1use crate::{
2 agent_profile::AgentProfile,
3 context::{AgentContext, AgentContextHandle, ContextLoadResult, LoadedContext},
4 thread_store::{
5 SerializedCrease, SerializedLanguageModel, SerializedMessage, SerializedMessageSegment,
6 SerializedThread, SerializedToolResult, SerializedToolUse, SharedProjectContext,
7 ThreadStore,
8 },
9 tool_use::{PendingToolUse, ToolUse, ToolUseMetadata, ToolUseState},
10};
11use action_log::ActionLog;
12use agent_settings::{AgentProfileId, AgentSettings, CompletionMode, SUMMARIZE_THREAD_PROMPT};
13use anyhow::{Result, anyhow};
14use assistant_tool::{AnyToolCard, Tool, ToolWorkingSet};
15use chrono::{DateTime, Utc};
16use client::{ModelRequestUsage, RequestUsage};
17use cloud_llm_client::{CompletionIntent, CompletionRequestStatus, Plan, UsageLimit};
18use collections::HashMap;
19use futures::{FutureExt, StreamExt as _, future::Shared};
20use git::repository::DiffType;
21use gpui::{
22 AnyWindowHandle, App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task,
23 WeakEntity, Window,
24};
25use http_client::StatusCode;
26use language_model::{
27 ConfiguredModel, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,
28 LanguageModelExt as _, LanguageModelId, LanguageModelRegistry, LanguageModelRequest,
29 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
30 LanguageModelToolResultContent, LanguageModelToolUse, LanguageModelToolUseId, MessageContent,
31 ModelRequestLimitReachedError, PaymentRequiredError, Role, SelectedModel, StopReason,
32 TokenUsage,
33};
34use postage::stream::Stream as _;
35use project::{
36 Project,
37 git_store::{GitStore, GitStoreCheckpoint, RepositoryState},
38};
39use prompt_store::{ModelContext, PromptBuilder};
40use schemars::JsonSchema;
41use serde::{Deserialize, Serialize};
42use settings::Settings;
43use std::{
44 io::Write,
45 ops::Range,
46 sync::Arc,
47 time::{Duration, Instant},
48};
49use thiserror::Error;
50use util::{ResultExt as _, post_inc};
51use uuid::Uuid;
52
53const MAX_RETRY_ATTEMPTS: u8 = 4;
54const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
55
56#[derive(Debug, Clone)]
57enum RetryStrategy {
58 ExponentialBackoff {
59 initial_delay: Duration,
60 max_attempts: u8,
61 },
62 Fixed {
63 delay: Duration,
64 max_attempts: u8,
65 },
66}
67
68#[derive(
69 Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize, JsonSchema,
70)]
71pub struct ThreadId(Arc<str>);
72
73impl ThreadId {
74 pub fn new() -> Self {
75 Self(Uuid::new_v4().to_string().into())
76 }
77}
78
79impl std::fmt::Display for ThreadId {
80 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
81 write!(f, "{}", self.0)
82 }
83}
84
85impl From<&str> for ThreadId {
86 fn from(value: &str) -> Self {
87 Self(value.into())
88 }
89}
90
91/// The ID of the user prompt that initiated a request.
92///
93/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
94#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
95pub struct PromptId(Arc<str>);
96
97impl PromptId {
98 pub fn new() -> Self {
99 Self(Uuid::new_v4().to_string().into())
100 }
101}
102
103impl std::fmt::Display for PromptId {
104 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
105 write!(f, "{}", self.0)
106 }
107}
108
109#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Copy, Serialize, Deserialize)]
110pub struct MessageId(pub(crate) usize);
111
112impl MessageId {
113 fn post_inc(&mut self) -> Self {
114 Self(post_inc(&mut self.0))
115 }
116
117 pub fn as_usize(&self) -> usize {
118 self.0
119 }
120}
121
122/// Stored information that can be used to resurrect a context crease when creating an editor for a past message.
123#[derive(Clone, Debug)]
124pub struct MessageCrease {
125 pub range: Range<usize>,
126 pub icon_path: SharedString,
127 pub label: SharedString,
128 /// None for a deserialized message, Some otherwise.
129 pub context: Option<AgentContextHandle>,
130}
131
132/// A message in a [`Thread`].
133#[derive(Debug, Clone)]
134pub struct Message {
135 pub id: MessageId,
136 pub role: Role,
137 pub segments: Vec<MessageSegment>,
138 pub loaded_context: LoadedContext,
139 pub creases: Vec<MessageCrease>,
140 pub is_hidden: bool,
141 pub ui_only: bool,
142}
143
144impl Message {
145 /// Returns whether the message contains any meaningful text that should be displayed
146 /// The model sometimes runs tool without producing any text or just a marker ([`USING_TOOL_MARKER`])
147 pub fn should_display_content(&self) -> bool {
148 self.segments.iter().all(|segment| segment.should_display())
149 }
150
151 pub fn push_thinking(&mut self, text: &str, signature: Option<String>) {
152 if let Some(MessageSegment::Thinking {
153 text: segment,
154 signature: current_signature,
155 }) = self.segments.last_mut()
156 {
157 if let Some(signature) = signature {
158 *current_signature = Some(signature);
159 }
160 segment.push_str(text);
161 } else {
162 self.segments.push(MessageSegment::Thinking {
163 text: text.to_string(),
164 signature,
165 });
166 }
167 }
168
169 pub fn push_redacted_thinking(&mut self, data: String) {
170 self.segments.push(MessageSegment::RedactedThinking(data));
171 }
172
173 pub fn push_text(&mut self, text: &str) {
174 if let Some(MessageSegment::Text(segment)) = self.segments.last_mut() {
175 segment.push_str(text);
176 } else {
177 self.segments.push(MessageSegment::Text(text.to_string()));
178 }
179 }
180
181 pub fn to_string(&self) -> String {
182 let mut result = String::new();
183
184 if !self.loaded_context.text.is_empty() {
185 result.push_str(&self.loaded_context.text);
186 }
187
188 for segment in &self.segments {
189 match segment {
190 MessageSegment::Text(text) => result.push_str(text),
191 MessageSegment::Thinking { text, .. } => {
192 result.push_str("<think>\n");
193 result.push_str(text);
194 result.push_str("\n</think>");
195 }
196 MessageSegment::RedactedThinking(_) => {}
197 }
198 }
199
200 result
201 }
202}
203
204#[derive(Debug, Clone, PartialEq, Eq)]
205pub enum MessageSegment {
206 Text(String),
207 Thinking {
208 text: String,
209 signature: Option<String>,
210 },
211 RedactedThinking(String),
212}
213
214impl MessageSegment {
215 pub fn should_display(&self) -> bool {
216 match self {
217 Self::Text(text) => text.is_empty(),
218 Self::Thinking { text, .. } => text.is_empty(),
219 Self::RedactedThinking(_) => false,
220 }
221 }
222
223 pub fn text(&self) -> Option<&str> {
224 match self {
225 MessageSegment::Text(text) => Some(text),
226 _ => None,
227 }
228 }
229}
230
231#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
232pub struct ProjectSnapshot {
233 pub worktree_snapshots: Vec<WorktreeSnapshot>,
234 pub unsaved_buffer_paths: Vec<String>,
235 pub timestamp: DateTime<Utc>,
236}
237
238#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
239pub struct WorktreeSnapshot {
240 pub worktree_path: String,
241 pub git_state: Option<GitState>,
242}
243
244#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
245pub struct GitState {
246 pub remote_url: Option<String>,
247 pub head_sha: Option<String>,
248 pub current_branch: Option<String>,
249 pub diff: Option<String>,
250}
251
252#[derive(Clone, Debug)]
253pub struct ThreadCheckpoint {
254 message_id: MessageId,
255 git_checkpoint: GitStoreCheckpoint,
256}
257
258#[derive(Copy, Clone, Debug, PartialEq, Eq)]
259pub enum ThreadFeedback {
260 Positive,
261 Negative,
262}
263
264pub enum LastRestoreCheckpoint {
265 Pending {
266 message_id: MessageId,
267 },
268 Error {
269 message_id: MessageId,
270 error: String,
271 },
272}
273
274impl LastRestoreCheckpoint {
275 pub fn message_id(&self) -> MessageId {
276 match self {
277 LastRestoreCheckpoint::Pending { message_id } => *message_id,
278 LastRestoreCheckpoint::Error { message_id, .. } => *message_id,
279 }
280 }
281}
282
283#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
284pub enum DetailedSummaryState {
285 #[default]
286 NotGenerated,
287 Generating {
288 message_id: MessageId,
289 },
290 Generated {
291 text: SharedString,
292 message_id: MessageId,
293 },
294}
295
296impl DetailedSummaryState {
297 fn text(&self) -> Option<SharedString> {
298 if let Self::Generated { text, .. } = self {
299 Some(text.clone())
300 } else {
301 None
302 }
303 }
304}
305
306#[derive(Default, Debug)]
307pub struct TotalTokenUsage {
308 pub total: u64,
309 pub max: u64,
310}
311
312impl TotalTokenUsage {
313 pub fn ratio(&self) -> TokenUsageRatio {
314 #[cfg(debug_assertions)]
315 let warning_threshold: f32 = std::env::var("ZED_THREAD_WARNING_THRESHOLD")
316 .unwrap_or("0.8".to_string())
317 .parse()
318 .unwrap();
319 #[cfg(not(debug_assertions))]
320 let warning_threshold: f32 = 0.8;
321
322 // When the maximum is unknown because there is no selected model,
323 // avoid showing the token limit warning.
324 if self.max == 0 {
325 TokenUsageRatio::Normal
326 } else if self.total >= self.max {
327 TokenUsageRatio::Exceeded
328 } else if self.total as f32 / self.max as f32 >= warning_threshold {
329 TokenUsageRatio::Warning
330 } else {
331 TokenUsageRatio::Normal
332 }
333 }
334
335 pub fn add(&self, tokens: u64) -> TotalTokenUsage {
336 TotalTokenUsage {
337 total: self.total + tokens,
338 max: self.max,
339 }
340 }
341}
342
343#[derive(Debug, Default, PartialEq, Eq)]
344pub enum TokenUsageRatio {
345 #[default]
346 Normal,
347 Warning,
348 Exceeded,
349}
350
351#[derive(Debug, Clone, Copy)]
352pub enum QueueState {
353 Sending,
354 Queued { position: usize },
355 Started,
356}
357
358/// A thread of conversation with the LLM.
359pub struct Thread {
360 id: ThreadId,
361 updated_at: DateTime<Utc>,
362 summary: ThreadSummary,
363 pending_summary: Task<Option<()>>,
364 detailed_summary_task: Task<Option<()>>,
365 detailed_summary_tx: postage::watch::Sender<DetailedSummaryState>,
366 detailed_summary_rx: postage::watch::Receiver<DetailedSummaryState>,
367 completion_mode: agent_settings::CompletionMode,
368 messages: Vec<Message>,
369 next_message_id: MessageId,
370 last_prompt_id: PromptId,
371 project_context: SharedProjectContext,
372 checkpoints_by_message: HashMap<MessageId, ThreadCheckpoint>,
373 completion_count: usize,
374 pending_completions: Vec<PendingCompletion>,
375 project: Entity<Project>,
376 prompt_builder: Arc<PromptBuilder>,
377 tools: Entity<ToolWorkingSet>,
378 tool_use: ToolUseState,
379 action_log: Entity<ActionLog>,
380 last_restore_checkpoint: Option<LastRestoreCheckpoint>,
381 pending_checkpoint: Option<ThreadCheckpoint>,
382 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
383 request_token_usage: Vec<TokenUsage>,
384 cumulative_token_usage: TokenUsage,
385 exceeded_window_error: Option<ExceededWindowError>,
386 tool_use_limit_reached: bool,
387 feedback: Option<ThreadFeedback>,
388 retry_state: Option<RetryState>,
389 message_feedback: HashMap<MessageId, ThreadFeedback>,
390 last_received_chunk_at: Option<Instant>,
391 request_callback: Option<
392 Box<dyn FnMut(&LanguageModelRequest, &[Result<LanguageModelCompletionEvent, String>])>,
393 >,
394 remaining_turns: u32,
395 configured_model: Option<ConfiguredModel>,
396 profile: AgentProfile,
397 last_error_context: Option<(Arc<dyn LanguageModel>, CompletionIntent)>,
398}
399
400#[derive(Clone, Debug)]
401struct RetryState {
402 attempt: u8,
403 max_attempts: u8,
404 intent: CompletionIntent,
405}
406
407#[derive(Clone, Debug, PartialEq, Eq)]
408pub enum ThreadSummary {
409 Pending,
410 Generating,
411 Ready(SharedString),
412 Error,
413}
414
415impl ThreadSummary {
416 pub const DEFAULT: SharedString = SharedString::new_static("New Thread");
417
418 pub fn or_default(&self) -> SharedString {
419 self.unwrap_or(Self::DEFAULT)
420 }
421
422 pub fn unwrap_or(&self, message: impl Into<SharedString>) -> SharedString {
423 self.ready().unwrap_or_else(|| message.into())
424 }
425
426 pub fn ready(&self) -> Option<SharedString> {
427 match self {
428 ThreadSummary::Ready(summary) => Some(summary.clone()),
429 ThreadSummary::Pending | ThreadSummary::Generating | ThreadSummary::Error => None,
430 }
431 }
432}
433
434#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
435pub struct ExceededWindowError {
436 /// Model used when last message exceeded context window
437 model_id: LanguageModelId,
438 /// Token count including last message
439 token_count: u64,
440}
441
442impl Thread {
443 pub fn new(
444 project: Entity<Project>,
445 tools: Entity<ToolWorkingSet>,
446 prompt_builder: Arc<PromptBuilder>,
447 system_prompt: SharedProjectContext,
448 cx: &mut Context<Self>,
449 ) -> Self {
450 let (detailed_summary_tx, detailed_summary_rx) = postage::watch::channel();
451 let configured_model = LanguageModelRegistry::read_global(cx).default_model();
452 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
453
454 Self {
455 id: ThreadId::new(),
456 updated_at: Utc::now(),
457 summary: ThreadSummary::Pending,
458 pending_summary: Task::ready(None),
459 detailed_summary_task: Task::ready(None),
460 detailed_summary_tx,
461 detailed_summary_rx,
462 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
463 messages: Vec::new(),
464 next_message_id: MessageId(0),
465 last_prompt_id: PromptId::new(),
466 project_context: system_prompt,
467 checkpoints_by_message: HashMap::default(),
468 completion_count: 0,
469 pending_completions: Vec::new(),
470 project: project.clone(),
471 prompt_builder,
472 tools: tools.clone(),
473 last_restore_checkpoint: None,
474 pending_checkpoint: None,
475 tool_use: ToolUseState::new(tools.clone()),
476 action_log: cx.new(|_| ActionLog::new(project.clone())),
477 initial_project_snapshot: {
478 let project_snapshot = Self::project_snapshot(project, cx);
479 cx.foreground_executor()
480 .spawn(async move { Some(project_snapshot.await) })
481 .shared()
482 },
483 request_token_usage: Vec::new(),
484 cumulative_token_usage: TokenUsage::default(),
485 exceeded_window_error: None,
486 tool_use_limit_reached: false,
487 feedback: None,
488 retry_state: None,
489 message_feedback: HashMap::default(),
490 last_error_context: None,
491 last_received_chunk_at: None,
492 request_callback: None,
493 remaining_turns: u32::MAX,
494 configured_model: configured_model.clone(),
495 profile: AgentProfile::new(profile_id, tools),
496 }
497 }
498
499 pub fn deserialize(
500 id: ThreadId,
501 serialized: SerializedThread,
502 project: Entity<Project>,
503 tools: Entity<ToolWorkingSet>,
504 prompt_builder: Arc<PromptBuilder>,
505 project_context: SharedProjectContext,
506 window: Option<&mut Window>, // None in headless mode
507 cx: &mut Context<Self>,
508 ) -> Self {
509 let next_message_id = MessageId(
510 serialized
511 .messages
512 .last()
513 .map(|message| message.id.0 + 1)
514 .unwrap_or(0),
515 );
516 let tool_use = ToolUseState::from_serialized_messages(
517 tools.clone(),
518 &serialized.messages,
519 project.clone(),
520 window,
521 cx,
522 );
523 let (detailed_summary_tx, detailed_summary_rx) =
524 postage::watch::channel_with(serialized.detailed_summary_state);
525
526 let configured_model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
527 serialized
528 .model
529 .and_then(|model| {
530 let model = SelectedModel {
531 provider: model.provider.clone().into(),
532 model: model.model.clone().into(),
533 };
534 registry.select_model(&model, cx)
535 })
536 .or_else(|| registry.default_model())
537 });
538
539 let completion_mode = serialized
540 .completion_mode
541 .unwrap_or_else(|| AgentSettings::get_global(cx).preferred_completion_mode);
542 let profile_id = serialized
543 .profile
544 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
545
546 Self {
547 id,
548 updated_at: serialized.updated_at,
549 summary: ThreadSummary::Ready(serialized.summary),
550 pending_summary: Task::ready(None),
551 detailed_summary_task: Task::ready(None),
552 detailed_summary_tx,
553 detailed_summary_rx,
554 completion_mode,
555 retry_state: None,
556 messages: serialized
557 .messages
558 .into_iter()
559 .map(|message| Message {
560 id: message.id,
561 role: message.role,
562 segments: message
563 .segments
564 .into_iter()
565 .map(|segment| match segment {
566 SerializedMessageSegment::Text { text } => MessageSegment::Text(text),
567 SerializedMessageSegment::Thinking { text, signature } => {
568 MessageSegment::Thinking { text, signature }
569 }
570 SerializedMessageSegment::RedactedThinking { data } => {
571 MessageSegment::RedactedThinking(data)
572 }
573 })
574 .collect(),
575 loaded_context: LoadedContext {
576 contexts: Vec::new(),
577 text: message.context,
578 images: Vec::new(),
579 },
580 creases: message
581 .creases
582 .into_iter()
583 .map(|crease| MessageCrease {
584 range: crease.start..crease.end,
585 icon_path: crease.icon_path,
586 label: crease.label,
587 context: None,
588 })
589 .collect(),
590 is_hidden: message.is_hidden,
591 ui_only: false, // UI-only messages are not persisted
592 })
593 .collect(),
594 next_message_id,
595 last_prompt_id: PromptId::new(),
596 project_context,
597 checkpoints_by_message: HashMap::default(),
598 completion_count: 0,
599 pending_completions: Vec::new(),
600 last_restore_checkpoint: None,
601 pending_checkpoint: None,
602 project: project.clone(),
603 prompt_builder,
604 tools: tools.clone(),
605 tool_use,
606 action_log: cx.new(|_| ActionLog::new(project)),
607 initial_project_snapshot: Task::ready(serialized.initial_project_snapshot).shared(),
608 request_token_usage: serialized.request_token_usage,
609 cumulative_token_usage: serialized.cumulative_token_usage,
610 exceeded_window_error: None,
611 tool_use_limit_reached: serialized.tool_use_limit_reached,
612 feedback: None,
613 message_feedback: HashMap::default(),
614 last_error_context: None,
615 last_received_chunk_at: None,
616 request_callback: None,
617 remaining_turns: u32::MAX,
618 configured_model,
619 profile: AgentProfile::new(profile_id, tools),
620 }
621 }
622
623 pub fn set_request_callback(
624 &mut self,
625 callback: impl 'static
626 + FnMut(&LanguageModelRequest, &[Result<LanguageModelCompletionEvent, String>]),
627 ) {
628 self.request_callback = Some(Box::new(callback));
629 }
630
631 pub fn id(&self) -> &ThreadId {
632 &self.id
633 }
634
635 pub fn profile(&self) -> &AgentProfile {
636 &self.profile
637 }
638
639 pub fn set_profile(&mut self, id: AgentProfileId, cx: &mut Context<Self>) {
640 if &id != self.profile.id() {
641 self.profile = AgentProfile::new(id, self.tools.clone());
642 cx.emit(ThreadEvent::ProfileChanged);
643 }
644 }
645
646 pub fn is_empty(&self) -> bool {
647 self.messages.is_empty()
648 }
649
650 pub fn updated_at(&self) -> DateTime<Utc> {
651 self.updated_at
652 }
653
654 pub fn touch_updated_at(&mut self) {
655 self.updated_at = Utc::now();
656 }
657
658 pub fn advance_prompt_id(&mut self) {
659 self.last_prompt_id = PromptId::new();
660 }
661
662 pub fn project_context(&self) -> SharedProjectContext {
663 self.project_context.clone()
664 }
665
666 pub fn get_or_init_configured_model(&mut self, cx: &App) -> Option<ConfiguredModel> {
667 if self.configured_model.is_none() {
668 self.configured_model = LanguageModelRegistry::read_global(cx).default_model();
669 }
670 self.configured_model.clone()
671 }
672
673 pub fn configured_model(&self) -> Option<ConfiguredModel> {
674 self.configured_model.clone()
675 }
676
677 pub fn set_configured_model(&mut self, model: Option<ConfiguredModel>, cx: &mut Context<Self>) {
678 self.configured_model = model;
679 cx.notify();
680 }
681
682 pub fn summary(&self) -> &ThreadSummary {
683 &self.summary
684 }
685
686 pub fn set_summary(&mut self, new_summary: impl Into<SharedString>, cx: &mut Context<Self>) {
687 let current_summary = match &self.summary {
688 ThreadSummary::Pending | ThreadSummary::Generating => return,
689 ThreadSummary::Ready(summary) => summary,
690 ThreadSummary::Error => &ThreadSummary::DEFAULT,
691 };
692
693 let mut new_summary = new_summary.into();
694
695 if new_summary.is_empty() {
696 new_summary = ThreadSummary::DEFAULT;
697 }
698
699 if current_summary != &new_summary {
700 self.summary = ThreadSummary::Ready(new_summary);
701 cx.emit(ThreadEvent::SummaryChanged);
702 }
703 }
704
705 pub fn completion_mode(&self) -> CompletionMode {
706 self.completion_mode
707 }
708
709 pub fn set_completion_mode(&mut self, mode: CompletionMode) {
710 self.completion_mode = mode;
711 }
712
713 pub fn message(&self, id: MessageId) -> Option<&Message> {
714 let index = self
715 .messages
716 .binary_search_by(|message| message.id.cmp(&id))
717 .ok()?;
718
719 self.messages.get(index)
720 }
721
722 pub fn messages(&self) -> impl ExactSizeIterator<Item = &Message> {
723 self.messages.iter()
724 }
725
726 pub fn is_generating(&self) -> bool {
727 !self.pending_completions.is_empty() || !self.all_tools_finished()
728 }
729
730 /// Indicates whether streaming of language model events is stale.
731 /// When `is_generating()` is false, this method returns `None`.
732 pub fn is_generation_stale(&self) -> Option<bool> {
733 const STALE_THRESHOLD: u128 = 250;
734
735 self.last_received_chunk_at
736 .map(|instant| instant.elapsed().as_millis() > STALE_THRESHOLD)
737 }
738
739 fn received_chunk(&mut self) {
740 self.last_received_chunk_at = Some(Instant::now());
741 }
742
743 pub fn queue_state(&self) -> Option<QueueState> {
744 self.pending_completions
745 .first()
746 .map(|pending_completion| pending_completion.queue_state)
747 }
748
749 pub fn tools(&self) -> &Entity<ToolWorkingSet> {
750 &self.tools
751 }
752
753 pub fn pending_tool(&self, id: &LanguageModelToolUseId) -> Option<&PendingToolUse> {
754 self.tool_use
755 .pending_tool_uses()
756 .into_iter()
757 .find(|tool_use| &tool_use.id == id)
758 }
759
760 pub fn tools_needing_confirmation(&self) -> impl Iterator<Item = &PendingToolUse> {
761 self.tool_use
762 .pending_tool_uses()
763 .into_iter()
764 .filter(|tool_use| tool_use.status.needs_confirmation())
765 }
766
767 pub fn has_pending_tool_uses(&self) -> bool {
768 !self.tool_use.pending_tool_uses().is_empty()
769 }
770
771 pub fn checkpoint_for_message(&self, id: MessageId) -> Option<ThreadCheckpoint> {
772 self.checkpoints_by_message.get(&id).cloned()
773 }
774
775 pub fn restore_checkpoint(
776 &mut self,
777 checkpoint: ThreadCheckpoint,
778 cx: &mut Context<Self>,
779 ) -> Task<Result<()>> {
780 self.last_restore_checkpoint = Some(LastRestoreCheckpoint::Pending {
781 message_id: checkpoint.message_id,
782 });
783 cx.emit(ThreadEvent::CheckpointChanged);
784 cx.notify();
785
786 let git_store = self.project().read(cx).git_store().clone();
787 let restore = git_store.update(cx, |git_store, cx| {
788 git_store.restore_checkpoint(checkpoint.git_checkpoint.clone(), cx)
789 });
790
791 cx.spawn(async move |this, cx| {
792 let result = restore.await;
793 this.update(cx, |this, cx| {
794 if let Err(err) = result.as_ref() {
795 this.last_restore_checkpoint = Some(LastRestoreCheckpoint::Error {
796 message_id: checkpoint.message_id,
797 error: err.to_string(),
798 });
799 } else {
800 this.truncate(checkpoint.message_id, cx);
801 this.last_restore_checkpoint = None;
802 }
803 this.pending_checkpoint = None;
804 cx.emit(ThreadEvent::CheckpointChanged);
805 cx.notify();
806 })?;
807 result
808 })
809 }
810
811 fn finalize_pending_checkpoint(&mut self, cx: &mut Context<Self>) {
812 let pending_checkpoint = if self.is_generating() {
813 return;
814 } else if let Some(checkpoint) = self.pending_checkpoint.take() {
815 checkpoint
816 } else {
817 return;
818 };
819
820 self.finalize_checkpoint(pending_checkpoint, cx);
821 }
822
823 fn finalize_checkpoint(
824 &mut self,
825 pending_checkpoint: ThreadCheckpoint,
826 cx: &mut Context<Self>,
827 ) {
828 let git_store = self.project.read(cx).git_store().clone();
829 let final_checkpoint = git_store.update(cx, |git_store, cx| git_store.checkpoint(cx));
830 cx.spawn(async move |this, cx| match final_checkpoint.await {
831 Ok(final_checkpoint) => {
832 let equal = git_store
833 .update(cx, |store, cx| {
834 store.compare_checkpoints(
835 pending_checkpoint.git_checkpoint.clone(),
836 final_checkpoint.clone(),
837 cx,
838 )
839 })?
840 .await
841 .unwrap_or(false);
842
843 this.update(cx, |this, cx| {
844 this.pending_checkpoint = if equal {
845 Some(pending_checkpoint)
846 } else {
847 this.insert_checkpoint(pending_checkpoint, cx);
848 Some(ThreadCheckpoint {
849 message_id: this.next_message_id,
850 git_checkpoint: final_checkpoint,
851 })
852 }
853 })?;
854
855 Ok(())
856 }
857 Err(_) => this.update(cx, |this, cx| {
858 this.insert_checkpoint(pending_checkpoint, cx)
859 }),
860 })
861 .detach();
862 }
863
864 fn insert_checkpoint(&mut self, checkpoint: ThreadCheckpoint, cx: &mut Context<Self>) {
865 self.checkpoints_by_message
866 .insert(checkpoint.message_id, checkpoint);
867 cx.emit(ThreadEvent::CheckpointChanged);
868 cx.notify();
869 }
870
871 pub fn last_restore_checkpoint(&self) -> Option<&LastRestoreCheckpoint> {
872 self.last_restore_checkpoint.as_ref()
873 }
874
875 pub fn truncate(&mut self, message_id: MessageId, cx: &mut Context<Self>) {
876 let Some(message_ix) = self
877 .messages
878 .iter()
879 .rposition(|message| message.id == message_id)
880 else {
881 return;
882 };
883 for deleted_message in self.messages.drain(message_ix..) {
884 self.checkpoints_by_message.remove(&deleted_message.id);
885 }
886 cx.notify();
887 }
888
889 pub fn context_for_message(&self, id: MessageId) -> impl Iterator<Item = &AgentContext> {
890 self.messages
891 .iter()
892 .find(|message| message.id == id)
893 .into_iter()
894 .flat_map(|message| message.loaded_context.contexts.iter())
895 }
896
897 pub fn is_turn_end(&self, ix: usize) -> bool {
898 if self.messages.is_empty() {
899 return false;
900 }
901
902 if !self.is_generating() && ix == self.messages.len() - 1 {
903 return true;
904 }
905
906 let Some(message) = self.messages.get(ix) else {
907 return false;
908 };
909
910 if message.role != Role::Assistant {
911 return false;
912 }
913
914 self.messages
915 .get(ix + 1)
916 .and_then(|message| {
917 self.message(message.id)
918 .map(|next_message| next_message.role == Role::User && !next_message.is_hidden)
919 })
920 .unwrap_or(false)
921 }
922
923 pub fn tool_use_limit_reached(&self) -> bool {
924 self.tool_use_limit_reached
925 }
926
927 /// Returns whether all of the tool uses have finished running.
928 pub fn all_tools_finished(&self) -> bool {
929 // If the only pending tool uses left are the ones with errors, then
930 // that means that we've finished running all of the pending tools.
931 self.tool_use
932 .pending_tool_uses()
933 .iter()
934 .all(|pending_tool_use| pending_tool_use.status.is_error())
935 }
936
937 /// Returns whether any pending tool uses may perform edits
938 pub fn has_pending_edit_tool_uses(&self) -> bool {
939 self.tool_use
940 .pending_tool_uses()
941 .iter()
942 .filter(|pending_tool_use| !pending_tool_use.status.is_error())
943 .any(|pending_tool_use| pending_tool_use.may_perform_edits)
944 }
945
946 pub fn tool_uses_for_message(&self, id: MessageId, cx: &App) -> Vec<ToolUse> {
947 self.tool_use.tool_uses_for_message(id, &self.project, cx)
948 }
949
950 pub fn tool_results_for_message(
951 &self,
952 assistant_message_id: MessageId,
953 ) -> Vec<&LanguageModelToolResult> {
954 self.tool_use.tool_results_for_message(assistant_message_id)
955 }
956
957 pub fn tool_result(&self, id: &LanguageModelToolUseId) -> Option<&LanguageModelToolResult> {
958 self.tool_use.tool_result(id)
959 }
960
961 pub fn output_for_tool(&self, id: &LanguageModelToolUseId) -> Option<&Arc<str>> {
962 match &self.tool_use.tool_result(id)?.content {
963 LanguageModelToolResultContent::Text(text) => Some(text),
964 LanguageModelToolResultContent::Image(_) => {
965 // TODO: We should display image
966 None
967 }
968 }
969 }
970
971 pub fn card_for_tool(&self, id: &LanguageModelToolUseId) -> Option<AnyToolCard> {
972 self.tool_use.tool_result_card(id).cloned()
973 }
974
975 /// Return tools that are both enabled and supported by the model
976 pub fn available_tools(
977 &self,
978 cx: &App,
979 model: Arc<dyn LanguageModel>,
980 ) -> Vec<LanguageModelRequestTool> {
981 if model.supports_tools() {
982 self.profile
983 .enabled_tools(cx)
984 .into_iter()
985 .filter_map(|(name, tool)| {
986 // Skip tools that cannot be supported
987 let input_schema = tool.input_schema(model.tool_input_format()).ok()?;
988 Some(LanguageModelRequestTool {
989 name: name.into(),
990 description: tool.description(),
991 input_schema,
992 })
993 })
994 .collect()
995 } else {
996 Vec::default()
997 }
998 }
999
1000 pub fn insert_user_message(
1001 &mut self,
1002 text: impl Into<String>,
1003 loaded_context: ContextLoadResult,
1004 git_checkpoint: Option<GitStoreCheckpoint>,
1005 creases: Vec<MessageCrease>,
1006 cx: &mut Context<Self>,
1007 ) -> MessageId {
1008 if !loaded_context.referenced_buffers.is_empty() {
1009 self.action_log.update(cx, |log, cx| {
1010 for buffer in loaded_context.referenced_buffers {
1011 log.buffer_read(buffer, cx);
1012 }
1013 });
1014 }
1015
1016 let message_id = self.insert_message(
1017 Role::User,
1018 vec![MessageSegment::Text(text.into())],
1019 loaded_context.loaded_context,
1020 creases,
1021 false,
1022 cx,
1023 );
1024
1025 if let Some(git_checkpoint) = git_checkpoint {
1026 self.pending_checkpoint = Some(ThreadCheckpoint {
1027 message_id,
1028 git_checkpoint,
1029 });
1030 }
1031
1032 message_id
1033 }
1034
1035 pub fn insert_invisible_continue_message(&mut self, cx: &mut Context<Self>) -> MessageId {
1036 let id = self.insert_message(
1037 Role::User,
1038 vec![MessageSegment::Text("Continue where you left off".into())],
1039 LoadedContext::default(),
1040 vec![],
1041 true,
1042 cx,
1043 );
1044 self.pending_checkpoint = None;
1045
1046 id
1047 }
1048
1049 pub fn insert_assistant_message(
1050 &mut self,
1051 segments: Vec<MessageSegment>,
1052 cx: &mut Context<Self>,
1053 ) -> MessageId {
1054 self.insert_message(
1055 Role::Assistant,
1056 segments,
1057 LoadedContext::default(),
1058 Vec::new(),
1059 false,
1060 cx,
1061 )
1062 }
1063
1064 pub fn insert_message(
1065 &mut self,
1066 role: Role,
1067 segments: Vec<MessageSegment>,
1068 loaded_context: LoadedContext,
1069 creases: Vec<MessageCrease>,
1070 is_hidden: bool,
1071 cx: &mut Context<Self>,
1072 ) -> MessageId {
1073 let id = self.next_message_id.post_inc();
1074 self.messages.push(Message {
1075 id,
1076 role,
1077 segments,
1078 loaded_context,
1079 creases,
1080 is_hidden,
1081 ui_only: false,
1082 });
1083 self.touch_updated_at();
1084 cx.emit(ThreadEvent::MessageAdded(id));
1085 id
1086 }
1087
1088 pub fn edit_message(
1089 &mut self,
1090 id: MessageId,
1091 new_role: Role,
1092 new_segments: Vec<MessageSegment>,
1093 creases: Vec<MessageCrease>,
1094 loaded_context: Option<LoadedContext>,
1095 checkpoint: Option<GitStoreCheckpoint>,
1096 cx: &mut Context<Self>,
1097 ) -> bool {
1098 let Some(message) = self.messages.iter_mut().find(|message| message.id == id) else {
1099 return false;
1100 };
1101 message.role = new_role;
1102 message.segments = new_segments;
1103 message.creases = creases;
1104 if let Some(context) = loaded_context {
1105 message.loaded_context = context;
1106 }
1107 if let Some(git_checkpoint) = checkpoint {
1108 self.checkpoints_by_message.insert(
1109 id,
1110 ThreadCheckpoint {
1111 message_id: id,
1112 git_checkpoint,
1113 },
1114 );
1115 }
1116 self.touch_updated_at();
1117 cx.emit(ThreadEvent::MessageEdited(id));
1118 true
1119 }
1120
1121 pub fn delete_message(&mut self, id: MessageId, cx: &mut Context<Self>) -> bool {
1122 let Some(index) = self.messages.iter().position(|message| message.id == id) else {
1123 return false;
1124 };
1125 self.messages.remove(index);
1126 self.touch_updated_at();
1127 cx.emit(ThreadEvent::MessageDeleted(id));
1128 true
1129 }
1130
1131 /// Returns the representation of this [`Thread`] in a textual form.
1132 ///
1133 /// This is the representation we use when attaching a thread as context to another thread.
1134 pub fn text(&self) -> String {
1135 let mut text = String::new();
1136
1137 for message in &self.messages {
1138 text.push_str(match message.role {
1139 language_model::Role::User => "User:",
1140 language_model::Role::Assistant => "Agent:",
1141 language_model::Role::System => "System:",
1142 });
1143 text.push('\n');
1144
1145 for segment in &message.segments {
1146 match segment {
1147 MessageSegment::Text(content) => text.push_str(content),
1148 MessageSegment::Thinking { text: content, .. } => {
1149 text.push_str(&format!("<think>{}</think>", content))
1150 }
1151 MessageSegment::RedactedThinking(_) => {}
1152 }
1153 }
1154 text.push('\n');
1155 }
1156
1157 text
1158 }
1159
1160 /// Serializes this thread into a format for storage or telemetry.
1161 pub fn serialize(&self, cx: &mut Context<Self>) -> Task<Result<SerializedThread>> {
1162 let initial_project_snapshot = self.initial_project_snapshot.clone();
1163 cx.spawn(async move |this, cx| {
1164 let initial_project_snapshot = initial_project_snapshot.await;
1165 this.read_with(cx, |this, cx| SerializedThread {
1166 version: SerializedThread::VERSION.to_string(),
1167 summary: this.summary().or_default(),
1168 updated_at: this.updated_at(),
1169 messages: this
1170 .messages()
1171 .filter(|message| !message.ui_only)
1172 .map(|message| SerializedMessage {
1173 id: message.id,
1174 role: message.role,
1175 segments: message
1176 .segments
1177 .iter()
1178 .map(|segment| match segment {
1179 MessageSegment::Text(text) => {
1180 SerializedMessageSegment::Text { text: text.clone() }
1181 }
1182 MessageSegment::Thinking { text, signature } => {
1183 SerializedMessageSegment::Thinking {
1184 text: text.clone(),
1185 signature: signature.clone(),
1186 }
1187 }
1188 MessageSegment::RedactedThinking(data) => {
1189 SerializedMessageSegment::RedactedThinking {
1190 data: data.clone(),
1191 }
1192 }
1193 })
1194 .collect(),
1195 tool_uses: this
1196 .tool_uses_for_message(message.id, cx)
1197 .into_iter()
1198 .map(|tool_use| SerializedToolUse {
1199 id: tool_use.id,
1200 name: tool_use.name,
1201 input: tool_use.input,
1202 })
1203 .collect(),
1204 tool_results: this
1205 .tool_results_for_message(message.id)
1206 .into_iter()
1207 .map(|tool_result| SerializedToolResult {
1208 tool_use_id: tool_result.tool_use_id.clone(),
1209 is_error: tool_result.is_error,
1210 content: tool_result.content.clone(),
1211 output: tool_result.output.clone(),
1212 })
1213 .collect(),
1214 context: message.loaded_context.text.clone(),
1215 creases: message
1216 .creases
1217 .iter()
1218 .map(|crease| SerializedCrease {
1219 start: crease.range.start,
1220 end: crease.range.end,
1221 icon_path: crease.icon_path.clone(),
1222 label: crease.label.clone(),
1223 })
1224 .collect(),
1225 is_hidden: message.is_hidden,
1226 })
1227 .collect(),
1228 initial_project_snapshot,
1229 cumulative_token_usage: this.cumulative_token_usage,
1230 request_token_usage: this.request_token_usage.clone(),
1231 detailed_summary_state: this.detailed_summary_rx.borrow().clone(),
1232 exceeded_window_error: this.exceeded_window_error.clone(),
1233 model: this
1234 .configured_model
1235 .as_ref()
1236 .map(|model| SerializedLanguageModel {
1237 provider: model.provider.id().0.to_string(),
1238 model: model.model.id().0.to_string(),
1239 }),
1240 completion_mode: Some(this.completion_mode),
1241 tool_use_limit_reached: this.tool_use_limit_reached,
1242 profile: Some(this.profile.id().clone()),
1243 })
1244 })
1245 }
1246
1247 pub fn remaining_turns(&self) -> u32 {
1248 self.remaining_turns
1249 }
1250
1251 pub fn set_remaining_turns(&mut self, remaining_turns: u32) {
1252 self.remaining_turns = remaining_turns;
1253 }
1254
1255 pub fn send_to_model(
1256 &mut self,
1257 model: Arc<dyn LanguageModel>,
1258 intent: CompletionIntent,
1259 window: Option<AnyWindowHandle>,
1260 cx: &mut Context<Self>,
1261 ) {
1262 if self.remaining_turns == 0 {
1263 return;
1264 }
1265
1266 self.remaining_turns -= 1;
1267
1268 self.flush_notifications(model.clone(), intent, cx);
1269
1270 let _checkpoint = self.finalize_pending_checkpoint(cx);
1271 self.stream_completion(
1272 self.to_completion_request(model.clone(), intent, cx),
1273 model,
1274 intent,
1275 window,
1276 cx,
1277 );
1278 }
1279
1280 pub fn retry_last_completion(
1281 &mut self,
1282 window: Option<AnyWindowHandle>,
1283 cx: &mut Context<Self>,
1284 ) {
1285 // Clear any existing error state
1286 self.retry_state = None;
1287
1288 // Use the last error context if available, otherwise fall back to configured model
1289 let (model, intent) = if let Some((model, intent)) = self.last_error_context.take() {
1290 (model, intent)
1291 } else if let Some(configured_model) = self.configured_model.as_ref() {
1292 let model = configured_model.model.clone();
1293 let intent = if self.has_pending_tool_uses() {
1294 CompletionIntent::ToolResults
1295 } else {
1296 CompletionIntent::UserPrompt
1297 };
1298 (model, intent)
1299 } else if let Some(configured_model) = self.get_or_init_configured_model(cx) {
1300 let model = configured_model.model.clone();
1301 let intent = if self.has_pending_tool_uses() {
1302 CompletionIntent::ToolResults
1303 } else {
1304 CompletionIntent::UserPrompt
1305 };
1306 (model, intent)
1307 } else {
1308 return;
1309 };
1310
1311 self.send_to_model(model, intent, window, cx);
1312 }
1313
1314 pub fn enable_burn_mode_and_retry(
1315 &mut self,
1316 window: Option<AnyWindowHandle>,
1317 cx: &mut Context<Self>,
1318 ) {
1319 self.completion_mode = CompletionMode::Burn;
1320 cx.emit(ThreadEvent::ProfileChanged);
1321 self.retry_last_completion(window, cx);
1322 }
1323
1324 pub fn used_tools_since_last_user_message(&self) -> bool {
1325 for message in self.messages.iter().rev() {
1326 if self.tool_use.message_has_tool_results(message.id) {
1327 return true;
1328 } else if message.role == Role::User {
1329 return false;
1330 }
1331 }
1332
1333 false
1334 }
1335
1336 pub fn to_completion_request(
1337 &self,
1338 model: Arc<dyn LanguageModel>,
1339 intent: CompletionIntent,
1340 cx: &mut Context<Self>,
1341 ) -> LanguageModelRequest {
1342 let mut request = LanguageModelRequest {
1343 thread_id: Some(self.id.to_string()),
1344 prompt_id: Some(self.last_prompt_id.to_string()),
1345 intent: Some(intent),
1346 mode: None,
1347 messages: vec![],
1348 tools: Vec::new(),
1349 tool_choice: None,
1350 stop: Vec::new(),
1351 temperature: AgentSettings::temperature_for_model(&model, cx),
1352 thinking_allowed: true,
1353 };
1354
1355 let available_tools = self.available_tools(cx, model.clone());
1356 let available_tool_names = available_tools
1357 .iter()
1358 .map(|tool| tool.name.clone())
1359 .collect();
1360
1361 let model_context = &ModelContext {
1362 available_tools: available_tool_names,
1363 };
1364
1365 if let Some(project_context) = self.project_context.borrow().as_ref() {
1366 match self
1367 .prompt_builder
1368 .generate_assistant_system_prompt(project_context, model_context)
1369 {
1370 Err(err) => {
1371 let message = format!("{err:?}").into();
1372 log::error!("{message}");
1373 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1374 header: "Error generating system prompt".into(),
1375 message,
1376 }));
1377 }
1378 Ok(system_prompt) => {
1379 request.messages.push(LanguageModelRequestMessage {
1380 role: Role::System,
1381 content: vec![MessageContent::Text(system_prompt)],
1382 cache: true,
1383 });
1384 }
1385 }
1386 } else {
1387 let message = "Context for system prompt unexpectedly not ready.".into();
1388 log::error!("{message}");
1389 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1390 header: "Error generating system prompt".into(),
1391 message,
1392 }));
1393 }
1394
1395 let mut message_ix_to_cache = None;
1396 for message in &self.messages {
1397 // ui_only messages are for the UI only, not for the model
1398 if message.ui_only {
1399 continue;
1400 }
1401
1402 let mut request_message = LanguageModelRequestMessage {
1403 role: message.role,
1404 content: Vec::new(),
1405 cache: false,
1406 };
1407
1408 message
1409 .loaded_context
1410 .add_to_request_message(&mut request_message);
1411
1412 for segment in &message.segments {
1413 match segment {
1414 MessageSegment::Text(text) => {
1415 let text = text.trim_end();
1416 if !text.is_empty() {
1417 request_message
1418 .content
1419 .push(MessageContent::Text(text.into()));
1420 }
1421 }
1422 MessageSegment::Thinking { text, signature } => {
1423 if !text.is_empty() {
1424 request_message.content.push(MessageContent::Thinking {
1425 text: text.into(),
1426 signature: signature.clone(),
1427 });
1428 }
1429 }
1430 MessageSegment::RedactedThinking(data) => {
1431 request_message
1432 .content
1433 .push(MessageContent::RedactedThinking(data.clone()));
1434 }
1435 };
1436 }
1437
1438 let mut cache_message = true;
1439 let mut tool_results_message = LanguageModelRequestMessage {
1440 role: Role::User,
1441 content: Vec::new(),
1442 cache: false,
1443 };
1444 for (tool_use, tool_result) in self.tool_use.tool_results(message.id) {
1445 if let Some(tool_result) = tool_result {
1446 request_message
1447 .content
1448 .push(MessageContent::ToolUse(tool_use.clone()));
1449 tool_results_message
1450 .content
1451 .push(MessageContent::ToolResult(LanguageModelToolResult {
1452 tool_use_id: tool_use.id.clone(),
1453 tool_name: tool_result.tool_name.clone(),
1454 is_error: tool_result.is_error,
1455 content: if tool_result.content.is_empty() {
1456 // Surprisingly, the API fails if we return an empty string here.
1457 // It thinks we are sending a tool use without a tool result.
1458 "<Tool returned an empty string>".into()
1459 } else {
1460 tool_result.content.clone()
1461 },
1462 output: None,
1463 }));
1464 } else {
1465 cache_message = false;
1466 log::debug!(
1467 "skipped tool use {:?} because it is still pending",
1468 tool_use
1469 );
1470 }
1471 }
1472
1473 if cache_message {
1474 message_ix_to_cache = Some(request.messages.len());
1475 }
1476 request.messages.push(request_message);
1477
1478 if !tool_results_message.content.is_empty() {
1479 if cache_message {
1480 message_ix_to_cache = Some(request.messages.len());
1481 }
1482 request.messages.push(tool_results_message);
1483 }
1484 }
1485
1486 // https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
1487 if let Some(message_ix_to_cache) = message_ix_to_cache {
1488 request.messages[message_ix_to_cache].cache = true;
1489 }
1490
1491 request.tools = available_tools;
1492 request.mode = if model.supports_burn_mode() {
1493 Some(self.completion_mode.into())
1494 } else {
1495 Some(CompletionMode::Normal.into())
1496 };
1497
1498 request
1499 }
1500
1501 fn to_summarize_request(
1502 &self,
1503 model: &Arc<dyn LanguageModel>,
1504 intent: CompletionIntent,
1505 added_user_message: String,
1506 cx: &App,
1507 ) -> LanguageModelRequest {
1508 let mut request = LanguageModelRequest {
1509 thread_id: None,
1510 prompt_id: None,
1511 intent: Some(intent),
1512 mode: None,
1513 messages: vec![],
1514 tools: Vec::new(),
1515 tool_choice: None,
1516 stop: Vec::new(),
1517 temperature: AgentSettings::temperature_for_model(model, cx),
1518 thinking_allowed: false,
1519 };
1520
1521 for message in &self.messages {
1522 let mut request_message = LanguageModelRequestMessage {
1523 role: message.role,
1524 content: Vec::new(),
1525 cache: false,
1526 };
1527
1528 for segment in &message.segments {
1529 match segment {
1530 MessageSegment::Text(text) => request_message
1531 .content
1532 .push(MessageContent::Text(text.clone())),
1533 MessageSegment::Thinking { .. } => {}
1534 MessageSegment::RedactedThinking(_) => {}
1535 }
1536 }
1537
1538 if request_message.content.is_empty() {
1539 continue;
1540 }
1541
1542 request.messages.push(request_message);
1543 }
1544
1545 request.messages.push(LanguageModelRequestMessage {
1546 role: Role::User,
1547 content: vec![MessageContent::Text(added_user_message)],
1548 cache: false,
1549 });
1550
1551 request
1552 }
1553
1554 /// Insert auto-generated notifications (if any) to the thread
1555 fn flush_notifications(
1556 &mut self,
1557 model: Arc<dyn LanguageModel>,
1558 intent: CompletionIntent,
1559 cx: &mut Context<Self>,
1560 ) {
1561 match intent {
1562 CompletionIntent::UserPrompt | CompletionIntent::ToolResults => {
1563 if let Some(pending_tool_use) = self.attach_tracked_files_state(model, cx) {
1564 cx.emit(ThreadEvent::ToolFinished {
1565 tool_use_id: pending_tool_use.id.clone(),
1566 pending_tool_use: Some(pending_tool_use),
1567 });
1568 }
1569 }
1570 CompletionIntent::ThreadSummarization
1571 | CompletionIntent::ThreadContextSummarization
1572 | CompletionIntent::CreateFile
1573 | CompletionIntent::EditFile
1574 | CompletionIntent::InlineAssist
1575 | CompletionIntent::TerminalInlineAssist
1576 | CompletionIntent::GenerateGitCommitMessage => {}
1577 };
1578 }
1579
1580 fn attach_tracked_files_state(
1581 &mut self,
1582 model: Arc<dyn LanguageModel>,
1583 cx: &mut App,
1584 ) -> Option<PendingToolUse> {
1585 // Represent notification as a simulated `project_notifications` tool call
1586 let tool_name = Arc::from("project_notifications");
1587 let tool = self.tools.read(cx).tool(&tool_name, cx)?;
1588
1589 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
1590 return None;
1591 }
1592
1593 if self
1594 .action_log
1595 .update(cx, |log, cx| log.unnotified_user_edits(cx).is_none())
1596 {
1597 return None;
1598 }
1599
1600 let input = serde_json::json!({});
1601 let request = Arc::new(LanguageModelRequest::default()); // unused
1602 let window = None;
1603 let tool_result = tool.run(
1604 input,
1605 request,
1606 self.project.clone(),
1607 self.action_log.clone(),
1608 model.clone(),
1609 window,
1610 cx,
1611 );
1612
1613 let tool_use_id =
1614 LanguageModelToolUseId::from(format!("project_notifications_{}", self.messages.len()));
1615
1616 let tool_use = LanguageModelToolUse {
1617 id: tool_use_id.clone(),
1618 name: tool_name.clone(),
1619 raw_input: "{}".to_string(),
1620 input: serde_json::json!({}),
1621 is_input_complete: true,
1622 };
1623
1624 let tool_output = cx.background_executor().block(tool_result.output);
1625
1626 // Attach a project_notification tool call to the latest existing
1627 // Assistant message. We cannot create a new Assistant message
1628 // because thinking models require a `thinking` block that we
1629 // cannot mock. We cannot send a notification as a normal
1630 // (non-tool-use) User message because this distracts Agent
1631 // too much.
1632 let tool_message_id = self
1633 .messages
1634 .iter()
1635 .enumerate()
1636 .rfind(|(_, message)| message.role == Role::Assistant)
1637 .map(|(_, message)| message.id)?;
1638
1639 let tool_use_metadata = ToolUseMetadata {
1640 model: model.clone(),
1641 thread_id: self.id.clone(),
1642 prompt_id: self.last_prompt_id.clone(),
1643 };
1644
1645 self.tool_use
1646 .request_tool_use(tool_message_id, tool_use, tool_use_metadata.clone(), cx);
1647
1648 self.tool_use.insert_tool_output(
1649 tool_use_id.clone(),
1650 tool_name,
1651 tool_output,
1652 self.configured_model.as_ref(),
1653 self.completion_mode,
1654 )
1655 }
1656
1657 pub fn stream_completion(
1658 &mut self,
1659 request: LanguageModelRequest,
1660 model: Arc<dyn LanguageModel>,
1661 intent: CompletionIntent,
1662 window: Option<AnyWindowHandle>,
1663 cx: &mut Context<Self>,
1664 ) {
1665 self.tool_use_limit_reached = false;
1666
1667 let pending_completion_id = post_inc(&mut self.completion_count);
1668 let mut request_callback_parameters = if self.request_callback.is_some() {
1669 Some((request.clone(), Vec::new()))
1670 } else {
1671 None
1672 };
1673 let prompt_id = self.last_prompt_id.clone();
1674 let tool_use_metadata = ToolUseMetadata {
1675 model: model.clone(),
1676 thread_id: self.id.clone(),
1677 prompt_id: prompt_id.clone(),
1678 };
1679
1680 let completion_mode = request
1681 .mode
1682 .unwrap_or(cloud_llm_client::CompletionMode::Normal);
1683
1684 self.last_received_chunk_at = Some(Instant::now());
1685
1686 let task = cx.spawn(async move |thread, cx| {
1687 let stream_completion_future = model.stream_completion(request, cx);
1688 let initial_token_usage =
1689 thread.read_with(cx, |thread, _cx| thread.cumulative_token_usage);
1690 let stream_completion = async {
1691 let mut events = stream_completion_future.await?;
1692
1693 let mut stop_reason = StopReason::EndTurn;
1694 let mut current_token_usage = TokenUsage::default();
1695
1696 thread
1697 .update(cx, |_thread, cx| {
1698 cx.emit(ThreadEvent::NewRequest);
1699 })
1700 .ok();
1701
1702 let mut request_assistant_message_id = None;
1703
1704 while let Some(event) = events.next().await {
1705 if let Some((_, response_events)) = request_callback_parameters.as_mut() {
1706 response_events
1707 .push(event.as_ref().map_err(|error| error.to_string()).cloned());
1708 }
1709
1710 thread.update(cx, |thread, cx| {
1711 match event? {
1712 LanguageModelCompletionEvent::StartMessage { .. } => {
1713 request_assistant_message_id =
1714 Some(thread.insert_assistant_message(
1715 vec![MessageSegment::Text(String::new())],
1716 cx,
1717 ));
1718 }
1719 LanguageModelCompletionEvent::Stop(reason) => {
1720 stop_reason = reason;
1721 }
1722 LanguageModelCompletionEvent::UsageUpdate(token_usage) => {
1723 thread.update_token_usage_at_last_message(token_usage);
1724 thread.cumulative_token_usage = thread.cumulative_token_usage
1725 + token_usage
1726 - current_token_usage;
1727 current_token_usage = token_usage;
1728 }
1729 LanguageModelCompletionEvent::Text(chunk) => {
1730 thread.received_chunk();
1731
1732 cx.emit(ThreadEvent::ReceivedTextChunk);
1733 if let Some(last_message) = thread.messages.last_mut() {
1734 if last_message.role == Role::Assistant
1735 && !thread.tool_use.has_tool_results(last_message.id)
1736 {
1737 last_message.push_text(&chunk);
1738 cx.emit(ThreadEvent::StreamedAssistantText(
1739 last_message.id,
1740 chunk,
1741 ));
1742 } else {
1743 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1744 // of a new Assistant response.
1745 //
1746 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1747 // will result in duplicating the text of the chunk in the rendered Markdown.
1748 request_assistant_message_id =
1749 Some(thread.insert_assistant_message(
1750 vec![MessageSegment::Text(chunk.to_string())],
1751 cx,
1752 ));
1753 };
1754 }
1755 }
1756 LanguageModelCompletionEvent::Thinking {
1757 text: chunk,
1758 signature,
1759 } => {
1760 thread.received_chunk();
1761
1762 if let Some(last_message) = thread.messages.last_mut() {
1763 if last_message.role == Role::Assistant
1764 && !thread.tool_use.has_tool_results(last_message.id)
1765 {
1766 last_message.push_thinking(&chunk, signature);
1767 cx.emit(ThreadEvent::StreamedAssistantThinking(
1768 last_message.id,
1769 chunk,
1770 ));
1771 } else {
1772 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1773 // of a new Assistant response.
1774 //
1775 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1776 // will result in duplicating the text of the chunk in the rendered Markdown.
1777 request_assistant_message_id =
1778 Some(thread.insert_assistant_message(
1779 vec![MessageSegment::Thinking {
1780 text: chunk.to_string(),
1781 signature,
1782 }],
1783 cx,
1784 ));
1785 };
1786 }
1787 }
1788 LanguageModelCompletionEvent::RedactedThinking { data } => {
1789 thread.received_chunk();
1790
1791 if let Some(last_message) = thread.messages.last_mut() {
1792 if last_message.role == Role::Assistant
1793 && !thread.tool_use.has_tool_results(last_message.id)
1794 {
1795 last_message.push_redacted_thinking(data);
1796 } else {
1797 request_assistant_message_id =
1798 Some(thread.insert_assistant_message(
1799 vec![MessageSegment::RedactedThinking(data)],
1800 cx,
1801 ));
1802 };
1803 }
1804 }
1805 LanguageModelCompletionEvent::ToolUse(tool_use) => {
1806 let last_assistant_message_id = request_assistant_message_id
1807 .unwrap_or_else(|| {
1808 let new_assistant_message_id =
1809 thread.insert_assistant_message(vec![], cx);
1810 request_assistant_message_id =
1811 Some(new_assistant_message_id);
1812 new_assistant_message_id
1813 });
1814
1815 let tool_use_id = tool_use.id.clone();
1816 let streamed_input = if tool_use.is_input_complete {
1817 None
1818 } else {
1819 Some(tool_use.input.clone())
1820 };
1821
1822 let ui_text = thread.tool_use.request_tool_use(
1823 last_assistant_message_id,
1824 tool_use,
1825 tool_use_metadata.clone(),
1826 cx,
1827 );
1828
1829 if let Some(input) = streamed_input {
1830 cx.emit(ThreadEvent::StreamedToolUse {
1831 tool_use_id,
1832 ui_text,
1833 input,
1834 });
1835 }
1836 }
1837 LanguageModelCompletionEvent::ToolUseJsonParseError {
1838 id,
1839 tool_name,
1840 raw_input: invalid_input_json,
1841 json_parse_error,
1842 } => {
1843 thread.receive_invalid_tool_json(
1844 id,
1845 tool_name,
1846 invalid_input_json,
1847 json_parse_error,
1848 window,
1849 cx,
1850 );
1851 }
1852 LanguageModelCompletionEvent::StatusUpdate(status_update) => {
1853 if let Some(completion) = thread
1854 .pending_completions
1855 .iter_mut()
1856 .find(|completion| completion.id == pending_completion_id)
1857 {
1858 match status_update {
1859 CompletionRequestStatus::Queued { position } => {
1860 completion.queue_state =
1861 QueueState::Queued { position };
1862 }
1863 CompletionRequestStatus::Started => {
1864 completion.queue_state = QueueState::Started;
1865 }
1866 CompletionRequestStatus::Failed {
1867 code,
1868 message,
1869 request_id: _,
1870 retry_after,
1871 } => {
1872 return Err(
1873 LanguageModelCompletionError::from_cloud_failure(
1874 model.upstream_provider_name(),
1875 code,
1876 message,
1877 retry_after.map(Duration::from_secs_f64),
1878 ),
1879 );
1880 }
1881 CompletionRequestStatus::UsageUpdated { amount, limit } => {
1882 thread.update_model_request_usage(
1883 amount as u32,
1884 limit,
1885 cx,
1886 );
1887 }
1888 CompletionRequestStatus::ToolUseLimitReached => {
1889 thread.tool_use_limit_reached = true;
1890 cx.emit(ThreadEvent::ToolUseLimitReached);
1891 }
1892 }
1893 }
1894 }
1895 }
1896
1897 thread.touch_updated_at();
1898 cx.emit(ThreadEvent::StreamedCompletion);
1899 cx.notify();
1900
1901 Ok(())
1902 })??;
1903
1904 smol::future::yield_now().await;
1905 }
1906
1907 thread.update(cx, |thread, cx| {
1908 thread.last_received_chunk_at = None;
1909 thread
1910 .pending_completions
1911 .retain(|completion| completion.id != pending_completion_id);
1912
1913 // If there is a response without tool use, summarize the message. Otherwise,
1914 // allow two tool uses before summarizing.
1915 if matches!(thread.summary, ThreadSummary::Pending)
1916 && thread.messages.len() >= 2
1917 && (!thread.has_pending_tool_uses() || thread.messages.len() >= 6)
1918 {
1919 thread.summarize(cx);
1920 }
1921 })?;
1922
1923 anyhow::Ok(stop_reason)
1924 };
1925
1926 let result = stream_completion.await;
1927 let mut retry_scheduled = false;
1928
1929 thread
1930 .update(cx, |thread, cx| {
1931 thread.finalize_pending_checkpoint(cx);
1932 match result.as_ref() {
1933 Ok(stop_reason) => {
1934 match stop_reason {
1935 StopReason::ToolUse => {
1936 let tool_uses =
1937 thread.use_pending_tools(window, model.clone(), cx);
1938 cx.emit(ThreadEvent::UsePendingTools { tool_uses });
1939 }
1940 StopReason::EndTurn | StopReason::MaxTokens => {
1941 thread.project.update(cx, |project, cx| {
1942 project.set_agent_location(None, cx);
1943 });
1944 }
1945 StopReason::Refusal => {
1946 thread.project.update(cx, |project, cx| {
1947 project.set_agent_location(None, cx);
1948 });
1949
1950 // Remove the turn that was refused.
1951 //
1952 // https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/handle-streaming-refusals#reset-context-after-refusal
1953 {
1954 let mut messages_to_remove = Vec::new();
1955
1956 for (ix, message) in
1957 thread.messages.iter().enumerate().rev()
1958 {
1959 messages_to_remove.push(message.id);
1960
1961 if message.role == Role::User {
1962 if ix == 0 {
1963 break;
1964 }
1965
1966 if let Some(prev_message) =
1967 thread.messages.get(ix - 1)
1968 && prev_message.role == Role::Assistant {
1969 break;
1970 }
1971 }
1972 }
1973
1974 for message_id in messages_to_remove {
1975 thread.delete_message(message_id, cx);
1976 }
1977 }
1978
1979 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1980 header: "Language model refusal".into(),
1981 message:
1982 "Model refused to generate content for safety reasons."
1983 .into(),
1984 }));
1985 }
1986 }
1987
1988 // We successfully completed, so cancel any remaining retries.
1989 thread.retry_state = None;
1990 }
1991 Err(error) => {
1992 thread.project.update(cx, |project, cx| {
1993 project.set_agent_location(None, cx);
1994 });
1995
1996 if error.is::<PaymentRequiredError>() {
1997 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
1998 } else if let Some(error) =
1999 error.downcast_ref::<ModelRequestLimitReachedError>()
2000 {
2001 cx.emit(ThreadEvent::ShowError(
2002 ThreadError::ModelRequestLimitReached { plan: error.plan },
2003 ));
2004 } else if let Some(completion_error) =
2005 error.downcast_ref::<LanguageModelCompletionError>()
2006 {
2007 match &completion_error {
2008 LanguageModelCompletionError::PromptTooLarge {
2009 tokens, ..
2010 } => {
2011 let tokens = tokens.unwrap_or_else(|| {
2012 // We didn't get an exact token count from the API, so fall back on our estimate.
2013 thread
2014 .total_token_usage()
2015 .map(|usage| usage.total)
2016 .unwrap_or(0)
2017 // We know the context window was exceeded in practice, so if our estimate was
2018 // lower than max tokens, the estimate was wrong; return that we exceeded by 1.
2019 .max(
2020 model
2021 .max_token_count_for_mode(completion_mode)
2022 .saturating_add(1),
2023 )
2024 });
2025 thread.exceeded_window_error = Some(ExceededWindowError {
2026 model_id: model.id(),
2027 token_count: tokens,
2028 });
2029 cx.notify();
2030 }
2031 _ => {
2032 if let Some(retry_strategy) =
2033 Thread::get_retry_strategy(completion_error)
2034 {
2035 log::info!(
2036 "Retrying with {:?} for language model completion error {:?}",
2037 retry_strategy,
2038 completion_error
2039 );
2040
2041 retry_scheduled = thread
2042 .handle_retryable_error_with_delay(
2043 completion_error,
2044 Some(retry_strategy),
2045 model.clone(),
2046 intent,
2047 window,
2048 cx,
2049 );
2050 }
2051 }
2052 }
2053 }
2054
2055 if !retry_scheduled {
2056 thread.cancel_last_completion(window, cx);
2057 }
2058 }
2059 }
2060
2061 if !retry_scheduled {
2062 cx.emit(ThreadEvent::Stopped(result.map_err(Arc::new)));
2063 }
2064
2065 if let Some((request_callback, (request, response_events))) = thread
2066 .request_callback
2067 .as_mut()
2068 .zip(request_callback_parameters.as_ref())
2069 {
2070 request_callback(request, response_events);
2071 }
2072
2073 if let Ok(initial_usage) = initial_token_usage {
2074 let usage = thread.cumulative_token_usage - initial_usage;
2075
2076 telemetry::event!(
2077 "Assistant Thread Completion",
2078 thread_id = thread.id().to_string(),
2079 prompt_id = prompt_id,
2080 model = model.telemetry_id(),
2081 model_provider = model.provider_id().to_string(),
2082 input_tokens = usage.input_tokens,
2083 output_tokens = usage.output_tokens,
2084 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2085 cache_read_input_tokens = usage.cache_read_input_tokens,
2086 );
2087 }
2088 })
2089 .ok();
2090 });
2091
2092 self.pending_completions.push(PendingCompletion {
2093 id: pending_completion_id,
2094 queue_state: QueueState::Sending,
2095 _task: task,
2096 });
2097 }
2098
2099 pub fn summarize(&mut self, cx: &mut Context<Self>) {
2100 let Some(model) = LanguageModelRegistry::read_global(cx).thread_summary_model() else {
2101 println!("No thread summary model");
2102 return;
2103 };
2104
2105 if !model.provider.is_authenticated(cx) {
2106 return;
2107 }
2108
2109 let request = self.to_summarize_request(
2110 &model.model,
2111 CompletionIntent::ThreadSummarization,
2112 SUMMARIZE_THREAD_PROMPT.into(),
2113 cx,
2114 );
2115
2116 self.summary = ThreadSummary::Generating;
2117
2118 self.pending_summary = cx.spawn(async move |this, cx| {
2119 let result = async {
2120 let mut messages = model.model.stream_completion(request, cx).await?;
2121
2122 let mut new_summary = String::new();
2123 while let Some(event) = messages.next().await {
2124 let Ok(event) = event else {
2125 continue;
2126 };
2127 let text = match event {
2128 LanguageModelCompletionEvent::Text(text) => text,
2129 LanguageModelCompletionEvent::StatusUpdate(
2130 CompletionRequestStatus::UsageUpdated { amount, limit },
2131 ) => {
2132 this.update(cx, |thread, cx| {
2133 thread.update_model_request_usage(amount as u32, limit, cx);
2134 })?;
2135 continue;
2136 }
2137 _ => continue,
2138 };
2139
2140 let mut lines = text.lines();
2141 new_summary.extend(lines.next());
2142
2143 // Stop if the LLM generated multiple lines.
2144 if lines.next().is_some() {
2145 break;
2146 }
2147 }
2148
2149 anyhow::Ok(new_summary)
2150 }
2151 .await;
2152
2153 this.update(cx, |this, cx| {
2154 match result {
2155 Ok(new_summary) => {
2156 if new_summary.is_empty() {
2157 this.summary = ThreadSummary::Error;
2158 } else {
2159 this.summary = ThreadSummary::Ready(new_summary.into());
2160 }
2161 }
2162 Err(err) => {
2163 this.summary = ThreadSummary::Error;
2164 log::error!("Failed to generate thread summary: {}", err);
2165 }
2166 }
2167 cx.emit(ThreadEvent::SummaryGenerated);
2168 })
2169 .log_err()?;
2170
2171 Some(())
2172 });
2173 }
2174
2175 fn get_retry_strategy(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2176 use LanguageModelCompletionError::*;
2177
2178 // General strategy here:
2179 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2180 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2181 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2182 match error {
2183 HttpResponseError {
2184 status_code: StatusCode::TOO_MANY_REQUESTS,
2185 ..
2186 } => Some(RetryStrategy::ExponentialBackoff {
2187 initial_delay: BASE_RETRY_DELAY,
2188 max_attempts: MAX_RETRY_ATTEMPTS,
2189 }),
2190 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2191 Some(RetryStrategy::Fixed {
2192 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2193 max_attempts: MAX_RETRY_ATTEMPTS,
2194 })
2195 }
2196 UpstreamProviderError {
2197 status,
2198 retry_after,
2199 ..
2200 } => match *status {
2201 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2202 Some(RetryStrategy::Fixed {
2203 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2204 max_attempts: MAX_RETRY_ATTEMPTS,
2205 })
2206 }
2207 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2208 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2209 // Internal Server Error could be anything, retry up to 3 times.
2210 max_attempts: 3,
2211 }),
2212 status => {
2213 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2214 // but we frequently get them in practice. See https://http.dev/529
2215 if status.as_u16() == 529 {
2216 Some(RetryStrategy::Fixed {
2217 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2218 max_attempts: MAX_RETRY_ATTEMPTS,
2219 })
2220 } else {
2221 Some(RetryStrategy::Fixed {
2222 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2223 max_attempts: 2,
2224 })
2225 }
2226 }
2227 },
2228 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2229 delay: BASE_RETRY_DELAY,
2230 max_attempts: 3,
2231 }),
2232 ApiReadResponseError { .. }
2233 | HttpSend { .. }
2234 | DeserializeResponse { .. }
2235 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2236 delay: BASE_RETRY_DELAY,
2237 max_attempts: 3,
2238 }),
2239 // Retrying these errors definitely shouldn't help.
2240 HttpResponseError {
2241 status_code:
2242 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2243 ..
2244 }
2245 | AuthenticationError { .. }
2246 | PermissionError { .. }
2247 | NoApiKey { .. }
2248 | ApiEndpointNotFound { .. }
2249 | PromptTooLarge { .. } => None,
2250 // These errors might be transient, so retry them
2251 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2252 delay: BASE_RETRY_DELAY,
2253 max_attempts: 1,
2254 }),
2255 // Retry all other 4xx and 5xx errors once.
2256 HttpResponseError { status_code, .. }
2257 if status_code.is_client_error() || status_code.is_server_error() =>
2258 {
2259 Some(RetryStrategy::Fixed {
2260 delay: BASE_RETRY_DELAY,
2261 max_attempts: 3,
2262 })
2263 }
2264 Other(err)
2265 if err.is::<PaymentRequiredError>()
2266 || err.is::<ModelRequestLimitReachedError>() =>
2267 {
2268 // Retrying won't help for Payment Required or Model Request Limit errors (where
2269 // the user must upgrade to usage-based billing to get more requests, or else wait
2270 // for a significant amount of time for the request limit to reset).
2271 None
2272 }
2273 // Conservatively assume that any other errors are non-retryable
2274 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2275 delay: BASE_RETRY_DELAY,
2276 max_attempts: 2,
2277 }),
2278 }
2279 }
2280
2281 fn handle_retryable_error_with_delay(
2282 &mut self,
2283 error: &LanguageModelCompletionError,
2284 strategy: Option<RetryStrategy>,
2285 model: Arc<dyn LanguageModel>,
2286 intent: CompletionIntent,
2287 window: Option<AnyWindowHandle>,
2288 cx: &mut Context<Self>,
2289 ) -> bool {
2290 // Store context for the Retry button
2291 self.last_error_context = Some((model.clone(), intent));
2292
2293 // Only auto-retry if Burn Mode is enabled
2294 if self.completion_mode != CompletionMode::Burn {
2295 // Show error with retry options
2296 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2297 message: format!(
2298 "{}\n\nTo automatically retry when similar errors happen, enable Burn Mode.",
2299 error
2300 )
2301 .into(),
2302 can_enable_burn_mode: true,
2303 }));
2304 return false;
2305 }
2306
2307 let Some(strategy) = strategy.or_else(|| Self::get_retry_strategy(error)) else {
2308 return false;
2309 };
2310
2311 let max_attempts = match &strategy {
2312 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2313 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2314 };
2315
2316 let retry_state = self.retry_state.get_or_insert(RetryState {
2317 attempt: 0,
2318 max_attempts,
2319 intent,
2320 });
2321
2322 retry_state.attempt += 1;
2323 let attempt = retry_state.attempt;
2324 let max_attempts = retry_state.max_attempts;
2325 let intent = retry_state.intent;
2326
2327 if attempt <= max_attempts {
2328 let delay = match &strategy {
2329 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2330 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2331 Duration::from_secs(delay_secs)
2332 }
2333 RetryStrategy::Fixed { delay, .. } => *delay,
2334 };
2335
2336 // Add a transient message to inform the user
2337 let delay_secs = delay.as_secs();
2338 let retry_message = if max_attempts == 1 {
2339 format!("{error}. Retrying in {delay_secs} seconds...")
2340 } else {
2341 format!(
2342 "{error}. Retrying (attempt {attempt} of {max_attempts}) \
2343 in {delay_secs} seconds..."
2344 )
2345 };
2346 log::warn!(
2347 "Retrying completion request (attempt {attempt} of {max_attempts}) \
2348 in {delay_secs} seconds: {error:?}",
2349 );
2350
2351 // Add a UI-only message instead of a regular message
2352 let id = self.next_message_id.post_inc();
2353 self.messages.push(Message {
2354 id,
2355 role: Role::System,
2356 segments: vec![MessageSegment::Text(retry_message)],
2357 loaded_context: LoadedContext::default(),
2358 creases: Vec::new(),
2359 is_hidden: false,
2360 ui_only: true,
2361 });
2362 cx.emit(ThreadEvent::MessageAdded(id));
2363
2364 // Schedule the retry
2365 let thread_handle = cx.entity().downgrade();
2366
2367 cx.spawn(async move |_thread, cx| {
2368 cx.background_executor().timer(delay).await;
2369
2370 thread_handle
2371 .update(cx, |thread, cx| {
2372 // Retry the completion
2373 thread.send_to_model(model, intent, window, cx);
2374 })
2375 .log_err();
2376 })
2377 .detach();
2378
2379 true
2380 } else {
2381 // Max retries exceeded
2382 self.retry_state = None;
2383
2384 // Stop generating since we're giving up on retrying.
2385 self.pending_completions.clear();
2386
2387 // Show error alongside a Retry button, but no
2388 // Enable Burn Mode button (since it's already enabled)
2389 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2390 message: format!("Failed after retrying: {}", error).into(),
2391 can_enable_burn_mode: false,
2392 }));
2393
2394 false
2395 }
2396 }
2397
2398 pub fn start_generating_detailed_summary_if_needed(
2399 &mut self,
2400 thread_store: WeakEntity<ThreadStore>,
2401 cx: &mut Context<Self>,
2402 ) {
2403 let Some(last_message_id) = self.messages.last().map(|message| message.id) else {
2404 return;
2405 };
2406
2407 match &*self.detailed_summary_rx.borrow() {
2408 DetailedSummaryState::Generating { message_id, .. }
2409 | DetailedSummaryState::Generated { message_id, .. }
2410 if *message_id == last_message_id =>
2411 {
2412 // Already up-to-date
2413 return;
2414 }
2415 _ => {}
2416 }
2417
2418 let Some(ConfiguredModel { model, provider }) =
2419 LanguageModelRegistry::read_global(cx).thread_summary_model()
2420 else {
2421 return;
2422 };
2423
2424 if !provider.is_authenticated(cx) {
2425 return;
2426 }
2427
2428 let added_user_message = include_str!("./prompts/summarize_thread_detailed_prompt.txt");
2429
2430 let request = self.to_summarize_request(
2431 &model,
2432 CompletionIntent::ThreadContextSummarization,
2433 added_user_message.into(),
2434 cx,
2435 );
2436
2437 *self.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generating {
2438 message_id: last_message_id,
2439 };
2440
2441 // Replace the detailed summarization task if there is one, cancelling it. It would probably
2442 // be better to allow the old task to complete, but this would require logic for choosing
2443 // which result to prefer (the old task could complete after the new one, resulting in a
2444 // stale summary).
2445 self.detailed_summary_task = cx.spawn(async move |thread, cx| {
2446 let stream = model.stream_completion_text(request, cx);
2447 let Some(mut messages) = stream.await.log_err() else {
2448 thread
2449 .update(cx, |thread, _cx| {
2450 *thread.detailed_summary_tx.borrow_mut() =
2451 DetailedSummaryState::NotGenerated;
2452 })
2453 .ok()?;
2454 return None;
2455 };
2456
2457 let mut new_detailed_summary = String::new();
2458
2459 while let Some(chunk) = messages.stream.next().await {
2460 if let Some(chunk) = chunk.log_err() {
2461 new_detailed_summary.push_str(&chunk);
2462 }
2463 }
2464
2465 thread
2466 .update(cx, |thread, _cx| {
2467 *thread.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generated {
2468 text: new_detailed_summary.into(),
2469 message_id: last_message_id,
2470 };
2471 })
2472 .ok()?;
2473
2474 // Save thread so its summary can be reused later
2475 if let Some(thread) = thread.upgrade()
2476 && let Ok(Ok(save_task)) = cx.update(|cx| {
2477 thread_store
2478 .update(cx, |thread_store, cx| thread_store.save_thread(&thread, cx))
2479 })
2480 {
2481 save_task.await.log_err();
2482 }
2483
2484 Some(())
2485 });
2486 }
2487
2488 pub async fn wait_for_detailed_summary_or_text(
2489 this: &Entity<Self>,
2490 cx: &mut AsyncApp,
2491 ) -> Option<SharedString> {
2492 let mut detailed_summary_rx = this
2493 .read_with(cx, |this, _cx| this.detailed_summary_rx.clone())
2494 .ok()?;
2495 loop {
2496 match detailed_summary_rx.recv().await? {
2497 DetailedSummaryState::Generating { .. } => {}
2498 DetailedSummaryState::NotGenerated => {
2499 return this.read_with(cx, |this, _cx| this.text().into()).ok();
2500 }
2501 DetailedSummaryState::Generated { text, .. } => return Some(text),
2502 }
2503 }
2504 }
2505
2506 pub fn latest_detailed_summary_or_text(&self) -> SharedString {
2507 self.detailed_summary_rx
2508 .borrow()
2509 .text()
2510 .unwrap_or_else(|| self.text().into())
2511 }
2512
2513 pub fn is_generating_detailed_summary(&self) -> bool {
2514 matches!(
2515 &*self.detailed_summary_rx.borrow(),
2516 DetailedSummaryState::Generating { .. }
2517 )
2518 }
2519
2520 pub fn use_pending_tools(
2521 &mut self,
2522 window: Option<AnyWindowHandle>,
2523 model: Arc<dyn LanguageModel>,
2524 cx: &mut Context<Self>,
2525 ) -> Vec<PendingToolUse> {
2526 let request =
2527 Arc::new(self.to_completion_request(model.clone(), CompletionIntent::ToolResults, cx));
2528 let pending_tool_uses = self
2529 .tool_use
2530 .pending_tool_uses()
2531 .into_iter()
2532 .filter(|tool_use| tool_use.status.is_idle())
2533 .cloned()
2534 .collect::<Vec<_>>();
2535
2536 for tool_use in pending_tool_uses.iter() {
2537 self.use_pending_tool(tool_use.clone(), request.clone(), model.clone(), window, cx);
2538 }
2539
2540 pending_tool_uses
2541 }
2542
2543 fn use_pending_tool(
2544 &mut self,
2545 tool_use: PendingToolUse,
2546 request: Arc<LanguageModelRequest>,
2547 model: Arc<dyn LanguageModel>,
2548 window: Option<AnyWindowHandle>,
2549 cx: &mut Context<Self>,
2550 ) {
2551 let Some(tool) = self.tools.read(cx).tool(&tool_use.name, cx) else {
2552 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2553 };
2554
2555 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
2556 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2557 }
2558
2559 if tool.needs_confirmation(&tool_use.input, &self.project, cx)
2560 && !AgentSettings::get_global(cx).always_allow_tool_actions
2561 {
2562 self.tool_use.confirm_tool_use(
2563 tool_use.id,
2564 tool_use.ui_text,
2565 tool_use.input,
2566 request,
2567 tool,
2568 );
2569 cx.emit(ThreadEvent::ToolConfirmationNeeded);
2570 } else {
2571 self.run_tool(
2572 tool_use.id,
2573 tool_use.ui_text,
2574 tool_use.input,
2575 request,
2576 tool,
2577 model,
2578 window,
2579 cx,
2580 );
2581 }
2582 }
2583
2584 pub fn handle_hallucinated_tool_use(
2585 &mut self,
2586 tool_use_id: LanguageModelToolUseId,
2587 hallucinated_tool_name: Arc<str>,
2588 window: Option<AnyWindowHandle>,
2589 cx: &mut Context<Thread>,
2590 ) {
2591 let available_tools = self.profile.enabled_tools(cx);
2592
2593 let tool_list = available_tools
2594 .iter()
2595 .map(|(name, tool)| format!("- {}: {}", name, tool.description()))
2596 .collect::<Vec<_>>()
2597 .join("\n");
2598
2599 let error_message = format!(
2600 "The tool '{}' doesn't exist or is not enabled. Available tools:\n{}",
2601 hallucinated_tool_name, tool_list
2602 );
2603
2604 let pending_tool_use = self.tool_use.insert_tool_output(
2605 tool_use_id.clone(),
2606 hallucinated_tool_name,
2607 Err(anyhow!("Missing tool call: {error_message}")),
2608 self.configured_model.as_ref(),
2609 self.completion_mode,
2610 );
2611
2612 cx.emit(ThreadEvent::MissingToolUse {
2613 tool_use_id: tool_use_id.clone(),
2614 ui_text: error_message.into(),
2615 });
2616
2617 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2618 }
2619
2620 pub fn receive_invalid_tool_json(
2621 &mut self,
2622 tool_use_id: LanguageModelToolUseId,
2623 tool_name: Arc<str>,
2624 invalid_json: Arc<str>,
2625 error: String,
2626 window: Option<AnyWindowHandle>,
2627 cx: &mut Context<Thread>,
2628 ) {
2629 log::error!("The model returned invalid input JSON: {invalid_json}");
2630
2631 let pending_tool_use = self.tool_use.insert_tool_output(
2632 tool_use_id.clone(),
2633 tool_name,
2634 Err(anyhow!("Error parsing input JSON: {error}")),
2635 self.configured_model.as_ref(),
2636 self.completion_mode,
2637 );
2638 let ui_text = if let Some(pending_tool_use) = &pending_tool_use {
2639 pending_tool_use.ui_text.clone()
2640 } else {
2641 log::error!(
2642 "There was no pending tool use for tool use {tool_use_id}, even though it finished (with invalid input JSON)."
2643 );
2644 format!("Unknown tool {}", tool_use_id).into()
2645 };
2646
2647 cx.emit(ThreadEvent::InvalidToolInput {
2648 tool_use_id: tool_use_id.clone(),
2649 ui_text,
2650 invalid_input_json: invalid_json,
2651 });
2652
2653 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2654 }
2655
2656 pub fn run_tool(
2657 &mut self,
2658 tool_use_id: LanguageModelToolUseId,
2659 ui_text: impl Into<SharedString>,
2660 input: serde_json::Value,
2661 request: Arc<LanguageModelRequest>,
2662 tool: Arc<dyn Tool>,
2663 model: Arc<dyn LanguageModel>,
2664 window: Option<AnyWindowHandle>,
2665 cx: &mut Context<Thread>,
2666 ) {
2667 let task =
2668 self.spawn_tool_use(tool_use_id.clone(), request, input, tool, model, window, cx);
2669 self.tool_use
2670 .run_pending_tool(tool_use_id, ui_text.into(), task);
2671 }
2672
2673 fn spawn_tool_use(
2674 &mut self,
2675 tool_use_id: LanguageModelToolUseId,
2676 request: Arc<LanguageModelRequest>,
2677 input: serde_json::Value,
2678 tool: Arc<dyn Tool>,
2679 model: Arc<dyn LanguageModel>,
2680 window: Option<AnyWindowHandle>,
2681 cx: &mut Context<Thread>,
2682 ) -> Task<()> {
2683 let tool_name: Arc<str> = tool.name().into();
2684
2685 let tool_result = tool.run(
2686 input,
2687 request,
2688 self.project.clone(),
2689 self.action_log.clone(),
2690 model,
2691 window,
2692 cx,
2693 );
2694
2695 // Store the card separately if it exists
2696 if let Some(card) = tool_result.card.clone() {
2697 self.tool_use
2698 .insert_tool_result_card(tool_use_id.clone(), card);
2699 }
2700
2701 cx.spawn({
2702 async move |thread: WeakEntity<Thread>, cx| {
2703 let output = tool_result.output.await;
2704
2705 thread
2706 .update(cx, |thread, cx| {
2707 let pending_tool_use = thread.tool_use.insert_tool_output(
2708 tool_use_id.clone(),
2709 tool_name,
2710 output,
2711 thread.configured_model.as_ref(),
2712 thread.completion_mode,
2713 );
2714 thread.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2715 })
2716 .ok();
2717 }
2718 })
2719 }
2720
2721 fn tool_finished(
2722 &mut self,
2723 tool_use_id: LanguageModelToolUseId,
2724 pending_tool_use: Option<PendingToolUse>,
2725 canceled: bool,
2726 window: Option<AnyWindowHandle>,
2727 cx: &mut Context<Self>,
2728 ) {
2729 if self.all_tools_finished()
2730 && let Some(ConfiguredModel { model, .. }) = self.configured_model.as_ref()
2731 && !canceled
2732 {
2733 self.send_to_model(model.clone(), CompletionIntent::ToolResults, window, cx);
2734 }
2735
2736 cx.emit(ThreadEvent::ToolFinished {
2737 tool_use_id,
2738 pending_tool_use,
2739 });
2740 }
2741
2742 /// Cancels the last pending completion, if there are any pending.
2743 ///
2744 /// Returns whether a completion was canceled.
2745 pub fn cancel_last_completion(
2746 &mut self,
2747 window: Option<AnyWindowHandle>,
2748 cx: &mut Context<Self>,
2749 ) -> bool {
2750 let mut canceled = self.pending_completions.pop().is_some() || self.retry_state.is_some();
2751
2752 self.retry_state = None;
2753
2754 for pending_tool_use in self.tool_use.cancel_pending() {
2755 canceled = true;
2756 self.tool_finished(
2757 pending_tool_use.id.clone(),
2758 Some(pending_tool_use),
2759 true,
2760 window,
2761 cx,
2762 );
2763 }
2764
2765 if canceled {
2766 cx.emit(ThreadEvent::CompletionCanceled);
2767
2768 // When canceled, we always want to insert the checkpoint.
2769 // (We skip over finalize_pending_checkpoint, because it
2770 // would conclude we didn't have anything to insert here.)
2771 if let Some(checkpoint) = self.pending_checkpoint.take() {
2772 self.insert_checkpoint(checkpoint, cx);
2773 }
2774 } else {
2775 self.finalize_pending_checkpoint(cx);
2776 }
2777
2778 canceled
2779 }
2780
2781 /// Signals that any in-progress editing should be canceled.
2782 ///
2783 /// This method is used to notify listeners (like ActiveThread) that
2784 /// they should cancel any editing operations.
2785 pub fn cancel_editing(&mut self, cx: &mut Context<Self>) {
2786 cx.emit(ThreadEvent::CancelEditing);
2787 }
2788
2789 pub fn feedback(&self) -> Option<ThreadFeedback> {
2790 self.feedback
2791 }
2792
2793 pub fn message_feedback(&self, message_id: MessageId) -> Option<ThreadFeedback> {
2794 self.message_feedback.get(&message_id).copied()
2795 }
2796
2797 pub fn report_message_feedback(
2798 &mut self,
2799 message_id: MessageId,
2800 feedback: ThreadFeedback,
2801 cx: &mut Context<Self>,
2802 ) -> Task<Result<()>> {
2803 if self.message_feedback.get(&message_id) == Some(&feedback) {
2804 return Task::ready(Ok(()));
2805 }
2806
2807 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2808 let serialized_thread = self.serialize(cx);
2809 let thread_id = self.id().clone();
2810 let client = self.project.read(cx).client();
2811
2812 let enabled_tool_names: Vec<String> = self
2813 .profile
2814 .enabled_tools(cx)
2815 .iter()
2816 .map(|(name, _)| name.clone().into())
2817 .collect();
2818
2819 self.message_feedback.insert(message_id, feedback);
2820
2821 cx.notify();
2822
2823 let message_content = self
2824 .message(message_id)
2825 .map(|msg| msg.to_string())
2826 .unwrap_or_default();
2827
2828 cx.background_spawn(async move {
2829 let final_project_snapshot = final_project_snapshot.await;
2830 let serialized_thread = serialized_thread.await?;
2831 let thread_data =
2832 serde_json::to_value(serialized_thread).unwrap_or_else(|_| serde_json::Value::Null);
2833
2834 let rating = match feedback {
2835 ThreadFeedback::Positive => "positive",
2836 ThreadFeedback::Negative => "negative",
2837 };
2838 telemetry::event!(
2839 "Assistant Thread Rated",
2840 rating,
2841 thread_id,
2842 enabled_tool_names,
2843 message_id = message_id.0,
2844 message_content,
2845 thread_data,
2846 final_project_snapshot
2847 );
2848 client.telemetry().flush_events().await;
2849
2850 Ok(())
2851 })
2852 }
2853
2854 pub fn report_feedback(
2855 &mut self,
2856 feedback: ThreadFeedback,
2857 cx: &mut Context<Self>,
2858 ) -> Task<Result<()>> {
2859 let last_assistant_message_id = self
2860 .messages
2861 .iter()
2862 .rev()
2863 .find(|msg| msg.role == Role::Assistant)
2864 .map(|msg| msg.id);
2865
2866 if let Some(message_id) = last_assistant_message_id {
2867 self.report_message_feedback(message_id, feedback, cx)
2868 } else {
2869 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2870 let serialized_thread = self.serialize(cx);
2871 let thread_id = self.id().clone();
2872 let client = self.project.read(cx).client();
2873 self.feedback = Some(feedback);
2874 cx.notify();
2875
2876 cx.background_spawn(async move {
2877 let final_project_snapshot = final_project_snapshot.await;
2878 let serialized_thread = serialized_thread.await?;
2879 let thread_data = serde_json::to_value(serialized_thread)
2880 .unwrap_or_else(|_| serde_json::Value::Null);
2881
2882 let rating = match feedback {
2883 ThreadFeedback::Positive => "positive",
2884 ThreadFeedback::Negative => "negative",
2885 };
2886 telemetry::event!(
2887 "Assistant Thread Rated",
2888 rating,
2889 thread_id,
2890 thread_data,
2891 final_project_snapshot
2892 );
2893 client.telemetry().flush_events().await;
2894
2895 Ok(())
2896 })
2897 }
2898 }
2899
2900 /// Create a snapshot of the current project state including git information and unsaved buffers.
2901 fn project_snapshot(
2902 project: Entity<Project>,
2903 cx: &mut Context<Self>,
2904 ) -> Task<Arc<ProjectSnapshot>> {
2905 let git_store = project.read(cx).git_store().clone();
2906 let worktree_snapshots: Vec<_> = project
2907 .read(cx)
2908 .visible_worktrees(cx)
2909 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
2910 .collect();
2911
2912 cx.spawn(async move |_, cx| {
2913 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
2914
2915 let mut unsaved_buffers = Vec::new();
2916 cx.update(|app_cx| {
2917 let buffer_store = project.read(app_cx).buffer_store();
2918 for buffer_handle in buffer_store.read(app_cx).buffers() {
2919 let buffer = buffer_handle.read(app_cx);
2920 if buffer.is_dirty()
2921 && let Some(file) = buffer.file()
2922 {
2923 let path = file.path().to_string_lossy().to_string();
2924 unsaved_buffers.push(path);
2925 }
2926 }
2927 })
2928 .ok();
2929
2930 Arc::new(ProjectSnapshot {
2931 worktree_snapshots,
2932 unsaved_buffer_paths: unsaved_buffers,
2933 timestamp: Utc::now(),
2934 })
2935 })
2936 }
2937
2938 fn worktree_snapshot(
2939 worktree: Entity<project::Worktree>,
2940 git_store: Entity<GitStore>,
2941 cx: &App,
2942 ) -> Task<WorktreeSnapshot> {
2943 cx.spawn(async move |cx| {
2944 // Get worktree path and snapshot
2945 let worktree_info = cx.update(|app_cx| {
2946 let worktree = worktree.read(app_cx);
2947 let path = worktree.abs_path().to_string_lossy().to_string();
2948 let snapshot = worktree.snapshot();
2949 (path, snapshot)
2950 });
2951
2952 let Ok((worktree_path, _snapshot)) = worktree_info else {
2953 return WorktreeSnapshot {
2954 worktree_path: String::new(),
2955 git_state: None,
2956 };
2957 };
2958
2959 let git_state = git_store
2960 .update(cx, |git_store, cx| {
2961 git_store
2962 .repositories()
2963 .values()
2964 .find(|repo| {
2965 repo.read(cx)
2966 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
2967 .is_some()
2968 })
2969 .cloned()
2970 })
2971 .ok()
2972 .flatten()
2973 .map(|repo| {
2974 repo.update(cx, |repo, _| {
2975 let current_branch =
2976 repo.branch.as_ref().map(|branch| branch.name().to_owned());
2977 repo.send_job(None, |state, _| async move {
2978 let RepositoryState::Local { backend, .. } = state else {
2979 return GitState {
2980 remote_url: None,
2981 head_sha: None,
2982 current_branch,
2983 diff: None,
2984 };
2985 };
2986
2987 let remote_url = backend.remote_url("origin");
2988 let head_sha = backend.head_sha().await;
2989 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
2990
2991 GitState {
2992 remote_url,
2993 head_sha,
2994 current_branch,
2995 diff,
2996 }
2997 })
2998 })
2999 });
3000
3001 let git_state = match git_state {
3002 Some(git_state) => match git_state.ok() {
3003 Some(git_state) => git_state.await.ok(),
3004 None => None,
3005 },
3006 None => None,
3007 };
3008
3009 WorktreeSnapshot {
3010 worktree_path,
3011 git_state,
3012 }
3013 })
3014 }
3015
3016 pub fn to_markdown(&self, cx: &App) -> Result<String> {
3017 let mut markdown = Vec::new();
3018
3019 let summary = self.summary().or_default();
3020 writeln!(markdown, "# {summary}\n")?;
3021
3022 for message in self.messages() {
3023 writeln!(
3024 markdown,
3025 "## {role}\n",
3026 role = match message.role {
3027 Role::User => "User",
3028 Role::Assistant => "Agent",
3029 Role::System => "System",
3030 }
3031 )?;
3032
3033 if !message.loaded_context.text.is_empty() {
3034 writeln!(markdown, "{}", message.loaded_context.text)?;
3035 }
3036
3037 if !message.loaded_context.images.is_empty() {
3038 writeln!(
3039 markdown,
3040 "\n{} images attached as context.\n",
3041 message.loaded_context.images.len()
3042 )?;
3043 }
3044
3045 for segment in &message.segments {
3046 match segment {
3047 MessageSegment::Text(text) => writeln!(markdown, "{}\n", text)?,
3048 MessageSegment::Thinking { text, .. } => {
3049 writeln!(markdown, "<think>\n{}\n</think>\n", text)?
3050 }
3051 MessageSegment::RedactedThinking(_) => {}
3052 }
3053 }
3054
3055 for tool_use in self.tool_uses_for_message(message.id, cx) {
3056 writeln!(
3057 markdown,
3058 "**Use Tool: {} ({})**",
3059 tool_use.name, tool_use.id
3060 )?;
3061 writeln!(markdown, "```json")?;
3062 writeln!(
3063 markdown,
3064 "{}",
3065 serde_json::to_string_pretty(&tool_use.input)?
3066 )?;
3067 writeln!(markdown, "```")?;
3068 }
3069
3070 for tool_result in self.tool_results_for_message(message.id) {
3071 write!(markdown, "\n**Tool Results: {}", tool_result.tool_use_id)?;
3072 if tool_result.is_error {
3073 write!(markdown, " (Error)")?;
3074 }
3075
3076 writeln!(markdown, "**\n")?;
3077 match &tool_result.content {
3078 LanguageModelToolResultContent::Text(text) => {
3079 writeln!(markdown, "{text}")?;
3080 }
3081 LanguageModelToolResultContent::Image(image) => {
3082 writeln!(markdown, "", image.source)?;
3083 }
3084 }
3085
3086 if let Some(output) = tool_result.output.as_ref() {
3087 writeln!(
3088 markdown,
3089 "\n\nDebug Output:\n\n```json\n{}\n```\n",
3090 serde_json::to_string_pretty(output)?
3091 )?;
3092 }
3093 }
3094 }
3095
3096 Ok(String::from_utf8_lossy(&markdown).to_string())
3097 }
3098
3099 pub fn keep_edits_in_range(
3100 &mut self,
3101 buffer: Entity<language::Buffer>,
3102 buffer_range: Range<language::Anchor>,
3103 cx: &mut Context<Self>,
3104 ) {
3105 self.action_log.update(cx, |action_log, cx| {
3106 action_log.keep_edits_in_range(buffer, buffer_range, cx)
3107 });
3108 }
3109
3110 pub fn keep_all_edits(&mut self, cx: &mut Context<Self>) {
3111 self.action_log
3112 .update(cx, |action_log, cx| action_log.keep_all_edits(cx));
3113 }
3114
3115 pub fn reject_edits_in_ranges(
3116 &mut self,
3117 buffer: Entity<language::Buffer>,
3118 buffer_ranges: Vec<Range<language::Anchor>>,
3119 cx: &mut Context<Self>,
3120 ) -> Task<Result<()>> {
3121 self.action_log.update(cx, |action_log, cx| {
3122 action_log.reject_edits_in_ranges(buffer, buffer_ranges, cx)
3123 })
3124 }
3125
3126 pub fn action_log(&self) -> &Entity<ActionLog> {
3127 &self.action_log
3128 }
3129
3130 pub fn project(&self) -> &Entity<Project> {
3131 &self.project
3132 }
3133
3134 pub fn cumulative_token_usage(&self) -> TokenUsage {
3135 self.cumulative_token_usage
3136 }
3137
3138 pub fn token_usage_up_to_message(&self, message_id: MessageId) -> TotalTokenUsage {
3139 let Some(model) = self.configured_model.as_ref() else {
3140 return TotalTokenUsage::default();
3141 };
3142
3143 let max = model
3144 .model
3145 .max_token_count_for_mode(self.completion_mode().into());
3146
3147 let index = self
3148 .messages
3149 .iter()
3150 .position(|msg| msg.id == message_id)
3151 .unwrap_or(0);
3152
3153 if index == 0 {
3154 return TotalTokenUsage { total: 0, max };
3155 }
3156
3157 let token_usage = &self
3158 .request_token_usage
3159 .get(index - 1)
3160 .cloned()
3161 .unwrap_or_default();
3162
3163 TotalTokenUsage {
3164 total: token_usage.total_tokens(),
3165 max,
3166 }
3167 }
3168
3169 pub fn total_token_usage(&self) -> Option<TotalTokenUsage> {
3170 let model = self.configured_model.as_ref()?;
3171
3172 let max = model
3173 .model
3174 .max_token_count_for_mode(self.completion_mode().into());
3175
3176 if let Some(exceeded_error) = &self.exceeded_window_error
3177 && model.model.id() == exceeded_error.model_id
3178 {
3179 return Some(TotalTokenUsage {
3180 total: exceeded_error.token_count,
3181 max,
3182 });
3183 }
3184
3185 let total = self
3186 .token_usage_at_last_message()
3187 .unwrap_or_default()
3188 .total_tokens();
3189
3190 Some(TotalTokenUsage { total, max })
3191 }
3192
3193 fn token_usage_at_last_message(&self) -> Option<TokenUsage> {
3194 self.request_token_usage
3195 .get(self.messages.len().saturating_sub(1))
3196 .or_else(|| self.request_token_usage.last())
3197 .cloned()
3198 }
3199
3200 fn update_token_usage_at_last_message(&mut self, token_usage: TokenUsage) {
3201 let placeholder = self.token_usage_at_last_message().unwrap_or_default();
3202 self.request_token_usage
3203 .resize(self.messages.len(), placeholder);
3204
3205 if let Some(last) = self.request_token_usage.last_mut() {
3206 *last = token_usage;
3207 }
3208 }
3209
3210 fn update_model_request_usage(&self, amount: u32, limit: UsageLimit, cx: &mut Context<Self>) {
3211 self.project
3212 .read(cx)
3213 .user_store()
3214 .update(cx, |user_store, cx| {
3215 user_store.update_model_request_usage(
3216 ModelRequestUsage(RequestUsage {
3217 amount: amount as i32,
3218 limit,
3219 }),
3220 cx,
3221 )
3222 });
3223 }
3224
3225 pub fn deny_tool_use(
3226 &mut self,
3227 tool_use_id: LanguageModelToolUseId,
3228 tool_name: Arc<str>,
3229 window: Option<AnyWindowHandle>,
3230 cx: &mut Context<Self>,
3231 ) {
3232 let err = Err(anyhow::anyhow!(
3233 "Permission to run tool action denied by user"
3234 ));
3235
3236 self.tool_use.insert_tool_output(
3237 tool_use_id.clone(),
3238 tool_name,
3239 err,
3240 self.configured_model.as_ref(),
3241 self.completion_mode,
3242 );
3243 self.tool_finished(tool_use_id.clone(), None, true, window, cx);
3244 }
3245}
3246
3247#[derive(Debug, Clone, Error)]
3248pub enum ThreadError {
3249 #[error("Payment required")]
3250 PaymentRequired,
3251 #[error("Model request limit reached")]
3252 ModelRequestLimitReached { plan: Plan },
3253 #[error("Message {header}: {message}")]
3254 Message {
3255 header: SharedString,
3256 message: SharedString,
3257 },
3258 #[error("Retryable error: {message}")]
3259 RetryableError {
3260 message: SharedString,
3261 can_enable_burn_mode: bool,
3262 },
3263}
3264
3265#[derive(Debug, Clone)]
3266pub enum ThreadEvent {
3267 ShowError(ThreadError),
3268 StreamedCompletion,
3269 ReceivedTextChunk,
3270 NewRequest,
3271 StreamedAssistantText(MessageId, String),
3272 StreamedAssistantThinking(MessageId, String),
3273 StreamedToolUse {
3274 tool_use_id: LanguageModelToolUseId,
3275 ui_text: Arc<str>,
3276 input: serde_json::Value,
3277 },
3278 MissingToolUse {
3279 tool_use_id: LanguageModelToolUseId,
3280 ui_text: Arc<str>,
3281 },
3282 InvalidToolInput {
3283 tool_use_id: LanguageModelToolUseId,
3284 ui_text: Arc<str>,
3285 invalid_input_json: Arc<str>,
3286 },
3287 Stopped(Result<StopReason, Arc<anyhow::Error>>),
3288 MessageAdded(MessageId),
3289 MessageEdited(MessageId),
3290 MessageDeleted(MessageId),
3291 SummaryGenerated,
3292 SummaryChanged,
3293 UsePendingTools {
3294 tool_uses: Vec<PendingToolUse>,
3295 },
3296 ToolFinished {
3297 #[allow(unused)]
3298 tool_use_id: LanguageModelToolUseId,
3299 /// The pending tool use that corresponds to this tool.
3300 pending_tool_use: Option<PendingToolUse>,
3301 },
3302 CheckpointChanged,
3303 ToolConfirmationNeeded,
3304 ToolUseLimitReached,
3305 CancelEditing,
3306 CompletionCanceled,
3307 ProfileChanged,
3308}
3309
3310impl EventEmitter<ThreadEvent> for Thread {}
3311
3312struct PendingCompletion {
3313 id: usize,
3314 queue_state: QueueState,
3315 _task: Task<()>,
3316}
3317
3318#[cfg(test)]
3319mod tests {
3320 use super::*;
3321 use crate::{
3322 context::load_context, context_store::ContextStore, thread_store, thread_store::ThreadStore,
3323 };
3324
3325 // Test-specific constants
3326 const TEST_RATE_LIMIT_RETRY_SECS: u64 = 30;
3327 use agent_settings::{AgentProfileId, AgentSettings, LanguageModelParameters};
3328 use assistant_tool::ToolRegistry;
3329 use assistant_tools;
3330 use futures::StreamExt;
3331 use futures::future::BoxFuture;
3332 use futures::stream::BoxStream;
3333 use gpui::TestAppContext;
3334 use http_client;
3335 use language_model::fake_provider::{FakeLanguageModel, FakeLanguageModelProvider};
3336 use language_model::{
3337 LanguageModelCompletionError, LanguageModelName, LanguageModelProviderId,
3338 LanguageModelProviderName, LanguageModelToolChoice,
3339 };
3340 use parking_lot::Mutex;
3341 use project::{FakeFs, Project};
3342 use prompt_store::PromptBuilder;
3343 use serde_json::json;
3344 use settings::{Settings, SettingsStore};
3345 use std::sync::Arc;
3346 use std::time::Duration;
3347 use theme::ThemeSettings;
3348 use util::path;
3349 use workspace::Workspace;
3350
3351 #[gpui::test]
3352 async fn test_message_with_context(cx: &mut TestAppContext) {
3353 init_test_settings(cx);
3354
3355 let project = create_test_project(
3356 cx,
3357 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3358 )
3359 .await;
3360
3361 let (_workspace, _thread_store, thread, context_store, model) =
3362 setup_test_environment(cx, project.clone()).await;
3363
3364 add_file_to_context(&project, &context_store, "test/code.rs", cx)
3365 .await
3366 .unwrap();
3367
3368 let context =
3369 context_store.read_with(cx, |store, _| store.context().next().cloned().unwrap());
3370 let loaded_context = cx
3371 .update(|cx| load_context(vec![context], &project, &None, cx))
3372 .await;
3373
3374 // Insert user message with context
3375 let message_id = thread.update(cx, |thread, cx| {
3376 thread.insert_user_message(
3377 "Please explain this code",
3378 loaded_context,
3379 None,
3380 Vec::new(),
3381 cx,
3382 )
3383 });
3384
3385 // Check content and context in message object
3386 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3387
3388 // Use different path format strings based on platform for the test
3389 #[cfg(windows)]
3390 let path_part = r"test\code.rs";
3391 #[cfg(not(windows))]
3392 let path_part = "test/code.rs";
3393
3394 let expected_context = format!(
3395 r#"
3396<context>
3397The following items were attached by the user. They are up-to-date and don't need to be re-read.
3398
3399<files>
3400```rs {path_part}
3401fn main() {{
3402 println!("Hello, world!");
3403}}
3404```
3405</files>
3406</context>
3407"#
3408 );
3409
3410 assert_eq!(message.role, Role::User);
3411 assert_eq!(message.segments.len(), 1);
3412 assert_eq!(
3413 message.segments[0],
3414 MessageSegment::Text("Please explain this code".to_string())
3415 );
3416 assert_eq!(message.loaded_context.text, expected_context);
3417
3418 // Check message in request
3419 let request = thread.update(cx, |thread, cx| {
3420 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3421 });
3422
3423 assert_eq!(request.messages.len(), 2);
3424 let expected_full_message = format!("{}Please explain this code", expected_context);
3425 assert_eq!(request.messages[1].string_contents(), expected_full_message);
3426 }
3427
3428 #[gpui::test]
3429 async fn test_only_include_new_contexts(cx: &mut TestAppContext) {
3430 init_test_settings(cx);
3431
3432 let project = create_test_project(
3433 cx,
3434 json!({
3435 "file1.rs": "fn function1() {}\n",
3436 "file2.rs": "fn function2() {}\n",
3437 "file3.rs": "fn function3() {}\n",
3438 "file4.rs": "fn function4() {}\n",
3439 }),
3440 )
3441 .await;
3442
3443 let (_, _thread_store, thread, context_store, model) =
3444 setup_test_environment(cx, project.clone()).await;
3445
3446 // First message with context 1
3447 add_file_to_context(&project, &context_store, "test/file1.rs", cx)
3448 .await
3449 .unwrap();
3450 let new_contexts = context_store.update(cx, |store, cx| {
3451 store.new_context_for_thread(thread.read(cx), None)
3452 });
3453 assert_eq!(new_contexts.len(), 1);
3454 let loaded_context = cx
3455 .update(|cx| load_context(new_contexts, &project, &None, cx))
3456 .await;
3457 let message1_id = thread.update(cx, |thread, cx| {
3458 thread.insert_user_message("Message 1", loaded_context, None, Vec::new(), cx)
3459 });
3460
3461 // Second message with contexts 1 and 2 (context 1 should be skipped as it's already included)
3462 add_file_to_context(&project, &context_store, "test/file2.rs", cx)
3463 .await
3464 .unwrap();
3465 let new_contexts = context_store.update(cx, |store, cx| {
3466 store.new_context_for_thread(thread.read(cx), None)
3467 });
3468 assert_eq!(new_contexts.len(), 1);
3469 let loaded_context = cx
3470 .update(|cx| load_context(new_contexts, &project, &None, cx))
3471 .await;
3472 let message2_id = thread.update(cx, |thread, cx| {
3473 thread.insert_user_message("Message 2", loaded_context, None, Vec::new(), cx)
3474 });
3475
3476 // Third message with all three contexts (contexts 1 and 2 should be skipped)
3477 //
3478 add_file_to_context(&project, &context_store, "test/file3.rs", cx)
3479 .await
3480 .unwrap();
3481 let new_contexts = context_store.update(cx, |store, cx| {
3482 store.new_context_for_thread(thread.read(cx), None)
3483 });
3484 assert_eq!(new_contexts.len(), 1);
3485 let loaded_context = cx
3486 .update(|cx| load_context(new_contexts, &project, &None, cx))
3487 .await;
3488 let message3_id = thread.update(cx, |thread, cx| {
3489 thread.insert_user_message("Message 3", loaded_context, None, Vec::new(), cx)
3490 });
3491
3492 // Check what contexts are included in each message
3493 let (message1, message2, message3) = thread.read_with(cx, |thread, _| {
3494 (
3495 thread.message(message1_id).unwrap().clone(),
3496 thread.message(message2_id).unwrap().clone(),
3497 thread.message(message3_id).unwrap().clone(),
3498 )
3499 });
3500
3501 // First message should include context 1
3502 assert!(message1.loaded_context.text.contains("file1.rs"));
3503
3504 // Second message should include only context 2 (not 1)
3505 assert!(!message2.loaded_context.text.contains("file1.rs"));
3506 assert!(message2.loaded_context.text.contains("file2.rs"));
3507
3508 // Third message should include only context 3 (not 1 or 2)
3509 assert!(!message3.loaded_context.text.contains("file1.rs"));
3510 assert!(!message3.loaded_context.text.contains("file2.rs"));
3511 assert!(message3.loaded_context.text.contains("file3.rs"));
3512
3513 // Check entire request to make sure all contexts are properly included
3514 let request = thread.update(cx, |thread, cx| {
3515 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3516 });
3517
3518 // The request should contain all 3 messages
3519 assert_eq!(request.messages.len(), 4);
3520
3521 // Check that the contexts are properly formatted in each message
3522 assert!(request.messages[1].string_contents().contains("file1.rs"));
3523 assert!(!request.messages[1].string_contents().contains("file2.rs"));
3524 assert!(!request.messages[1].string_contents().contains("file3.rs"));
3525
3526 assert!(!request.messages[2].string_contents().contains("file1.rs"));
3527 assert!(request.messages[2].string_contents().contains("file2.rs"));
3528 assert!(!request.messages[2].string_contents().contains("file3.rs"));
3529
3530 assert!(!request.messages[3].string_contents().contains("file1.rs"));
3531 assert!(!request.messages[3].string_contents().contains("file2.rs"));
3532 assert!(request.messages[3].string_contents().contains("file3.rs"));
3533
3534 add_file_to_context(&project, &context_store, "test/file4.rs", cx)
3535 .await
3536 .unwrap();
3537 let new_contexts = context_store.update(cx, |store, cx| {
3538 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3539 });
3540 assert_eq!(new_contexts.len(), 3);
3541 let loaded_context = cx
3542 .update(|cx| load_context(new_contexts, &project, &None, cx))
3543 .await
3544 .loaded_context;
3545
3546 assert!(!loaded_context.text.contains("file1.rs"));
3547 assert!(loaded_context.text.contains("file2.rs"));
3548 assert!(loaded_context.text.contains("file3.rs"));
3549 assert!(loaded_context.text.contains("file4.rs"));
3550
3551 let new_contexts = context_store.update(cx, |store, cx| {
3552 // Remove file4.rs
3553 store.remove_context(&loaded_context.contexts[2].handle(), cx);
3554 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3555 });
3556 assert_eq!(new_contexts.len(), 2);
3557 let loaded_context = cx
3558 .update(|cx| load_context(new_contexts, &project, &None, cx))
3559 .await
3560 .loaded_context;
3561
3562 assert!(!loaded_context.text.contains("file1.rs"));
3563 assert!(loaded_context.text.contains("file2.rs"));
3564 assert!(loaded_context.text.contains("file3.rs"));
3565 assert!(!loaded_context.text.contains("file4.rs"));
3566
3567 let new_contexts = context_store.update(cx, |store, cx| {
3568 // Remove file3.rs
3569 store.remove_context(&loaded_context.contexts[1].handle(), cx);
3570 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3571 });
3572 assert_eq!(new_contexts.len(), 1);
3573 let loaded_context = cx
3574 .update(|cx| load_context(new_contexts, &project, &None, cx))
3575 .await
3576 .loaded_context;
3577
3578 assert!(!loaded_context.text.contains("file1.rs"));
3579 assert!(loaded_context.text.contains("file2.rs"));
3580 assert!(!loaded_context.text.contains("file3.rs"));
3581 assert!(!loaded_context.text.contains("file4.rs"));
3582 }
3583
3584 #[gpui::test]
3585 async fn test_message_without_files(cx: &mut TestAppContext) {
3586 init_test_settings(cx);
3587
3588 let project = create_test_project(
3589 cx,
3590 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3591 )
3592 .await;
3593
3594 let (_, _thread_store, thread, _context_store, model) =
3595 setup_test_environment(cx, project.clone()).await;
3596
3597 // Insert user message without any context (empty context vector)
3598 let message_id = thread.update(cx, |thread, cx| {
3599 thread.insert_user_message(
3600 "What is the best way to learn Rust?",
3601 ContextLoadResult::default(),
3602 None,
3603 Vec::new(),
3604 cx,
3605 )
3606 });
3607
3608 // Check content and context in message object
3609 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3610
3611 // Context should be empty when no files are included
3612 assert_eq!(message.role, Role::User);
3613 assert_eq!(message.segments.len(), 1);
3614 assert_eq!(
3615 message.segments[0],
3616 MessageSegment::Text("What is the best way to learn Rust?".to_string())
3617 );
3618 assert_eq!(message.loaded_context.text, "");
3619
3620 // Check message in request
3621 let request = thread.update(cx, |thread, cx| {
3622 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3623 });
3624
3625 assert_eq!(request.messages.len(), 2);
3626 assert_eq!(
3627 request.messages[1].string_contents(),
3628 "What is the best way to learn Rust?"
3629 );
3630
3631 // Add second message, also without context
3632 let message2_id = thread.update(cx, |thread, cx| {
3633 thread.insert_user_message(
3634 "Are there any good books?",
3635 ContextLoadResult::default(),
3636 None,
3637 Vec::new(),
3638 cx,
3639 )
3640 });
3641
3642 let message2 =
3643 thread.read_with(cx, |thread, _| thread.message(message2_id).unwrap().clone());
3644 assert_eq!(message2.loaded_context.text, "");
3645
3646 // Check that both messages appear in the request
3647 let request = thread.update(cx, |thread, cx| {
3648 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3649 });
3650
3651 assert_eq!(request.messages.len(), 3);
3652 assert_eq!(
3653 request.messages[1].string_contents(),
3654 "What is the best way to learn Rust?"
3655 );
3656 assert_eq!(
3657 request.messages[2].string_contents(),
3658 "Are there any good books?"
3659 );
3660 }
3661
3662 #[gpui::test]
3663 #[ignore] // turn this test on when project_notifications tool is re-enabled
3664 async fn test_stale_buffer_notification(cx: &mut TestAppContext) {
3665 init_test_settings(cx);
3666
3667 let project = create_test_project(
3668 cx,
3669 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3670 )
3671 .await;
3672
3673 let (_workspace, _thread_store, thread, context_store, model) =
3674 setup_test_environment(cx, project.clone()).await;
3675
3676 // Add a buffer to the context. This will be a tracked buffer
3677 let buffer = add_file_to_context(&project, &context_store, "test/code.rs", cx)
3678 .await
3679 .unwrap();
3680
3681 let context = context_store
3682 .read_with(cx, |store, _| store.context().next().cloned())
3683 .unwrap();
3684 let loaded_context = cx
3685 .update(|cx| load_context(vec![context], &project, &None, cx))
3686 .await;
3687
3688 // Insert user message and assistant response
3689 thread.update(cx, |thread, cx| {
3690 thread.insert_user_message("Explain this code", loaded_context, None, Vec::new(), cx);
3691 thread.insert_assistant_message(
3692 vec![MessageSegment::Text("This code prints 42.".into())],
3693 cx,
3694 );
3695 });
3696 cx.run_until_parked();
3697
3698 // We shouldn't have a stale buffer notification yet
3699 let notifications = thread.read_with(cx, |thread, _| {
3700 find_tool_uses(thread, "project_notifications")
3701 });
3702 assert!(
3703 notifications.is_empty(),
3704 "Should not have stale buffer notification before buffer is modified"
3705 );
3706
3707 // Modify the buffer
3708 buffer.update(cx, |buffer, cx| {
3709 buffer.edit(
3710 [(1..1, "\n println!(\"Added a new line\");\n")],
3711 None,
3712 cx,
3713 );
3714 });
3715
3716 // Insert another user message
3717 thread.update(cx, |thread, cx| {
3718 thread.insert_user_message(
3719 "What does the code do now?",
3720 ContextLoadResult::default(),
3721 None,
3722 Vec::new(),
3723 cx,
3724 )
3725 });
3726 cx.run_until_parked();
3727
3728 // Check for the stale buffer warning
3729 thread.update(cx, |thread, cx| {
3730 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3731 });
3732 cx.run_until_parked();
3733
3734 let notifications = thread.read_with(cx, |thread, _cx| {
3735 find_tool_uses(thread, "project_notifications")
3736 });
3737
3738 let [notification] = notifications.as_slice() else {
3739 panic!("Should have a `project_notifications` tool use");
3740 };
3741
3742 let Some(notification_content) = notification.content.to_str() else {
3743 panic!("`project_notifications` should return text");
3744 };
3745
3746 assert!(notification_content.contains("These files have changed since the last read:"));
3747 assert!(notification_content.contains("code.rs"));
3748
3749 // Insert another user message and flush notifications again
3750 thread.update(cx, |thread, cx| {
3751 thread.insert_user_message(
3752 "Can you tell me more?",
3753 ContextLoadResult::default(),
3754 None,
3755 Vec::new(),
3756 cx,
3757 )
3758 });
3759
3760 thread.update(cx, |thread, cx| {
3761 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3762 });
3763 cx.run_until_parked();
3764
3765 // There should be no new notifications (we already flushed one)
3766 let notifications = thread.read_with(cx, |thread, _cx| {
3767 find_tool_uses(thread, "project_notifications")
3768 });
3769
3770 assert_eq!(
3771 notifications.len(),
3772 1,
3773 "Should still have only one notification after second flush - no duplicates"
3774 );
3775 }
3776
3777 fn find_tool_uses(thread: &Thread, tool_name: &str) -> Vec<LanguageModelToolResult> {
3778 thread
3779 .messages()
3780 .flat_map(|message| {
3781 thread
3782 .tool_results_for_message(message.id)
3783 .into_iter()
3784 .filter(|result| result.tool_name == tool_name.into())
3785 .cloned()
3786 .collect::<Vec<_>>()
3787 })
3788 .collect()
3789 }
3790
3791 #[gpui::test]
3792 async fn test_storing_profile_setting_per_thread(cx: &mut TestAppContext) {
3793 init_test_settings(cx);
3794
3795 let project = create_test_project(
3796 cx,
3797 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3798 )
3799 .await;
3800
3801 let (_workspace, thread_store, thread, _context_store, _model) =
3802 setup_test_environment(cx, project.clone()).await;
3803
3804 // Check that we are starting with the default profile
3805 let profile = cx.read(|cx| thread.read(cx).profile.clone());
3806 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3807 assert_eq!(
3808 profile,
3809 AgentProfile::new(AgentProfileId::default(), tool_set)
3810 );
3811 }
3812
3813 #[gpui::test]
3814 async fn test_serializing_thread_profile(cx: &mut TestAppContext) {
3815 init_test_settings(cx);
3816
3817 let project = create_test_project(
3818 cx,
3819 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3820 )
3821 .await;
3822
3823 let (_workspace, thread_store, thread, _context_store, _model) =
3824 setup_test_environment(cx, project.clone()).await;
3825
3826 // Profile gets serialized with default values
3827 let serialized = thread
3828 .update(cx, |thread, cx| thread.serialize(cx))
3829 .await
3830 .unwrap();
3831
3832 assert_eq!(serialized.profile, Some(AgentProfileId::default()));
3833
3834 let deserialized = cx.update(|cx| {
3835 thread.update(cx, |thread, cx| {
3836 Thread::deserialize(
3837 thread.id.clone(),
3838 serialized,
3839 thread.project.clone(),
3840 thread.tools.clone(),
3841 thread.prompt_builder.clone(),
3842 thread.project_context.clone(),
3843 None,
3844 cx,
3845 )
3846 })
3847 });
3848 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3849
3850 assert_eq!(
3851 deserialized.profile,
3852 AgentProfile::new(AgentProfileId::default(), tool_set)
3853 );
3854 }
3855
3856 #[gpui::test]
3857 async fn test_temperature_setting(cx: &mut TestAppContext) {
3858 init_test_settings(cx);
3859
3860 let project = create_test_project(
3861 cx,
3862 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3863 )
3864 .await;
3865
3866 let (_workspace, _thread_store, thread, _context_store, model) =
3867 setup_test_environment(cx, project.clone()).await;
3868
3869 // Both model and provider
3870 cx.update(|cx| {
3871 AgentSettings::override_global(
3872 AgentSettings {
3873 model_parameters: vec![LanguageModelParameters {
3874 provider: Some(model.provider_id().0.to_string().into()),
3875 model: Some(model.id().0.clone()),
3876 temperature: Some(0.66),
3877 }],
3878 ..AgentSettings::get_global(cx).clone()
3879 },
3880 cx,
3881 );
3882 });
3883
3884 let request = thread.update(cx, |thread, cx| {
3885 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3886 });
3887 assert_eq!(request.temperature, Some(0.66));
3888
3889 // Only model
3890 cx.update(|cx| {
3891 AgentSettings::override_global(
3892 AgentSettings {
3893 model_parameters: vec![LanguageModelParameters {
3894 provider: None,
3895 model: Some(model.id().0.clone()),
3896 temperature: Some(0.66),
3897 }],
3898 ..AgentSettings::get_global(cx).clone()
3899 },
3900 cx,
3901 );
3902 });
3903
3904 let request = thread.update(cx, |thread, cx| {
3905 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3906 });
3907 assert_eq!(request.temperature, Some(0.66));
3908
3909 // Only provider
3910 cx.update(|cx| {
3911 AgentSettings::override_global(
3912 AgentSettings {
3913 model_parameters: vec![LanguageModelParameters {
3914 provider: Some(model.provider_id().0.to_string().into()),
3915 model: None,
3916 temperature: Some(0.66),
3917 }],
3918 ..AgentSettings::get_global(cx).clone()
3919 },
3920 cx,
3921 );
3922 });
3923
3924 let request = thread.update(cx, |thread, cx| {
3925 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3926 });
3927 assert_eq!(request.temperature, Some(0.66));
3928
3929 // Same model name, different provider
3930 cx.update(|cx| {
3931 AgentSettings::override_global(
3932 AgentSettings {
3933 model_parameters: vec![LanguageModelParameters {
3934 provider: Some("anthropic".into()),
3935 model: Some(model.id().0.clone()),
3936 temperature: Some(0.66),
3937 }],
3938 ..AgentSettings::get_global(cx).clone()
3939 },
3940 cx,
3941 );
3942 });
3943
3944 let request = thread.update(cx, |thread, cx| {
3945 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3946 });
3947 assert_eq!(request.temperature, None);
3948 }
3949
3950 #[gpui::test]
3951 async fn test_thread_summary(cx: &mut TestAppContext) {
3952 init_test_settings(cx);
3953
3954 let project = create_test_project(cx, json!({})).await;
3955
3956 let (_, _thread_store, thread, _context_store, model) =
3957 setup_test_environment(cx, project.clone()).await;
3958
3959 // Initial state should be pending
3960 thread.read_with(cx, |thread, _| {
3961 assert!(matches!(thread.summary(), ThreadSummary::Pending));
3962 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
3963 });
3964
3965 // Manually setting the summary should not be allowed in this state
3966 thread.update(cx, |thread, cx| {
3967 thread.set_summary("This should not work", cx);
3968 });
3969
3970 thread.read_with(cx, |thread, _| {
3971 assert!(matches!(thread.summary(), ThreadSummary::Pending));
3972 });
3973
3974 // Send a message
3975 thread.update(cx, |thread, cx| {
3976 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
3977 thread.send_to_model(
3978 model.clone(),
3979 CompletionIntent::ThreadSummarization,
3980 None,
3981 cx,
3982 );
3983 });
3984
3985 let fake_model = model.as_fake();
3986 simulate_successful_response(fake_model, cx);
3987
3988 // Should start generating summary when there are >= 2 messages
3989 thread.read_with(cx, |thread, _| {
3990 assert_eq!(*thread.summary(), ThreadSummary::Generating);
3991 });
3992
3993 // Should not be able to set the summary while generating
3994 thread.update(cx, |thread, cx| {
3995 thread.set_summary("This should not work either", cx);
3996 });
3997
3998 thread.read_with(cx, |thread, _| {
3999 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4000 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4001 });
4002
4003 cx.run_until_parked();
4004 fake_model.send_last_completion_stream_text_chunk("Brief");
4005 fake_model.send_last_completion_stream_text_chunk(" Introduction");
4006 fake_model.end_last_completion_stream();
4007 cx.run_until_parked();
4008
4009 // Summary should be set
4010 thread.read_with(cx, |thread, _| {
4011 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4012 assert_eq!(thread.summary().or_default(), "Brief Introduction");
4013 });
4014
4015 // Now we should be able to set a summary
4016 thread.update(cx, |thread, cx| {
4017 thread.set_summary("Brief Intro", cx);
4018 });
4019
4020 thread.read_with(cx, |thread, _| {
4021 assert_eq!(thread.summary().or_default(), "Brief Intro");
4022 });
4023
4024 // Test setting an empty summary (should default to DEFAULT)
4025 thread.update(cx, |thread, cx| {
4026 thread.set_summary("", cx);
4027 });
4028
4029 thread.read_with(cx, |thread, _| {
4030 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4031 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4032 });
4033 }
4034
4035 #[gpui::test]
4036 async fn test_thread_summary_error_set_manually(cx: &mut TestAppContext) {
4037 init_test_settings(cx);
4038
4039 let project = create_test_project(cx, json!({})).await;
4040
4041 let (_, _thread_store, thread, _context_store, model) =
4042 setup_test_environment(cx, project.clone()).await;
4043
4044 test_summarize_error(&model, &thread, cx);
4045
4046 // Now we should be able to set a summary
4047 thread.update(cx, |thread, cx| {
4048 thread.set_summary("Brief Intro", cx);
4049 });
4050
4051 thread.read_with(cx, |thread, _| {
4052 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4053 assert_eq!(thread.summary().or_default(), "Brief Intro");
4054 });
4055 }
4056
4057 #[gpui::test]
4058 async fn test_thread_summary_error_retry(cx: &mut TestAppContext) {
4059 init_test_settings(cx);
4060
4061 let project = create_test_project(cx, json!({})).await;
4062
4063 let (_, _thread_store, thread, _context_store, model) =
4064 setup_test_environment(cx, project.clone()).await;
4065
4066 test_summarize_error(&model, &thread, cx);
4067
4068 // Sending another message should not trigger another summarize request
4069 thread.update(cx, |thread, cx| {
4070 thread.insert_user_message(
4071 "How are you?",
4072 ContextLoadResult::default(),
4073 None,
4074 vec![],
4075 cx,
4076 );
4077 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4078 });
4079
4080 let fake_model = model.as_fake();
4081 simulate_successful_response(fake_model, cx);
4082
4083 thread.read_with(cx, |thread, _| {
4084 // State is still Error, not Generating
4085 assert!(matches!(thread.summary(), ThreadSummary::Error));
4086 });
4087
4088 // But the summarize request can be invoked manually
4089 thread.update(cx, |thread, cx| {
4090 thread.summarize(cx);
4091 });
4092
4093 thread.read_with(cx, |thread, _| {
4094 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4095 });
4096
4097 cx.run_until_parked();
4098 fake_model.send_last_completion_stream_text_chunk("A successful summary");
4099 fake_model.end_last_completion_stream();
4100 cx.run_until_parked();
4101
4102 thread.read_with(cx, |thread, _| {
4103 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4104 assert_eq!(thread.summary().or_default(), "A successful summary");
4105 });
4106 }
4107
4108 // Helper to create a model that returns errors
4109 enum TestError {
4110 Overloaded,
4111 InternalServerError,
4112 }
4113
4114 struct ErrorInjector {
4115 inner: Arc<FakeLanguageModel>,
4116 error_type: TestError,
4117 }
4118
4119 impl ErrorInjector {
4120 fn new(error_type: TestError) -> Self {
4121 Self {
4122 inner: Arc::new(FakeLanguageModel::default()),
4123 error_type,
4124 }
4125 }
4126 }
4127
4128 impl LanguageModel for ErrorInjector {
4129 fn id(&self) -> LanguageModelId {
4130 self.inner.id()
4131 }
4132
4133 fn name(&self) -> LanguageModelName {
4134 self.inner.name()
4135 }
4136
4137 fn provider_id(&self) -> LanguageModelProviderId {
4138 self.inner.provider_id()
4139 }
4140
4141 fn provider_name(&self) -> LanguageModelProviderName {
4142 self.inner.provider_name()
4143 }
4144
4145 fn supports_tools(&self) -> bool {
4146 self.inner.supports_tools()
4147 }
4148
4149 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4150 self.inner.supports_tool_choice(choice)
4151 }
4152
4153 fn supports_images(&self) -> bool {
4154 self.inner.supports_images()
4155 }
4156
4157 fn telemetry_id(&self) -> String {
4158 self.inner.telemetry_id()
4159 }
4160
4161 fn max_token_count(&self) -> u64 {
4162 self.inner.max_token_count()
4163 }
4164
4165 fn count_tokens(
4166 &self,
4167 request: LanguageModelRequest,
4168 cx: &App,
4169 ) -> BoxFuture<'static, Result<u64>> {
4170 self.inner.count_tokens(request, cx)
4171 }
4172
4173 fn stream_completion(
4174 &self,
4175 _request: LanguageModelRequest,
4176 _cx: &AsyncApp,
4177 ) -> BoxFuture<
4178 'static,
4179 Result<
4180 BoxStream<
4181 'static,
4182 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4183 >,
4184 LanguageModelCompletionError,
4185 >,
4186 > {
4187 let error = match self.error_type {
4188 TestError::Overloaded => LanguageModelCompletionError::ServerOverloaded {
4189 provider: self.provider_name(),
4190 retry_after: None,
4191 },
4192 TestError::InternalServerError => {
4193 LanguageModelCompletionError::ApiInternalServerError {
4194 provider: self.provider_name(),
4195 message: "I'm a teapot orbiting the sun".to_string(),
4196 }
4197 }
4198 };
4199 async move {
4200 let stream = futures::stream::once(async move { Err(error) });
4201 Ok(stream.boxed())
4202 }
4203 .boxed()
4204 }
4205
4206 fn as_fake(&self) -> &FakeLanguageModel {
4207 &self.inner
4208 }
4209 }
4210
4211 #[gpui::test]
4212 async fn test_retry_on_overloaded_error(cx: &mut TestAppContext) {
4213 init_test_settings(cx);
4214
4215 let project = create_test_project(cx, json!({})).await;
4216 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4217
4218 // Enable Burn Mode to allow retries
4219 thread.update(cx, |thread, _| {
4220 thread.set_completion_mode(CompletionMode::Burn);
4221 });
4222
4223 // Create model that returns overloaded error
4224 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4225
4226 // Insert a user message
4227 thread.update(cx, |thread, cx| {
4228 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4229 });
4230
4231 // Start completion
4232 thread.update(cx, |thread, cx| {
4233 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4234 });
4235
4236 cx.run_until_parked();
4237
4238 thread.read_with(cx, |thread, _| {
4239 assert!(thread.retry_state.is_some(), "Should have retry state");
4240 let retry_state = thread.retry_state.as_ref().unwrap();
4241 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4242 assert_eq!(
4243 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
4244 "Should retry MAX_RETRY_ATTEMPTS times for overloaded errors"
4245 );
4246 });
4247
4248 // Check that a retry message was added
4249 thread.read_with(cx, |thread, _| {
4250 let mut messages = thread.messages();
4251 assert!(
4252 messages.any(|msg| {
4253 msg.role == Role::System
4254 && msg.ui_only
4255 && msg.segments.iter().any(|seg| {
4256 if let MessageSegment::Text(text) = seg {
4257 text.contains("overloaded")
4258 && text
4259 .contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS))
4260 } else {
4261 false
4262 }
4263 })
4264 }),
4265 "Should have added a system retry message"
4266 );
4267 });
4268
4269 let retry_count = thread.update(cx, |thread, _| {
4270 thread
4271 .messages
4272 .iter()
4273 .filter(|m| {
4274 m.ui_only
4275 && m.segments.iter().any(|s| {
4276 if let MessageSegment::Text(text) = s {
4277 text.contains("Retrying") && text.contains("seconds")
4278 } else {
4279 false
4280 }
4281 })
4282 })
4283 .count()
4284 });
4285
4286 assert_eq!(retry_count, 1, "Should have one retry message");
4287 }
4288
4289 #[gpui::test]
4290 async fn test_retry_on_internal_server_error(cx: &mut TestAppContext) {
4291 init_test_settings(cx);
4292
4293 let project = create_test_project(cx, json!({})).await;
4294 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4295
4296 // Enable Burn Mode to allow retries
4297 thread.update(cx, |thread, _| {
4298 thread.set_completion_mode(CompletionMode::Burn);
4299 });
4300
4301 // Create model that returns internal server error
4302 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4303
4304 // Insert a user message
4305 thread.update(cx, |thread, cx| {
4306 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4307 });
4308
4309 // Start completion
4310 thread.update(cx, |thread, cx| {
4311 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4312 });
4313
4314 cx.run_until_parked();
4315
4316 // Check retry state on thread
4317 thread.read_with(cx, |thread, _| {
4318 assert!(thread.retry_state.is_some(), "Should have retry state");
4319 let retry_state = thread.retry_state.as_ref().unwrap();
4320 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4321 assert_eq!(
4322 retry_state.max_attempts, 3,
4323 "Should have correct max attempts"
4324 );
4325 });
4326
4327 // Check that a retry message was added with provider name
4328 thread.read_with(cx, |thread, _| {
4329 let mut messages = thread.messages();
4330 assert!(
4331 messages.any(|msg| {
4332 msg.role == Role::System
4333 && msg.ui_only
4334 && msg.segments.iter().any(|seg| {
4335 if let MessageSegment::Text(text) = seg {
4336 text.contains("internal")
4337 && text.contains("Fake")
4338 && text.contains("Retrying")
4339 && text.contains("attempt 1 of 3")
4340 && text.contains("seconds")
4341 } else {
4342 false
4343 }
4344 })
4345 }),
4346 "Should have added a system retry message with provider name"
4347 );
4348 });
4349
4350 // Count retry messages
4351 let retry_count = thread.update(cx, |thread, _| {
4352 thread
4353 .messages
4354 .iter()
4355 .filter(|m| {
4356 m.ui_only
4357 && m.segments.iter().any(|s| {
4358 if let MessageSegment::Text(text) = s {
4359 text.contains("Retrying") && text.contains("seconds")
4360 } else {
4361 false
4362 }
4363 })
4364 })
4365 .count()
4366 });
4367
4368 assert_eq!(retry_count, 1, "Should have one retry message");
4369 }
4370
4371 #[gpui::test]
4372 async fn test_exponential_backoff_on_retries(cx: &mut TestAppContext) {
4373 init_test_settings(cx);
4374
4375 let project = create_test_project(cx, json!({})).await;
4376 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4377
4378 // Enable Burn Mode to allow retries
4379 thread.update(cx, |thread, _| {
4380 thread.set_completion_mode(CompletionMode::Burn);
4381 });
4382
4383 // Create model that returns internal server error
4384 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4385
4386 // Insert a user message
4387 thread.update(cx, |thread, cx| {
4388 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4389 });
4390
4391 // Track retry events and completion count
4392 // Track completion events
4393 let completion_count = Arc::new(Mutex::new(0));
4394 let completion_count_clone = completion_count.clone();
4395
4396 let _subscription = thread.update(cx, |_, cx| {
4397 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4398 if let ThreadEvent::NewRequest = event {
4399 *completion_count_clone.lock() += 1;
4400 }
4401 })
4402 });
4403
4404 // First attempt
4405 thread.update(cx, |thread, cx| {
4406 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4407 });
4408 cx.run_until_parked();
4409
4410 // Should have scheduled first retry - count retry messages
4411 let retry_count = thread.update(cx, |thread, _| {
4412 thread
4413 .messages
4414 .iter()
4415 .filter(|m| {
4416 m.ui_only
4417 && m.segments.iter().any(|s| {
4418 if let MessageSegment::Text(text) = s {
4419 text.contains("Retrying") && text.contains("seconds")
4420 } else {
4421 false
4422 }
4423 })
4424 })
4425 .count()
4426 });
4427 assert_eq!(retry_count, 1, "Should have scheduled first retry");
4428
4429 // Check retry state
4430 thread.read_with(cx, |thread, _| {
4431 assert!(thread.retry_state.is_some(), "Should have retry state");
4432 let retry_state = thread.retry_state.as_ref().unwrap();
4433 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4434 assert_eq!(
4435 retry_state.max_attempts, 3,
4436 "Internal server errors should retry up to 3 times"
4437 );
4438 });
4439
4440 // Advance clock for first retry
4441 cx.executor().advance_clock(BASE_RETRY_DELAY);
4442 cx.run_until_parked();
4443
4444 // Advance clock for second retry
4445 cx.executor().advance_clock(BASE_RETRY_DELAY);
4446 cx.run_until_parked();
4447
4448 // Advance clock for third retry
4449 cx.executor().advance_clock(BASE_RETRY_DELAY);
4450 cx.run_until_parked();
4451
4452 // Should have completed all retries - count retry messages
4453 let retry_count = thread.update(cx, |thread, _| {
4454 thread
4455 .messages
4456 .iter()
4457 .filter(|m| {
4458 m.ui_only
4459 && m.segments.iter().any(|s| {
4460 if let MessageSegment::Text(text) = s {
4461 text.contains("Retrying") && text.contains("seconds")
4462 } else {
4463 false
4464 }
4465 })
4466 })
4467 .count()
4468 });
4469 assert_eq!(
4470 retry_count, 3,
4471 "Should have 3 retries for internal server errors"
4472 );
4473
4474 // For internal server errors, we retry 3 times and then give up
4475 // Check that retry_state is cleared after all retries
4476 thread.read_with(cx, |thread, _| {
4477 assert!(
4478 thread.retry_state.is_none(),
4479 "Retry state should be cleared after all retries"
4480 );
4481 });
4482
4483 // Verify total attempts (1 initial + 3 retries)
4484 assert_eq!(
4485 *completion_count.lock(),
4486 4,
4487 "Should have attempted once plus 3 retries"
4488 );
4489 }
4490
4491 #[gpui::test]
4492 async fn test_max_retries_exceeded(cx: &mut TestAppContext) {
4493 init_test_settings(cx);
4494
4495 let project = create_test_project(cx, json!({})).await;
4496 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4497
4498 // Enable Burn Mode to allow retries
4499 thread.update(cx, |thread, _| {
4500 thread.set_completion_mode(CompletionMode::Burn);
4501 });
4502
4503 // Create model that returns overloaded error
4504 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4505
4506 // Insert a user message
4507 thread.update(cx, |thread, cx| {
4508 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4509 });
4510
4511 // Track events
4512 let stopped_with_error = Arc::new(Mutex::new(false));
4513 let stopped_with_error_clone = stopped_with_error.clone();
4514
4515 let _subscription = thread.update(cx, |_, cx| {
4516 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4517 if let ThreadEvent::Stopped(Err(_)) = event {
4518 *stopped_with_error_clone.lock() = true;
4519 }
4520 })
4521 });
4522
4523 // Start initial completion
4524 thread.update(cx, |thread, cx| {
4525 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4526 });
4527 cx.run_until_parked();
4528
4529 // Advance through all retries
4530 for _ in 0..MAX_RETRY_ATTEMPTS {
4531 cx.executor().advance_clock(BASE_RETRY_DELAY);
4532 cx.run_until_parked();
4533 }
4534
4535 let retry_count = thread.update(cx, |thread, _| {
4536 thread
4537 .messages
4538 .iter()
4539 .filter(|m| {
4540 m.ui_only
4541 && m.segments.iter().any(|s| {
4542 if let MessageSegment::Text(text) = s {
4543 text.contains("Retrying") && text.contains("seconds")
4544 } else {
4545 false
4546 }
4547 })
4548 })
4549 .count()
4550 });
4551
4552 // After max retries, should emit Stopped(Err(...)) event
4553 assert_eq!(
4554 retry_count, MAX_RETRY_ATTEMPTS as usize,
4555 "Should have attempted MAX_RETRY_ATTEMPTS retries for overloaded errors"
4556 );
4557 assert!(
4558 *stopped_with_error.lock(),
4559 "Should emit Stopped(Err(...)) event after max retries exceeded"
4560 );
4561
4562 // Retry state should be cleared
4563 thread.read_with(cx, |thread, _| {
4564 assert!(
4565 thread.retry_state.is_none(),
4566 "Retry state should be cleared after max retries"
4567 );
4568
4569 // Verify we have the expected number of retry messages
4570 let retry_messages = thread
4571 .messages
4572 .iter()
4573 .filter(|msg| msg.ui_only && msg.role == Role::System)
4574 .count();
4575 assert_eq!(
4576 retry_messages, MAX_RETRY_ATTEMPTS as usize,
4577 "Should have MAX_RETRY_ATTEMPTS retry messages for overloaded errors"
4578 );
4579 });
4580 }
4581
4582 #[gpui::test]
4583 async fn test_retry_message_removed_on_retry(cx: &mut TestAppContext) {
4584 init_test_settings(cx);
4585
4586 let project = create_test_project(cx, json!({})).await;
4587 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4588
4589 // Enable Burn Mode to allow retries
4590 thread.update(cx, |thread, _| {
4591 thread.set_completion_mode(CompletionMode::Burn);
4592 });
4593
4594 // We'll use a wrapper to switch behavior after first failure
4595 struct RetryTestModel {
4596 inner: Arc<FakeLanguageModel>,
4597 failed_once: Arc<Mutex<bool>>,
4598 }
4599
4600 impl LanguageModel for RetryTestModel {
4601 fn id(&self) -> LanguageModelId {
4602 self.inner.id()
4603 }
4604
4605 fn name(&self) -> LanguageModelName {
4606 self.inner.name()
4607 }
4608
4609 fn provider_id(&self) -> LanguageModelProviderId {
4610 self.inner.provider_id()
4611 }
4612
4613 fn provider_name(&self) -> LanguageModelProviderName {
4614 self.inner.provider_name()
4615 }
4616
4617 fn supports_tools(&self) -> bool {
4618 self.inner.supports_tools()
4619 }
4620
4621 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4622 self.inner.supports_tool_choice(choice)
4623 }
4624
4625 fn supports_images(&self) -> bool {
4626 self.inner.supports_images()
4627 }
4628
4629 fn telemetry_id(&self) -> String {
4630 self.inner.telemetry_id()
4631 }
4632
4633 fn max_token_count(&self) -> u64 {
4634 self.inner.max_token_count()
4635 }
4636
4637 fn count_tokens(
4638 &self,
4639 request: LanguageModelRequest,
4640 cx: &App,
4641 ) -> BoxFuture<'static, Result<u64>> {
4642 self.inner.count_tokens(request, cx)
4643 }
4644
4645 fn stream_completion(
4646 &self,
4647 request: LanguageModelRequest,
4648 cx: &AsyncApp,
4649 ) -> BoxFuture<
4650 'static,
4651 Result<
4652 BoxStream<
4653 'static,
4654 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4655 >,
4656 LanguageModelCompletionError,
4657 >,
4658 > {
4659 if !*self.failed_once.lock() {
4660 *self.failed_once.lock() = true;
4661 let provider = self.provider_name();
4662 // Return error on first attempt
4663 let stream = futures::stream::once(async move {
4664 Err(LanguageModelCompletionError::ServerOverloaded {
4665 provider,
4666 retry_after: None,
4667 })
4668 });
4669 async move { Ok(stream.boxed()) }.boxed()
4670 } else {
4671 // Succeed on retry
4672 self.inner.stream_completion(request, cx)
4673 }
4674 }
4675
4676 fn as_fake(&self) -> &FakeLanguageModel {
4677 &self.inner
4678 }
4679 }
4680
4681 let model = Arc::new(RetryTestModel {
4682 inner: Arc::new(FakeLanguageModel::default()),
4683 failed_once: Arc::new(Mutex::new(false)),
4684 });
4685
4686 // Insert a user message
4687 thread.update(cx, |thread, cx| {
4688 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4689 });
4690
4691 // Track message deletions
4692 // Track when retry completes successfully
4693 let retry_completed = Arc::new(Mutex::new(false));
4694 let retry_completed_clone = retry_completed.clone();
4695
4696 let _subscription = thread.update(cx, |_, cx| {
4697 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4698 if let ThreadEvent::StreamedCompletion = event {
4699 *retry_completed_clone.lock() = true;
4700 }
4701 })
4702 });
4703
4704 // Start completion
4705 thread.update(cx, |thread, cx| {
4706 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4707 });
4708 cx.run_until_parked();
4709
4710 // Get the retry message ID
4711 let retry_message_id = thread.read_with(cx, |thread, _| {
4712 thread
4713 .messages()
4714 .find(|msg| msg.role == Role::System && msg.ui_only)
4715 .map(|msg| msg.id)
4716 .expect("Should have a retry message")
4717 });
4718
4719 // Wait for retry
4720 cx.executor().advance_clock(BASE_RETRY_DELAY);
4721 cx.run_until_parked();
4722
4723 // Stream some successful content
4724 let fake_model = model.as_fake();
4725 // After the retry, there should be a new pending completion
4726 let pending = fake_model.pending_completions();
4727 assert!(
4728 !pending.is_empty(),
4729 "Should have a pending completion after retry"
4730 );
4731 fake_model.send_completion_stream_text_chunk(&pending[0], "Success!");
4732 fake_model.end_completion_stream(&pending[0]);
4733 cx.run_until_parked();
4734
4735 // Check that the retry completed successfully
4736 assert!(
4737 *retry_completed.lock(),
4738 "Retry should have completed successfully"
4739 );
4740
4741 // Retry message should still exist but be marked as ui_only
4742 thread.read_with(cx, |thread, _| {
4743 let retry_msg = thread
4744 .message(retry_message_id)
4745 .expect("Retry message should still exist");
4746 assert!(retry_msg.ui_only, "Retry message should be ui_only");
4747 assert_eq!(
4748 retry_msg.role,
4749 Role::System,
4750 "Retry message should have System role"
4751 );
4752 });
4753 }
4754
4755 #[gpui::test]
4756 async fn test_successful_completion_clears_retry_state(cx: &mut TestAppContext) {
4757 init_test_settings(cx);
4758
4759 let project = create_test_project(cx, json!({})).await;
4760 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4761
4762 // Enable Burn Mode to allow retries
4763 thread.update(cx, |thread, _| {
4764 thread.set_completion_mode(CompletionMode::Burn);
4765 });
4766
4767 // Create a model that fails once then succeeds
4768 struct FailOnceModel {
4769 inner: Arc<FakeLanguageModel>,
4770 failed_once: Arc<Mutex<bool>>,
4771 }
4772
4773 impl LanguageModel for FailOnceModel {
4774 fn id(&self) -> LanguageModelId {
4775 self.inner.id()
4776 }
4777
4778 fn name(&self) -> LanguageModelName {
4779 self.inner.name()
4780 }
4781
4782 fn provider_id(&self) -> LanguageModelProviderId {
4783 self.inner.provider_id()
4784 }
4785
4786 fn provider_name(&self) -> LanguageModelProviderName {
4787 self.inner.provider_name()
4788 }
4789
4790 fn supports_tools(&self) -> bool {
4791 self.inner.supports_tools()
4792 }
4793
4794 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4795 self.inner.supports_tool_choice(choice)
4796 }
4797
4798 fn supports_images(&self) -> bool {
4799 self.inner.supports_images()
4800 }
4801
4802 fn telemetry_id(&self) -> String {
4803 self.inner.telemetry_id()
4804 }
4805
4806 fn max_token_count(&self) -> u64 {
4807 self.inner.max_token_count()
4808 }
4809
4810 fn count_tokens(
4811 &self,
4812 request: LanguageModelRequest,
4813 cx: &App,
4814 ) -> BoxFuture<'static, Result<u64>> {
4815 self.inner.count_tokens(request, cx)
4816 }
4817
4818 fn stream_completion(
4819 &self,
4820 request: LanguageModelRequest,
4821 cx: &AsyncApp,
4822 ) -> BoxFuture<
4823 'static,
4824 Result<
4825 BoxStream<
4826 'static,
4827 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4828 >,
4829 LanguageModelCompletionError,
4830 >,
4831 > {
4832 if !*self.failed_once.lock() {
4833 *self.failed_once.lock() = true;
4834 let provider = self.provider_name();
4835 // Return error on first attempt
4836 let stream = futures::stream::once(async move {
4837 Err(LanguageModelCompletionError::ServerOverloaded {
4838 provider,
4839 retry_after: None,
4840 })
4841 });
4842 async move { Ok(stream.boxed()) }.boxed()
4843 } else {
4844 // Succeed on retry
4845 self.inner.stream_completion(request, cx)
4846 }
4847 }
4848 }
4849
4850 let fail_once_model = Arc::new(FailOnceModel {
4851 inner: Arc::new(FakeLanguageModel::default()),
4852 failed_once: Arc::new(Mutex::new(false)),
4853 });
4854
4855 // Insert a user message
4856 thread.update(cx, |thread, cx| {
4857 thread.insert_user_message(
4858 "Test message",
4859 ContextLoadResult::default(),
4860 None,
4861 vec![],
4862 cx,
4863 );
4864 });
4865
4866 // Start completion with fail-once model
4867 thread.update(cx, |thread, cx| {
4868 thread.send_to_model(
4869 fail_once_model.clone(),
4870 CompletionIntent::UserPrompt,
4871 None,
4872 cx,
4873 );
4874 });
4875
4876 cx.run_until_parked();
4877
4878 // Verify retry state exists after first failure
4879 thread.read_with(cx, |thread, _| {
4880 assert!(
4881 thread.retry_state.is_some(),
4882 "Should have retry state after failure"
4883 );
4884 });
4885
4886 // Wait for retry delay
4887 cx.executor().advance_clock(BASE_RETRY_DELAY);
4888 cx.run_until_parked();
4889
4890 // The retry should now use our FailOnceModel which should succeed
4891 // We need to help the FakeLanguageModel complete the stream
4892 let inner_fake = fail_once_model.inner.clone();
4893
4894 // Wait a bit for the retry to start
4895 cx.run_until_parked();
4896
4897 // Check for pending completions and complete them
4898 if let Some(pending) = inner_fake.pending_completions().first() {
4899 inner_fake.send_completion_stream_text_chunk(pending, "Success!");
4900 inner_fake.end_completion_stream(pending);
4901 }
4902 cx.run_until_parked();
4903
4904 thread.read_with(cx, |thread, _| {
4905 assert!(
4906 thread.retry_state.is_none(),
4907 "Retry state should be cleared after successful completion"
4908 );
4909
4910 let has_assistant_message = thread
4911 .messages
4912 .iter()
4913 .any(|msg| msg.role == Role::Assistant && !msg.ui_only);
4914 assert!(
4915 has_assistant_message,
4916 "Should have an assistant message after successful retry"
4917 );
4918 });
4919 }
4920
4921 #[gpui::test]
4922 async fn test_rate_limit_retry_single_attempt(cx: &mut TestAppContext) {
4923 init_test_settings(cx);
4924
4925 let project = create_test_project(cx, json!({})).await;
4926 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4927
4928 // Enable Burn Mode to allow retries
4929 thread.update(cx, |thread, _| {
4930 thread.set_completion_mode(CompletionMode::Burn);
4931 });
4932
4933 // Create a model that returns rate limit error with retry_after
4934 struct RateLimitModel {
4935 inner: Arc<FakeLanguageModel>,
4936 }
4937
4938 impl LanguageModel for RateLimitModel {
4939 fn id(&self) -> LanguageModelId {
4940 self.inner.id()
4941 }
4942
4943 fn name(&self) -> LanguageModelName {
4944 self.inner.name()
4945 }
4946
4947 fn provider_id(&self) -> LanguageModelProviderId {
4948 self.inner.provider_id()
4949 }
4950
4951 fn provider_name(&self) -> LanguageModelProviderName {
4952 self.inner.provider_name()
4953 }
4954
4955 fn supports_tools(&self) -> bool {
4956 self.inner.supports_tools()
4957 }
4958
4959 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4960 self.inner.supports_tool_choice(choice)
4961 }
4962
4963 fn supports_images(&self) -> bool {
4964 self.inner.supports_images()
4965 }
4966
4967 fn telemetry_id(&self) -> String {
4968 self.inner.telemetry_id()
4969 }
4970
4971 fn max_token_count(&self) -> u64 {
4972 self.inner.max_token_count()
4973 }
4974
4975 fn count_tokens(
4976 &self,
4977 request: LanguageModelRequest,
4978 cx: &App,
4979 ) -> BoxFuture<'static, Result<u64>> {
4980 self.inner.count_tokens(request, cx)
4981 }
4982
4983 fn stream_completion(
4984 &self,
4985 _request: LanguageModelRequest,
4986 _cx: &AsyncApp,
4987 ) -> BoxFuture<
4988 'static,
4989 Result<
4990 BoxStream<
4991 'static,
4992 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4993 >,
4994 LanguageModelCompletionError,
4995 >,
4996 > {
4997 let provider = self.provider_name();
4998 async move {
4999 let stream = futures::stream::once(async move {
5000 Err(LanguageModelCompletionError::RateLimitExceeded {
5001 provider,
5002 retry_after: Some(Duration::from_secs(TEST_RATE_LIMIT_RETRY_SECS)),
5003 })
5004 });
5005 Ok(stream.boxed())
5006 }
5007 .boxed()
5008 }
5009
5010 fn as_fake(&self) -> &FakeLanguageModel {
5011 &self.inner
5012 }
5013 }
5014
5015 let model = Arc::new(RateLimitModel {
5016 inner: Arc::new(FakeLanguageModel::default()),
5017 });
5018
5019 // Insert a user message
5020 thread.update(cx, |thread, cx| {
5021 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5022 });
5023
5024 // Start completion
5025 thread.update(cx, |thread, cx| {
5026 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5027 });
5028
5029 cx.run_until_parked();
5030
5031 let retry_count = thread.update(cx, |thread, _| {
5032 thread
5033 .messages
5034 .iter()
5035 .filter(|m| {
5036 m.ui_only
5037 && m.segments.iter().any(|s| {
5038 if let MessageSegment::Text(text) = s {
5039 text.contains("rate limit exceeded")
5040 } else {
5041 false
5042 }
5043 })
5044 })
5045 .count()
5046 });
5047 assert_eq!(retry_count, 1, "Should have scheduled one retry");
5048
5049 thread.read_with(cx, |thread, _| {
5050 assert!(
5051 thread.retry_state.is_some(),
5052 "Rate limit errors should set retry_state"
5053 );
5054 if let Some(retry_state) = &thread.retry_state {
5055 assert_eq!(
5056 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
5057 "Rate limit errors should use MAX_RETRY_ATTEMPTS"
5058 );
5059 }
5060 });
5061
5062 // Verify we have one retry message
5063 thread.read_with(cx, |thread, _| {
5064 let retry_messages = thread
5065 .messages
5066 .iter()
5067 .filter(|msg| {
5068 msg.ui_only
5069 && msg.segments.iter().any(|seg| {
5070 if let MessageSegment::Text(text) = seg {
5071 text.contains("rate limit exceeded")
5072 } else {
5073 false
5074 }
5075 })
5076 })
5077 .count();
5078 assert_eq!(
5079 retry_messages, 1,
5080 "Should have one rate limit retry message"
5081 );
5082 });
5083
5084 // Check that retry message doesn't include attempt count
5085 thread.read_with(cx, |thread, _| {
5086 let retry_message = thread
5087 .messages
5088 .iter()
5089 .find(|msg| msg.role == Role::System && msg.ui_only)
5090 .expect("Should have a retry message");
5091
5092 // Check that the message contains attempt count since we use retry_state
5093 if let Some(MessageSegment::Text(text)) = retry_message.segments.first() {
5094 assert!(
5095 text.contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS)),
5096 "Rate limit retry message should contain attempt count with MAX_RETRY_ATTEMPTS"
5097 );
5098 assert!(
5099 text.contains("Retrying"),
5100 "Rate limit retry message should contain retry text"
5101 );
5102 }
5103 });
5104 }
5105
5106 #[gpui::test]
5107 async fn test_ui_only_messages_not_sent_to_model(cx: &mut TestAppContext) {
5108 init_test_settings(cx);
5109
5110 let project = create_test_project(cx, json!({})).await;
5111 let (_, _, thread, _, model) = setup_test_environment(cx, project.clone()).await;
5112
5113 // Insert a regular user message
5114 thread.update(cx, |thread, cx| {
5115 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5116 });
5117
5118 // Insert a UI-only message (like our retry notifications)
5119 thread.update(cx, |thread, cx| {
5120 let id = thread.next_message_id.post_inc();
5121 thread.messages.push(Message {
5122 id,
5123 role: Role::System,
5124 segments: vec![MessageSegment::Text(
5125 "This is a UI-only message that should not be sent to the model".to_string(),
5126 )],
5127 loaded_context: LoadedContext::default(),
5128 creases: Vec::new(),
5129 is_hidden: true,
5130 ui_only: true,
5131 });
5132 cx.emit(ThreadEvent::MessageAdded(id));
5133 });
5134
5135 // Insert another regular message
5136 thread.update(cx, |thread, cx| {
5137 thread.insert_user_message(
5138 "How are you?",
5139 ContextLoadResult::default(),
5140 None,
5141 vec![],
5142 cx,
5143 );
5144 });
5145
5146 // Generate the completion request
5147 let request = thread.update(cx, |thread, cx| {
5148 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
5149 });
5150
5151 // Verify that the request only contains non-UI-only messages
5152 // Should have system prompt + 2 user messages, but not the UI-only message
5153 let user_messages: Vec<_> = request
5154 .messages
5155 .iter()
5156 .filter(|msg| msg.role == Role::User)
5157 .collect();
5158 assert_eq!(
5159 user_messages.len(),
5160 2,
5161 "Should have exactly 2 user messages"
5162 );
5163
5164 // Verify the UI-only content is not present anywhere in the request
5165 let request_text = request
5166 .messages
5167 .iter()
5168 .flat_map(|msg| &msg.content)
5169 .filter_map(|content| match content {
5170 MessageContent::Text(text) => Some(text.as_str()),
5171 _ => None,
5172 })
5173 .collect::<String>();
5174
5175 assert!(
5176 !request_text.contains("UI-only message"),
5177 "UI-only message content should not be in the request"
5178 );
5179
5180 // Verify the thread still has all 3 messages (including UI-only)
5181 thread.read_with(cx, |thread, _| {
5182 assert_eq!(
5183 thread.messages().count(),
5184 3,
5185 "Thread should have 3 messages"
5186 );
5187 assert_eq!(
5188 thread.messages().filter(|m| m.ui_only).count(),
5189 1,
5190 "Thread should have 1 UI-only message"
5191 );
5192 });
5193
5194 // Verify that UI-only messages are not serialized
5195 let serialized = thread
5196 .update(cx, |thread, cx| thread.serialize(cx))
5197 .await
5198 .unwrap();
5199 assert_eq!(
5200 serialized.messages.len(),
5201 2,
5202 "Serialized thread should only have 2 messages (no UI-only)"
5203 );
5204 }
5205
5206 #[gpui::test]
5207 async fn test_no_retry_without_burn_mode(cx: &mut TestAppContext) {
5208 init_test_settings(cx);
5209
5210 let project = create_test_project(cx, json!({})).await;
5211 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5212
5213 // Ensure we're in Normal mode (not Burn mode)
5214 thread.update(cx, |thread, _| {
5215 thread.set_completion_mode(CompletionMode::Normal);
5216 });
5217
5218 // Track error events
5219 let error_events = Arc::new(Mutex::new(Vec::new()));
5220 let error_events_clone = error_events.clone();
5221
5222 let _subscription = thread.update(cx, |_, cx| {
5223 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
5224 if let ThreadEvent::ShowError(error) = event {
5225 error_events_clone.lock().push(error.clone());
5226 }
5227 })
5228 });
5229
5230 // Create model that returns overloaded error
5231 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5232
5233 // Insert a user message
5234 thread.update(cx, |thread, cx| {
5235 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5236 });
5237
5238 // Start completion
5239 thread.update(cx, |thread, cx| {
5240 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5241 });
5242
5243 cx.run_until_parked();
5244
5245 // Verify no retry state was created
5246 thread.read_with(cx, |thread, _| {
5247 assert!(
5248 thread.retry_state.is_none(),
5249 "Should not have retry state in Normal mode"
5250 );
5251 });
5252
5253 // Check that a retryable error was reported
5254 let errors = error_events.lock();
5255 assert!(!errors.is_empty(), "Should have received an error event");
5256
5257 if let ThreadError::RetryableError {
5258 message: _,
5259 can_enable_burn_mode,
5260 } = &errors[0]
5261 {
5262 assert!(
5263 *can_enable_burn_mode,
5264 "Error should indicate burn mode can be enabled"
5265 );
5266 } else {
5267 panic!("Expected RetryableError, got {:?}", errors[0]);
5268 }
5269
5270 // Verify the thread is no longer generating
5271 thread.read_with(cx, |thread, _| {
5272 assert!(
5273 !thread.is_generating(),
5274 "Should not be generating after error without retry"
5275 );
5276 });
5277 }
5278
5279 #[gpui::test]
5280 async fn test_retry_canceled_on_stop(cx: &mut TestAppContext) {
5281 init_test_settings(cx);
5282
5283 let project = create_test_project(cx, json!({})).await;
5284 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5285
5286 // Enable Burn Mode to allow retries
5287 thread.update(cx, |thread, _| {
5288 thread.set_completion_mode(CompletionMode::Burn);
5289 });
5290
5291 // Create model that returns overloaded error
5292 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5293
5294 // Insert a user message
5295 thread.update(cx, |thread, cx| {
5296 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5297 });
5298
5299 // Start completion
5300 thread.update(cx, |thread, cx| {
5301 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5302 });
5303
5304 cx.run_until_parked();
5305
5306 // Verify retry was scheduled by checking for retry message
5307 let has_retry_message = thread.read_with(cx, |thread, _| {
5308 thread.messages.iter().any(|m| {
5309 m.ui_only
5310 && m.segments.iter().any(|s| {
5311 if let MessageSegment::Text(text) = s {
5312 text.contains("Retrying") && text.contains("seconds")
5313 } else {
5314 false
5315 }
5316 })
5317 })
5318 });
5319 assert!(has_retry_message, "Should have scheduled a retry");
5320
5321 // Cancel the completion before the retry happens
5322 thread.update(cx, |thread, cx| {
5323 thread.cancel_last_completion(None, cx);
5324 });
5325
5326 cx.run_until_parked();
5327
5328 // The retry should not have happened - no pending completions
5329 let fake_model = model.as_fake();
5330 assert_eq!(
5331 fake_model.pending_completions().len(),
5332 0,
5333 "Should have no pending completions after cancellation"
5334 );
5335
5336 // Verify the retry was canceled by checking retry state
5337 thread.read_with(cx, |thread, _| {
5338 if let Some(retry_state) = &thread.retry_state {
5339 panic!(
5340 "retry_state should be cleared after cancellation, but found: attempt={}, max_attempts={}, intent={:?}",
5341 retry_state.attempt, retry_state.max_attempts, retry_state.intent
5342 );
5343 }
5344 });
5345 }
5346
5347 fn test_summarize_error(
5348 model: &Arc<dyn LanguageModel>,
5349 thread: &Entity<Thread>,
5350 cx: &mut TestAppContext,
5351 ) {
5352 thread.update(cx, |thread, cx| {
5353 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
5354 thread.send_to_model(
5355 model.clone(),
5356 CompletionIntent::ThreadSummarization,
5357 None,
5358 cx,
5359 );
5360 });
5361
5362 let fake_model = model.as_fake();
5363 simulate_successful_response(fake_model, cx);
5364
5365 thread.read_with(cx, |thread, _| {
5366 assert!(matches!(thread.summary(), ThreadSummary::Generating));
5367 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5368 });
5369
5370 // Simulate summary request ending
5371 cx.run_until_parked();
5372 fake_model.end_last_completion_stream();
5373 cx.run_until_parked();
5374
5375 // State is set to Error and default message
5376 thread.read_with(cx, |thread, _| {
5377 assert!(matches!(thread.summary(), ThreadSummary::Error));
5378 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5379 });
5380 }
5381
5382 fn simulate_successful_response(fake_model: &FakeLanguageModel, cx: &mut TestAppContext) {
5383 cx.run_until_parked();
5384 fake_model.send_last_completion_stream_text_chunk("Assistant response");
5385 fake_model.end_last_completion_stream();
5386 cx.run_until_parked();
5387 }
5388
5389 fn init_test_settings(cx: &mut TestAppContext) {
5390 cx.update(|cx| {
5391 let settings_store = SettingsStore::test(cx);
5392 cx.set_global(settings_store);
5393 language::init(cx);
5394 Project::init_settings(cx);
5395 AgentSettings::register(cx);
5396 prompt_store::init(cx);
5397 thread_store::init(cx);
5398 workspace::init_settings(cx);
5399 language_model::init_settings(cx);
5400 ThemeSettings::register(cx);
5401 ToolRegistry::default_global(cx);
5402 assistant_tool::init(cx);
5403
5404 let http_client = Arc::new(http_client::HttpClientWithUrl::new(
5405 http_client::FakeHttpClient::with_200_response(),
5406 "http://localhost".to_string(),
5407 None,
5408 ));
5409 assistant_tools::init(http_client, cx);
5410 });
5411 }
5412
5413 // Helper to create a test project with test files
5414 async fn create_test_project(
5415 cx: &mut TestAppContext,
5416 files: serde_json::Value,
5417 ) -> Entity<Project> {
5418 let fs = FakeFs::new(cx.executor());
5419 fs.insert_tree(path!("/test"), files).await;
5420 Project::test(fs, [path!("/test").as_ref()], cx).await
5421 }
5422
5423 async fn setup_test_environment(
5424 cx: &mut TestAppContext,
5425 project: Entity<Project>,
5426 ) -> (
5427 Entity<Workspace>,
5428 Entity<ThreadStore>,
5429 Entity<Thread>,
5430 Entity<ContextStore>,
5431 Arc<dyn LanguageModel>,
5432 ) {
5433 let (workspace, cx) =
5434 cx.add_window_view(|window, cx| Workspace::test_new(project.clone(), window, cx));
5435
5436 let thread_store = cx
5437 .update(|_, cx| {
5438 ThreadStore::load(
5439 project.clone(),
5440 cx.new(|_| ToolWorkingSet::default()),
5441 None,
5442 Arc::new(PromptBuilder::new(None).unwrap()),
5443 cx,
5444 )
5445 })
5446 .await
5447 .unwrap();
5448
5449 let thread = thread_store.update(cx, |store, cx| store.create_thread(cx));
5450 let context_store = cx.new(|_cx| ContextStore::new(project.downgrade(), None));
5451
5452 let provider = Arc::new(FakeLanguageModelProvider::default());
5453 let model = provider.test_model();
5454 let model: Arc<dyn LanguageModel> = Arc::new(model);
5455
5456 cx.update(|_, cx| {
5457 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
5458 registry.set_default_model(
5459 Some(ConfiguredModel {
5460 provider: provider.clone(),
5461 model: model.clone(),
5462 }),
5463 cx,
5464 );
5465 registry.set_thread_summary_model(
5466 Some(ConfiguredModel {
5467 provider,
5468 model: model.clone(),
5469 }),
5470 cx,
5471 );
5472 })
5473 });
5474
5475 (workspace, thread_store, thread, context_store, model)
5476 }
5477
5478 async fn add_file_to_context(
5479 project: &Entity<Project>,
5480 context_store: &Entity<ContextStore>,
5481 path: &str,
5482 cx: &mut TestAppContext,
5483 ) -> Result<Entity<language::Buffer>> {
5484 let buffer_path = project
5485 .read_with(cx, |project, cx| project.find_project_path(path, cx))
5486 .unwrap();
5487
5488 let buffer = project
5489 .update(cx, |project, cx| {
5490 project.open_buffer(buffer_path.clone(), cx)
5491 })
5492 .await
5493 .unwrap();
5494
5495 context_store.update(cx, |context_store, cx| {
5496 context_store.add_file_from_buffer(&buffer_path, buffer.clone(), false, cx);
5497 });
5498
5499 Ok(buffer)
5500 }
5501}