1use crate::{
2 agent_profile::AgentProfile,
3 context::{AgentContext, AgentContextHandle, ContextLoadResult, LoadedContext},
4 thread_store::{
5 SerializedCrease, SerializedLanguageModel, SerializedMessage, SerializedMessageSegment,
6 SerializedThread, SerializedToolResult, SerializedToolUse, SharedProjectContext,
7 ThreadStore,
8 },
9 tool_use::{PendingToolUse, ToolUse, ToolUseMetadata, ToolUseState},
10};
11use action_log::ActionLog;
12use agent_settings::{AgentProfileId, AgentSettings, CompletionMode, SUMMARIZE_THREAD_PROMPT};
13use anyhow::{Result, anyhow};
14use assistant_tool::{AnyToolCard, Tool, ToolWorkingSet};
15use chrono::{DateTime, Utc};
16use client::{ModelRequestUsage, RequestUsage};
17use cloud_llm_client::{CompletionIntent, CompletionRequestStatus, Plan, UsageLimit};
18use collections::HashMap;
19use feature_flags::{self, FeatureFlagAppExt};
20use futures::{FutureExt, StreamExt as _, future::Shared};
21use git::repository::DiffType;
22use gpui::{
23 AnyWindowHandle, App, AppContext, AsyncApp, Context, Entity, EventEmitter, SharedString, Task,
24 WeakEntity, Window,
25};
26use http_client::StatusCode;
27use language_model::{
28 ConfiguredModel, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,
29 LanguageModelExt as _, LanguageModelId, LanguageModelRegistry, LanguageModelRequest,
30 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
31 LanguageModelToolResultContent, LanguageModelToolUse, LanguageModelToolUseId, MessageContent,
32 ModelRequestLimitReachedError, PaymentRequiredError, Role, SelectedModel, StopReason,
33 TokenUsage,
34};
35use postage::stream::Stream as _;
36use project::{
37 Project,
38 git_store::{GitStore, GitStoreCheckpoint, RepositoryState},
39};
40use prompt_store::{ModelContext, PromptBuilder};
41use schemars::JsonSchema;
42use serde::{Deserialize, Serialize};
43use settings::Settings;
44use std::{
45 io::Write,
46 ops::Range,
47 sync::Arc,
48 time::{Duration, Instant},
49};
50use thiserror::Error;
51use util::{ResultExt as _, post_inc};
52use uuid::Uuid;
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, &self.project, cx)
946 }
947
948 pub fn tool_results_for_message(
949 &self,
950 assistant_message_id: MessageId,
951 ) -> Vec<&LanguageModelToolResult> {
952 self.tool_use.tool_results_for_message(assistant_message_id)
953 }
954
955 pub fn tool_result(&self, id: &LanguageModelToolUseId) -> Option<&LanguageModelToolResult> {
956 self.tool_use.tool_result(id)
957 }
958
959 pub fn output_for_tool(&self, id: &LanguageModelToolUseId) -> Option<&Arc<str>> {
960 match &self.tool_use.tool_result(id)?.content {
961 LanguageModelToolResultContent::Text(text) => Some(text),
962 LanguageModelToolResultContent::Image(_) => {
963 // TODO: We should display image
964 None
965 }
966 }
967 }
968
969 pub fn card_for_tool(&self, id: &LanguageModelToolUseId) -> Option<AnyToolCard> {
970 self.tool_use.tool_result_card(id).cloned()
971 }
972
973 /// Return tools that are both enabled and supported by the model
974 pub fn available_tools(
975 &self,
976 cx: &App,
977 model: Arc<dyn LanguageModel>,
978 ) -> Vec<LanguageModelRequestTool> {
979 if model.supports_tools() {
980 self.profile
981 .enabled_tools(cx)
982 .into_iter()
983 .filter_map(|(name, tool)| {
984 // Skip tools that cannot be supported
985 let input_schema = tool.input_schema(model.tool_input_format()).ok()?;
986 Some(LanguageModelRequestTool {
987 name: name.into(),
988 description: tool.description(),
989 input_schema,
990 })
991 })
992 .collect()
993 } else {
994 Vec::default()
995 }
996 }
997
998 pub fn insert_user_message(
999 &mut self,
1000 text: impl Into<String>,
1001 loaded_context: ContextLoadResult,
1002 git_checkpoint: Option<GitStoreCheckpoint>,
1003 creases: Vec<MessageCrease>,
1004 cx: &mut Context<Self>,
1005 ) -> MessageId {
1006 if !loaded_context.referenced_buffers.is_empty() {
1007 self.action_log.update(cx, |log, cx| {
1008 for buffer in loaded_context.referenced_buffers {
1009 log.buffer_read(buffer, cx);
1010 }
1011 });
1012 }
1013
1014 let message_id = self.insert_message(
1015 Role::User,
1016 vec![MessageSegment::Text(text.into())],
1017 loaded_context.loaded_context,
1018 creases,
1019 false,
1020 cx,
1021 );
1022
1023 if let Some(git_checkpoint) = git_checkpoint {
1024 self.pending_checkpoint = Some(ThreadCheckpoint {
1025 message_id,
1026 git_checkpoint,
1027 });
1028 }
1029
1030 self.auto_capture_telemetry(cx);
1031
1032 message_id
1033 }
1034
1035 pub fn insert_invisible_continue_message(&mut self, cx: &mut Context<Self>) -> MessageId {
1036 let id = self.insert_message(
1037 Role::User,
1038 vec![MessageSegment::Text("Continue where you left off".into())],
1039 LoadedContext::default(),
1040 vec![],
1041 true,
1042 cx,
1043 );
1044 self.pending_checkpoint = None;
1045
1046 id
1047 }
1048
1049 pub fn insert_assistant_message(
1050 &mut self,
1051 segments: Vec<MessageSegment>,
1052 cx: &mut Context<Self>,
1053 ) -> MessageId {
1054 self.insert_message(
1055 Role::Assistant,
1056 segments,
1057 LoadedContext::default(),
1058 Vec::new(),
1059 false,
1060 cx,
1061 )
1062 }
1063
1064 pub fn insert_message(
1065 &mut self,
1066 role: Role,
1067 segments: Vec<MessageSegment>,
1068 loaded_context: LoadedContext,
1069 creases: Vec<MessageCrease>,
1070 is_hidden: bool,
1071 cx: &mut Context<Self>,
1072 ) -> MessageId {
1073 let id = self.next_message_id.post_inc();
1074 self.messages.push(Message {
1075 id,
1076 role,
1077 segments,
1078 loaded_context,
1079 creases,
1080 is_hidden,
1081 ui_only: false,
1082 });
1083 self.touch_updated_at();
1084 cx.emit(ThreadEvent::MessageAdded(id));
1085 id
1086 }
1087
1088 pub fn edit_message(
1089 &mut self,
1090 id: MessageId,
1091 new_role: Role,
1092 new_segments: Vec<MessageSegment>,
1093 creases: Vec<MessageCrease>,
1094 loaded_context: Option<LoadedContext>,
1095 checkpoint: Option<GitStoreCheckpoint>,
1096 cx: &mut Context<Self>,
1097 ) -> bool {
1098 let Some(message) = self.messages.iter_mut().find(|message| message.id == id) else {
1099 return false;
1100 };
1101 message.role = new_role;
1102 message.segments = new_segments;
1103 message.creases = creases;
1104 if let Some(context) = loaded_context {
1105 message.loaded_context = context;
1106 }
1107 if let Some(git_checkpoint) = checkpoint {
1108 self.checkpoints_by_message.insert(
1109 id,
1110 ThreadCheckpoint {
1111 message_id: id,
1112 git_checkpoint,
1113 },
1114 );
1115 }
1116 self.touch_updated_at();
1117 cx.emit(ThreadEvent::MessageEdited(id));
1118 true
1119 }
1120
1121 pub fn delete_message(&mut self, id: MessageId, cx: &mut Context<Self>) -> bool {
1122 let Some(index) = self.messages.iter().position(|message| message.id == id) else {
1123 return false;
1124 };
1125 self.messages.remove(index);
1126 self.touch_updated_at();
1127 cx.emit(ThreadEvent::MessageDeleted(id));
1128 true
1129 }
1130
1131 /// Returns the representation of this [`Thread`] in a textual form.
1132 ///
1133 /// This is the representation we use when attaching a thread as context to another thread.
1134 pub fn text(&self) -> String {
1135 let mut text = String::new();
1136
1137 for message in &self.messages {
1138 text.push_str(match message.role {
1139 language_model::Role::User => "User:",
1140 language_model::Role::Assistant => "Agent:",
1141 language_model::Role::System => "System:",
1142 });
1143 text.push('\n');
1144
1145 for segment in &message.segments {
1146 match segment {
1147 MessageSegment::Text(content) => text.push_str(content),
1148 MessageSegment::Thinking { text: content, .. } => {
1149 text.push_str(&format!("<think>{}</think>", content))
1150 }
1151 MessageSegment::RedactedThinking(_) => {}
1152 }
1153 }
1154 text.push('\n');
1155 }
1156
1157 text
1158 }
1159
1160 /// Serializes this thread into a format for storage or telemetry.
1161 pub fn serialize(&self, cx: &mut Context<Self>) -> Task<Result<SerializedThread>> {
1162 let initial_project_snapshot = self.initial_project_snapshot.clone();
1163 cx.spawn(async move |this, cx| {
1164 let initial_project_snapshot = initial_project_snapshot.await;
1165 this.read_with(cx, |this, cx| SerializedThread {
1166 version: SerializedThread::VERSION.to_string(),
1167 summary: this.summary().or_default(),
1168 updated_at: this.updated_at(),
1169 messages: this
1170 .messages()
1171 .filter(|message| !message.ui_only)
1172 .map(|message| SerializedMessage {
1173 id: message.id,
1174 role: message.role,
1175 segments: message
1176 .segments
1177 .iter()
1178 .map(|segment| match segment {
1179 MessageSegment::Text(text) => {
1180 SerializedMessageSegment::Text { text: text.clone() }
1181 }
1182 MessageSegment::Thinking { text, signature } => {
1183 SerializedMessageSegment::Thinking {
1184 text: text.clone(),
1185 signature: signature.clone(),
1186 }
1187 }
1188 MessageSegment::RedactedThinking(data) => {
1189 SerializedMessageSegment::RedactedThinking {
1190 data: data.clone(),
1191 }
1192 }
1193 })
1194 .collect(),
1195 tool_uses: this
1196 .tool_uses_for_message(message.id, cx)
1197 .into_iter()
1198 .map(|tool_use| SerializedToolUse {
1199 id: tool_use.id,
1200 name: tool_use.name,
1201 input: tool_use.input,
1202 })
1203 .collect(),
1204 tool_results: this
1205 .tool_results_for_message(message.id)
1206 .into_iter()
1207 .map(|tool_result| SerializedToolResult {
1208 tool_use_id: tool_result.tool_use_id.clone(),
1209 is_error: tool_result.is_error,
1210 content: tool_result.content.clone(),
1211 output: tool_result.output.clone(),
1212 })
1213 .collect(),
1214 context: message.loaded_context.text.clone(),
1215 creases: message
1216 .creases
1217 .iter()
1218 .map(|crease| SerializedCrease {
1219 start: crease.range.start,
1220 end: crease.range.end,
1221 icon_path: crease.icon_path.clone(),
1222 label: crease.label.clone(),
1223 })
1224 .collect(),
1225 is_hidden: message.is_hidden,
1226 })
1227 .collect(),
1228 initial_project_snapshot,
1229 cumulative_token_usage: this.cumulative_token_usage,
1230 request_token_usage: this.request_token_usage.clone(),
1231 detailed_summary_state: this.detailed_summary_rx.borrow().clone(),
1232 exceeded_window_error: this.exceeded_window_error.clone(),
1233 model: this
1234 .configured_model
1235 .as_ref()
1236 .map(|model| SerializedLanguageModel {
1237 provider: model.provider.id().0.to_string(),
1238 model: model.model.id().0.to_string(),
1239 }),
1240 completion_mode: Some(this.completion_mode),
1241 tool_use_limit_reached: this.tool_use_limit_reached,
1242 profile: Some(this.profile.id().clone()),
1243 })
1244 })
1245 }
1246
1247 pub fn remaining_turns(&self) -> u32 {
1248 self.remaining_turns
1249 }
1250
1251 pub fn set_remaining_turns(&mut self, remaining_turns: u32) {
1252 self.remaining_turns = remaining_turns;
1253 }
1254
1255 pub fn send_to_model(
1256 &mut self,
1257 model: Arc<dyn LanguageModel>,
1258 intent: CompletionIntent,
1259 window: Option<AnyWindowHandle>,
1260 cx: &mut Context<Self>,
1261 ) {
1262 if self.remaining_turns == 0 {
1263 return;
1264 }
1265
1266 self.remaining_turns -= 1;
1267
1268 self.flush_notifications(model.clone(), intent, cx);
1269
1270 let _checkpoint = self.finalize_pending_checkpoint(cx);
1271 self.stream_completion(
1272 self.to_completion_request(model.clone(), intent, cx),
1273 model,
1274 intent,
1275 window,
1276 cx,
1277 );
1278 }
1279
1280 pub fn retry_last_completion(
1281 &mut self,
1282 window: Option<AnyWindowHandle>,
1283 cx: &mut Context<Self>,
1284 ) {
1285 // Clear any existing error state
1286 self.retry_state = None;
1287
1288 // Use the last error context if available, otherwise fall back to configured model
1289 let (model, intent) = if let Some((model, intent)) = self.last_error_context.take() {
1290 (model, intent)
1291 } else if let Some(configured_model) = self.configured_model.as_ref() {
1292 let model = configured_model.model.clone();
1293 let intent = if self.has_pending_tool_uses() {
1294 CompletionIntent::ToolResults
1295 } else {
1296 CompletionIntent::UserPrompt
1297 };
1298 (model, intent)
1299 } else if let Some(configured_model) = self.get_or_init_configured_model(cx) {
1300 let model = configured_model.model.clone();
1301 let intent = if self.has_pending_tool_uses() {
1302 CompletionIntent::ToolResults
1303 } else {
1304 CompletionIntent::UserPrompt
1305 };
1306 (model, intent)
1307 } else {
1308 return;
1309 };
1310
1311 self.send_to_model(model, intent, window, cx);
1312 }
1313
1314 pub fn enable_burn_mode_and_retry(
1315 &mut self,
1316 window: Option<AnyWindowHandle>,
1317 cx: &mut Context<Self>,
1318 ) {
1319 self.completion_mode = CompletionMode::Burn;
1320 cx.emit(ThreadEvent::ProfileChanged);
1321 self.retry_last_completion(window, cx);
1322 }
1323
1324 pub fn used_tools_since_last_user_message(&self) -> bool {
1325 for message in self.messages.iter().rev() {
1326 if self.tool_use.message_has_tool_results(message.id) {
1327 return true;
1328 } else if message.role == Role::User {
1329 return false;
1330 }
1331 }
1332
1333 false
1334 }
1335
1336 pub fn to_completion_request(
1337 &self,
1338 model: Arc<dyn LanguageModel>,
1339 intent: CompletionIntent,
1340 cx: &mut Context<Self>,
1341 ) -> LanguageModelRequest {
1342 let mut request = LanguageModelRequest {
1343 thread_id: Some(self.id.to_string()),
1344 prompt_id: Some(self.last_prompt_id.to_string()),
1345 intent: Some(intent),
1346 mode: None,
1347 messages: vec![],
1348 tools: Vec::new(),
1349 tool_choice: None,
1350 stop: Vec::new(),
1351 temperature: AgentSettings::temperature_for_model(&model, cx),
1352 thinking_allowed: true,
1353 };
1354
1355 let available_tools = self.available_tools(cx, model.clone());
1356 let available_tool_names = available_tools
1357 .iter()
1358 .map(|tool| tool.name.clone())
1359 .collect();
1360
1361 let model_context = &ModelContext {
1362 available_tools: available_tool_names,
1363 };
1364
1365 if let Some(project_context) = self.project_context.borrow().as_ref() {
1366 match self
1367 .prompt_builder
1368 .generate_assistant_system_prompt(project_context, model_context)
1369 {
1370 Err(err) => {
1371 let message = format!("{err:?}").into();
1372 log::error!("{message}");
1373 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1374 header: "Error generating system prompt".into(),
1375 message,
1376 }));
1377 }
1378 Ok(system_prompt) => {
1379 request.messages.push(LanguageModelRequestMessage {
1380 role: Role::System,
1381 content: vec![MessageContent::Text(system_prompt)],
1382 cache: true,
1383 });
1384 }
1385 }
1386 } else {
1387 let message = "Context for system prompt unexpectedly not ready.".into();
1388 log::error!("{message}");
1389 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1390 header: "Error generating system prompt".into(),
1391 message,
1392 }));
1393 }
1394
1395 let mut message_ix_to_cache = None;
1396 for message in &self.messages {
1397 // ui_only messages are for the UI only, not for the model
1398 if message.ui_only {
1399 continue;
1400 }
1401
1402 let mut request_message = LanguageModelRequestMessage {
1403 role: message.role,
1404 content: Vec::new(),
1405 cache: false,
1406 };
1407
1408 message
1409 .loaded_context
1410 .add_to_request_message(&mut request_message);
1411
1412 for segment in &message.segments {
1413 match segment {
1414 MessageSegment::Text(text) => {
1415 let text = text.trim_end();
1416 if !text.is_empty() {
1417 request_message
1418 .content
1419 .push(MessageContent::Text(text.into()));
1420 }
1421 }
1422 MessageSegment::Thinking { text, signature } => {
1423 if !text.is_empty() {
1424 request_message.content.push(MessageContent::Thinking {
1425 text: text.into(),
1426 signature: signature.clone(),
1427 });
1428 }
1429 }
1430 MessageSegment::RedactedThinking(data) => {
1431 request_message
1432 .content
1433 .push(MessageContent::RedactedThinking(data.clone()));
1434 }
1435 };
1436 }
1437
1438 let mut cache_message = true;
1439 let mut tool_results_message = LanguageModelRequestMessage {
1440 role: Role::User,
1441 content: Vec::new(),
1442 cache: false,
1443 };
1444 for (tool_use, tool_result) in self.tool_use.tool_results(message.id) {
1445 if let Some(tool_result) = tool_result {
1446 request_message
1447 .content
1448 .push(MessageContent::ToolUse(tool_use.clone()));
1449 tool_results_message
1450 .content
1451 .push(MessageContent::ToolResult(LanguageModelToolResult {
1452 tool_use_id: tool_use.id.clone(),
1453 tool_name: tool_result.tool_name.clone(),
1454 is_error: tool_result.is_error,
1455 content: if tool_result.content.is_empty() {
1456 // Surprisingly, the API fails if we return an empty string here.
1457 // It thinks we are sending a tool use without a tool result.
1458 "<Tool returned an empty string>".into()
1459 } else {
1460 tool_result.content.clone()
1461 },
1462 output: None,
1463 }));
1464 } else {
1465 cache_message = false;
1466 log::debug!(
1467 "skipped tool use {:?} because it is still pending",
1468 tool_use
1469 );
1470 }
1471 }
1472
1473 if cache_message {
1474 message_ix_to_cache = Some(request.messages.len());
1475 }
1476 request.messages.push(request_message);
1477
1478 if !tool_results_message.content.is_empty() {
1479 if cache_message {
1480 message_ix_to_cache = Some(request.messages.len());
1481 }
1482 request.messages.push(tool_results_message);
1483 }
1484 }
1485
1486 // https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
1487 if let Some(message_ix_to_cache) = message_ix_to_cache {
1488 request.messages[message_ix_to_cache].cache = true;
1489 }
1490
1491 request.tools = available_tools;
1492 request.mode = if model.supports_burn_mode() {
1493 Some(self.completion_mode.into())
1494 } else {
1495 Some(CompletionMode::Normal.into())
1496 };
1497
1498 request
1499 }
1500
1501 fn to_summarize_request(
1502 &self,
1503 model: &Arc<dyn LanguageModel>,
1504 intent: CompletionIntent,
1505 added_user_message: String,
1506 cx: &App,
1507 ) -> LanguageModelRequest {
1508 let mut request = LanguageModelRequest {
1509 thread_id: None,
1510 prompt_id: None,
1511 intent: Some(intent),
1512 mode: None,
1513 messages: vec![],
1514 tools: Vec::new(),
1515 tool_choice: None,
1516 stop: Vec::new(),
1517 temperature: AgentSettings::temperature_for_model(model, cx),
1518 thinking_allowed: false,
1519 };
1520
1521 for message in &self.messages {
1522 let mut request_message = LanguageModelRequestMessage {
1523 role: message.role,
1524 content: Vec::new(),
1525 cache: false,
1526 };
1527
1528 for segment in &message.segments {
1529 match segment {
1530 MessageSegment::Text(text) => request_message
1531 .content
1532 .push(MessageContent::Text(text.clone())),
1533 MessageSegment::Thinking { .. } => {}
1534 MessageSegment::RedactedThinking(_) => {}
1535 }
1536 }
1537
1538 if request_message.content.is_empty() {
1539 continue;
1540 }
1541
1542 request.messages.push(request_message);
1543 }
1544
1545 request.messages.push(LanguageModelRequestMessage {
1546 role: Role::User,
1547 content: vec![MessageContent::Text(added_user_message)],
1548 cache: false,
1549 });
1550
1551 request
1552 }
1553
1554 /// Insert auto-generated notifications (if any) to the thread
1555 fn flush_notifications(
1556 &mut self,
1557 model: Arc<dyn LanguageModel>,
1558 intent: CompletionIntent,
1559 cx: &mut Context<Self>,
1560 ) {
1561 match intent {
1562 CompletionIntent::UserPrompt | CompletionIntent::ToolResults => {
1563 if let Some(pending_tool_use) = self.attach_tracked_files_state(model, cx) {
1564 cx.emit(ThreadEvent::ToolFinished {
1565 tool_use_id: pending_tool_use.id.clone(),
1566 pending_tool_use: Some(pending_tool_use),
1567 });
1568 }
1569 }
1570 CompletionIntent::ThreadSummarization
1571 | CompletionIntent::ThreadContextSummarization
1572 | CompletionIntent::CreateFile
1573 | CompletionIntent::EditFile
1574 | CompletionIntent::InlineAssist
1575 | CompletionIntent::TerminalInlineAssist
1576 | CompletionIntent::GenerateGitCommitMessage => {}
1577 };
1578 }
1579
1580 fn attach_tracked_files_state(
1581 &mut self,
1582 model: Arc<dyn LanguageModel>,
1583 cx: &mut App,
1584 ) -> Option<PendingToolUse> {
1585 // Represent notification as a simulated `project_notifications` tool call
1586 let tool_name = Arc::from("project_notifications");
1587 let tool = self.tools.read(cx).tool(&tool_name, cx)?;
1588
1589 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
1590 return None;
1591 }
1592
1593 if self
1594 .action_log
1595 .update(cx, |log, cx| log.unnotified_user_edits(cx).is_none())
1596 {
1597 return None;
1598 }
1599
1600 let input = serde_json::json!({});
1601 let request = Arc::new(LanguageModelRequest::default()); // unused
1602 let window = None;
1603 let tool_result = tool.run(
1604 input,
1605 request,
1606 self.project.clone(),
1607 self.action_log.clone(),
1608 model.clone(),
1609 window,
1610 cx,
1611 );
1612
1613 let tool_use_id =
1614 LanguageModelToolUseId::from(format!("project_notifications_{}", self.messages.len()));
1615
1616 let tool_use = LanguageModelToolUse {
1617 id: tool_use_id.clone(),
1618 name: tool_name.clone(),
1619 raw_input: "{}".to_string(),
1620 input: serde_json::json!({}),
1621 is_input_complete: true,
1622 };
1623
1624 let tool_output = cx.background_executor().block(tool_result.output);
1625
1626 // Attach a project_notification tool call to the latest existing
1627 // Assistant message. We cannot create a new Assistant message
1628 // because thinking models require a `thinking` block that we
1629 // cannot mock. We cannot send a notification as a normal
1630 // (non-tool-use) User message because this distracts Agent
1631 // too much.
1632 let tool_message_id = self
1633 .messages
1634 .iter()
1635 .enumerate()
1636 .rfind(|(_, message)| message.role == Role::Assistant)
1637 .map(|(_, message)| message.id)?;
1638
1639 let tool_use_metadata = ToolUseMetadata {
1640 model: model.clone(),
1641 thread_id: self.id.clone(),
1642 prompt_id: self.last_prompt_id.clone(),
1643 };
1644
1645 self.tool_use
1646 .request_tool_use(tool_message_id, tool_use, tool_use_metadata.clone(), cx);
1647
1648 let pending_tool_use = self.tool_use.insert_tool_output(
1649 tool_use_id.clone(),
1650 tool_name,
1651 tool_output,
1652 self.configured_model.as_ref(),
1653 self.completion_mode,
1654 );
1655
1656 pending_tool_use
1657 }
1658
1659 pub fn stream_completion(
1660 &mut self,
1661 request: LanguageModelRequest,
1662 model: Arc<dyn LanguageModel>,
1663 intent: CompletionIntent,
1664 window: Option<AnyWindowHandle>,
1665 cx: &mut Context<Self>,
1666 ) {
1667 self.tool_use_limit_reached = false;
1668
1669 let pending_completion_id = post_inc(&mut self.completion_count);
1670 let mut request_callback_parameters = if self.request_callback.is_some() {
1671 Some((request.clone(), Vec::new()))
1672 } else {
1673 None
1674 };
1675 let prompt_id = self.last_prompt_id.clone();
1676 let tool_use_metadata = ToolUseMetadata {
1677 model: model.clone(),
1678 thread_id: self.id.clone(),
1679 prompt_id: prompt_id.clone(),
1680 };
1681
1682 let completion_mode = request
1683 .mode
1684 .unwrap_or(cloud_llm_client::CompletionMode::Normal);
1685
1686 self.last_received_chunk_at = Some(Instant::now());
1687
1688 let task = cx.spawn(async move |thread, cx| {
1689 let stream_completion_future = model.stream_completion(request, &cx);
1690 let initial_token_usage =
1691 thread.read_with(cx, |thread, _cx| thread.cumulative_token_usage);
1692 let stream_completion = async {
1693 let mut events = stream_completion_future.await?;
1694
1695 let mut stop_reason = StopReason::EndTurn;
1696 let mut current_token_usage = TokenUsage::default();
1697
1698 thread
1699 .update(cx, |_thread, cx| {
1700 cx.emit(ThreadEvent::NewRequest);
1701 })
1702 .ok();
1703
1704 let mut request_assistant_message_id = None;
1705
1706 while let Some(event) = events.next().await {
1707 if let Some((_, response_events)) = request_callback_parameters.as_mut() {
1708 response_events
1709 .push(event.as_ref().map_err(|error| error.to_string()).cloned());
1710 }
1711
1712 thread.update(cx, |thread, cx| {
1713 match event? {
1714 LanguageModelCompletionEvent::StartMessage { .. } => {
1715 request_assistant_message_id =
1716 Some(thread.insert_assistant_message(
1717 vec![MessageSegment::Text(String::new())],
1718 cx,
1719 ));
1720 }
1721 LanguageModelCompletionEvent::Stop(reason) => {
1722 stop_reason = reason;
1723 }
1724 LanguageModelCompletionEvent::UsageUpdate(token_usage) => {
1725 thread.update_token_usage_at_last_message(token_usage);
1726 thread.cumulative_token_usage = thread.cumulative_token_usage
1727 + token_usage
1728 - current_token_usage;
1729 current_token_usage = token_usage;
1730 }
1731 LanguageModelCompletionEvent::Text(chunk) => {
1732 thread.received_chunk();
1733
1734 cx.emit(ThreadEvent::ReceivedTextChunk);
1735 if let Some(last_message) = thread.messages.last_mut() {
1736 if last_message.role == Role::Assistant
1737 && !thread.tool_use.has_tool_results(last_message.id)
1738 {
1739 last_message.push_text(&chunk);
1740 cx.emit(ThreadEvent::StreamedAssistantText(
1741 last_message.id,
1742 chunk,
1743 ));
1744 } else {
1745 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1746 // of a new Assistant response.
1747 //
1748 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1749 // will result in duplicating the text of the chunk in the rendered Markdown.
1750 request_assistant_message_id =
1751 Some(thread.insert_assistant_message(
1752 vec![MessageSegment::Text(chunk.to_string())],
1753 cx,
1754 ));
1755 };
1756 }
1757 }
1758 LanguageModelCompletionEvent::Thinking {
1759 text: chunk,
1760 signature,
1761 } => {
1762 thread.received_chunk();
1763
1764 if let Some(last_message) = thread.messages.last_mut() {
1765 if last_message.role == Role::Assistant
1766 && !thread.tool_use.has_tool_results(last_message.id)
1767 {
1768 last_message.push_thinking(&chunk, signature);
1769 cx.emit(ThreadEvent::StreamedAssistantThinking(
1770 last_message.id,
1771 chunk,
1772 ));
1773 } else {
1774 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1775 // of a new Assistant response.
1776 //
1777 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1778 // will result in duplicating the text of the chunk in the rendered Markdown.
1779 request_assistant_message_id =
1780 Some(thread.insert_assistant_message(
1781 vec![MessageSegment::Thinking {
1782 text: chunk.to_string(),
1783 signature,
1784 }],
1785 cx,
1786 ));
1787 };
1788 }
1789 }
1790 LanguageModelCompletionEvent::RedactedThinking { data } => {
1791 thread.received_chunk();
1792
1793 if let Some(last_message) = thread.messages.last_mut() {
1794 if last_message.role == Role::Assistant
1795 && !thread.tool_use.has_tool_results(last_message.id)
1796 {
1797 last_message.push_redacted_thinking(data);
1798 } else {
1799 request_assistant_message_id =
1800 Some(thread.insert_assistant_message(
1801 vec![MessageSegment::RedactedThinking(data)],
1802 cx,
1803 ));
1804 };
1805 }
1806 }
1807 LanguageModelCompletionEvent::ToolUse(tool_use) => {
1808 let last_assistant_message_id = request_assistant_message_id
1809 .unwrap_or_else(|| {
1810 let new_assistant_message_id =
1811 thread.insert_assistant_message(vec![], cx);
1812 request_assistant_message_id =
1813 Some(new_assistant_message_id);
1814 new_assistant_message_id
1815 });
1816
1817 let tool_use_id = tool_use.id.clone();
1818 let streamed_input = if tool_use.is_input_complete {
1819 None
1820 } else {
1821 Some((&tool_use.input).clone())
1822 };
1823
1824 let ui_text = thread.tool_use.request_tool_use(
1825 last_assistant_message_id,
1826 tool_use,
1827 tool_use_metadata.clone(),
1828 cx,
1829 );
1830
1831 if let Some(input) = streamed_input {
1832 cx.emit(ThreadEvent::StreamedToolUse {
1833 tool_use_id,
1834 ui_text,
1835 input,
1836 });
1837 }
1838 }
1839 LanguageModelCompletionEvent::ToolUseJsonParseError {
1840 id,
1841 tool_name,
1842 raw_input: invalid_input_json,
1843 json_parse_error,
1844 } => {
1845 thread.receive_invalid_tool_json(
1846 id,
1847 tool_name,
1848 invalid_input_json,
1849 json_parse_error,
1850 window,
1851 cx,
1852 );
1853 }
1854 LanguageModelCompletionEvent::StatusUpdate(status_update) => {
1855 if let Some(completion) = thread
1856 .pending_completions
1857 .iter_mut()
1858 .find(|completion| completion.id == pending_completion_id)
1859 {
1860 match status_update {
1861 CompletionRequestStatus::Queued { position } => {
1862 completion.queue_state =
1863 QueueState::Queued { position };
1864 }
1865 CompletionRequestStatus::Started => {
1866 completion.queue_state = QueueState::Started;
1867 }
1868 CompletionRequestStatus::Failed {
1869 code,
1870 message,
1871 request_id: _,
1872 retry_after,
1873 } => {
1874 return Err(
1875 LanguageModelCompletionError::from_cloud_failure(
1876 model.upstream_provider_name(),
1877 code,
1878 message,
1879 retry_after.map(Duration::from_secs_f64),
1880 ),
1881 );
1882 }
1883 CompletionRequestStatus::UsageUpdated { amount, limit } => {
1884 thread.update_model_request_usage(
1885 amount as u32,
1886 limit,
1887 cx,
1888 );
1889 }
1890 CompletionRequestStatus::ToolUseLimitReached => {
1891 thread.tool_use_limit_reached = true;
1892 cx.emit(ThreadEvent::ToolUseLimitReached);
1893 }
1894 }
1895 }
1896 }
1897 }
1898
1899 thread.touch_updated_at();
1900 cx.emit(ThreadEvent::StreamedCompletion);
1901 cx.notify();
1902
1903 thread.auto_capture_telemetry(cx);
1904 Ok(())
1905 })??;
1906
1907 smol::future::yield_now().await;
1908 }
1909
1910 thread.update(cx, |thread, cx| {
1911 thread.last_received_chunk_at = None;
1912 thread
1913 .pending_completions
1914 .retain(|completion| completion.id != pending_completion_id);
1915
1916 // If there is a response without tool use, summarize the message. Otherwise,
1917 // allow two tool uses before summarizing.
1918 if matches!(thread.summary, ThreadSummary::Pending)
1919 && thread.messages.len() >= 2
1920 && (!thread.has_pending_tool_uses() || thread.messages.len() >= 6)
1921 {
1922 thread.summarize(cx);
1923 }
1924 })?;
1925
1926 anyhow::Ok(stop_reason)
1927 };
1928
1929 let result = stream_completion.await;
1930 let mut retry_scheduled = false;
1931
1932 thread
1933 .update(cx, |thread, cx| {
1934 thread.finalize_pending_checkpoint(cx);
1935 match result.as_ref() {
1936 Ok(stop_reason) => {
1937 match stop_reason {
1938 StopReason::ToolUse => {
1939 let tool_uses =
1940 thread.use_pending_tools(window, model.clone(), cx);
1941 cx.emit(ThreadEvent::UsePendingTools { tool_uses });
1942 }
1943 StopReason::EndTurn | StopReason::MaxTokens => {
1944 thread.project.update(cx, |project, cx| {
1945 project.set_agent_location(None, cx);
1946 });
1947 }
1948 StopReason::Refusal => {
1949 thread.project.update(cx, |project, cx| {
1950 project.set_agent_location(None, cx);
1951 });
1952
1953 // Remove the turn that was refused.
1954 //
1955 // https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/handle-streaming-refusals#reset-context-after-refusal
1956 {
1957 let mut messages_to_remove = Vec::new();
1958
1959 for (ix, message) in
1960 thread.messages.iter().enumerate().rev()
1961 {
1962 messages_to_remove.push(message.id);
1963
1964 if message.role == Role::User {
1965 if ix == 0 {
1966 break;
1967 }
1968
1969 if let Some(prev_message) =
1970 thread.messages.get(ix - 1)
1971 {
1972 if prev_message.role == Role::Assistant {
1973 break;
1974 }
1975 }
1976 }
1977 }
1978
1979 for message_id in messages_to_remove {
1980 thread.delete_message(message_id, cx);
1981 }
1982 }
1983
1984 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1985 header: "Language model refusal".into(),
1986 message:
1987 "Model refused to generate content for safety reasons."
1988 .into(),
1989 }));
1990 }
1991 }
1992
1993 // We successfully completed, so cancel any remaining retries.
1994 thread.retry_state = None;
1995 }
1996 Err(error) => {
1997 thread.project.update(cx, |project, cx| {
1998 project.set_agent_location(None, cx);
1999 });
2000
2001 if error.is::<PaymentRequiredError>() {
2002 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
2003 } else if let Some(error) =
2004 error.downcast_ref::<ModelRequestLimitReachedError>()
2005 {
2006 cx.emit(ThreadEvent::ShowError(
2007 ThreadError::ModelRequestLimitReached { plan: error.plan },
2008 ));
2009 } else if let Some(completion_error) =
2010 error.downcast_ref::<LanguageModelCompletionError>()
2011 {
2012 match &completion_error {
2013 LanguageModelCompletionError::PromptTooLarge {
2014 tokens, ..
2015 } => {
2016 let tokens = tokens.unwrap_or_else(|| {
2017 // We didn't get an exact token count from the API, so fall back on our estimate.
2018 thread
2019 .total_token_usage()
2020 .map(|usage| usage.total)
2021 .unwrap_or(0)
2022 // We know the context window was exceeded in practice, so if our estimate was
2023 // lower than max tokens, the estimate was wrong; return that we exceeded by 1.
2024 .max(
2025 model
2026 .max_token_count_for_mode(completion_mode)
2027 .saturating_add(1),
2028 )
2029 });
2030 thread.exceeded_window_error = Some(ExceededWindowError {
2031 model_id: model.id(),
2032 token_count: tokens,
2033 });
2034 cx.notify();
2035 }
2036 _ => {
2037 if let Some(retry_strategy) =
2038 Thread::get_retry_strategy(completion_error)
2039 {
2040 log::info!(
2041 "Retrying with {:?} for language model completion error {:?}",
2042 retry_strategy,
2043 completion_error
2044 );
2045
2046 retry_scheduled = thread
2047 .handle_retryable_error_with_delay(
2048 &completion_error,
2049 Some(retry_strategy),
2050 model.clone(),
2051 intent,
2052 window,
2053 cx,
2054 );
2055 }
2056 }
2057 }
2058 }
2059
2060 if !retry_scheduled {
2061 thread.cancel_last_completion(window, cx);
2062 }
2063 }
2064 }
2065
2066 if !retry_scheduled {
2067 cx.emit(ThreadEvent::Stopped(result.map_err(Arc::new)));
2068 }
2069
2070 if let Some((request_callback, (request, response_events))) = thread
2071 .request_callback
2072 .as_mut()
2073 .zip(request_callback_parameters.as_ref())
2074 {
2075 request_callback(request, response_events);
2076 }
2077
2078 thread.auto_capture_telemetry(cx);
2079
2080 if let Ok(initial_usage) = initial_token_usage {
2081 let usage = thread.cumulative_token_usage - initial_usage;
2082
2083 telemetry::event!(
2084 "Assistant Thread Completion",
2085 thread_id = thread.id().to_string(),
2086 prompt_id = prompt_id,
2087 model = model.telemetry_id(),
2088 model_provider = model.provider_id().to_string(),
2089 input_tokens = usage.input_tokens,
2090 output_tokens = usage.output_tokens,
2091 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2092 cache_read_input_tokens = usage.cache_read_input_tokens,
2093 );
2094 }
2095 })
2096 .ok();
2097 });
2098
2099 self.pending_completions.push(PendingCompletion {
2100 id: pending_completion_id,
2101 queue_state: QueueState::Sending,
2102 _task: task,
2103 });
2104 }
2105
2106 pub fn summarize(&mut self, cx: &mut Context<Self>) {
2107 let Some(model) = LanguageModelRegistry::read_global(cx).thread_summary_model() else {
2108 println!("No thread summary model");
2109 return;
2110 };
2111
2112 if !model.provider.is_authenticated(cx) {
2113 return;
2114 }
2115
2116 let request = self.to_summarize_request(
2117 &model.model,
2118 CompletionIntent::ThreadSummarization,
2119 SUMMARIZE_THREAD_PROMPT.into(),
2120 cx,
2121 );
2122
2123 self.summary = ThreadSummary::Generating;
2124
2125 self.pending_summary = cx.spawn(async move |this, cx| {
2126 let result = async {
2127 let mut messages = model.model.stream_completion(request, &cx).await?;
2128
2129 let mut new_summary = String::new();
2130 while let Some(event) = messages.next().await {
2131 let Ok(event) = event else {
2132 continue;
2133 };
2134 let text = match event {
2135 LanguageModelCompletionEvent::Text(text) => text,
2136 LanguageModelCompletionEvent::StatusUpdate(
2137 CompletionRequestStatus::UsageUpdated { amount, limit },
2138 ) => {
2139 this.update(cx, |thread, cx| {
2140 thread.update_model_request_usage(amount as u32, limit, cx);
2141 })?;
2142 continue;
2143 }
2144 _ => continue,
2145 };
2146
2147 let mut lines = text.lines();
2148 new_summary.extend(lines.next());
2149
2150 // Stop if the LLM generated multiple lines.
2151 if lines.next().is_some() {
2152 break;
2153 }
2154 }
2155
2156 anyhow::Ok(new_summary)
2157 }
2158 .await;
2159
2160 this.update(cx, |this, cx| {
2161 match result {
2162 Ok(new_summary) => {
2163 if new_summary.is_empty() {
2164 this.summary = ThreadSummary::Error;
2165 } else {
2166 this.summary = ThreadSummary::Ready(new_summary.into());
2167 }
2168 }
2169 Err(err) => {
2170 this.summary = ThreadSummary::Error;
2171 log::error!("Failed to generate thread summary: {}", err);
2172 }
2173 }
2174 cx.emit(ThreadEvent::SummaryGenerated);
2175 })
2176 .log_err()?;
2177
2178 Some(())
2179 });
2180 }
2181
2182 fn get_retry_strategy(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2183 use LanguageModelCompletionError::*;
2184
2185 // General strategy here:
2186 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2187 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2188 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2189 match error {
2190 HttpResponseError {
2191 status_code: StatusCode::TOO_MANY_REQUESTS,
2192 ..
2193 } => Some(RetryStrategy::ExponentialBackoff {
2194 initial_delay: BASE_RETRY_DELAY,
2195 max_attempts: MAX_RETRY_ATTEMPTS,
2196 }),
2197 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2198 Some(RetryStrategy::Fixed {
2199 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2200 max_attempts: MAX_RETRY_ATTEMPTS,
2201 })
2202 }
2203 UpstreamProviderError {
2204 status,
2205 retry_after,
2206 ..
2207 } => match *status {
2208 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2209 Some(RetryStrategy::Fixed {
2210 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2211 max_attempts: MAX_RETRY_ATTEMPTS,
2212 })
2213 }
2214 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2215 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2216 // Internal Server Error could be anything, retry up to 3 times.
2217 max_attempts: 3,
2218 }),
2219 status => {
2220 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2221 // but we frequently get them in practice. See https://http.dev/529
2222 if status.as_u16() == 529 {
2223 Some(RetryStrategy::Fixed {
2224 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2225 max_attempts: MAX_RETRY_ATTEMPTS,
2226 })
2227 } else {
2228 Some(RetryStrategy::Fixed {
2229 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2230 max_attempts: 2,
2231 })
2232 }
2233 }
2234 },
2235 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2236 delay: BASE_RETRY_DELAY,
2237 max_attempts: 3,
2238 }),
2239 ApiReadResponseError { .. }
2240 | HttpSend { .. }
2241 | DeserializeResponse { .. }
2242 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2243 delay: BASE_RETRY_DELAY,
2244 max_attempts: 3,
2245 }),
2246 // Retrying these errors definitely shouldn't help.
2247 HttpResponseError {
2248 status_code:
2249 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2250 ..
2251 }
2252 | AuthenticationError { .. }
2253 | PermissionError { .. }
2254 | NoApiKey { .. }
2255 | ApiEndpointNotFound { .. }
2256 | PromptTooLarge { .. } => None,
2257 // These errors might be transient, so retry them
2258 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2259 delay: BASE_RETRY_DELAY,
2260 max_attempts: 1,
2261 }),
2262 // Retry all other 4xx and 5xx errors once.
2263 HttpResponseError { status_code, .. }
2264 if status_code.is_client_error() || status_code.is_server_error() =>
2265 {
2266 Some(RetryStrategy::Fixed {
2267 delay: BASE_RETRY_DELAY,
2268 max_attempts: 3,
2269 })
2270 }
2271 // Conservatively assume that any other errors are non-retryable
2272 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2273 delay: BASE_RETRY_DELAY,
2274 max_attempts: 2,
2275 }),
2276 }
2277 }
2278
2279 fn handle_retryable_error_with_delay(
2280 &mut self,
2281 error: &LanguageModelCompletionError,
2282 strategy: Option<RetryStrategy>,
2283 model: Arc<dyn LanguageModel>,
2284 intent: CompletionIntent,
2285 window: Option<AnyWindowHandle>,
2286 cx: &mut Context<Self>,
2287 ) -> bool {
2288 // Store context for the Retry button
2289 self.last_error_context = Some((model.clone(), intent));
2290
2291 // Only auto-retry if Burn Mode is enabled
2292 if self.completion_mode != CompletionMode::Burn {
2293 // Show error with retry options
2294 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2295 message: format!(
2296 "{}\n\nTo automatically retry when similar errors happen, enable Burn Mode.",
2297 error
2298 )
2299 .into(),
2300 can_enable_burn_mode: true,
2301 }));
2302 return false;
2303 }
2304
2305 let Some(strategy) = strategy.or_else(|| Self::get_retry_strategy(error)) else {
2306 return false;
2307 };
2308
2309 let max_attempts = match &strategy {
2310 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2311 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2312 };
2313
2314 let retry_state = self.retry_state.get_or_insert(RetryState {
2315 attempt: 0,
2316 max_attempts,
2317 intent,
2318 });
2319
2320 retry_state.attempt += 1;
2321 let attempt = retry_state.attempt;
2322 let max_attempts = retry_state.max_attempts;
2323 let intent = retry_state.intent;
2324
2325 if attempt <= max_attempts {
2326 let delay = match &strategy {
2327 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2328 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2329 Duration::from_secs(delay_secs)
2330 }
2331 RetryStrategy::Fixed { delay, .. } => *delay,
2332 };
2333
2334 // Add a transient message to inform the user
2335 let delay_secs = delay.as_secs();
2336 let retry_message = if max_attempts == 1 {
2337 format!("{error}. Retrying in {delay_secs} seconds...")
2338 } else {
2339 format!(
2340 "{error}. Retrying (attempt {attempt} of {max_attempts}) \
2341 in {delay_secs} seconds..."
2342 )
2343 };
2344 log::warn!(
2345 "Retrying completion request (attempt {attempt} of {max_attempts}) \
2346 in {delay_secs} seconds: {error:?}",
2347 );
2348
2349 // Add a UI-only message instead of a regular message
2350 let id = self.next_message_id.post_inc();
2351 self.messages.push(Message {
2352 id,
2353 role: Role::System,
2354 segments: vec![MessageSegment::Text(retry_message)],
2355 loaded_context: LoadedContext::default(),
2356 creases: Vec::new(),
2357 is_hidden: false,
2358 ui_only: true,
2359 });
2360 cx.emit(ThreadEvent::MessageAdded(id));
2361
2362 // Schedule the retry
2363 let thread_handle = cx.entity().downgrade();
2364
2365 cx.spawn(async move |_thread, cx| {
2366 cx.background_executor().timer(delay).await;
2367
2368 thread_handle
2369 .update(cx, |thread, cx| {
2370 // Retry the completion
2371 thread.send_to_model(model, intent, window, cx);
2372 })
2373 .log_err();
2374 })
2375 .detach();
2376
2377 true
2378 } else {
2379 // Max retries exceeded
2380 self.retry_state = None;
2381
2382 // Stop generating since we're giving up on retrying.
2383 self.pending_completions.clear();
2384
2385 // Show error alongside a Retry button, but no
2386 // Enable Burn Mode button (since it's already enabled)
2387 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2388 message: format!("Failed after retrying: {}", error).into(),
2389 can_enable_burn_mode: false,
2390 }));
2391
2392 false
2393 }
2394 }
2395
2396 pub fn start_generating_detailed_summary_if_needed(
2397 &mut self,
2398 thread_store: WeakEntity<ThreadStore>,
2399 cx: &mut Context<Self>,
2400 ) {
2401 let Some(last_message_id) = self.messages.last().map(|message| message.id) else {
2402 return;
2403 };
2404
2405 match &*self.detailed_summary_rx.borrow() {
2406 DetailedSummaryState::Generating { message_id, .. }
2407 | DetailedSummaryState::Generated { message_id, .. }
2408 if *message_id == last_message_id =>
2409 {
2410 // Already up-to-date
2411 return;
2412 }
2413 _ => {}
2414 }
2415
2416 let Some(ConfiguredModel { model, provider }) =
2417 LanguageModelRegistry::read_global(cx).thread_summary_model()
2418 else {
2419 return;
2420 };
2421
2422 if !provider.is_authenticated(cx) {
2423 return;
2424 }
2425
2426 let added_user_message = include_str!("./prompts/summarize_thread_detailed_prompt.txt");
2427
2428 let request = self.to_summarize_request(
2429 &model,
2430 CompletionIntent::ThreadContextSummarization,
2431 added_user_message.into(),
2432 cx,
2433 );
2434
2435 *self.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generating {
2436 message_id: last_message_id,
2437 };
2438
2439 // Replace the detailed summarization task if there is one, cancelling it. It would probably
2440 // be better to allow the old task to complete, but this would require logic for choosing
2441 // which result to prefer (the old task could complete after the new one, resulting in a
2442 // stale summary).
2443 self.detailed_summary_task = cx.spawn(async move |thread, cx| {
2444 let stream = model.stream_completion_text(request, &cx);
2445 let Some(mut messages) = stream.await.log_err() else {
2446 thread
2447 .update(cx, |thread, _cx| {
2448 *thread.detailed_summary_tx.borrow_mut() =
2449 DetailedSummaryState::NotGenerated;
2450 })
2451 .ok()?;
2452 return None;
2453 };
2454
2455 let mut new_detailed_summary = String::new();
2456
2457 while let Some(chunk) = messages.stream.next().await {
2458 if let Some(chunk) = chunk.log_err() {
2459 new_detailed_summary.push_str(&chunk);
2460 }
2461 }
2462
2463 thread
2464 .update(cx, |thread, _cx| {
2465 *thread.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generated {
2466 text: new_detailed_summary.into(),
2467 message_id: last_message_id,
2468 };
2469 })
2470 .ok()?;
2471
2472 // Save thread so its summary can be reused later
2473 if let Some(thread) = thread.upgrade() {
2474 if let Ok(Ok(save_task)) = cx.update(|cx| {
2475 thread_store
2476 .update(cx, |thread_store, cx| thread_store.save_thread(&thread, cx))
2477 }) {
2478 save_task.await.log_err();
2479 }
2480 }
2481
2482 Some(())
2483 });
2484 }
2485
2486 pub async fn wait_for_detailed_summary_or_text(
2487 this: &Entity<Self>,
2488 cx: &mut AsyncApp,
2489 ) -> Option<SharedString> {
2490 let mut detailed_summary_rx = this
2491 .read_with(cx, |this, _cx| this.detailed_summary_rx.clone())
2492 .ok()?;
2493 loop {
2494 match detailed_summary_rx.recv().await? {
2495 DetailedSummaryState::Generating { .. } => {}
2496 DetailedSummaryState::NotGenerated => {
2497 return this.read_with(cx, |this, _cx| this.text().into()).ok();
2498 }
2499 DetailedSummaryState::Generated { text, .. } => return Some(text),
2500 }
2501 }
2502 }
2503
2504 pub fn latest_detailed_summary_or_text(&self) -> SharedString {
2505 self.detailed_summary_rx
2506 .borrow()
2507 .text()
2508 .unwrap_or_else(|| self.text().into())
2509 }
2510
2511 pub fn is_generating_detailed_summary(&self) -> bool {
2512 matches!(
2513 &*self.detailed_summary_rx.borrow(),
2514 DetailedSummaryState::Generating { .. }
2515 )
2516 }
2517
2518 pub fn use_pending_tools(
2519 &mut self,
2520 window: Option<AnyWindowHandle>,
2521 model: Arc<dyn LanguageModel>,
2522 cx: &mut Context<Self>,
2523 ) -> Vec<PendingToolUse> {
2524 self.auto_capture_telemetry(cx);
2525 let request =
2526 Arc::new(self.to_completion_request(model.clone(), CompletionIntent::ToolResults, cx));
2527 let pending_tool_uses = self
2528 .tool_use
2529 .pending_tool_uses()
2530 .into_iter()
2531 .filter(|tool_use| tool_use.status.is_idle())
2532 .cloned()
2533 .collect::<Vec<_>>();
2534
2535 for tool_use in pending_tool_uses.iter() {
2536 self.use_pending_tool(tool_use.clone(), request.clone(), model.clone(), window, cx);
2537 }
2538
2539 pending_tool_uses
2540 }
2541
2542 fn use_pending_tool(
2543 &mut self,
2544 tool_use: PendingToolUse,
2545 request: Arc<LanguageModelRequest>,
2546 model: Arc<dyn LanguageModel>,
2547 window: Option<AnyWindowHandle>,
2548 cx: &mut Context<Self>,
2549 ) {
2550 let Some(tool) = self.tools.read(cx).tool(&tool_use.name, cx) else {
2551 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2552 };
2553
2554 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
2555 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2556 }
2557
2558 if tool.needs_confirmation(&tool_use.input, &self.project, cx)
2559 && !AgentSettings::get_global(cx).always_allow_tool_actions
2560 {
2561 self.tool_use.confirm_tool_use(
2562 tool_use.id,
2563 tool_use.ui_text,
2564 tool_use.input,
2565 request,
2566 tool,
2567 );
2568 cx.emit(ThreadEvent::ToolConfirmationNeeded);
2569 } else {
2570 self.run_tool(
2571 tool_use.id,
2572 tool_use.ui_text,
2573 tool_use.input,
2574 request,
2575 tool,
2576 model,
2577 window,
2578 cx,
2579 );
2580 }
2581 }
2582
2583 pub fn handle_hallucinated_tool_use(
2584 &mut self,
2585 tool_use_id: LanguageModelToolUseId,
2586 hallucinated_tool_name: Arc<str>,
2587 window: Option<AnyWindowHandle>,
2588 cx: &mut Context<Thread>,
2589 ) {
2590 let available_tools = self.profile.enabled_tools(cx);
2591
2592 let tool_list = available_tools
2593 .iter()
2594 .map(|(name, tool)| format!("- {}: {}", name, tool.description()))
2595 .collect::<Vec<_>>()
2596 .join("\n");
2597
2598 let error_message = format!(
2599 "The tool '{}' doesn't exist or is not enabled. Available tools:\n{}",
2600 hallucinated_tool_name, tool_list
2601 );
2602
2603 let pending_tool_use = self.tool_use.insert_tool_output(
2604 tool_use_id.clone(),
2605 hallucinated_tool_name,
2606 Err(anyhow!("Missing tool call: {error_message}")),
2607 self.configured_model.as_ref(),
2608 self.completion_mode,
2609 );
2610
2611 cx.emit(ThreadEvent::MissingToolUse {
2612 tool_use_id: tool_use_id.clone(),
2613 ui_text: error_message.into(),
2614 });
2615
2616 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2617 }
2618
2619 pub fn receive_invalid_tool_json(
2620 &mut self,
2621 tool_use_id: LanguageModelToolUseId,
2622 tool_name: Arc<str>,
2623 invalid_json: Arc<str>,
2624 error: String,
2625 window: Option<AnyWindowHandle>,
2626 cx: &mut Context<Thread>,
2627 ) {
2628 log::error!("The model returned invalid input JSON: {invalid_json}");
2629
2630 let pending_tool_use = self.tool_use.insert_tool_output(
2631 tool_use_id.clone(),
2632 tool_name,
2633 Err(anyhow!("Error parsing input JSON: {error}")),
2634 self.configured_model.as_ref(),
2635 self.completion_mode,
2636 );
2637 let ui_text = if let Some(pending_tool_use) = &pending_tool_use {
2638 pending_tool_use.ui_text.clone()
2639 } else {
2640 log::error!(
2641 "There was no pending tool use for tool use {tool_use_id}, even though it finished (with invalid input JSON)."
2642 );
2643 format!("Unknown tool {}", tool_use_id).into()
2644 };
2645
2646 cx.emit(ThreadEvent::InvalidToolInput {
2647 tool_use_id: tool_use_id.clone(),
2648 ui_text,
2649 invalid_input_json: invalid_json,
2650 });
2651
2652 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2653 }
2654
2655 pub fn run_tool(
2656 &mut self,
2657 tool_use_id: LanguageModelToolUseId,
2658 ui_text: impl Into<SharedString>,
2659 input: serde_json::Value,
2660 request: Arc<LanguageModelRequest>,
2661 tool: Arc<dyn Tool>,
2662 model: Arc<dyn LanguageModel>,
2663 window: Option<AnyWindowHandle>,
2664 cx: &mut Context<Thread>,
2665 ) {
2666 let task =
2667 self.spawn_tool_use(tool_use_id.clone(), request, input, tool, model, window, cx);
2668 self.tool_use
2669 .run_pending_tool(tool_use_id, ui_text.into(), task);
2670 }
2671
2672 fn spawn_tool_use(
2673 &mut self,
2674 tool_use_id: LanguageModelToolUseId,
2675 request: Arc<LanguageModelRequest>,
2676 input: serde_json::Value,
2677 tool: Arc<dyn Tool>,
2678 model: Arc<dyn LanguageModel>,
2679 window: Option<AnyWindowHandle>,
2680 cx: &mut Context<Thread>,
2681 ) -> Task<()> {
2682 let tool_name: Arc<str> = tool.name().into();
2683
2684 let tool_result = tool.run(
2685 input,
2686 request,
2687 self.project.clone(),
2688 self.action_log.clone(),
2689 model,
2690 window,
2691 cx,
2692 );
2693
2694 // Store the card separately if it exists
2695 if let Some(card) = tool_result.card.clone() {
2696 self.tool_use
2697 .insert_tool_result_card(tool_use_id.clone(), card);
2698 }
2699
2700 cx.spawn({
2701 async move |thread: WeakEntity<Thread>, cx| {
2702 let output = tool_result.output.await;
2703
2704 thread
2705 .update(cx, |thread, cx| {
2706 let pending_tool_use = thread.tool_use.insert_tool_output(
2707 tool_use_id.clone(),
2708 tool_name,
2709 output,
2710 thread.configured_model.as_ref(),
2711 thread.completion_mode,
2712 );
2713 thread.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2714 })
2715 .ok();
2716 }
2717 })
2718 }
2719
2720 fn tool_finished(
2721 &mut self,
2722 tool_use_id: LanguageModelToolUseId,
2723 pending_tool_use: Option<PendingToolUse>,
2724 canceled: bool,
2725 window: Option<AnyWindowHandle>,
2726 cx: &mut Context<Self>,
2727 ) {
2728 if self.all_tools_finished() {
2729 if let Some(ConfiguredModel { model, .. }) = self.configured_model.as_ref() {
2730 if !canceled {
2731 self.send_to_model(model.clone(), CompletionIntent::ToolResults, window, cx);
2732 }
2733 self.auto_capture_telemetry(cx);
2734 }
2735 }
2736
2737 cx.emit(ThreadEvent::ToolFinished {
2738 tool_use_id,
2739 pending_tool_use,
2740 });
2741 }
2742
2743 /// Cancels the last pending completion, if there are any pending.
2744 ///
2745 /// Returns whether a completion was canceled.
2746 pub fn cancel_last_completion(
2747 &mut self,
2748 window: Option<AnyWindowHandle>,
2749 cx: &mut Context<Self>,
2750 ) -> bool {
2751 let mut canceled = self.pending_completions.pop().is_some() || self.retry_state.is_some();
2752
2753 self.retry_state = None;
2754
2755 for pending_tool_use in self.tool_use.cancel_pending() {
2756 canceled = true;
2757 self.tool_finished(
2758 pending_tool_use.id.clone(),
2759 Some(pending_tool_use),
2760 true,
2761 window,
2762 cx,
2763 );
2764 }
2765
2766 if canceled {
2767 cx.emit(ThreadEvent::CompletionCanceled);
2768
2769 // When canceled, we always want to insert the checkpoint.
2770 // (We skip over finalize_pending_checkpoint, because it
2771 // would conclude we didn't have anything to insert here.)
2772 if let Some(checkpoint) = self.pending_checkpoint.take() {
2773 self.insert_checkpoint(checkpoint, cx);
2774 }
2775 } else {
2776 self.finalize_pending_checkpoint(cx);
2777 }
2778
2779 canceled
2780 }
2781
2782 /// Signals that any in-progress editing should be canceled.
2783 ///
2784 /// This method is used to notify listeners (like ActiveThread) that
2785 /// they should cancel any editing operations.
2786 pub fn cancel_editing(&mut self, cx: &mut Context<Self>) {
2787 cx.emit(ThreadEvent::CancelEditing);
2788 }
2789
2790 pub fn feedback(&self) -> Option<ThreadFeedback> {
2791 self.feedback
2792 }
2793
2794 pub fn message_feedback(&self, message_id: MessageId) -> Option<ThreadFeedback> {
2795 self.message_feedback.get(&message_id).copied()
2796 }
2797
2798 pub fn report_message_feedback(
2799 &mut self,
2800 message_id: MessageId,
2801 feedback: ThreadFeedback,
2802 cx: &mut Context<Self>,
2803 ) -> Task<Result<()>> {
2804 if self.message_feedback.get(&message_id) == Some(&feedback) {
2805 return Task::ready(Ok(()));
2806 }
2807
2808 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2809 let serialized_thread = self.serialize(cx);
2810 let thread_id = self.id().clone();
2811 let client = self.project.read(cx).client();
2812
2813 let enabled_tool_names: Vec<String> = self
2814 .profile
2815 .enabled_tools(cx)
2816 .iter()
2817 .map(|(name, _)| name.clone().into())
2818 .collect();
2819
2820 self.message_feedback.insert(message_id, feedback);
2821
2822 cx.notify();
2823
2824 let message_content = self
2825 .message(message_id)
2826 .map(|msg| msg.to_string())
2827 .unwrap_or_default();
2828
2829 cx.background_spawn(async move {
2830 let final_project_snapshot = final_project_snapshot.await;
2831 let serialized_thread = serialized_thread.await?;
2832 let thread_data =
2833 serde_json::to_value(serialized_thread).unwrap_or_else(|_| serde_json::Value::Null);
2834
2835 let rating = match feedback {
2836 ThreadFeedback::Positive => "positive",
2837 ThreadFeedback::Negative => "negative",
2838 };
2839 telemetry::event!(
2840 "Assistant Thread Rated",
2841 rating,
2842 thread_id,
2843 enabled_tool_names,
2844 message_id = message_id.0,
2845 message_content,
2846 thread_data,
2847 final_project_snapshot
2848 );
2849 client.telemetry().flush_events().await;
2850
2851 Ok(())
2852 })
2853 }
2854
2855 pub fn report_feedback(
2856 &mut self,
2857 feedback: ThreadFeedback,
2858 cx: &mut Context<Self>,
2859 ) -> Task<Result<()>> {
2860 let last_assistant_message_id = self
2861 .messages
2862 .iter()
2863 .rev()
2864 .find(|msg| msg.role == Role::Assistant)
2865 .map(|msg| msg.id);
2866
2867 if let Some(message_id) = last_assistant_message_id {
2868 self.report_message_feedback(message_id, feedback, cx)
2869 } else {
2870 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2871 let serialized_thread = self.serialize(cx);
2872 let thread_id = self.id().clone();
2873 let client = self.project.read(cx).client();
2874 self.feedback = Some(feedback);
2875 cx.notify();
2876
2877 cx.background_spawn(async move {
2878 let final_project_snapshot = final_project_snapshot.await;
2879 let serialized_thread = serialized_thread.await?;
2880 let thread_data = serde_json::to_value(serialized_thread)
2881 .unwrap_or_else(|_| serde_json::Value::Null);
2882
2883 let rating = match feedback {
2884 ThreadFeedback::Positive => "positive",
2885 ThreadFeedback::Negative => "negative",
2886 };
2887 telemetry::event!(
2888 "Assistant Thread Rated",
2889 rating,
2890 thread_id,
2891 thread_data,
2892 final_project_snapshot
2893 );
2894 client.telemetry().flush_events().await;
2895
2896 Ok(())
2897 })
2898 }
2899 }
2900
2901 /// Create a snapshot of the current project state including git information and unsaved buffers.
2902 fn project_snapshot(
2903 project: Entity<Project>,
2904 cx: &mut Context<Self>,
2905 ) -> Task<Arc<ProjectSnapshot>> {
2906 let git_store = project.read(cx).git_store().clone();
2907 let worktree_snapshots: Vec<_> = project
2908 .read(cx)
2909 .visible_worktrees(cx)
2910 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
2911 .collect();
2912
2913 cx.spawn(async move |_, cx| {
2914 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
2915
2916 let mut unsaved_buffers = Vec::new();
2917 cx.update(|app_cx| {
2918 let buffer_store = project.read(app_cx).buffer_store();
2919 for buffer_handle in buffer_store.read(app_cx).buffers() {
2920 let buffer = buffer_handle.read(app_cx);
2921 if buffer.is_dirty() {
2922 if let Some(file) = buffer.file() {
2923 let path = file.path().to_string_lossy().to_string();
2924 unsaved_buffers.push(path);
2925 }
2926 }
2927 }
2928 })
2929 .ok();
2930
2931 Arc::new(ProjectSnapshot {
2932 worktree_snapshots,
2933 unsaved_buffer_paths: unsaved_buffers,
2934 timestamp: Utc::now(),
2935 })
2936 })
2937 }
2938
2939 fn worktree_snapshot(
2940 worktree: Entity<project::Worktree>,
2941 git_store: Entity<GitStore>,
2942 cx: &App,
2943 ) -> Task<WorktreeSnapshot> {
2944 cx.spawn(async move |cx| {
2945 // Get worktree path and snapshot
2946 let worktree_info = cx.update(|app_cx| {
2947 let worktree = worktree.read(app_cx);
2948 let path = worktree.abs_path().to_string_lossy().to_string();
2949 let snapshot = worktree.snapshot();
2950 (path, snapshot)
2951 });
2952
2953 let Ok((worktree_path, _snapshot)) = worktree_info else {
2954 return WorktreeSnapshot {
2955 worktree_path: String::new(),
2956 git_state: None,
2957 };
2958 };
2959
2960 let git_state = git_store
2961 .update(cx, |git_store, cx| {
2962 git_store
2963 .repositories()
2964 .values()
2965 .find(|repo| {
2966 repo.read(cx)
2967 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
2968 .is_some()
2969 })
2970 .cloned()
2971 })
2972 .ok()
2973 .flatten()
2974 .map(|repo| {
2975 repo.update(cx, |repo, _| {
2976 let current_branch =
2977 repo.branch.as_ref().map(|branch| branch.name().to_owned());
2978 repo.send_job(None, |state, _| async move {
2979 let RepositoryState::Local { backend, .. } = state else {
2980 return GitState {
2981 remote_url: None,
2982 head_sha: None,
2983 current_branch,
2984 diff: None,
2985 };
2986 };
2987
2988 let remote_url = backend.remote_url("origin");
2989 let head_sha = backend.head_sha().await;
2990 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
2991
2992 GitState {
2993 remote_url,
2994 head_sha,
2995 current_branch,
2996 diff,
2997 }
2998 })
2999 })
3000 });
3001
3002 let git_state = match git_state {
3003 Some(git_state) => match git_state.ok() {
3004 Some(git_state) => git_state.await.ok(),
3005 None => None,
3006 },
3007 None => None,
3008 };
3009
3010 WorktreeSnapshot {
3011 worktree_path,
3012 git_state,
3013 }
3014 })
3015 }
3016
3017 pub fn to_markdown(&self, cx: &App) -> Result<String> {
3018 let mut markdown = Vec::new();
3019
3020 let summary = self.summary().or_default();
3021 writeln!(markdown, "# {summary}\n")?;
3022
3023 for message in self.messages() {
3024 writeln!(
3025 markdown,
3026 "## {role}\n",
3027 role = match message.role {
3028 Role::User => "User",
3029 Role::Assistant => "Agent",
3030 Role::System => "System",
3031 }
3032 )?;
3033
3034 if !message.loaded_context.text.is_empty() {
3035 writeln!(markdown, "{}", message.loaded_context.text)?;
3036 }
3037
3038 if !message.loaded_context.images.is_empty() {
3039 writeln!(
3040 markdown,
3041 "\n{} images attached as context.\n",
3042 message.loaded_context.images.len()
3043 )?;
3044 }
3045
3046 for segment in &message.segments {
3047 match segment {
3048 MessageSegment::Text(text) => writeln!(markdown, "{}\n", text)?,
3049 MessageSegment::Thinking { text, .. } => {
3050 writeln!(markdown, "<think>\n{}\n</think>\n", text)?
3051 }
3052 MessageSegment::RedactedThinking(_) => {}
3053 }
3054 }
3055
3056 for tool_use in self.tool_uses_for_message(message.id, cx) {
3057 writeln!(
3058 markdown,
3059 "**Use Tool: {} ({})**",
3060 tool_use.name, tool_use.id
3061 )?;
3062 writeln!(markdown, "```json")?;
3063 writeln!(
3064 markdown,
3065 "{}",
3066 serde_json::to_string_pretty(&tool_use.input)?
3067 )?;
3068 writeln!(markdown, "```")?;
3069 }
3070
3071 for tool_result in self.tool_results_for_message(message.id) {
3072 write!(markdown, "\n**Tool Results: {}", tool_result.tool_use_id)?;
3073 if tool_result.is_error {
3074 write!(markdown, " (Error)")?;
3075 }
3076
3077 writeln!(markdown, "**\n")?;
3078 match &tool_result.content {
3079 LanguageModelToolResultContent::Text(text) => {
3080 writeln!(markdown, "{text}")?;
3081 }
3082 LanguageModelToolResultContent::Image(image) => {
3083 writeln!(markdown, "", image.source)?;
3084 }
3085 }
3086
3087 if let Some(output) = tool_result.output.as_ref() {
3088 writeln!(
3089 markdown,
3090 "\n\nDebug Output:\n\n```json\n{}\n```\n",
3091 serde_json::to_string_pretty(output)?
3092 )?;
3093 }
3094 }
3095 }
3096
3097 Ok(String::from_utf8_lossy(&markdown).to_string())
3098 }
3099
3100 pub fn keep_edits_in_range(
3101 &mut self,
3102 buffer: Entity<language::Buffer>,
3103 buffer_range: Range<language::Anchor>,
3104 cx: &mut Context<Self>,
3105 ) {
3106 self.action_log.update(cx, |action_log, cx| {
3107 action_log.keep_edits_in_range(buffer, buffer_range, cx)
3108 });
3109 }
3110
3111 pub fn keep_all_edits(&mut self, cx: &mut Context<Self>) {
3112 self.action_log
3113 .update(cx, |action_log, cx| action_log.keep_all_edits(cx));
3114 }
3115
3116 pub fn reject_edits_in_ranges(
3117 &mut self,
3118 buffer: Entity<language::Buffer>,
3119 buffer_ranges: Vec<Range<language::Anchor>>,
3120 cx: &mut Context<Self>,
3121 ) -> Task<Result<()>> {
3122 self.action_log.update(cx, |action_log, cx| {
3123 action_log.reject_edits_in_ranges(buffer, buffer_ranges, cx)
3124 })
3125 }
3126
3127 pub fn action_log(&self) -> &Entity<ActionLog> {
3128 &self.action_log
3129 }
3130
3131 pub fn project(&self) -> &Entity<Project> {
3132 &self.project
3133 }
3134
3135 pub fn auto_capture_telemetry(&mut self, cx: &mut Context<Self>) {
3136 if !cx.has_flag::<feature_flags::ThreadAutoCaptureFeatureFlag>() {
3137 return;
3138 }
3139
3140 let now = Instant::now();
3141 if let Some(last) = self.last_auto_capture_at {
3142 if now.duration_since(last).as_secs() < 10 {
3143 return;
3144 }
3145 }
3146
3147 self.last_auto_capture_at = Some(now);
3148
3149 let thread_id = self.id().clone();
3150 let github_login = self
3151 .project
3152 .read(cx)
3153 .user_store()
3154 .read(cx)
3155 .current_user()
3156 .map(|user| user.github_login.clone());
3157 let client = self.project.read(cx).client();
3158 let serialize_task = self.serialize(cx);
3159
3160 cx.background_executor()
3161 .spawn(async move {
3162 if let Ok(serialized_thread) = serialize_task.await {
3163 if let Ok(thread_data) = serde_json::to_value(serialized_thread) {
3164 telemetry::event!(
3165 "Agent Thread Auto-Captured",
3166 thread_id = thread_id.to_string(),
3167 thread_data = thread_data,
3168 auto_capture_reason = "tracked_user",
3169 github_login = github_login
3170 );
3171
3172 client.telemetry().flush_events().await;
3173 }
3174 }
3175 })
3176 .detach();
3177 }
3178
3179 pub fn cumulative_token_usage(&self) -> TokenUsage {
3180 self.cumulative_token_usage
3181 }
3182
3183 pub fn token_usage_up_to_message(&self, message_id: MessageId) -> TotalTokenUsage {
3184 let Some(model) = self.configured_model.as_ref() else {
3185 return TotalTokenUsage::default();
3186 };
3187
3188 let max = model
3189 .model
3190 .max_token_count_for_mode(self.completion_mode().into());
3191
3192 let index = self
3193 .messages
3194 .iter()
3195 .position(|msg| msg.id == message_id)
3196 .unwrap_or(0);
3197
3198 if index == 0 {
3199 return TotalTokenUsage { total: 0, max };
3200 }
3201
3202 let token_usage = &self
3203 .request_token_usage
3204 .get(index - 1)
3205 .cloned()
3206 .unwrap_or_default();
3207
3208 TotalTokenUsage {
3209 total: token_usage.total_tokens(),
3210 max,
3211 }
3212 }
3213
3214 pub fn total_token_usage(&self) -> Option<TotalTokenUsage> {
3215 let model = self.configured_model.as_ref()?;
3216
3217 let max = model
3218 .model
3219 .max_token_count_for_mode(self.completion_mode().into());
3220
3221 if let Some(exceeded_error) = &self.exceeded_window_error {
3222 if model.model.id() == exceeded_error.model_id {
3223 return Some(TotalTokenUsage {
3224 total: exceeded_error.token_count,
3225 max,
3226 });
3227 }
3228 }
3229
3230 let total = self
3231 .token_usage_at_last_message()
3232 .unwrap_or_default()
3233 .total_tokens();
3234
3235 Some(TotalTokenUsage { total, max })
3236 }
3237
3238 fn token_usage_at_last_message(&self) -> Option<TokenUsage> {
3239 self.request_token_usage
3240 .get(self.messages.len().saturating_sub(1))
3241 .or_else(|| self.request_token_usage.last())
3242 .cloned()
3243 }
3244
3245 fn update_token_usage_at_last_message(&mut self, token_usage: TokenUsage) {
3246 let placeholder = self.token_usage_at_last_message().unwrap_or_default();
3247 self.request_token_usage
3248 .resize(self.messages.len(), placeholder);
3249
3250 if let Some(last) = self.request_token_usage.last_mut() {
3251 *last = token_usage;
3252 }
3253 }
3254
3255 fn update_model_request_usage(&self, amount: u32, limit: UsageLimit, cx: &mut Context<Self>) {
3256 self.project
3257 .read(cx)
3258 .user_store()
3259 .update(cx, |user_store, cx| {
3260 user_store.update_model_request_usage(
3261 ModelRequestUsage(RequestUsage {
3262 amount: amount as i32,
3263 limit,
3264 }),
3265 cx,
3266 )
3267 });
3268 }
3269
3270 pub fn deny_tool_use(
3271 &mut self,
3272 tool_use_id: LanguageModelToolUseId,
3273 tool_name: Arc<str>,
3274 window: Option<AnyWindowHandle>,
3275 cx: &mut Context<Self>,
3276 ) {
3277 let err = Err(anyhow::anyhow!(
3278 "Permission to run tool action denied by user"
3279 ));
3280
3281 self.tool_use.insert_tool_output(
3282 tool_use_id.clone(),
3283 tool_name,
3284 err,
3285 self.configured_model.as_ref(),
3286 self.completion_mode,
3287 );
3288 self.tool_finished(tool_use_id.clone(), None, true, window, cx);
3289 }
3290}
3291
3292#[derive(Debug, Clone, Error)]
3293pub enum ThreadError {
3294 #[error("Payment required")]
3295 PaymentRequired,
3296 #[error("Model request limit reached")]
3297 ModelRequestLimitReached { plan: Plan },
3298 #[error("Message {header}: {message}")]
3299 Message {
3300 header: SharedString,
3301 message: SharedString,
3302 },
3303 #[error("Retryable error: {message}")]
3304 RetryableError {
3305 message: SharedString,
3306 can_enable_burn_mode: bool,
3307 },
3308}
3309
3310#[derive(Debug, Clone)]
3311pub enum ThreadEvent {
3312 ShowError(ThreadError),
3313 StreamedCompletion,
3314 ReceivedTextChunk,
3315 NewRequest,
3316 StreamedAssistantText(MessageId, String),
3317 StreamedAssistantThinking(MessageId, String),
3318 StreamedToolUse {
3319 tool_use_id: LanguageModelToolUseId,
3320 ui_text: Arc<str>,
3321 input: serde_json::Value,
3322 },
3323 MissingToolUse {
3324 tool_use_id: LanguageModelToolUseId,
3325 ui_text: Arc<str>,
3326 },
3327 InvalidToolInput {
3328 tool_use_id: LanguageModelToolUseId,
3329 ui_text: Arc<str>,
3330 invalid_input_json: Arc<str>,
3331 },
3332 Stopped(Result<StopReason, Arc<anyhow::Error>>),
3333 MessageAdded(MessageId),
3334 MessageEdited(MessageId),
3335 MessageDeleted(MessageId),
3336 SummaryGenerated,
3337 SummaryChanged,
3338 UsePendingTools {
3339 tool_uses: Vec<PendingToolUse>,
3340 },
3341 ToolFinished {
3342 #[allow(unused)]
3343 tool_use_id: LanguageModelToolUseId,
3344 /// The pending tool use that corresponds to this tool.
3345 pending_tool_use: Option<PendingToolUse>,
3346 },
3347 CheckpointChanged,
3348 ToolConfirmationNeeded,
3349 ToolUseLimitReached,
3350 CancelEditing,
3351 CompletionCanceled,
3352 ProfileChanged,
3353}
3354
3355impl EventEmitter<ThreadEvent> for Thread {}
3356
3357struct PendingCompletion {
3358 id: usize,
3359 queue_state: QueueState,
3360 _task: Task<()>,
3361}
3362
3363#[cfg(test)]
3364mod tests {
3365 use super::*;
3366 use crate::{
3367 context::load_context, context_store::ContextStore, thread_store, thread_store::ThreadStore,
3368 };
3369
3370 // Test-specific constants
3371 const TEST_RATE_LIMIT_RETRY_SECS: u64 = 30;
3372 use agent_settings::{AgentProfileId, AgentSettings, LanguageModelParameters};
3373 use assistant_tool::ToolRegistry;
3374 use assistant_tools;
3375 use futures::StreamExt;
3376 use futures::future::BoxFuture;
3377 use futures::stream::BoxStream;
3378 use gpui::TestAppContext;
3379 use http_client;
3380 use language_model::fake_provider::{FakeLanguageModel, FakeLanguageModelProvider};
3381 use language_model::{
3382 LanguageModelCompletionError, LanguageModelName, LanguageModelProviderId,
3383 LanguageModelProviderName, LanguageModelToolChoice,
3384 };
3385 use parking_lot::Mutex;
3386 use project::{FakeFs, Project};
3387 use prompt_store::PromptBuilder;
3388 use serde_json::json;
3389 use settings::{Settings, SettingsStore};
3390 use std::sync::Arc;
3391 use std::time::Duration;
3392 use theme::ThemeSettings;
3393 use util::path;
3394 use workspace::Workspace;
3395
3396 #[gpui::test]
3397 async fn test_message_with_context(cx: &mut TestAppContext) {
3398 init_test_settings(cx);
3399
3400 let project = create_test_project(
3401 cx,
3402 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3403 )
3404 .await;
3405
3406 let (_workspace, _thread_store, thread, context_store, model) =
3407 setup_test_environment(cx, project.clone()).await;
3408
3409 add_file_to_context(&project, &context_store, "test/code.rs", cx)
3410 .await
3411 .unwrap();
3412
3413 let context =
3414 context_store.read_with(cx, |store, _| store.context().next().cloned().unwrap());
3415 let loaded_context = cx
3416 .update(|cx| load_context(vec![context], &project, &None, cx))
3417 .await;
3418
3419 // Insert user message with context
3420 let message_id = thread.update(cx, |thread, cx| {
3421 thread.insert_user_message(
3422 "Please explain this code",
3423 loaded_context,
3424 None,
3425 Vec::new(),
3426 cx,
3427 )
3428 });
3429
3430 // Check content and context in message object
3431 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3432
3433 // Use different path format strings based on platform for the test
3434 #[cfg(windows)]
3435 let path_part = r"test\code.rs";
3436 #[cfg(not(windows))]
3437 let path_part = "test/code.rs";
3438
3439 let expected_context = format!(
3440 r#"
3441<context>
3442The following items were attached by the user. They are up-to-date and don't need to be re-read.
3443
3444<files>
3445```rs {path_part}
3446fn main() {{
3447 println!("Hello, world!");
3448}}
3449```
3450</files>
3451</context>
3452"#
3453 );
3454
3455 assert_eq!(message.role, Role::User);
3456 assert_eq!(message.segments.len(), 1);
3457 assert_eq!(
3458 message.segments[0],
3459 MessageSegment::Text("Please explain this code".to_string())
3460 );
3461 assert_eq!(message.loaded_context.text, expected_context);
3462
3463 // Check message in request
3464 let request = thread.update(cx, |thread, cx| {
3465 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3466 });
3467
3468 assert_eq!(request.messages.len(), 2);
3469 let expected_full_message = format!("{}Please explain this code", expected_context);
3470 assert_eq!(request.messages[1].string_contents(), expected_full_message);
3471 }
3472
3473 #[gpui::test]
3474 async fn test_only_include_new_contexts(cx: &mut TestAppContext) {
3475 init_test_settings(cx);
3476
3477 let project = create_test_project(
3478 cx,
3479 json!({
3480 "file1.rs": "fn function1() {}\n",
3481 "file2.rs": "fn function2() {}\n",
3482 "file3.rs": "fn function3() {}\n",
3483 "file4.rs": "fn function4() {}\n",
3484 }),
3485 )
3486 .await;
3487
3488 let (_, _thread_store, thread, context_store, model) =
3489 setup_test_environment(cx, project.clone()).await;
3490
3491 // First message with context 1
3492 add_file_to_context(&project, &context_store, "test/file1.rs", cx)
3493 .await
3494 .unwrap();
3495 let new_contexts = context_store.update(cx, |store, cx| {
3496 store.new_context_for_thread(thread.read(cx), None)
3497 });
3498 assert_eq!(new_contexts.len(), 1);
3499 let loaded_context = cx
3500 .update(|cx| load_context(new_contexts, &project, &None, cx))
3501 .await;
3502 let message1_id = thread.update(cx, |thread, cx| {
3503 thread.insert_user_message("Message 1", loaded_context, None, Vec::new(), cx)
3504 });
3505
3506 // Second message with contexts 1 and 2 (context 1 should be skipped as it's already included)
3507 add_file_to_context(&project, &context_store, "test/file2.rs", cx)
3508 .await
3509 .unwrap();
3510 let new_contexts = context_store.update(cx, |store, cx| {
3511 store.new_context_for_thread(thread.read(cx), None)
3512 });
3513 assert_eq!(new_contexts.len(), 1);
3514 let loaded_context = cx
3515 .update(|cx| load_context(new_contexts, &project, &None, cx))
3516 .await;
3517 let message2_id = thread.update(cx, |thread, cx| {
3518 thread.insert_user_message("Message 2", loaded_context, None, Vec::new(), cx)
3519 });
3520
3521 // Third message with all three contexts (contexts 1 and 2 should be skipped)
3522 //
3523 add_file_to_context(&project, &context_store, "test/file3.rs", cx)
3524 .await
3525 .unwrap();
3526 let new_contexts = context_store.update(cx, |store, cx| {
3527 store.new_context_for_thread(thread.read(cx), None)
3528 });
3529 assert_eq!(new_contexts.len(), 1);
3530 let loaded_context = cx
3531 .update(|cx| load_context(new_contexts, &project, &None, cx))
3532 .await;
3533 let message3_id = thread.update(cx, |thread, cx| {
3534 thread.insert_user_message("Message 3", loaded_context, None, Vec::new(), cx)
3535 });
3536
3537 // Check what contexts are included in each message
3538 let (message1, message2, message3) = thread.read_with(cx, |thread, _| {
3539 (
3540 thread.message(message1_id).unwrap().clone(),
3541 thread.message(message2_id).unwrap().clone(),
3542 thread.message(message3_id).unwrap().clone(),
3543 )
3544 });
3545
3546 // First message should include context 1
3547 assert!(message1.loaded_context.text.contains("file1.rs"));
3548
3549 // Second message should include only context 2 (not 1)
3550 assert!(!message2.loaded_context.text.contains("file1.rs"));
3551 assert!(message2.loaded_context.text.contains("file2.rs"));
3552
3553 // Third message should include only context 3 (not 1 or 2)
3554 assert!(!message3.loaded_context.text.contains("file1.rs"));
3555 assert!(!message3.loaded_context.text.contains("file2.rs"));
3556 assert!(message3.loaded_context.text.contains("file3.rs"));
3557
3558 // Check entire request to make sure all contexts are properly included
3559 let request = thread.update(cx, |thread, cx| {
3560 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3561 });
3562
3563 // The request should contain all 3 messages
3564 assert_eq!(request.messages.len(), 4);
3565
3566 // Check that the contexts are properly formatted in each message
3567 assert!(request.messages[1].string_contents().contains("file1.rs"));
3568 assert!(!request.messages[1].string_contents().contains("file2.rs"));
3569 assert!(!request.messages[1].string_contents().contains("file3.rs"));
3570
3571 assert!(!request.messages[2].string_contents().contains("file1.rs"));
3572 assert!(request.messages[2].string_contents().contains("file2.rs"));
3573 assert!(!request.messages[2].string_contents().contains("file3.rs"));
3574
3575 assert!(!request.messages[3].string_contents().contains("file1.rs"));
3576 assert!(!request.messages[3].string_contents().contains("file2.rs"));
3577 assert!(request.messages[3].string_contents().contains("file3.rs"));
3578
3579 add_file_to_context(&project, &context_store, "test/file4.rs", cx)
3580 .await
3581 .unwrap();
3582 let new_contexts = context_store.update(cx, |store, cx| {
3583 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3584 });
3585 assert_eq!(new_contexts.len(), 3);
3586 let loaded_context = cx
3587 .update(|cx| load_context(new_contexts, &project, &None, cx))
3588 .await
3589 .loaded_context;
3590
3591 assert!(!loaded_context.text.contains("file1.rs"));
3592 assert!(loaded_context.text.contains("file2.rs"));
3593 assert!(loaded_context.text.contains("file3.rs"));
3594 assert!(loaded_context.text.contains("file4.rs"));
3595
3596 let new_contexts = context_store.update(cx, |store, cx| {
3597 // Remove file4.rs
3598 store.remove_context(&loaded_context.contexts[2].handle(), cx);
3599 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3600 });
3601 assert_eq!(new_contexts.len(), 2);
3602 let loaded_context = cx
3603 .update(|cx| load_context(new_contexts, &project, &None, cx))
3604 .await
3605 .loaded_context;
3606
3607 assert!(!loaded_context.text.contains("file1.rs"));
3608 assert!(loaded_context.text.contains("file2.rs"));
3609 assert!(loaded_context.text.contains("file3.rs"));
3610 assert!(!loaded_context.text.contains("file4.rs"));
3611
3612 let new_contexts = context_store.update(cx, |store, cx| {
3613 // Remove file3.rs
3614 store.remove_context(&loaded_context.contexts[1].handle(), cx);
3615 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3616 });
3617 assert_eq!(new_contexts.len(), 1);
3618 let loaded_context = cx
3619 .update(|cx| load_context(new_contexts, &project, &None, cx))
3620 .await
3621 .loaded_context;
3622
3623 assert!(!loaded_context.text.contains("file1.rs"));
3624 assert!(loaded_context.text.contains("file2.rs"));
3625 assert!(!loaded_context.text.contains("file3.rs"));
3626 assert!(!loaded_context.text.contains("file4.rs"));
3627 }
3628
3629 #[gpui::test]
3630 async fn test_message_without_files(cx: &mut TestAppContext) {
3631 init_test_settings(cx);
3632
3633 let project = create_test_project(
3634 cx,
3635 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3636 )
3637 .await;
3638
3639 let (_, _thread_store, thread, _context_store, model) =
3640 setup_test_environment(cx, project.clone()).await;
3641
3642 // Insert user message without any context (empty context vector)
3643 let message_id = thread.update(cx, |thread, cx| {
3644 thread.insert_user_message(
3645 "What is the best way to learn Rust?",
3646 ContextLoadResult::default(),
3647 None,
3648 Vec::new(),
3649 cx,
3650 )
3651 });
3652
3653 // Check content and context in message object
3654 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3655
3656 // Context should be empty when no files are included
3657 assert_eq!(message.role, Role::User);
3658 assert_eq!(message.segments.len(), 1);
3659 assert_eq!(
3660 message.segments[0],
3661 MessageSegment::Text("What is the best way to learn Rust?".to_string())
3662 );
3663 assert_eq!(message.loaded_context.text, "");
3664
3665 // Check message in request
3666 let request = thread.update(cx, |thread, cx| {
3667 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3668 });
3669
3670 assert_eq!(request.messages.len(), 2);
3671 assert_eq!(
3672 request.messages[1].string_contents(),
3673 "What is the best way to learn Rust?"
3674 );
3675
3676 // Add second message, also without context
3677 let message2_id = thread.update(cx, |thread, cx| {
3678 thread.insert_user_message(
3679 "Are there any good books?",
3680 ContextLoadResult::default(),
3681 None,
3682 Vec::new(),
3683 cx,
3684 )
3685 });
3686
3687 let message2 =
3688 thread.read_with(cx, |thread, _| thread.message(message2_id).unwrap().clone());
3689 assert_eq!(message2.loaded_context.text, "");
3690
3691 // Check that both messages appear in the request
3692 let request = thread.update(cx, |thread, cx| {
3693 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3694 });
3695
3696 assert_eq!(request.messages.len(), 3);
3697 assert_eq!(
3698 request.messages[1].string_contents(),
3699 "What is the best way to learn Rust?"
3700 );
3701 assert_eq!(
3702 request.messages[2].string_contents(),
3703 "Are there any good books?"
3704 );
3705 }
3706
3707 #[gpui::test]
3708 #[ignore] // turn this test on when project_notifications tool is re-enabled
3709 async fn test_stale_buffer_notification(cx: &mut TestAppContext) {
3710 init_test_settings(cx);
3711
3712 let project = create_test_project(
3713 cx,
3714 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3715 )
3716 .await;
3717
3718 let (_workspace, _thread_store, thread, context_store, model) =
3719 setup_test_environment(cx, project.clone()).await;
3720
3721 // Add a buffer to the context. This will be a tracked buffer
3722 let buffer = add_file_to_context(&project, &context_store, "test/code.rs", cx)
3723 .await
3724 .unwrap();
3725
3726 let context = context_store
3727 .read_with(cx, |store, _| store.context().next().cloned())
3728 .unwrap();
3729 let loaded_context = cx
3730 .update(|cx| load_context(vec![context], &project, &None, cx))
3731 .await;
3732
3733 // Insert user message and assistant response
3734 thread.update(cx, |thread, cx| {
3735 thread.insert_user_message("Explain this code", loaded_context, None, Vec::new(), cx);
3736 thread.insert_assistant_message(
3737 vec![MessageSegment::Text("This code prints 42.".into())],
3738 cx,
3739 );
3740 });
3741 cx.run_until_parked();
3742
3743 // We shouldn't have a stale buffer notification yet
3744 let notifications = thread.read_with(cx, |thread, _| {
3745 find_tool_uses(thread, "project_notifications")
3746 });
3747 assert!(
3748 notifications.is_empty(),
3749 "Should not have stale buffer notification before buffer is modified"
3750 );
3751
3752 // Modify the buffer
3753 buffer.update(cx, |buffer, cx| {
3754 buffer.edit(
3755 [(1..1, "\n println!(\"Added a new line\");\n")],
3756 None,
3757 cx,
3758 );
3759 });
3760
3761 // Insert another user message
3762 thread.update(cx, |thread, cx| {
3763 thread.insert_user_message(
3764 "What does the code do now?",
3765 ContextLoadResult::default(),
3766 None,
3767 Vec::new(),
3768 cx,
3769 )
3770 });
3771 cx.run_until_parked();
3772
3773 // Check for the stale buffer warning
3774 thread.update(cx, |thread, cx| {
3775 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3776 });
3777 cx.run_until_parked();
3778
3779 let notifications = thread.read_with(cx, |thread, _cx| {
3780 find_tool_uses(thread, "project_notifications")
3781 });
3782
3783 let [notification] = notifications.as_slice() else {
3784 panic!("Should have a `project_notifications` tool use");
3785 };
3786
3787 let Some(notification_content) = notification.content.to_str() else {
3788 panic!("`project_notifications` should return text");
3789 };
3790
3791 assert!(notification_content.contains("These files have changed since the last read:"));
3792 assert!(notification_content.contains("code.rs"));
3793
3794 // Insert another user message and flush notifications again
3795 thread.update(cx, |thread, cx| {
3796 thread.insert_user_message(
3797 "Can you tell me more?",
3798 ContextLoadResult::default(),
3799 None,
3800 Vec::new(),
3801 cx,
3802 )
3803 });
3804
3805 thread.update(cx, |thread, cx| {
3806 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3807 });
3808 cx.run_until_parked();
3809
3810 // There should be no new notifications (we already flushed one)
3811 let notifications = thread.read_with(cx, |thread, _cx| {
3812 find_tool_uses(thread, "project_notifications")
3813 });
3814
3815 assert_eq!(
3816 notifications.len(),
3817 1,
3818 "Should still have only one notification after second flush - no duplicates"
3819 );
3820 }
3821
3822 fn find_tool_uses(thread: &Thread, tool_name: &str) -> Vec<LanguageModelToolResult> {
3823 thread
3824 .messages()
3825 .flat_map(|message| {
3826 thread
3827 .tool_results_for_message(message.id)
3828 .into_iter()
3829 .filter(|result| result.tool_name == tool_name.into())
3830 .cloned()
3831 .collect::<Vec<_>>()
3832 })
3833 .collect()
3834 }
3835
3836 #[gpui::test]
3837 async fn test_storing_profile_setting_per_thread(cx: &mut TestAppContext) {
3838 init_test_settings(cx);
3839
3840 let project = create_test_project(
3841 cx,
3842 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3843 )
3844 .await;
3845
3846 let (_workspace, thread_store, thread, _context_store, _model) =
3847 setup_test_environment(cx, project.clone()).await;
3848
3849 // Check that we are starting with the default profile
3850 let profile = cx.read(|cx| thread.read(cx).profile.clone());
3851 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3852 assert_eq!(
3853 profile,
3854 AgentProfile::new(AgentProfileId::default(), tool_set)
3855 );
3856 }
3857
3858 #[gpui::test]
3859 async fn test_serializing_thread_profile(cx: &mut TestAppContext) {
3860 init_test_settings(cx);
3861
3862 let project = create_test_project(
3863 cx,
3864 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3865 )
3866 .await;
3867
3868 let (_workspace, thread_store, thread, _context_store, _model) =
3869 setup_test_environment(cx, project.clone()).await;
3870
3871 // Profile gets serialized with default values
3872 let serialized = thread
3873 .update(cx, |thread, cx| thread.serialize(cx))
3874 .await
3875 .unwrap();
3876
3877 assert_eq!(serialized.profile, Some(AgentProfileId::default()));
3878
3879 let deserialized = cx.update(|cx| {
3880 thread.update(cx, |thread, cx| {
3881 Thread::deserialize(
3882 thread.id.clone(),
3883 serialized,
3884 thread.project.clone(),
3885 thread.tools.clone(),
3886 thread.prompt_builder.clone(),
3887 thread.project_context.clone(),
3888 None,
3889 cx,
3890 )
3891 })
3892 });
3893 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3894
3895 assert_eq!(
3896 deserialized.profile,
3897 AgentProfile::new(AgentProfileId::default(), tool_set)
3898 );
3899 }
3900
3901 #[gpui::test]
3902 async fn test_temperature_setting(cx: &mut TestAppContext) {
3903 init_test_settings(cx);
3904
3905 let project = create_test_project(
3906 cx,
3907 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3908 )
3909 .await;
3910
3911 let (_workspace, _thread_store, thread, _context_store, model) =
3912 setup_test_environment(cx, project.clone()).await;
3913
3914 // Both model and provider
3915 cx.update(|cx| {
3916 AgentSettings::override_global(
3917 AgentSettings {
3918 model_parameters: vec![LanguageModelParameters {
3919 provider: Some(model.provider_id().0.to_string().into()),
3920 model: Some(model.id().0.clone()),
3921 temperature: Some(0.66),
3922 }],
3923 ..AgentSettings::get_global(cx).clone()
3924 },
3925 cx,
3926 );
3927 });
3928
3929 let request = thread.update(cx, |thread, cx| {
3930 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3931 });
3932 assert_eq!(request.temperature, Some(0.66));
3933
3934 // Only model
3935 cx.update(|cx| {
3936 AgentSettings::override_global(
3937 AgentSettings {
3938 model_parameters: vec![LanguageModelParameters {
3939 provider: None,
3940 model: Some(model.id().0.clone()),
3941 temperature: Some(0.66),
3942 }],
3943 ..AgentSettings::get_global(cx).clone()
3944 },
3945 cx,
3946 );
3947 });
3948
3949 let request = thread.update(cx, |thread, cx| {
3950 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3951 });
3952 assert_eq!(request.temperature, Some(0.66));
3953
3954 // Only provider
3955 cx.update(|cx| {
3956 AgentSettings::override_global(
3957 AgentSettings {
3958 model_parameters: vec![LanguageModelParameters {
3959 provider: Some(model.provider_id().0.to_string().into()),
3960 model: None,
3961 temperature: Some(0.66),
3962 }],
3963 ..AgentSettings::get_global(cx).clone()
3964 },
3965 cx,
3966 );
3967 });
3968
3969 let request = thread.update(cx, |thread, cx| {
3970 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3971 });
3972 assert_eq!(request.temperature, Some(0.66));
3973
3974 // Same model name, different provider
3975 cx.update(|cx| {
3976 AgentSettings::override_global(
3977 AgentSettings {
3978 model_parameters: vec![LanguageModelParameters {
3979 provider: Some("anthropic".into()),
3980 model: Some(model.id().0.clone()),
3981 temperature: Some(0.66),
3982 }],
3983 ..AgentSettings::get_global(cx).clone()
3984 },
3985 cx,
3986 );
3987 });
3988
3989 let request = thread.update(cx, |thread, cx| {
3990 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3991 });
3992 assert_eq!(request.temperature, None);
3993 }
3994
3995 #[gpui::test]
3996 async fn test_thread_summary(cx: &mut TestAppContext) {
3997 init_test_settings(cx);
3998
3999 let project = create_test_project(cx, json!({})).await;
4000
4001 let (_, _thread_store, thread, _context_store, model) =
4002 setup_test_environment(cx, project.clone()).await;
4003
4004 // Initial state should be pending
4005 thread.read_with(cx, |thread, _| {
4006 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4007 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4008 });
4009
4010 // Manually setting the summary should not be allowed in this state
4011 thread.update(cx, |thread, cx| {
4012 thread.set_summary("This should not work", cx);
4013 });
4014
4015 thread.read_with(cx, |thread, _| {
4016 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4017 });
4018
4019 // Send a message
4020 thread.update(cx, |thread, cx| {
4021 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
4022 thread.send_to_model(
4023 model.clone(),
4024 CompletionIntent::ThreadSummarization,
4025 None,
4026 cx,
4027 );
4028 });
4029
4030 let fake_model = model.as_fake();
4031 simulate_successful_response(&fake_model, cx);
4032
4033 // Should start generating summary when there are >= 2 messages
4034 thread.read_with(cx, |thread, _| {
4035 assert_eq!(*thread.summary(), ThreadSummary::Generating);
4036 });
4037
4038 // Should not be able to set the summary while generating
4039 thread.update(cx, |thread, cx| {
4040 thread.set_summary("This should not work either", cx);
4041 });
4042
4043 thread.read_with(cx, |thread, _| {
4044 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4045 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4046 });
4047
4048 cx.run_until_parked();
4049 fake_model.send_last_completion_stream_text_chunk("Brief");
4050 fake_model.send_last_completion_stream_text_chunk(" Introduction");
4051 fake_model.end_last_completion_stream();
4052 cx.run_until_parked();
4053
4054 // Summary should be set
4055 thread.read_with(cx, |thread, _| {
4056 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4057 assert_eq!(thread.summary().or_default(), "Brief Introduction");
4058 });
4059
4060 // Now we should be able to set a summary
4061 thread.update(cx, |thread, cx| {
4062 thread.set_summary("Brief Intro", cx);
4063 });
4064
4065 thread.read_with(cx, |thread, _| {
4066 assert_eq!(thread.summary().or_default(), "Brief Intro");
4067 });
4068
4069 // Test setting an empty summary (should default to DEFAULT)
4070 thread.update(cx, |thread, cx| {
4071 thread.set_summary("", cx);
4072 });
4073
4074 thread.read_with(cx, |thread, _| {
4075 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4076 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4077 });
4078 }
4079
4080 #[gpui::test]
4081 async fn test_thread_summary_error_set_manually(cx: &mut TestAppContext) {
4082 init_test_settings(cx);
4083
4084 let project = create_test_project(cx, json!({})).await;
4085
4086 let (_, _thread_store, thread, _context_store, model) =
4087 setup_test_environment(cx, project.clone()).await;
4088
4089 test_summarize_error(&model, &thread, cx);
4090
4091 // Now we should be able to set a summary
4092 thread.update(cx, |thread, cx| {
4093 thread.set_summary("Brief Intro", cx);
4094 });
4095
4096 thread.read_with(cx, |thread, _| {
4097 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4098 assert_eq!(thread.summary().or_default(), "Brief Intro");
4099 });
4100 }
4101
4102 #[gpui::test]
4103 async fn test_thread_summary_error_retry(cx: &mut TestAppContext) {
4104 init_test_settings(cx);
4105
4106 let project = create_test_project(cx, json!({})).await;
4107
4108 let (_, _thread_store, thread, _context_store, model) =
4109 setup_test_environment(cx, project.clone()).await;
4110
4111 test_summarize_error(&model, &thread, cx);
4112
4113 // Sending another message should not trigger another summarize request
4114 thread.update(cx, |thread, cx| {
4115 thread.insert_user_message(
4116 "How are you?",
4117 ContextLoadResult::default(),
4118 None,
4119 vec![],
4120 cx,
4121 );
4122 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4123 });
4124
4125 let fake_model = model.as_fake();
4126 simulate_successful_response(&fake_model, cx);
4127
4128 thread.read_with(cx, |thread, _| {
4129 // State is still Error, not Generating
4130 assert!(matches!(thread.summary(), ThreadSummary::Error));
4131 });
4132
4133 // But the summarize request can be invoked manually
4134 thread.update(cx, |thread, cx| {
4135 thread.summarize(cx);
4136 });
4137
4138 thread.read_with(cx, |thread, _| {
4139 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4140 });
4141
4142 cx.run_until_parked();
4143 fake_model.send_last_completion_stream_text_chunk("A successful summary");
4144 fake_model.end_last_completion_stream();
4145 cx.run_until_parked();
4146
4147 thread.read_with(cx, |thread, _| {
4148 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4149 assert_eq!(thread.summary().or_default(), "A successful summary");
4150 });
4151 }
4152
4153 // Helper to create a model that returns errors
4154 enum TestError {
4155 Overloaded,
4156 InternalServerError,
4157 }
4158
4159 struct ErrorInjector {
4160 inner: Arc<FakeLanguageModel>,
4161 error_type: TestError,
4162 }
4163
4164 impl ErrorInjector {
4165 fn new(error_type: TestError) -> Self {
4166 Self {
4167 inner: Arc::new(FakeLanguageModel::default()),
4168 error_type,
4169 }
4170 }
4171 }
4172
4173 impl LanguageModel for ErrorInjector {
4174 fn id(&self) -> LanguageModelId {
4175 self.inner.id()
4176 }
4177
4178 fn name(&self) -> LanguageModelName {
4179 self.inner.name()
4180 }
4181
4182 fn provider_id(&self) -> LanguageModelProviderId {
4183 self.inner.provider_id()
4184 }
4185
4186 fn provider_name(&self) -> LanguageModelProviderName {
4187 self.inner.provider_name()
4188 }
4189
4190 fn supports_tools(&self) -> bool {
4191 self.inner.supports_tools()
4192 }
4193
4194 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4195 self.inner.supports_tool_choice(choice)
4196 }
4197
4198 fn supports_images(&self) -> bool {
4199 self.inner.supports_images()
4200 }
4201
4202 fn telemetry_id(&self) -> String {
4203 self.inner.telemetry_id()
4204 }
4205
4206 fn max_token_count(&self) -> u64 {
4207 self.inner.max_token_count()
4208 }
4209
4210 fn count_tokens(
4211 &self,
4212 request: LanguageModelRequest,
4213 cx: &App,
4214 ) -> BoxFuture<'static, Result<u64>> {
4215 self.inner.count_tokens(request, cx)
4216 }
4217
4218 fn stream_completion(
4219 &self,
4220 _request: LanguageModelRequest,
4221 _cx: &AsyncApp,
4222 ) -> BoxFuture<
4223 'static,
4224 Result<
4225 BoxStream<
4226 'static,
4227 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4228 >,
4229 LanguageModelCompletionError,
4230 >,
4231 > {
4232 let error = match self.error_type {
4233 TestError::Overloaded => LanguageModelCompletionError::ServerOverloaded {
4234 provider: self.provider_name(),
4235 retry_after: None,
4236 },
4237 TestError::InternalServerError => {
4238 LanguageModelCompletionError::ApiInternalServerError {
4239 provider: self.provider_name(),
4240 message: "I'm a teapot orbiting the sun".to_string(),
4241 }
4242 }
4243 };
4244 async move {
4245 let stream = futures::stream::once(async move { Err(error) });
4246 Ok(stream.boxed())
4247 }
4248 .boxed()
4249 }
4250
4251 fn as_fake(&self) -> &FakeLanguageModel {
4252 &self.inner
4253 }
4254 }
4255
4256 #[gpui::test]
4257 async fn test_retry_on_overloaded_error(cx: &mut TestAppContext) {
4258 init_test_settings(cx);
4259
4260 let project = create_test_project(cx, json!({})).await;
4261 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4262
4263 // Enable Burn Mode to allow retries
4264 thread.update(cx, |thread, _| {
4265 thread.set_completion_mode(CompletionMode::Burn);
4266 });
4267
4268 // Create model that returns overloaded error
4269 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4270
4271 // Insert a user message
4272 thread.update(cx, |thread, cx| {
4273 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4274 });
4275
4276 // Start completion
4277 thread.update(cx, |thread, cx| {
4278 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4279 });
4280
4281 cx.run_until_parked();
4282
4283 thread.read_with(cx, |thread, _| {
4284 assert!(thread.retry_state.is_some(), "Should have retry state");
4285 let retry_state = thread.retry_state.as_ref().unwrap();
4286 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4287 assert_eq!(
4288 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
4289 "Should retry MAX_RETRY_ATTEMPTS times for overloaded errors"
4290 );
4291 });
4292
4293 // Check that a retry message was added
4294 thread.read_with(cx, |thread, _| {
4295 let mut messages = thread.messages();
4296 assert!(
4297 messages.any(|msg| {
4298 msg.role == Role::System
4299 && msg.ui_only
4300 && msg.segments.iter().any(|seg| {
4301 if let MessageSegment::Text(text) = seg {
4302 text.contains("overloaded")
4303 && text
4304 .contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS))
4305 } else {
4306 false
4307 }
4308 })
4309 }),
4310 "Should have added a system retry message"
4311 );
4312 });
4313
4314 let retry_count = thread.update(cx, |thread, _| {
4315 thread
4316 .messages
4317 .iter()
4318 .filter(|m| {
4319 m.ui_only
4320 && m.segments.iter().any(|s| {
4321 if let MessageSegment::Text(text) = s {
4322 text.contains("Retrying") && text.contains("seconds")
4323 } else {
4324 false
4325 }
4326 })
4327 })
4328 .count()
4329 });
4330
4331 assert_eq!(retry_count, 1, "Should have one retry message");
4332 }
4333
4334 #[gpui::test]
4335 async fn test_retry_on_internal_server_error(cx: &mut TestAppContext) {
4336 init_test_settings(cx);
4337
4338 let project = create_test_project(cx, json!({})).await;
4339 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4340
4341 // Enable Burn Mode to allow retries
4342 thread.update(cx, |thread, _| {
4343 thread.set_completion_mode(CompletionMode::Burn);
4344 });
4345
4346 // Create model that returns internal server error
4347 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4348
4349 // Insert a user message
4350 thread.update(cx, |thread, cx| {
4351 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4352 });
4353
4354 // Start completion
4355 thread.update(cx, |thread, cx| {
4356 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4357 });
4358
4359 cx.run_until_parked();
4360
4361 // Check retry state on thread
4362 thread.read_with(cx, |thread, _| {
4363 assert!(thread.retry_state.is_some(), "Should have retry state");
4364 let retry_state = thread.retry_state.as_ref().unwrap();
4365 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4366 assert_eq!(
4367 retry_state.max_attempts, 3,
4368 "Should have correct max attempts"
4369 );
4370 });
4371
4372 // Check that a retry message was added with provider name
4373 thread.read_with(cx, |thread, _| {
4374 let mut messages = thread.messages();
4375 assert!(
4376 messages.any(|msg| {
4377 msg.role == Role::System
4378 && msg.ui_only
4379 && msg.segments.iter().any(|seg| {
4380 if let MessageSegment::Text(text) = seg {
4381 text.contains("internal")
4382 && text.contains("Fake")
4383 && text.contains("Retrying")
4384 && text.contains("attempt 1 of 3")
4385 && text.contains("seconds")
4386 } else {
4387 false
4388 }
4389 })
4390 }),
4391 "Should have added a system retry message with provider name"
4392 );
4393 });
4394
4395 // Count retry messages
4396 let retry_count = thread.update(cx, |thread, _| {
4397 thread
4398 .messages
4399 .iter()
4400 .filter(|m| {
4401 m.ui_only
4402 && m.segments.iter().any(|s| {
4403 if let MessageSegment::Text(text) = s {
4404 text.contains("Retrying") && text.contains("seconds")
4405 } else {
4406 false
4407 }
4408 })
4409 })
4410 .count()
4411 });
4412
4413 assert_eq!(retry_count, 1, "Should have one retry message");
4414 }
4415
4416 #[gpui::test]
4417 async fn test_exponential_backoff_on_retries(cx: &mut TestAppContext) {
4418 init_test_settings(cx);
4419
4420 let project = create_test_project(cx, json!({})).await;
4421 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4422
4423 // Enable Burn Mode to allow retries
4424 thread.update(cx, |thread, _| {
4425 thread.set_completion_mode(CompletionMode::Burn);
4426 });
4427
4428 // Create model that returns internal server error
4429 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4430
4431 // Insert a user message
4432 thread.update(cx, |thread, cx| {
4433 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4434 });
4435
4436 // Track retry events and completion count
4437 // Track completion events
4438 let completion_count = Arc::new(Mutex::new(0));
4439 let completion_count_clone = completion_count.clone();
4440
4441 let _subscription = thread.update(cx, |_, cx| {
4442 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4443 if let ThreadEvent::NewRequest = event {
4444 *completion_count_clone.lock() += 1;
4445 }
4446 })
4447 });
4448
4449 // First attempt
4450 thread.update(cx, |thread, cx| {
4451 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4452 });
4453 cx.run_until_parked();
4454
4455 // Should have scheduled first retry - count retry messages
4456 let retry_count = thread.update(cx, |thread, _| {
4457 thread
4458 .messages
4459 .iter()
4460 .filter(|m| {
4461 m.ui_only
4462 && m.segments.iter().any(|s| {
4463 if let MessageSegment::Text(text) = s {
4464 text.contains("Retrying") && text.contains("seconds")
4465 } else {
4466 false
4467 }
4468 })
4469 })
4470 .count()
4471 });
4472 assert_eq!(retry_count, 1, "Should have scheduled first retry");
4473
4474 // Check retry state
4475 thread.read_with(cx, |thread, _| {
4476 assert!(thread.retry_state.is_some(), "Should have retry state");
4477 let retry_state = thread.retry_state.as_ref().unwrap();
4478 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4479 assert_eq!(
4480 retry_state.max_attempts, 3,
4481 "Internal server errors should retry up to 3 times"
4482 );
4483 });
4484
4485 // Advance clock for first retry
4486 cx.executor().advance_clock(BASE_RETRY_DELAY);
4487 cx.run_until_parked();
4488
4489 // Advance clock for second retry
4490 cx.executor().advance_clock(BASE_RETRY_DELAY);
4491 cx.run_until_parked();
4492
4493 // Advance clock for third retry
4494 cx.executor().advance_clock(BASE_RETRY_DELAY);
4495 cx.run_until_parked();
4496
4497 // Should have completed all retries - count retry messages
4498 let retry_count = thread.update(cx, |thread, _| {
4499 thread
4500 .messages
4501 .iter()
4502 .filter(|m| {
4503 m.ui_only
4504 && m.segments.iter().any(|s| {
4505 if let MessageSegment::Text(text) = s {
4506 text.contains("Retrying") && text.contains("seconds")
4507 } else {
4508 false
4509 }
4510 })
4511 })
4512 .count()
4513 });
4514 assert_eq!(
4515 retry_count, 3,
4516 "Should have 3 retries for internal server errors"
4517 );
4518
4519 // For internal server errors, we retry 3 times and then give up
4520 // Check that retry_state is cleared after all retries
4521 thread.read_with(cx, |thread, _| {
4522 assert!(
4523 thread.retry_state.is_none(),
4524 "Retry state should be cleared after all retries"
4525 );
4526 });
4527
4528 // Verify total attempts (1 initial + 3 retries)
4529 assert_eq!(
4530 *completion_count.lock(),
4531 4,
4532 "Should have attempted once plus 3 retries"
4533 );
4534 }
4535
4536 #[gpui::test]
4537 async fn test_max_retries_exceeded(cx: &mut TestAppContext) {
4538 init_test_settings(cx);
4539
4540 let project = create_test_project(cx, json!({})).await;
4541 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4542
4543 // Enable Burn Mode to allow retries
4544 thread.update(cx, |thread, _| {
4545 thread.set_completion_mode(CompletionMode::Burn);
4546 });
4547
4548 // Create model that returns overloaded error
4549 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4550
4551 // Insert a user message
4552 thread.update(cx, |thread, cx| {
4553 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4554 });
4555
4556 // Track events
4557 let stopped_with_error = Arc::new(Mutex::new(false));
4558 let stopped_with_error_clone = stopped_with_error.clone();
4559
4560 let _subscription = thread.update(cx, |_, cx| {
4561 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4562 if let ThreadEvent::Stopped(Err(_)) = event {
4563 *stopped_with_error_clone.lock() = true;
4564 }
4565 })
4566 });
4567
4568 // Start initial completion
4569 thread.update(cx, |thread, cx| {
4570 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4571 });
4572 cx.run_until_parked();
4573
4574 // Advance through all retries
4575 for _ in 0..MAX_RETRY_ATTEMPTS {
4576 cx.executor().advance_clock(BASE_RETRY_DELAY);
4577 cx.run_until_parked();
4578 }
4579
4580 let retry_count = thread.update(cx, |thread, _| {
4581 thread
4582 .messages
4583 .iter()
4584 .filter(|m| {
4585 m.ui_only
4586 && m.segments.iter().any(|s| {
4587 if let MessageSegment::Text(text) = s {
4588 text.contains("Retrying") && text.contains("seconds")
4589 } else {
4590 false
4591 }
4592 })
4593 })
4594 .count()
4595 });
4596
4597 // After max retries, should emit Stopped(Err(...)) event
4598 assert_eq!(
4599 retry_count, MAX_RETRY_ATTEMPTS as usize,
4600 "Should have attempted MAX_RETRY_ATTEMPTS retries for overloaded errors"
4601 );
4602 assert!(
4603 *stopped_with_error.lock(),
4604 "Should emit Stopped(Err(...)) event after max retries exceeded"
4605 );
4606
4607 // Retry state should be cleared
4608 thread.read_with(cx, |thread, _| {
4609 assert!(
4610 thread.retry_state.is_none(),
4611 "Retry state should be cleared after max retries"
4612 );
4613
4614 // Verify we have the expected number of retry messages
4615 let retry_messages = thread
4616 .messages
4617 .iter()
4618 .filter(|msg| msg.ui_only && msg.role == Role::System)
4619 .count();
4620 assert_eq!(
4621 retry_messages, MAX_RETRY_ATTEMPTS as usize,
4622 "Should have MAX_RETRY_ATTEMPTS retry messages for overloaded errors"
4623 );
4624 });
4625 }
4626
4627 #[gpui::test]
4628 async fn test_retry_message_removed_on_retry(cx: &mut TestAppContext) {
4629 init_test_settings(cx);
4630
4631 let project = create_test_project(cx, json!({})).await;
4632 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4633
4634 // Enable Burn Mode to allow retries
4635 thread.update(cx, |thread, _| {
4636 thread.set_completion_mode(CompletionMode::Burn);
4637 });
4638
4639 // We'll use a wrapper to switch behavior after first failure
4640 struct RetryTestModel {
4641 inner: Arc<FakeLanguageModel>,
4642 failed_once: Arc<Mutex<bool>>,
4643 }
4644
4645 impl LanguageModel for RetryTestModel {
4646 fn id(&self) -> LanguageModelId {
4647 self.inner.id()
4648 }
4649
4650 fn name(&self) -> LanguageModelName {
4651 self.inner.name()
4652 }
4653
4654 fn provider_id(&self) -> LanguageModelProviderId {
4655 self.inner.provider_id()
4656 }
4657
4658 fn provider_name(&self) -> LanguageModelProviderName {
4659 self.inner.provider_name()
4660 }
4661
4662 fn supports_tools(&self) -> bool {
4663 self.inner.supports_tools()
4664 }
4665
4666 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4667 self.inner.supports_tool_choice(choice)
4668 }
4669
4670 fn supports_images(&self) -> bool {
4671 self.inner.supports_images()
4672 }
4673
4674 fn telemetry_id(&self) -> String {
4675 self.inner.telemetry_id()
4676 }
4677
4678 fn max_token_count(&self) -> u64 {
4679 self.inner.max_token_count()
4680 }
4681
4682 fn count_tokens(
4683 &self,
4684 request: LanguageModelRequest,
4685 cx: &App,
4686 ) -> BoxFuture<'static, Result<u64>> {
4687 self.inner.count_tokens(request, cx)
4688 }
4689
4690 fn stream_completion(
4691 &self,
4692 request: LanguageModelRequest,
4693 cx: &AsyncApp,
4694 ) -> BoxFuture<
4695 'static,
4696 Result<
4697 BoxStream<
4698 'static,
4699 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4700 >,
4701 LanguageModelCompletionError,
4702 >,
4703 > {
4704 if !*self.failed_once.lock() {
4705 *self.failed_once.lock() = true;
4706 let provider = self.provider_name();
4707 // Return error on first attempt
4708 let stream = futures::stream::once(async move {
4709 Err(LanguageModelCompletionError::ServerOverloaded {
4710 provider,
4711 retry_after: None,
4712 })
4713 });
4714 async move { Ok(stream.boxed()) }.boxed()
4715 } else {
4716 // Succeed on retry
4717 self.inner.stream_completion(request, cx)
4718 }
4719 }
4720
4721 fn as_fake(&self) -> &FakeLanguageModel {
4722 &self.inner
4723 }
4724 }
4725
4726 let model = Arc::new(RetryTestModel {
4727 inner: Arc::new(FakeLanguageModel::default()),
4728 failed_once: Arc::new(Mutex::new(false)),
4729 });
4730
4731 // Insert a user message
4732 thread.update(cx, |thread, cx| {
4733 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4734 });
4735
4736 // Track message deletions
4737 // Track when retry completes successfully
4738 let retry_completed = Arc::new(Mutex::new(false));
4739 let retry_completed_clone = retry_completed.clone();
4740
4741 let _subscription = thread.update(cx, |_, cx| {
4742 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4743 if let ThreadEvent::StreamedCompletion = event {
4744 *retry_completed_clone.lock() = true;
4745 }
4746 })
4747 });
4748
4749 // Start completion
4750 thread.update(cx, |thread, cx| {
4751 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4752 });
4753 cx.run_until_parked();
4754
4755 // Get the retry message ID
4756 let retry_message_id = thread.read_with(cx, |thread, _| {
4757 thread
4758 .messages()
4759 .find(|msg| msg.role == Role::System && msg.ui_only)
4760 .map(|msg| msg.id)
4761 .expect("Should have a retry message")
4762 });
4763
4764 // Wait for retry
4765 cx.executor().advance_clock(BASE_RETRY_DELAY);
4766 cx.run_until_parked();
4767
4768 // Stream some successful content
4769 let fake_model = model.as_fake();
4770 // After the retry, there should be a new pending completion
4771 let pending = fake_model.pending_completions();
4772 assert!(
4773 !pending.is_empty(),
4774 "Should have a pending completion after retry"
4775 );
4776 fake_model.send_completion_stream_text_chunk(&pending[0], "Success!");
4777 fake_model.end_completion_stream(&pending[0]);
4778 cx.run_until_parked();
4779
4780 // Check that the retry completed successfully
4781 assert!(
4782 *retry_completed.lock(),
4783 "Retry should have completed successfully"
4784 );
4785
4786 // Retry message should still exist but be marked as ui_only
4787 thread.read_with(cx, |thread, _| {
4788 let retry_msg = thread
4789 .message(retry_message_id)
4790 .expect("Retry message should still exist");
4791 assert!(retry_msg.ui_only, "Retry message should be ui_only");
4792 assert_eq!(
4793 retry_msg.role,
4794 Role::System,
4795 "Retry message should have System role"
4796 );
4797 });
4798 }
4799
4800 #[gpui::test]
4801 async fn test_successful_completion_clears_retry_state(cx: &mut TestAppContext) {
4802 init_test_settings(cx);
4803
4804 let project = create_test_project(cx, json!({})).await;
4805 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4806
4807 // Enable Burn Mode to allow retries
4808 thread.update(cx, |thread, _| {
4809 thread.set_completion_mode(CompletionMode::Burn);
4810 });
4811
4812 // Create a model that fails once then succeeds
4813 struct FailOnceModel {
4814 inner: Arc<FakeLanguageModel>,
4815 failed_once: Arc<Mutex<bool>>,
4816 }
4817
4818 impl LanguageModel for FailOnceModel {
4819 fn id(&self) -> LanguageModelId {
4820 self.inner.id()
4821 }
4822
4823 fn name(&self) -> LanguageModelName {
4824 self.inner.name()
4825 }
4826
4827 fn provider_id(&self) -> LanguageModelProviderId {
4828 self.inner.provider_id()
4829 }
4830
4831 fn provider_name(&self) -> LanguageModelProviderName {
4832 self.inner.provider_name()
4833 }
4834
4835 fn supports_tools(&self) -> bool {
4836 self.inner.supports_tools()
4837 }
4838
4839 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4840 self.inner.supports_tool_choice(choice)
4841 }
4842
4843 fn supports_images(&self) -> bool {
4844 self.inner.supports_images()
4845 }
4846
4847 fn telemetry_id(&self) -> String {
4848 self.inner.telemetry_id()
4849 }
4850
4851 fn max_token_count(&self) -> u64 {
4852 self.inner.max_token_count()
4853 }
4854
4855 fn count_tokens(
4856 &self,
4857 request: LanguageModelRequest,
4858 cx: &App,
4859 ) -> BoxFuture<'static, Result<u64>> {
4860 self.inner.count_tokens(request, cx)
4861 }
4862
4863 fn stream_completion(
4864 &self,
4865 request: LanguageModelRequest,
4866 cx: &AsyncApp,
4867 ) -> BoxFuture<
4868 'static,
4869 Result<
4870 BoxStream<
4871 'static,
4872 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4873 >,
4874 LanguageModelCompletionError,
4875 >,
4876 > {
4877 if !*self.failed_once.lock() {
4878 *self.failed_once.lock() = true;
4879 let provider = self.provider_name();
4880 // Return error on first attempt
4881 let stream = futures::stream::once(async move {
4882 Err(LanguageModelCompletionError::ServerOverloaded {
4883 provider,
4884 retry_after: None,
4885 })
4886 });
4887 async move { Ok(stream.boxed()) }.boxed()
4888 } else {
4889 // Succeed on retry
4890 self.inner.stream_completion(request, cx)
4891 }
4892 }
4893 }
4894
4895 let fail_once_model = Arc::new(FailOnceModel {
4896 inner: Arc::new(FakeLanguageModel::default()),
4897 failed_once: Arc::new(Mutex::new(false)),
4898 });
4899
4900 // Insert a user message
4901 thread.update(cx, |thread, cx| {
4902 thread.insert_user_message(
4903 "Test message",
4904 ContextLoadResult::default(),
4905 None,
4906 vec![],
4907 cx,
4908 );
4909 });
4910
4911 // Start completion with fail-once model
4912 thread.update(cx, |thread, cx| {
4913 thread.send_to_model(
4914 fail_once_model.clone(),
4915 CompletionIntent::UserPrompt,
4916 None,
4917 cx,
4918 );
4919 });
4920
4921 cx.run_until_parked();
4922
4923 // Verify retry state exists after first failure
4924 thread.read_with(cx, |thread, _| {
4925 assert!(
4926 thread.retry_state.is_some(),
4927 "Should have retry state after failure"
4928 );
4929 });
4930
4931 // Wait for retry delay
4932 cx.executor().advance_clock(BASE_RETRY_DELAY);
4933 cx.run_until_parked();
4934
4935 // The retry should now use our FailOnceModel which should succeed
4936 // We need to help the FakeLanguageModel complete the stream
4937 let inner_fake = fail_once_model.inner.clone();
4938
4939 // Wait a bit for the retry to start
4940 cx.run_until_parked();
4941
4942 // Check for pending completions and complete them
4943 if let Some(pending) = inner_fake.pending_completions().first() {
4944 inner_fake.send_completion_stream_text_chunk(pending, "Success!");
4945 inner_fake.end_completion_stream(pending);
4946 }
4947 cx.run_until_parked();
4948
4949 thread.read_with(cx, |thread, _| {
4950 assert!(
4951 thread.retry_state.is_none(),
4952 "Retry state should be cleared after successful completion"
4953 );
4954
4955 let has_assistant_message = thread
4956 .messages
4957 .iter()
4958 .any(|msg| msg.role == Role::Assistant && !msg.ui_only);
4959 assert!(
4960 has_assistant_message,
4961 "Should have an assistant message after successful retry"
4962 );
4963 });
4964 }
4965
4966 #[gpui::test]
4967 async fn test_rate_limit_retry_single_attempt(cx: &mut TestAppContext) {
4968 init_test_settings(cx);
4969
4970 let project = create_test_project(cx, json!({})).await;
4971 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4972
4973 // Enable Burn Mode to allow retries
4974 thread.update(cx, |thread, _| {
4975 thread.set_completion_mode(CompletionMode::Burn);
4976 });
4977
4978 // Create a model that returns rate limit error with retry_after
4979 struct RateLimitModel {
4980 inner: Arc<FakeLanguageModel>,
4981 }
4982
4983 impl LanguageModel for RateLimitModel {
4984 fn id(&self) -> LanguageModelId {
4985 self.inner.id()
4986 }
4987
4988 fn name(&self) -> LanguageModelName {
4989 self.inner.name()
4990 }
4991
4992 fn provider_id(&self) -> LanguageModelProviderId {
4993 self.inner.provider_id()
4994 }
4995
4996 fn provider_name(&self) -> LanguageModelProviderName {
4997 self.inner.provider_name()
4998 }
4999
5000 fn supports_tools(&self) -> bool {
5001 self.inner.supports_tools()
5002 }
5003
5004 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
5005 self.inner.supports_tool_choice(choice)
5006 }
5007
5008 fn supports_images(&self) -> bool {
5009 self.inner.supports_images()
5010 }
5011
5012 fn telemetry_id(&self) -> String {
5013 self.inner.telemetry_id()
5014 }
5015
5016 fn max_token_count(&self) -> u64 {
5017 self.inner.max_token_count()
5018 }
5019
5020 fn count_tokens(
5021 &self,
5022 request: LanguageModelRequest,
5023 cx: &App,
5024 ) -> BoxFuture<'static, Result<u64>> {
5025 self.inner.count_tokens(request, cx)
5026 }
5027
5028 fn stream_completion(
5029 &self,
5030 _request: LanguageModelRequest,
5031 _cx: &AsyncApp,
5032 ) -> BoxFuture<
5033 'static,
5034 Result<
5035 BoxStream<
5036 'static,
5037 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
5038 >,
5039 LanguageModelCompletionError,
5040 >,
5041 > {
5042 let provider = self.provider_name();
5043 async move {
5044 let stream = futures::stream::once(async move {
5045 Err(LanguageModelCompletionError::RateLimitExceeded {
5046 provider,
5047 retry_after: Some(Duration::from_secs(TEST_RATE_LIMIT_RETRY_SECS)),
5048 })
5049 });
5050 Ok(stream.boxed())
5051 }
5052 .boxed()
5053 }
5054
5055 fn as_fake(&self) -> &FakeLanguageModel {
5056 &self.inner
5057 }
5058 }
5059
5060 let model = Arc::new(RateLimitModel {
5061 inner: Arc::new(FakeLanguageModel::default()),
5062 });
5063
5064 // Insert a user message
5065 thread.update(cx, |thread, cx| {
5066 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5067 });
5068
5069 // Start completion
5070 thread.update(cx, |thread, cx| {
5071 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5072 });
5073
5074 cx.run_until_parked();
5075
5076 let retry_count = thread.update(cx, |thread, _| {
5077 thread
5078 .messages
5079 .iter()
5080 .filter(|m| {
5081 m.ui_only
5082 && m.segments.iter().any(|s| {
5083 if let MessageSegment::Text(text) = s {
5084 text.contains("rate limit exceeded")
5085 } else {
5086 false
5087 }
5088 })
5089 })
5090 .count()
5091 });
5092 assert_eq!(retry_count, 1, "Should have scheduled one retry");
5093
5094 thread.read_with(cx, |thread, _| {
5095 assert!(
5096 thread.retry_state.is_some(),
5097 "Rate limit errors should set retry_state"
5098 );
5099 if let Some(retry_state) = &thread.retry_state {
5100 assert_eq!(
5101 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
5102 "Rate limit errors should use MAX_RETRY_ATTEMPTS"
5103 );
5104 }
5105 });
5106
5107 // Verify we have one retry message
5108 thread.read_with(cx, |thread, _| {
5109 let retry_messages = thread
5110 .messages
5111 .iter()
5112 .filter(|msg| {
5113 msg.ui_only
5114 && msg.segments.iter().any(|seg| {
5115 if let MessageSegment::Text(text) = seg {
5116 text.contains("rate limit exceeded")
5117 } else {
5118 false
5119 }
5120 })
5121 })
5122 .count();
5123 assert_eq!(
5124 retry_messages, 1,
5125 "Should have one rate limit retry message"
5126 );
5127 });
5128
5129 // Check that retry message doesn't include attempt count
5130 thread.read_with(cx, |thread, _| {
5131 let retry_message = thread
5132 .messages
5133 .iter()
5134 .find(|msg| msg.role == Role::System && msg.ui_only)
5135 .expect("Should have a retry message");
5136
5137 // Check that the message contains attempt count since we use retry_state
5138 if let Some(MessageSegment::Text(text)) = retry_message.segments.first() {
5139 assert!(
5140 text.contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS)),
5141 "Rate limit retry message should contain attempt count with MAX_RETRY_ATTEMPTS"
5142 );
5143 assert!(
5144 text.contains("Retrying"),
5145 "Rate limit retry message should contain retry text"
5146 );
5147 }
5148 });
5149 }
5150
5151 #[gpui::test]
5152 async fn test_ui_only_messages_not_sent_to_model(cx: &mut TestAppContext) {
5153 init_test_settings(cx);
5154
5155 let project = create_test_project(cx, json!({})).await;
5156 let (_, _, thread, _, model) = setup_test_environment(cx, project.clone()).await;
5157
5158 // Insert a regular user message
5159 thread.update(cx, |thread, cx| {
5160 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5161 });
5162
5163 // Insert a UI-only message (like our retry notifications)
5164 thread.update(cx, |thread, cx| {
5165 let id = thread.next_message_id.post_inc();
5166 thread.messages.push(Message {
5167 id,
5168 role: Role::System,
5169 segments: vec![MessageSegment::Text(
5170 "This is a UI-only message that should not be sent to the model".to_string(),
5171 )],
5172 loaded_context: LoadedContext::default(),
5173 creases: Vec::new(),
5174 is_hidden: true,
5175 ui_only: true,
5176 });
5177 cx.emit(ThreadEvent::MessageAdded(id));
5178 });
5179
5180 // Insert another regular message
5181 thread.update(cx, |thread, cx| {
5182 thread.insert_user_message(
5183 "How are you?",
5184 ContextLoadResult::default(),
5185 None,
5186 vec![],
5187 cx,
5188 );
5189 });
5190
5191 // Generate the completion request
5192 let request = thread.update(cx, |thread, cx| {
5193 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
5194 });
5195
5196 // Verify that the request only contains non-UI-only messages
5197 // Should have system prompt + 2 user messages, but not the UI-only message
5198 let user_messages: Vec<_> = request
5199 .messages
5200 .iter()
5201 .filter(|msg| msg.role == Role::User)
5202 .collect();
5203 assert_eq!(
5204 user_messages.len(),
5205 2,
5206 "Should have exactly 2 user messages"
5207 );
5208
5209 // Verify the UI-only content is not present anywhere in the request
5210 let request_text = request
5211 .messages
5212 .iter()
5213 .flat_map(|msg| &msg.content)
5214 .filter_map(|content| match content {
5215 MessageContent::Text(text) => Some(text.as_str()),
5216 _ => None,
5217 })
5218 .collect::<String>();
5219
5220 assert!(
5221 !request_text.contains("UI-only message"),
5222 "UI-only message content should not be in the request"
5223 );
5224
5225 // Verify the thread still has all 3 messages (including UI-only)
5226 thread.read_with(cx, |thread, _| {
5227 assert_eq!(
5228 thread.messages().count(),
5229 3,
5230 "Thread should have 3 messages"
5231 );
5232 assert_eq!(
5233 thread.messages().filter(|m| m.ui_only).count(),
5234 1,
5235 "Thread should have 1 UI-only message"
5236 );
5237 });
5238
5239 // Verify that UI-only messages are not serialized
5240 let serialized = thread
5241 .update(cx, |thread, cx| thread.serialize(cx))
5242 .await
5243 .unwrap();
5244 assert_eq!(
5245 serialized.messages.len(),
5246 2,
5247 "Serialized thread should only have 2 messages (no UI-only)"
5248 );
5249 }
5250
5251 #[gpui::test]
5252 async fn test_no_retry_without_burn_mode(cx: &mut TestAppContext) {
5253 init_test_settings(cx);
5254
5255 let project = create_test_project(cx, json!({})).await;
5256 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5257
5258 // Ensure we're in Normal mode (not Burn mode)
5259 thread.update(cx, |thread, _| {
5260 thread.set_completion_mode(CompletionMode::Normal);
5261 });
5262
5263 // Track error events
5264 let error_events = Arc::new(Mutex::new(Vec::new()));
5265 let error_events_clone = error_events.clone();
5266
5267 let _subscription = thread.update(cx, |_, cx| {
5268 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
5269 if let ThreadEvent::ShowError(error) = event {
5270 error_events_clone.lock().push(error.clone());
5271 }
5272 })
5273 });
5274
5275 // Create model that returns overloaded error
5276 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5277
5278 // Insert a user message
5279 thread.update(cx, |thread, cx| {
5280 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5281 });
5282
5283 // Start completion
5284 thread.update(cx, |thread, cx| {
5285 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5286 });
5287
5288 cx.run_until_parked();
5289
5290 // Verify no retry state was created
5291 thread.read_with(cx, |thread, _| {
5292 assert!(
5293 thread.retry_state.is_none(),
5294 "Should not have retry state in Normal mode"
5295 );
5296 });
5297
5298 // Check that a retryable error was reported
5299 let errors = error_events.lock();
5300 assert!(!errors.is_empty(), "Should have received an error event");
5301
5302 if let ThreadError::RetryableError {
5303 message: _,
5304 can_enable_burn_mode,
5305 } = &errors[0]
5306 {
5307 assert!(
5308 *can_enable_burn_mode,
5309 "Error should indicate burn mode can be enabled"
5310 );
5311 } else {
5312 panic!("Expected RetryableError, got {:?}", errors[0]);
5313 }
5314
5315 // Verify the thread is no longer generating
5316 thread.read_with(cx, |thread, _| {
5317 assert!(
5318 !thread.is_generating(),
5319 "Should not be generating after error without retry"
5320 );
5321 });
5322 }
5323
5324 #[gpui::test]
5325 async fn test_retry_cancelled_on_stop(cx: &mut TestAppContext) {
5326 init_test_settings(cx);
5327
5328 let project = create_test_project(cx, json!({})).await;
5329 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5330
5331 // Enable Burn Mode to allow retries
5332 thread.update(cx, |thread, _| {
5333 thread.set_completion_mode(CompletionMode::Burn);
5334 });
5335
5336 // Create model that returns overloaded error
5337 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5338
5339 // Insert a user message
5340 thread.update(cx, |thread, cx| {
5341 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5342 });
5343
5344 // Start completion
5345 thread.update(cx, |thread, cx| {
5346 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5347 });
5348
5349 cx.run_until_parked();
5350
5351 // Verify retry was scheduled by checking for retry message
5352 let has_retry_message = thread.read_with(cx, |thread, _| {
5353 thread.messages.iter().any(|m| {
5354 m.ui_only
5355 && m.segments.iter().any(|s| {
5356 if let MessageSegment::Text(text) = s {
5357 text.contains("Retrying") && text.contains("seconds")
5358 } else {
5359 false
5360 }
5361 })
5362 })
5363 });
5364 assert!(has_retry_message, "Should have scheduled a retry");
5365
5366 // Cancel the completion before the retry happens
5367 thread.update(cx, |thread, cx| {
5368 thread.cancel_last_completion(None, cx);
5369 });
5370
5371 cx.run_until_parked();
5372
5373 // The retry should not have happened - no pending completions
5374 let fake_model = model.as_fake();
5375 assert_eq!(
5376 fake_model.pending_completions().len(),
5377 0,
5378 "Should have no pending completions after cancellation"
5379 );
5380
5381 // Verify the retry was cancelled by checking retry state
5382 thread.read_with(cx, |thread, _| {
5383 if let Some(retry_state) = &thread.retry_state {
5384 panic!(
5385 "retry_state should be cleared after cancellation, but found: attempt={}, max_attempts={}, intent={:?}",
5386 retry_state.attempt, retry_state.max_attempts, retry_state.intent
5387 );
5388 }
5389 });
5390 }
5391
5392 fn test_summarize_error(
5393 model: &Arc<dyn LanguageModel>,
5394 thread: &Entity<Thread>,
5395 cx: &mut TestAppContext,
5396 ) {
5397 thread.update(cx, |thread, cx| {
5398 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
5399 thread.send_to_model(
5400 model.clone(),
5401 CompletionIntent::ThreadSummarization,
5402 None,
5403 cx,
5404 );
5405 });
5406
5407 let fake_model = model.as_fake();
5408 simulate_successful_response(&fake_model, cx);
5409
5410 thread.read_with(cx, |thread, _| {
5411 assert!(matches!(thread.summary(), ThreadSummary::Generating));
5412 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5413 });
5414
5415 // Simulate summary request ending
5416 cx.run_until_parked();
5417 fake_model.end_last_completion_stream();
5418 cx.run_until_parked();
5419
5420 // State is set to Error and default message
5421 thread.read_with(cx, |thread, _| {
5422 assert!(matches!(thread.summary(), ThreadSummary::Error));
5423 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5424 });
5425 }
5426
5427 fn simulate_successful_response(fake_model: &FakeLanguageModel, cx: &mut TestAppContext) {
5428 cx.run_until_parked();
5429 fake_model.send_last_completion_stream_text_chunk("Assistant response");
5430 fake_model.end_last_completion_stream();
5431 cx.run_until_parked();
5432 }
5433
5434 fn init_test_settings(cx: &mut TestAppContext) {
5435 cx.update(|cx| {
5436 let settings_store = SettingsStore::test(cx);
5437 cx.set_global(settings_store);
5438 language::init(cx);
5439 Project::init_settings(cx);
5440 AgentSettings::register(cx);
5441 prompt_store::init(cx);
5442 thread_store::init(cx);
5443 workspace::init_settings(cx);
5444 language_model::init_settings(cx);
5445 ThemeSettings::register(cx);
5446 ToolRegistry::default_global(cx);
5447 assistant_tool::init(cx);
5448
5449 let http_client = Arc::new(http_client::HttpClientWithUrl::new(
5450 http_client::FakeHttpClient::with_200_response(),
5451 "http://localhost".to_string(),
5452 None,
5453 ));
5454 assistant_tools::init(http_client, cx);
5455 });
5456 }
5457
5458 // Helper to create a test project with test files
5459 async fn create_test_project(
5460 cx: &mut TestAppContext,
5461 files: serde_json::Value,
5462 ) -> Entity<Project> {
5463 let fs = FakeFs::new(cx.executor());
5464 fs.insert_tree(path!("/test"), files).await;
5465 Project::test(fs, [path!("/test").as_ref()], cx).await
5466 }
5467
5468 async fn setup_test_environment(
5469 cx: &mut TestAppContext,
5470 project: Entity<Project>,
5471 ) -> (
5472 Entity<Workspace>,
5473 Entity<ThreadStore>,
5474 Entity<Thread>,
5475 Entity<ContextStore>,
5476 Arc<dyn LanguageModel>,
5477 ) {
5478 let (workspace, cx) =
5479 cx.add_window_view(|window, cx| Workspace::test_new(project.clone(), window, cx));
5480
5481 let thread_store = cx
5482 .update(|_, cx| {
5483 ThreadStore::load(
5484 project.clone(),
5485 cx.new(|_| ToolWorkingSet::default()),
5486 None,
5487 Arc::new(PromptBuilder::new(None).unwrap()),
5488 cx,
5489 )
5490 })
5491 .await
5492 .unwrap();
5493
5494 let thread = thread_store.update(cx, |store, cx| store.create_thread(cx));
5495 let context_store = cx.new(|_cx| ContextStore::new(project.downgrade(), None));
5496
5497 let provider = Arc::new(FakeLanguageModelProvider::default());
5498 let model = provider.test_model();
5499 let model: Arc<dyn LanguageModel> = Arc::new(model);
5500
5501 cx.update(|_, cx| {
5502 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
5503 registry.set_default_model(
5504 Some(ConfiguredModel {
5505 provider: provider.clone(),
5506 model: model.clone(),
5507 }),
5508 cx,
5509 );
5510 registry.set_thread_summary_model(
5511 Some(ConfiguredModel {
5512 provider,
5513 model: model.clone(),
5514 }),
5515 cx,
5516 );
5517 })
5518 });
5519
5520 (workspace, thread_store, thread, context_store, model)
5521 }
5522
5523 async fn add_file_to_context(
5524 project: &Entity<Project>,
5525 context_store: &Entity<ContextStore>,
5526 path: &str,
5527 cx: &mut TestAppContext,
5528 ) -> Result<Entity<language::Buffer>> {
5529 let buffer_path = project
5530 .read_with(cx, |project, cx| project.find_project_path(path, cx))
5531 .unwrap();
5532
5533 let buffer = project
5534 .update(cx, |project, cx| {
5535 project.open_buffer(buffer_path.clone(), cx)
5536 })
5537 .await
5538 .unwrap();
5539
5540 context_store.update(cx, |context_store, cx| {
5541 context_store.add_file_from_buffer(&buffer_path, buffer.clone(), false, cx);
5542 });
5543
5544 Ok(buffer)
5545 }
5546}