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 _, debug_panic, 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 let action_log = self.action_log.read(cx);
1586
1587 if !action_log.has_unnotified_user_edits() {
1588 return None;
1589 }
1590
1591 // Represent notification as a simulated `project_notifications` tool call
1592 let tool_name = Arc::from("project_notifications");
1593 let Some(tool) = self.tools.read(cx).tool(&tool_name, cx) else {
1594 debug_panic!("`project_notifications` tool not found");
1595 return None;
1596 };
1597
1598 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
1599 return None;
1600 }
1601
1602 let input = serde_json::json!({});
1603 let request = Arc::new(LanguageModelRequest::default()); // unused
1604 let window = None;
1605 let tool_result = tool.run(
1606 input,
1607 request,
1608 self.project.clone(),
1609 self.action_log.clone(),
1610 model.clone(),
1611 window,
1612 cx,
1613 );
1614
1615 let tool_use_id =
1616 LanguageModelToolUseId::from(format!("project_notifications_{}", self.messages.len()));
1617
1618 let tool_use = LanguageModelToolUse {
1619 id: tool_use_id.clone(),
1620 name: tool_name.clone(),
1621 raw_input: "{}".to_string(),
1622 input: serde_json::json!({}),
1623 is_input_complete: true,
1624 };
1625
1626 let tool_output = cx.background_executor().block(tool_result.output);
1627
1628 // Attach a project_notification tool call to the latest existing
1629 // Assistant message. We cannot create a new Assistant message
1630 // because thinking models require a `thinking` block that we
1631 // cannot mock. We cannot send a notification as a normal
1632 // (non-tool-use) User message because this distracts Agent
1633 // too much.
1634 let tool_message_id = self
1635 .messages
1636 .iter()
1637 .enumerate()
1638 .rfind(|(_, message)| message.role == Role::Assistant)
1639 .map(|(_, message)| message.id)?;
1640
1641 let tool_use_metadata = ToolUseMetadata {
1642 model: model.clone(),
1643 thread_id: self.id.clone(),
1644 prompt_id: self.last_prompt_id.clone(),
1645 };
1646
1647 self.tool_use
1648 .request_tool_use(tool_message_id, tool_use, tool_use_metadata.clone(), cx);
1649
1650 let pending_tool_use = self.tool_use.insert_tool_output(
1651 tool_use_id.clone(),
1652 tool_name,
1653 tool_output,
1654 self.configured_model.as_ref(),
1655 self.completion_mode,
1656 );
1657
1658 pending_tool_use
1659 }
1660
1661 pub fn stream_completion(
1662 &mut self,
1663 request: LanguageModelRequest,
1664 model: Arc<dyn LanguageModel>,
1665 intent: CompletionIntent,
1666 window: Option<AnyWindowHandle>,
1667 cx: &mut Context<Self>,
1668 ) {
1669 self.tool_use_limit_reached = false;
1670
1671 let pending_completion_id = post_inc(&mut self.completion_count);
1672 let mut request_callback_parameters = if self.request_callback.is_some() {
1673 Some((request.clone(), Vec::new()))
1674 } else {
1675 None
1676 };
1677 let prompt_id = self.last_prompt_id.clone();
1678 let tool_use_metadata = ToolUseMetadata {
1679 model: model.clone(),
1680 thread_id: self.id.clone(),
1681 prompt_id: prompt_id.clone(),
1682 };
1683
1684 let completion_mode = request
1685 .mode
1686 .unwrap_or(zed_llm_client::CompletionMode::Normal);
1687
1688 self.last_received_chunk_at = Some(Instant::now());
1689
1690 let task = cx.spawn(async move |thread, cx| {
1691 let stream_completion_future = model.stream_completion(request, &cx);
1692 let initial_token_usage =
1693 thread.read_with(cx, |thread, _cx| thread.cumulative_token_usage);
1694 let stream_completion = async {
1695 let mut events = stream_completion_future.await?;
1696
1697 let mut stop_reason = StopReason::EndTurn;
1698 let mut current_token_usage = TokenUsage::default();
1699
1700 thread
1701 .update(cx, |_thread, cx| {
1702 cx.emit(ThreadEvent::NewRequest);
1703 })
1704 .ok();
1705
1706 let mut request_assistant_message_id = None;
1707
1708 while let Some(event) = events.next().await {
1709 if let Some((_, response_events)) = request_callback_parameters.as_mut() {
1710 response_events
1711 .push(event.as_ref().map_err(|error| error.to_string()).cloned());
1712 }
1713
1714 thread.update(cx, |thread, cx| {
1715 match event? {
1716 LanguageModelCompletionEvent::StartMessage { .. } => {
1717 request_assistant_message_id =
1718 Some(thread.insert_assistant_message(
1719 vec![MessageSegment::Text(String::new())],
1720 cx,
1721 ));
1722 }
1723 LanguageModelCompletionEvent::Stop(reason) => {
1724 stop_reason = reason;
1725 }
1726 LanguageModelCompletionEvent::UsageUpdate(token_usage) => {
1727 thread.update_token_usage_at_last_message(token_usage);
1728 thread.cumulative_token_usage = thread.cumulative_token_usage
1729 + token_usage
1730 - current_token_usage;
1731 current_token_usage = token_usage;
1732 }
1733 LanguageModelCompletionEvent::Text(chunk) => {
1734 thread.received_chunk();
1735
1736 cx.emit(ThreadEvent::ReceivedTextChunk);
1737 if let Some(last_message) = thread.messages.last_mut() {
1738 if last_message.role == Role::Assistant
1739 && !thread.tool_use.has_tool_results(last_message.id)
1740 {
1741 last_message.push_text(&chunk);
1742 cx.emit(ThreadEvent::StreamedAssistantText(
1743 last_message.id,
1744 chunk,
1745 ));
1746 } else {
1747 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1748 // of a new Assistant response.
1749 //
1750 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1751 // will result in duplicating the text of the chunk in the rendered Markdown.
1752 request_assistant_message_id =
1753 Some(thread.insert_assistant_message(
1754 vec![MessageSegment::Text(chunk.to_string())],
1755 cx,
1756 ));
1757 };
1758 }
1759 }
1760 LanguageModelCompletionEvent::Thinking {
1761 text: chunk,
1762 signature,
1763 } => {
1764 thread.received_chunk();
1765
1766 if let Some(last_message) = thread.messages.last_mut() {
1767 if last_message.role == Role::Assistant
1768 && !thread.tool_use.has_tool_results(last_message.id)
1769 {
1770 last_message.push_thinking(&chunk, signature);
1771 cx.emit(ThreadEvent::StreamedAssistantThinking(
1772 last_message.id,
1773 chunk,
1774 ));
1775 } else {
1776 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1777 // of a new Assistant response.
1778 //
1779 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1780 // will result in duplicating the text of the chunk in the rendered Markdown.
1781 request_assistant_message_id =
1782 Some(thread.insert_assistant_message(
1783 vec![MessageSegment::Thinking {
1784 text: chunk.to_string(),
1785 signature,
1786 }],
1787 cx,
1788 ));
1789 };
1790 }
1791 }
1792 LanguageModelCompletionEvent::RedactedThinking { data } => {
1793 thread.received_chunk();
1794
1795 if let Some(last_message) = thread.messages.last_mut() {
1796 if last_message.role == Role::Assistant
1797 && !thread.tool_use.has_tool_results(last_message.id)
1798 {
1799 last_message.push_redacted_thinking(data);
1800 } else {
1801 request_assistant_message_id =
1802 Some(thread.insert_assistant_message(
1803 vec![MessageSegment::RedactedThinking(data)],
1804 cx,
1805 ));
1806 };
1807 }
1808 }
1809 LanguageModelCompletionEvent::ToolUse(tool_use) => {
1810 let last_assistant_message_id = request_assistant_message_id
1811 .unwrap_or_else(|| {
1812 let new_assistant_message_id =
1813 thread.insert_assistant_message(vec![], cx);
1814 request_assistant_message_id =
1815 Some(new_assistant_message_id);
1816 new_assistant_message_id
1817 });
1818
1819 let tool_use_id = tool_use.id.clone();
1820 let streamed_input = if tool_use.is_input_complete {
1821 None
1822 } else {
1823 Some((&tool_use.input).clone())
1824 };
1825
1826 let ui_text = thread.tool_use.request_tool_use(
1827 last_assistant_message_id,
1828 tool_use,
1829 tool_use_metadata.clone(),
1830 cx,
1831 );
1832
1833 if let Some(input) = streamed_input {
1834 cx.emit(ThreadEvent::StreamedToolUse {
1835 tool_use_id,
1836 ui_text,
1837 input,
1838 });
1839 }
1840 }
1841 LanguageModelCompletionEvent::ToolUseJsonParseError {
1842 id,
1843 tool_name,
1844 raw_input: invalid_input_json,
1845 json_parse_error,
1846 } => {
1847 thread.receive_invalid_tool_json(
1848 id,
1849 tool_name,
1850 invalid_input_json,
1851 json_parse_error,
1852 window,
1853 cx,
1854 );
1855 }
1856 LanguageModelCompletionEvent::StatusUpdate(status_update) => {
1857 if let Some(completion) = thread
1858 .pending_completions
1859 .iter_mut()
1860 .find(|completion| completion.id == pending_completion_id)
1861 {
1862 match status_update {
1863 CompletionRequestStatus::Queued { position } => {
1864 completion.queue_state =
1865 QueueState::Queued { position };
1866 }
1867 CompletionRequestStatus::Started => {
1868 completion.queue_state = QueueState::Started;
1869 }
1870 CompletionRequestStatus::Failed {
1871 code,
1872 message,
1873 request_id: _,
1874 retry_after,
1875 } => {
1876 return Err(
1877 LanguageModelCompletionError::from_cloud_failure(
1878 model.upstream_provider_name(),
1879 code,
1880 message,
1881 retry_after.map(Duration::from_secs_f64),
1882 ),
1883 );
1884 }
1885 CompletionRequestStatus::UsageUpdated { amount, limit } => {
1886 thread.update_model_request_usage(
1887 amount as u32,
1888 limit,
1889 cx,
1890 );
1891 }
1892 CompletionRequestStatus::ToolUseLimitReached => {
1893 thread.tool_use_limit_reached = true;
1894 cx.emit(ThreadEvent::ToolUseLimitReached);
1895 }
1896 }
1897 }
1898 }
1899 }
1900
1901 thread.touch_updated_at();
1902 cx.emit(ThreadEvent::StreamedCompletion);
1903 cx.notify();
1904
1905 thread.auto_capture_telemetry(cx);
1906 Ok(())
1907 })??;
1908
1909 smol::future::yield_now().await;
1910 }
1911
1912 thread.update(cx, |thread, cx| {
1913 thread.last_received_chunk_at = None;
1914 thread
1915 .pending_completions
1916 .retain(|completion| completion.id != pending_completion_id);
1917
1918 // If there is a response without tool use, summarize the message. Otherwise,
1919 // allow two tool uses before summarizing.
1920 if matches!(thread.summary, ThreadSummary::Pending)
1921 && thread.messages.len() >= 2
1922 && (!thread.has_pending_tool_uses() || thread.messages.len() >= 6)
1923 {
1924 thread.summarize(cx);
1925 }
1926 })?;
1927
1928 anyhow::Ok(stop_reason)
1929 };
1930
1931 let result = stream_completion.await;
1932 let mut retry_scheduled = false;
1933
1934 thread
1935 .update(cx, |thread, cx| {
1936 thread.finalize_pending_checkpoint(cx);
1937 match result.as_ref() {
1938 Ok(stop_reason) => {
1939 match stop_reason {
1940 StopReason::ToolUse => {
1941 let tool_uses =
1942 thread.use_pending_tools(window, model.clone(), cx);
1943 cx.emit(ThreadEvent::UsePendingTools { tool_uses });
1944 }
1945 StopReason::EndTurn | StopReason::MaxTokens => {
1946 thread.project.update(cx, |project, cx| {
1947 project.set_agent_location(None, cx);
1948 });
1949 }
1950 StopReason::Refusal => {
1951 thread.project.update(cx, |project, cx| {
1952 project.set_agent_location(None, cx);
1953 });
1954
1955 // Remove the turn that was refused.
1956 //
1957 // https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/handle-streaming-refusals#reset-context-after-refusal
1958 {
1959 let mut messages_to_remove = Vec::new();
1960
1961 for (ix, message) in
1962 thread.messages.iter().enumerate().rev()
1963 {
1964 messages_to_remove.push(message.id);
1965
1966 if message.role == Role::User {
1967 if ix == 0 {
1968 break;
1969 }
1970
1971 if let Some(prev_message) =
1972 thread.messages.get(ix - 1)
1973 {
1974 if prev_message.role == Role::Assistant {
1975 break;
1976 }
1977 }
1978 }
1979 }
1980
1981 for message_id in messages_to_remove {
1982 thread.delete_message(message_id, cx);
1983 }
1984 }
1985
1986 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1987 header: "Language model refusal".into(),
1988 message:
1989 "Model refused to generate content for safety reasons."
1990 .into(),
1991 }));
1992 }
1993 }
1994
1995 // We successfully completed, so cancel any remaining retries.
1996 thread.retry_state = None;
1997 }
1998 Err(error) => {
1999 thread.project.update(cx, |project, cx| {
2000 project.set_agent_location(None, cx);
2001 });
2002
2003 if error.is::<PaymentRequiredError>() {
2004 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
2005 } else if let Some(error) =
2006 error.downcast_ref::<ModelRequestLimitReachedError>()
2007 {
2008 cx.emit(ThreadEvent::ShowError(
2009 ThreadError::ModelRequestLimitReached { plan: error.plan },
2010 ));
2011 } else if let Some(completion_error) =
2012 error.downcast_ref::<LanguageModelCompletionError>()
2013 {
2014 match &completion_error {
2015 LanguageModelCompletionError::PromptTooLarge {
2016 tokens, ..
2017 } => {
2018 let tokens = tokens.unwrap_or_else(|| {
2019 // We didn't get an exact token count from the API, so fall back on our estimate.
2020 thread
2021 .total_token_usage()
2022 .map(|usage| usage.total)
2023 .unwrap_or(0)
2024 // We know the context window was exceeded in practice, so if our estimate was
2025 // lower than max tokens, the estimate was wrong; return that we exceeded by 1.
2026 .max(
2027 model
2028 .max_token_count_for_mode(completion_mode)
2029 .saturating_add(1),
2030 )
2031 });
2032 thread.exceeded_window_error = Some(ExceededWindowError {
2033 model_id: model.id(),
2034 token_count: tokens,
2035 });
2036 cx.notify();
2037 }
2038 _ => {
2039 if let Some(retry_strategy) =
2040 Thread::get_retry_strategy(completion_error)
2041 {
2042 retry_scheduled = thread
2043 .handle_retryable_error_with_delay(
2044 &completion_error,
2045 Some(retry_strategy),
2046 model.clone(),
2047 intent,
2048 window,
2049 cx,
2050 );
2051 }
2052 }
2053 }
2054 }
2055
2056 if !retry_scheduled {
2057 thread.cancel_last_completion(window, cx);
2058 }
2059 }
2060 }
2061
2062 if !retry_scheduled {
2063 cx.emit(ThreadEvent::Stopped(result.map_err(Arc::new)));
2064 }
2065
2066 if let Some((request_callback, (request, response_events))) = thread
2067 .request_callback
2068 .as_mut()
2069 .zip(request_callback_parameters.as_ref())
2070 {
2071 request_callback(request, response_events);
2072 }
2073
2074 thread.auto_capture_telemetry(cx);
2075
2076 if let Ok(initial_usage) = initial_token_usage {
2077 let usage = thread.cumulative_token_usage - initial_usage;
2078
2079 telemetry::event!(
2080 "Assistant Thread Completion",
2081 thread_id = thread.id().to_string(),
2082 prompt_id = prompt_id,
2083 model = model.telemetry_id(),
2084 model_provider = model.provider_id().to_string(),
2085 input_tokens = usage.input_tokens,
2086 output_tokens = usage.output_tokens,
2087 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2088 cache_read_input_tokens = usage.cache_read_input_tokens,
2089 );
2090 }
2091 })
2092 .ok();
2093 });
2094
2095 self.pending_completions.push(PendingCompletion {
2096 id: pending_completion_id,
2097 queue_state: QueueState::Sending,
2098 _task: task,
2099 });
2100 }
2101
2102 pub fn summarize(&mut self, cx: &mut Context<Self>) {
2103 let Some(model) = LanguageModelRegistry::read_global(cx).thread_summary_model() else {
2104 println!("No thread summary model");
2105 return;
2106 };
2107
2108 if !model.provider.is_authenticated(cx) {
2109 return;
2110 }
2111
2112 let added_user_message = include_str!("./prompts/summarize_thread_prompt.txt");
2113
2114 let request = self.to_summarize_request(
2115 &model.model,
2116 CompletionIntent::ThreadSummarization,
2117 added_user_message.into(),
2118 cx,
2119 );
2120
2121 self.summary = ThreadSummary::Generating;
2122
2123 self.pending_summary = cx.spawn(async move |this, cx| {
2124 let result = async {
2125 let mut messages = model.model.stream_completion(request, &cx).await?;
2126
2127 let mut new_summary = String::new();
2128 while let Some(event) = messages.next().await {
2129 let Ok(event) = event else {
2130 continue;
2131 };
2132 let text = match event {
2133 LanguageModelCompletionEvent::Text(text) => text,
2134 LanguageModelCompletionEvent::StatusUpdate(
2135 CompletionRequestStatus::UsageUpdated { amount, limit },
2136 ) => {
2137 this.update(cx, |thread, cx| {
2138 thread.update_model_request_usage(amount as u32, limit, cx);
2139 })?;
2140 continue;
2141 }
2142 _ => continue,
2143 };
2144
2145 let mut lines = text.lines();
2146 new_summary.extend(lines.next());
2147
2148 // Stop if the LLM generated multiple lines.
2149 if lines.next().is_some() {
2150 break;
2151 }
2152 }
2153
2154 anyhow::Ok(new_summary)
2155 }
2156 .await;
2157
2158 this.update(cx, |this, cx| {
2159 match result {
2160 Ok(new_summary) => {
2161 if new_summary.is_empty() {
2162 this.summary = ThreadSummary::Error;
2163 } else {
2164 this.summary = ThreadSummary::Ready(new_summary.into());
2165 }
2166 }
2167 Err(err) => {
2168 this.summary = ThreadSummary::Error;
2169 log::error!("Failed to generate thread summary: {}", err);
2170 }
2171 }
2172 cx.emit(ThreadEvent::SummaryGenerated);
2173 })
2174 .log_err()?;
2175
2176 Some(())
2177 });
2178 }
2179
2180 fn get_retry_strategy(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2181 use LanguageModelCompletionError::*;
2182
2183 // General strategy here:
2184 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2185 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2186 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2187 match error {
2188 HttpResponseError {
2189 status_code: StatusCode::TOO_MANY_REQUESTS,
2190 ..
2191 } => Some(RetryStrategy::ExponentialBackoff {
2192 initial_delay: BASE_RETRY_DELAY,
2193 max_attempts: MAX_RETRY_ATTEMPTS,
2194 }),
2195 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2196 Some(RetryStrategy::Fixed {
2197 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2198 max_attempts: MAX_RETRY_ATTEMPTS,
2199 })
2200 }
2201 UpstreamProviderError {
2202 status,
2203 retry_after,
2204 ..
2205 } => match *status {
2206 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2207 Some(RetryStrategy::Fixed {
2208 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2209 max_attempts: MAX_RETRY_ATTEMPTS,
2210 })
2211 }
2212 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2213 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2214 // Internal Server Error could be anything, retry up to 3 times.
2215 max_attempts: 3,
2216 }),
2217 status => {
2218 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2219 // but we frequently get them in practice. See https://http.dev/529
2220 if status.as_u16() == 529 {
2221 Some(RetryStrategy::Fixed {
2222 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2223 max_attempts: MAX_RETRY_ATTEMPTS,
2224 })
2225 } else {
2226 Some(RetryStrategy::Fixed {
2227 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2228 max_attempts: 2,
2229 })
2230 }
2231 }
2232 },
2233 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2234 delay: BASE_RETRY_DELAY,
2235 max_attempts: 3,
2236 }),
2237 ApiReadResponseError { .. }
2238 | HttpSend { .. }
2239 | DeserializeResponse { .. }
2240 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2241 delay: BASE_RETRY_DELAY,
2242 max_attempts: 3,
2243 }),
2244 // Retrying these errors definitely shouldn't help.
2245 HttpResponseError {
2246 status_code:
2247 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2248 ..
2249 }
2250 | AuthenticationError { .. }
2251 | PermissionError { .. } => None,
2252 // These errors might be transient, so retry them
2253 SerializeRequest { .. }
2254 | BuildRequestBody { .. }
2255 | PromptTooLarge { .. }
2256 | ApiEndpointNotFound { .. }
2257 | NoApiKey { .. } => Some(RetryStrategy::Fixed {
2258 delay: BASE_RETRY_DELAY,
2259 max_attempts: 2,
2260 }),
2261 // Retry all other 4xx and 5xx errors once.
2262 HttpResponseError { status_code, .. }
2263 if status_code.is_client_error() || status_code.is_server_error() =>
2264 {
2265 Some(RetryStrategy::Fixed {
2266 delay: BASE_RETRY_DELAY,
2267 max_attempts: 3,
2268 })
2269 }
2270 // Conservatively assume that any other errors are non-retryable
2271 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2272 delay: BASE_RETRY_DELAY,
2273 max_attempts: 2,
2274 }),
2275 }
2276 }
2277
2278 fn handle_retryable_error_with_delay(
2279 &mut self,
2280 error: &LanguageModelCompletionError,
2281 strategy: Option<RetryStrategy>,
2282 model: Arc<dyn LanguageModel>,
2283 intent: CompletionIntent,
2284 window: Option<AnyWindowHandle>,
2285 cx: &mut Context<Self>,
2286 ) -> bool {
2287 // Store context for the Retry button
2288 self.last_error_context = Some((model.clone(), intent));
2289
2290 // Only auto-retry if Burn Mode is enabled
2291 if self.completion_mode != CompletionMode::Burn {
2292 // Show error with retry options
2293 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2294 message: format!(
2295 "{}\n\nTo automatically retry when similar errors happen, enable Burn Mode.",
2296 error
2297 )
2298 .into(),
2299 can_enable_burn_mode: true,
2300 }));
2301 return false;
2302 }
2303
2304 let Some(strategy) = strategy.or_else(|| Self::get_retry_strategy(error)) else {
2305 return false;
2306 };
2307
2308 let max_attempts = match &strategy {
2309 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2310 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2311 };
2312
2313 let retry_state = self.retry_state.get_or_insert(RetryState {
2314 attempt: 0,
2315 max_attempts,
2316 intent,
2317 });
2318
2319 retry_state.attempt += 1;
2320 let attempt = retry_state.attempt;
2321 let max_attempts = retry_state.max_attempts;
2322 let intent = retry_state.intent;
2323
2324 if attempt <= max_attempts {
2325 let delay = match &strategy {
2326 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2327 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2328 Duration::from_secs(delay_secs)
2329 }
2330 RetryStrategy::Fixed { delay, .. } => *delay,
2331 };
2332
2333 // Add a transient message to inform the user
2334 let delay_secs = delay.as_secs();
2335 let retry_message = if max_attempts == 1 {
2336 format!("{error}. Retrying in {delay_secs} seconds...")
2337 } else {
2338 format!(
2339 "{error}. Retrying (attempt {attempt} of {max_attempts}) \
2340 in {delay_secs} seconds..."
2341 )
2342 };
2343 log::warn!(
2344 "Retrying completion request (attempt {attempt} of {max_attempts}) \
2345 in {delay_secs} seconds: {error:?}",
2346 );
2347
2348 // Add a UI-only message instead of a regular message
2349 let id = self.next_message_id.post_inc();
2350 self.messages.push(Message {
2351 id,
2352 role: Role::System,
2353 segments: vec![MessageSegment::Text(retry_message)],
2354 loaded_context: LoadedContext::default(),
2355 creases: Vec::new(),
2356 is_hidden: false,
2357 ui_only: true,
2358 });
2359 cx.emit(ThreadEvent::MessageAdded(id));
2360
2361 // Schedule the retry
2362 let thread_handle = cx.entity().downgrade();
2363
2364 cx.spawn(async move |_thread, cx| {
2365 cx.background_executor().timer(delay).await;
2366
2367 thread_handle
2368 .update(cx, |thread, cx| {
2369 // Retry the completion
2370 thread.send_to_model(model, intent, window, cx);
2371 })
2372 .log_err();
2373 })
2374 .detach();
2375
2376 true
2377 } else {
2378 // Max retries exceeded
2379 self.retry_state = None;
2380
2381 // Stop generating since we're giving up on retrying.
2382 self.pending_completions.clear();
2383
2384 // Show error alongside a Retry button, but no
2385 // Enable Burn Mode button (since it's already enabled)
2386 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2387 message: format!("Failed after retrying: {}", error).into(),
2388 can_enable_burn_mode: false,
2389 }));
2390
2391 false
2392 }
2393 }
2394
2395 pub fn start_generating_detailed_summary_if_needed(
2396 &mut self,
2397 thread_store: WeakEntity<ThreadStore>,
2398 cx: &mut Context<Self>,
2399 ) {
2400 let Some(last_message_id) = self.messages.last().map(|message| message.id) else {
2401 return;
2402 };
2403
2404 match &*self.detailed_summary_rx.borrow() {
2405 DetailedSummaryState::Generating { message_id, .. }
2406 | DetailedSummaryState::Generated { message_id, .. }
2407 if *message_id == last_message_id =>
2408 {
2409 // Already up-to-date
2410 return;
2411 }
2412 _ => {}
2413 }
2414
2415 let Some(ConfiguredModel { model, provider }) =
2416 LanguageModelRegistry::read_global(cx).thread_summary_model()
2417 else {
2418 return;
2419 };
2420
2421 if !provider.is_authenticated(cx) {
2422 return;
2423 }
2424
2425 let added_user_message = include_str!("./prompts/summarize_thread_detailed_prompt.txt");
2426
2427 let request = self.to_summarize_request(
2428 &model,
2429 CompletionIntent::ThreadContextSummarization,
2430 added_user_message.into(),
2431 cx,
2432 );
2433
2434 *self.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generating {
2435 message_id: last_message_id,
2436 };
2437
2438 // Replace the detailed summarization task if there is one, cancelling it. It would probably
2439 // be better to allow the old task to complete, but this would require logic for choosing
2440 // which result to prefer (the old task could complete after the new one, resulting in a
2441 // stale summary).
2442 self.detailed_summary_task = cx.spawn(async move |thread, cx| {
2443 let stream = model.stream_completion_text(request, &cx);
2444 let Some(mut messages) = stream.await.log_err() else {
2445 thread
2446 .update(cx, |thread, _cx| {
2447 *thread.detailed_summary_tx.borrow_mut() =
2448 DetailedSummaryState::NotGenerated;
2449 })
2450 .ok()?;
2451 return None;
2452 };
2453
2454 let mut new_detailed_summary = String::new();
2455
2456 while let Some(chunk) = messages.stream.next().await {
2457 if let Some(chunk) = chunk.log_err() {
2458 new_detailed_summary.push_str(&chunk);
2459 }
2460 }
2461
2462 thread
2463 .update(cx, |thread, _cx| {
2464 *thread.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generated {
2465 text: new_detailed_summary.into(),
2466 message_id: last_message_id,
2467 };
2468 })
2469 .ok()?;
2470
2471 // Save thread so its summary can be reused later
2472 if let Some(thread) = thread.upgrade() {
2473 if let Ok(Ok(save_task)) = cx.update(|cx| {
2474 thread_store
2475 .update(cx, |thread_store, cx| thread_store.save_thread(&thread, cx))
2476 }) {
2477 save_task.await.log_err();
2478 }
2479 }
2480
2481 Some(())
2482 });
2483 }
2484
2485 pub async fn wait_for_detailed_summary_or_text(
2486 this: &Entity<Self>,
2487 cx: &mut AsyncApp,
2488 ) -> Option<SharedString> {
2489 let mut detailed_summary_rx = this
2490 .read_with(cx, |this, _cx| this.detailed_summary_rx.clone())
2491 .ok()?;
2492 loop {
2493 match detailed_summary_rx.recv().await? {
2494 DetailedSummaryState::Generating { .. } => {}
2495 DetailedSummaryState::NotGenerated => {
2496 return this.read_with(cx, |this, _cx| this.text().into()).ok();
2497 }
2498 DetailedSummaryState::Generated { text, .. } => return Some(text),
2499 }
2500 }
2501 }
2502
2503 pub fn latest_detailed_summary_or_text(&self) -> SharedString {
2504 self.detailed_summary_rx
2505 .borrow()
2506 .text()
2507 .unwrap_or_else(|| self.text().into())
2508 }
2509
2510 pub fn is_generating_detailed_summary(&self) -> bool {
2511 matches!(
2512 &*self.detailed_summary_rx.borrow(),
2513 DetailedSummaryState::Generating { .. }
2514 )
2515 }
2516
2517 pub fn use_pending_tools(
2518 &mut self,
2519 window: Option<AnyWindowHandle>,
2520 model: Arc<dyn LanguageModel>,
2521 cx: &mut Context<Self>,
2522 ) -> Vec<PendingToolUse> {
2523 self.auto_capture_telemetry(cx);
2524 let request =
2525 Arc::new(self.to_completion_request(model.clone(), CompletionIntent::ToolResults, cx));
2526 let pending_tool_uses = self
2527 .tool_use
2528 .pending_tool_uses()
2529 .into_iter()
2530 .filter(|tool_use| tool_use.status.is_idle())
2531 .cloned()
2532 .collect::<Vec<_>>();
2533
2534 for tool_use in pending_tool_uses.iter() {
2535 self.use_pending_tool(tool_use.clone(), request.clone(), model.clone(), window, cx);
2536 }
2537
2538 pending_tool_uses
2539 }
2540
2541 fn use_pending_tool(
2542 &mut self,
2543 tool_use: PendingToolUse,
2544 request: Arc<LanguageModelRequest>,
2545 model: Arc<dyn LanguageModel>,
2546 window: Option<AnyWindowHandle>,
2547 cx: &mut Context<Self>,
2548 ) {
2549 let Some(tool) = self.tools.read(cx).tool(&tool_use.name, cx) else {
2550 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2551 };
2552
2553 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
2554 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2555 }
2556
2557 if tool.needs_confirmation(&tool_use.input, cx)
2558 && !AgentSettings::get_global(cx).always_allow_tool_actions
2559 {
2560 self.tool_use.confirm_tool_use(
2561 tool_use.id,
2562 tool_use.ui_text,
2563 tool_use.input,
2564 request,
2565 tool,
2566 );
2567 cx.emit(ThreadEvent::ToolConfirmationNeeded);
2568 } else {
2569 self.run_tool(
2570 tool_use.id,
2571 tool_use.ui_text,
2572 tool_use.input,
2573 request,
2574 tool,
2575 model,
2576 window,
2577 cx,
2578 );
2579 }
2580 }
2581
2582 pub fn handle_hallucinated_tool_use(
2583 &mut self,
2584 tool_use_id: LanguageModelToolUseId,
2585 hallucinated_tool_name: Arc<str>,
2586 window: Option<AnyWindowHandle>,
2587 cx: &mut Context<Thread>,
2588 ) {
2589 let available_tools = self.profile.enabled_tools(cx);
2590
2591 let tool_list = available_tools
2592 .iter()
2593 .map(|(name, tool)| format!("- {}: {}", name, tool.description()))
2594 .collect::<Vec<_>>()
2595 .join("\n");
2596
2597 let error_message = format!(
2598 "The tool '{}' doesn't exist or is not enabled. Available tools:\n{}",
2599 hallucinated_tool_name, tool_list
2600 );
2601
2602 let pending_tool_use = self.tool_use.insert_tool_output(
2603 tool_use_id.clone(),
2604 hallucinated_tool_name,
2605 Err(anyhow!("Missing tool call: {error_message}")),
2606 self.configured_model.as_ref(),
2607 self.completion_mode,
2608 );
2609
2610 cx.emit(ThreadEvent::MissingToolUse {
2611 tool_use_id: tool_use_id.clone(),
2612 ui_text: error_message.into(),
2613 });
2614
2615 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2616 }
2617
2618 pub fn receive_invalid_tool_json(
2619 &mut self,
2620 tool_use_id: LanguageModelToolUseId,
2621 tool_name: Arc<str>,
2622 invalid_json: Arc<str>,
2623 error: String,
2624 window: Option<AnyWindowHandle>,
2625 cx: &mut Context<Thread>,
2626 ) {
2627 log::error!("The model returned invalid input JSON: {invalid_json}");
2628
2629 let pending_tool_use = self.tool_use.insert_tool_output(
2630 tool_use_id.clone(),
2631 tool_name,
2632 Err(anyhow!("Error parsing input JSON: {error}")),
2633 self.configured_model.as_ref(),
2634 self.completion_mode,
2635 );
2636 let ui_text = if let Some(pending_tool_use) = &pending_tool_use {
2637 pending_tool_use.ui_text.clone()
2638 } else {
2639 log::error!(
2640 "There was no pending tool use for tool use {tool_use_id}, even though it finished (with invalid input JSON)."
2641 );
2642 format!("Unknown tool {}", tool_use_id).into()
2643 };
2644
2645 cx.emit(ThreadEvent::InvalidToolInput {
2646 tool_use_id: tool_use_id.clone(),
2647 ui_text,
2648 invalid_input_json: invalid_json,
2649 });
2650
2651 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2652 }
2653
2654 pub fn run_tool(
2655 &mut self,
2656 tool_use_id: LanguageModelToolUseId,
2657 ui_text: impl Into<SharedString>,
2658 input: serde_json::Value,
2659 request: Arc<LanguageModelRequest>,
2660 tool: Arc<dyn Tool>,
2661 model: Arc<dyn LanguageModel>,
2662 window: Option<AnyWindowHandle>,
2663 cx: &mut Context<Thread>,
2664 ) {
2665 let task =
2666 self.spawn_tool_use(tool_use_id.clone(), request, input, tool, model, window, cx);
2667 self.tool_use
2668 .run_pending_tool(tool_use_id, ui_text.into(), task);
2669 }
2670
2671 fn spawn_tool_use(
2672 &mut self,
2673 tool_use_id: LanguageModelToolUseId,
2674 request: Arc<LanguageModelRequest>,
2675 input: serde_json::Value,
2676 tool: Arc<dyn Tool>,
2677 model: Arc<dyn LanguageModel>,
2678 window: Option<AnyWindowHandle>,
2679 cx: &mut Context<Thread>,
2680 ) -> Task<()> {
2681 let tool_name: Arc<str> = tool.name().into();
2682
2683 let tool_result = tool.run(
2684 input,
2685 request,
2686 self.project.clone(),
2687 self.action_log.clone(),
2688 model,
2689 window,
2690 cx,
2691 );
2692
2693 // Store the card separately if it exists
2694 if let Some(card) = tool_result.card.clone() {
2695 self.tool_use
2696 .insert_tool_result_card(tool_use_id.clone(), card);
2697 }
2698
2699 cx.spawn({
2700 async move |thread: WeakEntity<Thread>, cx| {
2701 let output = tool_result.output.await;
2702
2703 thread
2704 .update(cx, |thread, cx| {
2705 let pending_tool_use = thread.tool_use.insert_tool_output(
2706 tool_use_id.clone(),
2707 tool_name,
2708 output,
2709 thread.configured_model.as_ref(),
2710 thread.completion_mode,
2711 );
2712 thread.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2713 })
2714 .ok();
2715 }
2716 })
2717 }
2718
2719 fn tool_finished(
2720 &mut self,
2721 tool_use_id: LanguageModelToolUseId,
2722 pending_tool_use: Option<PendingToolUse>,
2723 canceled: bool,
2724 window: Option<AnyWindowHandle>,
2725 cx: &mut Context<Self>,
2726 ) {
2727 if self.all_tools_finished() {
2728 if let Some(ConfiguredModel { model, .. }) = self.configured_model.as_ref() {
2729 if !canceled {
2730 self.send_to_model(model.clone(), CompletionIntent::ToolResults, window, cx);
2731 }
2732 self.auto_capture_telemetry(cx);
2733 }
2734 }
2735
2736 cx.emit(ThreadEvent::ToolFinished {
2737 tool_use_id,
2738 pending_tool_use,
2739 });
2740 }
2741
2742 /// Cancels the last pending completion, if there are any pending.
2743 ///
2744 /// Returns whether a completion was canceled.
2745 pub fn cancel_last_completion(
2746 &mut self,
2747 window: Option<AnyWindowHandle>,
2748 cx: &mut Context<Self>,
2749 ) -> bool {
2750 let mut canceled = self.pending_completions.pop().is_some() || self.retry_state.is_some();
2751
2752 self.retry_state = None;
2753
2754 for pending_tool_use in self.tool_use.cancel_pending() {
2755 canceled = true;
2756 self.tool_finished(
2757 pending_tool_use.id.clone(),
2758 Some(pending_tool_use),
2759 true,
2760 window,
2761 cx,
2762 );
2763 }
2764
2765 if canceled {
2766 cx.emit(ThreadEvent::CompletionCanceled);
2767
2768 // When canceled, we always want to insert the checkpoint.
2769 // (We skip over finalize_pending_checkpoint, because it
2770 // would conclude we didn't have anything to insert here.)
2771 if let Some(checkpoint) = self.pending_checkpoint.take() {
2772 self.insert_checkpoint(checkpoint, cx);
2773 }
2774 } else {
2775 self.finalize_pending_checkpoint(cx);
2776 }
2777
2778 canceled
2779 }
2780
2781 /// Signals that any in-progress editing should be canceled.
2782 ///
2783 /// This method is used to notify listeners (like ActiveThread) that
2784 /// they should cancel any editing operations.
2785 pub fn cancel_editing(&mut self, cx: &mut Context<Self>) {
2786 cx.emit(ThreadEvent::CancelEditing);
2787 }
2788
2789 pub fn feedback(&self) -> Option<ThreadFeedback> {
2790 self.feedback
2791 }
2792
2793 pub fn message_feedback(&self, message_id: MessageId) -> Option<ThreadFeedback> {
2794 self.message_feedback.get(&message_id).copied()
2795 }
2796
2797 pub fn report_message_feedback(
2798 &mut self,
2799 message_id: MessageId,
2800 feedback: ThreadFeedback,
2801 cx: &mut Context<Self>,
2802 ) -> Task<Result<()>> {
2803 if self.message_feedback.get(&message_id) == Some(&feedback) {
2804 return Task::ready(Ok(()));
2805 }
2806
2807 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2808 let serialized_thread = self.serialize(cx);
2809 let thread_id = self.id().clone();
2810 let client = self.project.read(cx).client();
2811
2812 let enabled_tool_names: Vec<String> = self
2813 .profile
2814 .enabled_tools(cx)
2815 .iter()
2816 .map(|(name, _)| name.clone().into())
2817 .collect();
2818
2819 self.message_feedback.insert(message_id, feedback);
2820
2821 cx.notify();
2822
2823 let message_content = self
2824 .message(message_id)
2825 .map(|msg| msg.to_string())
2826 .unwrap_or_default();
2827
2828 cx.background_spawn(async move {
2829 let final_project_snapshot = final_project_snapshot.await;
2830 let serialized_thread = serialized_thread.await?;
2831 let thread_data =
2832 serde_json::to_value(serialized_thread).unwrap_or_else(|_| serde_json::Value::Null);
2833
2834 let rating = match feedback {
2835 ThreadFeedback::Positive => "positive",
2836 ThreadFeedback::Negative => "negative",
2837 };
2838 telemetry::event!(
2839 "Assistant Thread Rated",
2840 rating,
2841 thread_id,
2842 enabled_tool_names,
2843 message_id = message_id.0,
2844 message_content,
2845 thread_data,
2846 final_project_snapshot
2847 );
2848 client.telemetry().flush_events().await;
2849
2850 Ok(())
2851 })
2852 }
2853
2854 pub fn report_feedback(
2855 &mut self,
2856 feedback: ThreadFeedback,
2857 cx: &mut Context<Self>,
2858 ) -> Task<Result<()>> {
2859 let last_assistant_message_id = self
2860 .messages
2861 .iter()
2862 .rev()
2863 .find(|msg| msg.role == Role::Assistant)
2864 .map(|msg| msg.id);
2865
2866 if let Some(message_id) = last_assistant_message_id {
2867 self.report_message_feedback(message_id, feedback, cx)
2868 } else {
2869 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2870 let serialized_thread = self.serialize(cx);
2871 let thread_id = self.id().clone();
2872 let client = self.project.read(cx).client();
2873 self.feedback = Some(feedback);
2874 cx.notify();
2875
2876 cx.background_spawn(async move {
2877 let final_project_snapshot = final_project_snapshot.await;
2878 let serialized_thread = serialized_thread.await?;
2879 let thread_data = serde_json::to_value(serialized_thread)
2880 .unwrap_or_else(|_| serde_json::Value::Null);
2881
2882 let rating = match feedback {
2883 ThreadFeedback::Positive => "positive",
2884 ThreadFeedback::Negative => "negative",
2885 };
2886 telemetry::event!(
2887 "Assistant Thread Rated",
2888 rating,
2889 thread_id,
2890 thread_data,
2891 final_project_snapshot
2892 );
2893 client.telemetry().flush_events().await;
2894
2895 Ok(())
2896 })
2897 }
2898 }
2899
2900 /// Create a snapshot of the current project state including git information and unsaved buffers.
2901 fn project_snapshot(
2902 project: Entity<Project>,
2903 cx: &mut Context<Self>,
2904 ) -> Task<Arc<ProjectSnapshot>> {
2905 let git_store = project.read(cx).git_store().clone();
2906 let worktree_snapshots: Vec<_> = project
2907 .read(cx)
2908 .visible_worktrees(cx)
2909 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
2910 .collect();
2911
2912 cx.spawn(async move |_, cx| {
2913 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
2914
2915 let mut unsaved_buffers = Vec::new();
2916 cx.update(|app_cx| {
2917 let buffer_store = project.read(app_cx).buffer_store();
2918 for buffer_handle in buffer_store.read(app_cx).buffers() {
2919 let buffer = buffer_handle.read(app_cx);
2920 if buffer.is_dirty() {
2921 if let Some(file) = buffer.file() {
2922 let path = file.path().to_string_lossy().to_string();
2923 unsaved_buffers.push(path);
2924 }
2925 }
2926 }
2927 })
2928 .ok();
2929
2930 Arc::new(ProjectSnapshot {
2931 worktree_snapshots,
2932 unsaved_buffer_paths: unsaved_buffers,
2933 timestamp: Utc::now(),
2934 })
2935 })
2936 }
2937
2938 fn worktree_snapshot(
2939 worktree: Entity<project::Worktree>,
2940 git_store: Entity<GitStore>,
2941 cx: &App,
2942 ) -> Task<WorktreeSnapshot> {
2943 cx.spawn(async move |cx| {
2944 // Get worktree path and snapshot
2945 let worktree_info = cx.update(|app_cx| {
2946 let worktree = worktree.read(app_cx);
2947 let path = worktree.abs_path().to_string_lossy().to_string();
2948 let snapshot = worktree.snapshot();
2949 (path, snapshot)
2950 });
2951
2952 let Ok((worktree_path, _snapshot)) = worktree_info else {
2953 return WorktreeSnapshot {
2954 worktree_path: String::new(),
2955 git_state: None,
2956 };
2957 };
2958
2959 let git_state = git_store
2960 .update(cx, |git_store, cx| {
2961 git_store
2962 .repositories()
2963 .values()
2964 .find(|repo| {
2965 repo.read(cx)
2966 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
2967 .is_some()
2968 })
2969 .cloned()
2970 })
2971 .ok()
2972 .flatten()
2973 .map(|repo| {
2974 repo.update(cx, |repo, _| {
2975 let current_branch =
2976 repo.branch.as_ref().map(|branch| branch.name().to_owned());
2977 repo.send_job(None, |state, _| async move {
2978 let RepositoryState::Local { backend, .. } = state else {
2979 return GitState {
2980 remote_url: None,
2981 head_sha: None,
2982 current_branch,
2983 diff: None,
2984 };
2985 };
2986
2987 let remote_url = backend.remote_url("origin");
2988 let head_sha = backend.head_sha().await;
2989 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
2990
2991 GitState {
2992 remote_url,
2993 head_sha,
2994 current_branch,
2995 diff,
2996 }
2997 })
2998 })
2999 });
3000
3001 let git_state = match git_state {
3002 Some(git_state) => match git_state.ok() {
3003 Some(git_state) => git_state.await.ok(),
3004 None => None,
3005 },
3006 None => None,
3007 };
3008
3009 WorktreeSnapshot {
3010 worktree_path,
3011 git_state,
3012 }
3013 })
3014 }
3015
3016 pub fn to_markdown(&self, cx: &App) -> Result<String> {
3017 let mut markdown = Vec::new();
3018
3019 let summary = self.summary().or_default();
3020 writeln!(markdown, "# {summary}\n")?;
3021
3022 for message in self.messages() {
3023 writeln!(
3024 markdown,
3025 "## {role}\n",
3026 role = match message.role {
3027 Role::User => "User",
3028 Role::Assistant => "Agent",
3029 Role::System => "System",
3030 }
3031 )?;
3032
3033 if !message.loaded_context.text.is_empty() {
3034 writeln!(markdown, "{}", message.loaded_context.text)?;
3035 }
3036
3037 if !message.loaded_context.images.is_empty() {
3038 writeln!(
3039 markdown,
3040 "\n{} images attached as context.\n",
3041 message.loaded_context.images.len()
3042 )?;
3043 }
3044
3045 for segment in &message.segments {
3046 match segment {
3047 MessageSegment::Text(text) => writeln!(markdown, "{}\n", text)?,
3048 MessageSegment::Thinking { text, .. } => {
3049 writeln!(markdown, "<think>\n{}\n</think>\n", text)?
3050 }
3051 MessageSegment::RedactedThinking(_) => {}
3052 }
3053 }
3054
3055 for tool_use in self.tool_uses_for_message(message.id, cx) {
3056 writeln!(
3057 markdown,
3058 "**Use Tool: {} ({})**",
3059 tool_use.name, tool_use.id
3060 )?;
3061 writeln!(markdown, "```json")?;
3062 writeln!(
3063 markdown,
3064 "{}",
3065 serde_json::to_string_pretty(&tool_use.input)?
3066 )?;
3067 writeln!(markdown, "```")?;
3068 }
3069
3070 for tool_result in self.tool_results_for_message(message.id) {
3071 write!(markdown, "\n**Tool Results: {}", tool_result.tool_use_id)?;
3072 if tool_result.is_error {
3073 write!(markdown, " (Error)")?;
3074 }
3075
3076 writeln!(markdown, "**\n")?;
3077 match &tool_result.content {
3078 LanguageModelToolResultContent::Text(text) => {
3079 writeln!(markdown, "{text}")?;
3080 }
3081 LanguageModelToolResultContent::Image(image) => {
3082 writeln!(markdown, "", image.source)?;
3083 }
3084 }
3085
3086 if let Some(output) = tool_result.output.as_ref() {
3087 writeln!(
3088 markdown,
3089 "\n\nDebug Output:\n\n```json\n{}\n```\n",
3090 serde_json::to_string_pretty(output)?
3091 )?;
3092 }
3093 }
3094 }
3095
3096 Ok(String::from_utf8_lossy(&markdown).to_string())
3097 }
3098
3099 pub fn keep_edits_in_range(
3100 &mut self,
3101 buffer: Entity<language::Buffer>,
3102 buffer_range: Range<language::Anchor>,
3103 cx: &mut Context<Self>,
3104 ) {
3105 self.action_log.update(cx, |action_log, cx| {
3106 action_log.keep_edits_in_range(buffer, buffer_range, cx)
3107 });
3108 }
3109
3110 pub fn keep_all_edits(&mut self, cx: &mut Context<Self>) {
3111 self.action_log
3112 .update(cx, |action_log, cx| action_log.keep_all_edits(cx));
3113 }
3114
3115 pub fn reject_edits_in_ranges(
3116 &mut self,
3117 buffer: Entity<language::Buffer>,
3118 buffer_ranges: Vec<Range<language::Anchor>>,
3119 cx: &mut Context<Self>,
3120 ) -> Task<Result<()>> {
3121 self.action_log.update(cx, |action_log, cx| {
3122 action_log.reject_edits_in_ranges(buffer, buffer_ranges, cx)
3123 })
3124 }
3125
3126 pub fn action_log(&self) -> &Entity<ActionLog> {
3127 &self.action_log
3128 }
3129
3130 pub fn project(&self) -> &Entity<Project> {
3131 &self.project
3132 }
3133
3134 pub fn auto_capture_telemetry(&mut self, cx: &mut Context<Self>) {
3135 if !cx.has_flag::<feature_flags::ThreadAutoCaptureFeatureFlag>() {
3136 return;
3137 }
3138
3139 let now = Instant::now();
3140 if let Some(last) = self.last_auto_capture_at {
3141 if now.duration_since(last).as_secs() < 10 {
3142 return;
3143 }
3144 }
3145
3146 self.last_auto_capture_at = Some(now);
3147
3148 let thread_id = self.id().clone();
3149 let github_login = self
3150 .project
3151 .read(cx)
3152 .user_store()
3153 .read(cx)
3154 .current_user()
3155 .map(|user| user.github_login.clone());
3156 let client = self.project.read(cx).client();
3157 let serialize_task = self.serialize(cx);
3158
3159 cx.background_executor()
3160 .spawn(async move {
3161 if let Ok(serialized_thread) = serialize_task.await {
3162 if let Ok(thread_data) = serde_json::to_value(serialized_thread) {
3163 telemetry::event!(
3164 "Agent Thread Auto-Captured",
3165 thread_id = thread_id.to_string(),
3166 thread_data = thread_data,
3167 auto_capture_reason = "tracked_user",
3168 github_login = github_login
3169 );
3170
3171 client.telemetry().flush_events().await;
3172 }
3173 }
3174 })
3175 .detach();
3176 }
3177
3178 pub fn cumulative_token_usage(&self) -> TokenUsage {
3179 self.cumulative_token_usage
3180 }
3181
3182 pub fn token_usage_up_to_message(&self, message_id: MessageId) -> TotalTokenUsage {
3183 let Some(model) = self.configured_model.as_ref() else {
3184 return TotalTokenUsage::default();
3185 };
3186
3187 let max = model
3188 .model
3189 .max_token_count_for_mode(self.completion_mode().into());
3190
3191 let index = self
3192 .messages
3193 .iter()
3194 .position(|msg| msg.id == message_id)
3195 .unwrap_or(0);
3196
3197 if index == 0 {
3198 return TotalTokenUsage { total: 0, max };
3199 }
3200
3201 let token_usage = &self
3202 .request_token_usage
3203 .get(index - 1)
3204 .cloned()
3205 .unwrap_or_default();
3206
3207 TotalTokenUsage {
3208 total: token_usage.total_tokens(),
3209 max,
3210 }
3211 }
3212
3213 pub fn total_token_usage(&self) -> Option<TotalTokenUsage> {
3214 let model = self.configured_model.as_ref()?;
3215
3216 let max = model
3217 .model
3218 .max_token_count_for_mode(self.completion_mode().into());
3219
3220 if let Some(exceeded_error) = &self.exceeded_window_error {
3221 if model.model.id() == exceeded_error.model_id {
3222 return Some(TotalTokenUsage {
3223 total: exceeded_error.token_count,
3224 max,
3225 });
3226 }
3227 }
3228
3229 let total = self
3230 .token_usage_at_last_message()
3231 .unwrap_or_default()
3232 .total_tokens();
3233
3234 Some(TotalTokenUsage { total, max })
3235 }
3236
3237 fn token_usage_at_last_message(&self) -> Option<TokenUsage> {
3238 self.request_token_usage
3239 .get(self.messages.len().saturating_sub(1))
3240 .or_else(|| self.request_token_usage.last())
3241 .cloned()
3242 }
3243
3244 fn update_token_usage_at_last_message(&mut self, token_usage: TokenUsage) {
3245 let placeholder = self.token_usage_at_last_message().unwrap_or_default();
3246 self.request_token_usage
3247 .resize(self.messages.len(), placeholder);
3248
3249 if let Some(last) = self.request_token_usage.last_mut() {
3250 *last = token_usage;
3251 }
3252 }
3253
3254 fn update_model_request_usage(&self, amount: u32, limit: UsageLimit, cx: &mut Context<Self>) {
3255 self.project.update(cx, |project, cx| {
3256 project.user_store().update(cx, |user_store, cx| {
3257 user_store.update_model_request_usage(
3258 ModelRequestUsage(RequestUsage {
3259 amount: amount as i32,
3260 limit,
3261 }),
3262 cx,
3263 )
3264 })
3265 });
3266 }
3267
3268 pub fn deny_tool_use(
3269 &mut self,
3270 tool_use_id: LanguageModelToolUseId,
3271 tool_name: Arc<str>,
3272 window: Option<AnyWindowHandle>,
3273 cx: &mut Context<Self>,
3274 ) {
3275 let err = Err(anyhow::anyhow!(
3276 "Permission to run tool action denied by user"
3277 ));
3278
3279 self.tool_use.insert_tool_output(
3280 tool_use_id.clone(),
3281 tool_name,
3282 err,
3283 self.configured_model.as_ref(),
3284 self.completion_mode,
3285 );
3286 self.tool_finished(tool_use_id.clone(), None, true, window, cx);
3287 }
3288}
3289
3290#[derive(Debug, Clone, Error)]
3291pub enum ThreadError {
3292 #[error("Payment required")]
3293 PaymentRequired,
3294 #[error("Model request limit reached")]
3295 ModelRequestLimitReached { plan: Plan },
3296 #[error("Message {header}: {message}")]
3297 Message {
3298 header: SharedString,
3299 message: SharedString,
3300 },
3301 #[error("Retryable error: {message}")]
3302 RetryableError {
3303 message: SharedString,
3304 can_enable_burn_mode: bool,
3305 },
3306}
3307
3308#[derive(Debug, Clone)]
3309pub enum ThreadEvent {
3310 ShowError(ThreadError),
3311 StreamedCompletion,
3312 ReceivedTextChunk,
3313 NewRequest,
3314 StreamedAssistantText(MessageId, String),
3315 StreamedAssistantThinking(MessageId, String),
3316 StreamedToolUse {
3317 tool_use_id: LanguageModelToolUseId,
3318 ui_text: Arc<str>,
3319 input: serde_json::Value,
3320 },
3321 MissingToolUse {
3322 tool_use_id: LanguageModelToolUseId,
3323 ui_text: Arc<str>,
3324 },
3325 InvalidToolInput {
3326 tool_use_id: LanguageModelToolUseId,
3327 ui_text: Arc<str>,
3328 invalid_input_json: Arc<str>,
3329 },
3330 Stopped(Result<StopReason, Arc<anyhow::Error>>),
3331 MessageAdded(MessageId),
3332 MessageEdited(MessageId),
3333 MessageDeleted(MessageId),
3334 SummaryGenerated,
3335 SummaryChanged,
3336 UsePendingTools {
3337 tool_uses: Vec<PendingToolUse>,
3338 },
3339 ToolFinished {
3340 #[allow(unused)]
3341 tool_use_id: LanguageModelToolUseId,
3342 /// The pending tool use that corresponds to this tool.
3343 pending_tool_use: Option<PendingToolUse>,
3344 },
3345 CheckpointChanged,
3346 ToolConfirmationNeeded,
3347 ToolUseLimitReached,
3348 CancelEditing,
3349 CompletionCanceled,
3350 ProfileChanged,
3351}
3352
3353impl EventEmitter<ThreadEvent> for Thread {}
3354
3355struct PendingCompletion {
3356 id: usize,
3357 queue_state: QueueState,
3358 _task: Task<()>,
3359}
3360
3361#[cfg(test)]
3362mod tests {
3363 use super::*;
3364 use crate::{
3365 context::load_context, context_store::ContextStore, thread_store, thread_store::ThreadStore,
3366 };
3367
3368 // Test-specific constants
3369 const TEST_RATE_LIMIT_RETRY_SECS: u64 = 30;
3370 use agent_settings::{AgentProfileId, AgentSettings, LanguageModelParameters};
3371 use assistant_tool::ToolRegistry;
3372 use assistant_tools;
3373 use futures::StreamExt;
3374 use futures::future::BoxFuture;
3375 use futures::stream::BoxStream;
3376 use gpui::TestAppContext;
3377 use http_client;
3378 use language_model::fake_provider::{FakeLanguageModel, FakeLanguageModelProvider};
3379 use language_model::{
3380 LanguageModelCompletionError, LanguageModelName, LanguageModelProviderId,
3381 LanguageModelProviderName, LanguageModelToolChoice,
3382 };
3383 use parking_lot::Mutex;
3384 use project::{FakeFs, Project};
3385 use prompt_store::PromptBuilder;
3386 use serde_json::json;
3387 use settings::{Settings, SettingsStore};
3388 use std::sync::Arc;
3389 use std::time::Duration;
3390 use theme::ThemeSettings;
3391 use util::path;
3392 use workspace::Workspace;
3393
3394 #[gpui::test]
3395 async fn test_message_with_context(cx: &mut TestAppContext) {
3396 init_test_settings(cx);
3397
3398 let project = create_test_project(
3399 cx,
3400 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3401 )
3402 .await;
3403
3404 let (_workspace, _thread_store, thread, context_store, model) =
3405 setup_test_environment(cx, project.clone()).await;
3406
3407 add_file_to_context(&project, &context_store, "test/code.rs", cx)
3408 .await
3409 .unwrap();
3410
3411 let context =
3412 context_store.read_with(cx, |store, _| store.context().next().cloned().unwrap());
3413 let loaded_context = cx
3414 .update(|cx| load_context(vec![context], &project, &None, cx))
3415 .await;
3416
3417 // Insert user message with context
3418 let message_id = thread.update(cx, |thread, cx| {
3419 thread.insert_user_message(
3420 "Please explain this code",
3421 loaded_context,
3422 None,
3423 Vec::new(),
3424 cx,
3425 )
3426 });
3427
3428 // Check content and context in message object
3429 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3430
3431 // Use different path format strings based on platform for the test
3432 #[cfg(windows)]
3433 let path_part = r"test\code.rs";
3434 #[cfg(not(windows))]
3435 let path_part = "test/code.rs";
3436
3437 let expected_context = format!(
3438 r#"
3439<context>
3440The following items were attached by the user. They are up-to-date and don't need to be re-read.
3441
3442<files>
3443```rs {path_part}
3444fn main() {{
3445 println!("Hello, world!");
3446}}
3447```
3448</files>
3449</context>
3450"#
3451 );
3452
3453 assert_eq!(message.role, Role::User);
3454 assert_eq!(message.segments.len(), 1);
3455 assert_eq!(
3456 message.segments[0],
3457 MessageSegment::Text("Please explain this code".to_string())
3458 );
3459 assert_eq!(message.loaded_context.text, expected_context);
3460
3461 // Check message in request
3462 let request = thread.update(cx, |thread, cx| {
3463 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3464 });
3465
3466 assert_eq!(request.messages.len(), 2);
3467 let expected_full_message = format!("{}Please explain this code", expected_context);
3468 assert_eq!(request.messages[1].string_contents(), expected_full_message);
3469 }
3470
3471 #[gpui::test]
3472 async fn test_only_include_new_contexts(cx: &mut TestAppContext) {
3473 init_test_settings(cx);
3474
3475 let project = create_test_project(
3476 cx,
3477 json!({
3478 "file1.rs": "fn function1() {}\n",
3479 "file2.rs": "fn function2() {}\n",
3480 "file3.rs": "fn function3() {}\n",
3481 "file4.rs": "fn function4() {}\n",
3482 }),
3483 )
3484 .await;
3485
3486 let (_, _thread_store, thread, context_store, model) =
3487 setup_test_environment(cx, project.clone()).await;
3488
3489 // First message with context 1
3490 add_file_to_context(&project, &context_store, "test/file1.rs", cx)
3491 .await
3492 .unwrap();
3493 let new_contexts = context_store.update(cx, |store, cx| {
3494 store.new_context_for_thread(thread.read(cx), None)
3495 });
3496 assert_eq!(new_contexts.len(), 1);
3497 let loaded_context = cx
3498 .update(|cx| load_context(new_contexts, &project, &None, cx))
3499 .await;
3500 let message1_id = thread.update(cx, |thread, cx| {
3501 thread.insert_user_message("Message 1", loaded_context, None, Vec::new(), cx)
3502 });
3503
3504 // Second message with contexts 1 and 2 (context 1 should be skipped as it's already included)
3505 add_file_to_context(&project, &context_store, "test/file2.rs", cx)
3506 .await
3507 .unwrap();
3508 let new_contexts = context_store.update(cx, |store, cx| {
3509 store.new_context_for_thread(thread.read(cx), None)
3510 });
3511 assert_eq!(new_contexts.len(), 1);
3512 let loaded_context = cx
3513 .update(|cx| load_context(new_contexts, &project, &None, cx))
3514 .await;
3515 let message2_id = thread.update(cx, |thread, cx| {
3516 thread.insert_user_message("Message 2", loaded_context, None, Vec::new(), cx)
3517 });
3518
3519 // Third message with all three contexts (contexts 1 and 2 should be skipped)
3520 //
3521 add_file_to_context(&project, &context_store, "test/file3.rs", cx)
3522 .await
3523 .unwrap();
3524 let new_contexts = context_store.update(cx, |store, cx| {
3525 store.new_context_for_thread(thread.read(cx), None)
3526 });
3527 assert_eq!(new_contexts.len(), 1);
3528 let loaded_context = cx
3529 .update(|cx| load_context(new_contexts, &project, &None, cx))
3530 .await;
3531 let message3_id = thread.update(cx, |thread, cx| {
3532 thread.insert_user_message("Message 3", loaded_context, None, Vec::new(), cx)
3533 });
3534
3535 // Check what contexts are included in each message
3536 let (message1, message2, message3) = thread.read_with(cx, |thread, _| {
3537 (
3538 thread.message(message1_id).unwrap().clone(),
3539 thread.message(message2_id).unwrap().clone(),
3540 thread.message(message3_id).unwrap().clone(),
3541 )
3542 });
3543
3544 // First message should include context 1
3545 assert!(message1.loaded_context.text.contains("file1.rs"));
3546
3547 // Second message should include only context 2 (not 1)
3548 assert!(!message2.loaded_context.text.contains("file1.rs"));
3549 assert!(message2.loaded_context.text.contains("file2.rs"));
3550
3551 // Third message should include only context 3 (not 1 or 2)
3552 assert!(!message3.loaded_context.text.contains("file1.rs"));
3553 assert!(!message3.loaded_context.text.contains("file2.rs"));
3554 assert!(message3.loaded_context.text.contains("file3.rs"));
3555
3556 // Check entire request to make sure all contexts are properly included
3557 let request = thread.update(cx, |thread, cx| {
3558 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3559 });
3560
3561 // The request should contain all 3 messages
3562 assert_eq!(request.messages.len(), 4);
3563
3564 // Check that the contexts are properly formatted in each message
3565 assert!(request.messages[1].string_contents().contains("file1.rs"));
3566 assert!(!request.messages[1].string_contents().contains("file2.rs"));
3567 assert!(!request.messages[1].string_contents().contains("file3.rs"));
3568
3569 assert!(!request.messages[2].string_contents().contains("file1.rs"));
3570 assert!(request.messages[2].string_contents().contains("file2.rs"));
3571 assert!(!request.messages[2].string_contents().contains("file3.rs"));
3572
3573 assert!(!request.messages[3].string_contents().contains("file1.rs"));
3574 assert!(!request.messages[3].string_contents().contains("file2.rs"));
3575 assert!(request.messages[3].string_contents().contains("file3.rs"));
3576
3577 add_file_to_context(&project, &context_store, "test/file4.rs", cx)
3578 .await
3579 .unwrap();
3580 let new_contexts = context_store.update(cx, |store, cx| {
3581 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3582 });
3583 assert_eq!(new_contexts.len(), 3);
3584 let loaded_context = cx
3585 .update(|cx| load_context(new_contexts, &project, &None, cx))
3586 .await
3587 .loaded_context;
3588
3589 assert!(!loaded_context.text.contains("file1.rs"));
3590 assert!(loaded_context.text.contains("file2.rs"));
3591 assert!(loaded_context.text.contains("file3.rs"));
3592 assert!(loaded_context.text.contains("file4.rs"));
3593
3594 let new_contexts = context_store.update(cx, |store, cx| {
3595 // Remove file4.rs
3596 store.remove_context(&loaded_context.contexts[2].handle(), cx);
3597 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3598 });
3599 assert_eq!(new_contexts.len(), 2);
3600 let loaded_context = cx
3601 .update(|cx| load_context(new_contexts, &project, &None, cx))
3602 .await
3603 .loaded_context;
3604
3605 assert!(!loaded_context.text.contains("file1.rs"));
3606 assert!(loaded_context.text.contains("file2.rs"));
3607 assert!(loaded_context.text.contains("file3.rs"));
3608 assert!(!loaded_context.text.contains("file4.rs"));
3609
3610 let new_contexts = context_store.update(cx, |store, cx| {
3611 // Remove file3.rs
3612 store.remove_context(&loaded_context.contexts[1].handle(), cx);
3613 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3614 });
3615 assert_eq!(new_contexts.len(), 1);
3616 let loaded_context = cx
3617 .update(|cx| load_context(new_contexts, &project, &None, cx))
3618 .await
3619 .loaded_context;
3620
3621 assert!(!loaded_context.text.contains("file1.rs"));
3622 assert!(loaded_context.text.contains("file2.rs"));
3623 assert!(!loaded_context.text.contains("file3.rs"));
3624 assert!(!loaded_context.text.contains("file4.rs"));
3625 }
3626
3627 #[gpui::test]
3628 async fn test_message_without_files(cx: &mut TestAppContext) {
3629 init_test_settings(cx);
3630
3631 let project = create_test_project(
3632 cx,
3633 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3634 )
3635 .await;
3636
3637 let (_, _thread_store, thread, _context_store, model) =
3638 setup_test_environment(cx, project.clone()).await;
3639
3640 // Insert user message without any context (empty context vector)
3641 let message_id = thread.update(cx, |thread, cx| {
3642 thread.insert_user_message(
3643 "What is the best way to learn Rust?",
3644 ContextLoadResult::default(),
3645 None,
3646 Vec::new(),
3647 cx,
3648 )
3649 });
3650
3651 // Check content and context in message object
3652 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3653
3654 // Context should be empty when no files are included
3655 assert_eq!(message.role, Role::User);
3656 assert_eq!(message.segments.len(), 1);
3657 assert_eq!(
3658 message.segments[0],
3659 MessageSegment::Text("What is the best way to learn Rust?".to_string())
3660 );
3661 assert_eq!(message.loaded_context.text, "");
3662
3663 // Check message in request
3664 let request = thread.update(cx, |thread, cx| {
3665 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3666 });
3667
3668 assert_eq!(request.messages.len(), 2);
3669 assert_eq!(
3670 request.messages[1].string_contents(),
3671 "What is the best way to learn Rust?"
3672 );
3673
3674 // Add second message, also without context
3675 let message2_id = thread.update(cx, |thread, cx| {
3676 thread.insert_user_message(
3677 "Are there any good books?",
3678 ContextLoadResult::default(),
3679 None,
3680 Vec::new(),
3681 cx,
3682 )
3683 });
3684
3685 let message2 =
3686 thread.read_with(cx, |thread, _| thread.message(message2_id).unwrap().clone());
3687 assert_eq!(message2.loaded_context.text, "");
3688
3689 // Check that both messages appear in the request
3690 let request = thread.update(cx, |thread, cx| {
3691 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3692 });
3693
3694 assert_eq!(request.messages.len(), 3);
3695 assert_eq!(
3696 request.messages[1].string_contents(),
3697 "What is the best way to learn Rust?"
3698 );
3699 assert_eq!(
3700 request.messages[2].string_contents(),
3701 "Are there any good books?"
3702 );
3703 }
3704
3705 #[gpui::test]
3706 #[ignore] // turn this test on when project_notifications tool is re-enabled
3707 async fn test_stale_buffer_notification(cx: &mut TestAppContext) {
3708 init_test_settings(cx);
3709
3710 let project = create_test_project(
3711 cx,
3712 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3713 )
3714 .await;
3715
3716 let (_workspace, _thread_store, thread, context_store, model) =
3717 setup_test_environment(cx, project.clone()).await;
3718
3719 // Add a buffer to the context. This will be a tracked buffer
3720 let buffer = add_file_to_context(&project, &context_store, "test/code.rs", cx)
3721 .await
3722 .unwrap();
3723
3724 let context = context_store
3725 .read_with(cx, |store, _| store.context().next().cloned())
3726 .unwrap();
3727 let loaded_context = cx
3728 .update(|cx| load_context(vec![context], &project, &None, cx))
3729 .await;
3730
3731 // Insert user message and assistant response
3732 thread.update(cx, |thread, cx| {
3733 thread.insert_user_message("Explain this code", loaded_context, None, Vec::new(), cx);
3734 thread.insert_assistant_message(
3735 vec![MessageSegment::Text("This code prints 42.".into())],
3736 cx,
3737 );
3738 });
3739 cx.run_until_parked();
3740
3741 // We shouldn't have a stale buffer notification yet
3742 let notifications = thread.read_with(cx, |thread, _| {
3743 find_tool_uses(thread, "project_notifications")
3744 });
3745 assert!(
3746 notifications.is_empty(),
3747 "Should not have stale buffer notification before buffer is modified"
3748 );
3749
3750 // Modify the buffer
3751 buffer.update(cx, |buffer, cx| {
3752 buffer.edit(
3753 [(1..1, "\n println!(\"Added a new line\");\n")],
3754 None,
3755 cx,
3756 );
3757 });
3758
3759 // Insert another user message
3760 thread.update(cx, |thread, cx| {
3761 thread.insert_user_message(
3762 "What does the code do now?",
3763 ContextLoadResult::default(),
3764 None,
3765 Vec::new(),
3766 cx,
3767 )
3768 });
3769 cx.run_until_parked();
3770
3771 // Check for the stale buffer warning
3772 thread.update(cx, |thread, cx| {
3773 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3774 });
3775 cx.run_until_parked();
3776
3777 let notifications = thread.read_with(cx, |thread, _cx| {
3778 find_tool_uses(thread, "project_notifications")
3779 });
3780
3781 let [notification] = notifications.as_slice() else {
3782 panic!("Should have a `project_notifications` tool use");
3783 };
3784
3785 let Some(notification_content) = notification.content.to_str() else {
3786 panic!("`project_notifications` should return text");
3787 };
3788
3789 assert!(notification_content.contains("These files have changed since the last read:"));
3790 assert!(notification_content.contains("code.rs"));
3791
3792 // Insert another user message and flush notifications again
3793 thread.update(cx, |thread, cx| {
3794 thread.insert_user_message(
3795 "Can you tell me more?",
3796 ContextLoadResult::default(),
3797 None,
3798 Vec::new(),
3799 cx,
3800 )
3801 });
3802
3803 thread.update(cx, |thread, cx| {
3804 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3805 });
3806 cx.run_until_parked();
3807
3808 // There should be no new notifications (we already flushed one)
3809 let notifications = thread.read_with(cx, |thread, _cx| {
3810 find_tool_uses(thread, "project_notifications")
3811 });
3812
3813 assert_eq!(
3814 notifications.len(),
3815 1,
3816 "Should still have only one notification after second flush - no duplicates"
3817 );
3818 }
3819
3820 fn find_tool_uses(thread: &Thread, tool_name: &str) -> Vec<LanguageModelToolResult> {
3821 thread
3822 .messages()
3823 .flat_map(|message| {
3824 thread
3825 .tool_results_for_message(message.id)
3826 .into_iter()
3827 .filter(|result| result.tool_name == tool_name.into())
3828 .cloned()
3829 .collect::<Vec<_>>()
3830 })
3831 .collect()
3832 }
3833
3834 #[gpui::test]
3835 async fn test_storing_profile_setting_per_thread(cx: &mut TestAppContext) {
3836 init_test_settings(cx);
3837
3838 let project = create_test_project(
3839 cx,
3840 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3841 )
3842 .await;
3843
3844 let (_workspace, thread_store, thread, _context_store, _model) =
3845 setup_test_environment(cx, project.clone()).await;
3846
3847 // Check that we are starting with the default profile
3848 let profile = cx.read(|cx| thread.read(cx).profile.clone());
3849 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3850 assert_eq!(
3851 profile,
3852 AgentProfile::new(AgentProfileId::default(), tool_set)
3853 );
3854 }
3855
3856 #[gpui::test]
3857 async fn test_serializing_thread_profile(cx: &mut TestAppContext) {
3858 init_test_settings(cx);
3859
3860 let project = create_test_project(
3861 cx,
3862 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3863 )
3864 .await;
3865
3866 let (_workspace, thread_store, thread, _context_store, _model) =
3867 setup_test_environment(cx, project.clone()).await;
3868
3869 // Profile gets serialized with default values
3870 let serialized = thread
3871 .update(cx, |thread, cx| thread.serialize(cx))
3872 .await
3873 .unwrap();
3874
3875 assert_eq!(serialized.profile, Some(AgentProfileId::default()));
3876
3877 let deserialized = cx.update(|cx| {
3878 thread.update(cx, |thread, cx| {
3879 Thread::deserialize(
3880 thread.id.clone(),
3881 serialized,
3882 thread.project.clone(),
3883 thread.tools.clone(),
3884 thread.prompt_builder.clone(),
3885 thread.project_context.clone(),
3886 None,
3887 cx,
3888 )
3889 })
3890 });
3891 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3892
3893 assert_eq!(
3894 deserialized.profile,
3895 AgentProfile::new(AgentProfileId::default(), tool_set)
3896 );
3897 }
3898
3899 #[gpui::test]
3900 async fn test_temperature_setting(cx: &mut TestAppContext) {
3901 init_test_settings(cx);
3902
3903 let project = create_test_project(
3904 cx,
3905 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3906 )
3907 .await;
3908
3909 let (_workspace, _thread_store, thread, _context_store, model) =
3910 setup_test_environment(cx, project.clone()).await;
3911
3912 // Both model and provider
3913 cx.update(|cx| {
3914 AgentSettings::override_global(
3915 AgentSettings {
3916 model_parameters: vec![LanguageModelParameters {
3917 provider: Some(model.provider_id().0.to_string().into()),
3918 model: Some(model.id().0.clone()),
3919 temperature: Some(0.66),
3920 }],
3921 ..AgentSettings::get_global(cx).clone()
3922 },
3923 cx,
3924 );
3925 });
3926
3927 let request = thread.update(cx, |thread, cx| {
3928 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3929 });
3930 assert_eq!(request.temperature, Some(0.66));
3931
3932 // Only model
3933 cx.update(|cx| {
3934 AgentSettings::override_global(
3935 AgentSettings {
3936 model_parameters: vec![LanguageModelParameters {
3937 provider: None,
3938 model: Some(model.id().0.clone()),
3939 temperature: Some(0.66),
3940 }],
3941 ..AgentSettings::get_global(cx).clone()
3942 },
3943 cx,
3944 );
3945 });
3946
3947 let request = thread.update(cx, |thread, cx| {
3948 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3949 });
3950 assert_eq!(request.temperature, Some(0.66));
3951
3952 // Only provider
3953 cx.update(|cx| {
3954 AgentSettings::override_global(
3955 AgentSettings {
3956 model_parameters: vec![LanguageModelParameters {
3957 provider: Some(model.provider_id().0.to_string().into()),
3958 model: None,
3959 temperature: Some(0.66),
3960 }],
3961 ..AgentSettings::get_global(cx).clone()
3962 },
3963 cx,
3964 );
3965 });
3966
3967 let request = thread.update(cx, |thread, cx| {
3968 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3969 });
3970 assert_eq!(request.temperature, Some(0.66));
3971
3972 // Same model name, different provider
3973 cx.update(|cx| {
3974 AgentSettings::override_global(
3975 AgentSettings {
3976 model_parameters: vec![LanguageModelParameters {
3977 provider: Some("anthropic".into()),
3978 model: Some(model.id().0.clone()),
3979 temperature: Some(0.66),
3980 }],
3981 ..AgentSettings::get_global(cx).clone()
3982 },
3983 cx,
3984 );
3985 });
3986
3987 let request = thread.update(cx, |thread, cx| {
3988 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3989 });
3990 assert_eq!(request.temperature, None);
3991 }
3992
3993 #[gpui::test]
3994 async fn test_thread_summary(cx: &mut TestAppContext) {
3995 init_test_settings(cx);
3996
3997 let project = create_test_project(cx, json!({})).await;
3998
3999 let (_, _thread_store, thread, _context_store, model) =
4000 setup_test_environment(cx, project.clone()).await;
4001
4002 // Initial state should be pending
4003 thread.read_with(cx, |thread, _| {
4004 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4005 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4006 });
4007
4008 // Manually setting the summary should not be allowed in this state
4009 thread.update(cx, |thread, cx| {
4010 thread.set_summary("This should not work", cx);
4011 });
4012
4013 thread.read_with(cx, |thread, _| {
4014 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4015 });
4016
4017 // Send a message
4018 thread.update(cx, |thread, cx| {
4019 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
4020 thread.send_to_model(
4021 model.clone(),
4022 CompletionIntent::ThreadSummarization,
4023 None,
4024 cx,
4025 );
4026 });
4027
4028 let fake_model = model.as_fake();
4029 simulate_successful_response(&fake_model, cx);
4030
4031 // Should start generating summary when there are >= 2 messages
4032 thread.read_with(cx, |thread, _| {
4033 assert_eq!(*thread.summary(), ThreadSummary::Generating);
4034 });
4035
4036 // Should not be able to set the summary while generating
4037 thread.update(cx, |thread, cx| {
4038 thread.set_summary("This should not work either", cx);
4039 });
4040
4041 thread.read_with(cx, |thread, _| {
4042 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4043 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4044 });
4045
4046 cx.run_until_parked();
4047 fake_model.stream_last_completion_response("Brief");
4048 fake_model.stream_last_completion_response(" Introduction");
4049 fake_model.end_last_completion_stream();
4050 cx.run_until_parked();
4051
4052 // Summary should be set
4053 thread.read_with(cx, |thread, _| {
4054 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4055 assert_eq!(thread.summary().or_default(), "Brief Introduction");
4056 });
4057
4058 // Now we should be able to set a summary
4059 thread.update(cx, |thread, cx| {
4060 thread.set_summary("Brief Intro", cx);
4061 });
4062
4063 thread.read_with(cx, |thread, _| {
4064 assert_eq!(thread.summary().or_default(), "Brief Intro");
4065 });
4066
4067 // Test setting an empty summary (should default to DEFAULT)
4068 thread.update(cx, |thread, cx| {
4069 thread.set_summary("", cx);
4070 });
4071
4072 thread.read_with(cx, |thread, _| {
4073 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4074 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4075 });
4076 }
4077
4078 #[gpui::test]
4079 async fn test_thread_summary_error_set_manually(cx: &mut TestAppContext) {
4080 init_test_settings(cx);
4081
4082 let project = create_test_project(cx, json!({})).await;
4083
4084 let (_, _thread_store, thread, _context_store, model) =
4085 setup_test_environment(cx, project.clone()).await;
4086
4087 test_summarize_error(&model, &thread, cx);
4088
4089 // Now we should be able to set a summary
4090 thread.update(cx, |thread, cx| {
4091 thread.set_summary("Brief Intro", cx);
4092 });
4093
4094 thread.read_with(cx, |thread, _| {
4095 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4096 assert_eq!(thread.summary().or_default(), "Brief Intro");
4097 });
4098 }
4099
4100 #[gpui::test]
4101 async fn test_thread_summary_error_retry(cx: &mut TestAppContext) {
4102 init_test_settings(cx);
4103
4104 let project = create_test_project(cx, json!({})).await;
4105
4106 let (_, _thread_store, thread, _context_store, model) =
4107 setup_test_environment(cx, project.clone()).await;
4108
4109 test_summarize_error(&model, &thread, cx);
4110
4111 // Sending another message should not trigger another summarize request
4112 thread.update(cx, |thread, cx| {
4113 thread.insert_user_message(
4114 "How are you?",
4115 ContextLoadResult::default(),
4116 None,
4117 vec![],
4118 cx,
4119 );
4120 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4121 });
4122
4123 let fake_model = model.as_fake();
4124 simulate_successful_response(&fake_model, cx);
4125
4126 thread.read_with(cx, |thread, _| {
4127 // State is still Error, not Generating
4128 assert!(matches!(thread.summary(), ThreadSummary::Error));
4129 });
4130
4131 // But the summarize request can be invoked manually
4132 thread.update(cx, |thread, cx| {
4133 thread.summarize(cx);
4134 });
4135
4136 thread.read_with(cx, |thread, _| {
4137 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4138 });
4139
4140 cx.run_until_parked();
4141 fake_model.stream_last_completion_response("A successful summary");
4142 fake_model.end_last_completion_stream();
4143 cx.run_until_parked();
4144
4145 thread.read_with(cx, |thread, _| {
4146 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4147 assert_eq!(thread.summary().or_default(), "A successful summary");
4148 });
4149 }
4150
4151 // Helper to create a model that returns errors
4152 enum TestError {
4153 Overloaded,
4154 InternalServerError,
4155 }
4156
4157 struct ErrorInjector {
4158 inner: Arc<FakeLanguageModel>,
4159 error_type: TestError,
4160 }
4161
4162 impl ErrorInjector {
4163 fn new(error_type: TestError) -> Self {
4164 Self {
4165 inner: Arc::new(FakeLanguageModel::default()),
4166 error_type,
4167 }
4168 }
4169 }
4170
4171 impl LanguageModel for ErrorInjector {
4172 fn id(&self) -> LanguageModelId {
4173 self.inner.id()
4174 }
4175
4176 fn name(&self) -> LanguageModelName {
4177 self.inner.name()
4178 }
4179
4180 fn provider_id(&self) -> LanguageModelProviderId {
4181 self.inner.provider_id()
4182 }
4183
4184 fn provider_name(&self) -> LanguageModelProviderName {
4185 self.inner.provider_name()
4186 }
4187
4188 fn supports_tools(&self) -> bool {
4189 self.inner.supports_tools()
4190 }
4191
4192 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4193 self.inner.supports_tool_choice(choice)
4194 }
4195
4196 fn supports_images(&self) -> bool {
4197 self.inner.supports_images()
4198 }
4199
4200 fn telemetry_id(&self) -> String {
4201 self.inner.telemetry_id()
4202 }
4203
4204 fn max_token_count(&self) -> u64 {
4205 self.inner.max_token_count()
4206 }
4207
4208 fn count_tokens(
4209 &self,
4210 request: LanguageModelRequest,
4211 cx: &App,
4212 ) -> BoxFuture<'static, Result<u64>> {
4213 self.inner.count_tokens(request, cx)
4214 }
4215
4216 fn stream_completion(
4217 &self,
4218 _request: LanguageModelRequest,
4219 _cx: &AsyncApp,
4220 ) -> BoxFuture<
4221 'static,
4222 Result<
4223 BoxStream<
4224 'static,
4225 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4226 >,
4227 LanguageModelCompletionError,
4228 >,
4229 > {
4230 let error = match self.error_type {
4231 TestError::Overloaded => LanguageModelCompletionError::ServerOverloaded {
4232 provider: self.provider_name(),
4233 retry_after: None,
4234 },
4235 TestError::InternalServerError => {
4236 LanguageModelCompletionError::ApiInternalServerError {
4237 provider: self.provider_name(),
4238 message: "I'm a teapot orbiting the sun".to_string(),
4239 }
4240 }
4241 };
4242 async move {
4243 let stream = futures::stream::once(async move { Err(error) });
4244 Ok(stream.boxed())
4245 }
4246 .boxed()
4247 }
4248
4249 fn as_fake(&self) -> &FakeLanguageModel {
4250 &self.inner
4251 }
4252 }
4253
4254 #[gpui::test]
4255 async fn test_retry_on_overloaded_error(cx: &mut TestAppContext) {
4256 init_test_settings(cx);
4257
4258 let project = create_test_project(cx, json!({})).await;
4259 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4260
4261 // Enable Burn Mode to allow retries
4262 thread.update(cx, |thread, _| {
4263 thread.set_completion_mode(CompletionMode::Burn);
4264 });
4265
4266 // Create model that returns overloaded error
4267 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4268
4269 // Insert a user message
4270 thread.update(cx, |thread, cx| {
4271 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4272 });
4273
4274 // Start completion
4275 thread.update(cx, |thread, cx| {
4276 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4277 });
4278
4279 cx.run_until_parked();
4280
4281 thread.read_with(cx, |thread, _| {
4282 assert!(thread.retry_state.is_some(), "Should have retry state");
4283 let retry_state = thread.retry_state.as_ref().unwrap();
4284 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4285 assert_eq!(
4286 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
4287 "Should retry MAX_RETRY_ATTEMPTS times for overloaded errors"
4288 );
4289 });
4290
4291 // Check that a retry message was added
4292 thread.read_with(cx, |thread, _| {
4293 let mut messages = thread.messages();
4294 assert!(
4295 messages.any(|msg| {
4296 msg.role == Role::System
4297 && msg.ui_only
4298 && msg.segments.iter().any(|seg| {
4299 if let MessageSegment::Text(text) = seg {
4300 text.contains("overloaded")
4301 && text
4302 .contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS))
4303 } else {
4304 false
4305 }
4306 })
4307 }),
4308 "Should have added a system retry message"
4309 );
4310 });
4311
4312 let retry_count = thread.update(cx, |thread, _| {
4313 thread
4314 .messages
4315 .iter()
4316 .filter(|m| {
4317 m.ui_only
4318 && m.segments.iter().any(|s| {
4319 if let MessageSegment::Text(text) = s {
4320 text.contains("Retrying") && text.contains("seconds")
4321 } else {
4322 false
4323 }
4324 })
4325 })
4326 .count()
4327 });
4328
4329 assert_eq!(retry_count, 1, "Should have one retry message");
4330 }
4331
4332 #[gpui::test]
4333 async fn test_retry_on_internal_server_error(cx: &mut TestAppContext) {
4334 init_test_settings(cx);
4335
4336 let project = create_test_project(cx, json!({})).await;
4337 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4338
4339 // Enable Burn Mode to allow retries
4340 thread.update(cx, |thread, _| {
4341 thread.set_completion_mode(CompletionMode::Burn);
4342 });
4343
4344 // Create model that returns internal server error
4345 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4346
4347 // Insert a user message
4348 thread.update(cx, |thread, cx| {
4349 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4350 });
4351
4352 // Start completion
4353 thread.update(cx, |thread, cx| {
4354 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4355 });
4356
4357 cx.run_until_parked();
4358
4359 // Check retry state on thread
4360 thread.read_with(cx, |thread, _| {
4361 assert!(thread.retry_state.is_some(), "Should have retry state");
4362 let retry_state = thread.retry_state.as_ref().unwrap();
4363 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4364 assert_eq!(
4365 retry_state.max_attempts, 3,
4366 "Should have correct max attempts"
4367 );
4368 });
4369
4370 // Check that a retry message was added with provider name
4371 thread.read_with(cx, |thread, _| {
4372 let mut messages = thread.messages();
4373 assert!(
4374 messages.any(|msg| {
4375 msg.role == Role::System
4376 && msg.ui_only
4377 && msg.segments.iter().any(|seg| {
4378 if let MessageSegment::Text(text) = seg {
4379 text.contains("internal")
4380 && text.contains("Fake")
4381 && text.contains("Retrying")
4382 && text.contains("attempt 1 of 3")
4383 && text.contains("seconds")
4384 } else {
4385 false
4386 }
4387 })
4388 }),
4389 "Should have added a system retry message with provider name"
4390 );
4391 });
4392
4393 // Count retry messages
4394 let retry_count = thread.update(cx, |thread, _| {
4395 thread
4396 .messages
4397 .iter()
4398 .filter(|m| {
4399 m.ui_only
4400 && m.segments.iter().any(|s| {
4401 if let MessageSegment::Text(text) = s {
4402 text.contains("Retrying") && text.contains("seconds")
4403 } else {
4404 false
4405 }
4406 })
4407 })
4408 .count()
4409 });
4410
4411 assert_eq!(retry_count, 1, "Should have one retry message");
4412 }
4413
4414 #[gpui::test]
4415 async fn test_exponential_backoff_on_retries(cx: &mut TestAppContext) {
4416 init_test_settings(cx);
4417
4418 let project = create_test_project(cx, json!({})).await;
4419 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4420
4421 // Enable Burn Mode to allow retries
4422 thread.update(cx, |thread, _| {
4423 thread.set_completion_mode(CompletionMode::Burn);
4424 });
4425
4426 // Create model that returns internal server error
4427 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4428
4429 // Insert a user message
4430 thread.update(cx, |thread, cx| {
4431 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4432 });
4433
4434 // Track retry events and completion count
4435 // Track completion events
4436 let completion_count = Arc::new(Mutex::new(0));
4437 let completion_count_clone = completion_count.clone();
4438
4439 let _subscription = thread.update(cx, |_, cx| {
4440 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4441 if let ThreadEvent::NewRequest = event {
4442 *completion_count_clone.lock() += 1;
4443 }
4444 })
4445 });
4446
4447 // First attempt
4448 thread.update(cx, |thread, cx| {
4449 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4450 });
4451 cx.run_until_parked();
4452
4453 // Should have scheduled first retry - count retry messages
4454 let retry_count = thread.update(cx, |thread, _| {
4455 thread
4456 .messages
4457 .iter()
4458 .filter(|m| {
4459 m.ui_only
4460 && m.segments.iter().any(|s| {
4461 if let MessageSegment::Text(text) = s {
4462 text.contains("Retrying") && text.contains("seconds")
4463 } else {
4464 false
4465 }
4466 })
4467 })
4468 .count()
4469 });
4470 assert_eq!(retry_count, 1, "Should have scheduled first retry");
4471
4472 // Check retry state
4473 thread.read_with(cx, |thread, _| {
4474 assert!(thread.retry_state.is_some(), "Should have retry state");
4475 let retry_state = thread.retry_state.as_ref().unwrap();
4476 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4477 assert_eq!(
4478 retry_state.max_attempts, 3,
4479 "Internal server errors should retry up to 3 times"
4480 );
4481 });
4482
4483 // Advance clock for first retry
4484 cx.executor().advance_clock(BASE_RETRY_DELAY);
4485 cx.run_until_parked();
4486
4487 // Advance clock for second retry
4488 cx.executor().advance_clock(BASE_RETRY_DELAY);
4489 cx.run_until_parked();
4490
4491 // Advance clock for third retry
4492 cx.executor().advance_clock(BASE_RETRY_DELAY);
4493 cx.run_until_parked();
4494
4495 // Should have completed all retries - count retry messages
4496 let retry_count = thread.update(cx, |thread, _| {
4497 thread
4498 .messages
4499 .iter()
4500 .filter(|m| {
4501 m.ui_only
4502 && m.segments.iter().any(|s| {
4503 if let MessageSegment::Text(text) = s {
4504 text.contains("Retrying") && text.contains("seconds")
4505 } else {
4506 false
4507 }
4508 })
4509 })
4510 .count()
4511 });
4512 assert_eq!(
4513 retry_count, 3,
4514 "Should have 3 retries for internal server errors"
4515 );
4516
4517 // For internal server errors, we retry 3 times and then give up
4518 // Check that retry_state is cleared after all retries
4519 thread.read_with(cx, |thread, _| {
4520 assert!(
4521 thread.retry_state.is_none(),
4522 "Retry state should be cleared after all retries"
4523 );
4524 });
4525
4526 // Verify total attempts (1 initial + 3 retries)
4527 assert_eq!(
4528 *completion_count.lock(),
4529 4,
4530 "Should have attempted once plus 3 retries"
4531 );
4532 }
4533
4534 #[gpui::test]
4535 async fn test_max_retries_exceeded(cx: &mut TestAppContext) {
4536 init_test_settings(cx);
4537
4538 let project = create_test_project(cx, json!({})).await;
4539 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4540
4541 // Enable Burn Mode to allow retries
4542 thread.update(cx, |thread, _| {
4543 thread.set_completion_mode(CompletionMode::Burn);
4544 });
4545
4546 // Create model that returns overloaded error
4547 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4548
4549 // Insert a user message
4550 thread.update(cx, |thread, cx| {
4551 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4552 });
4553
4554 // Track events
4555 let stopped_with_error = Arc::new(Mutex::new(false));
4556 let stopped_with_error_clone = stopped_with_error.clone();
4557
4558 let _subscription = thread.update(cx, |_, cx| {
4559 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4560 if let ThreadEvent::Stopped(Err(_)) = event {
4561 *stopped_with_error_clone.lock() = true;
4562 }
4563 })
4564 });
4565
4566 // Start initial completion
4567 thread.update(cx, |thread, cx| {
4568 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4569 });
4570 cx.run_until_parked();
4571
4572 // Advance through all retries
4573 for _ in 0..MAX_RETRY_ATTEMPTS {
4574 cx.executor().advance_clock(BASE_RETRY_DELAY);
4575 cx.run_until_parked();
4576 }
4577
4578 let retry_count = thread.update(cx, |thread, _| {
4579 thread
4580 .messages
4581 .iter()
4582 .filter(|m| {
4583 m.ui_only
4584 && m.segments.iter().any(|s| {
4585 if let MessageSegment::Text(text) = s {
4586 text.contains("Retrying") && text.contains("seconds")
4587 } else {
4588 false
4589 }
4590 })
4591 })
4592 .count()
4593 });
4594
4595 // After max retries, should emit Stopped(Err(...)) event
4596 assert_eq!(
4597 retry_count, MAX_RETRY_ATTEMPTS as usize,
4598 "Should have attempted MAX_RETRY_ATTEMPTS retries for overloaded errors"
4599 );
4600 assert!(
4601 *stopped_with_error.lock(),
4602 "Should emit Stopped(Err(...)) event after max retries exceeded"
4603 );
4604
4605 // Retry state should be cleared
4606 thread.read_with(cx, |thread, _| {
4607 assert!(
4608 thread.retry_state.is_none(),
4609 "Retry state should be cleared after max retries"
4610 );
4611
4612 // Verify we have the expected number of retry messages
4613 let retry_messages = thread
4614 .messages
4615 .iter()
4616 .filter(|msg| msg.ui_only && msg.role == Role::System)
4617 .count();
4618 assert_eq!(
4619 retry_messages, MAX_RETRY_ATTEMPTS as usize,
4620 "Should have MAX_RETRY_ATTEMPTS retry messages for overloaded errors"
4621 );
4622 });
4623 }
4624
4625 #[gpui::test]
4626 async fn test_retry_message_removed_on_retry(cx: &mut TestAppContext) {
4627 init_test_settings(cx);
4628
4629 let project = create_test_project(cx, json!({})).await;
4630 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4631
4632 // Enable Burn Mode to allow retries
4633 thread.update(cx, |thread, _| {
4634 thread.set_completion_mode(CompletionMode::Burn);
4635 });
4636
4637 // We'll use a wrapper to switch behavior after first failure
4638 struct RetryTestModel {
4639 inner: Arc<FakeLanguageModel>,
4640 failed_once: Arc<Mutex<bool>>,
4641 }
4642
4643 impl LanguageModel for RetryTestModel {
4644 fn id(&self) -> LanguageModelId {
4645 self.inner.id()
4646 }
4647
4648 fn name(&self) -> LanguageModelName {
4649 self.inner.name()
4650 }
4651
4652 fn provider_id(&self) -> LanguageModelProviderId {
4653 self.inner.provider_id()
4654 }
4655
4656 fn provider_name(&self) -> LanguageModelProviderName {
4657 self.inner.provider_name()
4658 }
4659
4660 fn supports_tools(&self) -> bool {
4661 self.inner.supports_tools()
4662 }
4663
4664 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4665 self.inner.supports_tool_choice(choice)
4666 }
4667
4668 fn supports_images(&self) -> bool {
4669 self.inner.supports_images()
4670 }
4671
4672 fn telemetry_id(&self) -> String {
4673 self.inner.telemetry_id()
4674 }
4675
4676 fn max_token_count(&self) -> u64 {
4677 self.inner.max_token_count()
4678 }
4679
4680 fn count_tokens(
4681 &self,
4682 request: LanguageModelRequest,
4683 cx: &App,
4684 ) -> BoxFuture<'static, Result<u64>> {
4685 self.inner.count_tokens(request, cx)
4686 }
4687
4688 fn stream_completion(
4689 &self,
4690 request: LanguageModelRequest,
4691 cx: &AsyncApp,
4692 ) -> BoxFuture<
4693 'static,
4694 Result<
4695 BoxStream<
4696 'static,
4697 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4698 >,
4699 LanguageModelCompletionError,
4700 >,
4701 > {
4702 if !*self.failed_once.lock() {
4703 *self.failed_once.lock() = true;
4704 let provider = self.provider_name();
4705 // Return error on first attempt
4706 let stream = futures::stream::once(async move {
4707 Err(LanguageModelCompletionError::ServerOverloaded {
4708 provider,
4709 retry_after: None,
4710 })
4711 });
4712 async move { Ok(stream.boxed()) }.boxed()
4713 } else {
4714 // Succeed on retry
4715 self.inner.stream_completion(request, cx)
4716 }
4717 }
4718
4719 fn as_fake(&self) -> &FakeLanguageModel {
4720 &self.inner
4721 }
4722 }
4723
4724 let model = Arc::new(RetryTestModel {
4725 inner: Arc::new(FakeLanguageModel::default()),
4726 failed_once: Arc::new(Mutex::new(false)),
4727 });
4728
4729 // Insert a user message
4730 thread.update(cx, |thread, cx| {
4731 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4732 });
4733
4734 // Track message deletions
4735 // Track when retry completes successfully
4736 let retry_completed = Arc::new(Mutex::new(false));
4737 let retry_completed_clone = retry_completed.clone();
4738
4739 let _subscription = thread.update(cx, |_, cx| {
4740 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4741 if let ThreadEvent::StreamedCompletion = event {
4742 *retry_completed_clone.lock() = true;
4743 }
4744 })
4745 });
4746
4747 // Start completion
4748 thread.update(cx, |thread, cx| {
4749 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4750 });
4751 cx.run_until_parked();
4752
4753 // Get the retry message ID
4754 let retry_message_id = thread.read_with(cx, |thread, _| {
4755 thread
4756 .messages()
4757 .find(|msg| msg.role == Role::System && msg.ui_only)
4758 .map(|msg| msg.id)
4759 .expect("Should have a retry message")
4760 });
4761
4762 // Wait for retry
4763 cx.executor().advance_clock(BASE_RETRY_DELAY);
4764 cx.run_until_parked();
4765
4766 // Stream some successful content
4767 let fake_model = model.as_fake();
4768 // After the retry, there should be a new pending completion
4769 let pending = fake_model.pending_completions();
4770 assert!(
4771 !pending.is_empty(),
4772 "Should have a pending completion after retry"
4773 );
4774 fake_model.stream_completion_response(&pending[0], "Success!");
4775 fake_model.end_completion_stream(&pending[0]);
4776 cx.run_until_parked();
4777
4778 // Check that the retry completed successfully
4779 assert!(
4780 *retry_completed.lock(),
4781 "Retry should have completed successfully"
4782 );
4783
4784 // Retry message should still exist but be marked as ui_only
4785 thread.read_with(cx, |thread, _| {
4786 let retry_msg = thread
4787 .message(retry_message_id)
4788 .expect("Retry message should still exist");
4789 assert!(retry_msg.ui_only, "Retry message should be ui_only");
4790 assert_eq!(
4791 retry_msg.role,
4792 Role::System,
4793 "Retry message should have System role"
4794 );
4795 });
4796 }
4797
4798 #[gpui::test]
4799 async fn test_successful_completion_clears_retry_state(cx: &mut TestAppContext) {
4800 init_test_settings(cx);
4801
4802 let project = create_test_project(cx, json!({})).await;
4803 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4804
4805 // Enable Burn Mode to allow retries
4806 thread.update(cx, |thread, _| {
4807 thread.set_completion_mode(CompletionMode::Burn);
4808 });
4809
4810 // Create a model that fails once then succeeds
4811 struct FailOnceModel {
4812 inner: Arc<FakeLanguageModel>,
4813 failed_once: Arc<Mutex<bool>>,
4814 }
4815
4816 impl LanguageModel for FailOnceModel {
4817 fn id(&self) -> LanguageModelId {
4818 self.inner.id()
4819 }
4820
4821 fn name(&self) -> LanguageModelName {
4822 self.inner.name()
4823 }
4824
4825 fn provider_id(&self) -> LanguageModelProviderId {
4826 self.inner.provider_id()
4827 }
4828
4829 fn provider_name(&self) -> LanguageModelProviderName {
4830 self.inner.provider_name()
4831 }
4832
4833 fn supports_tools(&self) -> bool {
4834 self.inner.supports_tools()
4835 }
4836
4837 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4838 self.inner.supports_tool_choice(choice)
4839 }
4840
4841 fn supports_images(&self) -> bool {
4842 self.inner.supports_images()
4843 }
4844
4845 fn telemetry_id(&self) -> String {
4846 self.inner.telemetry_id()
4847 }
4848
4849 fn max_token_count(&self) -> u64 {
4850 self.inner.max_token_count()
4851 }
4852
4853 fn count_tokens(
4854 &self,
4855 request: LanguageModelRequest,
4856 cx: &App,
4857 ) -> BoxFuture<'static, Result<u64>> {
4858 self.inner.count_tokens(request, cx)
4859 }
4860
4861 fn stream_completion(
4862 &self,
4863 request: LanguageModelRequest,
4864 cx: &AsyncApp,
4865 ) -> BoxFuture<
4866 'static,
4867 Result<
4868 BoxStream<
4869 'static,
4870 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4871 >,
4872 LanguageModelCompletionError,
4873 >,
4874 > {
4875 if !*self.failed_once.lock() {
4876 *self.failed_once.lock() = true;
4877 let provider = self.provider_name();
4878 // Return error on first attempt
4879 let stream = futures::stream::once(async move {
4880 Err(LanguageModelCompletionError::ServerOverloaded {
4881 provider,
4882 retry_after: None,
4883 })
4884 });
4885 async move { Ok(stream.boxed()) }.boxed()
4886 } else {
4887 // Succeed on retry
4888 self.inner.stream_completion(request, cx)
4889 }
4890 }
4891 }
4892
4893 let fail_once_model = Arc::new(FailOnceModel {
4894 inner: Arc::new(FakeLanguageModel::default()),
4895 failed_once: Arc::new(Mutex::new(false)),
4896 });
4897
4898 // Insert a user message
4899 thread.update(cx, |thread, cx| {
4900 thread.insert_user_message(
4901 "Test message",
4902 ContextLoadResult::default(),
4903 None,
4904 vec![],
4905 cx,
4906 );
4907 });
4908
4909 // Start completion with fail-once model
4910 thread.update(cx, |thread, cx| {
4911 thread.send_to_model(
4912 fail_once_model.clone(),
4913 CompletionIntent::UserPrompt,
4914 None,
4915 cx,
4916 );
4917 });
4918
4919 cx.run_until_parked();
4920
4921 // Verify retry state exists after first failure
4922 thread.read_with(cx, |thread, _| {
4923 assert!(
4924 thread.retry_state.is_some(),
4925 "Should have retry state after failure"
4926 );
4927 });
4928
4929 // Wait for retry delay
4930 cx.executor().advance_clock(BASE_RETRY_DELAY);
4931 cx.run_until_parked();
4932
4933 // The retry should now use our FailOnceModel which should succeed
4934 // We need to help the FakeLanguageModel complete the stream
4935 let inner_fake = fail_once_model.inner.clone();
4936
4937 // Wait a bit for the retry to start
4938 cx.run_until_parked();
4939
4940 // Check for pending completions and complete them
4941 if let Some(pending) = inner_fake.pending_completions().first() {
4942 inner_fake.stream_completion_response(pending, "Success!");
4943 inner_fake.end_completion_stream(pending);
4944 }
4945 cx.run_until_parked();
4946
4947 thread.read_with(cx, |thread, _| {
4948 assert!(
4949 thread.retry_state.is_none(),
4950 "Retry state should be cleared after successful completion"
4951 );
4952
4953 let has_assistant_message = thread
4954 .messages
4955 .iter()
4956 .any(|msg| msg.role == Role::Assistant && !msg.ui_only);
4957 assert!(
4958 has_assistant_message,
4959 "Should have an assistant message after successful retry"
4960 );
4961 });
4962 }
4963
4964 #[gpui::test]
4965 async fn test_rate_limit_retry_single_attempt(cx: &mut TestAppContext) {
4966 init_test_settings(cx);
4967
4968 let project = create_test_project(cx, json!({})).await;
4969 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4970
4971 // Enable Burn Mode to allow retries
4972 thread.update(cx, |thread, _| {
4973 thread.set_completion_mode(CompletionMode::Burn);
4974 });
4975
4976 // Create a model that returns rate limit error with retry_after
4977 struct RateLimitModel {
4978 inner: Arc<FakeLanguageModel>,
4979 }
4980
4981 impl LanguageModel for RateLimitModel {
4982 fn id(&self) -> LanguageModelId {
4983 self.inner.id()
4984 }
4985
4986 fn name(&self) -> LanguageModelName {
4987 self.inner.name()
4988 }
4989
4990 fn provider_id(&self) -> LanguageModelProviderId {
4991 self.inner.provider_id()
4992 }
4993
4994 fn provider_name(&self) -> LanguageModelProviderName {
4995 self.inner.provider_name()
4996 }
4997
4998 fn supports_tools(&self) -> bool {
4999 self.inner.supports_tools()
5000 }
5001
5002 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
5003 self.inner.supports_tool_choice(choice)
5004 }
5005
5006 fn supports_images(&self) -> bool {
5007 self.inner.supports_images()
5008 }
5009
5010 fn telemetry_id(&self) -> String {
5011 self.inner.telemetry_id()
5012 }
5013
5014 fn max_token_count(&self) -> u64 {
5015 self.inner.max_token_count()
5016 }
5017
5018 fn count_tokens(
5019 &self,
5020 request: LanguageModelRequest,
5021 cx: &App,
5022 ) -> BoxFuture<'static, Result<u64>> {
5023 self.inner.count_tokens(request, cx)
5024 }
5025
5026 fn stream_completion(
5027 &self,
5028 _request: LanguageModelRequest,
5029 _cx: &AsyncApp,
5030 ) -> BoxFuture<
5031 'static,
5032 Result<
5033 BoxStream<
5034 'static,
5035 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
5036 >,
5037 LanguageModelCompletionError,
5038 >,
5039 > {
5040 let provider = self.provider_name();
5041 async move {
5042 let stream = futures::stream::once(async move {
5043 Err(LanguageModelCompletionError::RateLimitExceeded {
5044 provider,
5045 retry_after: Some(Duration::from_secs(TEST_RATE_LIMIT_RETRY_SECS)),
5046 })
5047 });
5048 Ok(stream.boxed())
5049 }
5050 .boxed()
5051 }
5052
5053 fn as_fake(&self) -> &FakeLanguageModel {
5054 &self.inner
5055 }
5056 }
5057
5058 let model = Arc::new(RateLimitModel {
5059 inner: Arc::new(FakeLanguageModel::default()),
5060 });
5061
5062 // Insert a user message
5063 thread.update(cx, |thread, cx| {
5064 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5065 });
5066
5067 // Start completion
5068 thread.update(cx, |thread, cx| {
5069 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5070 });
5071
5072 cx.run_until_parked();
5073
5074 let retry_count = thread.update(cx, |thread, _| {
5075 thread
5076 .messages
5077 .iter()
5078 .filter(|m| {
5079 m.ui_only
5080 && m.segments.iter().any(|s| {
5081 if let MessageSegment::Text(text) = s {
5082 text.contains("rate limit exceeded")
5083 } else {
5084 false
5085 }
5086 })
5087 })
5088 .count()
5089 });
5090 assert_eq!(retry_count, 1, "Should have scheduled one retry");
5091
5092 thread.read_with(cx, |thread, _| {
5093 assert!(
5094 thread.retry_state.is_some(),
5095 "Rate limit errors should set retry_state"
5096 );
5097 if let Some(retry_state) = &thread.retry_state {
5098 assert_eq!(
5099 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
5100 "Rate limit errors should use MAX_RETRY_ATTEMPTS"
5101 );
5102 }
5103 });
5104
5105 // Verify we have one retry message
5106 thread.read_with(cx, |thread, _| {
5107 let retry_messages = thread
5108 .messages
5109 .iter()
5110 .filter(|msg| {
5111 msg.ui_only
5112 && msg.segments.iter().any(|seg| {
5113 if let MessageSegment::Text(text) = seg {
5114 text.contains("rate limit exceeded")
5115 } else {
5116 false
5117 }
5118 })
5119 })
5120 .count();
5121 assert_eq!(
5122 retry_messages, 1,
5123 "Should have one rate limit retry message"
5124 );
5125 });
5126
5127 // Check that retry message doesn't include attempt count
5128 thread.read_with(cx, |thread, _| {
5129 let retry_message = thread
5130 .messages
5131 .iter()
5132 .find(|msg| msg.role == Role::System && msg.ui_only)
5133 .expect("Should have a retry message");
5134
5135 // Check that the message contains attempt count since we use retry_state
5136 if let Some(MessageSegment::Text(text)) = retry_message.segments.first() {
5137 assert!(
5138 text.contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS)),
5139 "Rate limit retry message should contain attempt count with MAX_RETRY_ATTEMPTS"
5140 );
5141 assert!(
5142 text.contains("Retrying"),
5143 "Rate limit retry message should contain retry text"
5144 );
5145 }
5146 });
5147 }
5148
5149 #[gpui::test]
5150 async fn test_ui_only_messages_not_sent_to_model(cx: &mut TestAppContext) {
5151 init_test_settings(cx);
5152
5153 let project = create_test_project(cx, json!({})).await;
5154 let (_, _, thread, _, model) = setup_test_environment(cx, project.clone()).await;
5155
5156 // Insert a regular user message
5157 thread.update(cx, |thread, cx| {
5158 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5159 });
5160
5161 // Insert a UI-only message (like our retry notifications)
5162 thread.update(cx, |thread, cx| {
5163 let id = thread.next_message_id.post_inc();
5164 thread.messages.push(Message {
5165 id,
5166 role: Role::System,
5167 segments: vec![MessageSegment::Text(
5168 "This is a UI-only message that should not be sent to the model".to_string(),
5169 )],
5170 loaded_context: LoadedContext::default(),
5171 creases: Vec::new(),
5172 is_hidden: true,
5173 ui_only: true,
5174 });
5175 cx.emit(ThreadEvent::MessageAdded(id));
5176 });
5177
5178 // Insert another regular message
5179 thread.update(cx, |thread, cx| {
5180 thread.insert_user_message(
5181 "How are you?",
5182 ContextLoadResult::default(),
5183 None,
5184 vec![],
5185 cx,
5186 );
5187 });
5188
5189 // Generate the completion request
5190 let request = thread.update(cx, |thread, cx| {
5191 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
5192 });
5193
5194 // Verify that the request only contains non-UI-only messages
5195 // Should have system prompt + 2 user messages, but not the UI-only message
5196 let user_messages: Vec<_> = request
5197 .messages
5198 .iter()
5199 .filter(|msg| msg.role == Role::User)
5200 .collect();
5201 assert_eq!(
5202 user_messages.len(),
5203 2,
5204 "Should have exactly 2 user messages"
5205 );
5206
5207 // Verify the UI-only content is not present anywhere in the request
5208 let request_text = request
5209 .messages
5210 .iter()
5211 .flat_map(|msg| &msg.content)
5212 .filter_map(|content| match content {
5213 MessageContent::Text(text) => Some(text.as_str()),
5214 _ => None,
5215 })
5216 .collect::<String>();
5217
5218 assert!(
5219 !request_text.contains("UI-only message"),
5220 "UI-only message content should not be in the request"
5221 );
5222
5223 // Verify the thread still has all 3 messages (including UI-only)
5224 thread.read_with(cx, |thread, _| {
5225 assert_eq!(
5226 thread.messages().count(),
5227 3,
5228 "Thread should have 3 messages"
5229 );
5230 assert_eq!(
5231 thread.messages().filter(|m| m.ui_only).count(),
5232 1,
5233 "Thread should have 1 UI-only message"
5234 );
5235 });
5236
5237 // Verify that UI-only messages are not serialized
5238 let serialized = thread
5239 .update(cx, |thread, cx| thread.serialize(cx))
5240 .await
5241 .unwrap();
5242 assert_eq!(
5243 serialized.messages.len(),
5244 2,
5245 "Serialized thread should only have 2 messages (no UI-only)"
5246 );
5247 }
5248
5249 #[gpui::test]
5250 async fn test_no_retry_without_burn_mode(cx: &mut TestAppContext) {
5251 init_test_settings(cx);
5252
5253 let project = create_test_project(cx, json!({})).await;
5254 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5255
5256 // Ensure we're in Normal mode (not Burn mode)
5257 thread.update(cx, |thread, _| {
5258 thread.set_completion_mode(CompletionMode::Normal);
5259 });
5260
5261 // Track error events
5262 let error_events = Arc::new(Mutex::new(Vec::new()));
5263 let error_events_clone = error_events.clone();
5264
5265 let _subscription = thread.update(cx, |_, cx| {
5266 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
5267 if let ThreadEvent::ShowError(error) = event {
5268 error_events_clone.lock().push(error.clone());
5269 }
5270 })
5271 });
5272
5273 // Create model that returns overloaded error
5274 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5275
5276 // Insert a user message
5277 thread.update(cx, |thread, cx| {
5278 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5279 });
5280
5281 // Start completion
5282 thread.update(cx, |thread, cx| {
5283 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5284 });
5285
5286 cx.run_until_parked();
5287
5288 // Verify no retry state was created
5289 thread.read_with(cx, |thread, _| {
5290 assert!(
5291 thread.retry_state.is_none(),
5292 "Should not have retry state in Normal mode"
5293 );
5294 });
5295
5296 // Check that a retryable error was reported
5297 let errors = error_events.lock();
5298 assert!(!errors.is_empty(), "Should have received an error event");
5299
5300 if let ThreadError::RetryableError {
5301 message: _,
5302 can_enable_burn_mode,
5303 } = &errors[0]
5304 {
5305 assert!(
5306 *can_enable_burn_mode,
5307 "Error should indicate burn mode can be enabled"
5308 );
5309 } else {
5310 panic!("Expected RetryableError, got {:?}", errors[0]);
5311 }
5312
5313 // Verify the thread is no longer generating
5314 thread.read_with(cx, |thread, _| {
5315 assert!(
5316 !thread.is_generating(),
5317 "Should not be generating after error without retry"
5318 );
5319 });
5320 }
5321
5322 #[gpui::test]
5323 async fn test_retry_cancelled_on_stop(cx: &mut TestAppContext) {
5324 init_test_settings(cx);
5325
5326 let project = create_test_project(cx, json!({})).await;
5327 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5328
5329 // Enable Burn Mode to allow retries
5330 thread.update(cx, |thread, _| {
5331 thread.set_completion_mode(CompletionMode::Burn);
5332 });
5333
5334 // Create model that returns overloaded error
5335 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5336
5337 // Insert a user message
5338 thread.update(cx, |thread, cx| {
5339 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5340 });
5341
5342 // Start completion
5343 thread.update(cx, |thread, cx| {
5344 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5345 });
5346
5347 cx.run_until_parked();
5348
5349 // Verify retry was scheduled by checking for retry message
5350 let has_retry_message = thread.read_with(cx, |thread, _| {
5351 thread.messages.iter().any(|m| {
5352 m.ui_only
5353 && m.segments.iter().any(|s| {
5354 if let MessageSegment::Text(text) = s {
5355 text.contains("Retrying") && text.contains("seconds")
5356 } else {
5357 false
5358 }
5359 })
5360 })
5361 });
5362 assert!(has_retry_message, "Should have scheduled a retry");
5363
5364 // Cancel the completion before the retry happens
5365 thread.update(cx, |thread, cx| {
5366 thread.cancel_last_completion(None, cx);
5367 });
5368
5369 cx.run_until_parked();
5370
5371 // The retry should not have happened - no pending completions
5372 let fake_model = model.as_fake();
5373 assert_eq!(
5374 fake_model.pending_completions().len(),
5375 0,
5376 "Should have no pending completions after cancellation"
5377 );
5378
5379 // Verify the retry was cancelled by checking retry state
5380 thread.read_with(cx, |thread, _| {
5381 if let Some(retry_state) = &thread.retry_state {
5382 panic!(
5383 "retry_state should be cleared after cancellation, but found: attempt={}, max_attempts={}, intent={:?}",
5384 retry_state.attempt, retry_state.max_attempts, retry_state.intent
5385 );
5386 }
5387 });
5388 }
5389
5390 fn test_summarize_error(
5391 model: &Arc<dyn LanguageModel>,
5392 thread: &Entity<Thread>,
5393 cx: &mut TestAppContext,
5394 ) {
5395 thread.update(cx, |thread, cx| {
5396 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
5397 thread.send_to_model(
5398 model.clone(),
5399 CompletionIntent::ThreadSummarization,
5400 None,
5401 cx,
5402 );
5403 });
5404
5405 let fake_model = model.as_fake();
5406 simulate_successful_response(&fake_model, cx);
5407
5408 thread.read_with(cx, |thread, _| {
5409 assert!(matches!(thread.summary(), ThreadSummary::Generating));
5410 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5411 });
5412
5413 // Simulate summary request ending
5414 cx.run_until_parked();
5415 fake_model.end_last_completion_stream();
5416 cx.run_until_parked();
5417
5418 // State is set to Error and default message
5419 thread.read_with(cx, |thread, _| {
5420 assert!(matches!(thread.summary(), ThreadSummary::Error));
5421 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5422 });
5423 }
5424
5425 fn simulate_successful_response(fake_model: &FakeLanguageModel, cx: &mut TestAppContext) {
5426 cx.run_until_parked();
5427 fake_model.stream_last_completion_response("Assistant response");
5428 fake_model.end_last_completion_stream();
5429 cx.run_until_parked();
5430 }
5431
5432 fn init_test_settings(cx: &mut TestAppContext) {
5433 cx.update(|cx| {
5434 let settings_store = SettingsStore::test(cx);
5435 cx.set_global(settings_store);
5436 language::init(cx);
5437 Project::init_settings(cx);
5438 AgentSettings::register(cx);
5439 prompt_store::init(cx);
5440 thread_store::init(cx);
5441 workspace::init_settings(cx);
5442 language_model::init_settings(cx);
5443 ThemeSettings::register(cx);
5444 ToolRegistry::default_global(cx);
5445 assistant_tool::init(cx);
5446
5447 let http_client = Arc::new(http_client::HttpClientWithUrl::new(
5448 http_client::FakeHttpClient::with_200_response(),
5449 "http://localhost".to_string(),
5450 None,
5451 ));
5452 assistant_tools::init(http_client, cx);
5453 });
5454 }
5455
5456 // Helper to create a test project with test files
5457 async fn create_test_project(
5458 cx: &mut TestAppContext,
5459 files: serde_json::Value,
5460 ) -> Entity<Project> {
5461 let fs = FakeFs::new(cx.executor());
5462 fs.insert_tree(path!("/test"), files).await;
5463 Project::test(fs, [path!("/test").as_ref()], cx).await
5464 }
5465
5466 async fn setup_test_environment(
5467 cx: &mut TestAppContext,
5468 project: Entity<Project>,
5469 ) -> (
5470 Entity<Workspace>,
5471 Entity<ThreadStore>,
5472 Entity<Thread>,
5473 Entity<ContextStore>,
5474 Arc<dyn LanguageModel>,
5475 ) {
5476 let (workspace, cx) =
5477 cx.add_window_view(|window, cx| Workspace::test_new(project.clone(), window, cx));
5478
5479 let thread_store = cx
5480 .update(|_, cx| {
5481 ThreadStore::load(
5482 project.clone(),
5483 cx.new(|_| ToolWorkingSet::default()),
5484 None,
5485 Arc::new(PromptBuilder::new(None).unwrap()),
5486 cx,
5487 )
5488 })
5489 .await
5490 .unwrap();
5491
5492 let thread = thread_store.update(cx, |store, cx| store.create_thread(cx));
5493 let context_store = cx.new(|_cx| ContextStore::new(project.downgrade(), None));
5494
5495 let provider = Arc::new(FakeLanguageModelProvider);
5496 let model = provider.test_model();
5497 let model: Arc<dyn LanguageModel> = Arc::new(model);
5498
5499 cx.update(|_, cx| {
5500 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
5501 registry.set_default_model(
5502 Some(ConfiguredModel {
5503 provider: provider.clone(),
5504 model: model.clone(),
5505 }),
5506 cx,
5507 );
5508 registry.set_thread_summary_model(
5509 Some(ConfiguredModel {
5510 provider,
5511 model: model.clone(),
5512 }),
5513 cx,
5514 );
5515 })
5516 });
5517
5518 (workspace, thread_store, thread, context_store, model)
5519 }
5520
5521 async fn add_file_to_context(
5522 project: &Entity<Project>,
5523 context_store: &Entity<ContextStore>,
5524 path: &str,
5525 cx: &mut TestAppContext,
5526 ) -> Result<Entity<language::Buffer>> {
5527 let buffer_path = project
5528 .read_with(cx, |project, cx| project.find_project_path(path, cx))
5529 .unwrap();
5530
5531 let buffer = project
5532 .update(cx, |project, cx| {
5533 project.open_buffer(buffer_path.clone(), cx)
5534 })
5535 .await
5536 .unwrap();
5537
5538 context_store.update(cx, |context_store, cx| {
5539 context_store.add_file_from_buffer(&buffer_path, buffer.clone(), false, cx);
5540 });
5541
5542 Ok(buffer)
5543 }
5544}