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 agent_settings::{AgentProfileId, AgentSettings, CompletionMode};
12use anyhow::{Result, anyhow};
13use assistant_tool::{ActionLog, AnyToolCard, Tool, ToolWorkingSet};
14use chrono::{DateTime, Utc};
15use client::{ModelRequestUsage, RequestUsage};
16use collections::HashMap;
17use feature_flags::{self, FeatureFlagAppExt};
18use futures::{FutureExt, StreamExt as _, future::Shared};
19use git::repository::DiffType;
20use gpui::{
21 AnyWindowHandle, App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task,
22 WeakEntity, Window,
23};
24use http_client::StatusCode;
25use language_model::{
26 ConfiguredModel, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,
27 LanguageModelExt as _, LanguageModelId, LanguageModelRegistry, LanguageModelRequest,
28 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
29 LanguageModelToolResultContent, LanguageModelToolUse, LanguageModelToolUseId, MessageContent,
30 ModelRequestLimitReachedError, PaymentRequiredError, Role, SelectedModel, StopReason,
31 TokenUsage,
32};
33use postage::stream::Stream as _;
34use project::{
35 Project,
36 git_store::{GitStore, GitStoreCheckpoint, RepositoryState},
37};
38use prompt_store::{ModelContext, PromptBuilder};
39use proto::Plan;
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;
52use zed_llm_client::{CompletionIntent, CompletionRequestStatus, UsageLimit};
53
54const MAX_RETRY_ATTEMPTS: u8 = 4;
55const BASE_RETRY_DELAY: Duration = Duration::from_secs(5);
56
57#[derive(Debug, Clone)]
58enum RetryStrategy {
59 ExponentialBackoff {
60 initial_delay: Duration,
61 max_attempts: u8,
62 },
63 Fixed {
64 delay: Duration,
65 max_attempts: u8,
66 },
67}
68
69#[derive(
70 Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize, JsonSchema,
71)]
72pub struct ThreadId(Arc<str>);
73
74impl ThreadId {
75 pub fn new() -> Self {
76 Self(Uuid::new_v4().to_string().into())
77 }
78}
79
80impl std::fmt::Display for ThreadId {
81 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
82 write!(f, "{}", self.0)
83 }
84}
85
86impl From<&str> for ThreadId {
87 fn from(value: &str) -> Self {
88 Self(value.into())
89 }
90}
91
92/// The ID of the user prompt that initiated a request.
93///
94/// This equates to the user physically submitting a message to the model (e.g., by pressing the Enter key).
95#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
96pub struct PromptId(Arc<str>);
97
98impl PromptId {
99 pub fn new() -> Self {
100 Self(Uuid::new_v4().to_string().into())
101 }
102}
103
104impl std::fmt::Display for PromptId {
105 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
106 write!(f, "{}", self.0)
107 }
108}
109
110#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Copy, Serialize, Deserialize)]
111pub struct MessageId(pub(crate) usize);
112
113impl MessageId {
114 fn post_inc(&mut self) -> Self {
115 Self(post_inc(&mut self.0))
116 }
117
118 pub fn as_usize(&self) -> usize {
119 self.0
120 }
121}
122
123/// Stored information that can be used to resurrect a context crease when creating an editor for a past message.
124#[derive(Clone, Debug)]
125pub struct MessageCrease {
126 pub range: Range<usize>,
127 pub icon_path: SharedString,
128 pub label: SharedString,
129 /// None for a deserialized message, Some otherwise.
130 pub context: Option<AgentContextHandle>,
131}
132
133/// A message in a [`Thread`].
134#[derive(Debug, Clone)]
135pub struct Message {
136 pub id: MessageId,
137 pub role: Role,
138 pub segments: Vec<MessageSegment>,
139 pub loaded_context: LoadedContext,
140 pub creases: Vec<MessageCrease>,
141 pub is_hidden: bool,
142 pub ui_only: bool,
143}
144
145impl Message {
146 /// Returns whether the message contains any meaningful text that should be displayed
147 /// The model sometimes runs tool without producing any text or just a marker ([`USING_TOOL_MARKER`])
148 pub fn should_display_content(&self) -> bool {
149 self.segments.iter().all(|segment| segment.should_display())
150 }
151
152 pub fn push_thinking(&mut self, text: &str, signature: Option<String>) {
153 if let Some(MessageSegment::Thinking {
154 text: segment,
155 signature: current_signature,
156 }) = self.segments.last_mut()
157 {
158 if let Some(signature) = signature {
159 *current_signature = Some(signature);
160 }
161 segment.push_str(text);
162 } else {
163 self.segments.push(MessageSegment::Thinking {
164 text: text.to_string(),
165 signature,
166 });
167 }
168 }
169
170 pub fn push_redacted_thinking(&mut self, data: String) {
171 self.segments.push(MessageSegment::RedactedThinking(data));
172 }
173
174 pub fn push_text(&mut self, text: &str) {
175 if let Some(MessageSegment::Text(segment)) = self.segments.last_mut() {
176 segment.push_str(text);
177 } else {
178 self.segments.push(MessageSegment::Text(text.to_string()));
179 }
180 }
181
182 pub fn to_string(&self) -> String {
183 let mut result = String::new();
184
185 if !self.loaded_context.text.is_empty() {
186 result.push_str(&self.loaded_context.text);
187 }
188
189 for segment in &self.segments {
190 match segment {
191 MessageSegment::Text(text) => result.push_str(text),
192 MessageSegment::Thinking { text, .. } => {
193 result.push_str("<think>\n");
194 result.push_str(text);
195 result.push_str("\n</think>");
196 }
197 MessageSegment::RedactedThinking(_) => {}
198 }
199 }
200
201 result
202 }
203}
204
205#[derive(Debug, Clone, PartialEq, Eq)]
206pub enum MessageSegment {
207 Text(String),
208 Thinking {
209 text: String,
210 signature: Option<String>,
211 },
212 RedactedThinking(String),
213}
214
215impl MessageSegment {
216 pub fn should_display(&self) -> bool {
217 match self {
218 Self::Text(text) => text.is_empty(),
219 Self::Thinking { text, .. } => text.is_empty(),
220 Self::RedactedThinking(_) => false,
221 }
222 }
223
224 pub fn text(&self) -> Option<&str> {
225 match self {
226 MessageSegment::Text(text) => Some(text),
227 _ => None,
228 }
229 }
230}
231
232#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
233pub struct ProjectSnapshot {
234 pub worktree_snapshots: Vec<WorktreeSnapshot>,
235 pub unsaved_buffer_paths: Vec<String>,
236 pub timestamp: DateTime<Utc>,
237}
238
239#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
240pub struct WorktreeSnapshot {
241 pub worktree_path: String,
242 pub git_state: Option<GitState>,
243}
244
245#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
246pub struct GitState {
247 pub remote_url: Option<String>,
248 pub head_sha: Option<String>,
249 pub current_branch: Option<String>,
250 pub diff: Option<String>,
251}
252
253#[derive(Clone, Debug)]
254pub struct ThreadCheckpoint {
255 message_id: MessageId,
256 git_checkpoint: GitStoreCheckpoint,
257}
258
259#[derive(Copy, Clone, Debug, PartialEq, Eq)]
260pub enum ThreadFeedback {
261 Positive,
262 Negative,
263}
264
265pub enum LastRestoreCheckpoint {
266 Pending {
267 message_id: MessageId,
268 },
269 Error {
270 message_id: MessageId,
271 error: String,
272 },
273}
274
275impl LastRestoreCheckpoint {
276 pub fn message_id(&self) -> MessageId {
277 match self {
278 LastRestoreCheckpoint::Pending { message_id } => *message_id,
279 LastRestoreCheckpoint::Error { message_id, .. } => *message_id,
280 }
281 }
282}
283
284#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
285pub enum DetailedSummaryState {
286 #[default]
287 NotGenerated,
288 Generating {
289 message_id: MessageId,
290 },
291 Generated {
292 text: SharedString,
293 message_id: MessageId,
294 },
295}
296
297impl DetailedSummaryState {
298 fn text(&self) -> Option<SharedString> {
299 if let Self::Generated { text, .. } = self {
300 Some(text.clone())
301 } else {
302 None
303 }
304 }
305}
306
307#[derive(Default, Debug)]
308pub struct TotalTokenUsage {
309 pub total: u64,
310 pub max: u64,
311}
312
313impl TotalTokenUsage {
314 pub fn ratio(&self) -> TokenUsageRatio {
315 #[cfg(debug_assertions)]
316 let warning_threshold: f32 = std::env::var("ZED_THREAD_WARNING_THRESHOLD")
317 .unwrap_or("0.8".to_string())
318 .parse()
319 .unwrap();
320 #[cfg(not(debug_assertions))]
321 let warning_threshold: f32 = 0.8;
322
323 // When the maximum is unknown because there is no selected model,
324 // avoid showing the token limit warning.
325 if self.max == 0 {
326 TokenUsageRatio::Normal
327 } else if self.total >= self.max {
328 TokenUsageRatio::Exceeded
329 } else if self.total as f32 / self.max as f32 >= warning_threshold {
330 TokenUsageRatio::Warning
331 } else {
332 TokenUsageRatio::Normal
333 }
334 }
335
336 pub fn add(&self, tokens: u64) -> TotalTokenUsage {
337 TotalTokenUsage {
338 total: self.total + tokens,
339 max: self.max,
340 }
341 }
342}
343
344#[derive(Debug, Default, PartialEq, Eq)]
345pub enum TokenUsageRatio {
346 #[default]
347 Normal,
348 Warning,
349 Exceeded,
350}
351
352#[derive(Debug, Clone, Copy)]
353pub enum QueueState {
354 Sending,
355 Queued { position: usize },
356 Started,
357}
358
359/// A thread of conversation with the LLM.
360pub struct Thread {
361 id: ThreadId,
362 updated_at: DateTime<Utc>,
363 summary: ThreadSummary,
364 pending_summary: Task<Option<()>>,
365 detailed_summary_task: Task<Option<()>>,
366 detailed_summary_tx: postage::watch::Sender<DetailedSummaryState>,
367 detailed_summary_rx: postage::watch::Receiver<DetailedSummaryState>,
368 completion_mode: agent_settings::CompletionMode,
369 messages: Vec<Message>,
370 next_message_id: MessageId,
371 last_prompt_id: PromptId,
372 project_context: SharedProjectContext,
373 checkpoints_by_message: HashMap<MessageId, ThreadCheckpoint>,
374 completion_count: usize,
375 pending_completions: Vec<PendingCompletion>,
376 project: Entity<Project>,
377 prompt_builder: Arc<PromptBuilder>,
378 tools: Entity<ToolWorkingSet>,
379 tool_use: ToolUseState,
380 action_log: Entity<ActionLog>,
381 last_restore_checkpoint: Option<LastRestoreCheckpoint>,
382 pending_checkpoint: Option<ThreadCheckpoint>,
383 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
384 request_token_usage: Vec<TokenUsage>,
385 cumulative_token_usage: TokenUsage,
386 exceeded_window_error: Option<ExceededWindowError>,
387 tool_use_limit_reached: bool,
388 feedback: Option<ThreadFeedback>,
389 retry_state: Option<RetryState>,
390 message_feedback: HashMap<MessageId, ThreadFeedback>,
391 last_auto_capture_at: Option<Instant>,
392 last_received_chunk_at: Option<Instant>,
393 request_callback: Option<
394 Box<dyn FnMut(&LanguageModelRequest, &[Result<LanguageModelCompletionEvent, String>])>,
395 >,
396 remaining_turns: u32,
397 configured_model: Option<ConfiguredModel>,
398 profile: AgentProfile,
399 last_error_context: Option<(Arc<dyn LanguageModel>, CompletionIntent)>,
400}
401
402#[derive(Clone, Debug)]
403struct RetryState {
404 attempt: u8,
405 max_attempts: u8,
406 intent: CompletionIntent,
407}
408
409#[derive(Clone, Debug, PartialEq, Eq)]
410pub enum ThreadSummary {
411 Pending,
412 Generating,
413 Ready(SharedString),
414 Error,
415}
416
417impl ThreadSummary {
418 pub const DEFAULT: SharedString = SharedString::new_static("New Thread");
419
420 pub fn or_default(&self) -> SharedString {
421 self.unwrap_or(Self::DEFAULT)
422 }
423
424 pub fn unwrap_or(&self, message: impl Into<SharedString>) -> SharedString {
425 self.ready().unwrap_or_else(|| message.into())
426 }
427
428 pub fn ready(&self) -> Option<SharedString> {
429 match self {
430 ThreadSummary::Ready(summary) => Some(summary.clone()),
431 ThreadSummary::Pending | ThreadSummary::Generating | ThreadSummary::Error => None,
432 }
433 }
434}
435
436#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
437pub struct ExceededWindowError {
438 /// Model used when last message exceeded context window
439 model_id: LanguageModelId,
440 /// Token count including last message
441 token_count: u64,
442}
443
444impl Thread {
445 pub fn new(
446 project: Entity<Project>,
447 tools: Entity<ToolWorkingSet>,
448 prompt_builder: Arc<PromptBuilder>,
449 system_prompt: SharedProjectContext,
450 cx: &mut Context<Self>,
451 ) -> Self {
452 let (detailed_summary_tx, detailed_summary_rx) = postage::watch::channel();
453 let configured_model = LanguageModelRegistry::read_global(cx).default_model();
454 let profile_id = AgentSettings::get_global(cx).default_profile.clone();
455
456 Self {
457 id: ThreadId::new(),
458 updated_at: Utc::now(),
459 summary: ThreadSummary::Pending,
460 pending_summary: Task::ready(None),
461 detailed_summary_task: Task::ready(None),
462 detailed_summary_tx,
463 detailed_summary_rx,
464 completion_mode: AgentSettings::get_global(cx).preferred_completion_mode,
465 messages: Vec::new(),
466 next_message_id: MessageId(0),
467 last_prompt_id: PromptId::new(),
468 project_context: system_prompt,
469 checkpoints_by_message: HashMap::default(),
470 completion_count: 0,
471 pending_completions: Vec::new(),
472 project: project.clone(),
473 prompt_builder,
474 tools: tools.clone(),
475 last_restore_checkpoint: None,
476 pending_checkpoint: None,
477 tool_use: ToolUseState::new(tools.clone()),
478 action_log: cx.new(|_| ActionLog::new(project.clone())),
479 initial_project_snapshot: {
480 let project_snapshot = Self::project_snapshot(project, cx);
481 cx.foreground_executor()
482 .spawn(async move { Some(project_snapshot.await) })
483 .shared()
484 },
485 request_token_usage: Vec::new(),
486 cumulative_token_usage: TokenUsage::default(),
487 exceeded_window_error: None,
488 tool_use_limit_reached: false,
489 feedback: None,
490 retry_state: None,
491 message_feedback: HashMap::default(),
492 last_auto_capture_at: None,
493 last_error_context: None,
494 last_received_chunk_at: None,
495 request_callback: None,
496 remaining_turns: u32::MAX,
497 configured_model: configured_model.clone(),
498 profile: AgentProfile::new(profile_id, tools),
499 }
500 }
501
502 pub fn deserialize(
503 id: ThreadId,
504 serialized: SerializedThread,
505 project: Entity<Project>,
506 tools: Entity<ToolWorkingSet>,
507 prompt_builder: Arc<PromptBuilder>,
508 project_context: SharedProjectContext,
509 window: Option<&mut Window>, // None in headless mode
510 cx: &mut Context<Self>,
511 ) -> Self {
512 let next_message_id = MessageId(
513 serialized
514 .messages
515 .last()
516 .map(|message| message.id.0 + 1)
517 .unwrap_or(0),
518 );
519 let tool_use = ToolUseState::from_serialized_messages(
520 tools.clone(),
521 &serialized.messages,
522 project.clone(),
523 window,
524 cx,
525 );
526 let (detailed_summary_tx, detailed_summary_rx) =
527 postage::watch::channel_with(serialized.detailed_summary_state);
528
529 let configured_model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
530 serialized
531 .model
532 .and_then(|model| {
533 let model = SelectedModel {
534 provider: model.provider.clone().into(),
535 model: model.model.clone().into(),
536 };
537 registry.select_model(&model, cx)
538 })
539 .or_else(|| registry.default_model())
540 });
541
542 let completion_mode = serialized
543 .completion_mode
544 .unwrap_or_else(|| AgentSettings::get_global(cx).preferred_completion_mode);
545 let profile_id = serialized
546 .profile
547 .unwrap_or_else(|| AgentSettings::get_global(cx).default_profile.clone());
548
549 Self {
550 id,
551 updated_at: serialized.updated_at,
552 summary: ThreadSummary::Ready(serialized.summary),
553 pending_summary: Task::ready(None),
554 detailed_summary_task: Task::ready(None),
555 detailed_summary_tx,
556 detailed_summary_rx,
557 completion_mode,
558 retry_state: None,
559 messages: serialized
560 .messages
561 .into_iter()
562 .map(|message| Message {
563 id: message.id,
564 role: message.role,
565 segments: message
566 .segments
567 .into_iter()
568 .map(|segment| match segment {
569 SerializedMessageSegment::Text { text } => MessageSegment::Text(text),
570 SerializedMessageSegment::Thinking { text, signature } => {
571 MessageSegment::Thinking { text, signature }
572 }
573 SerializedMessageSegment::RedactedThinking { data } => {
574 MessageSegment::RedactedThinking(data)
575 }
576 })
577 .collect(),
578 loaded_context: LoadedContext {
579 contexts: Vec::new(),
580 text: message.context,
581 images: Vec::new(),
582 },
583 creases: message
584 .creases
585 .into_iter()
586 .map(|crease| MessageCrease {
587 range: crease.start..crease.end,
588 icon_path: crease.icon_path,
589 label: crease.label,
590 context: None,
591 })
592 .collect(),
593 is_hidden: message.is_hidden,
594 ui_only: false, // UI-only messages are not persisted
595 })
596 .collect(),
597 next_message_id,
598 last_prompt_id: PromptId::new(),
599 project_context,
600 checkpoints_by_message: HashMap::default(),
601 completion_count: 0,
602 pending_completions: Vec::new(),
603 last_restore_checkpoint: None,
604 pending_checkpoint: None,
605 project: project.clone(),
606 prompt_builder,
607 tools: tools.clone(),
608 tool_use,
609 action_log: cx.new(|_| ActionLog::new(project)),
610 initial_project_snapshot: Task::ready(serialized.initial_project_snapshot).shared(),
611 request_token_usage: serialized.request_token_usage,
612 cumulative_token_usage: serialized.cumulative_token_usage,
613 exceeded_window_error: None,
614 tool_use_limit_reached: serialized.tool_use_limit_reached,
615 feedback: None,
616 message_feedback: HashMap::default(),
617 last_auto_capture_at: None,
618 last_error_context: None,
619 last_received_chunk_at: None,
620 request_callback: None,
621 remaining_turns: u32::MAX,
622 configured_model,
623 profile: AgentProfile::new(profile_id, tools),
624 }
625 }
626
627 pub fn set_request_callback(
628 &mut self,
629 callback: impl 'static
630 + FnMut(&LanguageModelRequest, &[Result<LanguageModelCompletionEvent, String>]),
631 ) {
632 self.request_callback = Some(Box::new(callback));
633 }
634
635 pub fn id(&self) -> &ThreadId {
636 &self.id
637 }
638
639 pub fn profile(&self) -> &AgentProfile {
640 &self.profile
641 }
642
643 pub fn set_profile(&mut self, id: AgentProfileId, cx: &mut Context<Self>) {
644 if &id != self.profile.id() {
645 self.profile = AgentProfile::new(id, self.tools.clone());
646 cx.emit(ThreadEvent::ProfileChanged);
647 }
648 }
649
650 pub fn is_empty(&self) -> bool {
651 self.messages.is_empty()
652 }
653
654 pub fn updated_at(&self) -> DateTime<Utc> {
655 self.updated_at
656 }
657
658 pub fn touch_updated_at(&mut self) {
659 self.updated_at = Utc::now();
660 }
661
662 pub fn advance_prompt_id(&mut self) {
663 self.last_prompt_id = PromptId::new();
664 }
665
666 pub fn project_context(&self) -> SharedProjectContext {
667 self.project_context.clone()
668 }
669
670 pub fn get_or_init_configured_model(&mut self, cx: &App) -> Option<ConfiguredModel> {
671 if self.configured_model.is_none() {
672 self.configured_model = LanguageModelRegistry::read_global(cx).default_model();
673 }
674 self.configured_model.clone()
675 }
676
677 pub fn configured_model(&self) -> Option<ConfiguredModel> {
678 self.configured_model.clone()
679 }
680
681 pub fn set_configured_model(&mut self, model: Option<ConfiguredModel>, cx: &mut Context<Self>) {
682 self.configured_model = model;
683 cx.notify();
684 }
685
686 pub fn summary(&self) -> &ThreadSummary {
687 &self.summary
688 }
689
690 pub fn set_summary(&mut self, new_summary: impl Into<SharedString>, cx: &mut Context<Self>) {
691 let current_summary = match &self.summary {
692 ThreadSummary::Pending | ThreadSummary::Generating => return,
693 ThreadSummary::Ready(summary) => summary,
694 ThreadSummary::Error => &ThreadSummary::DEFAULT,
695 };
696
697 let mut new_summary = new_summary.into();
698
699 if new_summary.is_empty() {
700 new_summary = ThreadSummary::DEFAULT;
701 }
702
703 if current_summary != &new_summary {
704 self.summary = ThreadSummary::Ready(new_summary);
705 cx.emit(ThreadEvent::SummaryChanged);
706 }
707 }
708
709 pub fn completion_mode(&self) -> CompletionMode {
710 self.completion_mode
711 }
712
713 pub fn set_completion_mode(&mut self, mode: CompletionMode) {
714 self.completion_mode = mode;
715 }
716
717 pub fn message(&self, id: MessageId) -> Option<&Message> {
718 let index = self
719 .messages
720 .binary_search_by(|message| message.id.cmp(&id))
721 .ok()?;
722
723 self.messages.get(index)
724 }
725
726 pub fn messages(&self) -> impl ExactSizeIterator<Item = &Message> {
727 self.messages.iter()
728 }
729
730 pub fn is_generating(&self) -> bool {
731 !self.pending_completions.is_empty() || !self.all_tools_finished()
732 }
733
734 /// Indicates whether streaming of language model events is stale.
735 /// When `is_generating()` is false, this method returns `None`.
736 pub fn is_generation_stale(&self) -> Option<bool> {
737 const STALE_THRESHOLD: u128 = 250;
738
739 self.last_received_chunk_at
740 .map(|instant| instant.elapsed().as_millis() > STALE_THRESHOLD)
741 }
742
743 fn received_chunk(&mut self) {
744 self.last_received_chunk_at = Some(Instant::now());
745 }
746
747 pub fn queue_state(&self) -> Option<QueueState> {
748 self.pending_completions
749 .first()
750 .map(|pending_completion| pending_completion.queue_state)
751 }
752
753 pub fn tools(&self) -> &Entity<ToolWorkingSet> {
754 &self.tools
755 }
756
757 pub fn pending_tool(&self, id: &LanguageModelToolUseId) -> Option<&PendingToolUse> {
758 self.tool_use
759 .pending_tool_uses()
760 .into_iter()
761 .find(|tool_use| &tool_use.id == id)
762 }
763
764 pub fn tools_needing_confirmation(&self) -> impl Iterator<Item = &PendingToolUse> {
765 self.tool_use
766 .pending_tool_uses()
767 .into_iter()
768 .filter(|tool_use| tool_use.status.needs_confirmation())
769 }
770
771 pub fn has_pending_tool_uses(&self) -> bool {
772 !self.tool_use.pending_tool_uses().is_empty()
773 }
774
775 pub fn checkpoint_for_message(&self, id: MessageId) -> Option<ThreadCheckpoint> {
776 self.checkpoints_by_message.get(&id).cloned()
777 }
778
779 pub fn restore_checkpoint(
780 &mut self,
781 checkpoint: ThreadCheckpoint,
782 cx: &mut Context<Self>,
783 ) -> Task<Result<()>> {
784 self.last_restore_checkpoint = Some(LastRestoreCheckpoint::Pending {
785 message_id: checkpoint.message_id,
786 });
787 cx.emit(ThreadEvent::CheckpointChanged);
788 cx.notify();
789
790 let git_store = self.project().read(cx).git_store().clone();
791 let restore = git_store.update(cx, |git_store, cx| {
792 git_store.restore_checkpoint(checkpoint.git_checkpoint.clone(), cx)
793 });
794
795 cx.spawn(async move |this, cx| {
796 let result = restore.await;
797 this.update(cx, |this, cx| {
798 if let Err(err) = result.as_ref() {
799 this.last_restore_checkpoint = Some(LastRestoreCheckpoint::Error {
800 message_id: checkpoint.message_id,
801 error: err.to_string(),
802 });
803 } else {
804 this.truncate(checkpoint.message_id, cx);
805 this.last_restore_checkpoint = None;
806 }
807 this.pending_checkpoint = None;
808 cx.emit(ThreadEvent::CheckpointChanged);
809 cx.notify();
810 })?;
811 result
812 })
813 }
814
815 fn finalize_pending_checkpoint(&mut self, cx: &mut Context<Self>) {
816 let pending_checkpoint = if self.is_generating() {
817 return;
818 } else if let Some(checkpoint) = self.pending_checkpoint.take() {
819 checkpoint
820 } else {
821 return;
822 };
823
824 self.finalize_checkpoint(pending_checkpoint, cx);
825 }
826
827 fn finalize_checkpoint(
828 &mut self,
829 pending_checkpoint: ThreadCheckpoint,
830 cx: &mut Context<Self>,
831 ) {
832 let git_store = self.project.read(cx).git_store().clone();
833 let final_checkpoint = git_store.update(cx, |git_store, cx| git_store.checkpoint(cx));
834 cx.spawn(async move |this, cx| match final_checkpoint.await {
835 Ok(final_checkpoint) => {
836 let equal = git_store
837 .update(cx, |store, cx| {
838 store.compare_checkpoints(
839 pending_checkpoint.git_checkpoint.clone(),
840 final_checkpoint.clone(),
841 cx,
842 )
843 })?
844 .await
845 .unwrap_or(false);
846
847 if !equal {
848 this.update(cx, |this, cx| {
849 this.insert_checkpoint(pending_checkpoint, cx)
850 })?;
851 }
852
853 Ok(())
854 }
855 Err(_) => this.update(cx, |this, cx| {
856 this.insert_checkpoint(pending_checkpoint, cx)
857 }),
858 })
859 .detach();
860 }
861
862 fn insert_checkpoint(&mut self, checkpoint: ThreadCheckpoint, cx: &mut Context<Self>) {
863 self.checkpoints_by_message
864 .insert(checkpoint.message_id, checkpoint);
865 cx.emit(ThreadEvent::CheckpointChanged);
866 cx.notify();
867 }
868
869 pub fn last_restore_checkpoint(&self) -> Option<&LastRestoreCheckpoint> {
870 self.last_restore_checkpoint.as_ref()
871 }
872
873 pub fn truncate(&mut self, message_id: MessageId, cx: &mut Context<Self>) {
874 let Some(message_ix) = self
875 .messages
876 .iter()
877 .rposition(|message| message.id == message_id)
878 else {
879 return;
880 };
881 for deleted_message in self.messages.drain(message_ix..) {
882 self.checkpoints_by_message.remove(&deleted_message.id);
883 }
884 cx.notify();
885 }
886
887 pub fn context_for_message(&self, id: MessageId) -> impl Iterator<Item = &AgentContext> {
888 self.messages
889 .iter()
890 .find(|message| message.id == id)
891 .into_iter()
892 .flat_map(|message| message.loaded_context.contexts.iter())
893 }
894
895 pub fn is_turn_end(&self, ix: usize) -> bool {
896 if self.messages.is_empty() {
897 return false;
898 }
899
900 if !self.is_generating() && ix == self.messages.len() - 1 {
901 return true;
902 }
903
904 let Some(message) = self.messages.get(ix) else {
905 return false;
906 };
907
908 if message.role != Role::Assistant {
909 return false;
910 }
911
912 self.messages
913 .get(ix + 1)
914 .and_then(|message| {
915 self.message(message.id)
916 .map(|next_message| next_message.role == Role::User && !next_message.is_hidden)
917 })
918 .unwrap_or(false)
919 }
920
921 pub fn tool_use_limit_reached(&self) -> bool {
922 self.tool_use_limit_reached
923 }
924
925 /// Returns whether all of the tool uses have finished running.
926 pub fn all_tools_finished(&self) -> bool {
927 // If the only pending tool uses left are the ones with errors, then
928 // that means that we've finished running all of the pending tools.
929 self.tool_use
930 .pending_tool_uses()
931 .iter()
932 .all(|pending_tool_use| pending_tool_use.status.is_error())
933 }
934
935 /// Returns whether any pending tool uses may perform edits
936 pub fn has_pending_edit_tool_uses(&self) -> bool {
937 self.tool_use
938 .pending_tool_uses()
939 .iter()
940 .filter(|pending_tool_use| !pending_tool_use.status.is_error())
941 .any(|pending_tool_use| pending_tool_use.may_perform_edits)
942 }
943
944 pub fn tool_uses_for_message(&self, id: MessageId, cx: &App) -> Vec<ToolUse> {
945 self.tool_use.tool_uses_for_message(id, cx)
946 }
947
948 pub fn tool_results_for_message(
949 &self,
950 assistant_message_id: MessageId,
951 ) -> Vec<&LanguageModelToolResult> {
952 self.tool_use.tool_results_for_message(assistant_message_id)
953 }
954
955 pub fn tool_result(&self, id: &LanguageModelToolUseId) -> Option<&LanguageModelToolResult> {
956 self.tool_use.tool_result(id)
957 }
958
959 pub fn output_for_tool(&self, id: &LanguageModelToolUseId) -> Option<&Arc<str>> {
960 match &self.tool_use.tool_result(id)?.content {
961 LanguageModelToolResultContent::Text(text) => Some(text),
962 LanguageModelToolResultContent::Image(_) => {
963 // TODO: We should display image
964 None
965 }
966 }
967 }
968
969 pub fn card_for_tool(&self, id: &LanguageModelToolUseId) -> Option<AnyToolCard> {
970 self.tool_use.tool_result_card(id).cloned()
971 }
972
973 /// Return tools that are both enabled and supported by the model
974 pub fn available_tools(
975 &self,
976 cx: &App,
977 model: Arc<dyn LanguageModel>,
978 ) -> Vec<LanguageModelRequestTool> {
979 if model.supports_tools() {
980 self.profile
981 .enabled_tools(cx)
982 .into_iter()
983 .filter_map(|(name, tool)| {
984 // Skip tools that cannot be supported
985 let input_schema = tool.input_schema(model.tool_input_format()).ok()?;
986 Some(LanguageModelRequestTool {
987 name: name.into(),
988 description: tool.description(),
989 input_schema,
990 })
991 })
992 .collect()
993 } else {
994 Vec::default()
995 }
996 }
997
998 pub fn insert_user_message(
999 &mut self,
1000 text: impl Into<String>,
1001 loaded_context: ContextLoadResult,
1002 git_checkpoint: Option<GitStoreCheckpoint>,
1003 creases: Vec<MessageCrease>,
1004 cx: &mut Context<Self>,
1005 ) -> MessageId {
1006 if !loaded_context.referenced_buffers.is_empty() {
1007 self.action_log.update(cx, |log, cx| {
1008 for buffer in loaded_context.referenced_buffers {
1009 log.buffer_read(buffer, cx);
1010 }
1011 });
1012 }
1013
1014 let message_id = self.insert_message(
1015 Role::User,
1016 vec![MessageSegment::Text(text.into())],
1017 loaded_context.loaded_context,
1018 creases,
1019 false,
1020 cx,
1021 );
1022
1023 if let Some(git_checkpoint) = git_checkpoint {
1024 self.pending_checkpoint = Some(ThreadCheckpoint {
1025 message_id,
1026 git_checkpoint,
1027 });
1028 }
1029
1030 self.auto_capture_telemetry(cx);
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 let pending_tool_use = 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 pending_tool_use
1657 }
1658
1659 pub fn stream_completion(
1660 &mut self,
1661 request: LanguageModelRequest,
1662 model: Arc<dyn LanguageModel>,
1663 intent: CompletionIntent,
1664 window: Option<AnyWindowHandle>,
1665 cx: &mut Context<Self>,
1666 ) {
1667 self.tool_use_limit_reached = false;
1668
1669 let pending_completion_id = post_inc(&mut self.completion_count);
1670 let mut request_callback_parameters = if self.request_callback.is_some() {
1671 Some((request.clone(), Vec::new()))
1672 } else {
1673 None
1674 };
1675 let prompt_id = self.last_prompt_id.clone();
1676 let tool_use_metadata = ToolUseMetadata {
1677 model: model.clone(),
1678 thread_id: self.id.clone(),
1679 prompt_id: prompt_id.clone(),
1680 };
1681
1682 let completion_mode = request
1683 .mode
1684 .unwrap_or(zed_llm_client::CompletionMode::Normal);
1685
1686 self.last_received_chunk_at = Some(Instant::now());
1687
1688 let task = cx.spawn(async move |thread, cx| {
1689 let stream_completion_future = model.stream_completion(request, &cx);
1690 let initial_token_usage =
1691 thread.read_with(cx, |thread, _cx| thread.cumulative_token_usage);
1692 let stream_completion = async {
1693 let mut events = stream_completion_future.await?;
1694
1695 let mut stop_reason = StopReason::EndTurn;
1696 let mut current_token_usage = TokenUsage::default();
1697
1698 thread
1699 .update(cx, |_thread, cx| {
1700 cx.emit(ThreadEvent::NewRequest);
1701 })
1702 .ok();
1703
1704 let mut request_assistant_message_id = None;
1705
1706 while let Some(event) = events.next().await {
1707 if let Some((_, response_events)) = request_callback_parameters.as_mut() {
1708 response_events
1709 .push(event.as_ref().map_err(|error| error.to_string()).cloned());
1710 }
1711
1712 thread.update(cx, |thread, cx| {
1713 match event? {
1714 LanguageModelCompletionEvent::StartMessage { .. } => {
1715 request_assistant_message_id =
1716 Some(thread.insert_assistant_message(
1717 vec![MessageSegment::Text(String::new())],
1718 cx,
1719 ));
1720 }
1721 LanguageModelCompletionEvent::Stop(reason) => {
1722 stop_reason = reason;
1723 }
1724 LanguageModelCompletionEvent::UsageUpdate(token_usage) => {
1725 thread.update_token_usage_at_last_message(token_usage);
1726 thread.cumulative_token_usage = thread.cumulative_token_usage
1727 + token_usage
1728 - current_token_usage;
1729 current_token_usage = token_usage;
1730 }
1731 LanguageModelCompletionEvent::Text(chunk) => {
1732 thread.received_chunk();
1733
1734 cx.emit(ThreadEvent::ReceivedTextChunk);
1735 if let Some(last_message) = thread.messages.last_mut() {
1736 if last_message.role == Role::Assistant
1737 && !thread.tool_use.has_tool_results(last_message.id)
1738 {
1739 last_message.push_text(&chunk);
1740 cx.emit(ThreadEvent::StreamedAssistantText(
1741 last_message.id,
1742 chunk,
1743 ));
1744 } else {
1745 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1746 // of a new Assistant response.
1747 //
1748 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1749 // will result in duplicating the text of the chunk in the rendered Markdown.
1750 request_assistant_message_id =
1751 Some(thread.insert_assistant_message(
1752 vec![MessageSegment::Text(chunk.to_string())],
1753 cx,
1754 ));
1755 };
1756 }
1757 }
1758 LanguageModelCompletionEvent::Thinking {
1759 text: chunk,
1760 signature,
1761 } => {
1762 thread.received_chunk();
1763
1764 if let Some(last_message) = thread.messages.last_mut() {
1765 if last_message.role == Role::Assistant
1766 && !thread.tool_use.has_tool_results(last_message.id)
1767 {
1768 last_message.push_thinking(&chunk, signature);
1769 cx.emit(ThreadEvent::StreamedAssistantThinking(
1770 last_message.id,
1771 chunk,
1772 ));
1773 } else {
1774 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1775 // of a new Assistant response.
1776 //
1777 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1778 // will result in duplicating the text of the chunk in the rendered Markdown.
1779 request_assistant_message_id =
1780 Some(thread.insert_assistant_message(
1781 vec![MessageSegment::Thinking {
1782 text: chunk.to_string(),
1783 signature,
1784 }],
1785 cx,
1786 ));
1787 };
1788 }
1789 }
1790 LanguageModelCompletionEvent::RedactedThinking { data } => {
1791 thread.received_chunk();
1792
1793 if let Some(last_message) = thread.messages.last_mut() {
1794 if last_message.role == Role::Assistant
1795 && !thread.tool_use.has_tool_results(last_message.id)
1796 {
1797 last_message.push_redacted_thinking(data);
1798 } else {
1799 request_assistant_message_id =
1800 Some(thread.insert_assistant_message(
1801 vec![MessageSegment::RedactedThinking(data)],
1802 cx,
1803 ));
1804 };
1805 }
1806 }
1807 LanguageModelCompletionEvent::ToolUse(tool_use) => {
1808 let last_assistant_message_id = request_assistant_message_id
1809 .unwrap_or_else(|| {
1810 let new_assistant_message_id =
1811 thread.insert_assistant_message(vec![], cx);
1812 request_assistant_message_id =
1813 Some(new_assistant_message_id);
1814 new_assistant_message_id
1815 });
1816
1817 let tool_use_id = tool_use.id.clone();
1818 let streamed_input = if tool_use.is_input_complete {
1819 None
1820 } else {
1821 Some((&tool_use.input).clone())
1822 };
1823
1824 let ui_text = thread.tool_use.request_tool_use(
1825 last_assistant_message_id,
1826 tool_use,
1827 tool_use_metadata.clone(),
1828 cx,
1829 );
1830
1831 if let Some(input) = streamed_input {
1832 cx.emit(ThreadEvent::StreamedToolUse {
1833 tool_use_id,
1834 ui_text,
1835 input,
1836 });
1837 }
1838 }
1839 LanguageModelCompletionEvent::ToolUseJsonParseError {
1840 id,
1841 tool_name,
1842 raw_input: invalid_input_json,
1843 json_parse_error,
1844 } => {
1845 thread.receive_invalid_tool_json(
1846 id,
1847 tool_name,
1848 invalid_input_json,
1849 json_parse_error,
1850 window,
1851 cx,
1852 );
1853 }
1854 LanguageModelCompletionEvent::StatusUpdate(status_update) => {
1855 if let Some(completion) = thread
1856 .pending_completions
1857 .iter_mut()
1858 .find(|completion| completion.id == pending_completion_id)
1859 {
1860 match status_update {
1861 CompletionRequestStatus::Queued { position } => {
1862 completion.queue_state =
1863 QueueState::Queued { position };
1864 }
1865 CompletionRequestStatus::Started => {
1866 completion.queue_state = QueueState::Started;
1867 }
1868 CompletionRequestStatus::Failed {
1869 code,
1870 message,
1871 request_id: _,
1872 retry_after,
1873 } => {
1874 return Err(
1875 LanguageModelCompletionError::from_cloud_failure(
1876 model.upstream_provider_name(),
1877 code,
1878 message,
1879 retry_after.map(Duration::from_secs_f64),
1880 ),
1881 );
1882 }
1883 CompletionRequestStatus::UsageUpdated { amount, limit } => {
1884 thread.update_model_request_usage(
1885 amount as u32,
1886 limit,
1887 cx,
1888 );
1889 }
1890 CompletionRequestStatus::ToolUseLimitReached => {
1891 thread.tool_use_limit_reached = true;
1892 cx.emit(ThreadEvent::ToolUseLimitReached);
1893 }
1894 }
1895 }
1896 }
1897 }
1898
1899 thread.touch_updated_at();
1900 cx.emit(ThreadEvent::StreamedCompletion);
1901 cx.notify();
1902
1903 thread.auto_capture_telemetry(cx);
1904 Ok(())
1905 })??;
1906
1907 smol::future::yield_now().await;
1908 }
1909
1910 thread.update(cx, |thread, cx| {
1911 thread.last_received_chunk_at = None;
1912 thread
1913 .pending_completions
1914 .retain(|completion| completion.id != pending_completion_id);
1915
1916 // If there is a response without tool use, summarize the message. Otherwise,
1917 // allow two tool uses before summarizing.
1918 if matches!(thread.summary, ThreadSummary::Pending)
1919 && thread.messages.len() >= 2
1920 && (!thread.has_pending_tool_uses() || thread.messages.len() >= 6)
1921 {
1922 thread.summarize(cx);
1923 }
1924 })?;
1925
1926 anyhow::Ok(stop_reason)
1927 };
1928
1929 let result = stream_completion.await;
1930 let mut retry_scheduled = false;
1931
1932 thread
1933 .update(cx, |thread, cx| {
1934 thread.finalize_pending_checkpoint(cx);
1935 match result.as_ref() {
1936 Ok(stop_reason) => {
1937 match stop_reason {
1938 StopReason::ToolUse => {
1939 let tool_uses =
1940 thread.use_pending_tools(window, model.clone(), cx);
1941 cx.emit(ThreadEvent::UsePendingTools { tool_uses });
1942 }
1943 StopReason::EndTurn | StopReason::MaxTokens => {
1944 thread.project.update(cx, |project, cx| {
1945 project.set_agent_location(None, cx);
1946 });
1947 }
1948 StopReason::Refusal => {
1949 thread.project.update(cx, |project, cx| {
1950 project.set_agent_location(None, cx);
1951 });
1952
1953 // Remove the turn that was refused.
1954 //
1955 // https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/handle-streaming-refusals#reset-context-after-refusal
1956 {
1957 let mut messages_to_remove = Vec::new();
1958
1959 for (ix, message) in
1960 thread.messages.iter().enumerate().rev()
1961 {
1962 messages_to_remove.push(message.id);
1963
1964 if message.role == Role::User {
1965 if ix == 0 {
1966 break;
1967 }
1968
1969 if let Some(prev_message) =
1970 thread.messages.get(ix - 1)
1971 {
1972 if prev_message.role == Role::Assistant {
1973 break;
1974 }
1975 }
1976 }
1977 }
1978
1979 for message_id in messages_to_remove {
1980 thread.delete_message(message_id, cx);
1981 }
1982 }
1983
1984 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1985 header: "Language model refusal".into(),
1986 message:
1987 "Model refused to generate content for safety reasons."
1988 .into(),
1989 }));
1990 }
1991 }
1992
1993 // We successfully completed, so cancel any remaining retries.
1994 thread.retry_state = None;
1995 }
1996 Err(error) => {
1997 thread.project.update(cx, |project, cx| {
1998 project.set_agent_location(None, cx);
1999 });
2000
2001 if error.is::<PaymentRequiredError>() {
2002 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
2003 } else if let Some(error) =
2004 error.downcast_ref::<ModelRequestLimitReachedError>()
2005 {
2006 cx.emit(ThreadEvent::ShowError(
2007 ThreadError::ModelRequestLimitReached { plan: error.plan },
2008 ));
2009 } else if let Some(completion_error) =
2010 error.downcast_ref::<LanguageModelCompletionError>()
2011 {
2012 match &completion_error {
2013 LanguageModelCompletionError::PromptTooLarge {
2014 tokens, ..
2015 } => {
2016 let tokens = tokens.unwrap_or_else(|| {
2017 // We didn't get an exact token count from the API, so fall back on our estimate.
2018 thread
2019 .total_token_usage()
2020 .map(|usage| usage.total)
2021 .unwrap_or(0)
2022 // We know the context window was exceeded in practice, so if our estimate was
2023 // lower than max tokens, the estimate was wrong; return that we exceeded by 1.
2024 .max(
2025 model
2026 .max_token_count_for_mode(completion_mode)
2027 .saturating_add(1),
2028 )
2029 });
2030 thread.exceeded_window_error = Some(ExceededWindowError {
2031 model_id: model.id(),
2032 token_count: tokens,
2033 });
2034 cx.notify();
2035 }
2036 _ => {
2037 if let Some(retry_strategy) =
2038 Thread::get_retry_strategy(completion_error)
2039 {
2040 retry_scheduled = thread
2041 .handle_retryable_error_with_delay(
2042 &completion_error,
2043 Some(retry_strategy),
2044 model.clone(),
2045 intent,
2046 window,
2047 cx,
2048 );
2049 }
2050 }
2051 }
2052 }
2053
2054 if !retry_scheduled {
2055 thread.cancel_last_completion(window, cx);
2056 }
2057 }
2058 }
2059
2060 if !retry_scheduled {
2061 cx.emit(ThreadEvent::Stopped(result.map_err(Arc::new)));
2062 }
2063
2064 if let Some((request_callback, (request, response_events))) = thread
2065 .request_callback
2066 .as_mut()
2067 .zip(request_callback_parameters.as_ref())
2068 {
2069 request_callback(request, response_events);
2070 }
2071
2072 thread.auto_capture_telemetry(cx);
2073
2074 if let Ok(initial_usage) = initial_token_usage {
2075 let usage = thread.cumulative_token_usage - initial_usage;
2076
2077 telemetry::event!(
2078 "Assistant Thread Completion",
2079 thread_id = thread.id().to_string(),
2080 prompt_id = prompt_id,
2081 model = model.telemetry_id(),
2082 model_provider = model.provider_id().to_string(),
2083 input_tokens = usage.input_tokens,
2084 output_tokens = usage.output_tokens,
2085 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2086 cache_read_input_tokens = usage.cache_read_input_tokens,
2087 );
2088 }
2089 })
2090 .ok();
2091 });
2092
2093 self.pending_completions.push(PendingCompletion {
2094 id: pending_completion_id,
2095 queue_state: QueueState::Sending,
2096 _task: task,
2097 });
2098 }
2099
2100 pub fn summarize(&mut self, cx: &mut Context<Self>) {
2101 let Some(model) = LanguageModelRegistry::read_global(cx).thread_summary_model() else {
2102 println!("No thread summary model");
2103 return;
2104 };
2105
2106 if !model.provider.is_authenticated(cx) {
2107 return;
2108 }
2109
2110 let added_user_message = include_str!("./prompts/summarize_thread_prompt.txt");
2111
2112 let request = self.to_summarize_request(
2113 &model.model,
2114 CompletionIntent::ThreadSummarization,
2115 added_user_message.into(),
2116 cx,
2117 );
2118
2119 self.summary = ThreadSummary::Generating;
2120
2121 self.pending_summary = cx.spawn(async move |this, cx| {
2122 let result = async {
2123 let mut messages = model.model.stream_completion(request, &cx).await?;
2124
2125 let mut new_summary = String::new();
2126 while let Some(event) = messages.next().await {
2127 let Ok(event) = event else {
2128 continue;
2129 };
2130 let text = match event {
2131 LanguageModelCompletionEvent::Text(text) => text,
2132 LanguageModelCompletionEvent::StatusUpdate(
2133 CompletionRequestStatus::UsageUpdated { amount, limit },
2134 ) => {
2135 this.update(cx, |thread, cx| {
2136 thread.update_model_request_usage(amount as u32, limit, cx);
2137 })?;
2138 continue;
2139 }
2140 _ => continue,
2141 };
2142
2143 let mut lines = text.lines();
2144 new_summary.extend(lines.next());
2145
2146 // Stop if the LLM generated multiple lines.
2147 if lines.next().is_some() {
2148 break;
2149 }
2150 }
2151
2152 anyhow::Ok(new_summary)
2153 }
2154 .await;
2155
2156 this.update(cx, |this, cx| {
2157 match result {
2158 Ok(new_summary) => {
2159 if new_summary.is_empty() {
2160 this.summary = ThreadSummary::Error;
2161 } else {
2162 this.summary = ThreadSummary::Ready(new_summary.into());
2163 }
2164 }
2165 Err(err) => {
2166 this.summary = ThreadSummary::Error;
2167 log::error!("Failed to generate thread summary: {}", err);
2168 }
2169 }
2170 cx.emit(ThreadEvent::SummaryGenerated);
2171 })
2172 .log_err()?;
2173
2174 Some(())
2175 });
2176 }
2177
2178 fn get_retry_strategy(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2179 use LanguageModelCompletionError::*;
2180
2181 // General strategy here:
2182 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2183 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2184 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2185 match error {
2186 HttpResponseError {
2187 status_code: StatusCode::TOO_MANY_REQUESTS,
2188 ..
2189 } => Some(RetryStrategy::ExponentialBackoff {
2190 initial_delay: BASE_RETRY_DELAY,
2191 max_attempts: MAX_RETRY_ATTEMPTS,
2192 }),
2193 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2194 Some(RetryStrategy::Fixed {
2195 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2196 max_attempts: MAX_RETRY_ATTEMPTS,
2197 })
2198 }
2199 UpstreamProviderError {
2200 status,
2201 retry_after,
2202 ..
2203 } => match *status {
2204 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2205 Some(RetryStrategy::Fixed {
2206 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2207 max_attempts: MAX_RETRY_ATTEMPTS,
2208 })
2209 }
2210 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2211 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2212 // Internal Server Error could be anything, retry up to 3 times.
2213 max_attempts: 3,
2214 }),
2215 status => {
2216 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2217 // but we frequently get them in practice. See https://http.dev/529
2218 if status.as_u16() == 529 {
2219 Some(RetryStrategy::Fixed {
2220 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2221 max_attempts: MAX_RETRY_ATTEMPTS,
2222 })
2223 } else {
2224 Some(RetryStrategy::Fixed {
2225 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2226 max_attempts: 2,
2227 })
2228 }
2229 }
2230 },
2231 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2232 delay: BASE_RETRY_DELAY,
2233 max_attempts: 3,
2234 }),
2235 ApiReadResponseError { .. }
2236 | HttpSend { .. }
2237 | DeserializeResponse { .. }
2238 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2239 delay: BASE_RETRY_DELAY,
2240 max_attempts: 3,
2241 }),
2242 // Retrying these errors definitely shouldn't help.
2243 HttpResponseError {
2244 status_code:
2245 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2246 ..
2247 }
2248 | AuthenticationError { .. }
2249 | PermissionError { .. } => None,
2250 // These errors might be transient, so retry them
2251 SerializeRequest { .. }
2252 | BuildRequestBody { .. }
2253 | PromptTooLarge { .. }
2254 | ApiEndpointNotFound { .. }
2255 | NoApiKey { .. } => Some(RetryStrategy::Fixed {
2256 delay: BASE_RETRY_DELAY,
2257 max_attempts: 2,
2258 }),
2259 // Retry all other 4xx and 5xx errors once.
2260 HttpResponseError { status_code, .. }
2261 if status_code.is_client_error() || status_code.is_server_error() =>
2262 {
2263 Some(RetryStrategy::Fixed {
2264 delay: BASE_RETRY_DELAY,
2265 max_attempts: 3,
2266 })
2267 }
2268 // Conservatively assume that any other errors are non-retryable
2269 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2270 delay: BASE_RETRY_DELAY,
2271 max_attempts: 2,
2272 }),
2273 }
2274 }
2275
2276 fn handle_retryable_error_with_delay(
2277 &mut self,
2278 error: &LanguageModelCompletionError,
2279 strategy: Option<RetryStrategy>,
2280 model: Arc<dyn LanguageModel>,
2281 intent: CompletionIntent,
2282 window: Option<AnyWindowHandle>,
2283 cx: &mut Context<Self>,
2284 ) -> bool {
2285 // Store context for the Retry button
2286 self.last_error_context = Some((model.clone(), intent));
2287
2288 // Only auto-retry if Burn Mode is enabled
2289 if self.completion_mode != CompletionMode::Burn {
2290 // Show error with retry options
2291 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2292 message: format!(
2293 "{}\n\nTo automatically retry when similar errors happen, enable Burn Mode.",
2294 error
2295 )
2296 .into(),
2297 can_enable_burn_mode: true,
2298 }));
2299 return false;
2300 }
2301
2302 let Some(strategy) = strategy.or_else(|| Self::get_retry_strategy(error)) else {
2303 return false;
2304 };
2305
2306 let max_attempts = match &strategy {
2307 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2308 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2309 };
2310
2311 let retry_state = self.retry_state.get_or_insert(RetryState {
2312 attempt: 0,
2313 max_attempts,
2314 intent,
2315 });
2316
2317 retry_state.attempt += 1;
2318 let attempt = retry_state.attempt;
2319 let max_attempts = retry_state.max_attempts;
2320 let intent = retry_state.intent;
2321
2322 if attempt <= max_attempts {
2323 let delay = match &strategy {
2324 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2325 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2326 Duration::from_secs(delay_secs)
2327 }
2328 RetryStrategy::Fixed { delay, .. } => *delay,
2329 };
2330
2331 // Add a transient message to inform the user
2332 let delay_secs = delay.as_secs();
2333 let retry_message = if max_attempts == 1 {
2334 format!("{error}. Retrying in {delay_secs} seconds...")
2335 } else {
2336 format!(
2337 "{error}. Retrying (attempt {attempt} of {max_attempts}) \
2338 in {delay_secs} seconds..."
2339 )
2340 };
2341 log::warn!(
2342 "Retrying completion request (attempt {attempt} of {max_attempts}) \
2343 in {delay_secs} seconds: {error:?}",
2344 );
2345
2346 // Add a UI-only message instead of a regular message
2347 let id = self.next_message_id.post_inc();
2348 self.messages.push(Message {
2349 id,
2350 role: Role::System,
2351 segments: vec![MessageSegment::Text(retry_message)],
2352 loaded_context: LoadedContext::default(),
2353 creases: Vec::new(),
2354 is_hidden: false,
2355 ui_only: true,
2356 });
2357 cx.emit(ThreadEvent::MessageAdded(id));
2358
2359 // Schedule the retry
2360 let thread_handle = cx.entity().downgrade();
2361
2362 cx.spawn(async move |_thread, cx| {
2363 cx.background_executor().timer(delay).await;
2364
2365 thread_handle
2366 .update(cx, |thread, cx| {
2367 // Retry the completion
2368 thread.send_to_model(model, intent, window, cx);
2369 })
2370 .log_err();
2371 })
2372 .detach();
2373
2374 true
2375 } else {
2376 // Max retries exceeded
2377 self.retry_state = None;
2378
2379 // Stop generating since we're giving up on retrying.
2380 self.pending_completions.clear();
2381
2382 // Show error alongside a Retry button, but no
2383 // Enable Burn Mode button (since it's already enabled)
2384 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2385 message: format!("Failed after retrying: {}", error).into(),
2386 can_enable_burn_mode: false,
2387 }));
2388
2389 false
2390 }
2391 }
2392
2393 pub fn start_generating_detailed_summary_if_needed(
2394 &mut self,
2395 thread_store: WeakEntity<ThreadStore>,
2396 cx: &mut Context<Self>,
2397 ) {
2398 let Some(last_message_id) = self.messages.last().map(|message| message.id) else {
2399 return;
2400 };
2401
2402 match &*self.detailed_summary_rx.borrow() {
2403 DetailedSummaryState::Generating { message_id, .. }
2404 | DetailedSummaryState::Generated { message_id, .. }
2405 if *message_id == last_message_id =>
2406 {
2407 // Already up-to-date
2408 return;
2409 }
2410 _ => {}
2411 }
2412
2413 let Some(ConfiguredModel { model, provider }) =
2414 LanguageModelRegistry::read_global(cx).thread_summary_model()
2415 else {
2416 return;
2417 };
2418
2419 if !provider.is_authenticated(cx) {
2420 return;
2421 }
2422
2423 let added_user_message = include_str!("./prompts/summarize_thread_detailed_prompt.txt");
2424
2425 let request = self.to_summarize_request(
2426 &model,
2427 CompletionIntent::ThreadContextSummarization,
2428 added_user_message.into(),
2429 cx,
2430 );
2431
2432 *self.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generating {
2433 message_id: last_message_id,
2434 };
2435
2436 // Replace the detailed summarization task if there is one, cancelling it. It would probably
2437 // be better to allow the old task to complete, but this would require logic for choosing
2438 // which result to prefer (the old task could complete after the new one, resulting in a
2439 // stale summary).
2440 self.detailed_summary_task = cx.spawn(async move |thread, cx| {
2441 let stream = model.stream_completion_text(request, &cx);
2442 let Some(mut messages) = stream.await.log_err() else {
2443 thread
2444 .update(cx, |thread, _cx| {
2445 *thread.detailed_summary_tx.borrow_mut() =
2446 DetailedSummaryState::NotGenerated;
2447 })
2448 .ok()?;
2449 return None;
2450 };
2451
2452 let mut new_detailed_summary = String::new();
2453
2454 while let Some(chunk) = messages.stream.next().await {
2455 if let Some(chunk) = chunk.log_err() {
2456 new_detailed_summary.push_str(&chunk);
2457 }
2458 }
2459
2460 thread
2461 .update(cx, |thread, _cx| {
2462 *thread.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generated {
2463 text: new_detailed_summary.into(),
2464 message_id: last_message_id,
2465 };
2466 })
2467 .ok()?;
2468
2469 // Save thread so its summary can be reused later
2470 if let Some(thread) = thread.upgrade() {
2471 if let Ok(Ok(save_task)) = cx.update(|cx| {
2472 thread_store
2473 .update(cx, |thread_store, cx| thread_store.save_thread(&thread, cx))
2474 }) {
2475 save_task.await.log_err();
2476 }
2477 }
2478
2479 Some(())
2480 });
2481 }
2482
2483 pub async fn wait_for_detailed_summary_or_text(
2484 this: &Entity<Self>,
2485 cx: &mut AsyncApp,
2486 ) -> Option<SharedString> {
2487 let mut detailed_summary_rx = this
2488 .read_with(cx, |this, _cx| this.detailed_summary_rx.clone())
2489 .ok()?;
2490 loop {
2491 match detailed_summary_rx.recv().await? {
2492 DetailedSummaryState::Generating { .. } => {}
2493 DetailedSummaryState::NotGenerated => {
2494 return this.read_with(cx, |this, _cx| this.text().into()).ok();
2495 }
2496 DetailedSummaryState::Generated { text, .. } => return Some(text),
2497 }
2498 }
2499 }
2500
2501 pub fn latest_detailed_summary_or_text(&self) -> SharedString {
2502 self.detailed_summary_rx
2503 .borrow()
2504 .text()
2505 .unwrap_or_else(|| self.text().into())
2506 }
2507
2508 pub fn is_generating_detailed_summary(&self) -> bool {
2509 matches!(
2510 &*self.detailed_summary_rx.borrow(),
2511 DetailedSummaryState::Generating { .. }
2512 )
2513 }
2514
2515 pub fn use_pending_tools(
2516 &mut self,
2517 window: Option<AnyWindowHandle>,
2518 model: Arc<dyn LanguageModel>,
2519 cx: &mut Context<Self>,
2520 ) -> Vec<PendingToolUse> {
2521 self.auto_capture_telemetry(cx);
2522 let request =
2523 Arc::new(self.to_completion_request(model.clone(), CompletionIntent::ToolResults, cx));
2524 let pending_tool_uses = self
2525 .tool_use
2526 .pending_tool_uses()
2527 .into_iter()
2528 .filter(|tool_use| tool_use.status.is_idle())
2529 .cloned()
2530 .collect::<Vec<_>>();
2531
2532 for tool_use in pending_tool_uses.iter() {
2533 self.use_pending_tool(tool_use.clone(), request.clone(), model.clone(), window, cx);
2534 }
2535
2536 pending_tool_uses
2537 }
2538
2539 fn use_pending_tool(
2540 &mut self,
2541 tool_use: PendingToolUse,
2542 request: Arc<LanguageModelRequest>,
2543 model: Arc<dyn LanguageModel>,
2544 window: Option<AnyWindowHandle>,
2545 cx: &mut Context<Self>,
2546 ) {
2547 let Some(tool) = self.tools.read(cx).tool(&tool_use.name, cx) else {
2548 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2549 };
2550
2551 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
2552 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2553 }
2554
2555 if tool.needs_confirmation(&tool_use.input, cx)
2556 && !AgentSettings::get_global(cx).always_allow_tool_actions
2557 {
2558 self.tool_use.confirm_tool_use(
2559 tool_use.id,
2560 tool_use.ui_text,
2561 tool_use.input,
2562 request,
2563 tool,
2564 );
2565 cx.emit(ThreadEvent::ToolConfirmationNeeded);
2566 } else {
2567 self.run_tool(
2568 tool_use.id,
2569 tool_use.ui_text,
2570 tool_use.input,
2571 request,
2572 tool,
2573 model,
2574 window,
2575 cx,
2576 );
2577 }
2578 }
2579
2580 pub fn handle_hallucinated_tool_use(
2581 &mut self,
2582 tool_use_id: LanguageModelToolUseId,
2583 hallucinated_tool_name: Arc<str>,
2584 window: Option<AnyWindowHandle>,
2585 cx: &mut Context<Thread>,
2586 ) {
2587 let available_tools = self.profile.enabled_tools(cx);
2588
2589 let tool_list = available_tools
2590 .iter()
2591 .map(|(name, tool)| format!("- {}: {}", name, tool.description()))
2592 .collect::<Vec<_>>()
2593 .join("\n");
2594
2595 let error_message = format!(
2596 "The tool '{}' doesn't exist or is not enabled. Available tools:\n{}",
2597 hallucinated_tool_name, tool_list
2598 );
2599
2600 let pending_tool_use = self.tool_use.insert_tool_output(
2601 tool_use_id.clone(),
2602 hallucinated_tool_name,
2603 Err(anyhow!("Missing tool call: {error_message}")),
2604 self.configured_model.as_ref(),
2605 self.completion_mode,
2606 );
2607
2608 cx.emit(ThreadEvent::MissingToolUse {
2609 tool_use_id: tool_use_id.clone(),
2610 ui_text: error_message.into(),
2611 });
2612
2613 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2614 }
2615
2616 pub fn receive_invalid_tool_json(
2617 &mut self,
2618 tool_use_id: LanguageModelToolUseId,
2619 tool_name: Arc<str>,
2620 invalid_json: Arc<str>,
2621 error: String,
2622 window: Option<AnyWindowHandle>,
2623 cx: &mut Context<Thread>,
2624 ) {
2625 log::error!("The model returned invalid input JSON: {invalid_json}");
2626
2627 let pending_tool_use = self.tool_use.insert_tool_output(
2628 tool_use_id.clone(),
2629 tool_name,
2630 Err(anyhow!("Error parsing input JSON: {error}")),
2631 self.configured_model.as_ref(),
2632 self.completion_mode,
2633 );
2634 let ui_text = if let Some(pending_tool_use) = &pending_tool_use {
2635 pending_tool_use.ui_text.clone()
2636 } else {
2637 log::error!(
2638 "There was no pending tool use for tool use {tool_use_id}, even though it finished (with invalid input JSON)."
2639 );
2640 format!("Unknown tool {}", tool_use_id).into()
2641 };
2642
2643 cx.emit(ThreadEvent::InvalidToolInput {
2644 tool_use_id: tool_use_id.clone(),
2645 ui_text,
2646 invalid_input_json: invalid_json,
2647 });
2648
2649 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2650 }
2651
2652 pub fn run_tool(
2653 &mut self,
2654 tool_use_id: LanguageModelToolUseId,
2655 ui_text: impl Into<SharedString>,
2656 input: serde_json::Value,
2657 request: Arc<LanguageModelRequest>,
2658 tool: Arc<dyn Tool>,
2659 model: Arc<dyn LanguageModel>,
2660 window: Option<AnyWindowHandle>,
2661 cx: &mut Context<Thread>,
2662 ) {
2663 let task =
2664 self.spawn_tool_use(tool_use_id.clone(), request, input, tool, model, window, cx);
2665 self.tool_use
2666 .run_pending_tool(tool_use_id, ui_text.into(), task);
2667 }
2668
2669 fn spawn_tool_use(
2670 &mut self,
2671 tool_use_id: LanguageModelToolUseId,
2672 request: Arc<LanguageModelRequest>,
2673 input: serde_json::Value,
2674 tool: Arc<dyn Tool>,
2675 model: Arc<dyn LanguageModel>,
2676 window: Option<AnyWindowHandle>,
2677 cx: &mut Context<Thread>,
2678 ) -> Task<()> {
2679 let tool_name: Arc<str> = tool.name().into();
2680
2681 let tool_result = tool.run(
2682 input,
2683 request,
2684 self.project.clone(),
2685 self.action_log.clone(),
2686 model,
2687 window,
2688 cx,
2689 );
2690
2691 // Store the card separately if it exists
2692 if let Some(card) = tool_result.card.clone() {
2693 self.tool_use
2694 .insert_tool_result_card(tool_use_id.clone(), card);
2695 }
2696
2697 cx.spawn({
2698 async move |thread: WeakEntity<Thread>, cx| {
2699 let output = tool_result.output.await;
2700
2701 thread
2702 .update(cx, |thread, cx| {
2703 let pending_tool_use = thread.tool_use.insert_tool_output(
2704 tool_use_id.clone(),
2705 tool_name,
2706 output,
2707 thread.configured_model.as_ref(),
2708 thread.completion_mode,
2709 );
2710 thread.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2711 })
2712 .ok();
2713 }
2714 })
2715 }
2716
2717 fn tool_finished(
2718 &mut self,
2719 tool_use_id: LanguageModelToolUseId,
2720 pending_tool_use: Option<PendingToolUse>,
2721 canceled: bool,
2722 window: Option<AnyWindowHandle>,
2723 cx: &mut Context<Self>,
2724 ) {
2725 if self.all_tools_finished() {
2726 if let Some(ConfiguredModel { model, .. }) = self.configured_model.as_ref() {
2727 if !canceled {
2728 self.send_to_model(model.clone(), CompletionIntent::ToolResults, window, cx);
2729 }
2730 self.auto_capture_telemetry(cx);
2731 }
2732 }
2733
2734 cx.emit(ThreadEvent::ToolFinished {
2735 tool_use_id,
2736 pending_tool_use,
2737 });
2738 }
2739
2740 /// Cancels the last pending completion, if there are any pending.
2741 ///
2742 /// Returns whether a completion was canceled.
2743 pub fn cancel_last_completion(
2744 &mut self,
2745 window: Option<AnyWindowHandle>,
2746 cx: &mut Context<Self>,
2747 ) -> bool {
2748 let mut canceled = self.pending_completions.pop().is_some() || self.retry_state.is_some();
2749
2750 self.retry_state = None;
2751
2752 for pending_tool_use in self.tool_use.cancel_pending() {
2753 canceled = true;
2754 self.tool_finished(
2755 pending_tool_use.id.clone(),
2756 Some(pending_tool_use),
2757 true,
2758 window,
2759 cx,
2760 );
2761 }
2762
2763 if canceled {
2764 cx.emit(ThreadEvent::CompletionCanceled);
2765
2766 // When canceled, we always want to insert the checkpoint.
2767 // (We skip over finalize_pending_checkpoint, because it
2768 // would conclude we didn't have anything to insert here.)
2769 if let Some(checkpoint) = self.pending_checkpoint.take() {
2770 self.insert_checkpoint(checkpoint, cx);
2771 }
2772 } else {
2773 self.finalize_pending_checkpoint(cx);
2774 }
2775
2776 canceled
2777 }
2778
2779 /// Signals that any in-progress editing should be canceled.
2780 ///
2781 /// This method is used to notify listeners (like ActiveThread) that
2782 /// they should cancel any editing operations.
2783 pub fn cancel_editing(&mut self, cx: &mut Context<Self>) {
2784 cx.emit(ThreadEvent::CancelEditing);
2785 }
2786
2787 pub fn feedback(&self) -> Option<ThreadFeedback> {
2788 self.feedback
2789 }
2790
2791 pub fn message_feedback(&self, message_id: MessageId) -> Option<ThreadFeedback> {
2792 self.message_feedback.get(&message_id).copied()
2793 }
2794
2795 pub fn report_message_feedback(
2796 &mut self,
2797 message_id: MessageId,
2798 feedback: ThreadFeedback,
2799 cx: &mut Context<Self>,
2800 ) -> Task<Result<()>> {
2801 if self.message_feedback.get(&message_id) == Some(&feedback) {
2802 return Task::ready(Ok(()));
2803 }
2804
2805 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2806 let serialized_thread = self.serialize(cx);
2807 let thread_id = self.id().clone();
2808 let client = self.project.read(cx).client();
2809
2810 let enabled_tool_names: Vec<String> = self
2811 .profile
2812 .enabled_tools(cx)
2813 .iter()
2814 .map(|(name, _)| name.clone().into())
2815 .collect();
2816
2817 self.message_feedback.insert(message_id, feedback);
2818
2819 cx.notify();
2820
2821 let message_content = self
2822 .message(message_id)
2823 .map(|msg| msg.to_string())
2824 .unwrap_or_default();
2825
2826 cx.background_spawn(async move {
2827 let final_project_snapshot = final_project_snapshot.await;
2828 let serialized_thread = serialized_thread.await?;
2829 let thread_data =
2830 serde_json::to_value(serialized_thread).unwrap_or_else(|_| serde_json::Value::Null);
2831
2832 let rating = match feedback {
2833 ThreadFeedback::Positive => "positive",
2834 ThreadFeedback::Negative => "negative",
2835 };
2836 telemetry::event!(
2837 "Assistant Thread Rated",
2838 rating,
2839 thread_id,
2840 enabled_tool_names,
2841 message_id = message_id.0,
2842 message_content,
2843 thread_data,
2844 final_project_snapshot
2845 );
2846 client.telemetry().flush_events().await;
2847
2848 Ok(())
2849 })
2850 }
2851
2852 pub fn report_feedback(
2853 &mut self,
2854 feedback: ThreadFeedback,
2855 cx: &mut Context<Self>,
2856 ) -> Task<Result<()>> {
2857 let last_assistant_message_id = self
2858 .messages
2859 .iter()
2860 .rev()
2861 .find(|msg| msg.role == Role::Assistant)
2862 .map(|msg| msg.id);
2863
2864 if let Some(message_id) = last_assistant_message_id {
2865 self.report_message_feedback(message_id, feedback, cx)
2866 } else {
2867 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2868 let serialized_thread = self.serialize(cx);
2869 let thread_id = self.id().clone();
2870 let client = self.project.read(cx).client();
2871 self.feedback = Some(feedback);
2872 cx.notify();
2873
2874 cx.background_spawn(async move {
2875 let final_project_snapshot = final_project_snapshot.await;
2876 let serialized_thread = serialized_thread.await?;
2877 let thread_data = serde_json::to_value(serialized_thread)
2878 .unwrap_or_else(|_| serde_json::Value::Null);
2879
2880 let rating = match feedback {
2881 ThreadFeedback::Positive => "positive",
2882 ThreadFeedback::Negative => "negative",
2883 };
2884 telemetry::event!(
2885 "Assistant Thread Rated",
2886 rating,
2887 thread_id,
2888 thread_data,
2889 final_project_snapshot
2890 );
2891 client.telemetry().flush_events().await;
2892
2893 Ok(())
2894 })
2895 }
2896 }
2897
2898 /// Create a snapshot of the current project state including git information and unsaved buffers.
2899 fn project_snapshot(
2900 project: Entity<Project>,
2901 cx: &mut Context<Self>,
2902 ) -> Task<Arc<ProjectSnapshot>> {
2903 let git_store = project.read(cx).git_store().clone();
2904 let worktree_snapshots: Vec<_> = project
2905 .read(cx)
2906 .visible_worktrees(cx)
2907 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
2908 .collect();
2909
2910 cx.spawn(async move |_, cx| {
2911 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
2912
2913 let mut unsaved_buffers = Vec::new();
2914 cx.update(|app_cx| {
2915 let buffer_store = project.read(app_cx).buffer_store();
2916 for buffer_handle in buffer_store.read(app_cx).buffers() {
2917 let buffer = buffer_handle.read(app_cx);
2918 if buffer.is_dirty() {
2919 if let Some(file) = buffer.file() {
2920 let path = file.path().to_string_lossy().to_string();
2921 unsaved_buffers.push(path);
2922 }
2923 }
2924 }
2925 })
2926 .ok();
2927
2928 Arc::new(ProjectSnapshot {
2929 worktree_snapshots,
2930 unsaved_buffer_paths: unsaved_buffers,
2931 timestamp: Utc::now(),
2932 })
2933 })
2934 }
2935
2936 fn worktree_snapshot(
2937 worktree: Entity<project::Worktree>,
2938 git_store: Entity<GitStore>,
2939 cx: &App,
2940 ) -> Task<WorktreeSnapshot> {
2941 cx.spawn(async move |cx| {
2942 // Get worktree path and snapshot
2943 let worktree_info = cx.update(|app_cx| {
2944 let worktree = worktree.read(app_cx);
2945 let path = worktree.abs_path().to_string_lossy().to_string();
2946 let snapshot = worktree.snapshot();
2947 (path, snapshot)
2948 });
2949
2950 let Ok((worktree_path, _snapshot)) = worktree_info else {
2951 return WorktreeSnapshot {
2952 worktree_path: String::new(),
2953 git_state: None,
2954 };
2955 };
2956
2957 let git_state = git_store
2958 .update(cx, |git_store, cx| {
2959 git_store
2960 .repositories()
2961 .values()
2962 .find(|repo| {
2963 repo.read(cx)
2964 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
2965 .is_some()
2966 })
2967 .cloned()
2968 })
2969 .ok()
2970 .flatten()
2971 .map(|repo| {
2972 repo.update(cx, |repo, _| {
2973 let current_branch =
2974 repo.branch.as_ref().map(|branch| branch.name().to_owned());
2975 repo.send_job(None, |state, _| async move {
2976 let RepositoryState::Local { backend, .. } = state else {
2977 return GitState {
2978 remote_url: None,
2979 head_sha: None,
2980 current_branch,
2981 diff: None,
2982 };
2983 };
2984
2985 let remote_url = backend.remote_url("origin");
2986 let head_sha = backend.head_sha().await;
2987 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
2988
2989 GitState {
2990 remote_url,
2991 head_sha,
2992 current_branch,
2993 diff,
2994 }
2995 })
2996 })
2997 });
2998
2999 let git_state = match git_state {
3000 Some(git_state) => match git_state.ok() {
3001 Some(git_state) => git_state.await.ok(),
3002 None => None,
3003 },
3004 None => None,
3005 };
3006
3007 WorktreeSnapshot {
3008 worktree_path,
3009 git_state,
3010 }
3011 })
3012 }
3013
3014 pub fn to_markdown(&self, cx: &App) -> Result<String> {
3015 let mut markdown = Vec::new();
3016
3017 let summary = self.summary().or_default();
3018 writeln!(markdown, "# {summary}\n")?;
3019
3020 for message in self.messages() {
3021 writeln!(
3022 markdown,
3023 "## {role}\n",
3024 role = match message.role {
3025 Role::User => "User",
3026 Role::Assistant => "Agent",
3027 Role::System => "System",
3028 }
3029 )?;
3030
3031 if !message.loaded_context.text.is_empty() {
3032 writeln!(markdown, "{}", message.loaded_context.text)?;
3033 }
3034
3035 if !message.loaded_context.images.is_empty() {
3036 writeln!(
3037 markdown,
3038 "\n{} images attached as context.\n",
3039 message.loaded_context.images.len()
3040 )?;
3041 }
3042
3043 for segment in &message.segments {
3044 match segment {
3045 MessageSegment::Text(text) => writeln!(markdown, "{}\n", text)?,
3046 MessageSegment::Thinking { text, .. } => {
3047 writeln!(markdown, "<think>\n{}\n</think>\n", text)?
3048 }
3049 MessageSegment::RedactedThinking(_) => {}
3050 }
3051 }
3052
3053 for tool_use in self.tool_uses_for_message(message.id, cx) {
3054 writeln!(
3055 markdown,
3056 "**Use Tool: {} ({})**",
3057 tool_use.name, tool_use.id
3058 )?;
3059 writeln!(markdown, "```json")?;
3060 writeln!(
3061 markdown,
3062 "{}",
3063 serde_json::to_string_pretty(&tool_use.input)?
3064 )?;
3065 writeln!(markdown, "```")?;
3066 }
3067
3068 for tool_result in self.tool_results_for_message(message.id) {
3069 write!(markdown, "\n**Tool Results: {}", tool_result.tool_use_id)?;
3070 if tool_result.is_error {
3071 write!(markdown, " (Error)")?;
3072 }
3073
3074 writeln!(markdown, "**\n")?;
3075 match &tool_result.content {
3076 LanguageModelToolResultContent::Text(text) => {
3077 writeln!(markdown, "{text}")?;
3078 }
3079 LanguageModelToolResultContent::Image(image) => {
3080 writeln!(markdown, "", image.source)?;
3081 }
3082 }
3083
3084 if let Some(output) = tool_result.output.as_ref() {
3085 writeln!(
3086 markdown,
3087 "\n\nDebug Output:\n\n```json\n{}\n```\n",
3088 serde_json::to_string_pretty(output)?
3089 )?;
3090 }
3091 }
3092 }
3093
3094 Ok(String::from_utf8_lossy(&markdown).to_string())
3095 }
3096
3097 pub fn keep_edits_in_range(
3098 &mut self,
3099 buffer: Entity<language::Buffer>,
3100 buffer_range: Range<language::Anchor>,
3101 cx: &mut Context<Self>,
3102 ) {
3103 self.action_log.update(cx, |action_log, cx| {
3104 action_log.keep_edits_in_range(buffer, buffer_range, cx)
3105 });
3106 }
3107
3108 pub fn keep_all_edits(&mut self, cx: &mut Context<Self>) {
3109 self.action_log
3110 .update(cx, |action_log, cx| action_log.keep_all_edits(cx));
3111 }
3112
3113 pub fn reject_edits_in_ranges(
3114 &mut self,
3115 buffer: Entity<language::Buffer>,
3116 buffer_ranges: Vec<Range<language::Anchor>>,
3117 cx: &mut Context<Self>,
3118 ) -> Task<Result<()>> {
3119 self.action_log.update(cx, |action_log, cx| {
3120 action_log.reject_edits_in_ranges(buffer, buffer_ranges, cx)
3121 })
3122 }
3123
3124 pub fn action_log(&self) -> &Entity<ActionLog> {
3125 &self.action_log
3126 }
3127
3128 pub fn project(&self) -> &Entity<Project> {
3129 &self.project
3130 }
3131
3132 pub fn auto_capture_telemetry(&mut self, cx: &mut Context<Self>) {
3133 if !cx.has_flag::<feature_flags::ThreadAutoCaptureFeatureFlag>() {
3134 return;
3135 }
3136
3137 let now = Instant::now();
3138 if let Some(last) = self.last_auto_capture_at {
3139 if now.duration_since(last).as_secs() < 10 {
3140 return;
3141 }
3142 }
3143
3144 self.last_auto_capture_at = Some(now);
3145
3146 let thread_id = self.id().clone();
3147 let github_login = self
3148 .project
3149 .read(cx)
3150 .user_store()
3151 .read(cx)
3152 .current_user()
3153 .map(|user| user.github_login.clone());
3154 let client = self.project.read(cx).client();
3155 let serialize_task = self.serialize(cx);
3156
3157 cx.background_executor()
3158 .spawn(async move {
3159 if let Ok(serialized_thread) = serialize_task.await {
3160 if let Ok(thread_data) = serde_json::to_value(serialized_thread) {
3161 telemetry::event!(
3162 "Agent Thread Auto-Captured",
3163 thread_id = thread_id.to_string(),
3164 thread_data = thread_data,
3165 auto_capture_reason = "tracked_user",
3166 github_login = github_login
3167 );
3168
3169 client.telemetry().flush_events().await;
3170 }
3171 }
3172 })
3173 .detach();
3174 }
3175
3176 pub fn cumulative_token_usage(&self) -> TokenUsage {
3177 self.cumulative_token_usage
3178 }
3179
3180 pub fn token_usage_up_to_message(&self, message_id: MessageId) -> TotalTokenUsage {
3181 let Some(model) = self.configured_model.as_ref() else {
3182 return TotalTokenUsage::default();
3183 };
3184
3185 let max = model
3186 .model
3187 .max_token_count_for_mode(self.completion_mode().into());
3188
3189 let index = self
3190 .messages
3191 .iter()
3192 .position(|msg| msg.id == message_id)
3193 .unwrap_or(0);
3194
3195 if index == 0 {
3196 return TotalTokenUsage { total: 0, max };
3197 }
3198
3199 let token_usage = &self
3200 .request_token_usage
3201 .get(index - 1)
3202 .cloned()
3203 .unwrap_or_default();
3204
3205 TotalTokenUsage {
3206 total: token_usage.total_tokens(),
3207 max,
3208 }
3209 }
3210
3211 pub fn total_token_usage(&self) -> Option<TotalTokenUsage> {
3212 let model = self.configured_model.as_ref()?;
3213
3214 let max = model
3215 .model
3216 .max_token_count_for_mode(self.completion_mode().into());
3217
3218 if let Some(exceeded_error) = &self.exceeded_window_error {
3219 if model.model.id() == exceeded_error.model_id {
3220 return Some(TotalTokenUsage {
3221 total: exceeded_error.token_count,
3222 max,
3223 });
3224 }
3225 }
3226
3227 let total = self
3228 .token_usage_at_last_message()
3229 .unwrap_or_default()
3230 .total_tokens();
3231
3232 Some(TotalTokenUsage { total, max })
3233 }
3234
3235 fn token_usage_at_last_message(&self) -> Option<TokenUsage> {
3236 self.request_token_usage
3237 .get(self.messages.len().saturating_sub(1))
3238 .or_else(|| self.request_token_usage.last())
3239 .cloned()
3240 }
3241
3242 fn update_token_usage_at_last_message(&mut self, token_usage: TokenUsage) {
3243 let placeholder = self.token_usage_at_last_message().unwrap_or_default();
3244 self.request_token_usage
3245 .resize(self.messages.len(), placeholder);
3246
3247 if let Some(last) = self.request_token_usage.last_mut() {
3248 *last = token_usage;
3249 }
3250 }
3251
3252 fn update_model_request_usage(&self, amount: u32, limit: UsageLimit, cx: &mut Context<Self>) {
3253 self.project.update(cx, |project, cx| {
3254 project.user_store().update(cx, |user_store, cx| {
3255 user_store.update_model_request_usage(
3256 ModelRequestUsage(RequestUsage {
3257 amount: amount as i32,
3258 limit,
3259 }),
3260 cx,
3261 )
3262 })
3263 });
3264 }
3265
3266 pub fn deny_tool_use(
3267 &mut self,
3268 tool_use_id: LanguageModelToolUseId,
3269 tool_name: Arc<str>,
3270 window: Option<AnyWindowHandle>,
3271 cx: &mut Context<Self>,
3272 ) {
3273 let err = Err(anyhow::anyhow!(
3274 "Permission to run tool action denied by user"
3275 ));
3276
3277 self.tool_use.insert_tool_output(
3278 tool_use_id.clone(),
3279 tool_name,
3280 err,
3281 self.configured_model.as_ref(),
3282 self.completion_mode,
3283 );
3284 self.tool_finished(tool_use_id.clone(), None, true, window, cx);
3285 }
3286}
3287
3288#[derive(Debug, Clone, Error)]
3289pub enum ThreadError {
3290 #[error("Payment required")]
3291 PaymentRequired,
3292 #[error("Model request limit reached")]
3293 ModelRequestLimitReached { plan: Plan },
3294 #[error("Message {header}: {message}")]
3295 Message {
3296 header: SharedString,
3297 message: SharedString,
3298 },
3299 #[error("Retryable error: {message}")]
3300 RetryableError {
3301 message: SharedString,
3302 can_enable_burn_mode: bool,
3303 },
3304}
3305
3306#[derive(Debug, Clone)]
3307pub enum ThreadEvent {
3308 ShowError(ThreadError),
3309 StreamedCompletion,
3310 ReceivedTextChunk,
3311 NewRequest,
3312 StreamedAssistantText(MessageId, String),
3313 StreamedAssistantThinking(MessageId, String),
3314 StreamedToolUse {
3315 tool_use_id: LanguageModelToolUseId,
3316 ui_text: Arc<str>,
3317 input: serde_json::Value,
3318 },
3319 MissingToolUse {
3320 tool_use_id: LanguageModelToolUseId,
3321 ui_text: Arc<str>,
3322 },
3323 InvalidToolInput {
3324 tool_use_id: LanguageModelToolUseId,
3325 ui_text: Arc<str>,
3326 invalid_input_json: Arc<str>,
3327 },
3328 Stopped(Result<StopReason, Arc<anyhow::Error>>),
3329 MessageAdded(MessageId),
3330 MessageEdited(MessageId),
3331 MessageDeleted(MessageId),
3332 SummaryGenerated,
3333 SummaryChanged,
3334 UsePendingTools {
3335 tool_uses: Vec<PendingToolUse>,
3336 },
3337 ToolFinished {
3338 #[allow(unused)]
3339 tool_use_id: LanguageModelToolUseId,
3340 /// The pending tool use that corresponds to this tool.
3341 pending_tool_use: Option<PendingToolUse>,
3342 },
3343 CheckpointChanged,
3344 ToolConfirmationNeeded,
3345 ToolUseLimitReached,
3346 CancelEditing,
3347 CompletionCanceled,
3348 ProfileChanged,
3349}
3350
3351impl EventEmitter<ThreadEvent> for Thread {}
3352
3353struct PendingCompletion {
3354 id: usize,
3355 queue_state: QueueState,
3356 _task: Task<()>,
3357}
3358
3359#[cfg(test)]
3360mod tests {
3361 use super::*;
3362 use crate::{
3363 context::load_context, context_store::ContextStore, thread_store, thread_store::ThreadStore,
3364 };
3365
3366 // Test-specific constants
3367 const TEST_RATE_LIMIT_RETRY_SECS: u64 = 30;
3368 use agent_settings::{AgentProfileId, AgentSettings, LanguageModelParameters};
3369 use assistant_tool::ToolRegistry;
3370 use assistant_tools;
3371 use futures::StreamExt;
3372 use futures::future::BoxFuture;
3373 use futures::stream::BoxStream;
3374 use gpui::TestAppContext;
3375 use http_client;
3376 use language_model::fake_provider::{FakeLanguageModel, FakeLanguageModelProvider};
3377 use language_model::{
3378 LanguageModelCompletionError, LanguageModelName, LanguageModelProviderId,
3379 LanguageModelProviderName, LanguageModelToolChoice,
3380 };
3381 use parking_lot::Mutex;
3382 use project::{FakeFs, Project};
3383 use prompt_store::PromptBuilder;
3384 use serde_json::json;
3385 use settings::{Settings, SettingsStore};
3386 use std::sync::Arc;
3387 use std::time::Duration;
3388 use theme::ThemeSettings;
3389 use util::path;
3390 use workspace::Workspace;
3391
3392 #[gpui::test]
3393 async fn test_message_with_context(cx: &mut TestAppContext) {
3394 init_test_settings(cx);
3395
3396 let project = create_test_project(
3397 cx,
3398 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3399 )
3400 .await;
3401
3402 let (_workspace, _thread_store, thread, context_store, model) =
3403 setup_test_environment(cx, project.clone()).await;
3404
3405 add_file_to_context(&project, &context_store, "test/code.rs", cx)
3406 .await
3407 .unwrap();
3408
3409 let context =
3410 context_store.read_with(cx, |store, _| store.context().next().cloned().unwrap());
3411 let loaded_context = cx
3412 .update(|cx| load_context(vec![context], &project, &None, cx))
3413 .await;
3414
3415 // Insert user message with context
3416 let message_id = thread.update(cx, |thread, cx| {
3417 thread.insert_user_message(
3418 "Please explain this code",
3419 loaded_context,
3420 None,
3421 Vec::new(),
3422 cx,
3423 )
3424 });
3425
3426 // Check content and context in message object
3427 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3428
3429 // Use different path format strings based on platform for the test
3430 #[cfg(windows)]
3431 let path_part = r"test\code.rs";
3432 #[cfg(not(windows))]
3433 let path_part = "test/code.rs";
3434
3435 let expected_context = format!(
3436 r#"
3437<context>
3438The following items were attached by the user. They are up-to-date and don't need to be re-read.
3439
3440<files>
3441```rs {path_part}
3442fn main() {{
3443 println!("Hello, world!");
3444}}
3445```
3446</files>
3447</context>
3448"#
3449 );
3450
3451 assert_eq!(message.role, Role::User);
3452 assert_eq!(message.segments.len(), 1);
3453 assert_eq!(
3454 message.segments[0],
3455 MessageSegment::Text("Please explain this code".to_string())
3456 );
3457 assert_eq!(message.loaded_context.text, expected_context);
3458
3459 // Check message in request
3460 let request = thread.update(cx, |thread, cx| {
3461 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3462 });
3463
3464 assert_eq!(request.messages.len(), 2);
3465 let expected_full_message = format!("{}Please explain this code", expected_context);
3466 assert_eq!(request.messages[1].string_contents(), expected_full_message);
3467 }
3468
3469 #[gpui::test]
3470 async fn test_only_include_new_contexts(cx: &mut TestAppContext) {
3471 init_test_settings(cx);
3472
3473 let project = create_test_project(
3474 cx,
3475 json!({
3476 "file1.rs": "fn function1() {}\n",
3477 "file2.rs": "fn function2() {}\n",
3478 "file3.rs": "fn function3() {}\n",
3479 "file4.rs": "fn function4() {}\n",
3480 }),
3481 )
3482 .await;
3483
3484 let (_, _thread_store, thread, context_store, model) =
3485 setup_test_environment(cx, project.clone()).await;
3486
3487 // First message with context 1
3488 add_file_to_context(&project, &context_store, "test/file1.rs", cx)
3489 .await
3490 .unwrap();
3491 let new_contexts = context_store.update(cx, |store, cx| {
3492 store.new_context_for_thread(thread.read(cx), None)
3493 });
3494 assert_eq!(new_contexts.len(), 1);
3495 let loaded_context = cx
3496 .update(|cx| load_context(new_contexts, &project, &None, cx))
3497 .await;
3498 let message1_id = thread.update(cx, |thread, cx| {
3499 thread.insert_user_message("Message 1", loaded_context, None, Vec::new(), cx)
3500 });
3501
3502 // Second message with contexts 1 and 2 (context 1 should be skipped as it's already included)
3503 add_file_to_context(&project, &context_store, "test/file2.rs", cx)
3504 .await
3505 .unwrap();
3506 let new_contexts = context_store.update(cx, |store, cx| {
3507 store.new_context_for_thread(thread.read(cx), None)
3508 });
3509 assert_eq!(new_contexts.len(), 1);
3510 let loaded_context = cx
3511 .update(|cx| load_context(new_contexts, &project, &None, cx))
3512 .await;
3513 let message2_id = thread.update(cx, |thread, cx| {
3514 thread.insert_user_message("Message 2", loaded_context, None, Vec::new(), cx)
3515 });
3516
3517 // Third message with all three contexts (contexts 1 and 2 should be skipped)
3518 //
3519 add_file_to_context(&project, &context_store, "test/file3.rs", cx)
3520 .await
3521 .unwrap();
3522 let new_contexts = context_store.update(cx, |store, cx| {
3523 store.new_context_for_thread(thread.read(cx), None)
3524 });
3525 assert_eq!(new_contexts.len(), 1);
3526 let loaded_context = cx
3527 .update(|cx| load_context(new_contexts, &project, &None, cx))
3528 .await;
3529 let message3_id = thread.update(cx, |thread, cx| {
3530 thread.insert_user_message("Message 3", loaded_context, None, Vec::new(), cx)
3531 });
3532
3533 // Check what contexts are included in each message
3534 let (message1, message2, message3) = thread.read_with(cx, |thread, _| {
3535 (
3536 thread.message(message1_id).unwrap().clone(),
3537 thread.message(message2_id).unwrap().clone(),
3538 thread.message(message3_id).unwrap().clone(),
3539 )
3540 });
3541
3542 // First message should include context 1
3543 assert!(message1.loaded_context.text.contains("file1.rs"));
3544
3545 // Second message should include only context 2 (not 1)
3546 assert!(!message2.loaded_context.text.contains("file1.rs"));
3547 assert!(message2.loaded_context.text.contains("file2.rs"));
3548
3549 // Third message should include only context 3 (not 1 or 2)
3550 assert!(!message3.loaded_context.text.contains("file1.rs"));
3551 assert!(!message3.loaded_context.text.contains("file2.rs"));
3552 assert!(message3.loaded_context.text.contains("file3.rs"));
3553
3554 // Check entire request to make sure all contexts are properly included
3555 let request = thread.update(cx, |thread, cx| {
3556 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3557 });
3558
3559 // The request should contain all 3 messages
3560 assert_eq!(request.messages.len(), 4);
3561
3562 // Check that the contexts are properly formatted in each message
3563 assert!(request.messages[1].string_contents().contains("file1.rs"));
3564 assert!(!request.messages[1].string_contents().contains("file2.rs"));
3565 assert!(!request.messages[1].string_contents().contains("file3.rs"));
3566
3567 assert!(!request.messages[2].string_contents().contains("file1.rs"));
3568 assert!(request.messages[2].string_contents().contains("file2.rs"));
3569 assert!(!request.messages[2].string_contents().contains("file3.rs"));
3570
3571 assert!(!request.messages[3].string_contents().contains("file1.rs"));
3572 assert!(!request.messages[3].string_contents().contains("file2.rs"));
3573 assert!(request.messages[3].string_contents().contains("file3.rs"));
3574
3575 add_file_to_context(&project, &context_store, "test/file4.rs", cx)
3576 .await
3577 .unwrap();
3578 let new_contexts = context_store.update(cx, |store, cx| {
3579 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3580 });
3581 assert_eq!(new_contexts.len(), 3);
3582 let loaded_context = cx
3583 .update(|cx| load_context(new_contexts, &project, &None, cx))
3584 .await
3585 .loaded_context;
3586
3587 assert!(!loaded_context.text.contains("file1.rs"));
3588 assert!(loaded_context.text.contains("file2.rs"));
3589 assert!(loaded_context.text.contains("file3.rs"));
3590 assert!(loaded_context.text.contains("file4.rs"));
3591
3592 let new_contexts = context_store.update(cx, |store, cx| {
3593 // Remove file4.rs
3594 store.remove_context(&loaded_context.contexts[2].handle(), cx);
3595 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3596 });
3597 assert_eq!(new_contexts.len(), 2);
3598 let loaded_context = cx
3599 .update(|cx| load_context(new_contexts, &project, &None, cx))
3600 .await
3601 .loaded_context;
3602
3603 assert!(!loaded_context.text.contains("file1.rs"));
3604 assert!(loaded_context.text.contains("file2.rs"));
3605 assert!(loaded_context.text.contains("file3.rs"));
3606 assert!(!loaded_context.text.contains("file4.rs"));
3607
3608 let new_contexts = context_store.update(cx, |store, cx| {
3609 // Remove file3.rs
3610 store.remove_context(&loaded_context.contexts[1].handle(), cx);
3611 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3612 });
3613 assert_eq!(new_contexts.len(), 1);
3614 let loaded_context = cx
3615 .update(|cx| load_context(new_contexts, &project, &None, cx))
3616 .await
3617 .loaded_context;
3618
3619 assert!(!loaded_context.text.contains("file1.rs"));
3620 assert!(loaded_context.text.contains("file2.rs"));
3621 assert!(!loaded_context.text.contains("file3.rs"));
3622 assert!(!loaded_context.text.contains("file4.rs"));
3623 }
3624
3625 #[gpui::test]
3626 async fn test_message_without_files(cx: &mut TestAppContext) {
3627 init_test_settings(cx);
3628
3629 let project = create_test_project(
3630 cx,
3631 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3632 )
3633 .await;
3634
3635 let (_, _thread_store, thread, _context_store, model) =
3636 setup_test_environment(cx, project.clone()).await;
3637
3638 // Insert user message without any context (empty context vector)
3639 let message_id = thread.update(cx, |thread, cx| {
3640 thread.insert_user_message(
3641 "What is the best way to learn Rust?",
3642 ContextLoadResult::default(),
3643 None,
3644 Vec::new(),
3645 cx,
3646 )
3647 });
3648
3649 // Check content and context in message object
3650 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3651
3652 // Context should be empty when no files are included
3653 assert_eq!(message.role, Role::User);
3654 assert_eq!(message.segments.len(), 1);
3655 assert_eq!(
3656 message.segments[0],
3657 MessageSegment::Text("What is the best way to learn Rust?".to_string())
3658 );
3659 assert_eq!(message.loaded_context.text, "");
3660
3661 // Check message in request
3662 let request = thread.update(cx, |thread, cx| {
3663 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3664 });
3665
3666 assert_eq!(request.messages.len(), 2);
3667 assert_eq!(
3668 request.messages[1].string_contents(),
3669 "What is the best way to learn Rust?"
3670 );
3671
3672 // Add second message, also without context
3673 let message2_id = thread.update(cx, |thread, cx| {
3674 thread.insert_user_message(
3675 "Are there any good books?",
3676 ContextLoadResult::default(),
3677 None,
3678 Vec::new(),
3679 cx,
3680 )
3681 });
3682
3683 let message2 =
3684 thread.read_with(cx, |thread, _| thread.message(message2_id).unwrap().clone());
3685 assert_eq!(message2.loaded_context.text, "");
3686
3687 // Check that both messages appear in the request
3688 let request = thread.update(cx, |thread, cx| {
3689 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3690 });
3691
3692 assert_eq!(request.messages.len(), 3);
3693 assert_eq!(
3694 request.messages[1].string_contents(),
3695 "What is the best way to learn Rust?"
3696 );
3697 assert_eq!(
3698 request.messages[2].string_contents(),
3699 "Are there any good books?"
3700 );
3701 }
3702
3703 #[gpui::test]
3704 #[ignore] // turn this test on when project_notifications tool is re-enabled
3705 async fn test_stale_buffer_notification(cx: &mut TestAppContext) {
3706 init_test_settings(cx);
3707
3708 let project = create_test_project(
3709 cx,
3710 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3711 )
3712 .await;
3713
3714 let (_workspace, _thread_store, thread, context_store, model) =
3715 setup_test_environment(cx, project.clone()).await;
3716
3717 // Add a buffer to the context. This will be a tracked buffer
3718 let buffer = add_file_to_context(&project, &context_store, "test/code.rs", cx)
3719 .await
3720 .unwrap();
3721
3722 let context = context_store
3723 .read_with(cx, |store, _| store.context().next().cloned())
3724 .unwrap();
3725 let loaded_context = cx
3726 .update(|cx| load_context(vec![context], &project, &None, cx))
3727 .await;
3728
3729 // Insert user message and assistant response
3730 thread.update(cx, |thread, cx| {
3731 thread.insert_user_message("Explain this code", loaded_context, None, Vec::new(), cx);
3732 thread.insert_assistant_message(
3733 vec![MessageSegment::Text("This code prints 42.".into())],
3734 cx,
3735 );
3736 });
3737 cx.run_until_parked();
3738
3739 // We shouldn't have a stale buffer notification yet
3740 let notifications = thread.read_with(cx, |thread, _| {
3741 find_tool_uses(thread, "project_notifications")
3742 });
3743 assert!(
3744 notifications.is_empty(),
3745 "Should not have stale buffer notification before buffer is modified"
3746 );
3747
3748 // Modify the buffer
3749 buffer.update(cx, |buffer, cx| {
3750 buffer.edit(
3751 [(1..1, "\n println!(\"Added a new line\");\n")],
3752 None,
3753 cx,
3754 );
3755 });
3756
3757 // Insert another user message
3758 thread.update(cx, |thread, cx| {
3759 thread.insert_user_message(
3760 "What does the code do now?",
3761 ContextLoadResult::default(),
3762 None,
3763 Vec::new(),
3764 cx,
3765 )
3766 });
3767 cx.run_until_parked();
3768
3769 // Check for the stale buffer warning
3770 thread.update(cx, |thread, cx| {
3771 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3772 });
3773 cx.run_until_parked();
3774
3775 let notifications = thread.read_with(cx, |thread, _cx| {
3776 find_tool_uses(thread, "project_notifications")
3777 });
3778
3779 let [notification] = notifications.as_slice() else {
3780 panic!("Should have a `project_notifications` tool use");
3781 };
3782
3783 let Some(notification_content) = notification.content.to_str() else {
3784 panic!("`project_notifications` should return text");
3785 };
3786
3787 assert!(notification_content.contains("These files have changed since the last read:"));
3788 assert!(notification_content.contains("code.rs"));
3789
3790 // Insert another user message and flush notifications again
3791 thread.update(cx, |thread, cx| {
3792 thread.insert_user_message(
3793 "Can you tell me more?",
3794 ContextLoadResult::default(),
3795 None,
3796 Vec::new(),
3797 cx,
3798 )
3799 });
3800
3801 thread.update(cx, |thread, cx| {
3802 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3803 });
3804 cx.run_until_parked();
3805
3806 // There should be no new notifications (we already flushed one)
3807 let notifications = thread.read_with(cx, |thread, _cx| {
3808 find_tool_uses(thread, "project_notifications")
3809 });
3810
3811 assert_eq!(
3812 notifications.len(),
3813 1,
3814 "Should still have only one notification after second flush - no duplicates"
3815 );
3816 }
3817
3818 fn find_tool_uses(thread: &Thread, tool_name: &str) -> Vec<LanguageModelToolResult> {
3819 thread
3820 .messages()
3821 .flat_map(|message| {
3822 thread
3823 .tool_results_for_message(message.id)
3824 .into_iter()
3825 .filter(|result| result.tool_name == tool_name.into())
3826 .cloned()
3827 .collect::<Vec<_>>()
3828 })
3829 .collect()
3830 }
3831
3832 #[gpui::test]
3833 async fn test_storing_profile_setting_per_thread(cx: &mut TestAppContext) {
3834 init_test_settings(cx);
3835
3836 let project = create_test_project(
3837 cx,
3838 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3839 )
3840 .await;
3841
3842 let (_workspace, thread_store, thread, _context_store, _model) =
3843 setup_test_environment(cx, project.clone()).await;
3844
3845 // Check that we are starting with the default profile
3846 let profile = cx.read(|cx| thread.read(cx).profile.clone());
3847 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3848 assert_eq!(
3849 profile,
3850 AgentProfile::new(AgentProfileId::default(), tool_set)
3851 );
3852 }
3853
3854 #[gpui::test]
3855 async fn test_serializing_thread_profile(cx: &mut TestAppContext) {
3856 init_test_settings(cx);
3857
3858 let project = create_test_project(
3859 cx,
3860 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3861 )
3862 .await;
3863
3864 let (_workspace, thread_store, thread, _context_store, _model) =
3865 setup_test_environment(cx, project.clone()).await;
3866
3867 // Profile gets serialized with default values
3868 let serialized = thread
3869 .update(cx, |thread, cx| thread.serialize(cx))
3870 .await
3871 .unwrap();
3872
3873 assert_eq!(serialized.profile, Some(AgentProfileId::default()));
3874
3875 let deserialized = cx.update(|cx| {
3876 thread.update(cx, |thread, cx| {
3877 Thread::deserialize(
3878 thread.id.clone(),
3879 serialized,
3880 thread.project.clone(),
3881 thread.tools.clone(),
3882 thread.prompt_builder.clone(),
3883 thread.project_context.clone(),
3884 None,
3885 cx,
3886 )
3887 })
3888 });
3889 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3890
3891 assert_eq!(
3892 deserialized.profile,
3893 AgentProfile::new(AgentProfileId::default(), tool_set)
3894 );
3895 }
3896
3897 #[gpui::test]
3898 async fn test_temperature_setting(cx: &mut TestAppContext) {
3899 init_test_settings(cx);
3900
3901 let project = create_test_project(
3902 cx,
3903 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3904 )
3905 .await;
3906
3907 let (_workspace, _thread_store, thread, _context_store, model) =
3908 setup_test_environment(cx, project.clone()).await;
3909
3910 // Both model and provider
3911 cx.update(|cx| {
3912 AgentSettings::override_global(
3913 AgentSettings {
3914 model_parameters: vec![LanguageModelParameters {
3915 provider: Some(model.provider_id().0.to_string().into()),
3916 model: Some(model.id().0.clone()),
3917 temperature: Some(0.66),
3918 }],
3919 ..AgentSettings::get_global(cx).clone()
3920 },
3921 cx,
3922 );
3923 });
3924
3925 let request = thread.update(cx, |thread, cx| {
3926 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3927 });
3928 assert_eq!(request.temperature, Some(0.66));
3929
3930 // Only model
3931 cx.update(|cx| {
3932 AgentSettings::override_global(
3933 AgentSettings {
3934 model_parameters: vec![LanguageModelParameters {
3935 provider: None,
3936 model: Some(model.id().0.clone()),
3937 temperature: Some(0.66),
3938 }],
3939 ..AgentSettings::get_global(cx).clone()
3940 },
3941 cx,
3942 );
3943 });
3944
3945 let request = thread.update(cx, |thread, cx| {
3946 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3947 });
3948 assert_eq!(request.temperature, Some(0.66));
3949
3950 // Only provider
3951 cx.update(|cx| {
3952 AgentSettings::override_global(
3953 AgentSettings {
3954 model_parameters: vec![LanguageModelParameters {
3955 provider: Some(model.provider_id().0.to_string().into()),
3956 model: None,
3957 temperature: Some(0.66),
3958 }],
3959 ..AgentSettings::get_global(cx).clone()
3960 },
3961 cx,
3962 );
3963 });
3964
3965 let request = thread.update(cx, |thread, cx| {
3966 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3967 });
3968 assert_eq!(request.temperature, Some(0.66));
3969
3970 // Same model name, different provider
3971 cx.update(|cx| {
3972 AgentSettings::override_global(
3973 AgentSettings {
3974 model_parameters: vec![LanguageModelParameters {
3975 provider: Some("anthropic".into()),
3976 model: Some(model.id().0.clone()),
3977 temperature: Some(0.66),
3978 }],
3979 ..AgentSettings::get_global(cx).clone()
3980 },
3981 cx,
3982 );
3983 });
3984
3985 let request = thread.update(cx, |thread, cx| {
3986 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3987 });
3988 assert_eq!(request.temperature, None);
3989 }
3990
3991 #[gpui::test]
3992 async fn test_thread_summary(cx: &mut TestAppContext) {
3993 init_test_settings(cx);
3994
3995 let project = create_test_project(cx, json!({})).await;
3996
3997 let (_, _thread_store, thread, _context_store, model) =
3998 setup_test_environment(cx, project.clone()).await;
3999
4000 // Initial state should be pending
4001 thread.read_with(cx, |thread, _| {
4002 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4003 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4004 });
4005
4006 // Manually setting the summary should not be allowed in this state
4007 thread.update(cx, |thread, cx| {
4008 thread.set_summary("This should not work", cx);
4009 });
4010
4011 thread.read_with(cx, |thread, _| {
4012 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4013 });
4014
4015 // Send a message
4016 thread.update(cx, |thread, cx| {
4017 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
4018 thread.send_to_model(
4019 model.clone(),
4020 CompletionIntent::ThreadSummarization,
4021 None,
4022 cx,
4023 );
4024 });
4025
4026 let fake_model = model.as_fake();
4027 simulate_successful_response(&fake_model, cx);
4028
4029 // Should start generating summary when there are >= 2 messages
4030 thread.read_with(cx, |thread, _| {
4031 assert_eq!(*thread.summary(), ThreadSummary::Generating);
4032 });
4033
4034 // Should not be able to set the summary while generating
4035 thread.update(cx, |thread, cx| {
4036 thread.set_summary("This should not work either", cx);
4037 });
4038
4039 thread.read_with(cx, |thread, _| {
4040 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4041 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4042 });
4043
4044 cx.run_until_parked();
4045 fake_model.stream_last_completion_response("Brief");
4046 fake_model.stream_last_completion_response(" Introduction");
4047 fake_model.end_last_completion_stream();
4048 cx.run_until_parked();
4049
4050 // Summary should be set
4051 thread.read_with(cx, |thread, _| {
4052 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4053 assert_eq!(thread.summary().or_default(), "Brief Introduction");
4054 });
4055
4056 // Now we should be able to set a summary
4057 thread.update(cx, |thread, cx| {
4058 thread.set_summary("Brief Intro", cx);
4059 });
4060
4061 thread.read_with(cx, |thread, _| {
4062 assert_eq!(thread.summary().or_default(), "Brief Intro");
4063 });
4064
4065 // Test setting an empty summary (should default to DEFAULT)
4066 thread.update(cx, |thread, cx| {
4067 thread.set_summary("", cx);
4068 });
4069
4070 thread.read_with(cx, |thread, _| {
4071 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4072 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4073 });
4074 }
4075
4076 #[gpui::test]
4077 async fn test_thread_summary_error_set_manually(cx: &mut TestAppContext) {
4078 init_test_settings(cx);
4079
4080 let project = create_test_project(cx, json!({})).await;
4081
4082 let (_, _thread_store, thread, _context_store, model) =
4083 setup_test_environment(cx, project.clone()).await;
4084
4085 test_summarize_error(&model, &thread, cx);
4086
4087 // Now we should be able to set a summary
4088 thread.update(cx, |thread, cx| {
4089 thread.set_summary("Brief Intro", cx);
4090 });
4091
4092 thread.read_with(cx, |thread, _| {
4093 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4094 assert_eq!(thread.summary().or_default(), "Brief Intro");
4095 });
4096 }
4097
4098 #[gpui::test]
4099 async fn test_thread_summary_error_retry(cx: &mut TestAppContext) {
4100 init_test_settings(cx);
4101
4102 let project = create_test_project(cx, json!({})).await;
4103
4104 let (_, _thread_store, thread, _context_store, model) =
4105 setup_test_environment(cx, project.clone()).await;
4106
4107 test_summarize_error(&model, &thread, cx);
4108
4109 // Sending another message should not trigger another summarize request
4110 thread.update(cx, |thread, cx| {
4111 thread.insert_user_message(
4112 "How are you?",
4113 ContextLoadResult::default(),
4114 None,
4115 vec![],
4116 cx,
4117 );
4118 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4119 });
4120
4121 let fake_model = model.as_fake();
4122 simulate_successful_response(&fake_model, cx);
4123
4124 thread.read_with(cx, |thread, _| {
4125 // State is still Error, not Generating
4126 assert!(matches!(thread.summary(), ThreadSummary::Error));
4127 });
4128
4129 // But the summarize request can be invoked manually
4130 thread.update(cx, |thread, cx| {
4131 thread.summarize(cx);
4132 });
4133
4134 thread.read_with(cx, |thread, _| {
4135 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4136 });
4137
4138 cx.run_until_parked();
4139 fake_model.stream_last_completion_response("A successful summary");
4140 fake_model.end_last_completion_stream();
4141 cx.run_until_parked();
4142
4143 thread.read_with(cx, |thread, _| {
4144 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4145 assert_eq!(thread.summary().or_default(), "A successful summary");
4146 });
4147 }
4148
4149 // Helper to create a model that returns errors
4150 enum TestError {
4151 Overloaded,
4152 InternalServerError,
4153 }
4154
4155 struct ErrorInjector {
4156 inner: Arc<FakeLanguageModel>,
4157 error_type: TestError,
4158 }
4159
4160 impl ErrorInjector {
4161 fn new(error_type: TestError) -> Self {
4162 Self {
4163 inner: Arc::new(FakeLanguageModel::default()),
4164 error_type,
4165 }
4166 }
4167 }
4168
4169 impl LanguageModel for ErrorInjector {
4170 fn id(&self) -> LanguageModelId {
4171 self.inner.id()
4172 }
4173
4174 fn name(&self) -> LanguageModelName {
4175 self.inner.name()
4176 }
4177
4178 fn provider_id(&self) -> LanguageModelProviderId {
4179 self.inner.provider_id()
4180 }
4181
4182 fn provider_name(&self) -> LanguageModelProviderName {
4183 self.inner.provider_name()
4184 }
4185
4186 fn supports_tools(&self) -> bool {
4187 self.inner.supports_tools()
4188 }
4189
4190 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4191 self.inner.supports_tool_choice(choice)
4192 }
4193
4194 fn supports_images(&self) -> bool {
4195 self.inner.supports_images()
4196 }
4197
4198 fn telemetry_id(&self) -> String {
4199 self.inner.telemetry_id()
4200 }
4201
4202 fn max_token_count(&self) -> u64 {
4203 self.inner.max_token_count()
4204 }
4205
4206 fn count_tokens(
4207 &self,
4208 request: LanguageModelRequest,
4209 cx: &App,
4210 ) -> BoxFuture<'static, Result<u64>> {
4211 self.inner.count_tokens(request, cx)
4212 }
4213
4214 fn stream_completion(
4215 &self,
4216 _request: LanguageModelRequest,
4217 _cx: &AsyncApp,
4218 ) -> BoxFuture<
4219 'static,
4220 Result<
4221 BoxStream<
4222 'static,
4223 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4224 >,
4225 LanguageModelCompletionError,
4226 >,
4227 > {
4228 let error = match self.error_type {
4229 TestError::Overloaded => LanguageModelCompletionError::ServerOverloaded {
4230 provider: self.provider_name(),
4231 retry_after: None,
4232 },
4233 TestError::InternalServerError => {
4234 LanguageModelCompletionError::ApiInternalServerError {
4235 provider: self.provider_name(),
4236 message: "I'm a teapot orbiting the sun".to_string(),
4237 }
4238 }
4239 };
4240 async move {
4241 let stream = futures::stream::once(async move { Err(error) });
4242 Ok(stream.boxed())
4243 }
4244 .boxed()
4245 }
4246
4247 fn as_fake(&self) -> &FakeLanguageModel {
4248 &self.inner
4249 }
4250 }
4251
4252 #[gpui::test]
4253 async fn test_retry_on_overloaded_error(cx: &mut TestAppContext) {
4254 init_test_settings(cx);
4255
4256 let project = create_test_project(cx, json!({})).await;
4257 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4258
4259 // Enable Burn Mode to allow retries
4260 thread.update(cx, |thread, _| {
4261 thread.set_completion_mode(CompletionMode::Burn);
4262 });
4263
4264 // Create model that returns overloaded error
4265 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4266
4267 // Insert a user message
4268 thread.update(cx, |thread, cx| {
4269 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4270 });
4271
4272 // Start completion
4273 thread.update(cx, |thread, cx| {
4274 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4275 });
4276
4277 cx.run_until_parked();
4278
4279 thread.read_with(cx, |thread, _| {
4280 assert!(thread.retry_state.is_some(), "Should have retry state");
4281 let retry_state = thread.retry_state.as_ref().unwrap();
4282 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4283 assert_eq!(
4284 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
4285 "Should retry MAX_RETRY_ATTEMPTS times for overloaded errors"
4286 );
4287 });
4288
4289 // Check that a retry message was added
4290 thread.read_with(cx, |thread, _| {
4291 let mut messages = thread.messages();
4292 assert!(
4293 messages.any(|msg| {
4294 msg.role == Role::System
4295 && msg.ui_only
4296 && msg.segments.iter().any(|seg| {
4297 if let MessageSegment::Text(text) = seg {
4298 text.contains("overloaded")
4299 && text
4300 .contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS))
4301 } else {
4302 false
4303 }
4304 })
4305 }),
4306 "Should have added a system retry message"
4307 );
4308 });
4309
4310 let retry_count = thread.update(cx, |thread, _| {
4311 thread
4312 .messages
4313 .iter()
4314 .filter(|m| {
4315 m.ui_only
4316 && m.segments.iter().any(|s| {
4317 if let MessageSegment::Text(text) = s {
4318 text.contains("Retrying") && text.contains("seconds")
4319 } else {
4320 false
4321 }
4322 })
4323 })
4324 .count()
4325 });
4326
4327 assert_eq!(retry_count, 1, "Should have one retry message");
4328 }
4329
4330 #[gpui::test]
4331 async fn test_retry_on_internal_server_error(cx: &mut TestAppContext) {
4332 init_test_settings(cx);
4333
4334 let project = create_test_project(cx, json!({})).await;
4335 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4336
4337 // Enable Burn Mode to allow retries
4338 thread.update(cx, |thread, _| {
4339 thread.set_completion_mode(CompletionMode::Burn);
4340 });
4341
4342 // Create model that returns internal server error
4343 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4344
4345 // Insert a user message
4346 thread.update(cx, |thread, cx| {
4347 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4348 });
4349
4350 // Start completion
4351 thread.update(cx, |thread, cx| {
4352 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4353 });
4354
4355 cx.run_until_parked();
4356
4357 // Check retry state on thread
4358 thread.read_with(cx, |thread, _| {
4359 assert!(thread.retry_state.is_some(), "Should have retry state");
4360 let retry_state = thread.retry_state.as_ref().unwrap();
4361 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4362 assert_eq!(
4363 retry_state.max_attempts, 3,
4364 "Should have correct max attempts"
4365 );
4366 });
4367
4368 // Check that a retry message was added with provider name
4369 thread.read_with(cx, |thread, _| {
4370 let mut messages = thread.messages();
4371 assert!(
4372 messages.any(|msg| {
4373 msg.role == Role::System
4374 && msg.ui_only
4375 && msg.segments.iter().any(|seg| {
4376 if let MessageSegment::Text(text) = seg {
4377 text.contains("internal")
4378 && text.contains("Fake")
4379 && text.contains("Retrying")
4380 && text.contains("attempt 1 of 3")
4381 && text.contains("seconds")
4382 } else {
4383 false
4384 }
4385 })
4386 }),
4387 "Should have added a system retry message with provider name"
4388 );
4389 });
4390
4391 // Count retry messages
4392 let retry_count = thread.update(cx, |thread, _| {
4393 thread
4394 .messages
4395 .iter()
4396 .filter(|m| {
4397 m.ui_only
4398 && m.segments.iter().any(|s| {
4399 if let MessageSegment::Text(text) = s {
4400 text.contains("Retrying") && text.contains("seconds")
4401 } else {
4402 false
4403 }
4404 })
4405 })
4406 .count()
4407 });
4408
4409 assert_eq!(retry_count, 1, "Should have one retry message");
4410 }
4411
4412 #[gpui::test]
4413 async fn test_exponential_backoff_on_retries(cx: &mut TestAppContext) {
4414 init_test_settings(cx);
4415
4416 let project = create_test_project(cx, json!({})).await;
4417 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4418
4419 // Enable Burn Mode to allow retries
4420 thread.update(cx, |thread, _| {
4421 thread.set_completion_mode(CompletionMode::Burn);
4422 });
4423
4424 // Create model that returns internal server error
4425 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4426
4427 // Insert a user message
4428 thread.update(cx, |thread, cx| {
4429 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4430 });
4431
4432 // Track retry events and completion count
4433 // Track completion events
4434 let completion_count = Arc::new(Mutex::new(0));
4435 let completion_count_clone = completion_count.clone();
4436
4437 let _subscription = thread.update(cx, |_, cx| {
4438 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4439 if let ThreadEvent::NewRequest = event {
4440 *completion_count_clone.lock() += 1;
4441 }
4442 })
4443 });
4444
4445 // First attempt
4446 thread.update(cx, |thread, cx| {
4447 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4448 });
4449 cx.run_until_parked();
4450
4451 // Should have scheduled first retry - count retry messages
4452 let retry_count = thread.update(cx, |thread, _| {
4453 thread
4454 .messages
4455 .iter()
4456 .filter(|m| {
4457 m.ui_only
4458 && m.segments.iter().any(|s| {
4459 if let MessageSegment::Text(text) = s {
4460 text.contains("Retrying") && text.contains("seconds")
4461 } else {
4462 false
4463 }
4464 })
4465 })
4466 .count()
4467 });
4468 assert_eq!(retry_count, 1, "Should have scheduled first retry");
4469
4470 // Check retry state
4471 thread.read_with(cx, |thread, _| {
4472 assert!(thread.retry_state.is_some(), "Should have retry state");
4473 let retry_state = thread.retry_state.as_ref().unwrap();
4474 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4475 assert_eq!(
4476 retry_state.max_attempts, 3,
4477 "Internal server errors should retry up to 3 times"
4478 );
4479 });
4480
4481 // Advance clock for first retry
4482 cx.executor().advance_clock(BASE_RETRY_DELAY);
4483 cx.run_until_parked();
4484
4485 // Advance clock for second retry
4486 cx.executor().advance_clock(BASE_RETRY_DELAY);
4487 cx.run_until_parked();
4488
4489 // Advance clock for third retry
4490 cx.executor().advance_clock(BASE_RETRY_DELAY);
4491 cx.run_until_parked();
4492
4493 // Should have completed all retries - count retry messages
4494 let retry_count = thread.update(cx, |thread, _| {
4495 thread
4496 .messages
4497 .iter()
4498 .filter(|m| {
4499 m.ui_only
4500 && m.segments.iter().any(|s| {
4501 if let MessageSegment::Text(text) = s {
4502 text.contains("Retrying") && text.contains("seconds")
4503 } else {
4504 false
4505 }
4506 })
4507 })
4508 .count()
4509 });
4510 assert_eq!(
4511 retry_count, 3,
4512 "Should have 3 retries for internal server errors"
4513 );
4514
4515 // For internal server errors, we retry 3 times and then give up
4516 // Check that retry_state is cleared after all retries
4517 thread.read_with(cx, |thread, _| {
4518 assert!(
4519 thread.retry_state.is_none(),
4520 "Retry state should be cleared after all retries"
4521 );
4522 });
4523
4524 // Verify total attempts (1 initial + 3 retries)
4525 assert_eq!(
4526 *completion_count.lock(),
4527 4,
4528 "Should have attempted once plus 3 retries"
4529 );
4530 }
4531
4532 #[gpui::test]
4533 async fn test_max_retries_exceeded(cx: &mut TestAppContext) {
4534 init_test_settings(cx);
4535
4536 let project = create_test_project(cx, json!({})).await;
4537 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4538
4539 // Enable Burn Mode to allow retries
4540 thread.update(cx, |thread, _| {
4541 thread.set_completion_mode(CompletionMode::Burn);
4542 });
4543
4544 // Create model that returns overloaded error
4545 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4546
4547 // Insert a user message
4548 thread.update(cx, |thread, cx| {
4549 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4550 });
4551
4552 // Track events
4553 let stopped_with_error = Arc::new(Mutex::new(false));
4554 let stopped_with_error_clone = stopped_with_error.clone();
4555
4556 let _subscription = thread.update(cx, |_, cx| {
4557 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4558 if let ThreadEvent::Stopped(Err(_)) = event {
4559 *stopped_with_error_clone.lock() = true;
4560 }
4561 })
4562 });
4563
4564 // Start initial completion
4565 thread.update(cx, |thread, cx| {
4566 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4567 });
4568 cx.run_until_parked();
4569
4570 // Advance through all retries
4571 for _ in 0..MAX_RETRY_ATTEMPTS {
4572 cx.executor().advance_clock(BASE_RETRY_DELAY);
4573 cx.run_until_parked();
4574 }
4575
4576 let retry_count = thread.update(cx, |thread, _| {
4577 thread
4578 .messages
4579 .iter()
4580 .filter(|m| {
4581 m.ui_only
4582 && m.segments.iter().any(|s| {
4583 if let MessageSegment::Text(text) = s {
4584 text.contains("Retrying") && text.contains("seconds")
4585 } else {
4586 false
4587 }
4588 })
4589 })
4590 .count()
4591 });
4592
4593 // After max retries, should emit Stopped(Err(...)) event
4594 assert_eq!(
4595 retry_count, MAX_RETRY_ATTEMPTS as usize,
4596 "Should have attempted MAX_RETRY_ATTEMPTS retries for overloaded errors"
4597 );
4598 assert!(
4599 *stopped_with_error.lock(),
4600 "Should emit Stopped(Err(...)) event after max retries exceeded"
4601 );
4602
4603 // Retry state should be cleared
4604 thread.read_with(cx, |thread, _| {
4605 assert!(
4606 thread.retry_state.is_none(),
4607 "Retry state should be cleared after max retries"
4608 );
4609
4610 // Verify we have the expected number of retry messages
4611 let retry_messages = thread
4612 .messages
4613 .iter()
4614 .filter(|msg| msg.ui_only && msg.role == Role::System)
4615 .count();
4616 assert_eq!(
4617 retry_messages, MAX_RETRY_ATTEMPTS as usize,
4618 "Should have MAX_RETRY_ATTEMPTS retry messages for overloaded errors"
4619 );
4620 });
4621 }
4622
4623 #[gpui::test]
4624 async fn test_retry_message_removed_on_retry(cx: &mut TestAppContext) {
4625 init_test_settings(cx);
4626
4627 let project = create_test_project(cx, json!({})).await;
4628 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4629
4630 // Enable Burn Mode to allow retries
4631 thread.update(cx, |thread, _| {
4632 thread.set_completion_mode(CompletionMode::Burn);
4633 });
4634
4635 // We'll use a wrapper to switch behavior after first failure
4636 struct RetryTestModel {
4637 inner: Arc<FakeLanguageModel>,
4638 failed_once: Arc<Mutex<bool>>,
4639 }
4640
4641 impl LanguageModel for RetryTestModel {
4642 fn id(&self) -> LanguageModelId {
4643 self.inner.id()
4644 }
4645
4646 fn name(&self) -> LanguageModelName {
4647 self.inner.name()
4648 }
4649
4650 fn provider_id(&self) -> LanguageModelProviderId {
4651 self.inner.provider_id()
4652 }
4653
4654 fn provider_name(&self) -> LanguageModelProviderName {
4655 self.inner.provider_name()
4656 }
4657
4658 fn supports_tools(&self) -> bool {
4659 self.inner.supports_tools()
4660 }
4661
4662 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4663 self.inner.supports_tool_choice(choice)
4664 }
4665
4666 fn supports_images(&self) -> bool {
4667 self.inner.supports_images()
4668 }
4669
4670 fn telemetry_id(&self) -> String {
4671 self.inner.telemetry_id()
4672 }
4673
4674 fn max_token_count(&self) -> u64 {
4675 self.inner.max_token_count()
4676 }
4677
4678 fn count_tokens(
4679 &self,
4680 request: LanguageModelRequest,
4681 cx: &App,
4682 ) -> BoxFuture<'static, Result<u64>> {
4683 self.inner.count_tokens(request, cx)
4684 }
4685
4686 fn stream_completion(
4687 &self,
4688 request: LanguageModelRequest,
4689 cx: &AsyncApp,
4690 ) -> BoxFuture<
4691 'static,
4692 Result<
4693 BoxStream<
4694 'static,
4695 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4696 >,
4697 LanguageModelCompletionError,
4698 >,
4699 > {
4700 if !*self.failed_once.lock() {
4701 *self.failed_once.lock() = true;
4702 let provider = self.provider_name();
4703 // Return error on first attempt
4704 let stream = futures::stream::once(async move {
4705 Err(LanguageModelCompletionError::ServerOverloaded {
4706 provider,
4707 retry_after: None,
4708 })
4709 });
4710 async move { Ok(stream.boxed()) }.boxed()
4711 } else {
4712 // Succeed on retry
4713 self.inner.stream_completion(request, cx)
4714 }
4715 }
4716
4717 fn as_fake(&self) -> &FakeLanguageModel {
4718 &self.inner
4719 }
4720 }
4721
4722 let model = Arc::new(RetryTestModel {
4723 inner: Arc::new(FakeLanguageModel::default()),
4724 failed_once: Arc::new(Mutex::new(false)),
4725 });
4726
4727 // Insert a user message
4728 thread.update(cx, |thread, cx| {
4729 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4730 });
4731
4732 // Track message deletions
4733 // Track when retry completes successfully
4734 let retry_completed = Arc::new(Mutex::new(false));
4735 let retry_completed_clone = retry_completed.clone();
4736
4737 let _subscription = thread.update(cx, |_, cx| {
4738 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4739 if let ThreadEvent::StreamedCompletion = event {
4740 *retry_completed_clone.lock() = true;
4741 }
4742 })
4743 });
4744
4745 // Start completion
4746 thread.update(cx, |thread, cx| {
4747 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4748 });
4749 cx.run_until_parked();
4750
4751 // Get the retry message ID
4752 let retry_message_id = thread.read_with(cx, |thread, _| {
4753 thread
4754 .messages()
4755 .find(|msg| msg.role == Role::System && msg.ui_only)
4756 .map(|msg| msg.id)
4757 .expect("Should have a retry message")
4758 });
4759
4760 // Wait for retry
4761 cx.executor().advance_clock(BASE_RETRY_DELAY);
4762 cx.run_until_parked();
4763
4764 // Stream some successful content
4765 let fake_model = model.as_fake();
4766 // After the retry, there should be a new pending completion
4767 let pending = fake_model.pending_completions();
4768 assert!(
4769 !pending.is_empty(),
4770 "Should have a pending completion after retry"
4771 );
4772 fake_model.stream_completion_response(&pending[0], "Success!");
4773 fake_model.end_completion_stream(&pending[0]);
4774 cx.run_until_parked();
4775
4776 // Check that the retry completed successfully
4777 assert!(
4778 *retry_completed.lock(),
4779 "Retry should have completed successfully"
4780 );
4781
4782 // Retry message should still exist but be marked as ui_only
4783 thread.read_with(cx, |thread, _| {
4784 let retry_msg = thread
4785 .message(retry_message_id)
4786 .expect("Retry message should still exist");
4787 assert!(retry_msg.ui_only, "Retry message should be ui_only");
4788 assert_eq!(
4789 retry_msg.role,
4790 Role::System,
4791 "Retry message should have System role"
4792 );
4793 });
4794 }
4795
4796 #[gpui::test]
4797 async fn test_successful_completion_clears_retry_state(cx: &mut TestAppContext) {
4798 init_test_settings(cx);
4799
4800 let project = create_test_project(cx, json!({})).await;
4801 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4802
4803 // Enable Burn Mode to allow retries
4804 thread.update(cx, |thread, _| {
4805 thread.set_completion_mode(CompletionMode::Burn);
4806 });
4807
4808 // Create a model that fails once then succeeds
4809 struct FailOnceModel {
4810 inner: Arc<FakeLanguageModel>,
4811 failed_once: Arc<Mutex<bool>>,
4812 }
4813
4814 impl LanguageModel for FailOnceModel {
4815 fn id(&self) -> LanguageModelId {
4816 self.inner.id()
4817 }
4818
4819 fn name(&self) -> LanguageModelName {
4820 self.inner.name()
4821 }
4822
4823 fn provider_id(&self) -> LanguageModelProviderId {
4824 self.inner.provider_id()
4825 }
4826
4827 fn provider_name(&self) -> LanguageModelProviderName {
4828 self.inner.provider_name()
4829 }
4830
4831 fn supports_tools(&self) -> bool {
4832 self.inner.supports_tools()
4833 }
4834
4835 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4836 self.inner.supports_tool_choice(choice)
4837 }
4838
4839 fn supports_images(&self) -> bool {
4840 self.inner.supports_images()
4841 }
4842
4843 fn telemetry_id(&self) -> String {
4844 self.inner.telemetry_id()
4845 }
4846
4847 fn max_token_count(&self) -> u64 {
4848 self.inner.max_token_count()
4849 }
4850
4851 fn count_tokens(
4852 &self,
4853 request: LanguageModelRequest,
4854 cx: &App,
4855 ) -> BoxFuture<'static, Result<u64>> {
4856 self.inner.count_tokens(request, cx)
4857 }
4858
4859 fn stream_completion(
4860 &self,
4861 request: LanguageModelRequest,
4862 cx: &AsyncApp,
4863 ) -> BoxFuture<
4864 'static,
4865 Result<
4866 BoxStream<
4867 'static,
4868 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4869 >,
4870 LanguageModelCompletionError,
4871 >,
4872 > {
4873 if !*self.failed_once.lock() {
4874 *self.failed_once.lock() = true;
4875 let provider = self.provider_name();
4876 // Return error on first attempt
4877 let stream = futures::stream::once(async move {
4878 Err(LanguageModelCompletionError::ServerOverloaded {
4879 provider,
4880 retry_after: None,
4881 })
4882 });
4883 async move { Ok(stream.boxed()) }.boxed()
4884 } else {
4885 // Succeed on retry
4886 self.inner.stream_completion(request, cx)
4887 }
4888 }
4889 }
4890
4891 let fail_once_model = Arc::new(FailOnceModel {
4892 inner: Arc::new(FakeLanguageModel::default()),
4893 failed_once: Arc::new(Mutex::new(false)),
4894 });
4895
4896 // Insert a user message
4897 thread.update(cx, |thread, cx| {
4898 thread.insert_user_message(
4899 "Test message",
4900 ContextLoadResult::default(),
4901 None,
4902 vec![],
4903 cx,
4904 );
4905 });
4906
4907 // Start completion with fail-once model
4908 thread.update(cx, |thread, cx| {
4909 thread.send_to_model(
4910 fail_once_model.clone(),
4911 CompletionIntent::UserPrompt,
4912 None,
4913 cx,
4914 );
4915 });
4916
4917 cx.run_until_parked();
4918
4919 // Verify retry state exists after first failure
4920 thread.read_with(cx, |thread, _| {
4921 assert!(
4922 thread.retry_state.is_some(),
4923 "Should have retry state after failure"
4924 );
4925 });
4926
4927 // Wait for retry delay
4928 cx.executor().advance_clock(BASE_RETRY_DELAY);
4929 cx.run_until_parked();
4930
4931 // The retry should now use our FailOnceModel which should succeed
4932 // We need to help the FakeLanguageModel complete the stream
4933 let inner_fake = fail_once_model.inner.clone();
4934
4935 // Wait a bit for the retry to start
4936 cx.run_until_parked();
4937
4938 // Check for pending completions and complete them
4939 if let Some(pending) = inner_fake.pending_completions().first() {
4940 inner_fake.stream_completion_response(pending, "Success!");
4941 inner_fake.end_completion_stream(pending);
4942 }
4943 cx.run_until_parked();
4944
4945 thread.read_with(cx, |thread, _| {
4946 assert!(
4947 thread.retry_state.is_none(),
4948 "Retry state should be cleared after successful completion"
4949 );
4950
4951 let has_assistant_message = thread
4952 .messages
4953 .iter()
4954 .any(|msg| msg.role == Role::Assistant && !msg.ui_only);
4955 assert!(
4956 has_assistant_message,
4957 "Should have an assistant message after successful retry"
4958 );
4959 });
4960 }
4961
4962 #[gpui::test]
4963 async fn test_rate_limit_retry_single_attempt(cx: &mut TestAppContext) {
4964 init_test_settings(cx);
4965
4966 let project = create_test_project(cx, json!({})).await;
4967 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4968
4969 // Enable Burn Mode to allow retries
4970 thread.update(cx, |thread, _| {
4971 thread.set_completion_mode(CompletionMode::Burn);
4972 });
4973
4974 // Create a model that returns rate limit error with retry_after
4975 struct RateLimitModel {
4976 inner: Arc<FakeLanguageModel>,
4977 }
4978
4979 impl LanguageModel for RateLimitModel {
4980 fn id(&self) -> LanguageModelId {
4981 self.inner.id()
4982 }
4983
4984 fn name(&self) -> LanguageModelName {
4985 self.inner.name()
4986 }
4987
4988 fn provider_id(&self) -> LanguageModelProviderId {
4989 self.inner.provider_id()
4990 }
4991
4992 fn provider_name(&self) -> LanguageModelProviderName {
4993 self.inner.provider_name()
4994 }
4995
4996 fn supports_tools(&self) -> bool {
4997 self.inner.supports_tools()
4998 }
4999
5000 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
5001 self.inner.supports_tool_choice(choice)
5002 }
5003
5004 fn supports_images(&self) -> bool {
5005 self.inner.supports_images()
5006 }
5007
5008 fn telemetry_id(&self) -> String {
5009 self.inner.telemetry_id()
5010 }
5011
5012 fn max_token_count(&self) -> u64 {
5013 self.inner.max_token_count()
5014 }
5015
5016 fn count_tokens(
5017 &self,
5018 request: LanguageModelRequest,
5019 cx: &App,
5020 ) -> BoxFuture<'static, Result<u64>> {
5021 self.inner.count_tokens(request, cx)
5022 }
5023
5024 fn stream_completion(
5025 &self,
5026 _request: LanguageModelRequest,
5027 _cx: &AsyncApp,
5028 ) -> BoxFuture<
5029 'static,
5030 Result<
5031 BoxStream<
5032 'static,
5033 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
5034 >,
5035 LanguageModelCompletionError,
5036 >,
5037 > {
5038 let provider = self.provider_name();
5039 async move {
5040 let stream = futures::stream::once(async move {
5041 Err(LanguageModelCompletionError::RateLimitExceeded {
5042 provider,
5043 retry_after: Some(Duration::from_secs(TEST_RATE_LIMIT_RETRY_SECS)),
5044 })
5045 });
5046 Ok(stream.boxed())
5047 }
5048 .boxed()
5049 }
5050
5051 fn as_fake(&self) -> &FakeLanguageModel {
5052 &self.inner
5053 }
5054 }
5055
5056 let model = Arc::new(RateLimitModel {
5057 inner: Arc::new(FakeLanguageModel::default()),
5058 });
5059
5060 // Insert a user message
5061 thread.update(cx, |thread, cx| {
5062 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5063 });
5064
5065 // Start completion
5066 thread.update(cx, |thread, cx| {
5067 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5068 });
5069
5070 cx.run_until_parked();
5071
5072 let retry_count = thread.update(cx, |thread, _| {
5073 thread
5074 .messages
5075 .iter()
5076 .filter(|m| {
5077 m.ui_only
5078 && m.segments.iter().any(|s| {
5079 if let MessageSegment::Text(text) = s {
5080 text.contains("rate limit exceeded")
5081 } else {
5082 false
5083 }
5084 })
5085 })
5086 .count()
5087 });
5088 assert_eq!(retry_count, 1, "Should have scheduled one retry");
5089
5090 thread.read_with(cx, |thread, _| {
5091 assert!(
5092 thread.retry_state.is_some(),
5093 "Rate limit errors should set retry_state"
5094 );
5095 if let Some(retry_state) = &thread.retry_state {
5096 assert_eq!(
5097 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
5098 "Rate limit errors should use MAX_RETRY_ATTEMPTS"
5099 );
5100 }
5101 });
5102
5103 // Verify we have one retry message
5104 thread.read_with(cx, |thread, _| {
5105 let retry_messages = thread
5106 .messages
5107 .iter()
5108 .filter(|msg| {
5109 msg.ui_only
5110 && msg.segments.iter().any(|seg| {
5111 if let MessageSegment::Text(text) = seg {
5112 text.contains("rate limit exceeded")
5113 } else {
5114 false
5115 }
5116 })
5117 })
5118 .count();
5119 assert_eq!(
5120 retry_messages, 1,
5121 "Should have one rate limit retry message"
5122 );
5123 });
5124
5125 // Check that retry message doesn't include attempt count
5126 thread.read_with(cx, |thread, _| {
5127 let retry_message = thread
5128 .messages
5129 .iter()
5130 .find(|msg| msg.role == Role::System && msg.ui_only)
5131 .expect("Should have a retry message");
5132
5133 // Check that the message contains attempt count since we use retry_state
5134 if let Some(MessageSegment::Text(text)) = retry_message.segments.first() {
5135 assert!(
5136 text.contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS)),
5137 "Rate limit retry message should contain attempt count with MAX_RETRY_ATTEMPTS"
5138 );
5139 assert!(
5140 text.contains("Retrying"),
5141 "Rate limit retry message should contain retry text"
5142 );
5143 }
5144 });
5145 }
5146
5147 #[gpui::test]
5148 async fn test_ui_only_messages_not_sent_to_model(cx: &mut TestAppContext) {
5149 init_test_settings(cx);
5150
5151 let project = create_test_project(cx, json!({})).await;
5152 let (_, _, thread, _, model) = setup_test_environment(cx, project.clone()).await;
5153
5154 // Insert a regular user message
5155 thread.update(cx, |thread, cx| {
5156 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5157 });
5158
5159 // Insert a UI-only message (like our retry notifications)
5160 thread.update(cx, |thread, cx| {
5161 let id = thread.next_message_id.post_inc();
5162 thread.messages.push(Message {
5163 id,
5164 role: Role::System,
5165 segments: vec![MessageSegment::Text(
5166 "This is a UI-only message that should not be sent to the model".to_string(),
5167 )],
5168 loaded_context: LoadedContext::default(),
5169 creases: Vec::new(),
5170 is_hidden: true,
5171 ui_only: true,
5172 });
5173 cx.emit(ThreadEvent::MessageAdded(id));
5174 });
5175
5176 // Insert another regular message
5177 thread.update(cx, |thread, cx| {
5178 thread.insert_user_message(
5179 "How are you?",
5180 ContextLoadResult::default(),
5181 None,
5182 vec![],
5183 cx,
5184 );
5185 });
5186
5187 // Generate the completion request
5188 let request = thread.update(cx, |thread, cx| {
5189 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
5190 });
5191
5192 // Verify that the request only contains non-UI-only messages
5193 // Should have system prompt + 2 user messages, but not the UI-only message
5194 let user_messages: Vec<_> = request
5195 .messages
5196 .iter()
5197 .filter(|msg| msg.role == Role::User)
5198 .collect();
5199 assert_eq!(
5200 user_messages.len(),
5201 2,
5202 "Should have exactly 2 user messages"
5203 );
5204
5205 // Verify the UI-only content is not present anywhere in the request
5206 let request_text = request
5207 .messages
5208 .iter()
5209 .flat_map(|msg| &msg.content)
5210 .filter_map(|content| match content {
5211 MessageContent::Text(text) => Some(text.as_str()),
5212 _ => None,
5213 })
5214 .collect::<String>();
5215
5216 assert!(
5217 !request_text.contains("UI-only message"),
5218 "UI-only message content should not be in the request"
5219 );
5220
5221 // Verify the thread still has all 3 messages (including UI-only)
5222 thread.read_with(cx, |thread, _| {
5223 assert_eq!(
5224 thread.messages().count(),
5225 3,
5226 "Thread should have 3 messages"
5227 );
5228 assert_eq!(
5229 thread.messages().filter(|m| m.ui_only).count(),
5230 1,
5231 "Thread should have 1 UI-only message"
5232 );
5233 });
5234
5235 // Verify that UI-only messages are not serialized
5236 let serialized = thread
5237 .update(cx, |thread, cx| thread.serialize(cx))
5238 .await
5239 .unwrap();
5240 assert_eq!(
5241 serialized.messages.len(),
5242 2,
5243 "Serialized thread should only have 2 messages (no UI-only)"
5244 );
5245 }
5246
5247 #[gpui::test]
5248 async fn test_no_retry_without_burn_mode(cx: &mut TestAppContext) {
5249 init_test_settings(cx);
5250
5251 let project = create_test_project(cx, json!({})).await;
5252 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5253
5254 // Ensure we're in Normal mode (not Burn mode)
5255 thread.update(cx, |thread, _| {
5256 thread.set_completion_mode(CompletionMode::Normal);
5257 });
5258
5259 // Track error events
5260 let error_events = Arc::new(Mutex::new(Vec::new()));
5261 let error_events_clone = error_events.clone();
5262
5263 let _subscription = thread.update(cx, |_, cx| {
5264 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
5265 if let ThreadEvent::ShowError(error) = event {
5266 error_events_clone.lock().push(error.clone());
5267 }
5268 })
5269 });
5270
5271 // Create model that returns overloaded error
5272 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5273
5274 // Insert a user message
5275 thread.update(cx, |thread, cx| {
5276 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5277 });
5278
5279 // Start completion
5280 thread.update(cx, |thread, cx| {
5281 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5282 });
5283
5284 cx.run_until_parked();
5285
5286 // Verify no retry state was created
5287 thread.read_with(cx, |thread, _| {
5288 assert!(
5289 thread.retry_state.is_none(),
5290 "Should not have retry state in Normal mode"
5291 );
5292 });
5293
5294 // Check that a retryable error was reported
5295 let errors = error_events.lock();
5296 assert!(!errors.is_empty(), "Should have received an error event");
5297
5298 if let ThreadError::RetryableError {
5299 message: _,
5300 can_enable_burn_mode,
5301 } = &errors[0]
5302 {
5303 assert!(
5304 *can_enable_burn_mode,
5305 "Error should indicate burn mode can be enabled"
5306 );
5307 } else {
5308 panic!("Expected RetryableError, got {:?}", errors[0]);
5309 }
5310
5311 // Verify the thread is no longer generating
5312 thread.read_with(cx, |thread, _| {
5313 assert!(
5314 !thread.is_generating(),
5315 "Should not be generating after error without retry"
5316 );
5317 });
5318 }
5319
5320 #[gpui::test]
5321 async fn test_retry_cancelled_on_stop(cx: &mut TestAppContext) {
5322 init_test_settings(cx);
5323
5324 let project = create_test_project(cx, json!({})).await;
5325 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5326
5327 // Enable Burn Mode to allow retries
5328 thread.update(cx, |thread, _| {
5329 thread.set_completion_mode(CompletionMode::Burn);
5330 });
5331
5332 // Create model that returns overloaded error
5333 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5334
5335 // Insert a user message
5336 thread.update(cx, |thread, cx| {
5337 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5338 });
5339
5340 // Start completion
5341 thread.update(cx, |thread, cx| {
5342 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5343 });
5344
5345 cx.run_until_parked();
5346
5347 // Verify retry was scheduled by checking for retry message
5348 let has_retry_message = thread.read_with(cx, |thread, _| {
5349 thread.messages.iter().any(|m| {
5350 m.ui_only
5351 && m.segments.iter().any(|s| {
5352 if let MessageSegment::Text(text) = s {
5353 text.contains("Retrying") && text.contains("seconds")
5354 } else {
5355 false
5356 }
5357 })
5358 })
5359 });
5360 assert!(has_retry_message, "Should have scheduled a retry");
5361
5362 // Cancel the completion before the retry happens
5363 thread.update(cx, |thread, cx| {
5364 thread.cancel_last_completion(None, cx);
5365 });
5366
5367 cx.run_until_parked();
5368
5369 // The retry should not have happened - no pending completions
5370 let fake_model = model.as_fake();
5371 assert_eq!(
5372 fake_model.pending_completions().len(),
5373 0,
5374 "Should have no pending completions after cancellation"
5375 );
5376
5377 // Verify the retry was cancelled by checking retry state
5378 thread.read_with(cx, |thread, _| {
5379 if let Some(retry_state) = &thread.retry_state {
5380 panic!(
5381 "retry_state should be cleared after cancellation, but found: attempt={}, max_attempts={}, intent={:?}",
5382 retry_state.attempt, retry_state.max_attempts, retry_state.intent
5383 );
5384 }
5385 });
5386 }
5387
5388 fn test_summarize_error(
5389 model: &Arc<dyn LanguageModel>,
5390 thread: &Entity<Thread>,
5391 cx: &mut TestAppContext,
5392 ) {
5393 thread.update(cx, |thread, cx| {
5394 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
5395 thread.send_to_model(
5396 model.clone(),
5397 CompletionIntent::ThreadSummarization,
5398 None,
5399 cx,
5400 );
5401 });
5402
5403 let fake_model = model.as_fake();
5404 simulate_successful_response(&fake_model, cx);
5405
5406 thread.read_with(cx, |thread, _| {
5407 assert!(matches!(thread.summary(), ThreadSummary::Generating));
5408 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5409 });
5410
5411 // Simulate summary request ending
5412 cx.run_until_parked();
5413 fake_model.end_last_completion_stream();
5414 cx.run_until_parked();
5415
5416 // State is set to Error and default message
5417 thread.read_with(cx, |thread, _| {
5418 assert!(matches!(thread.summary(), ThreadSummary::Error));
5419 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5420 });
5421 }
5422
5423 fn simulate_successful_response(fake_model: &FakeLanguageModel, cx: &mut TestAppContext) {
5424 cx.run_until_parked();
5425 fake_model.stream_last_completion_response("Assistant response");
5426 fake_model.end_last_completion_stream();
5427 cx.run_until_parked();
5428 }
5429
5430 fn init_test_settings(cx: &mut TestAppContext) {
5431 cx.update(|cx| {
5432 let settings_store = SettingsStore::test(cx);
5433 cx.set_global(settings_store);
5434 language::init(cx);
5435 Project::init_settings(cx);
5436 AgentSettings::register(cx);
5437 prompt_store::init(cx);
5438 thread_store::init(cx);
5439 workspace::init_settings(cx);
5440 language_model::init_settings(cx);
5441 ThemeSettings::register(cx);
5442 ToolRegistry::default_global(cx);
5443 assistant_tool::init(cx);
5444
5445 let http_client = Arc::new(http_client::HttpClientWithUrl::new(
5446 http_client::FakeHttpClient::with_200_response(),
5447 "http://localhost".to_string(),
5448 None,
5449 ));
5450 assistant_tools::init(http_client, cx);
5451 });
5452 }
5453
5454 // Helper to create a test project with test files
5455 async fn create_test_project(
5456 cx: &mut TestAppContext,
5457 files: serde_json::Value,
5458 ) -> Entity<Project> {
5459 let fs = FakeFs::new(cx.executor());
5460 fs.insert_tree(path!("/test"), files).await;
5461 Project::test(fs, [path!("/test").as_ref()], cx).await
5462 }
5463
5464 async fn setup_test_environment(
5465 cx: &mut TestAppContext,
5466 project: Entity<Project>,
5467 ) -> (
5468 Entity<Workspace>,
5469 Entity<ThreadStore>,
5470 Entity<Thread>,
5471 Entity<ContextStore>,
5472 Arc<dyn LanguageModel>,
5473 ) {
5474 let (workspace, cx) =
5475 cx.add_window_view(|window, cx| Workspace::test_new(project.clone(), window, cx));
5476
5477 let thread_store = cx
5478 .update(|_, cx| {
5479 ThreadStore::load(
5480 project.clone(),
5481 cx.new(|_| ToolWorkingSet::default()),
5482 None,
5483 Arc::new(PromptBuilder::new(None).unwrap()),
5484 cx,
5485 )
5486 })
5487 .await
5488 .unwrap();
5489
5490 let thread = thread_store.update(cx, |store, cx| store.create_thread(cx));
5491 let context_store = cx.new(|_cx| ContextStore::new(project.downgrade(), None));
5492
5493 let provider = Arc::new(FakeLanguageModelProvider::default());
5494 let model = provider.test_model();
5495 let model: Arc<dyn LanguageModel> = Arc::new(model);
5496
5497 cx.update(|_, cx| {
5498 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
5499 registry.set_default_model(
5500 Some(ConfiguredModel {
5501 provider: provider.clone(),
5502 model: model.clone(),
5503 }),
5504 cx,
5505 );
5506 registry.set_thread_summary_model(
5507 Some(ConfiguredModel {
5508 provider,
5509 model: model.clone(),
5510 }),
5511 cx,
5512 );
5513 })
5514 });
5515
5516 (workspace, thread_store, thread, context_store, model)
5517 }
5518
5519 async fn add_file_to_context(
5520 project: &Entity<Project>,
5521 context_store: &Entity<ContextStore>,
5522 path: &str,
5523 cx: &mut TestAppContext,
5524 ) -> Result<Entity<language::Buffer>> {
5525 let buffer_path = project
5526 .read_with(cx, |project, cx| project.find_project_path(path, cx))
5527 .unwrap();
5528
5529 let buffer = project
5530 .update(cx, |project, cx| {
5531 project.open_buffer(buffer_path.clone(), cx)
5532 })
5533 .await
5534 .unwrap();
5535
5536 context_store.update(cx, |context_store, cx| {
5537 context_store.add_file_from_buffer(&buffer_path, buffer.clone(), false, cx);
5538 });
5539
5540 Ok(buffer)
5541 }
5542}