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 dbg!("finalize_pending_checkpoint");
817 let pending_checkpoint = if self.is_generating() {
818 return;
819 } else if let Some(checkpoint) = self.pending_checkpoint.take() {
820 checkpoint
821 } else {
822 return;
823 };
824
825 self.finalize_checkpoint(pending_checkpoint, cx);
826 }
827
828 fn finalize_checkpoint(
829 &mut self,
830 pending_checkpoint: ThreadCheckpoint,
831 cx: &mut Context<Self>,
832 ) {
833 dbg!("finalize_checkpoint");
834 let git_store = self.project.read(cx).git_store().clone();
835 let final_checkpoint = git_store.update(cx, |git_store, cx| git_store.checkpoint(cx));
836 cx.spawn(async move |this, cx| match final_checkpoint.await {
837 Ok(final_checkpoint) => {
838 dbg!(&pending_checkpoint.git_checkpoint);
839 dbg!(&final_checkpoint);
840 let equal = git_store
841 .update(cx, |store, cx| {
842 store.compare_checkpoints(
843 pending_checkpoint.git_checkpoint.clone(),
844 final_checkpoint.clone(),
845 cx,
846 )
847 })?
848 .await
849 .unwrap_or(false);
850
851 if dbg!(!equal) {
852 this.update(cx, |this, cx| {
853 this.insert_checkpoint(pending_checkpoint, cx)
854 })?;
855 }
856
857 Ok(())
858 }
859 Err(_) => this.update(cx, |this, cx| {
860 this.insert_checkpoint(pending_checkpoint, cx)
861 }),
862 })
863 .detach();
864 }
865
866 fn insert_checkpoint(&mut self, checkpoint: ThreadCheckpoint, cx: &mut Context<Self>) {
867 dbg!("insert_checkpoint");
868 self.checkpoints_by_message
869 .insert(checkpoint.message_id, checkpoint);
870 cx.emit(ThreadEvent::CheckpointChanged);
871 cx.notify();
872 }
873
874 pub fn last_restore_checkpoint(&self) -> Option<&LastRestoreCheckpoint> {
875 dbg!();
876 self.last_restore_checkpoint.as_ref()
877 }
878
879 pub fn truncate(&mut self, message_id: MessageId, cx: &mut Context<Self>) {
880 let Some(message_ix) = self
881 .messages
882 .iter()
883 .rposition(|message| message.id == message_id)
884 else {
885 return;
886 };
887 for deleted_message in self.messages.drain(message_ix..) {
888 self.checkpoints_by_message.remove(&deleted_message.id);
889 }
890 cx.notify();
891 }
892
893 pub fn context_for_message(&self, id: MessageId) -> impl Iterator<Item = &AgentContext> {
894 self.messages
895 .iter()
896 .find(|message| message.id == id)
897 .into_iter()
898 .flat_map(|message| message.loaded_context.contexts.iter())
899 }
900
901 pub fn is_turn_end(&self, ix: usize) -> bool {
902 if self.messages.is_empty() {
903 return false;
904 }
905
906 if !self.is_generating() && ix == self.messages.len() - 1 {
907 return true;
908 }
909
910 let Some(message) = self.messages.get(ix) else {
911 return false;
912 };
913
914 if message.role != Role::Assistant {
915 return false;
916 }
917
918 self.messages
919 .get(ix + 1)
920 .and_then(|message| {
921 self.message(message.id)
922 .map(|next_message| next_message.role == Role::User && !next_message.is_hidden)
923 })
924 .unwrap_or(false)
925 }
926
927 pub fn tool_use_limit_reached(&self) -> bool {
928 self.tool_use_limit_reached
929 }
930
931 /// Returns whether all of the tool uses have finished running.
932 pub fn all_tools_finished(&self) -> bool {
933 // If the only pending tool uses left are the ones with errors, then
934 // that means that we've finished running all of the pending tools.
935 self.tool_use
936 .pending_tool_uses()
937 .iter()
938 .all(|pending_tool_use| pending_tool_use.status.is_error())
939 }
940
941 /// Returns whether any pending tool uses may perform edits
942 pub fn has_pending_edit_tool_uses(&self) -> bool {
943 self.tool_use
944 .pending_tool_uses()
945 .iter()
946 .filter(|pending_tool_use| !pending_tool_use.status.is_error())
947 .any(|pending_tool_use| pending_tool_use.may_perform_edits)
948 }
949
950 pub fn tool_uses_for_message(&self, id: MessageId, cx: &App) -> Vec<ToolUse> {
951 self.tool_use.tool_uses_for_message(id, &self.project, cx)
952 }
953
954 pub fn tool_results_for_message(
955 &self,
956 assistant_message_id: MessageId,
957 ) -> Vec<&LanguageModelToolResult> {
958 self.tool_use.tool_results_for_message(assistant_message_id)
959 }
960
961 pub fn tool_result(&self, id: &LanguageModelToolUseId) -> Option<&LanguageModelToolResult> {
962 self.tool_use.tool_result(id)
963 }
964
965 pub fn output_for_tool(&self, id: &LanguageModelToolUseId) -> Option<&Arc<str>> {
966 match &self.tool_use.tool_result(id)?.content {
967 LanguageModelToolResultContent::Text(text) => Some(text),
968 LanguageModelToolResultContent::Image(_) => {
969 // TODO: We should display image
970 None
971 }
972 }
973 }
974
975 pub fn card_for_tool(&self, id: &LanguageModelToolUseId) -> Option<AnyToolCard> {
976 self.tool_use.tool_result_card(id).cloned()
977 }
978
979 /// Return tools that are both enabled and supported by the model
980 pub fn available_tools(
981 &self,
982 cx: &App,
983 model: Arc<dyn LanguageModel>,
984 ) -> Vec<LanguageModelRequestTool> {
985 if model.supports_tools() {
986 self.profile
987 .enabled_tools(cx)
988 .into_iter()
989 .filter_map(|(name, tool)| {
990 // Skip tools that cannot be supported
991 let input_schema = tool.input_schema(model.tool_input_format()).ok()?;
992 Some(LanguageModelRequestTool {
993 name: name.into(),
994 description: tool.description(),
995 input_schema,
996 })
997 })
998 .collect()
999 } else {
1000 Vec::default()
1001 }
1002 }
1003
1004 pub fn insert_user_message(
1005 &mut self,
1006 text: impl Into<String>,
1007 loaded_context: ContextLoadResult,
1008 git_checkpoint: Option<GitStoreCheckpoint>,
1009 creases: Vec<MessageCrease>,
1010 cx: &mut Context<Self>,
1011 ) -> MessageId {
1012 if !loaded_context.referenced_buffers.is_empty() {
1013 self.action_log.update(cx, |log, cx| {
1014 for buffer in loaded_context.referenced_buffers {
1015 log.buffer_read(buffer, cx);
1016 }
1017 });
1018 }
1019
1020 let message_id = self.insert_message(
1021 Role::User,
1022 vec![MessageSegment::Text(text.into())],
1023 loaded_context.loaded_context,
1024 creases,
1025 false,
1026 cx,
1027 );
1028
1029 if let Some(git_checkpoint) = git_checkpoint {
1030 self.pending_checkpoint = Some(ThreadCheckpoint {
1031 message_id,
1032 git_checkpoint,
1033 });
1034 }
1035
1036 self.auto_capture_telemetry(cx);
1037
1038 message_id
1039 }
1040
1041 pub fn insert_invisible_continue_message(&mut self, cx: &mut Context<Self>) -> MessageId {
1042 let id = self.insert_message(
1043 Role::User,
1044 vec![MessageSegment::Text("Continue where you left off".into())],
1045 LoadedContext::default(),
1046 vec![],
1047 true,
1048 cx,
1049 );
1050 self.pending_checkpoint = None;
1051
1052 id
1053 }
1054
1055 pub fn insert_assistant_message(
1056 &mut self,
1057 segments: Vec<MessageSegment>,
1058 cx: &mut Context<Self>,
1059 ) -> MessageId {
1060 self.insert_message(
1061 Role::Assistant,
1062 segments,
1063 LoadedContext::default(),
1064 Vec::new(),
1065 false,
1066 cx,
1067 )
1068 }
1069
1070 pub fn insert_message(
1071 &mut self,
1072 role: Role,
1073 segments: Vec<MessageSegment>,
1074 loaded_context: LoadedContext,
1075 creases: Vec<MessageCrease>,
1076 is_hidden: bool,
1077 cx: &mut Context<Self>,
1078 ) -> MessageId {
1079 let id = self.next_message_id.post_inc();
1080 self.messages.push(Message {
1081 id,
1082 role,
1083 segments,
1084 loaded_context,
1085 creases,
1086 is_hidden,
1087 ui_only: false,
1088 });
1089 self.touch_updated_at();
1090 cx.emit(ThreadEvent::MessageAdded(id));
1091 id
1092 }
1093
1094 pub fn edit_message(
1095 &mut self,
1096 id: MessageId,
1097 new_role: Role,
1098 new_segments: Vec<MessageSegment>,
1099 creases: Vec<MessageCrease>,
1100 loaded_context: Option<LoadedContext>,
1101 checkpoint: Option<GitStoreCheckpoint>,
1102 cx: &mut Context<Self>,
1103 ) -> bool {
1104 let Some(message) = self.messages.iter_mut().find(|message| message.id == id) else {
1105 return false;
1106 };
1107 message.role = new_role;
1108 message.segments = new_segments;
1109 message.creases = creases;
1110 if let Some(context) = loaded_context {
1111 message.loaded_context = context;
1112 }
1113 if let Some(git_checkpoint) = checkpoint {
1114 self.checkpoints_by_message.insert(
1115 id,
1116 ThreadCheckpoint {
1117 message_id: id,
1118 git_checkpoint,
1119 },
1120 );
1121 }
1122 self.touch_updated_at();
1123 cx.emit(ThreadEvent::MessageEdited(id));
1124 true
1125 }
1126
1127 pub fn delete_message(&mut self, id: MessageId, cx: &mut Context<Self>) -> bool {
1128 let Some(index) = self.messages.iter().position(|message| message.id == id) else {
1129 return false;
1130 };
1131 self.messages.remove(index);
1132 self.touch_updated_at();
1133 cx.emit(ThreadEvent::MessageDeleted(id));
1134 true
1135 }
1136
1137 /// Returns the representation of this [`Thread`] in a textual form.
1138 ///
1139 /// This is the representation we use when attaching a thread as context to another thread.
1140 pub fn text(&self) -> String {
1141 let mut text = String::new();
1142
1143 for message in &self.messages {
1144 text.push_str(match message.role {
1145 language_model::Role::User => "User:",
1146 language_model::Role::Assistant => "Agent:",
1147 language_model::Role::System => "System:",
1148 });
1149 text.push('\n');
1150
1151 for segment in &message.segments {
1152 match segment {
1153 MessageSegment::Text(content) => text.push_str(content),
1154 MessageSegment::Thinking { text: content, .. } => {
1155 text.push_str(&format!("<think>{}</think>", content))
1156 }
1157 MessageSegment::RedactedThinking(_) => {}
1158 }
1159 }
1160 text.push('\n');
1161 }
1162
1163 text
1164 }
1165
1166 /// Serializes this thread into a format for storage or telemetry.
1167 pub fn serialize(&self, cx: &mut Context<Self>) -> Task<Result<SerializedThread>> {
1168 let initial_project_snapshot = self.initial_project_snapshot.clone();
1169 cx.spawn(async move |this, cx| {
1170 let initial_project_snapshot = initial_project_snapshot.await;
1171 this.read_with(cx, |this, cx| SerializedThread {
1172 version: SerializedThread::VERSION.to_string(),
1173 summary: this.summary().or_default(),
1174 updated_at: this.updated_at(),
1175 messages: this
1176 .messages()
1177 .filter(|message| !message.ui_only)
1178 .map(|message| SerializedMessage {
1179 id: message.id,
1180 role: message.role,
1181 segments: message
1182 .segments
1183 .iter()
1184 .map(|segment| match segment {
1185 MessageSegment::Text(text) => {
1186 SerializedMessageSegment::Text { text: text.clone() }
1187 }
1188 MessageSegment::Thinking { text, signature } => {
1189 SerializedMessageSegment::Thinking {
1190 text: text.clone(),
1191 signature: signature.clone(),
1192 }
1193 }
1194 MessageSegment::RedactedThinking(data) => {
1195 SerializedMessageSegment::RedactedThinking {
1196 data: data.clone(),
1197 }
1198 }
1199 })
1200 .collect(),
1201 tool_uses: this
1202 .tool_uses_for_message(message.id, cx)
1203 .into_iter()
1204 .map(|tool_use| SerializedToolUse {
1205 id: tool_use.id,
1206 name: tool_use.name,
1207 input: tool_use.input,
1208 })
1209 .collect(),
1210 tool_results: this
1211 .tool_results_for_message(message.id)
1212 .into_iter()
1213 .map(|tool_result| SerializedToolResult {
1214 tool_use_id: tool_result.tool_use_id.clone(),
1215 is_error: tool_result.is_error,
1216 content: tool_result.content.clone(),
1217 output: tool_result.output.clone(),
1218 })
1219 .collect(),
1220 context: message.loaded_context.text.clone(),
1221 creases: message
1222 .creases
1223 .iter()
1224 .map(|crease| SerializedCrease {
1225 start: crease.range.start,
1226 end: crease.range.end,
1227 icon_path: crease.icon_path.clone(),
1228 label: crease.label.clone(),
1229 })
1230 .collect(),
1231 is_hidden: message.is_hidden,
1232 })
1233 .collect(),
1234 initial_project_snapshot,
1235 cumulative_token_usage: this.cumulative_token_usage,
1236 request_token_usage: this.request_token_usage.clone(),
1237 detailed_summary_state: this.detailed_summary_rx.borrow().clone(),
1238 exceeded_window_error: this.exceeded_window_error.clone(),
1239 model: this
1240 .configured_model
1241 .as_ref()
1242 .map(|model| SerializedLanguageModel {
1243 provider: model.provider.id().0.to_string(),
1244 model: model.model.id().0.to_string(),
1245 }),
1246 completion_mode: Some(this.completion_mode),
1247 tool_use_limit_reached: this.tool_use_limit_reached,
1248 profile: Some(this.profile.id().clone()),
1249 })
1250 })
1251 }
1252
1253 pub fn remaining_turns(&self) -> u32 {
1254 self.remaining_turns
1255 }
1256
1257 pub fn set_remaining_turns(&mut self, remaining_turns: u32) {
1258 self.remaining_turns = remaining_turns;
1259 }
1260
1261 pub fn send_to_model(
1262 &mut self,
1263 model: Arc<dyn LanguageModel>,
1264 intent: CompletionIntent,
1265 window: Option<AnyWindowHandle>,
1266 cx: &mut Context<Self>,
1267 ) {
1268 if self.remaining_turns == 0 {
1269 return;
1270 }
1271
1272 self.remaining_turns -= 1;
1273
1274 self.flush_notifications(model.clone(), intent, cx);
1275
1276 let _checkpoint = self.finalize_pending_checkpoint(cx);
1277 self.stream_completion(
1278 self.to_completion_request(model.clone(), intent, cx),
1279 model,
1280 intent,
1281 window,
1282 cx,
1283 );
1284 }
1285
1286 pub fn retry_last_completion(
1287 &mut self,
1288 window: Option<AnyWindowHandle>,
1289 cx: &mut Context<Self>,
1290 ) {
1291 // Clear any existing error state
1292 self.retry_state = None;
1293
1294 // Use the last error context if available, otherwise fall back to configured model
1295 let (model, intent) = if let Some((model, intent)) = self.last_error_context.take() {
1296 (model, intent)
1297 } else if let Some(configured_model) = self.configured_model.as_ref() {
1298 let model = configured_model.model.clone();
1299 let intent = if self.has_pending_tool_uses() {
1300 CompletionIntent::ToolResults
1301 } else {
1302 CompletionIntent::UserPrompt
1303 };
1304 (model, intent)
1305 } else if let Some(configured_model) = self.get_or_init_configured_model(cx) {
1306 let model = configured_model.model.clone();
1307 let intent = if self.has_pending_tool_uses() {
1308 CompletionIntent::ToolResults
1309 } else {
1310 CompletionIntent::UserPrompt
1311 };
1312 (model, intent)
1313 } else {
1314 return;
1315 };
1316
1317 self.send_to_model(model, intent, window, cx);
1318 }
1319
1320 pub fn enable_burn_mode_and_retry(
1321 &mut self,
1322 window: Option<AnyWindowHandle>,
1323 cx: &mut Context<Self>,
1324 ) {
1325 self.completion_mode = CompletionMode::Burn;
1326 cx.emit(ThreadEvent::ProfileChanged);
1327 self.retry_last_completion(window, cx);
1328 }
1329
1330 pub fn used_tools_since_last_user_message(&self) -> bool {
1331 for message in self.messages.iter().rev() {
1332 if self.tool_use.message_has_tool_results(message.id) {
1333 return true;
1334 } else if message.role == Role::User {
1335 return false;
1336 }
1337 }
1338
1339 false
1340 }
1341
1342 pub fn to_completion_request(
1343 &self,
1344 model: Arc<dyn LanguageModel>,
1345 intent: CompletionIntent,
1346 cx: &mut Context<Self>,
1347 ) -> LanguageModelRequest {
1348 let mut request = LanguageModelRequest {
1349 thread_id: Some(self.id.to_string()),
1350 prompt_id: Some(self.last_prompt_id.to_string()),
1351 intent: Some(intent),
1352 mode: None,
1353 messages: vec![],
1354 tools: Vec::new(),
1355 tool_choice: None,
1356 stop: Vec::new(),
1357 temperature: AgentSettings::temperature_for_model(&model, cx),
1358 thinking_allowed: true,
1359 };
1360
1361 let available_tools = self.available_tools(cx, model.clone());
1362 let available_tool_names = available_tools
1363 .iter()
1364 .map(|tool| tool.name.clone())
1365 .collect();
1366
1367 let model_context = &ModelContext {
1368 available_tools: available_tool_names,
1369 };
1370
1371 if let Some(project_context) = self.project_context.borrow().as_ref() {
1372 match self
1373 .prompt_builder
1374 .generate_assistant_system_prompt(project_context, model_context)
1375 {
1376 Err(err) => {
1377 let message = format!("{err:?}").into();
1378 log::error!("{message}");
1379 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1380 header: "Error generating system prompt".into(),
1381 message,
1382 }));
1383 }
1384 Ok(system_prompt) => {
1385 request.messages.push(LanguageModelRequestMessage {
1386 role: Role::System,
1387 content: vec![MessageContent::Text(system_prompt)],
1388 cache: true,
1389 });
1390 }
1391 }
1392 } else {
1393 let message = "Context for system prompt unexpectedly not ready.".into();
1394 log::error!("{message}");
1395 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1396 header: "Error generating system prompt".into(),
1397 message,
1398 }));
1399 }
1400
1401 let mut message_ix_to_cache = None;
1402 for message in &self.messages {
1403 // ui_only messages are for the UI only, not for the model
1404 if message.ui_only {
1405 continue;
1406 }
1407
1408 let mut request_message = LanguageModelRequestMessage {
1409 role: message.role,
1410 content: Vec::new(),
1411 cache: false,
1412 };
1413
1414 message
1415 .loaded_context
1416 .add_to_request_message(&mut request_message);
1417
1418 for segment in &message.segments {
1419 match segment {
1420 MessageSegment::Text(text) => {
1421 let text = text.trim_end();
1422 if !text.is_empty() {
1423 request_message
1424 .content
1425 .push(MessageContent::Text(text.into()));
1426 }
1427 }
1428 MessageSegment::Thinking { text, signature } => {
1429 if !text.is_empty() {
1430 request_message.content.push(MessageContent::Thinking {
1431 text: text.into(),
1432 signature: signature.clone(),
1433 });
1434 }
1435 }
1436 MessageSegment::RedactedThinking(data) => {
1437 request_message
1438 .content
1439 .push(MessageContent::RedactedThinking(data.clone()));
1440 }
1441 };
1442 }
1443
1444 let mut cache_message = true;
1445 let mut tool_results_message = LanguageModelRequestMessage {
1446 role: Role::User,
1447 content: Vec::new(),
1448 cache: false,
1449 };
1450 for (tool_use, tool_result) in self.tool_use.tool_results(message.id) {
1451 if let Some(tool_result) = tool_result {
1452 request_message
1453 .content
1454 .push(MessageContent::ToolUse(tool_use.clone()));
1455 tool_results_message
1456 .content
1457 .push(MessageContent::ToolResult(LanguageModelToolResult {
1458 tool_use_id: tool_use.id.clone(),
1459 tool_name: tool_result.tool_name.clone(),
1460 is_error: tool_result.is_error,
1461 content: if tool_result.content.is_empty() {
1462 // Surprisingly, the API fails if we return an empty string here.
1463 // It thinks we are sending a tool use without a tool result.
1464 "<Tool returned an empty string>".into()
1465 } else {
1466 tool_result.content.clone()
1467 },
1468 output: None,
1469 }));
1470 } else {
1471 cache_message = false;
1472 log::debug!(
1473 "skipped tool use {:?} because it is still pending",
1474 tool_use
1475 );
1476 }
1477 }
1478
1479 if cache_message {
1480 message_ix_to_cache = Some(request.messages.len());
1481 }
1482 request.messages.push(request_message);
1483
1484 if !tool_results_message.content.is_empty() {
1485 if cache_message {
1486 message_ix_to_cache = Some(request.messages.len());
1487 }
1488 request.messages.push(tool_results_message);
1489 }
1490 }
1491
1492 // https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
1493 if let Some(message_ix_to_cache) = message_ix_to_cache {
1494 request.messages[message_ix_to_cache].cache = true;
1495 }
1496
1497 request.tools = available_tools;
1498 request.mode = if model.supports_burn_mode() {
1499 Some(self.completion_mode.into())
1500 } else {
1501 Some(CompletionMode::Normal.into())
1502 };
1503
1504 request
1505 }
1506
1507 fn to_summarize_request(
1508 &self,
1509 model: &Arc<dyn LanguageModel>,
1510 intent: CompletionIntent,
1511 added_user_message: String,
1512 cx: &App,
1513 ) -> LanguageModelRequest {
1514 let mut request = LanguageModelRequest {
1515 thread_id: None,
1516 prompt_id: None,
1517 intent: Some(intent),
1518 mode: None,
1519 messages: vec![],
1520 tools: Vec::new(),
1521 tool_choice: None,
1522 stop: Vec::new(),
1523 temperature: AgentSettings::temperature_for_model(model, cx),
1524 thinking_allowed: false,
1525 };
1526
1527 for message in &self.messages {
1528 let mut request_message = LanguageModelRequestMessage {
1529 role: message.role,
1530 content: Vec::new(),
1531 cache: false,
1532 };
1533
1534 for segment in &message.segments {
1535 match segment {
1536 MessageSegment::Text(text) => request_message
1537 .content
1538 .push(MessageContent::Text(text.clone())),
1539 MessageSegment::Thinking { .. } => {}
1540 MessageSegment::RedactedThinking(_) => {}
1541 }
1542 }
1543
1544 if request_message.content.is_empty() {
1545 continue;
1546 }
1547
1548 request.messages.push(request_message);
1549 }
1550
1551 request.messages.push(LanguageModelRequestMessage {
1552 role: Role::User,
1553 content: vec![MessageContent::Text(added_user_message)],
1554 cache: false,
1555 });
1556
1557 request
1558 }
1559
1560 /// Insert auto-generated notifications (if any) to the thread
1561 fn flush_notifications(
1562 &mut self,
1563 model: Arc<dyn LanguageModel>,
1564 intent: CompletionIntent,
1565 cx: &mut Context<Self>,
1566 ) {
1567 match intent {
1568 CompletionIntent::UserPrompt | CompletionIntent::ToolResults => {
1569 if let Some(pending_tool_use) = self.attach_tracked_files_state(model, cx) {
1570 cx.emit(ThreadEvent::ToolFinished {
1571 tool_use_id: pending_tool_use.id.clone(),
1572 pending_tool_use: Some(pending_tool_use),
1573 });
1574 }
1575 }
1576 CompletionIntent::ThreadSummarization
1577 | CompletionIntent::ThreadContextSummarization
1578 | CompletionIntent::CreateFile
1579 | CompletionIntent::EditFile
1580 | CompletionIntent::InlineAssist
1581 | CompletionIntent::TerminalInlineAssist
1582 | CompletionIntent::GenerateGitCommitMessage => {}
1583 };
1584 }
1585
1586 fn attach_tracked_files_state(
1587 &mut self,
1588 model: Arc<dyn LanguageModel>,
1589 cx: &mut App,
1590 ) -> Option<PendingToolUse> {
1591 // Represent notification as a simulated `project_notifications` tool call
1592 let tool_name = Arc::from("project_notifications");
1593 let tool = self.tools.read(cx).tool(&tool_name, cx)?;
1594
1595 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
1596 return None;
1597 }
1598
1599 if self
1600 .action_log
1601 .update(cx, |log, cx| log.unnotified_user_edits(cx).is_none())
1602 {
1603 return None;
1604 }
1605
1606 let input = serde_json::json!({});
1607 let request = Arc::new(LanguageModelRequest::default()); // unused
1608 let window = None;
1609 let tool_result = tool.run(
1610 input,
1611 request,
1612 self.project.clone(),
1613 self.action_log.clone(),
1614 model.clone(),
1615 window,
1616 cx,
1617 );
1618
1619 let tool_use_id =
1620 LanguageModelToolUseId::from(format!("project_notifications_{}", self.messages.len()));
1621
1622 let tool_use = LanguageModelToolUse {
1623 id: tool_use_id.clone(),
1624 name: tool_name.clone(),
1625 raw_input: "{}".to_string(),
1626 input: serde_json::json!({}),
1627 is_input_complete: true,
1628 };
1629
1630 let tool_output = cx.background_executor().block(tool_result.output);
1631
1632 // Attach a project_notification tool call to the latest existing
1633 // Assistant message. We cannot create a new Assistant message
1634 // because thinking models require a `thinking` block that we
1635 // cannot mock. We cannot send a notification as a normal
1636 // (non-tool-use) User message because this distracts Agent
1637 // too much.
1638 let tool_message_id = self
1639 .messages
1640 .iter()
1641 .enumerate()
1642 .rfind(|(_, message)| message.role == Role::Assistant)
1643 .map(|(_, message)| message.id)?;
1644
1645 let tool_use_metadata = ToolUseMetadata {
1646 model: model.clone(),
1647 thread_id: self.id.clone(),
1648 prompt_id: self.last_prompt_id.clone(),
1649 };
1650
1651 self.tool_use
1652 .request_tool_use(tool_message_id, tool_use, tool_use_metadata.clone(), cx);
1653
1654 let pending_tool_use = self.tool_use.insert_tool_output(
1655 tool_use_id.clone(),
1656 tool_name,
1657 tool_output,
1658 self.configured_model.as_ref(),
1659 self.completion_mode,
1660 );
1661
1662 pending_tool_use
1663 }
1664
1665 pub fn stream_completion(
1666 &mut self,
1667 request: LanguageModelRequest,
1668 model: Arc<dyn LanguageModel>,
1669 intent: CompletionIntent,
1670 window: Option<AnyWindowHandle>,
1671 cx: &mut Context<Self>,
1672 ) {
1673 self.tool_use_limit_reached = false;
1674
1675 let pending_completion_id = post_inc(&mut self.completion_count);
1676 let mut request_callback_parameters = if self.request_callback.is_some() {
1677 Some((request.clone(), Vec::new()))
1678 } else {
1679 None
1680 };
1681 let prompt_id = self.last_prompt_id.clone();
1682 let tool_use_metadata = ToolUseMetadata {
1683 model: model.clone(),
1684 thread_id: self.id.clone(),
1685 prompt_id: prompt_id.clone(),
1686 };
1687
1688 let completion_mode = request
1689 .mode
1690 .unwrap_or(cloud_llm_client::CompletionMode::Normal);
1691
1692 self.last_received_chunk_at = Some(Instant::now());
1693
1694 let task = cx.spawn(async move |thread, cx| {
1695 let stream_completion_future = model.stream_completion(request, &cx);
1696 let initial_token_usage =
1697 thread.read_with(cx, |thread, _cx| thread.cumulative_token_usage);
1698 let stream_completion = async {
1699 let mut events = stream_completion_future.await?;
1700
1701 let mut stop_reason = StopReason::EndTurn;
1702 let mut current_token_usage = TokenUsage::default();
1703
1704 thread
1705 .update(cx, |_thread, cx| {
1706 cx.emit(ThreadEvent::NewRequest);
1707 })
1708 .ok();
1709
1710 let mut request_assistant_message_id = None;
1711
1712 while let Some(event) = events.next().await {
1713 if let Some((_, response_events)) = request_callback_parameters.as_mut() {
1714 response_events
1715 .push(event.as_ref().map_err(|error| error.to_string()).cloned());
1716 }
1717
1718 thread.update(cx, |thread, cx| {
1719 match event? {
1720 LanguageModelCompletionEvent::StartMessage { .. } => {
1721 request_assistant_message_id =
1722 Some(thread.insert_assistant_message(
1723 vec![MessageSegment::Text(String::new())],
1724 cx,
1725 ));
1726 }
1727 LanguageModelCompletionEvent::Stop(reason) => {
1728 stop_reason = reason;
1729 }
1730 LanguageModelCompletionEvent::UsageUpdate(token_usage) => {
1731 thread.update_token_usage_at_last_message(token_usage);
1732 thread.cumulative_token_usage = thread.cumulative_token_usage
1733 + token_usage
1734 - current_token_usage;
1735 current_token_usage = token_usage;
1736 }
1737 LanguageModelCompletionEvent::Text(chunk) => {
1738 thread.received_chunk();
1739
1740 cx.emit(ThreadEvent::ReceivedTextChunk);
1741 if let Some(last_message) = thread.messages.last_mut() {
1742 if last_message.role == Role::Assistant
1743 && !thread.tool_use.has_tool_results(last_message.id)
1744 {
1745 last_message.push_text(&chunk);
1746 cx.emit(ThreadEvent::StreamedAssistantText(
1747 last_message.id,
1748 chunk,
1749 ));
1750 } else {
1751 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1752 // of a new Assistant response.
1753 //
1754 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1755 // will result in duplicating the text of the chunk in the rendered Markdown.
1756 request_assistant_message_id =
1757 Some(thread.insert_assistant_message(
1758 vec![MessageSegment::Text(chunk.to_string())],
1759 cx,
1760 ));
1761 };
1762 }
1763 }
1764 LanguageModelCompletionEvent::Thinking {
1765 text: chunk,
1766 signature,
1767 } => {
1768 thread.received_chunk();
1769
1770 if let Some(last_message) = thread.messages.last_mut() {
1771 if last_message.role == Role::Assistant
1772 && !thread.tool_use.has_tool_results(last_message.id)
1773 {
1774 last_message.push_thinking(&chunk, signature);
1775 cx.emit(ThreadEvent::StreamedAssistantThinking(
1776 last_message.id,
1777 chunk,
1778 ));
1779 } else {
1780 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1781 // of a new Assistant response.
1782 //
1783 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1784 // will result in duplicating the text of the chunk in the rendered Markdown.
1785 request_assistant_message_id =
1786 Some(thread.insert_assistant_message(
1787 vec![MessageSegment::Thinking {
1788 text: chunk.to_string(),
1789 signature,
1790 }],
1791 cx,
1792 ));
1793 };
1794 }
1795 }
1796 LanguageModelCompletionEvent::RedactedThinking { data } => {
1797 thread.received_chunk();
1798
1799 if let Some(last_message) = thread.messages.last_mut() {
1800 if last_message.role == Role::Assistant
1801 && !thread.tool_use.has_tool_results(last_message.id)
1802 {
1803 last_message.push_redacted_thinking(data);
1804 } else {
1805 request_assistant_message_id =
1806 Some(thread.insert_assistant_message(
1807 vec![MessageSegment::RedactedThinking(data)],
1808 cx,
1809 ));
1810 };
1811 }
1812 }
1813 LanguageModelCompletionEvent::ToolUse(tool_use) => {
1814 let last_assistant_message_id = request_assistant_message_id
1815 .unwrap_or_else(|| {
1816 let new_assistant_message_id =
1817 thread.insert_assistant_message(vec![], cx);
1818 request_assistant_message_id =
1819 Some(new_assistant_message_id);
1820 new_assistant_message_id
1821 });
1822
1823 let tool_use_id = tool_use.id.clone();
1824 let streamed_input = if tool_use.is_input_complete {
1825 None
1826 } else {
1827 Some((&tool_use.input).clone())
1828 };
1829
1830 let ui_text = thread.tool_use.request_tool_use(
1831 last_assistant_message_id,
1832 tool_use,
1833 tool_use_metadata.clone(),
1834 cx,
1835 );
1836
1837 if let Some(input) = streamed_input {
1838 cx.emit(ThreadEvent::StreamedToolUse {
1839 tool_use_id,
1840 ui_text,
1841 input,
1842 });
1843 }
1844 }
1845 LanguageModelCompletionEvent::ToolUseJsonParseError {
1846 id,
1847 tool_name,
1848 raw_input: invalid_input_json,
1849 json_parse_error,
1850 } => {
1851 thread.receive_invalid_tool_json(
1852 id,
1853 tool_name,
1854 invalid_input_json,
1855 json_parse_error,
1856 window,
1857 cx,
1858 );
1859 }
1860 LanguageModelCompletionEvent::StatusUpdate(status_update) => {
1861 if let Some(completion) = thread
1862 .pending_completions
1863 .iter_mut()
1864 .find(|completion| completion.id == pending_completion_id)
1865 {
1866 match status_update {
1867 CompletionRequestStatus::Queued { position } => {
1868 completion.queue_state =
1869 QueueState::Queued { position };
1870 }
1871 CompletionRequestStatus::Started => {
1872 completion.queue_state = QueueState::Started;
1873 }
1874 CompletionRequestStatus::Failed {
1875 code,
1876 message,
1877 request_id: _,
1878 retry_after,
1879 } => {
1880 return Err(
1881 LanguageModelCompletionError::from_cloud_failure(
1882 model.upstream_provider_name(),
1883 code,
1884 message,
1885 retry_after.map(Duration::from_secs_f64),
1886 ),
1887 );
1888 }
1889 CompletionRequestStatus::UsageUpdated { amount, limit } => {
1890 thread.update_model_request_usage(
1891 amount as u32,
1892 limit,
1893 cx,
1894 );
1895 }
1896 CompletionRequestStatus::ToolUseLimitReached => {
1897 thread.tool_use_limit_reached = true;
1898 cx.emit(ThreadEvent::ToolUseLimitReached);
1899 }
1900 }
1901 }
1902 }
1903 }
1904
1905 thread.touch_updated_at();
1906 cx.emit(ThreadEvent::StreamedCompletion);
1907 cx.notify();
1908
1909 thread.auto_capture_telemetry(cx);
1910 Ok(())
1911 })??;
1912
1913 smol::future::yield_now().await;
1914 }
1915
1916 thread.update(cx, |thread, cx| {
1917 thread.last_received_chunk_at = None;
1918 thread
1919 .pending_completions
1920 .retain(|completion| completion.id != pending_completion_id);
1921
1922 // If there is a response without tool use, summarize the message. Otherwise,
1923 // allow two tool uses before summarizing.
1924 if matches!(thread.summary, ThreadSummary::Pending)
1925 && thread.messages.len() >= 2
1926 && (!thread.has_pending_tool_uses() || thread.messages.len() >= 6)
1927 {
1928 thread.summarize(cx);
1929 }
1930 })?;
1931
1932 anyhow::Ok(stop_reason)
1933 };
1934
1935 let result = stream_completion.await;
1936 let mut retry_scheduled = false;
1937
1938 thread
1939 .update(cx, |thread, cx| {
1940 thread.finalize_pending_checkpoint(cx);
1941 match result.as_ref() {
1942 Ok(stop_reason) => {
1943 match stop_reason {
1944 StopReason::ToolUse => {
1945 let tool_uses =
1946 thread.use_pending_tools(window, model.clone(), cx);
1947 cx.emit(ThreadEvent::UsePendingTools { tool_uses });
1948 }
1949 StopReason::EndTurn | StopReason::MaxTokens => {
1950 thread.project.update(cx, |project, cx| {
1951 project.set_agent_location(None, cx);
1952 });
1953 }
1954 StopReason::Refusal => {
1955 thread.project.update(cx, |project, cx| {
1956 project.set_agent_location(None, cx);
1957 });
1958
1959 // Remove the turn that was refused.
1960 //
1961 // https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/handle-streaming-refusals#reset-context-after-refusal
1962 {
1963 let mut messages_to_remove = Vec::new();
1964
1965 for (ix, message) in
1966 thread.messages.iter().enumerate().rev()
1967 {
1968 messages_to_remove.push(message.id);
1969
1970 if message.role == Role::User {
1971 if ix == 0 {
1972 break;
1973 }
1974
1975 if let Some(prev_message) =
1976 thread.messages.get(ix - 1)
1977 {
1978 if prev_message.role == Role::Assistant {
1979 break;
1980 }
1981 }
1982 }
1983 }
1984
1985 for message_id in messages_to_remove {
1986 thread.delete_message(message_id, cx);
1987 }
1988 }
1989
1990 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1991 header: "Language model refusal".into(),
1992 message:
1993 "Model refused to generate content for safety reasons."
1994 .into(),
1995 }));
1996 }
1997 }
1998
1999 // We successfully completed, so cancel any remaining retries.
2000 thread.retry_state = None;
2001 }
2002 Err(error) => {
2003 thread.project.update(cx, |project, cx| {
2004 project.set_agent_location(None, cx);
2005 });
2006
2007 if error.is::<PaymentRequiredError>() {
2008 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
2009 } else if let Some(error) =
2010 error.downcast_ref::<ModelRequestLimitReachedError>()
2011 {
2012 cx.emit(ThreadEvent::ShowError(
2013 ThreadError::ModelRequestLimitReached { plan: error.plan },
2014 ));
2015 } else if let Some(completion_error) =
2016 error.downcast_ref::<LanguageModelCompletionError>()
2017 {
2018 match &completion_error {
2019 LanguageModelCompletionError::PromptTooLarge {
2020 tokens, ..
2021 } => {
2022 let tokens = tokens.unwrap_or_else(|| {
2023 // We didn't get an exact token count from the API, so fall back on our estimate.
2024 thread
2025 .total_token_usage()
2026 .map(|usage| usage.total)
2027 .unwrap_or(0)
2028 // We know the context window was exceeded in practice, so if our estimate was
2029 // lower than max tokens, the estimate was wrong; return that we exceeded by 1.
2030 .max(
2031 model
2032 .max_token_count_for_mode(completion_mode)
2033 .saturating_add(1),
2034 )
2035 });
2036 thread.exceeded_window_error = Some(ExceededWindowError {
2037 model_id: model.id(),
2038 token_count: tokens,
2039 });
2040 cx.notify();
2041 }
2042 _ => {
2043 if let Some(retry_strategy) =
2044 Thread::get_retry_strategy(completion_error)
2045 {
2046 log::info!(
2047 "Retrying with {:?} for language model completion error {:?}",
2048 retry_strategy,
2049 completion_error
2050 );
2051
2052 retry_scheduled = thread
2053 .handle_retryable_error_with_delay(
2054 &completion_error,
2055 Some(retry_strategy),
2056 model.clone(),
2057 intent,
2058 window,
2059 cx,
2060 );
2061 }
2062 }
2063 }
2064 }
2065
2066 if !retry_scheduled {
2067 thread.cancel_last_completion(window, cx);
2068 }
2069 }
2070 }
2071
2072 if !retry_scheduled {
2073 cx.emit(ThreadEvent::Stopped(result.map_err(Arc::new)));
2074 }
2075
2076 if let Some((request_callback, (request, response_events))) = thread
2077 .request_callback
2078 .as_mut()
2079 .zip(request_callback_parameters.as_ref())
2080 {
2081 request_callback(request, response_events);
2082 }
2083
2084 thread.auto_capture_telemetry(cx);
2085
2086 if let Ok(initial_usage) = initial_token_usage {
2087 let usage = thread.cumulative_token_usage - initial_usage;
2088
2089 telemetry::event!(
2090 "Assistant Thread Completion",
2091 thread_id = thread.id().to_string(),
2092 prompt_id = prompt_id,
2093 model = model.telemetry_id(),
2094 model_provider = model.provider_id().to_string(),
2095 input_tokens = usage.input_tokens,
2096 output_tokens = usage.output_tokens,
2097 cache_creation_input_tokens = usage.cache_creation_input_tokens,
2098 cache_read_input_tokens = usage.cache_read_input_tokens,
2099 );
2100 }
2101 })
2102 .ok();
2103 });
2104
2105 self.pending_completions.push(PendingCompletion {
2106 id: pending_completion_id,
2107 queue_state: QueueState::Sending,
2108 _task: task,
2109 });
2110 }
2111
2112 pub fn summarize(&mut self, cx: &mut Context<Self>) {
2113 let Some(model) = LanguageModelRegistry::read_global(cx).thread_summary_model() else {
2114 println!("No thread summary model");
2115 return;
2116 };
2117
2118 if !model.provider.is_authenticated(cx) {
2119 return;
2120 }
2121
2122 let request = self.to_summarize_request(
2123 &model.model,
2124 CompletionIntent::ThreadSummarization,
2125 SUMMARIZE_THREAD_PROMPT.into(),
2126 cx,
2127 );
2128
2129 self.summary = ThreadSummary::Generating;
2130
2131 self.pending_summary = cx.spawn(async move |this, cx| {
2132 let result = async {
2133 let mut messages = model.model.stream_completion(request, &cx).await?;
2134
2135 let mut new_summary = String::new();
2136 while let Some(event) = messages.next().await {
2137 let Ok(event) = event else {
2138 continue;
2139 };
2140 let text = match event {
2141 LanguageModelCompletionEvent::Text(text) => text,
2142 LanguageModelCompletionEvent::StatusUpdate(
2143 CompletionRequestStatus::UsageUpdated { amount, limit },
2144 ) => {
2145 this.update(cx, |thread, cx| {
2146 thread.update_model_request_usage(amount as u32, limit, cx);
2147 })?;
2148 continue;
2149 }
2150 _ => continue,
2151 };
2152
2153 let mut lines = text.lines();
2154 new_summary.extend(lines.next());
2155
2156 // Stop if the LLM generated multiple lines.
2157 if lines.next().is_some() {
2158 break;
2159 }
2160 }
2161
2162 anyhow::Ok(new_summary)
2163 }
2164 .await;
2165
2166 this.update(cx, |this, cx| {
2167 match result {
2168 Ok(new_summary) => {
2169 if new_summary.is_empty() {
2170 this.summary = ThreadSummary::Error;
2171 } else {
2172 this.summary = ThreadSummary::Ready(new_summary.into());
2173 }
2174 }
2175 Err(err) => {
2176 this.summary = ThreadSummary::Error;
2177 log::error!("Failed to generate thread summary: {}", err);
2178 }
2179 }
2180 cx.emit(ThreadEvent::SummaryGenerated);
2181 })
2182 .log_err()?;
2183
2184 Some(())
2185 });
2186 }
2187
2188 fn get_retry_strategy(error: &LanguageModelCompletionError) -> Option<RetryStrategy> {
2189 use LanguageModelCompletionError::*;
2190
2191 // General strategy here:
2192 // - If retrying won't help (e.g. invalid API key or payload too large), return None so we don't retry at all.
2193 // - If it's a time-based issue (e.g. server overloaded, rate limit exceeded), retry up to 4 times with exponential backoff.
2194 // - If it's an issue that *might* be fixed by retrying (e.g. internal server error), retry up to 3 times.
2195 match error {
2196 HttpResponseError {
2197 status_code: StatusCode::TOO_MANY_REQUESTS,
2198 ..
2199 } => Some(RetryStrategy::ExponentialBackoff {
2200 initial_delay: BASE_RETRY_DELAY,
2201 max_attempts: MAX_RETRY_ATTEMPTS,
2202 }),
2203 ServerOverloaded { retry_after, .. } | RateLimitExceeded { retry_after, .. } => {
2204 Some(RetryStrategy::Fixed {
2205 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2206 max_attempts: MAX_RETRY_ATTEMPTS,
2207 })
2208 }
2209 UpstreamProviderError {
2210 status,
2211 retry_after,
2212 ..
2213 } => match *status {
2214 StatusCode::TOO_MANY_REQUESTS | StatusCode::SERVICE_UNAVAILABLE => {
2215 Some(RetryStrategy::Fixed {
2216 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2217 max_attempts: MAX_RETRY_ATTEMPTS,
2218 })
2219 }
2220 StatusCode::INTERNAL_SERVER_ERROR => Some(RetryStrategy::Fixed {
2221 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2222 // Internal Server Error could be anything, retry up to 3 times.
2223 max_attempts: 3,
2224 }),
2225 status => {
2226 // There is no StatusCode variant for the unofficial HTTP 529 ("The service is overloaded"),
2227 // but we frequently get them in practice. See https://http.dev/529
2228 if status.as_u16() == 529 {
2229 Some(RetryStrategy::Fixed {
2230 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2231 max_attempts: MAX_RETRY_ATTEMPTS,
2232 })
2233 } else {
2234 Some(RetryStrategy::Fixed {
2235 delay: retry_after.unwrap_or(BASE_RETRY_DELAY),
2236 max_attempts: 2,
2237 })
2238 }
2239 }
2240 },
2241 ApiInternalServerError { .. } => Some(RetryStrategy::Fixed {
2242 delay: BASE_RETRY_DELAY,
2243 max_attempts: 3,
2244 }),
2245 ApiReadResponseError { .. }
2246 | HttpSend { .. }
2247 | DeserializeResponse { .. }
2248 | BadRequestFormat { .. } => Some(RetryStrategy::Fixed {
2249 delay: BASE_RETRY_DELAY,
2250 max_attempts: 3,
2251 }),
2252 // Retrying these errors definitely shouldn't help.
2253 HttpResponseError {
2254 status_code:
2255 StatusCode::PAYLOAD_TOO_LARGE | StatusCode::FORBIDDEN | StatusCode::UNAUTHORIZED,
2256 ..
2257 }
2258 | AuthenticationError { .. }
2259 | PermissionError { .. }
2260 | NoApiKey { .. }
2261 | ApiEndpointNotFound { .. }
2262 | PromptTooLarge { .. } => None,
2263 // These errors might be transient, so retry them
2264 SerializeRequest { .. } | BuildRequestBody { .. } => Some(RetryStrategy::Fixed {
2265 delay: BASE_RETRY_DELAY,
2266 max_attempts: 1,
2267 }),
2268 // Retry all other 4xx and 5xx errors once.
2269 HttpResponseError { status_code, .. }
2270 if status_code.is_client_error() || status_code.is_server_error() =>
2271 {
2272 Some(RetryStrategy::Fixed {
2273 delay: BASE_RETRY_DELAY,
2274 max_attempts: 3,
2275 })
2276 }
2277 // Conservatively assume that any other errors are non-retryable
2278 HttpResponseError { .. } | Other(..) => Some(RetryStrategy::Fixed {
2279 delay: BASE_RETRY_DELAY,
2280 max_attempts: 2,
2281 }),
2282 }
2283 }
2284
2285 fn handle_retryable_error_with_delay(
2286 &mut self,
2287 error: &LanguageModelCompletionError,
2288 strategy: Option<RetryStrategy>,
2289 model: Arc<dyn LanguageModel>,
2290 intent: CompletionIntent,
2291 window: Option<AnyWindowHandle>,
2292 cx: &mut Context<Self>,
2293 ) -> bool {
2294 // Store context for the Retry button
2295 self.last_error_context = Some((model.clone(), intent));
2296
2297 // Only auto-retry if Burn Mode is enabled
2298 if self.completion_mode != CompletionMode::Burn {
2299 // Show error with retry options
2300 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2301 message: format!(
2302 "{}\n\nTo automatically retry when similar errors happen, enable Burn Mode.",
2303 error
2304 )
2305 .into(),
2306 can_enable_burn_mode: true,
2307 }));
2308 return false;
2309 }
2310
2311 let Some(strategy) = strategy.or_else(|| Self::get_retry_strategy(error)) else {
2312 return false;
2313 };
2314
2315 let max_attempts = match &strategy {
2316 RetryStrategy::ExponentialBackoff { max_attempts, .. } => *max_attempts,
2317 RetryStrategy::Fixed { max_attempts, .. } => *max_attempts,
2318 };
2319
2320 let retry_state = self.retry_state.get_or_insert(RetryState {
2321 attempt: 0,
2322 max_attempts,
2323 intent,
2324 });
2325
2326 retry_state.attempt += 1;
2327 let attempt = retry_state.attempt;
2328 let max_attempts = retry_state.max_attempts;
2329 let intent = retry_state.intent;
2330
2331 if attempt <= max_attempts {
2332 let delay = match &strategy {
2333 RetryStrategy::ExponentialBackoff { initial_delay, .. } => {
2334 let delay_secs = initial_delay.as_secs() * 2u64.pow((attempt - 1) as u32);
2335 Duration::from_secs(delay_secs)
2336 }
2337 RetryStrategy::Fixed { delay, .. } => *delay,
2338 };
2339
2340 // Add a transient message to inform the user
2341 let delay_secs = delay.as_secs();
2342 let retry_message = if max_attempts == 1 {
2343 format!("{error}. Retrying in {delay_secs} seconds...")
2344 } else {
2345 format!(
2346 "{error}. Retrying (attempt {attempt} of {max_attempts}) \
2347 in {delay_secs} seconds..."
2348 )
2349 };
2350 log::warn!(
2351 "Retrying completion request (attempt {attempt} of {max_attempts}) \
2352 in {delay_secs} seconds: {error:?}",
2353 );
2354
2355 // Add a UI-only message instead of a regular message
2356 let id = self.next_message_id.post_inc();
2357 self.messages.push(Message {
2358 id,
2359 role: Role::System,
2360 segments: vec![MessageSegment::Text(retry_message)],
2361 loaded_context: LoadedContext::default(),
2362 creases: Vec::new(),
2363 is_hidden: false,
2364 ui_only: true,
2365 });
2366 cx.emit(ThreadEvent::MessageAdded(id));
2367
2368 // Schedule the retry
2369 let thread_handle = cx.entity().downgrade();
2370
2371 cx.spawn(async move |_thread, cx| {
2372 cx.background_executor().timer(delay).await;
2373
2374 thread_handle
2375 .update(cx, |thread, cx| {
2376 // Retry the completion
2377 thread.send_to_model(model, intent, window, cx);
2378 })
2379 .log_err();
2380 })
2381 .detach();
2382
2383 true
2384 } else {
2385 // Max retries exceeded
2386 self.retry_state = None;
2387
2388 // Stop generating since we're giving up on retrying.
2389 self.pending_completions.clear();
2390
2391 // Show error alongside a Retry button, but no
2392 // Enable Burn Mode button (since it's already enabled)
2393 cx.emit(ThreadEvent::ShowError(ThreadError::RetryableError {
2394 message: format!("Failed after retrying: {}", error).into(),
2395 can_enable_burn_mode: false,
2396 }));
2397
2398 false
2399 }
2400 }
2401
2402 pub fn start_generating_detailed_summary_if_needed(
2403 &mut self,
2404 thread_store: WeakEntity<ThreadStore>,
2405 cx: &mut Context<Self>,
2406 ) {
2407 let Some(last_message_id) = self.messages.last().map(|message| message.id) else {
2408 return;
2409 };
2410
2411 match &*self.detailed_summary_rx.borrow() {
2412 DetailedSummaryState::Generating { message_id, .. }
2413 | DetailedSummaryState::Generated { message_id, .. }
2414 if *message_id == last_message_id =>
2415 {
2416 // Already up-to-date
2417 return;
2418 }
2419 _ => {}
2420 }
2421
2422 let Some(ConfiguredModel { model, provider }) =
2423 LanguageModelRegistry::read_global(cx).thread_summary_model()
2424 else {
2425 return;
2426 };
2427
2428 if !provider.is_authenticated(cx) {
2429 return;
2430 }
2431
2432 let added_user_message = include_str!("./prompts/summarize_thread_detailed_prompt.txt");
2433
2434 let request = self.to_summarize_request(
2435 &model,
2436 CompletionIntent::ThreadContextSummarization,
2437 added_user_message.into(),
2438 cx,
2439 );
2440
2441 *self.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generating {
2442 message_id: last_message_id,
2443 };
2444
2445 // Replace the detailed summarization task if there is one, cancelling it. It would probably
2446 // be better to allow the old task to complete, but this would require logic for choosing
2447 // which result to prefer (the old task could complete after the new one, resulting in a
2448 // stale summary).
2449 self.detailed_summary_task = cx.spawn(async move |thread, cx| {
2450 let stream = model.stream_completion_text(request, &cx);
2451 let Some(mut messages) = stream.await.log_err() else {
2452 thread
2453 .update(cx, |thread, _cx| {
2454 *thread.detailed_summary_tx.borrow_mut() =
2455 DetailedSummaryState::NotGenerated;
2456 })
2457 .ok()?;
2458 return None;
2459 };
2460
2461 let mut new_detailed_summary = String::new();
2462
2463 while let Some(chunk) = messages.stream.next().await {
2464 if let Some(chunk) = chunk.log_err() {
2465 new_detailed_summary.push_str(&chunk);
2466 }
2467 }
2468
2469 thread
2470 .update(cx, |thread, _cx| {
2471 *thread.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generated {
2472 text: new_detailed_summary.into(),
2473 message_id: last_message_id,
2474 };
2475 })
2476 .ok()?;
2477
2478 // Save thread so its summary can be reused later
2479 if let Some(thread) = thread.upgrade() {
2480 if let Ok(Ok(save_task)) = cx.update(|cx| {
2481 thread_store
2482 .update(cx, |thread_store, cx| thread_store.save_thread(&thread, cx))
2483 }) {
2484 save_task.await.log_err();
2485 }
2486 }
2487
2488 Some(())
2489 });
2490 }
2491
2492 pub async fn wait_for_detailed_summary_or_text(
2493 this: &Entity<Self>,
2494 cx: &mut AsyncApp,
2495 ) -> Option<SharedString> {
2496 let mut detailed_summary_rx = this
2497 .read_with(cx, |this, _cx| this.detailed_summary_rx.clone())
2498 .ok()?;
2499 loop {
2500 match detailed_summary_rx.recv().await? {
2501 DetailedSummaryState::Generating { .. } => {}
2502 DetailedSummaryState::NotGenerated => {
2503 return this.read_with(cx, |this, _cx| this.text().into()).ok();
2504 }
2505 DetailedSummaryState::Generated { text, .. } => return Some(text),
2506 }
2507 }
2508 }
2509
2510 pub fn latest_detailed_summary_or_text(&self) -> SharedString {
2511 self.detailed_summary_rx
2512 .borrow()
2513 .text()
2514 .unwrap_or_else(|| self.text().into())
2515 }
2516
2517 pub fn is_generating_detailed_summary(&self) -> bool {
2518 matches!(
2519 &*self.detailed_summary_rx.borrow(),
2520 DetailedSummaryState::Generating { .. }
2521 )
2522 }
2523
2524 pub fn use_pending_tools(
2525 &mut self,
2526 window: Option<AnyWindowHandle>,
2527 model: Arc<dyn LanguageModel>,
2528 cx: &mut Context<Self>,
2529 ) -> Vec<PendingToolUse> {
2530 self.auto_capture_telemetry(cx);
2531 let request =
2532 Arc::new(self.to_completion_request(model.clone(), CompletionIntent::ToolResults, cx));
2533 let pending_tool_uses = self
2534 .tool_use
2535 .pending_tool_uses()
2536 .into_iter()
2537 .filter(|tool_use| tool_use.status.is_idle())
2538 .cloned()
2539 .collect::<Vec<_>>();
2540
2541 for tool_use in pending_tool_uses.iter() {
2542 self.use_pending_tool(tool_use.clone(), request.clone(), model.clone(), window, cx);
2543 }
2544
2545 pending_tool_uses
2546 }
2547
2548 fn use_pending_tool(
2549 &mut self,
2550 tool_use: PendingToolUse,
2551 request: Arc<LanguageModelRequest>,
2552 model: Arc<dyn LanguageModel>,
2553 window: Option<AnyWindowHandle>,
2554 cx: &mut Context<Self>,
2555 ) {
2556 let Some(tool) = self.tools.read(cx).tool(&tool_use.name, cx) else {
2557 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2558 };
2559
2560 if !self.profile.is_tool_enabled(tool.source(), tool.name(), cx) {
2561 return self.handle_hallucinated_tool_use(tool_use.id, tool_use.name, window, cx);
2562 }
2563
2564 if tool.needs_confirmation(&tool_use.input, &self.project, cx)
2565 && !AgentSettings::get_global(cx).always_allow_tool_actions
2566 {
2567 self.tool_use.confirm_tool_use(
2568 tool_use.id,
2569 tool_use.ui_text,
2570 tool_use.input,
2571 request,
2572 tool,
2573 );
2574 cx.emit(ThreadEvent::ToolConfirmationNeeded);
2575 } else {
2576 self.run_tool(
2577 tool_use.id,
2578 tool_use.ui_text,
2579 tool_use.input,
2580 request,
2581 tool,
2582 model,
2583 window,
2584 cx,
2585 );
2586 }
2587 }
2588
2589 pub fn handle_hallucinated_tool_use(
2590 &mut self,
2591 tool_use_id: LanguageModelToolUseId,
2592 hallucinated_tool_name: Arc<str>,
2593 window: Option<AnyWindowHandle>,
2594 cx: &mut Context<Thread>,
2595 ) {
2596 let available_tools = self.profile.enabled_tools(cx);
2597
2598 let tool_list = available_tools
2599 .iter()
2600 .map(|(name, tool)| format!("- {}: {}", name, tool.description()))
2601 .collect::<Vec<_>>()
2602 .join("\n");
2603
2604 let error_message = format!(
2605 "The tool '{}' doesn't exist or is not enabled. Available tools:\n{}",
2606 hallucinated_tool_name, tool_list
2607 );
2608
2609 let pending_tool_use = self.tool_use.insert_tool_output(
2610 tool_use_id.clone(),
2611 hallucinated_tool_name,
2612 Err(anyhow!("Missing tool call: {error_message}")),
2613 self.configured_model.as_ref(),
2614 self.completion_mode,
2615 );
2616
2617 cx.emit(ThreadEvent::MissingToolUse {
2618 tool_use_id: tool_use_id.clone(),
2619 ui_text: error_message.into(),
2620 });
2621
2622 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2623 }
2624
2625 pub fn receive_invalid_tool_json(
2626 &mut self,
2627 tool_use_id: LanguageModelToolUseId,
2628 tool_name: Arc<str>,
2629 invalid_json: Arc<str>,
2630 error: String,
2631 window: Option<AnyWindowHandle>,
2632 cx: &mut Context<Thread>,
2633 ) {
2634 log::error!("The model returned invalid input JSON: {invalid_json}");
2635
2636 let pending_tool_use = self.tool_use.insert_tool_output(
2637 tool_use_id.clone(),
2638 tool_name,
2639 Err(anyhow!("Error parsing input JSON: {error}")),
2640 self.configured_model.as_ref(),
2641 self.completion_mode,
2642 );
2643 let ui_text = if let Some(pending_tool_use) = &pending_tool_use {
2644 pending_tool_use.ui_text.clone()
2645 } else {
2646 log::error!(
2647 "There was no pending tool use for tool use {tool_use_id}, even though it finished (with invalid input JSON)."
2648 );
2649 format!("Unknown tool {}", tool_use_id).into()
2650 };
2651
2652 cx.emit(ThreadEvent::InvalidToolInput {
2653 tool_use_id: tool_use_id.clone(),
2654 ui_text,
2655 invalid_input_json: invalid_json,
2656 });
2657
2658 self.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2659 }
2660
2661 pub fn run_tool(
2662 &mut self,
2663 tool_use_id: LanguageModelToolUseId,
2664 ui_text: impl Into<SharedString>,
2665 input: serde_json::Value,
2666 request: Arc<LanguageModelRequest>,
2667 tool: Arc<dyn Tool>,
2668 model: Arc<dyn LanguageModel>,
2669 window: Option<AnyWindowHandle>,
2670 cx: &mut Context<Thread>,
2671 ) {
2672 let task =
2673 self.spawn_tool_use(tool_use_id.clone(), request, input, tool, model, window, cx);
2674 self.tool_use
2675 .run_pending_tool(tool_use_id, ui_text.into(), task);
2676 }
2677
2678 fn spawn_tool_use(
2679 &mut self,
2680 tool_use_id: LanguageModelToolUseId,
2681 request: Arc<LanguageModelRequest>,
2682 input: serde_json::Value,
2683 tool: Arc<dyn Tool>,
2684 model: Arc<dyn LanguageModel>,
2685 window: Option<AnyWindowHandle>,
2686 cx: &mut Context<Thread>,
2687 ) -> Task<()> {
2688 let tool_name: Arc<str> = tool.name().into();
2689
2690 let tool_result = tool.run(
2691 input,
2692 request,
2693 self.project.clone(),
2694 self.action_log.clone(),
2695 model,
2696 window,
2697 cx,
2698 );
2699
2700 // Store the card separately if it exists
2701 if let Some(card) = tool_result.card.clone() {
2702 self.tool_use
2703 .insert_tool_result_card(tool_use_id.clone(), card);
2704 }
2705
2706 cx.spawn({
2707 async move |thread: WeakEntity<Thread>, cx| {
2708 let output = tool_result.output.await;
2709
2710 thread
2711 .update(cx, |thread, cx| {
2712 let pending_tool_use = thread.tool_use.insert_tool_output(
2713 tool_use_id.clone(),
2714 tool_name,
2715 output,
2716 thread.configured_model.as_ref(),
2717 thread.completion_mode,
2718 );
2719 thread.tool_finished(tool_use_id, pending_tool_use, false, window, cx);
2720 })
2721 .ok();
2722 }
2723 })
2724 }
2725
2726 fn tool_finished(
2727 &mut self,
2728 tool_use_id: LanguageModelToolUseId,
2729 pending_tool_use: Option<PendingToolUse>,
2730 canceled: bool,
2731 window: Option<AnyWindowHandle>,
2732 cx: &mut Context<Self>,
2733 ) {
2734 if self.all_tools_finished() {
2735 if let Some(ConfiguredModel { model, .. }) = self.configured_model.as_ref() {
2736 if !canceled {
2737 self.send_to_model(model.clone(), CompletionIntent::ToolResults, window, cx);
2738 }
2739 self.auto_capture_telemetry(cx);
2740 }
2741 }
2742
2743 cx.emit(ThreadEvent::ToolFinished {
2744 tool_use_id,
2745 pending_tool_use,
2746 });
2747 }
2748
2749 /// Cancels the last pending completion, if there are any pending.
2750 ///
2751 /// Returns whether a completion was canceled.
2752 pub fn cancel_last_completion(
2753 &mut self,
2754 window: Option<AnyWindowHandle>,
2755 cx: &mut Context<Self>,
2756 ) -> bool {
2757 let mut canceled = self.pending_completions.pop().is_some() || self.retry_state.is_some();
2758
2759 self.retry_state = None;
2760
2761 for pending_tool_use in self.tool_use.cancel_pending() {
2762 canceled = true;
2763 self.tool_finished(
2764 pending_tool_use.id.clone(),
2765 Some(pending_tool_use),
2766 true,
2767 window,
2768 cx,
2769 );
2770 }
2771
2772 if canceled {
2773 cx.emit(ThreadEvent::CompletionCanceled);
2774
2775 // When canceled, we always want to insert the checkpoint.
2776 // (We skip over finalize_pending_checkpoint, because it
2777 // would conclude we didn't have anything to insert here.)
2778 if let Some(checkpoint) = self.pending_checkpoint.take() {
2779 self.insert_checkpoint(checkpoint, cx);
2780 }
2781 } else {
2782 self.finalize_pending_checkpoint(cx);
2783 }
2784
2785 canceled
2786 }
2787
2788 /// Signals that any in-progress editing should be canceled.
2789 ///
2790 /// This method is used to notify listeners (like ActiveThread) that
2791 /// they should cancel any editing operations.
2792 pub fn cancel_editing(&mut self, cx: &mut Context<Self>) {
2793 cx.emit(ThreadEvent::CancelEditing);
2794 }
2795
2796 pub fn feedback(&self) -> Option<ThreadFeedback> {
2797 self.feedback
2798 }
2799
2800 pub fn message_feedback(&self, message_id: MessageId) -> Option<ThreadFeedback> {
2801 self.message_feedback.get(&message_id).copied()
2802 }
2803
2804 pub fn report_message_feedback(
2805 &mut self,
2806 message_id: MessageId,
2807 feedback: ThreadFeedback,
2808 cx: &mut Context<Self>,
2809 ) -> Task<Result<()>> {
2810 if self.message_feedback.get(&message_id) == Some(&feedback) {
2811 return Task::ready(Ok(()));
2812 }
2813
2814 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2815 let serialized_thread = self.serialize(cx);
2816 let thread_id = self.id().clone();
2817 let client = self.project.read(cx).client();
2818
2819 let enabled_tool_names: Vec<String> = self
2820 .profile
2821 .enabled_tools(cx)
2822 .iter()
2823 .map(|(name, _)| name.clone().into())
2824 .collect();
2825
2826 self.message_feedback.insert(message_id, feedback);
2827
2828 cx.notify();
2829
2830 let message_content = self
2831 .message(message_id)
2832 .map(|msg| msg.to_string())
2833 .unwrap_or_default();
2834
2835 cx.background_spawn(async move {
2836 let final_project_snapshot = final_project_snapshot.await;
2837 let serialized_thread = serialized_thread.await?;
2838 let thread_data =
2839 serde_json::to_value(serialized_thread).unwrap_or_else(|_| serde_json::Value::Null);
2840
2841 let rating = match feedback {
2842 ThreadFeedback::Positive => "positive",
2843 ThreadFeedback::Negative => "negative",
2844 };
2845 telemetry::event!(
2846 "Assistant Thread Rated",
2847 rating,
2848 thread_id,
2849 enabled_tool_names,
2850 message_id = message_id.0,
2851 message_content,
2852 thread_data,
2853 final_project_snapshot
2854 );
2855 client.telemetry().flush_events().await;
2856
2857 Ok(())
2858 })
2859 }
2860
2861 pub fn report_feedback(
2862 &mut self,
2863 feedback: ThreadFeedback,
2864 cx: &mut Context<Self>,
2865 ) -> Task<Result<()>> {
2866 let last_assistant_message_id = self
2867 .messages
2868 .iter()
2869 .rev()
2870 .find(|msg| msg.role == Role::Assistant)
2871 .map(|msg| msg.id);
2872
2873 if let Some(message_id) = last_assistant_message_id {
2874 self.report_message_feedback(message_id, feedback, cx)
2875 } else {
2876 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
2877 let serialized_thread = self.serialize(cx);
2878 let thread_id = self.id().clone();
2879 let client = self.project.read(cx).client();
2880 self.feedback = Some(feedback);
2881 cx.notify();
2882
2883 cx.background_spawn(async move {
2884 let final_project_snapshot = final_project_snapshot.await;
2885 let serialized_thread = serialized_thread.await?;
2886 let thread_data = serde_json::to_value(serialized_thread)
2887 .unwrap_or_else(|_| serde_json::Value::Null);
2888
2889 let rating = match feedback {
2890 ThreadFeedback::Positive => "positive",
2891 ThreadFeedback::Negative => "negative",
2892 };
2893 telemetry::event!(
2894 "Assistant Thread Rated",
2895 rating,
2896 thread_id,
2897 thread_data,
2898 final_project_snapshot
2899 );
2900 client.telemetry().flush_events().await;
2901
2902 Ok(())
2903 })
2904 }
2905 }
2906
2907 /// Create a snapshot of the current project state including git information and unsaved buffers.
2908 fn project_snapshot(
2909 project: Entity<Project>,
2910 cx: &mut Context<Self>,
2911 ) -> Task<Arc<ProjectSnapshot>> {
2912 let git_store = project.read(cx).git_store().clone();
2913 let worktree_snapshots: Vec<_> = project
2914 .read(cx)
2915 .visible_worktrees(cx)
2916 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
2917 .collect();
2918
2919 cx.spawn(async move |_, cx| {
2920 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
2921
2922 let mut unsaved_buffers = Vec::new();
2923 cx.update(|app_cx| {
2924 let buffer_store = project.read(app_cx).buffer_store();
2925 for buffer_handle in buffer_store.read(app_cx).buffers() {
2926 let buffer = buffer_handle.read(app_cx);
2927 if buffer.is_dirty() {
2928 if let Some(file) = buffer.file() {
2929 let path = file.path().to_string_lossy().to_string();
2930 unsaved_buffers.push(path);
2931 }
2932 }
2933 }
2934 })
2935 .ok();
2936
2937 Arc::new(ProjectSnapshot {
2938 worktree_snapshots,
2939 unsaved_buffer_paths: unsaved_buffers,
2940 timestamp: Utc::now(),
2941 })
2942 })
2943 }
2944
2945 fn worktree_snapshot(
2946 worktree: Entity<project::Worktree>,
2947 git_store: Entity<GitStore>,
2948 cx: &App,
2949 ) -> Task<WorktreeSnapshot> {
2950 cx.spawn(async move |cx| {
2951 // Get worktree path and snapshot
2952 let worktree_info = cx.update(|app_cx| {
2953 let worktree = worktree.read(app_cx);
2954 let path = worktree.abs_path().to_string_lossy().to_string();
2955 let snapshot = worktree.snapshot();
2956 (path, snapshot)
2957 });
2958
2959 let Ok((worktree_path, _snapshot)) = worktree_info else {
2960 return WorktreeSnapshot {
2961 worktree_path: String::new(),
2962 git_state: None,
2963 };
2964 };
2965
2966 let git_state = git_store
2967 .update(cx, |git_store, cx| {
2968 git_store
2969 .repositories()
2970 .values()
2971 .find(|repo| {
2972 repo.read(cx)
2973 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
2974 .is_some()
2975 })
2976 .cloned()
2977 })
2978 .ok()
2979 .flatten()
2980 .map(|repo| {
2981 repo.update(cx, |repo, _| {
2982 let current_branch =
2983 repo.branch.as_ref().map(|branch| branch.name().to_owned());
2984 repo.send_job(None, |state, _| async move {
2985 let RepositoryState::Local { backend, .. } = state else {
2986 return GitState {
2987 remote_url: None,
2988 head_sha: None,
2989 current_branch,
2990 diff: None,
2991 };
2992 };
2993
2994 let remote_url = backend.remote_url("origin");
2995 let head_sha = backend.head_sha().await;
2996 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
2997
2998 GitState {
2999 remote_url,
3000 head_sha,
3001 current_branch,
3002 diff,
3003 }
3004 })
3005 })
3006 });
3007
3008 let git_state = match git_state {
3009 Some(git_state) => match git_state.ok() {
3010 Some(git_state) => git_state.await.ok(),
3011 None => None,
3012 },
3013 None => None,
3014 };
3015
3016 WorktreeSnapshot {
3017 worktree_path,
3018 git_state,
3019 }
3020 })
3021 }
3022
3023 pub fn to_markdown(&self, cx: &App) -> Result<String> {
3024 let mut markdown = Vec::new();
3025
3026 let summary = self.summary().or_default();
3027 writeln!(markdown, "# {summary}\n")?;
3028
3029 for message in self.messages() {
3030 writeln!(
3031 markdown,
3032 "## {role}\n",
3033 role = match message.role {
3034 Role::User => "User",
3035 Role::Assistant => "Agent",
3036 Role::System => "System",
3037 }
3038 )?;
3039
3040 if !message.loaded_context.text.is_empty() {
3041 writeln!(markdown, "{}", message.loaded_context.text)?;
3042 }
3043
3044 if !message.loaded_context.images.is_empty() {
3045 writeln!(
3046 markdown,
3047 "\n{} images attached as context.\n",
3048 message.loaded_context.images.len()
3049 )?;
3050 }
3051
3052 for segment in &message.segments {
3053 match segment {
3054 MessageSegment::Text(text) => writeln!(markdown, "{}\n", text)?,
3055 MessageSegment::Thinking { text, .. } => {
3056 writeln!(markdown, "<think>\n{}\n</think>\n", text)?
3057 }
3058 MessageSegment::RedactedThinking(_) => {}
3059 }
3060 }
3061
3062 for tool_use in self.tool_uses_for_message(message.id, cx) {
3063 writeln!(
3064 markdown,
3065 "**Use Tool: {} ({})**",
3066 tool_use.name, tool_use.id
3067 )?;
3068 writeln!(markdown, "```json")?;
3069 writeln!(
3070 markdown,
3071 "{}",
3072 serde_json::to_string_pretty(&tool_use.input)?
3073 )?;
3074 writeln!(markdown, "```")?;
3075 }
3076
3077 for tool_result in self.tool_results_for_message(message.id) {
3078 write!(markdown, "\n**Tool Results: {}", tool_result.tool_use_id)?;
3079 if tool_result.is_error {
3080 write!(markdown, " (Error)")?;
3081 }
3082
3083 writeln!(markdown, "**\n")?;
3084 match &tool_result.content {
3085 LanguageModelToolResultContent::Text(text) => {
3086 writeln!(markdown, "{text}")?;
3087 }
3088 LanguageModelToolResultContent::Image(image) => {
3089 writeln!(markdown, "", image.source)?;
3090 }
3091 }
3092
3093 if let Some(output) = tool_result.output.as_ref() {
3094 writeln!(
3095 markdown,
3096 "\n\nDebug Output:\n\n```json\n{}\n```\n",
3097 serde_json::to_string_pretty(output)?
3098 )?;
3099 }
3100 }
3101 }
3102
3103 Ok(String::from_utf8_lossy(&markdown).to_string())
3104 }
3105
3106 pub fn keep_edits_in_range(
3107 &mut self,
3108 buffer: Entity<language::Buffer>,
3109 buffer_range: Range<language::Anchor>,
3110 cx: &mut Context<Self>,
3111 ) {
3112 self.action_log.update(cx, |action_log, cx| {
3113 action_log.keep_edits_in_range(buffer, buffer_range, cx)
3114 });
3115 }
3116
3117 pub fn keep_all_edits(&mut self, cx: &mut Context<Self>) {
3118 self.action_log
3119 .update(cx, |action_log, cx| action_log.keep_all_edits(cx));
3120 }
3121
3122 pub fn reject_edits_in_ranges(
3123 &mut self,
3124 buffer: Entity<language::Buffer>,
3125 buffer_ranges: Vec<Range<language::Anchor>>,
3126 cx: &mut Context<Self>,
3127 ) -> Task<Result<()>> {
3128 self.action_log.update(cx, |action_log, cx| {
3129 action_log.reject_edits_in_ranges(buffer, buffer_ranges, cx)
3130 })
3131 }
3132
3133 pub fn action_log(&self) -> &Entity<ActionLog> {
3134 &self.action_log
3135 }
3136
3137 pub fn project(&self) -> &Entity<Project> {
3138 &self.project
3139 }
3140
3141 pub fn auto_capture_telemetry(&mut self, cx: &mut Context<Self>) {
3142 if !cx.has_flag::<feature_flags::ThreadAutoCaptureFeatureFlag>() {
3143 return;
3144 }
3145
3146 let now = Instant::now();
3147 if let Some(last) = self.last_auto_capture_at {
3148 if now.duration_since(last).as_secs() < 10 {
3149 return;
3150 }
3151 }
3152
3153 self.last_auto_capture_at = Some(now);
3154
3155 let thread_id = self.id().clone();
3156 let github_login = self
3157 .project
3158 .read(cx)
3159 .user_store()
3160 .read(cx)
3161 .current_user()
3162 .map(|user| user.github_login.clone());
3163 let client = self.project.read(cx).client();
3164 let serialize_task = self.serialize(cx);
3165
3166 cx.background_executor()
3167 .spawn(async move {
3168 if let Ok(serialized_thread) = serialize_task.await {
3169 if let Ok(thread_data) = serde_json::to_value(serialized_thread) {
3170 telemetry::event!(
3171 "Agent Thread Auto-Captured",
3172 thread_id = thread_id.to_string(),
3173 thread_data = thread_data,
3174 auto_capture_reason = "tracked_user",
3175 github_login = github_login
3176 );
3177
3178 client.telemetry().flush_events().await;
3179 }
3180 }
3181 })
3182 .detach();
3183 }
3184
3185 pub fn cumulative_token_usage(&self) -> TokenUsage {
3186 self.cumulative_token_usage
3187 }
3188
3189 pub fn token_usage_up_to_message(&self, message_id: MessageId) -> TotalTokenUsage {
3190 let Some(model) = self.configured_model.as_ref() else {
3191 return TotalTokenUsage::default();
3192 };
3193
3194 let max = model
3195 .model
3196 .max_token_count_for_mode(self.completion_mode().into());
3197
3198 let index = self
3199 .messages
3200 .iter()
3201 .position(|msg| msg.id == message_id)
3202 .unwrap_or(0);
3203
3204 if index == 0 {
3205 return TotalTokenUsage { total: 0, max };
3206 }
3207
3208 let token_usage = &self
3209 .request_token_usage
3210 .get(index - 1)
3211 .cloned()
3212 .unwrap_or_default();
3213
3214 TotalTokenUsage {
3215 total: token_usage.total_tokens(),
3216 max,
3217 }
3218 }
3219
3220 pub fn total_token_usage(&self) -> Option<TotalTokenUsage> {
3221 let model = self.configured_model.as_ref()?;
3222
3223 let max = model
3224 .model
3225 .max_token_count_for_mode(self.completion_mode().into());
3226
3227 if let Some(exceeded_error) = &self.exceeded_window_error {
3228 if model.model.id() == exceeded_error.model_id {
3229 return Some(TotalTokenUsage {
3230 total: exceeded_error.token_count,
3231 max,
3232 });
3233 }
3234 }
3235
3236 let total = self
3237 .token_usage_at_last_message()
3238 .unwrap_or_default()
3239 .total_tokens();
3240
3241 Some(TotalTokenUsage { total, max })
3242 }
3243
3244 fn token_usage_at_last_message(&self) -> Option<TokenUsage> {
3245 self.request_token_usage
3246 .get(self.messages.len().saturating_sub(1))
3247 .or_else(|| self.request_token_usage.last())
3248 .cloned()
3249 }
3250
3251 fn update_token_usage_at_last_message(&mut self, token_usage: TokenUsage) {
3252 let placeholder = self.token_usage_at_last_message().unwrap_or_default();
3253 self.request_token_usage
3254 .resize(self.messages.len(), placeholder);
3255
3256 if let Some(last) = self.request_token_usage.last_mut() {
3257 *last = token_usage;
3258 }
3259 }
3260
3261 fn update_model_request_usage(&self, amount: u32, limit: UsageLimit, cx: &mut Context<Self>) {
3262 self.project
3263 .read(cx)
3264 .user_store()
3265 .update(cx, |user_store, cx| {
3266 user_store.update_model_request_usage(
3267 ModelRequestUsage(RequestUsage {
3268 amount: amount as i32,
3269 limit,
3270 }),
3271 cx,
3272 )
3273 });
3274 }
3275
3276 pub fn deny_tool_use(
3277 &mut self,
3278 tool_use_id: LanguageModelToolUseId,
3279 tool_name: Arc<str>,
3280 window: Option<AnyWindowHandle>,
3281 cx: &mut Context<Self>,
3282 ) {
3283 let err = Err(anyhow::anyhow!(
3284 "Permission to run tool action denied by user"
3285 ));
3286
3287 self.tool_use.insert_tool_output(
3288 tool_use_id.clone(),
3289 tool_name,
3290 err,
3291 self.configured_model.as_ref(),
3292 self.completion_mode,
3293 );
3294 self.tool_finished(tool_use_id.clone(), None, true, window, cx);
3295 }
3296}
3297
3298#[derive(Debug, Clone, Error)]
3299pub enum ThreadError {
3300 #[error("Payment required")]
3301 PaymentRequired,
3302 #[error("Model request limit reached")]
3303 ModelRequestLimitReached { plan: Plan },
3304 #[error("Message {header}: {message}")]
3305 Message {
3306 header: SharedString,
3307 message: SharedString,
3308 },
3309 #[error("Retryable error: {message}")]
3310 RetryableError {
3311 message: SharedString,
3312 can_enable_burn_mode: bool,
3313 },
3314}
3315
3316#[derive(Debug, Clone)]
3317pub enum ThreadEvent {
3318 ShowError(ThreadError),
3319 StreamedCompletion,
3320 ReceivedTextChunk,
3321 NewRequest,
3322 StreamedAssistantText(MessageId, String),
3323 StreamedAssistantThinking(MessageId, String),
3324 StreamedToolUse {
3325 tool_use_id: LanguageModelToolUseId,
3326 ui_text: Arc<str>,
3327 input: serde_json::Value,
3328 },
3329 MissingToolUse {
3330 tool_use_id: LanguageModelToolUseId,
3331 ui_text: Arc<str>,
3332 },
3333 InvalidToolInput {
3334 tool_use_id: LanguageModelToolUseId,
3335 ui_text: Arc<str>,
3336 invalid_input_json: Arc<str>,
3337 },
3338 Stopped(Result<StopReason, Arc<anyhow::Error>>),
3339 MessageAdded(MessageId),
3340 MessageEdited(MessageId),
3341 MessageDeleted(MessageId),
3342 SummaryGenerated,
3343 SummaryChanged,
3344 UsePendingTools {
3345 tool_uses: Vec<PendingToolUse>,
3346 },
3347 ToolFinished {
3348 #[allow(unused)]
3349 tool_use_id: LanguageModelToolUseId,
3350 /// The pending tool use that corresponds to this tool.
3351 pending_tool_use: Option<PendingToolUse>,
3352 },
3353 CheckpointChanged,
3354 ToolConfirmationNeeded,
3355 ToolUseLimitReached,
3356 CancelEditing,
3357 CompletionCanceled,
3358 ProfileChanged,
3359}
3360
3361impl EventEmitter<ThreadEvent> for Thread {}
3362
3363struct PendingCompletion {
3364 id: usize,
3365 queue_state: QueueState,
3366 _task: Task<()>,
3367}
3368
3369#[cfg(test)]
3370mod tests {
3371 use super::*;
3372 use crate::{
3373 context::load_context, context_store::ContextStore, thread_store, thread_store::ThreadStore,
3374 };
3375
3376 // Test-specific constants
3377 const TEST_RATE_LIMIT_RETRY_SECS: u64 = 30;
3378 use agent_settings::{AgentProfileId, AgentSettings, LanguageModelParameters};
3379 use assistant_tool::ToolRegistry;
3380 use assistant_tools;
3381 use futures::StreamExt;
3382 use futures::future::BoxFuture;
3383 use futures::stream::BoxStream;
3384 use gpui::TestAppContext;
3385 use http_client;
3386 use language_model::fake_provider::{FakeLanguageModel, FakeLanguageModelProvider};
3387 use language_model::{
3388 LanguageModelCompletionError, LanguageModelName, LanguageModelProviderId,
3389 LanguageModelProviderName, LanguageModelToolChoice,
3390 };
3391 use parking_lot::Mutex;
3392 use project::{FakeFs, Project};
3393 use prompt_store::PromptBuilder;
3394 use serde_json::json;
3395 use settings::{Settings, SettingsStore};
3396 use std::sync::Arc;
3397 use std::time::Duration;
3398 use theme::ThemeSettings;
3399 use util::path;
3400 use workspace::Workspace;
3401
3402 #[gpui::test]
3403 async fn test_message_with_context(cx: &mut TestAppContext) {
3404 init_test_settings(cx);
3405
3406 let project = create_test_project(
3407 cx,
3408 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3409 )
3410 .await;
3411
3412 let (_workspace, _thread_store, thread, context_store, model) =
3413 setup_test_environment(cx, project.clone()).await;
3414
3415 add_file_to_context(&project, &context_store, "test/code.rs", cx)
3416 .await
3417 .unwrap();
3418
3419 let context =
3420 context_store.read_with(cx, |store, _| store.context().next().cloned().unwrap());
3421 let loaded_context = cx
3422 .update(|cx| load_context(vec![context], &project, &None, cx))
3423 .await;
3424
3425 // Insert user message with context
3426 let message_id = thread.update(cx, |thread, cx| {
3427 thread.insert_user_message(
3428 "Please explain this code",
3429 loaded_context,
3430 None,
3431 Vec::new(),
3432 cx,
3433 )
3434 });
3435
3436 // Check content and context in message object
3437 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3438
3439 // Use different path format strings based on platform for the test
3440 #[cfg(windows)]
3441 let path_part = r"test\code.rs";
3442 #[cfg(not(windows))]
3443 let path_part = "test/code.rs";
3444
3445 let expected_context = format!(
3446 r#"
3447<context>
3448The following items were attached by the user. They are up-to-date and don't need to be re-read.
3449
3450<files>
3451```rs {path_part}
3452fn main() {{
3453 println!("Hello, world!");
3454}}
3455```
3456</files>
3457</context>
3458"#
3459 );
3460
3461 assert_eq!(message.role, Role::User);
3462 assert_eq!(message.segments.len(), 1);
3463 assert_eq!(
3464 message.segments[0],
3465 MessageSegment::Text("Please explain this code".to_string())
3466 );
3467 assert_eq!(message.loaded_context.text, expected_context);
3468
3469 // Check message in request
3470 let request = thread.update(cx, |thread, cx| {
3471 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3472 });
3473
3474 assert_eq!(request.messages.len(), 2);
3475 let expected_full_message = format!("{}Please explain this code", expected_context);
3476 assert_eq!(request.messages[1].string_contents(), expected_full_message);
3477 }
3478
3479 #[gpui::test]
3480 async fn test_only_include_new_contexts(cx: &mut TestAppContext) {
3481 init_test_settings(cx);
3482
3483 let project = create_test_project(
3484 cx,
3485 json!({
3486 "file1.rs": "fn function1() {}\n",
3487 "file2.rs": "fn function2() {}\n",
3488 "file3.rs": "fn function3() {}\n",
3489 "file4.rs": "fn function4() {}\n",
3490 }),
3491 )
3492 .await;
3493
3494 let (_, _thread_store, thread, context_store, model) =
3495 setup_test_environment(cx, project.clone()).await;
3496
3497 // First message with context 1
3498 add_file_to_context(&project, &context_store, "test/file1.rs", cx)
3499 .await
3500 .unwrap();
3501 let new_contexts = context_store.update(cx, |store, cx| {
3502 store.new_context_for_thread(thread.read(cx), None)
3503 });
3504 assert_eq!(new_contexts.len(), 1);
3505 let loaded_context = cx
3506 .update(|cx| load_context(new_contexts, &project, &None, cx))
3507 .await;
3508 let message1_id = thread.update(cx, |thread, cx| {
3509 thread.insert_user_message("Message 1", loaded_context, None, Vec::new(), cx)
3510 });
3511
3512 // Second message with contexts 1 and 2 (context 1 should be skipped as it's already included)
3513 add_file_to_context(&project, &context_store, "test/file2.rs", cx)
3514 .await
3515 .unwrap();
3516 let new_contexts = context_store.update(cx, |store, cx| {
3517 store.new_context_for_thread(thread.read(cx), None)
3518 });
3519 assert_eq!(new_contexts.len(), 1);
3520 let loaded_context = cx
3521 .update(|cx| load_context(new_contexts, &project, &None, cx))
3522 .await;
3523 let message2_id = thread.update(cx, |thread, cx| {
3524 thread.insert_user_message("Message 2", loaded_context, None, Vec::new(), cx)
3525 });
3526
3527 // Third message with all three contexts (contexts 1 and 2 should be skipped)
3528 //
3529 add_file_to_context(&project, &context_store, "test/file3.rs", cx)
3530 .await
3531 .unwrap();
3532 let new_contexts = context_store.update(cx, |store, cx| {
3533 store.new_context_for_thread(thread.read(cx), None)
3534 });
3535 assert_eq!(new_contexts.len(), 1);
3536 let loaded_context = cx
3537 .update(|cx| load_context(new_contexts, &project, &None, cx))
3538 .await;
3539 let message3_id = thread.update(cx, |thread, cx| {
3540 thread.insert_user_message("Message 3", loaded_context, None, Vec::new(), cx)
3541 });
3542
3543 // Check what contexts are included in each message
3544 let (message1, message2, message3) = thread.read_with(cx, |thread, _| {
3545 (
3546 thread.message(message1_id).unwrap().clone(),
3547 thread.message(message2_id).unwrap().clone(),
3548 thread.message(message3_id).unwrap().clone(),
3549 )
3550 });
3551
3552 // First message should include context 1
3553 assert!(message1.loaded_context.text.contains("file1.rs"));
3554
3555 // Second message should include only context 2 (not 1)
3556 assert!(!message2.loaded_context.text.contains("file1.rs"));
3557 assert!(message2.loaded_context.text.contains("file2.rs"));
3558
3559 // Third message should include only context 3 (not 1 or 2)
3560 assert!(!message3.loaded_context.text.contains("file1.rs"));
3561 assert!(!message3.loaded_context.text.contains("file2.rs"));
3562 assert!(message3.loaded_context.text.contains("file3.rs"));
3563
3564 // Check entire request to make sure all contexts are properly included
3565 let request = thread.update(cx, |thread, cx| {
3566 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3567 });
3568
3569 // The request should contain all 3 messages
3570 assert_eq!(request.messages.len(), 4);
3571
3572 // Check that the contexts are properly formatted in each message
3573 assert!(request.messages[1].string_contents().contains("file1.rs"));
3574 assert!(!request.messages[1].string_contents().contains("file2.rs"));
3575 assert!(!request.messages[1].string_contents().contains("file3.rs"));
3576
3577 assert!(!request.messages[2].string_contents().contains("file1.rs"));
3578 assert!(request.messages[2].string_contents().contains("file2.rs"));
3579 assert!(!request.messages[2].string_contents().contains("file3.rs"));
3580
3581 assert!(!request.messages[3].string_contents().contains("file1.rs"));
3582 assert!(!request.messages[3].string_contents().contains("file2.rs"));
3583 assert!(request.messages[3].string_contents().contains("file3.rs"));
3584
3585 add_file_to_context(&project, &context_store, "test/file4.rs", cx)
3586 .await
3587 .unwrap();
3588 let new_contexts = context_store.update(cx, |store, cx| {
3589 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3590 });
3591 assert_eq!(new_contexts.len(), 3);
3592 let loaded_context = cx
3593 .update(|cx| load_context(new_contexts, &project, &None, cx))
3594 .await
3595 .loaded_context;
3596
3597 assert!(!loaded_context.text.contains("file1.rs"));
3598 assert!(loaded_context.text.contains("file2.rs"));
3599 assert!(loaded_context.text.contains("file3.rs"));
3600 assert!(loaded_context.text.contains("file4.rs"));
3601
3602 let new_contexts = context_store.update(cx, |store, cx| {
3603 // Remove file4.rs
3604 store.remove_context(&loaded_context.contexts[2].handle(), cx);
3605 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3606 });
3607 assert_eq!(new_contexts.len(), 2);
3608 let loaded_context = cx
3609 .update(|cx| load_context(new_contexts, &project, &None, cx))
3610 .await
3611 .loaded_context;
3612
3613 assert!(!loaded_context.text.contains("file1.rs"));
3614 assert!(loaded_context.text.contains("file2.rs"));
3615 assert!(loaded_context.text.contains("file3.rs"));
3616 assert!(!loaded_context.text.contains("file4.rs"));
3617
3618 let new_contexts = context_store.update(cx, |store, cx| {
3619 // Remove file3.rs
3620 store.remove_context(&loaded_context.contexts[1].handle(), cx);
3621 store.new_context_for_thread(thread.read(cx), Some(message2_id))
3622 });
3623 assert_eq!(new_contexts.len(), 1);
3624 let loaded_context = cx
3625 .update(|cx| load_context(new_contexts, &project, &None, cx))
3626 .await
3627 .loaded_context;
3628
3629 assert!(!loaded_context.text.contains("file1.rs"));
3630 assert!(loaded_context.text.contains("file2.rs"));
3631 assert!(!loaded_context.text.contains("file3.rs"));
3632 assert!(!loaded_context.text.contains("file4.rs"));
3633 }
3634
3635 #[gpui::test]
3636 async fn test_message_without_files(cx: &mut TestAppContext) {
3637 init_test_settings(cx);
3638
3639 let project = create_test_project(
3640 cx,
3641 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3642 )
3643 .await;
3644
3645 let (_, _thread_store, thread, _context_store, model) =
3646 setup_test_environment(cx, project.clone()).await;
3647
3648 // Insert user message without any context (empty context vector)
3649 let message_id = thread.update(cx, |thread, cx| {
3650 thread.insert_user_message(
3651 "What is the best way to learn Rust?",
3652 ContextLoadResult::default(),
3653 None,
3654 Vec::new(),
3655 cx,
3656 )
3657 });
3658
3659 // Check content and context in message object
3660 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
3661
3662 // Context should be empty when no files are included
3663 assert_eq!(message.role, Role::User);
3664 assert_eq!(message.segments.len(), 1);
3665 assert_eq!(
3666 message.segments[0],
3667 MessageSegment::Text("What is the best way to learn Rust?".to_string())
3668 );
3669 assert_eq!(message.loaded_context.text, "");
3670
3671 // Check message in request
3672 let request = thread.update(cx, |thread, cx| {
3673 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3674 });
3675
3676 assert_eq!(request.messages.len(), 2);
3677 assert_eq!(
3678 request.messages[1].string_contents(),
3679 "What is the best way to learn Rust?"
3680 );
3681
3682 // Add second message, also without context
3683 let message2_id = thread.update(cx, |thread, cx| {
3684 thread.insert_user_message(
3685 "Are there any good books?",
3686 ContextLoadResult::default(),
3687 None,
3688 Vec::new(),
3689 cx,
3690 )
3691 });
3692
3693 let message2 =
3694 thread.read_with(cx, |thread, _| thread.message(message2_id).unwrap().clone());
3695 assert_eq!(message2.loaded_context.text, "");
3696
3697 // Check that both messages appear in the request
3698 let request = thread.update(cx, |thread, cx| {
3699 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3700 });
3701
3702 assert_eq!(request.messages.len(), 3);
3703 assert_eq!(
3704 request.messages[1].string_contents(),
3705 "What is the best way to learn Rust?"
3706 );
3707 assert_eq!(
3708 request.messages[2].string_contents(),
3709 "Are there any good books?"
3710 );
3711 }
3712
3713 #[gpui::test]
3714 #[ignore] // turn this test on when project_notifications tool is re-enabled
3715 async fn test_stale_buffer_notification(cx: &mut TestAppContext) {
3716 init_test_settings(cx);
3717
3718 let project = create_test_project(
3719 cx,
3720 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3721 )
3722 .await;
3723
3724 let (_workspace, _thread_store, thread, context_store, model) =
3725 setup_test_environment(cx, project.clone()).await;
3726
3727 // Add a buffer to the context. This will be a tracked buffer
3728 let buffer = add_file_to_context(&project, &context_store, "test/code.rs", cx)
3729 .await
3730 .unwrap();
3731
3732 let context = context_store
3733 .read_with(cx, |store, _| store.context().next().cloned())
3734 .unwrap();
3735 let loaded_context = cx
3736 .update(|cx| load_context(vec![context], &project, &None, cx))
3737 .await;
3738
3739 // Insert user message and assistant response
3740 thread.update(cx, |thread, cx| {
3741 thread.insert_user_message("Explain this code", loaded_context, None, Vec::new(), cx);
3742 thread.insert_assistant_message(
3743 vec![MessageSegment::Text("This code prints 42.".into())],
3744 cx,
3745 );
3746 });
3747 cx.run_until_parked();
3748
3749 // We shouldn't have a stale buffer notification yet
3750 let notifications = thread.read_with(cx, |thread, _| {
3751 find_tool_uses(thread, "project_notifications")
3752 });
3753 assert!(
3754 notifications.is_empty(),
3755 "Should not have stale buffer notification before buffer is modified"
3756 );
3757
3758 // Modify the buffer
3759 buffer.update(cx, |buffer, cx| {
3760 buffer.edit(
3761 [(1..1, "\n println!(\"Added a new line\");\n")],
3762 None,
3763 cx,
3764 );
3765 });
3766
3767 // Insert another user message
3768 thread.update(cx, |thread, cx| {
3769 thread.insert_user_message(
3770 "What does the code do now?",
3771 ContextLoadResult::default(),
3772 None,
3773 Vec::new(),
3774 cx,
3775 )
3776 });
3777 cx.run_until_parked();
3778
3779 // Check for the stale buffer warning
3780 thread.update(cx, |thread, cx| {
3781 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3782 });
3783 cx.run_until_parked();
3784
3785 let notifications = thread.read_with(cx, |thread, _cx| {
3786 find_tool_uses(thread, "project_notifications")
3787 });
3788
3789 let [notification] = notifications.as_slice() else {
3790 panic!("Should have a `project_notifications` tool use");
3791 };
3792
3793 let Some(notification_content) = notification.content.to_str() else {
3794 panic!("`project_notifications` should return text");
3795 };
3796
3797 assert!(notification_content.contains("These files have changed since the last read:"));
3798 assert!(notification_content.contains("code.rs"));
3799
3800 // Insert another user message and flush notifications again
3801 thread.update(cx, |thread, cx| {
3802 thread.insert_user_message(
3803 "Can you tell me more?",
3804 ContextLoadResult::default(),
3805 None,
3806 Vec::new(),
3807 cx,
3808 )
3809 });
3810
3811 thread.update(cx, |thread, cx| {
3812 thread.flush_notifications(model.clone(), CompletionIntent::UserPrompt, cx)
3813 });
3814 cx.run_until_parked();
3815
3816 // There should be no new notifications (we already flushed one)
3817 let notifications = thread.read_with(cx, |thread, _cx| {
3818 find_tool_uses(thread, "project_notifications")
3819 });
3820
3821 assert_eq!(
3822 notifications.len(),
3823 1,
3824 "Should still have only one notification after second flush - no duplicates"
3825 );
3826 }
3827
3828 fn find_tool_uses(thread: &Thread, tool_name: &str) -> Vec<LanguageModelToolResult> {
3829 thread
3830 .messages()
3831 .flat_map(|message| {
3832 thread
3833 .tool_results_for_message(message.id)
3834 .into_iter()
3835 .filter(|result| result.tool_name == tool_name.into())
3836 .cloned()
3837 .collect::<Vec<_>>()
3838 })
3839 .collect()
3840 }
3841
3842 #[gpui::test]
3843 async fn test_storing_profile_setting_per_thread(cx: &mut TestAppContext) {
3844 init_test_settings(cx);
3845
3846 let project = create_test_project(
3847 cx,
3848 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3849 )
3850 .await;
3851
3852 let (_workspace, thread_store, thread, _context_store, _model) =
3853 setup_test_environment(cx, project.clone()).await;
3854
3855 // Check that we are starting with the default profile
3856 let profile = cx.read(|cx| thread.read(cx).profile.clone());
3857 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3858 assert_eq!(
3859 profile,
3860 AgentProfile::new(AgentProfileId::default(), tool_set)
3861 );
3862 }
3863
3864 #[gpui::test]
3865 async fn test_serializing_thread_profile(cx: &mut TestAppContext) {
3866 init_test_settings(cx);
3867
3868 let project = create_test_project(
3869 cx,
3870 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3871 )
3872 .await;
3873
3874 let (_workspace, thread_store, thread, _context_store, _model) =
3875 setup_test_environment(cx, project.clone()).await;
3876
3877 // Profile gets serialized with default values
3878 let serialized = thread
3879 .update(cx, |thread, cx| thread.serialize(cx))
3880 .await
3881 .unwrap();
3882
3883 assert_eq!(serialized.profile, Some(AgentProfileId::default()));
3884
3885 let deserialized = cx.update(|cx| {
3886 thread.update(cx, |thread, cx| {
3887 Thread::deserialize(
3888 thread.id.clone(),
3889 serialized,
3890 thread.project.clone(),
3891 thread.tools.clone(),
3892 thread.prompt_builder.clone(),
3893 thread.project_context.clone(),
3894 None,
3895 cx,
3896 )
3897 })
3898 });
3899 let tool_set = cx.read(|cx| thread_store.read(cx).tools());
3900
3901 assert_eq!(
3902 deserialized.profile,
3903 AgentProfile::new(AgentProfileId::default(), tool_set)
3904 );
3905 }
3906
3907 #[gpui::test]
3908 async fn test_temperature_setting(cx: &mut TestAppContext) {
3909 init_test_settings(cx);
3910
3911 let project = create_test_project(
3912 cx,
3913 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
3914 )
3915 .await;
3916
3917 let (_workspace, _thread_store, thread, _context_store, model) =
3918 setup_test_environment(cx, project.clone()).await;
3919
3920 // Both model and provider
3921 cx.update(|cx| {
3922 AgentSettings::override_global(
3923 AgentSettings {
3924 model_parameters: vec![LanguageModelParameters {
3925 provider: Some(model.provider_id().0.to_string().into()),
3926 model: Some(model.id().0.clone()),
3927 temperature: Some(0.66),
3928 }],
3929 ..AgentSettings::get_global(cx).clone()
3930 },
3931 cx,
3932 );
3933 });
3934
3935 let request = thread.update(cx, |thread, cx| {
3936 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3937 });
3938 assert_eq!(request.temperature, Some(0.66));
3939
3940 // Only model
3941 cx.update(|cx| {
3942 AgentSettings::override_global(
3943 AgentSettings {
3944 model_parameters: vec![LanguageModelParameters {
3945 provider: None,
3946 model: Some(model.id().0.clone()),
3947 temperature: Some(0.66),
3948 }],
3949 ..AgentSettings::get_global(cx).clone()
3950 },
3951 cx,
3952 );
3953 });
3954
3955 let request = thread.update(cx, |thread, cx| {
3956 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3957 });
3958 assert_eq!(request.temperature, Some(0.66));
3959
3960 // Only provider
3961 cx.update(|cx| {
3962 AgentSettings::override_global(
3963 AgentSettings {
3964 model_parameters: vec![LanguageModelParameters {
3965 provider: Some(model.provider_id().0.to_string().into()),
3966 model: None,
3967 temperature: Some(0.66),
3968 }],
3969 ..AgentSettings::get_global(cx).clone()
3970 },
3971 cx,
3972 );
3973 });
3974
3975 let request = thread.update(cx, |thread, cx| {
3976 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3977 });
3978 assert_eq!(request.temperature, Some(0.66));
3979
3980 // Same model name, different provider
3981 cx.update(|cx| {
3982 AgentSettings::override_global(
3983 AgentSettings {
3984 model_parameters: vec![LanguageModelParameters {
3985 provider: Some("anthropic".into()),
3986 model: Some(model.id().0.clone()),
3987 temperature: Some(0.66),
3988 }],
3989 ..AgentSettings::get_global(cx).clone()
3990 },
3991 cx,
3992 );
3993 });
3994
3995 let request = thread.update(cx, |thread, cx| {
3996 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
3997 });
3998 assert_eq!(request.temperature, None);
3999 }
4000
4001 #[gpui::test]
4002 async fn test_thread_summary(cx: &mut TestAppContext) {
4003 init_test_settings(cx);
4004
4005 let project = create_test_project(cx, json!({})).await;
4006
4007 let (_, _thread_store, thread, _context_store, model) =
4008 setup_test_environment(cx, project.clone()).await;
4009
4010 // Initial state should be pending
4011 thread.read_with(cx, |thread, _| {
4012 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4013 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4014 });
4015
4016 // Manually setting the summary should not be allowed in this state
4017 thread.update(cx, |thread, cx| {
4018 thread.set_summary("This should not work", cx);
4019 });
4020
4021 thread.read_with(cx, |thread, _| {
4022 assert!(matches!(thread.summary(), ThreadSummary::Pending));
4023 });
4024
4025 // Send a message
4026 thread.update(cx, |thread, cx| {
4027 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
4028 thread.send_to_model(
4029 model.clone(),
4030 CompletionIntent::ThreadSummarization,
4031 None,
4032 cx,
4033 );
4034 });
4035
4036 let fake_model = model.as_fake();
4037 simulate_successful_response(&fake_model, cx);
4038
4039 // Should start generating summary when there are >= 2 messages
4040 thread.read_with(cx, |thread, _| {
4041 assert_eq!(*thread.summary(), ThreadSummary::Generating);
4042 });
4043
4044 // Should not be able to set the summary while generating
4045 thread.update(cx, |thread, cx| {
4046 thread.set_summary("This should not work either", cx);
4047 });
4048
4049 thread.read_with(cx, |thread, _| {
4050 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4051 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4052 });
4053
4054 cx.run_until_parked();
4055 fake_model.send_last_completion_stream_text_chunk("Brief");
4056 fake_model.send_last_completion_stream_text_chunk(" Introduction");
4057 fake_model.end_last_completion_stream();
4058 cx.run_until_parked();
4059
4060 // Summary should be set
4061 thread.read_with(cx, |thread, _| {
4062 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4063 assert_eq!(thread.summary().or_default(), "Brief Introduction");
4064 });
4065
4066 // Now we should be able to set a summary
4067 thread.update(cx, |thread, cx| {
4068 thread.set_summary("Brief Intro", cx);
4069 });
4070
4071 thread.read_with(cx, |thread, _| {
4072 assert_eq!(thread.summary().or_default(), "Brief Intro");
4073 });
4074
4075 // Test setting an empty summary (should default to DEFAULT)
4076 thread.update(cx, |thread, cx| {
4077 thread.set_summary("", cx);
4078 });
4079
4080 thread.read_with(cx, |thread, _| {
4081 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4082 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
4083 });
4084 }
4085
4086 #[gpui::test]
4087 async fn test_thread_summary_error_set_manually(cx: &mut TestAppContext) {
4088 init_test_settings(cx);
4089
4090 let project = create_test_project(cx, json!({})).await;
4091
4092 let (_, _thread_store, thread, _context_store, model) =
4093 setup_test_environment(cx, project.clone()).await;
4094
4095 test_summarize_error(&model, &thread, cx);
4096
4097 // Now we should be able to set a summary
4098 thread.update(cx, |thread, cx| {
4099 thread.set_summary("Brief Intro", cx);
4100 });
4101
4102 thread.read_with(cx, |thread, _| {
4103 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4104 assert_eq!(thread.summary().or_default(), "Brief Intro");
4105 });
4106 }
4107
4108 #[gpui::test]
4109 async fn test_thread_summary_error_retry(cx: &mut TestAppContext) {
4110 init_test_settings(cx);
4111
4112 let project = create_test_project(cx, json!({})).await;
4113
4114 let (_, _thread_store, thread, _context_store, model) =
4115 setup_test_environment(cx, project.clone()).await;
4116
4117 test_summarize_error(&model, &thread, cx);
4118
4119 // Sending another message should not trigger another summarize request
4120 thread.update(cx, |thread, cx| {
4121 thread.insert_user_message(
4122 "How are you?",
4123 ContextLoadResult::default(),
4124 None,
4125 vec![],
4126 cx,
4127 );
4128 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4129 });
4130
4131 let fake_model = model.as_fake();
4132 simulate_successful_response(&fake_model, cx);
4133
4134 thread.read_with(cx, |thread, _| {
4135 // State is still Error, not Generating
4136 assert!(matches!(thread.summary(), ThreadSummary::Error));
4137 });
4138
4139 // But the summarize request can be invoked manually
4140 thread.update(cx, |thread, cx| {
4141 thread.summarize(cx);
4142 });
4143
4144 thread.read_with(cx, |thread, _| {
4145 assert!(matches!(thread.summary(), ThreadSummary::Generating));
4146 });
4147
4148 cx.run_until_parked();
4149 fake_model.send_last_completion_stream_text_chunk("A successful summary");
4150 fake_model.end_last_completion_stream();
4151 cx.run_until_parked();
4152
4153 thread.read_with(cx, |thread, _| {
4154 assert!(matches!(thread.summary(), ThreadSummary::Ready(_)));
4155 assert_eq!(thread.summary().or_default(), "A successful summary");
4156 });
4157 }
4158
4159 // Helper to create a model that returns errors
4160 enum TestError {
4161 Overloaded,
4162 InternalServerError,
4163 }
4164
4165 struct ErrorInjector {
4166 inner: Arc<FakeLanguageModel>,
4167 error_type: TestError,
4168 }
4169
4170 impl ErrorInjector {
4171 fn new(error_type: TestError) -> Self {
4172 Self {
4173 inner: Arc::new(FakeLanguageModel::default()),
4174 error_type,
4175 }
4176 }
4177 }
4178
4179 impl LanguageModel for ErrorInjector {
4180 fn id(&self) -> LanguageModelId {
4181 self.inner.id()
4182 }
4183
4184 fn name(&self) -> LanguageModelName {
4185 self.inner.name()
4186 }
4187
4188 fn provider_id(&self) -> LanguageModelProviderId {
4189 self.inner.provider_id()
4190 }
4191
4192 fn provider_name(&self) -> LanguageModelProviderName {
4193 self.inner.provider_name()
4194 }
4195
4196 fn supports_tools(&self) -> bool {
4197 self.inner.supports_tools()
4198 }
4199
4200 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4201 self.inner.supports_tool_choice(choice)
4202 }
4203
4204 fn supports_images(&self) -> bool {
4205 self.inner.supports_images()
4206 }
4207
4208 fn telemetry_id(&self) -> String {
4209 self.inner.telemetry_id()
4210 }
4211
4212 fn max_token_count(&self) -> u64 {
4213 self.inner.max_token_count()
4214 }
4215
4216 fn count_tokens(
4217 &self,
4218 request: LanguageModelRequest,
4219 cx: &App,
4220 ) -> BoxFuture<'static, Result<u64>> {
4221 self.inner.count_tokens(request, cx)
4222 }
4223
4224 fn stream_completion(
4225 &self,
4226 _request: LanguageModelRequest,
4227 _cx: &AsyncApp,
4228 ) -> BoxFuture<
4229 'static,
4230 Result<
4231 BoxStream<
4232 'static,
4233 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4234 >,
4235 LanguageModelCompletionError,
4236 >,
4237 > {
4238 let error = match self.error_type {
4239 TestError::Overloaded => LanguageModelCompletionError::ServerOverloaded {
4240 provider: self.provider_name(),
4241 retry_after: None,
4242 },
4243 TestError::InternalServerError => {
4244 LanguageModelCompletionError::ApiInternalServerError {
4245 provider: self.provider_name(),
4246 message: "I'm a teapot orbiting the sun".to_string(),
4247 }
4248 }
4249 };
4250 async move {
4251 let stream = futures::stream::once(async move { Err(error) });
4252 Ok(stream.boxed())
4253 }
4254 .boxed()
4255 }
4256
4257 fn as_fake(&self) -> &FakeLanguageModel {
4258 &self.inner
4259 }
4260 }
4261
4262 #[gpui::test]
4263 async fn test_retry_on_overloaded_error(cx: &mut TestAppContext) {
4264 init_test_settings(cx);
4265
4266 let project = create_test_project(cx, json!({})).await;
4267 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4268
4269 // Enable Burn Mode to allow retries
4270 thread.update(cx, |thread, _| {
4271 thread.set_completion_mode(CompletionMode::Burn);
4272 });
4273
4274 // Create model that returns overloaded error
4275 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4276
4277 // Insert a user message
4278 thread.update(cx, |thread, cx| {
4279 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4280 });
4281
4282 // Start completion
4283 thread.update(cx, |thread, cx| {
4284 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4285 });
4286
4287 cx.run_until_parked();
4288
4289 thread.read_with(cx, |thread, _| {
4290 assert!(thread.retry_state.is_some(), "Should have retry state");
4291 let retry_state = thread.retry_state.as_ref().unwrap();
4292 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4293 assert_eq!(
4294 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
4295 "Should retry MAX_RETRY_ATTEMPTS times for overloaded errors"
4296 );
4297 });
4298
4299 // Check that a retry message was added
4300 thread.read_with(cx, |thread, _| {
4301 let mut messages = thread.messages();
4302 assert!(
4303 messages.any(|msg| {
4304 msg.role == Role::System
4305 && msg.ui_only
4306 && msg.segments.iter().any(|seg| {
4307 if let MessageSegment::Text(text) = seg {
4308 text.contains("overloaded")
4309 && text
4310 .contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS))
4311 } else {
4312 false
4313 }
4314 })
4315 }),
4316 "Should have added a system retry message"
4317 );
4318 });
4319
4320 let retry_count = thread.update(cx, |thread, _| {
4321 thread
4322 .messages
4323 .iter()
4324 .filter(|m| {
4325 m.ui_only
4326 && m.segments.iter().any(|s| {
4327 if let MessageSegment::Text(text) = s {
4328 text.contains("Retrying") && text.contains("seconds")
4329 } else {
4330 false
4331 }
4332 })
4333 })
4334 .count()
4335 });
4336
4337 assert_eq!(retry_count, 1, "Should have one retry message");
4338 }
4339
4340 #[gpui::test]
4341 async fn test_retry_on_internal_server_error(cx: &mut TestAppContext) {
4342 init_test_settings(cx);
4343
4344 let project = create_test_project(cx, json!({})).await;
4345 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4346
4347 // Enable Burn Mode to allow retries
4348 thread.update(cx, |thread, _| {
4349 thread.set_completion_mode(CompletionMode::Burn);
4350 });
4351
4352 // Create model that returns internal server error
4353 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4354
4355 // Insert a user message
4356 thread.update(cx, |thread, cx| {
4357 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4358 });
4359
4360 // Start completion
4361 thread.update(cx, |thread, cx| {
4362 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4363 });
4364
4365 cx.run_until_parked();
4366
4367 // Check retry state on thread
4368 thread.read_with(cx, |thread, _| {
4369 assert!(thread.retry_state.is_some(), "Should have retry state");
4370 let retry_state = thread.retry_state.as_ref().unwrap();
4371 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4372 assert_eq!(
4373 retry_state.max_attempts, 3,
4374 "Should have correct max attempts"
4375 );
4376 });
4377
4378 // Check that a retry message was added with provider name
4379 thread.read_with(cx, |thread, _| {
4380 let mut messages = thread.messages();
4381 assert!(
4382 messages.any(|msg| {
4383 msg.role == Role::System
4384 && msg.ui_only
4385 && msg.segments.iter().any(|seg| {
4386 if let MessageSegment::Text(text) = seg {
4387 text.contains("internal")
4388 && text.contains("Fake")
4389 && text.contains("Retrying")
4390 && text.contains("attempt 1 of 3")
4391 && text.contains("seconds")
4392 } else {
4393 false
4394 }
4395 })
4396 }),
4397 "Should have added a system retry message with provider name"
4398 );
4399 });
4400
4401 // Count retry messages
4402 let retry_count = thread.update(cx, |thread, _| {
4403 thread
4404 .messages
4405 .iter()
4406 .filter(|m| {
4407 m.ui_only
4408 && m.segments.iter().any(|s| {
4409 if let MessageSegment::Text(text) = s {
4410 text.contains("Retrying") && text.contains("seconds")
4411 } else {
4412 false
4413 }
4414 })
4415 })
4416 .count()
4417 });
4418
4419 assert_eq!(retry_count, 1, "Should have one retry message");
4420 }
4421
4422 #[gpui::test]
4423 async fn test_exponential_backoff_on_retries(cx: &mut TestAppContext) {
4424 init_test_settings(cx);
4425
4426 let project = create_test_project(cx, json!({})).await;
4427 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4428
4429 // Enable Burn Mode to allow retries
4430 thread.update(cx, |thread, _| {
4431 thread.set_completion_mode(CompletionMode::Burn);
4432 });
4433
4434 // Create model that returns internal server error
4435 let model = Arc::new(ErrorInjector::new(TestError::InternalServerError));
4436
4437 // Insert a user message
4438 thread.update(cx, |thread, cx| {
4439 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4440 });
4441
4442 // Track retry events and completion count
4443 // Track completion events
4444 let completion_count = Arc::new(Mutex::new(0));
4445 let completion_count_clone = completion_count.clone();
4446
4447 let _subscription = thread.update(cx, |_, cx| {
4448 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4449 if let ThreadEvent::NewRequest = event {
4450 *completion_count_clone.lock() += 1;
4451 }
4452 })
4453 });
4454
4455 // First attempt
4456 thread.update(cx, |thread, cx| {
4457 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4458 });
4459 cx.run_until_parked();
4460
4461 // Should have scheduled first retry - count retry messages
4462 let retry_count = thread.update(cx, |thread, _| {
4463 thread
4464 .messages
4465 .iter()
4466 .filter(|m| {
4467 m.ui_only
4468 && m.segments.iter().any(|s| {
4469 if let MessageSegment::Text(text) = s {
4470 text.contains("Retrying") && text.contains("seconds")
4471 } else {
4472 false
4473 }
4474 })
4475 })
4476 .count()
4477 });
4478 assert_eq!(retry_count, 1, "Should have scheduled first retry");
4479
4480 // Check retry state
4481 thread.read_with(cx, |thread, _| {
4482 assert!(thread.retry_state.is_some(), "Should have retry state");
4483 let retry_state = thread.retry_state.as_ref().unwrap();
4484 assert_eq!(retry_state.attempt, 1, "Should be first retry attempt");
4485 assert_eq!(
4486 retry_state.max_attempts, 3,
4487 "Internal server errors should retry up to 3 times"
4488 );
4489 });
4490
4491 // Advance clock for first retry
4492 cx.executor().advance_clock(BASE_RETRY_DELAY);
4493 cx.run_until_parked();
4494
4495 // Advance clock for second retry
4496 cx.executor().advance_clock(BASE_RETRY_DELAY);
4497 cx.run_until_parked();
4498
4499 // Advance clock for third retry
4500 cx.executor().advance_clock(BASE_RETRY_DELAY);
4501 cx.run_until_parked();
4502
4503 // Should have completed all retries - count retry messages
4504 let retry_count = thread.update(cx, |thread, _| {
4505 thread
4506 .messages
4507 .iter()
4508 .filter(|m| {
4509 m.ui_only
4510 && m.segments.iter().any(|s| {
4511 if let MessageSegment::Text(text) = s {
4512 text.contains("Retrying") && text.contains("seconds")
4513 } else {
4514 false
4515 }
4516 })
4517 })
4518 .count()
4519 });
4520 assert_eq!(
4521 retry_count, 3,
4522 "Should have 3 retries for internal server errors"
4523 );
4524
4525 // For internal server errors, we retry 3 times and then give up
4526 // Check that retry_state is cleared after all retries
4527 thread.read_with(cx, |thread, _| {
4528 assert!(
4529 thread.retry_state.is_none(),
4530 "Retry state should be cleared after all retries"
4531 );
4532 });
4533
4534 // Verify total attempts (1 initial + 3 retries)
4535 assert_eq!(
4536 *completion_count.lock(),
4537 4,
4538 "Should have attempted once plus 3 retries"
4539 );
4540 }
4541
4542 #[gpui::test]
4543 async fn test_max_retries_exceeded(cx: &mut TestAppContext) {
4544 init_test_settings(cx);
4545
4546 let project = create_test_project(cx, json!({})).await;
4547 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4548
4549 // Enable Burn Mode to allow retries
4550 thread.update(cx, |thread, _| {
4551 thread.set_completion_mode(CompletionMode::Burn);
4552 });
4553
4554 // Create model that returns overloaded error
4555 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
4556
4557 // Insert a user message
4558 thread.update(cx, |thread, cx| {
4559 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4560 });
4561
4562 // Track events
4563 let stopped_with_error = Arc::new(Mutex::new(false));
4564 let stopped_with_error_clone = stopped_with_error.clone();
4565
4566 let _subscription = thread.update(cx, |_, cx| {
4567 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4568 if let ThreadEvent::Stopped(Err(_)) = event {
4569 *stopped_with_error_clone.lock() = true;
4570 }
4571 })
4572 });
4573
4574 // Start initial completion
4575 thread.update(cx, |thread, cx| {
4576 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4577 });
4578 cx.run_until_parked();
4579
4580 // Advance through all retries
4581 for _ in 0..MAX_RETRY_ATTEMPTS {
4582 cx.executor().advance_clock(BASE_RETRY_DELAY);
4583 cx.run_until_parked();
4584 }
4585
4586 let retry_count = thread.update(cx, |thread, _| {
4587 thread
4588 .messages
4589 .iter()
4590 .filter(|m| {
4591 m.ui_only
4592 && m.segments.iter().any(|s| {
4593 if let MessageSegment::Text(text) = s {
4594 text.contains("Retrying") && text.contains("seconds")
4595 } else {
4596 false
4597 }
4598 })
4599 })
4600 .count()
4601 });
4602
4603 // After max retries, should emit Stopped(Err(...)) event
4604 assert_eq!(
4605 retry_count, MAX_RETRY_ATTEMPTS as usize,
4606 "Should have attempted MAX_RETRY_ATTEMPTS retries for overloaded errors"
4607 );
4608 assert!(
4609 *stopped_with_error.lock(),
4610 "Should emit Stopped(Err(...)) event after max retries exceeded"
4611 );
4612
4613 // Retry state should be cleared
4614 thread.read_with(cx, |thread, _| {
4615 assert!(
4616 thread.retry_state.is_none(),
4617 "Retry state should be cleared after max retries"
4618 );
4619
4620 // Verify we have the expected number of retry messages
4621 let retry_messages = thread
4622 .messages
4623 .iter()
4624 .filter(|msg| msg.ui_only && msg.role == Role::System)
4625 .count();
4626 assert_eq!(
4627 retry_messages, MAX_RETRY_ATTEMPTS as usize,
4628 "Should have MAX_RETRY_ATTEMPTS retry messages for overloaded errors"
4629 );
4630 });
4631 }
4632
4633 #[gpui::test]
4634 async fn test_retry_message_removed_on_retry(cx: &mut TestAppContext) {
4635 init_test_settings(cx);
4636
4637 let project = create_test_project(cx, json!({})).await;
4638 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4639
4640 // Enable Burn Mode to allow retries
4641 thread.update(cx, |thread, _| {
4642 thread.set_completion_mode(CompletionMode::Burn);
4643 });
4644
4645 // We'll use a wrapper to switch behavior after first failure
4646 struct RetryTestModel {
4647 inner: Arc<FakeLanguageModel>,
4648 failed_once: Arc<Mutex<bool>>,
4649 }
4650
4651 impl LanguageModel for RetryTestModel {
4652 fn id(&self) -> LanguageModelId {
4653 self.inner.id()
4654 }
4655
4656 fn name(&self) -> LanguageModelName {
4657 self.inner.name()
4658 }
4659
4660 fn provider_id(&self) -> LanguageModelProviderId {
4661 self.inner.provider_id()
4662 }
4663
4664 fn provider_name(&self) -> LanguageModelProviderName {
4665 self.inner.provider_name()
4666 }
4667
4668 fn supports_tools(&self) -> bool {
4669 self.inner.supports_tools()
4670 }
4671
4672 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4673 self.inner.supports_tool_choice(choice)
4674 }
4675
4676 fn supports_images(&self) -> bool {
4677 self.inner.supports_images()
4678 }
4679
4680 fn telemetry_id(&self) -> String {
4681 self.inner.telemetry_id()
4682 }
4683
4684 fn max_token_count(&self) -> u64 {
4685 self.inner.max_token_count()
4686 }
4687
4688 fn count_tokens(
4689 &self,
4690 request: LanguageModelRequest,
4691 cx: &App,
4692 ) -> BoxFuture<'static, Result<u64>> {
4693 self.inner.count_tokens(request, cx)
4694 }
4695
4696 fn stream_completion(
4697 &self,
4698 request: LanguageModelRequest,
4699 cx: &AsyncApp,
4700 ) -> BoxFuture<
4701 'static,
4702 Result<
4703 BoxStream<
4704 'static,
4705 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4706 >,
4707 LanguageModelCompletionError,
4708 >,
4709 > {
4710 if !*self.failed_once.lock() {
4711 *self.failed_once.lock() = true;
4712 let provider = self.provider_name();
4713 // Return error on first attempt
4714 let stream = futures::stream::once(async move {
4715 Err(LanguageModelCompletionError::ServerOverloaded {
4716 provider,
4717 retry_after: None,
4718 })
4719 });
4720 async move { Ok(stream.boxed()) }.boxed()
4721 } else {
4722 // Succeed on retry
4723 self.inner.stream_completion(request, cx)
4724 }
4725 }
4726
4727 fn as_fake(&self) -> &FakeLanguageModel {
4728 &self.inner
4729 }
4730 }
4731
4732 let model = Arc::new(RetryTestModel {
4733 inner: Arc::new(FakeLanguageModel::default()),
4734 failed_once: Arc::new(Mutex::new(false)),
4735 });
4736
4737 // Insert a user message
4738 thread.update(cx, |thread, cx| {
4739 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
4740 });
4741
4742 // Track message deletions
4743 // Track when retry completes successfully
4744 let retry_completed = Arc::new(Mutex::new(false));
4745 let retry_completed_clone = retry_completed.clone();
4746
4747 let _subscription = thread.update(cx, |_, cx| {
4748 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
4749 if let ThreadEvent::StreamedCompletion = event {
4750 *retry_completed_clone.lock() = true;
4751 }
4752 })
4753 });
4754
4755 // Start completion
4756 thread.update(cx, |thread, cx| {
4757 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
4758 });
4759 cx.run_until_parked();
4760
4761 // Get the retry message ID
4762 let retry_message_id = thread.read_with(cx, |thread, _| {
4763 thread
4764 .messages()
4765 .find(|msg| msg.role == Role::System && msg.ui_only)
4766 .map(|msg| msg.id)
4767 .expect("Should have a retry message")
4768 });
4769
4770 // Wait for retry
4771 cx.executor().advance_clock(BASE_RETRY_DELAY);
4772 cx.run_until_parked();
4773
4774 // Stream some successful content
4775 let fake_model = model.as_fake();
4776 // After the retry, there should be a new pending completion
4777 let pending = fake_model.pending_completions();
4778 assert!(
4779 !pending.is_empty(),
4780 "Should have a pending completion after retry"
4781 );
4782 fake_model.send_completion_stream_text_chunk(&pending[0], "Success!");
4783 fake_model.end_completion_stream(&pending[0]);
4784 cx.run_until_parked();
4785
4786 // Check that the retry completed successfully
4787 assert!(
4788 *retry_completed.lock(),
4789 "Retry should have completed successfully"
4790 );
4791
4792 // Retry message should still exist but be marked as ui_only
4793 thread.read_with(cx, |thread, _| {
4794 let retry_msg = thread
4795 .message(retry_message_id)
4796 .expect("Retry message should still exist");
4797 assert!(retry_msg.ui_only, "Retry message should be ui_only");
4798 assert_eq!(
4799 retry_msg.role,
4800 Role::System,
4801 "Retry message should have System role"
4802 );
4803 });
4804 }
4805
4806 #[gpui::test]
4807 async fn test_successful_completion_clears_retry_state(cx: &mut TestAppContext) {
4808 init_test_settings(cx);
4809
4810 let project = create_test_project(cx, json!({})).await;
4811 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4812
4813 // Enable Burn Mode to allow retries
4814 thread.update(cx, |thread, _| {
4815 thread.set_completion_mode(CompletionMode::Burn);
4816 });
4817
4818 // Create a model that fails once then succeeds
4819 struct FailOnceModel {
4820 inner: Arc<FakeLanguageModel>,
4821 failed_once: Arc<Mutex<bool>>,
4822 }
4823
4824 impl LanguageModel for FailOnceModel {
4825 fn id(&self) -> LanguageModelId {
4826 self.inner.id()
4827 }
4828
4829 fn name(&self) -> LanguageModelName {
4830 self.inner.name()
4831 }
4832
4833 fn provider_id(&self) -> LanguageModelProviderId {
4834 self.inner.provider_id()
4835 }
4836
4837 fn provider_name(&self) -> LanguageModelProviderName {
4838 self.inner.provider_name()
4839 }
4840
4841 fn supports_tools(&self) -> bool {
4842 self.inner.supports_tools()
4843 }
4844
4845 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
4846 self.inner.supports_tool_choice(choice)
4847 }
4848
4849 fn supports_images(&self) -> bool {
4850 self.inner.supports_images()
4851 }
4852
4853 fn telemetry_id(&self) -> String {
4854 self.inner.telemetry_id()
4855 }
4856
4857 fn max_token_count(&self) -> u64 {
4858 self.inner.max_token_count()
4859 }
4860
4861 fn count_tokens(
4862 &self,
4863 request: LanguageModelRequest,
4864 cx: &App,
4865 ) -> BoxFuture<'static, Result<u64>> {
4866 self.inner.count_tokens(request, cx)
4867 }
4868
4869 fn stream_completion(
4870 &self,
4871 request: LanguageModelRequest,
4872 cx: &AsyncApp,
4873 ) -> BoxFuture<
4874 'static,
4875 Result<
4876 BoxStream<
4877 'static,
4878 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
4879 >,
4880 LanguageModelCompletionError,
4881 >,
4882 > {
4883 if !*self.failed_once.lock() {
4884 *self.failed_once.lock() = true;
4885 let provider = self.provider_name();
4886 // Return error on first attempt
4887 let stream = futures::stream::once(async move {
4888 Err(LanguageModelCompletionError::ServerOverloaded {
4889 provider,
4890 retry_after: None,
4891 })
4892 });
4893 async move { Ok(stream.boxed()) }.boxed()
4894 } else {
4895 // Succeed on retry
4896 self.inner.stream_completion(request, cx)
4897 }
4898 }
4899 }
4900
4901 let fail_once_model = Arc::new(FailOnceModel {
4902 inner: Arc::new(FakeLanguageModel::default()),
4903 failed_once: Arc::new(Mutex::new(false)),
4904 });
4905
4906 // Insert a user message
4907 thread.update(cx, |thread, cx| {
4908 thread.insert_user_message(
4909 "Test message",
4910 ContextLoadResult::default(),
4911 None,
4912 vec![],
4913 cx,
4914 );
4915 });
4916
4917 // Start completion with fail-once model
4918 thread.update(cx, |thread, cx| {
4919 thread.send_to_model(
4920 fail_once_model.clone(),
4921 CompletionIntent::UserPrompt,
4922 None,
4923 cx,
4924 );
4925 });
4926
4927 cx.run_until_parked();
4928
4929 // Verify retry state exists after first failure
4930 thread.read_with(cx, |thread, _| {
4931 assert!(
4932 thread.retry_state.is_some(),
4933 "Should have retry state after failure"
4934 );
4935 });
4936
4937 // Wait for retry delay
4938 cx.executor().advance_clock(BASE_RETRY_DELAY);
4939 cx.run_until_parked();
4940
4941 // The retry should now use our FailOnceModel which should succeed
4942 // We need to help the FakeLanguageModel complete the stream
4943 let inner_fake = fail_once_model.inner.clone();
4944
4945 // Wait a bit for the retry to start
4946 cx.run_until_parked();
4947
4948 // Check for pending completions and complete them
4949 if let Some(pending) = inner_fake.pending_completions().first() {
4950 inner_fake.send_completion_stream_text_chunk(pending, "Success!");
4951 inner_fake.end_completion_stream(pending);
4952 }
4953 cx.run_until_parked();
4954
4955 thread.read_with(cx, |thread, _| {
4956 assert!(
4957 thread.retry_state.is_none(),
4958 "Retry state should be cleared after successful completion"
4959 );
4960
4961 let has_assistant_message = thread
4962 .messages
4963 .iter()
4964 .any(|msg| msg.role == Role::Assistant && !msg.ui_only);
4965 assert!(
4966 has_assistant_message,
4967 "Should have an assistant message after successful retry"
4968 );
4969 });
4970 }
4971
4972 #[gpui::test]
4973 async fn test_rate_limit_retry_single_attempt(cx: &mut TestAppContext) {
4974 init_test_settings(cx);
4975
4976 let project = create_test_project(cx, json!({})).await;
4977 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
4978
4979 // Enable Burn Mode to allow retries
4980 thread.update(cx, |thread, _| {
4981 thread.set_completion_mode(CompletionMode::Burn);
4982 });
4983
4984 // Create a model that returns rate limit error with retry_after
4985 struct RateLimitModel {
4986 inner: Arc<FakeLanguageModel>,
4987 }
4988
4989 impl LanguageModel for RateLimitModel {
4990 fn id(&self) -> LanguageModelId {
4991 self.inner.id()
4992 }
4993
4994 fn name(&self) -> LanguageModelName {
4995 self.inner.name()
4996 }
4997
4998 fn provider_id(&self) -> LanguageModelProviderId {
4999 self.inner.provider_id()
5000 }
5001
5002 fn provider_name(&self) -> LanguageModelProviderName {
5003 self.inner.provider_name()
5004 }
5005
5006 fn supports_tools(&self) -> bool {
5007 self.inner.supports_tools()
5008 }
5009
5010 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
5011 self.inner.supports_tool_choice(choice)
5012 }
5013
5014 fn supports_images(&self) -> bool {
5015 self.inner.supports_images()
5016 }
5017
5018 fn telemetry_id(&self) -> String {
5019 self.inner.telemetry_id()
5020 }
5021
5022 fn max_token_count(&self) -> u64 {
5023 self.inner.max_token_count()
5024 }
5025
5026 fn count_tokens(
5027 &self,
5028 request: LanguageModelRequest,
5029 cx: &App,
5030 ) -> BoxFuture<'static, Result<u64>> {
5031 self.inner.count_tokens(request, cx)
5032 }
5033
5034 fn stream_completion(
5035 &self,
5036 _request: LanguageModelRequest,
5037 _cx: &AsyncApp,
5038 ) -> BoxFuture<
5039 'static,
5040 Result<
5041 BoxStream<
5042 'static,
5043 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
5044 >,
5045 LanguageModelCompletionError,
5046 >,
5047 > {
5048 let provider = self.provider_name();
5049 async move {
5050 let stream = futures::stream::once(async move {
5051 Err(LanguageModelCompletionError::RateLimitExceeded {
5052 provider,
5053 retry_after: Some(Duration::from_secs(TEST_RATE_LIMIT_RETRY_SECS)),
5054 })
5055 });
5056 Ok(stream.boxed())
5057 }
5058 .boxed()
5059 }
5060
5061 fn as_fake(&self) -> &FakeLanguageModel {
5062 &self.inner
5063 }
5064 }
5065
5066 let model = Arc::new(RateLimitModel {
5067 inner: Arc::new(FakeLanguageModel::default()),
5068 });
5069
5070 // Insert a user message
5071 thread.update(cx, |thread, cx| {
5072 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5073 });
5074
5075 // Start completion
5076 thread.update(cx, |thread, cx| {
5077 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5078 });
5079
5080 cx.run_until_parked();
5081
5082 let retry_count = thread.update(cx, |thread, _| {
5083 thread
5084 .messages
5085 .iter()
5086 .filter(|m| {
5087 m.ui_only
5088 && m.segments.iter().any(|s| {
5089 if let MessageSegment::Text(text) = s {
5090 text.contains("rate limit exceeded")
5091 } else {
5092 false
5093 }
5094 })
5095 })
5096 .count()
5097 });
5098 assert_eq!(retry_count, 1, "Should have scheduled one retry");
5099
5100 thread.read_with(cx, |thread, _| {
5101 assert!(
5102 thread.retry_state.is_some(),
5103 "Rate limit errors should set retry_state"
5104 );
5105 if let Some(retry_state) = &thread.retry_state {
5106 assert_eq!(
5107 retry_state.max_attempts, MAX_RETRY_ATTEMPTS,
5108 "Rate limit errors should use MAX_RETRY_ATTEMPTS"
5109 );
5110 }
5111 });
5112
5113 // Verify we have one retry message
5114 thread.read_with(cx, |thread, _| {
5115 let retry_messages = thread
5116 .messages
5117 .iter()
5118 .filter(|msg| {
5119 msg.ui_only
5120 && msg.segments.iter().any(|seg| {
5121 if let MessageSegment::Text(text) = seg {
5122 text.contains("rate limit exceeded")
5123 } else {
5124 false
5125 }
5126 })
5127 })
5128 .count();
5129 assert_eq!(
5130 retry_messages, 1,
5131 "Should have one rate limit retry message"
5132 );
5133 });
5134
5135 // Check that retry message doesn't include attempt count
5136 thread.read_with(cx, |thread, _| {
5137 let retry_message = thread
5138 .messages
5139 .iter()
5140 .find(|msg| msg.role == Role::System && msg.ui_only)
5141 .expect("Should have a retry message");
5142
5143 // Check that the message contains attempt count since we use retry_state
5144 if let Some(MessageSegment::Text(text)) = retry_message.segments.first() {
5145 assert!(
5146 text.contains(&format!("attempt 1 of {}", MAX_RETRY_ATTEMPTS)),
5147 "Rate limit retry message should contain attempt count with MAX_RETRY_ATTEMPTS"
5148 );
5149 assert!(
5150 text.contains("Retrying"),
5151 "Rate limit retry message should contain retry text"
5152 );
5153 }
5154 });
5155 }
5156
5157 #[gpui::test]
5158 async fn test_ui_only_messages_not_sent_to_model(cx: &mut TestAppContext) {
5159 init_test_settings(cx);
5160
5161 let project = create_test_project(cx, json!({})).await;
5162 let (_, _, thread, _, model) = setup_test_environment(cx, project.clone()).await;
5163
5164 // Insert a regular user message
5165 thread.update(cx, |thread, cx| {
5166 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5167 });
5168
5169 // Insert a UI-only message (like our retry notifications)
5170 thread.update(cx, |thread, cx| {
5171 let id = thread.next_message_id.post_inc();
5172 thread.messages.push(Message {
5173 id,
5174 role: Role::System,
5175 segments: vec![MessageSegment::Text(
5176 "This is a UI-only message that should not be sent to the model".to_string(),
5177 )],
5178 loaded_context: LoadedContext::default(),
5179 creases: Vec::new(),
5180 is_hidden: true,
5181 ui_only: true,
5182 });
5183 cx.emit(ThreadEvent::MessageAdded(id));
5184 });
5185
5186 // Insert another regular message
5187 thread.update(cx, |thread, cx| {
5188 thread.insert_user_message(
5189 "How are you?",
5190 ContextLoadResult::default(),
5191 None,
5192 vec![],
5193 cx,
5194 );
5195 });
5196
5197 // Generate the completion request
5198 let request = thread.update(cx, |thread, cx| {
5199 thread.to_completion_request(model.clone(), CompletionIntent::UserPrompt, cx)
5200 });
5201
5202 // Verify that the request only contains non-UI-only messages
5203 // Should have system prompt + 2 user messages, but not the UI-only message
5204 let user_messages: Vec<_> = request
5205 .messages
5206 .iter()
5207 .filter(|msg| msg.role == Role::User)
5208 .collect();
5209 assert_eq!(
5210 user_messages.len(),
5211 2,
5212 "Should have exactly 2 user messages"
5213 );
5214
5215 // Verify the UI-only content is not present anywhere in the request
5216 let request_text = request
5217 .messages
5218 .iter()
5219 .flat_map(|msg| &msg.content)
5220 .filter_map(|content| match content {
5221 MessageContent::Text(text) => Some(text.as_str()),
5222 _ => None,
5223 })
5224 .collect::<String>();
5225
5226 assert!(
5227 !request_text.contains("UI-only message"),
5228 "UI-only message content should not be in the request"
5229 );
5230
5231 // Verify the thread still has all 3 messages (including UI-only)
5232 thread.read_with(cx, |thread, _| {
5233 assert_eq!(
5234 thread.messages().count(),
5235 3,
5236 "Thread should have 3 messages"
5237 );
5238 assert_eq!(
5239 thread.messages().filter(|m| m.ui_only).count(),
5240 1,
5241 "Thread should have 1 UI-only message"
5242 );
5243 });
5244
5245 // Verify that UI-only messages are not serialized
5246 let serialized = thread
5247 .update(cx, |thread, cx| thread.serialize(cx))
5248 .await
5249 .unwrap();
5250 assert_eq!(
5251 serialized.messages.len(),
5252 2,
5253 "Serialized thread should only have 2 messages (no UI-only)"
5254 );
5255 }
5256
5257 #[gpui::test]
5258 async fn test_no_retry_without_burn_mode(cx: &mut TestAppContext) {
5259 init_test_settings(cx);
5260
5261 let project = create_test_project(cx, json!({})).await;
5262 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5263
5264 // Ensure we're in Normal mode (not Burn mode)
5265 thread.update(cx, |thread, _| {
5266 thread.set_completion_mode(CompletionMode::Normal);
5267 });
5268
5269 // Track error events
5270 let error_events = Arc::new(Mutex::new(Vec::new()));
5271 let error_events_clone = error_events.clone();
5272
5273 let _subscription = thread.update(cx, |_, cx| {
5274 cx.subscribe(&thread, move |_, _, event: &ThreadEvent, _| {
5275 if let ThreadEvent::ShowError(error) = event {
5276 error_events_clone.lock().push(error.clone());
5277 }
5278 })
5279 });
5280
5281 // Create model that returns overloaded error
5282 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5283
5284 // Insert a user message
5285 thread.update(cx, |thread, cx| {
5286 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5287 });
5288
5289 // Start completion
5290 thread.update(cx, |thread, cx| {
5291 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5292 });
5293
5294 cx.run_until_parked();
5295
5296 // Verify no retry state was created
5297 thread.read_with(cx, |thread, _| {
5298 assert!(
5299 thread.retry_state.is_none(),
5300 "Should not have retry state in Normal mode"
5301 );
5302 });
5303
5304 // Check that a retryable error was reported
5305 let errors = error_events.lock();
5306 assert!(!errors.is_empty(), "Should have received an error event");
5307
5308 if let ThreadError::RetryableError {
5309 message: _,
5310 can_enable_burn_mode,
5311 } = &errors[0]
5312 {
5313 assert!(
5314 *can_enable_burn_mode,
5315 "Error should indicate burn mode can be enabled"
5316 );
5317 } else {
5318 panic!("Expected RetryableError, got {:?}", errors[0]);
5319 }
5320
5321 // Verify the thread is no longer generating
5322 thread.read_with(cx, |thread, _| {
5323 assert!(
5324 !thread.is_generating(),
5325 "Should not be generating after error without retry"
5326 );
5327 });
5328 }
5329
5330 #[gpui::test]
5331 async fn test_retry_cancelled_on_stop(cx: &mut TestAppContext) {
5332 init_test_settings(cx);
5333
5334 let project = create_test_project(cx, json!({})).await;
5335 let (_, _, thread, _, _base_model) = setup_test_environment(cx, project.clone()).await;
5336
5337 // Enable Burn Mode to allow retries
5338 thread.update(cx, |thread, _| {
5339 thread.set_completion_mode(CompletionMode::Burn);
5340 });
5341
5342 // Create model that returns overloaded error
5343 let model = Arc::new(ErrorInjector::new(TestError::Overloaded));
5344
5345 // Insert a user message
5346 thread.update(cx, |thread, cx| {
5347 thread.insert_user_message("Hello!", ContextLoadResult::default(), None, vec![], cx);
5348 });
5349
5350 // Start completion
5351 thread.update(cx, |thread, cx| {
5352 thread.send_to_model(model.clone(), CompletionIntent::UserPrompt, None, cx);
5353 });
5354
5355 cx.run_until_parked();
5356
5357 // Verify retry was scheduled by checking for retry message
5358 let has_retry_message = thread.read_with(cx, |thread, _| {
5359 thread.messages.iter().any(|m| {
5360 m.ui_only
5361 && m.segments.iter().any(|s| {
5362 if let MessageSegment::Text(text) = s {
5363 text.contains("Retrying") && text.contains("seconds")
5364 } else {
5365 false
5366 }
5367 })
5368 })
5369 });
5370 assert!(has_retry_message, "Should have scheduled a retry");
5371
5372 // Cancel the completion before the retry happens
5373 thread.update(cx, |thread, cx| {
5374 thread.cancel_last_completion(None, cx);
5375 });
5376
5377 cx.run_until_parked();
5378
5379 // The retry should not have happened - no pending completions
5380 let fake_model = model.as_fake();
5381 assert_eq!(
5382 fake_model.pending_completions().len(),
5383 0,
5384 "Should have no pending completions after cancellation"
5385 );
5386
5387 // Verify the retry was cancelled by checking retry state
5388 thread.read_with(cx, |thread, _| {
5389 if let Some(retry_state) = &thread.retry_state {
5390 panic!(
5391 "retry_state should be cleared after cancellation, but found: attempt={}, max_attempts={}, intent={:?}",
5392 retry_state.attempt, retry_state.max_attempts, retry_state.intent
5393 );
5394 }
5395 });
5396 }
5397
5398 fn test_summarize_error(
5399 model: &Arc<dyn LanguageModel>,
5400 thread: &Entity<Thread>,
5401 cx: &mut TestAppContext,
5402 ) {
5403 thread.update(cx, |thread, cx| {
5404 thread.insert_user_message("Hi!", ContextLoadResult::default(), None, vec![], cx);
5405 thread.send_to_model(
5406 model.clone(),
5407 CompletionIntent::ThreadSummarization,
5408 None,
5409 cx,
5410 );
5411 });
5412
5413 let fake_model = model.as_fake();
5414 simulate_successful_response(&fake_model, cx);
5415
5416 thread.read_with(cx, |thread, _| {
5417 assert!(matches!(thread.summary(), ThreadSummary::Generating));
5418 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5419 });
5420
5421 // Simulate summary request ending
5422 cx.run_until_parked();
5423 fake_model.end_last_completion_stream();
5424 cx.run_until_parked();
5425
5426 // State is set to Error and default message
5427 thread.read_with(cx, |thread, _| {
5428 assert!(matches!(thread.summary(), ThreadSummary::Error));
5429 assert_eq!(thread.summary().or_default(), ThreadSummary::DEFAULT);
5430 });
5431 }
5432
5433 fn simulate_successful_response(fake_model: &FakeLanguageModel, cx: &mut TestAppContext) {
5434 cx.run_until_parked();
5435 fake_model.send_last_completion_stream_text_chunk("Assistant response");
5436 fake_model.end_last_completion_stream();
5437 cx.run_until_parked();
5438 }
5439
5440 fn init_test_settings(cx: &mut TestAppContext) {
5441 cx.update(|cx| {
5442 let settings_store = SettingsStore::test(cx);
5443 cx.set_global(settings_store);
5444 language::init(cx);
5445 Project::init_settings(cx);
5446 AgentSettings::register(cx);
5447 prompt_store::init(cx);
5448 thread_store::init(cx);
5449 workspace::init_settings(cx);
5450 language_model::init_settings(cx);
5451 ThemeSettings::register(cx);
5452 ToolRegistry::default_global(cx);
5453 assistant_tool::init(cx);
5454
5455 let http_client = Arc::new(http_client::HttpClientWithUrl::new(
5456 http_client::FakeHttpClient::with_200_response(),
5457 "http://localhost".to_string(),
5458 None,
5459 ));
5460 assistant_tools::init(http_client, cx);
5461 });
5462 }
5463
5464 // Helper to create a test project with test files
5465 async fn create_test_project(
5466 cx: &mut TestAppContext,
5467 files: serde_json::Value,
5468 ) -> Entity<Project> {
5469 let fs = FakeFs::new(cx.executor());
5470 fs.insert_tree(path!("/test"), files).await;
5471 Project::test(fs, [path!("/test").as_ref()], cx).await
5472 }
5473
5474 async fn setup_test_environment(
5475 cx: &mut TestAppContext,
5476 project: Entity<Project>,
5477 ) -> (
5478 Entity<Workspace>,
5479 Entity<ThreadStore>,
5480 Entity<Thread>,
5481 Entity<ContextStore>,
5482 Arc<dyn LanguageModel>,
5483 ) {
5484 let (workspace, cx) =
5485 cx.add_window_view(|window, cx| Workspace::test_new(project.clone(), window, cx));
5486
5487 let thread_store = cx
5488 .update(|_, cx| {
5489 ThreadStore::load(
5490 project.clone(),
5491 cx.new(|_| ToolWorkingSet::default()),
5492 None,
5493 Arc::new(PromptBuilder::new(None).unwrap()),
5494 cx,
5495 )
5496 })
5497 .await
5498 .unwrap();
5499
5500 let thread = thread_store.update(cx, |store, cx| store.create_thread(cx));
5501 let context_store = cx.new(|_cx| ContextStore::new(project.downgrade(), None));
5502
5503 let provider = Arc::new(FakeLanguageModelProvider::default());
5504 let model = provider.test_model();
5505 let model: Arc<dyn LanguageModel> = Arc::new(model);
5506
5507 cx.update(|_, cx| {
5508 LanguageModelRegistry::global(cx).update(cx, |registry, cx| {
5509 registry.set_default_model(
5510 Some(ConfiguredModel {
5511 provider: provider.clone(),
5512 model: model.clone(),
5513 }),
5514 cx,
5515 );
5516 registry.set_thread_summary_model(
5517 Some(ConfiguredModel {
5518 provider,
5519 model: model.clone(),
5520 }),
5521 cx,
5522 );
5523 })
5524 });
5525
5526 (workspace, thread_store, thread, context_store, model)
5527 }
5528
5529 async fn add_file_to_context(
5530 project: &Entity<Project>,
5531 context_store: &Entity<ContextStore>,
5532 path: &str,
5533 cx: &mut TestAppContext,
5534 ) -> Result<Entity<language::Buffer>> {
5535 let buffer_path = project
5536 .read_with(cx, |project, cx| project.find_project_path(path, cx))
5537 .unwrap();
5538
5539 let buffer = project
5540 .update(cx, |project, cx| {
5541 project.open_buffer(buffer_path.clone(), cx)
5542 })
5543 .await
5544 .unwrap();
5545
5546 context_store.update(cx, |context_store, cx| {
5547 context_store.add_file_from_buffer(&buffer_path, buffer.clone(), false, cx);
5548 });
5549
5550 Ok(buffer)
5551 }
5552}