continue") if you think that's the
right remedy, and additionally we show a "Retry and Enable Burn Mode"
button if you don't have it enabled.

<img width="484" height="260" alt="Screenshot 2025-07-17 at 6 25 27 PM"
src="https://github.com/user-attachments/assets/dc5bf1f6-8b11-4041-87aa-4f37c95ea9f0"
/>

<img width="478" height="307" alt="Screenshot 2025-07-17 at 6 22 36 PM"
src="https://github.com/user-attachments/assets/1ed6578a-1696-449d-96d1-e447d11959fa"
/>


Release Notes:

- Only auto-retry Agent requests when Burn Mode is enabled

crates/agent/src/thread.rs

@@ -396,6 +396,7 @@ pub struct Thread {
     remaining_turns: u32,
     configured_model: Option<ConfiguredModel>,
     profile: AgentProfile,
+    last_error_context: Option<(Arc<dyn LanguageModel>, CompletionIntent)>,
 }
 
 #[derive(Clone, Debug)]
@@ -489,10 +490,11 @@ impl Thread {
             retry_state: None,
             message_feedback: HashMap::default(),
             last_auto_capture_at: None,
+            last_error_context: None,
             last_received_chunk_at: None,
             request_callback: None,
             remaining_turns: u32::MAX,
-            configured_model,
+            configured_model: configured_model.clone(),
             profile: AgentProfile::new(profile_id, tools),
         }
     }
@@ -613,6 +615,7 @@ impl Thread {
             feedback: None,
             message_feedback: HashMap::default(),
             last_auto_capture_at: None,
+            last_error_context: None,
             last_received_chunk_at: None,
             request_callback: None,
             remaining_turns: u32::MAX,
@@ -1264,9 +1267,58 @@ impl Thread {
 
         self.flush_notifications(model.clone(), intent, cx);
 
-        let request = self.to_completion_request(model.clone(), intent, cx);
+        let _checkpoint = self.finalize_pending_checkpoint(cx);
+        self.stream_completion(
+            self.to_completion_request(model.clone(), intent, cx),
+            model,
+            intent,
+            window,
+            cx,
+        );
+    }
 
-        self.stream_completion(request, model, intent, window, cx);
+    pub fn retry_last_completion(
+        &mut self,
+        window: Option<AnyWindowHandle>,
+        cx: &mut Context<Self>,
+    ) {
+        // Clear any existing error state
+        self.retry_state = None;
+
+        // Use the last error context if available, otherwise fall back to configured model
+        let (model, intent) = if let Some((model, intent)) = self.last_error_context.take() {
+            (model, intent)
+        } else if let Some(configured_model) = self.configured_model.as_ref() {
+            let model = configured_model.model.clone();
+            let intent = if self.has_pending_tool_uses() {
+                CompletionIntent::ToolResults
+            } else {
+                CompletionIntent::UserPrompt
+            };
+            (model, intent)
+        } else if let Some(configured_model) = self.get_or_init_configured_model(cx) {
+            let model = configured_model.model.clone();
+            let intent = if self.has_pending_tool_uses() {
+                CompletionIntent::ToolResults
+            } else {
+                CompletionIntent::UserPrompt
+            };
+            (model, intent)
+        } else {
+            return;
+        };
+
+        self.send_to_model(model, intent, window, cx);
+    }
+
+    pub fn enable_burn_mode_and_retry(
+        &mut self,
+        window: Option<AnyWindowHandle>,
+        cx: &mut Context<Self>,
+    ) {
+        self.completion_mode = CompletionMode::Burn;
+        cx.emit(ThreadEvent::ProfileChanged);
+        self.retry_last_completion(window, cx);
     }
 
     pub fn used_tools_since_last_user_message(&self) -> bool {
@@ -2222,6 +2274,23 @@ impl Thread {
         window: Option<AnyWindowHandle>,
         cx: &mut Context<Self>,
     ) -> bool {
+        // Store context for the Retry button
+        self.last_error_context = Some((model.clone(), intent));
+
+        // Only auto-retry if Burn Mode is enabled
+        if self.completion_mode != CompletionMode::Burn {
+            // Show error with retry options
+            cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
+                message: format!(
+                    "{}\n\nTo automatically retry when similar errors happen, enable Burn Mode.",
+                    error
+                )
+                .into(),
+                can_enable_burn_mode: true,
+            }));
+            return false;
+        }
+
         let Some(strategy) = strategy.or_else(|| Self::get_retry_strategy(error)) else {
             return false;
         };
@@ -2302,6 +2371,13 @@ impl Thread {
             // Stop generating since we're giving up on retrying.
             self.pending_completions.clear();
 
+            // Show error alongside a Retry button, but no
+            // Enable Burn Mode button (since it's already enabled)
+            cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
+                message: format!("Failed after retrying: {}", error).into(),
+                can_enable_burn_mode: false,
+            }));
+
             false
         }
     }
@@ -3212,6 +3288,11 @@ pub enum ThreadError {
         header: SharedString,
         message: SharedString,
     },
+    #[error("Retryable error: {message}")]
+    RetryableError {
+        message: SharedString,
+        can_enable_burn_mode: bool,
+    },
 }
 
 #[derive(Debug, Clone)]
@@ -4167,6 +4248,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // Create model that returns overloaded error
         let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
 
@@ -4240,6 +4326,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // Create model that returns internal server error
         let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
 
@@ -4316,6 +4407,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // Create model that returns internal server error
         let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
 
@@ -4423,6 +4519,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // Create model that returns overloaded error
         let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
 
@@ -4509,6 +4610,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // We'll use a wrapper to switch behavior after first failure
         struct RetryTestModel {
             inner: Arc<FakeLanguageModel>,
@@ -4677,6 +4783,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // Create a model that fails once then succeeds
         struct FailOnceModel {
             inner: Arc<FakeLanguageModel>,
@@ -4838,6 +4949,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // Create a model that returns rate limit error with retry_after
         struct RateLimitModel {
             inner: Arc<FakeLanguageModel>,
@@ -5111,6 +5227,79 @@ fn main() {{
         );
     }
 
+    #[gpui::test]
+    async fn test_no_retry_without_burn_mode(cx: &mut TestAppContext) {
+        init_test_settings(cx);
+
+        let project = create_test_project(cx, json!({})).await;
+        let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
+
+        // Ensure we're in Normal mode (not Burn mode)
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Normal);
+        });
+
+        // Track error events
+        let error_events = Arc::new(Mutex::new(Vec::new()));
+        let error_events_clone = error_events.clone();
+
+        let _subscription = thread.update(cx, |_, cx| {
+            cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
+                if let ThreadEvent::ShowError(error) = event {
+                    error_events_clone.lock().push(error.clone());
+                }
+            })
+        });
+
+        // Create model that returns overloaded error
+        let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
+
+        // Insert a user message
+        thread.update(cx, |thread, cx| {
+            thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
+        });
+
+        // Start completion
+        thread.update(cx, |thread, cx| {
+            thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
+        });
+
+        cx.run_until_parked();
+
+        // Verify no retry state was created
+        thread.read_with(cx, |thread, _| {
+            assert!(
+                thread.retry_state.is_none(),
+                "Should not have retry state in Normal mode"
+            );
+        });
+
+        // Check that a retryable error was reported
+        let errors = error_events.lock();
+        assert!(!errors.is_empty(), "Should have received an error event");
+
+        if let ThreadError::RetryableError {
+            message: _,
+            can_enable_burn_mode,
+        } = &errors[0]
+        {
+            assert!(
+                *can_enable_burn_mode,
+                "Error should indicate burn mode can be enabled"
+            );
+        } else {
+            panic!("Expected RetryableError, got {:?}", errors[0]);
+        }
+
+        // Verify the thread is no longer generating
+        thread.read_with(cx, |thread, _| {
+            assert!(
+                !thread.is_generating(),
+                "Should not be generating after error without retry"
+            );
+        });
+    }
+
     #[gpui::test]
     async fn test_retry_cancelled_on_stop(cx: &mut TestAppContext) {
         init_test_settings(cx);
@@ -5118,6 +5307,11 @@ fn main() {{
         let project = create_test_project(cx, json!({})).await;
         let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
 
+        // Enable Burn Mode to allow retries
+        thread.update(cx, |thread, _| {
+            thread.set_completion_mode(CompletionMode::Burn);
+        });
+
         // Create model that returns overloaded error
         let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
 

crates/agent_ui/src/agent_panel.rs

@@ -65,8 +65,9 @@ use theme::ThemeSettings;
 use time::UtcOffset;
 use ui::utils::WithRemSize;
 use ui::{
-    Banner, Callout, CheckboxWithLabel, ContextMenu, ElevationIndex, KeyBinding, PopoverMenu,
+    Banner, Button, Callout, CheckboxWithLabel, ContextMenu, ElevationIndex, IconPosition,
-    PopoverMenuHandle, ProgressBar, Tab, Tooltip, Vector, VectorName, prelude::*,
+    KeyBinding, PopoverMenu, PopoverMenuHandle, ProgressBar, Tab, Tooltip, Vector, VectorName,
+    prelude::*,
 };
 use util::ResultExt as _;
 use workspace::{
@@ -2977,6 +2978,21 @@ impl AgentPanel {
             .size(IconSize::Small)
             .color(Color::Error);
 
+        let retry_button = Button::new("retry", "Retry")
+            .icon(IconName::RotateCw)
+            .icon_position(IconPosition::Start)
+            .on_click({
+                let thread = thread.clone();
+                move |_, window, cx| {
+                    thread.update(cx, |thread, cx| {
+                        thread.clear_last_error();
+                        thread.thread().update(cx, |thread, cx| {
+                            thread.retry_last_completion(Some(window.window_handle()), cx);
+                        });
+                    });
+                }
+            });
+
         div()
             .border_t_1()
             .border_color(cx.theme().colors().border)
@@ -2985,13 +3001,72 @@ impl AgentPanel {
                     .icon(icon)
                     .title(header)
                     .description(message.clone())
-                    .primary_action(self.dismiss_error_button(thread, cx))
+                    .primary_action(retry_button)
-                    .secondary_action(self.create_copy_button(message_with_header))
+                    .secondary_action(self.dismiss_error_button(thread, cx))
+                    .tertiary_action(self.create_copy_button(message_with_header))
                     .bg_color(self.error_callout_bg(cx)),
             )
             .into_any_element()
     }
 
+    fn render_retryable_error(
+        &self,
+        message: SharedString,
+        can_enable_burn_mode: bool,
+        thread: &Entity<ActiveThread>,
+        cx: &mut Context<Self>,
+    ) -> AnyElement {
+        let icon = Icon::new(IconName::XCircle)
+            .size(IconSize::Small)
+            .color(Color::Error);
+
+        let retry_button = Button::new("retry", "Retry")
+            .icon(IconName::RotateCw)
+            .icon_position(IconPosition::Start)
+            .on_click({
+                let thread = thread.clone();
+                move |_, window, cx| {
+                    thread.update(cx, |thread, cx| {
+                        thread.clear_last_error();
+                        thread.thread().update(cx, |thread, cx| {
+                            thread.retry_last_completion(Some(window.window_handle()), cx);
+                        });
+                    });
+                }
+            });
+
+        let mut callout = Callout::new()
+            .icon(icon)
+            .title("Error")
+            .description(message.clone())
+            .bg_color(self.error_callout_bg(cx))
+            .primary_action(retry_button);
+
+        if can_enable_burn_mode {
+            let burn_mode_button = Button::new("enable_burn_retry", "Enable Burn Mode and Retry")
+                .icon(IconName::ZedBurnMode)
+                .icon_position(IconPosition::Start)
+                .on_click({
+                    let thread = thread.clone();
+                    move |_, window, cx| {
+                        thread.update(cx, |thread, cx| {
+                            thread.clear_last_error();
+                            thread.thread().update(cx, |thread, cx| {
+                                thread.enable_burn_mode_and_retry(Some(window.window_handle()), cx);
+                            });
+                        });
+                    }
+                });
+            callout = callout.secondary_action(burn_mode_button);
+        }
+
+        div()
+            .border_t_1()
+            .border_color(cx.theme().colors().border)
+            .child(callout)
+            .into_any_element()
+    }
+
     fn render_prompt_editor(
         &self,
         context_editor: &Entity<TextThreadEditor>,
@@ -3233,6 +3308,15 @@ impl Render for AgentPanel {
                                     ThreadError::Message { header, message } => {
                                         self.render_error_message(header, message, thread, cx)
                                     }
+                                    ThreadError::RetryableError {
+                                        message,
+                                        can_enable_burn_mode,
+                                    } => self.render_retryable_error(
+                                        message,
+                                        can_enable_burn_mode,
+                                        thread,
+                                        cx,
+                                    ),
                                 })
                                 .into_any(),
                         )