1use std::fmt::Write as _;
2use std::io::Write;
3use std::ops::Range;
4use std::sync::Arc;
5
6use anyhow::{Context as _, Result, anyhow};
7use assistant_settings::AssistantSettings;
8use assistant_tool::{ActionLog, Tool, ToolWorkingSet};
9use chrono::{DateTime, Utc};
10use collections::{BTreeMap, HashMap};
11use fs::Fs;
12use futures::future::Shared;
13use futures::{FutureExt, StreamExt as _};
14use git::repository::DiffType;
15use gpui::{App, AppContext, Context, Entity, EventEmitter, SharedString, Task, WeakEntity};
16use language_model::{
17 LanguageModel, LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
18 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
19 LanguageModelToolUseId, MaxMonthlySpendReachedError, MessageContent, PaymentRequiredError,
20 Role, StopReason, TokenUsage,
21};
22use project::git_store::{GitStore, GitStoreCheckpoint, RepositoryState};
23use project::{Project, Worktree};
24use prompt_store::{
25 AssistantSystemPromptContext, PromptBuilder, RulesFile, WorktreeInfoForSystemPrompt,
26};
27use schemars::JsonSchema;
28use serde::{Deserialize, Serialize};
29use settings::Settings;
30use util::{ResultExt as _, TryFutureExt as _, maybe, post_inc};
31use uuid::Uuid;
32
33use crate::context::{AssistantContext, ContextId, format_context_as_string};
34use crate::thread_store::{
35 SerializedMessage, SerializedMessageSegment, SerializedThread, SerializedToolResult,
36 SerializedToolUse,
37};
38use crate::tool_use::{PendingToolUse, ToolUse, ToolUseState, USING_TOOL_MARKER};
39
40#[derive(Debug, Clone, Copy)]
41pub enum RequestKind {
42 Chat,
43 /// Used when summarizing a thread.
44 Summarize,
45}
46
47#[derive(
48 Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize, JsonSchema,
49)]
50pub struct ThreadId(Arc<str>);
51
52impl ThreadId {
53 pub fn new() -> Self {
54 Self(Uuid::new_v4().to_string().into())
55 }
56}
57
58impl std::fmt::Display for ThreadId {
59 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
60 write!(f, "{}", self.0)
61 }
62}
63
64impl From<&str> for ThreadId {
65 fn from(value: &str) -> Self {
66 Self(value.into())
67 }
68}
69
70#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Copy, Serialize, Deserialize)]
71pub struct MessageId(pub(crate) usize);
72
73impl MessageId {
74 fn post_inc(&mut self) -> Self {
75 Self(post_inc(&mut self.0))
76 }
77}
78
79/// A message in a [`Thread`].
80#[derive(Debug, Clone)]
81pub struct Message {
82 pub id: MessageId,
83 pub role: Role,
84 pub segments: Vec<MessageSegment>,
85 pub context: String,
86}
87
88impl Message {
89 /// Returns whether the message contains any meaningful text that should be displayed
90 /// The model sometimes runs tool without producing any text or just a marker ([`USING_TOOL_MARKER`])
91 pub fn should_display_content(&self) -> bool {
92 self.segments.iter().all(|segment| segment.should_display())
93 }
94
95 pub fn push_thinking(&mut self, text: &str) {
96 if let Some(MessageSegment::Thinking(segment)) = self.segments.last_mut() {
97 segment.push_str(text);
98 } else {
99 self.segments
100 .push(MessageSegment::Thinking(text.to_string()));
101 }
102 }
103
104 pub fn push_text(&mut self, text: &str) {
105 if let Some(MessageSegment::Text(segment)) = self.segments.last_mut() {
106 segment.push_str(text);
107 } else {
108 self.segments.push(MessageSegment::Text(text.to_string()));
109 }
110 }
111
112 pub fn to_string(&self) -> String {
113 let mut result = String::new();
114
115 if !self.context.is_empty() {
116 result.push_str(&self.context);
117 }
118
119 for segment in &self.segments {
120 match segment {
121 MessageSegment::Text(text) => result.push_str(text),
122 MessageSegment::Thinking(text) => {
123 result.push_str("<think>");
124 result.push_str(text);
125 result.push_str("</think>");
126 }
127 }
128 }
129
130 result
131 }
132}
133
134#[derive(Debug, Clone, PartialEq, Eq)]
135pub enum MessageSegment {
136 Text(String),
137 Thinking(String),
138}
139
140impl MessageSegment {
141 pub fn text_mut(&mut self) -> &mut String {
142 match self {
143 Self::Text(text) => text,
144 Self::Thinking(text) => text,
145 }
146 }
147
148 pub fn should_display(&self) -> bool {
149 // We add USING_TOOL_MARKER when making a request that includes tool uses
150 // without non-whitespace text around them, and this can cause the model
151 // to mimic the pattern, so we consider those segments not displayable.
152 match self {
153 Self::Text(text) => text.is_empty() || text.trim() == USING_TOOL_MARKER,
154 Self::Thinking(text) => text.is_empty() || text.trim() == USING_TOOL_MARKER,
155 }
156 }
157}
158
159#[derive(Debug, Clone, Serialize, Deserialize)]
160pub struct ProjectSnapshot {
161 pub worktree_snapshots: Vec<WorktreeSnapshot>,
162 pub unsaved_buffer_paths: Vec<String>,
163 pub timestamp: DateTime<Utc>,
164}
165
166#[derive(Debug, Clone, Serialize, Deserialize)]
167pub struct WorktreeSnapshot {
168 pub worktree_path: String,
169 pub git_state: Option<GitState>,
170}
171
172#[derive(Debug, Clone, Serialize, Deserialize)]
173pub struct GitState {
174 pub remote_url: Option<String>,
175 pub head_sha: Option<String>,
176 pub current_branch: Option<String>,
177 pub diff: Option<String>,
178}
179
180#[derive(Clone)]
181pub struct ThreadCheckpoint {
182 message_id: MessageId,
183 git_checkpoint: GitStoreCheckpoint,
184}
185
186#[derive(Copy, Clone, Debug)]
187pub enum ThreadFeedback {
188 Positive,
189 Negative,
190}
191
192pub enum LastRestoreCheckpoint {
193 Pending {
194 message_id: MessageId,
195 },
196 Error {
197 message_id: MessageId,
198 error: String,
199 },
200}
201
202impl LastRestoreCheckpoint {
203 pub fn message_id(&self) -> MessageId {
204 match self {
205 LastRestoreCheckpoint::Pending { message_id } => *message_id,
206 LastRestoreCheckpoint::Error { message_id, .. } => *message_id,
207 }
208 }
209}
210
211#[derive(Clone, Debug, Default, Serialize, Deserialize)]
212pub enum DetailedSummaryState {
213 #[default]
214 NotGenerated,
215 Generating {
216 message_id: MessageId,
217 },
218 Generated {
219 text: SharedString,
220 message_id: MessageId,
221 },
222}
223
224#[derive(Default)]
225pub struct TotalTokenUsage {
226 pub total: usize,
227 pub max: usize,
228 pub ratio: TokenUsageRatio,
229}
230
231#[derive(Default, PartialEq, Eq)]
232pub enum TokenUsageRatio {
233 #[default]
234 Normal,
235 Warning,
236 Exceeded,
237}
238
239/// A thread of conversation with the LLM.
240pub struct Thread {
241 id: ThreadId,
242 updated_at: DateTime<Utc>,
243 summary: Option<SharedString>,
244 pending_summary: Task<Option<()>>,
245 detailed_summary_state: DetailedSummaryState,
246 messages: Vec<Message>,
247 next_message_id: MessageId,
248 context: BTreeMap<ContextId, AssistantContext>,
249 context_by_message: HashMap<MessageId, Vec<ContextId>>,
250 system_prompt_context: Option<AssistantSystemPromptContext>,
251 checkpoints_by_message: HashMap<MessageId, ThreadCheckpoint>,
252 completion_count: usize,
253 pending_completions: Vec<PendingCompletion>,
254 project: Entity<Project>,
255 prompt_builder: Arc<PromptBuilder>,
256 tools: Arc<ToolWorkingSet>,
257 tool_use: ToolUseState,
258 action_log: Entity<ActionLog>,
259 last_restore_checkpoint: Option<LastRestoreCheckpoint>,
260 pending_checkpoint: Option<ThreadCheckpoint>,
261 initial_project_snapshot: Shared<Task<Option<Arc<ProjectSnapshot>>>>,
262 cumulative_token_usage: TokenUsage,
263 feedback: Option<ThreadFeedback>,
264}
265
266impl Thread {
267 pub fn new(
268 project: Entity<Project>,
269 tools: Arc<ToolWorkingSet>,
270 prompt_builder: Arc<PromptBuilder>,
271 cx: &mut Context<Self>,
272 ) -> Self {
273 Self {
274 id: ThreadId::new(),
275 updated_at: Utc::now(),
276 summary: None,
277 pending_summary: Task::ready(None),
278 detailed_summary_state: DetailedSummaryState::NotGenerated,
279 messages: Vec::new(),
280 next_message_id: MessageId(0),
281 context: BTreeMap::default(),
282 context_by_message: HashMap::default(),
283 system_prompt_context: None,
284 checkpoints_by_message: HashMap::default(),
285 completion_count: 0,
286 pending_completions: Vec::new(),
287 project: project.clone(),
288 prompt_builder,
289 tools: tools.clone(),
290 last_restore_checkpoint: None,
291 pending_checkpoint: None,
292 tool_use: ToolUseState::new(tools.clone()),
293 action_log: cx.new(|_| ActionLog::new()),
294 initial_project_snapshot: {
295 let project_snapshot = Self::project_snapshot(project, cx);
296 cx.foreground_executor()
297 .spawn(async move { Some(project_snapshot.await) })
298 .shared()
299 },
300 cumulative_token_usage: TokenUsage::default(),
301 feedback: None,
302 }
303 }
304
305 pub fn deserialize(
306 id: ThreadId,
307 serialized: SerializedThread,
308 project: Entity<Project>,
309 tools: Arc<ToolWorkingSet>,
310 prompt_builder: Arc<PromptBuilder>,
311 cx: &mut Context<Self>,
312 ) -> Self {
313 let next_message_id = MessageId(
314 serialized
315 .messages
316 .last()
317 .map(|message| message.id.0 + 1)
318 .unwrap_or(0),
319 );
320 let tool_use =
321 ToolUseState::from_serialized_messages(tools.clone(), &serialized.messages, |_| true);
322
323 Self {
324 id,
325 updated_at: serialized.updated_at,
326 summary: Some(serialized.summary),
327 pending_summary: Task::ready(None),
328 detailed_summary_state: serialized.detailed_summary_state,
329 messages: serialized
330 .messages
331 .into_iter()
332 .map(|message| Message {
333 id: message.id,
334 role: message.role,
335 segments: message
336 .segments
337 .into_iter()
338 .map(|segment| match segment {
339 SerializedMessageSegment::Text { text } => MessageSegment::Text(text),
340 SerializedMessageSegment::Thinking { text } => {
341 MessageSegment::Thinking(text)
342 }
343 })
344 .collect(),
345 context: message.context,
346 })
347 .collect(),
348 next_message_id,
349 context: BTreeMap::default(),
350 context_by_message: HashMap::default(),
351 system_prompt_context: None,
352 checkpoints_by_message: HashMap::default(),
353 completion_count: 0,
354 pending_completions: Vec::new(),
355 last_restore_checkpoint: None,
356 pending_checkpoint: None,
357 project,
358 prompt_builder,
359 tools,
360 tool_use,
361 action_log: cx.new(|_| ActionLog::new()),
362 initial_project_snapshot: Task::ready(serialized.initial_project_snapshot).shared(),
363 cumulative_token_usage: serialized.cumulative_token_usage,
364 feedback: None,
365 }
366 }
367
368 pub fn id(&self) -> &ThreadId {
369 &self.id
370 }
371
372 pub fn is_empty(&self) -> bool {
373 self.messages.is_empty()
374 }
375
376 pub fn updated_at(&self) -> DateTime<Utc> {
377 self.updated_at
378 }
379
380 pub fn touch_updated_at(&mut self) {
381 self.updated_at = Utc::now();
382 }
383
384 pub fn summary(&self) -> Option<SharedString> {
385 self.summary.clone()
386 }
387
388 pub fn summary_or_default(&self) -> SharedString {
389 const DEFAULT: SharedString = SharedString::new_static("New Thread");
390 self.summary.clone().unwrap_or(DEFAULT)
391 }
392
393 pub fn set_summary(&mut self, summary: impl Into<SharedString>, cx: &mut Context<Self>) {
394 self.summary = Some(summary.into());
395 cx.emit(ThreadEvent::SummaryChanged);
396 }
397
398 pub fn latest_detailed_summary_or_text(&self) -> SharedString {
399 self.latest_detailed_summary()
400 .unwrap_or_else(|| self.text().into())
401 }
402
403 fn latest_detailed_summary(&self) -> Option<SharedString> {
404 if let DetailedSummaryState::Generated { text, .. } = &self.detailed_summary_state {
405 Some(text.clone())
406 } else {
407 None
408 }
409 }
410
411 pub fn message(&self, id: MessageId) -> Option<&Message> {
412 self.messages.iter().find(|message| message.id == id)
413 }
414
415 pub fn messages(&self) -> impl Iterator<Item = &Message> {
416 self.messages.iter()
417 }
418
419 pub fn is_generating(&self) -> bool {
420 !self.pending_completions.is_empty() || !self.all_tools_finished()
421 }
422
423 pub fn tools(&self) -> &Arc<ToolWorkingSet> {
424 &self.tools
425 }
426
427 pub fn pending_tool(&self, id: &LanguageModelToolUseId) -> Option<&PendingToolUse> {
428 self.tool_use
429 .pending_tool_uses()
430 .into_iter()
431 .find(|tool_use| &tool_use.id == id)
432 }
433
434 pub fn tools_needing_confirmation(&self) -> impl Iterator<Item = &PendingToolUse> {
435 self.tool_use
436 .pending_tool_uses()
437 .into_iter()
438 .filter(|tool_use| tool_use.status.needs_confirmation())
439 }
440
441 pub fn has_pending_tool_uses(&self) -> bool {
442 !self.tool_use.pending_tool_uses().is_empty()
443 }
444
445 pub fn checkpoint_for_message(&self, id: MessageId) -> Option<ThreadCheckpoint> {
446 self.checkpoints_by_message.get(&id).cloned()
447 }
448
449 pub fn restore_checkpoint(
450 &mut self,
451 checkpoint: ThreadCheckpoint,
452 cx: &mut Context<Self>,
453 ) -> Task<Result<()>> {
454 self.last_restore_checkpoint = Some(LastRestoreCheckpoint::Pending {
455 message_id: checkpoint.message_id,
456 });
457 cx.emit(ThreadEvent::CheckpointChanged);
458 cx.notify();
459
460 let project = self.project.read(cx);
461 let restore = project
462 .git_store()
463 .read(cx)
464 .restore_checkpoint(checkpoint.git_checkpoint.clone(), cx);
465 cx.spawn(async move |this, cx| {
466 let result = restore.await;
467 this.update(cx, |this, cx| {
468 if let Err(err) = result.as_ref() {
469 this.last_restore_checkpoint = Some(LastRestoreCheckpoint::Error {
470 message_id: checkpoint.message_id,
471 error: err.to_string(),
472 });
473 } else {
474 this.truncate(checkpoint.message_id, cx);
475 this.last_restore_checkpoint = None;
476 }
477 this.pending_checkpoint = None;
478 cx.emit(ThreadEvent::CheckpointChanged);
479 cx.notify();
480 })?;
481 result
482 })
483 }
484
485 fn finalize_pending_checkpoint(&mut self, cx: &mut Context<Self>) {
486 let pending_checkpoint = if self.is_generating() {
487 return;
488 } else if let Some(checkpoint) = self.pending_checkpoint.take() {
489 checkpoint
490 } else {
491 return;
492 };
493
494 let git_store = self.project.read(cx).git_store().clone();
495 let final_checkpoint = git_store.read(cx).checkpoint(cx);
496 cx.spawn(async move |this, cx| match final_checkpoint.await {
497 Ok(final_checkpoint) => {
498 let equal = git_store
499 .read_with(cx, |store, cx| {
500 store.compare_checkpoints(
501 pending_checkpoint.git_checkpoint.clone(),
502 final_checkpoint.clone(),
503 cx,
504 )
505 })?
506 .await
507 .unwrap_or(false);
508
509 if equal {
510 git_store
511 .read_with(cx, |store, cx| {
512 store.delete_checkpoint(pending_checkpoint.git_checkpoint, cx)
513 })?
514 .detach();
515 } else {
516 this.update(cx, |this, cx| {
517 this.insert_checkpoint(pending_checkpoint, cx)
518 })?;
519 }
520
521 git_store
522 .read_with(cx, |store, cx| {
523 store.delete_checkpoint(final_checkpoint, cx)
524 })?
525 .detach();
526
527 Ok(())
528 }
529 Err(_) => this.update(cx, |this, cx| {
530 this.insert_checkpoint(pending_checkpoint, cx)
531 }),
532 })
533 .detach();
534 }
535
536 fn insert_checkpoint(&mut self, checkpoint: ThreadCheckpoint, cx: &mut Context<Self>) {
537 self.checkpoints_by_message
538 .insert(checkpoint.message_id, checkpoint);
539 cx.emit(ThreadEvent::CheckpointChanged);
540 cx.notify();
541 }
542
543 pub fn last_restore_checkpoint(&self) -> Option<&LastRestoreCheckpoint> {
544 self.last_restore_checkpoint.as_ref()
545 }
546
547 pub fn truncate(&mut self, message_id: MessageId, cx: &mut Context<Self>) {
548 let Some(message_ix) = self
549 .messages
550 .iter()
551 .rposition(|message| message.id == message_id)
552 else {
553 return;
554 };
555 for deleted_message in self.messages.drain(message_ix..) {
556 self.context_by_message.remove(&deleted_message.id);
557 self.checkpoints_by_message.remove(&deleted_message.id);
558 }
559 cx.notify();
560 }
561
562 pub fn context_for_message(&self, id: MessageId) -> impl Iterator<Item = &AssistantContext> {
563 self.context_by_message
564 .get(&id)
565 .into_iter()
566 .flat_map(|context| {
567 context
568 .iter()
569 .filter_map(|context_id| self.context.get(&context_id))
570 })
571 }
572
573 /// Returns whether all of the tool uses have finished running.
574 pub fn all_tools_finished(&self) -> bool {
575 // If the only pending tool uses left are the ones with errors, then
576 // that means that we've finished running all of the pending tools.
577 self.tool_use
578 .pending_tool_uses()
579 .iter()
580 .all(|tool_use| tool_use.status.is_error())
581 }
582
583 pub fn tool_uses_for_message(&self, id: MessageId, cx: &App) -> Vec<ToolUse> {
584 self.tool_use.tool_uses_for_message(id, cx)
585 }
586
587 pub fn tool_results_for_message(&self, id: MessageId) -> Vec<&LanguageModelToolResult> {
588 self.tool_use.tool_results_for_message(id)
589 }
590
591 pub fn tool_result(&self, id: &LanguageModelToolUseId) -> Option<&LanguageModelToolResult> {
592 self.tool_use.tool_result(id)
593 }
594
595 pub fn message_has_tool_results(&self, message_id: MessageId) -> bool {
596 self.tool_use.message_has_tool_results(message_id)
597 }
598
599 pub fn insert_user_message(
600 &mut self,
601 text: impl Into<String>,
602 context: Vec<AssistantContext>,
603 git_checkpoint: Option<GitStoreCheckpoint>,
604 cx: &mut Context<Self>,
605 ) -> MessageId {
606 let text = text.into();
607
608 let message_id = self.insert_message(Role::User, vec![MessageSegment::Text(text)], cx);
609
610 // Filter out contexts that have already been included in previous messages
611 let new_context: Vec<_> = context
612 .into_iter()
613 .filter(|ctx| !self.context.contains_key(&ctx.id()))
614 .collect();
615
616 if !new_context.is_empty() {
617 if let Some(context_string) = format_context_as_string(new_context.iter(), cx) {
618 if let Some(message) = self.messages.iter_mut().find(|m| m.id == message_id) {
619 message.context = context_string;
620 }
621 }
622
623 self.action_log.update(cx, |log, cx| {
624 // Track all buffers added as context
625 for ctx in &new_context {
626 match ctx {
627 AssistantContext::File(file_ctx) => {
628 log.buffer_added_as_context(file_ctx.context_buffer.buffer.clone(), cx);
629 }
630 AssistantContext::Directory(dir_ctx) => {
631 for context_buffer in &dir_ctx.context_buffers {
632 log.buffer_added_as_context(context_buffer.buffer.clone(), cx);
633 }
634 }
635 AssistantContext::Symbol(symbol_ctx) => {
636 log.buffer_added_as_context(
637 symbol_ctx.context_symbol.buffer.clone(),
638 cx,
639 );
640 }
641 AssistantContext::FetchedUrl(_) | AssistantContext::Thread(_) => {}
642 }
643 }
644 });
645 }
646
647 let context_ids = new_context
648 .iter()
649 .map(|context| context.id())
650 .collect::<Vec<_>>();
651 self.context.extend(
652 new_context
653 .into_iter()
654 .map(|context| (context.id(), context)),
655 );
656 self.context_by_message.insert(message_id, context_ids);
657
658 if let Some(git_checkpoint) = git_checkpoint {
659 self.pending_checkpoint = Some(ThreadCheckpoint {
660 message_id,
661 git_checkpoint,
662 });
663 }
664 message_id
665 }
666
667 pub fn insert_message(
668 &mut self,
669 role: Role,
670 segments: Vec<MessageSegment>,
671 cx: &mut Context<Self>,
672 ) -> MessageId {
673 let id = self.next_message_id.post_inc();
674 self.messages.push(Message {
675 id,
676 role,
677 segments,
678 context: String::new(),
679 });
680 self.touch_updated_at();
681 cx.emit(ThreadEvent::MessageAdded(id));
682 id
683 }
684
685 pub fn edit_message(
686 &mut self,
687 id: MessageId,
688 new_role: Role,
689 new_segments: Vec<MessageSegment>,
690 cx: &mut Context<Self>,
691 ) -> bool {
692 let Some(message) = self.messages.iter_mut().find(|message| message.id == id) else {
693 return false;
694 };
695 message.role = new_role;
696 message.segments = new_segments;
697 self.touch_updated_at();
698 cx.emit(ThreadEvent::MessageEdited(id));
699 true
700 }
701
702 pub fn delete_message(&mut self, id: MessageId, cx: &mut Context<Self>) -> bool {
703 let Some(index) = self.messages.iter().position(|message| message.id == id) else {
704 return false;
705 };
706 self.messages.remove(index);
707 self.context_by_message.remove(&id);
708 self.touch_updated_at();
709 cx.emit(ThreadEvent::MessageDeleted(id));
710 true
711 }
712
713 /// Returns the representation of this [`Thread`] in a textual form.
714 ///
715 /// This is the representation we use when attaching a thread as context to another thread.
716 pub fn text(&self) -> String {
717 let mut text = String::new();
718
719 for message in &self.messages {
720 text.push_str(match message.role {
721 language_model::Role::User => "User:",
722 language_model::Role::Assistant => "Assistant:",
723 language_model::Role::System => "System:",
724 });
725 text.push('\n');
726
727 for segment in &message.segments {
728 match segment {
729 MessageSegment::Text(content) => text.push_str(content),
730 MessageSegment::Thinking(content) => {
731 text.push_str(&format!("<think>{}</think>", content))
732 }
733 }
734 }
735 text.push('\n');
736 }
737
738 text
739 }
740
741 /// Serializes this thread into a format for storage or telemetry.
742 pub fn serialize(&self, cx: &mut Context<Self>) -> Task<Result<SerializedThread>> {
743 let initial_project_snapshot = self.initial_project_snapshot.clone();
744 cx.spawn(async move |this, cx| {
745 let initial_project_snapshot = initial_project_snapshot.await;
746 this.read_with(cx, |this, cx| SerializedThread {
747 version: SerializedThread::VERSION.to_string(),
748 summary: this.summary_or_default(),
749 updated_at: this.updated_at(),
750 messages: this
751 .messages()
752 .map(|message| SerializedMessage {
753 id: message.id,
754 role: message.role,
755 segments: message
756 .segments
757 .iter()
758 .map(|segment| match segment {
759 MessageSegment::Text(text) => {
760 SerializedMessageSegment::Text { text: text.clone() }
761 }
762 MessageSegment::Thinking(text) => {
763 SerializedMessageSegment::Thinking { text: text.clone() }
764 }
765 })
766 .collect(),
767 tool_uses: this
768 .tool_uses_for_message(message.id, cx)
769 .into_iter()
770 .map(|tool_use| SerializedToolUse {
771 id: tool_use.id,
772 name: tool_use.name,
773 input: tool_use.input,
774 })
775 .collect(),
776 tool_results: this
777 .tool_results_for_message(message.id)
778 .into_iter()
779 .map(|tool_result| SerializedToolResult {
780 tool_use_id: tool_result.tool_use_id.clone(),
781 is_error: tool_result.is_error,
782 content: tool_result.content.clone(),
783 })
784 .collect(),
785 context: message.context.clone(),
786 })
787 .collect(),
788 initial_project_snapshot,
789 cumulative_token_usage: this.cumulative_token_usage.clone(),
790 detailed_summary_state: this.detailed_summary_state.clone(),
791 })
792 })
793 }
794
795 pub fn set_system_prompt_context(&mut self, context: AssistantSystemPromptContext) {
796 self.system_prompt_context = Some(context);
797 }
798
799 pub fn system_prompt_context(&self) -> &Option<AssistantSystemPromptContext> {
800 &self.system_prompt_context
801 }
802
803 pub fn load_system_prompt_context(
804 &self,
805 cx: &App,
806 ) -> Task<(AssistantSystemPromptContext, Option<ThreadError>)> {
807 let project = self.project.read(cx);
808 let tasks = project
809 .visible_worktrees(cx)
810 .map(|worktree| {
811 Self::load_worktree_info_for_system_prompt(
812 project.fs().clone(),
813 worktree.read(cx),
814 cx,
815 )
816 })
817 .collect::<Vec<_>>();
818
819 cx.spawn(async |_cx| {
820 let results = futures::future::join_all(tasks).await;
821 let mut first_err = None;
822 let worktrees = results
823 .into_iter()
824 .map(|(worktree, err)| {
825 if first_err.is_none() && err.is_some() {
826 first_err = err;
827 }
828 worktree
829 })
830 .collect::<Vec<_>>();
831 (AssistantSystemPromptContext::new(worktrees), first_err)
832 })
833 }
834
835 fn load_worktree_info_for_system_prompt(
836 fs: Arc<dyn Fs>,
837 worktree: &Worktree,
838 cx: &App,
839 ) -> Task<(WorktreeInfoForSystemPrompt, Option<ThreadError>)> {
840 let root_name = worktree.root_name().into();
841 let abs_path = worktree.abs_path();
842
843 // Note that Cline supports `.clinerules` being a directory, but that is not currently
844 // supported. This doesn't seem to occur often in GitHub repositories.
845 const RULES_FILE_NAMES: [&'static str; 6] = [
846 ".rules",
847 ".cursorrules",
848 ".windsurfrules",
849 ".clinerules",
850 ".github/copilot-instructions.md",
851 "CLAUDE.md",
852 ];
853 let selected_rules_file = RULES_FILE_NAMES
854 .into_iter()
855 .filter_map(|name| {
856 worktree
857 .entry_for_path(name)
858 .filter(|entry| entry.is_file())
859 .map(|entry| (entry.path.clone(), worktree.absolutize(&entry.path)))
860 })
861 .next();
862
863 if let Some((rel_rules_path, abs_rules_path)) = selected_rules_file {
864 cx.spawn(async move |_| {
865 let rules_file_result = maybe!(async move {
866 let abs_rules_path = abs_rules_path?;
867 let text = fs.load(&abs_rules_path).await.with_context(|| {
868 format!("Failed to load assistant rules file {:?}", abs_rules_path)
869 })?;
870 anyhow::Ok(RulesFile {
871 rel_path: rel_rules_path,
872 abs_path: abs_rules_path.into(),
873 text: text.trim().to_string(),
874 })
875 })
876 .await;
877 let (rules_file, rules_file_error) = match rules_file_result {
878 Ok(rules_file) => (Some(rules_file), None),
879 Err(err) => (
880 None,
881 Some(ThreadError::Message {
882 header: "Error loading rules file".into(),
883 message: format!("{err}").into(),
884 }),
885 ),
886 };
887 let worktree_info = WorktreeInfoForSystemPrompt {
888 root_name,
889 abs_path,
890 rules_file,
891 };
892 (worktree_info, rules_file_error)
893 })
894 } else {
895 Task::ready((
896 WorktreeInfoForSystemPrompt {
897 root_name,
898 abs_path,
899 rules_file: None,
900 },
901 None,
902 ))
903 }
904 }
905
906 pub fn send_to_model(
907 &mut self,
908 model: Arc<dyn LanguageModel>,
909 request_kind: RequestKind,
910 cx: &mut Context<Self>,
911 ) {
912 let mut request = self.to_completion_request(request_kind, cx);
913 if model.supports_tools() {
914 request.tools = {
915 let mut tools = Vec::new();
916 tools.extend(self.tools().enabled_tools(cx).into_iter().map(|tool| {
917 LanguageModelRequestTool {
918 name: tool.name(),
919 description: tool.description(),
920 input_schema: tool.input_schema(model.tool_input_format()),
921 }
922 }));
923
924 tools
925 };
926 }
927
928 self.stream_completion(request, model, cx);
929 }
930
931 pub fn used_tools_since_last_user_message(&self) -> bool {
932 for message in self.messages.iter().rev() {
933 if self.tool_use.message_has_tool_results(message.id) {
934 return true;
935 } else if message.role == Role::User {
936 return false;
937 }
938 }
939
940 false
941 }
942
943 pub fn to_completion_request(
944 &self,
945 request_kind: RequestKind,
946 cx: &App,
947 ) -> LanguageModelRequest {
948 let mut request = LanguageModelRequest {
949 messages: vec![],
950 tools: Vec::new(),
951 stop: Vec::new(),
952 temperature: None,
953 };
954
955 if let Some(system_prompt_context) = self.system_prompt_context.as_ref() {
956 if let Some(system_prompt) = self
957 .prompt_builder
958 .generate_assistant_system_prompt(system_prompt_context)
959 .context("failed to generate assistant system prompt")
960 .log_err()
961 {
962 request.messages.push(LanguageModelRequestMessage {
963 role: Role::System,
964 content: vec![MessageContent::Text(system_prompt)],
965 cache: true,
966 });
967 }
968 } else {
969 log::error!("system_prompt_context not set.")
970 }
971
972 for message in &self.messages {
973 let mut request_message = LanguageModelRequestMessage {
974 role: message.role,
975 content: Vec::new(),
976 cache: false,
977 };
978
979 match request_kind {
980 RequestKind::Chat => {
981 self.tool_use
982 .attach_tool_results(message.id, &mut request_message);
983 }
984 RequestKind::Summarize => {
985 // We don't care about tool use during summarization.
986 if self.tool_use.message_has_tool_results(message.id) {
987 continue;
988 }
989 }
990 }
991
992 if !message.segments.is_empty() {
993 request_message
994 .content
995 .push(MessageContent::Text(message.to_string()));
996 }
997
998 match request_kind {
999 RequestKind::Chat => {
1000 self.tool_use
1001 .attach_tool_uses(message.id, &mut request_message);
1002 }
1003 RequestKind::Summarize => {
1004 // We don't care about tool use during summarization.
1005 }
1006 };
1007
1008 request.messages.push(request_message);
1009 }
1010
1011 // https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
1012 if let Some(last) = request.messages.last_mut() {
1013 last.cache = true;
1014 }
1015
1016 self.attached_tracked_files_state(&mut request.messages, cx);
1017
1018 request
1019 }
1020
1021 fn attached_tracked_files_state(
1022 &self,
1023 messages: &mut Vec<LanguageModelRequestMessage>,
1024 cx: &App,
1025 ) {
1026 const STALE_FILES_HEADER: &str = "These files changed since last read:";
1027
1028 let mut stale_message = String::new();
1029
1030 let action_log = self.action_log.read(cx);
1031
1032 for stale_file in action_log.stale_buffers(cx) {
1033 let Some(file) = stale_file.read(cx).file() else {
1034 continue;
1035 };
1036
1037 if stale_message.is_empty() {
1038 write!(&mut stale_message, "{}\n", STALE_FILES_HEADER).ok();
1039 }
1040
1041 writeln!(&mut stale_message, "- {}", file.path().display()).ok();
1042 }
1043
1044 let mut content = Vec::with_capacity(2);
1045
1046 if !stale_message.is_empty() {
1047 content.push(stale_message.into());
1048 }
1049
1050 if action_log.has_edited_files_since_project_diagnostics_check() {
1051 content.push(
1052 "\n\nWhen you're done making changes, make sure to check project diagnostics \
1053 and fix all errors AND warnings you introduced! \
1054 DO NOT mention you're going to do this until you're done."
1055 .into(),
1056 );
1057 }
1058
1059 if !content.is_empty() {
1060 let context_message = LanguageModelRequestMessage {
1061 role: Role::User,
1062 content,
1063 cache: false,
1064 };
1065
1066 messages.push(context_message);
1067 }
1068 }
1069
1070 pub fn stream_completion(
1071 &mut self,
1072 request: LanguageModelRequest,
1073 model: Arc<dyn LanguageModel>,
1074 cx: &mut Context<Self>,
1075 ) {
1076 let pending_completion_id = post_inc(&mut self.completion_count);
1077
1078 let task = cx.spawn(async move |thread, cx| {
1079 let stream = model.stream_completion(request, &cx);
1080 let initial_token_usage =
1081 thread.read_with(cx, |thread, _cx| thread.cumulative_token_usage.clone());
1082 let stream_completion = async {
1083 let mut events = stream.await?;
1084 let mut stop_reason = StopReason::EndTurn;
1085 let mut current_token_usage = TokenUsage::default();
1086
1087 while let Some(event) = events.next().await {
1088 let event = event?;
1089
1090 thread.update(cx, |thread, cx| {
1091 match event {
1092 LanguageModelCompletionEvent::StartMessage { .. } => {
1093 thread.insert_message(
1094 Role::Assistant,
1095 vec![MessageSegment::Text(String::new())],
1096 cx,
1097 );
1098 }
1099 LanguageModelCompletionEvent::Stop(reason) => {
1100 stop_reason = reason;
1101 }
1102 LanguageModelCompletionEvent::UsageUpdate(token_usage) => {
1103 thread.cumulative_token_usage =
1104 thread.cumulative_token_usage.clone() + token_usage.clone()
1105 - current_token_usage.clone();
1106 current_token_usage = token_usage;
1107 }
1108 LanguageModelCompletionEvent::Text(chunk) => {
1109 if let Some(last_message) = thread.messages.last_mut() {
1110 if last_message.role == Role::Assistant {
1111 last_message.push_text(&chunk);
1112 cx.emit(ThreadEvent::StreamedAssistantText(
1113 last_message.id,
1114 chunk,
1115 ));
1116 } else {
1117 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1118 // of a new Assistant response.
1119 //
1120 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1121 // will result in duplicating the text of the chunk in the rendered Markdown.
1122 thread.insert_message(
1123 Role::Assistant,
1124 vec![MessageSegment::Text(chunk.to_string())],
1125 cx,
1126 );
1127 };
1128 }
1129 }
1130 LanguageModelCompletionEvent::Thinking(chunk) => {
1131 if let Some(last_message) = thread.messages.last_mut() {
1132 if last_message.role == Role::Assistant {
1133 last_message.push_thinking(&chunk);
1134 cx.emit(ThreadEvent::StreamedAssistantThinking(
1135 last_message.id,
1136 chunk,
1137 ));
1138 } else {
1139 // If we won't have an Assistant message yet, assume this chunk marks the beginning
1140 // of a new Assistant response.
1141 //
1142 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
1143 // will result in duplicating the text of the chunk in the rendered Markdown.
1144 thread.insert_message(
1145 Role::Assistant,
1146 vec![MessageSegment::Thinking(chunk.to_string())],
1147 cx,
1148 );
1149 };
1150 }
1151 }
1152 LanguageModelCompletionEvent::ToolUse(tool_use) => {
1153 let last_assistant_message_id = thread
1154 .messages
1155 .iter_mut()
1156 .rfind(|message| message.role == Role::Assistant)
1157 .map(|message| message.id)
1158 .unwrap_or_else(|| {
1159 thread.insert_message(Role::Assistant, vec![], cx)
1160 });
1161
1162 thread.tool_use.request_tool_use(
1163 last_assistant_message_id,
1164 tool_use,
1165 cx,
1166 );
1167 }
1168 }
1169
1170 thread.touch_updated_at();
1171 cx.emit(ThreadEvent::StreamedCompletion);
1172 cx.notify();
1173 })?;
1174
1175 smol::future::yield_now().await;
1176 }
1177
1178 thread.update(cx, |thread, cx| {
1179 thread
1180 .pending_completions
1181 .retain(|completion| completion.id != pending_completion_id);
1182
1183 if thread.summary.is_none() && thread.messages.len() >= 2 {
1184 thread.summarize(cx);
1185 }
1186 })?;
1187
1188 anyhow::Ok(stop_reason)
1189 };
1190
1191 let result = stream_completion.await;
1192
1193 thread
1194 .update(cx, |thread, cx| {
1195 thread.finalize_pending_checkpoint(cx);
1196 match result.as_ref() {
1197 Ok(stop_reason) => match stop_reason {
1198 StopReason::ToolUse => {
1199 cx.emit(ThreadEvent::UsePendingTools);
1200 }
1201 StopReason::EndTurn => {}
1202 StopReason::MaxTokens => {}
1203 },
1204 Err(error) => {
1205 if error.is::<PaymentRequiredError>() {
1206 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
1207 } else if error.is::<MaxMonthlySpendReachedError>() {
1208 cx.emit(ThreadEvent::ShowError(
1209 ThreadError::MaxMonthlySpendReached,
1210 ));
1211 } else {
1212 let error_message = error
1213 .chain()
1214 .map(|err| err.to_string())
1215 .collect::<Vec<_>>()
1216 .join("\n");
1217 cx.emit(ThreadEvent::ShowError(ThreadError::Message {
1218 header: "Error interacting with language model".into(),
1219 message: SharedString::from(error_message.clone()),
1220 }));
1221 }
1222
1223 thread.cancel_last_completion(cx);
1224 }
1225 }
1226 cx.emit(ThreadEvent::DoneStreaming);
1227
1228 if let Ok(initial_usage) = initial_token_usage {
1229 let usage = thread.cumulative_token_usage.clone() - initial_usage;
1230
1231 telemetry::event!(
1232 "Assistant Thread Completion",
1233 thread_id = thread.id().to_string(),
1234 model = model.telemetry_id(),
1235 model_provider = model.provider_id().to_string(),
1236 input_tokens = usage.input_tokens,
1237 output_tokens = usage.output_tokens,
1238 cache_creation_input_tokens = usage.cache_creation_input_tokens,
1239 cache_read_input_tokens = usage.cache_read_input_tokens,
1240 );
1241 }
1242 })
1243 .ok();
1244 });
1245
1246 self.pending_completions.push(PendingCompletion {
1247 id: pending_completion_id,
1248 _task: task,
1249 });
1250 }
1251
1252 pub fn summarize(&mut self, cx: &mut Context<Self>) {
1253 let Some(provider) = LanguageModelRegistry::read_global(cx).active_provider() else {
1254 return;
1255 };
1256 let Some(model) = LanguageModelRegistry::read_global(cx).active_model() else {
1257 return;
1258 };
1259
1260 if !provider.is_authenticated(cx) {
1261 return;
1262 }
1263
1264 let mut request = self.to_completion_request(RequestKind::Summarize, cx);
1265 request.messages.push(LanguageModelRequestMessage {
1266 role: Role::User,
1267 content: vec![
1268 "Generate a concise 3-7 word title for this conversation, omitting punctuation. \
1269 Go straight to the title, without any preamble and prefix like `Here's a concise suggestion:...` or `Title:`. \
1270 If the conversation is about a specific subject, include it in the title. \
1271 Be descriptive. DO NOT speak in the first person."
1272 .into(),
1273 ],
1274 cache: false,
1275 });
1276
1277 self.pending_summary = cx.spawn(async move |this, cx| {
1278 async move {
1279 let stream = model.stream_completion_text(request, &cx);
1280 let mut messages = stream.await?;
1281
1282 let mut new_summary = String::new();
1283 while let Some(message) = messages.stream.next().await {
1284 let text = message?;
1285 let mut lines = text.lines();
1286 new_summary.extend(lines.next());
1287
1288 // Stop if the LLM generated multiple lines.
1289 if lines.next().is_some() {
1290 break;
1291 }
1292 }
1293
1294 this.update(cx, |this, cx| {
1295 if !new_summary.is_empty() {
1296 this.summary = Some(new_summary.into());
1297 }
1298
1299 cx.emit(ThreadEvent::SummaryChanged);
1300 })?;
1301
1302 anyhow::Ok(())
1303 }
1304 .log_err()
1305 .await
1306 });
1307 }
1308
1309 pub fn generate_detailed_summary(&mut self, cx: &mut Context<Self>) -> Option<Task<()>> {
1310 let last_message_id = self.messages.last().map(|message| message.id)?;
1311
1312 match &self.detailed_summary_state {
1313 DetailedSummaryState::Generating { message_id, .. }
1314 | DetailedSummaryState::Generated { message_id, .. }
1315 if *message_id == last_message_id =>
1316 {
1317 // Already up-to-date
1318 return None;
1319 }
1320 _ => {}
1321 }
1322
1323 let provider = LanguageModelRegistry::read_global(cx).active_provider()?;
1324 let model = LanguageModelRegistry::read_global(cx).active_model()?;
1325
1326 if !provider.is_authenticated(cx) {
1327 return None;
1328 }
1329
1330 let mut request = self.to_completion_request(RequestKind::Summarize, cx);
1331
1332 request.messages.push(LanguageModelRequestMessage {
1333 role: Role::User,
1334 content: vec![
1335 "Generate a detailed summary of this conversation. Include:\n\
1336 1. A brief overview of what was discussed\n\
1337 2. Key facts or information discovered\n\
1338 3. Outcomes or conclusions reached\n\
1339 4. Any action items or next steps if any\n\
1340 Format it in Markdown with headings and bullet points."
1341 .into(),
1342 ],
1343 cache: false,
1344 });
1345
1346 let task = cx.spawn(async move |thread, cx| {
1347 let stream = model.stream_completion_text(request, &cx);
1348 let Some(mut messages) = stream.await.log_err() else {
1349 thread
1350 .update(cx, |this, _cx| {
1351 this.detailed_summary_state = DetailedSummaryState::NotGenerated;
1352 })
1353 .log_err();
1354
1355 return;
1356 };
1357
1358 let mut new_detailed_summary = String::new();
1359
1360 while let Some(chunk) = messages.stream.next().await {
1361 if let Some(chunk) = chunk.log_err() {
1362 new_detailed_summary.push_str(&chunk);
1363 }
1364 }
1365
1366 thread
1367 .update(cx, |this, _cx| {
1368 this.detailed_summary_state = DetailedSummaryState::Generated {
1369 text: new_detailed_summary.into(),
1370 message_id: last_message_id,
1371 };
1372 })
1373 .log_err();
1374 });
1375
1376 self.detailed_summary_state = DetailedSummaryState::Generating {
1377 message_id: last_message_id,
1378 };
1379
1380 Some(task)
1381 }
1382
1383 pub fn is_generating_detailed_summary(&self) -> bool {
1384 matches!(
1385 self.detailed_summary_state,
1386 DetailedSummaryState::Generating { .. }
1387 )
1388 }
1389
1390 pub fn use_pending_tools(
1391 &mut self,
1392 cx: &mut Context<Self>,
1393 ) -> impl IntoIterator<Item = PendingToolUse> + use<> {
1394 let request = self.to_completion_request(RequestKind::Chat, cx);
1395 let messages = Arc::new(request.messages);
1396 let pending_tool_uses = self
1397 .tool_use
1398 .pending_tool_uses()
1399 .into_iter()
1400 .filter(|tool_use| tool_use.status.is_idle())
1401 .cloned()
1402 .collect::<Vec<_>>();
1403
1404 for tool_use in pending_tool_uses.iter() {
1405 if let Some(tool) = self.tools.tool(&tool_use.name, cx) {
1406 if tool.needs_confirmation()
1407 && !AssistantSettings::get_global(cx).always_allow_tool_actions
1408 {
1409 self.tool_use.confirm_tool_use(
1410 tool_use.id.clone(),
1411 tool_use.ui_text.clone(),
1412 tool_use.input.clone(),
1413 messages.clone(),
1414 tool,
1415 );
1416 cx.emit(ThreadEvent::ToolConfirmationNeeded);
1417 } else {
1418 self.run_tool(
1419 tool_use.id.clone(),
1420 tool_use.ui_text.clone(),
1421 tool_use.input.clone(),
1422 &messages,
1423 tool,
1424 cx,
1425 );
1426 }
1427 }
1428 }
1429
1430 pending_tool_uses
1431 }
1432
1433 pub fn run_tool(
1434 &mut self,
1435 tool_use_id: LanguageModelToolUseId,
1436 ui_text: impl Into<SharedString>,
1437 input: serde_json::Value,
1438 messages: &[LanguageModelRequestMessage],
1439 tool: Arc<dyn Tool>,
1440 cx: &mut Context<Thread>,
1441 ) {
1442 let task = self.spawn_tool_use(tool_use_id.clone(), messages, input, tool, cx);
1443 self.tool_use
1444 .run_pending_tool(tool_use_id, ui_text.into(), task);
1445 }
1446
1447 fn spawn_tool_use(
1448 &mut self,
1449 tool_use_id: LanguageModelToolUseId,
1450 messages: &[LanguageModelRequestMessage],
1451 input: serde_json::Value,
1452 tool: Arc<dyn Tool>,
1453 cx: &mut Context<Thread>,
1454 ) -> Task<()> {
1455 let tool_name: Arc<str> = tool.name().into();
1456
1457 let run_tool = if self.tools.is_disabled(&tool.source(), &tool_name) {
1458 Task::ready(Err(anyhow!("tool is disabled: {tool_name}")))
1459 } else {
1460 tool.run(
1461 input,
1462 messages,
1463 self.project.clone(),
1464 self.action_log.clone(),
1465 cx,
1466 )
1467 };
1468
1469 cx.spawn({
1470 async move |thread: WeakEntity<Thread>, cx| {
1471 let output = run_tool.await;
1472
1473 thread
1474 .update(cx, |thread, cx| {
1475 let pending_tool_use = thread.tool_use.insert_tool_output(
1476 tool_use_id.clone(),
1477 tool_name,
1478 output,
1479 );
1480
1481 cx.emit(ThreadEvent::ToolFinished {
1482 tool_use_id,
1483 pending_tool_use,
1484 canceled: false,
1485 });
1486 })
1487 .ok();
1488 }
1489 })
1490 }
1491
1492 pub fn attach_tool_results(&mut self, cx: &mut Context<Self>) {
1493 // Insert a user message to contain the tool results.
1494 self.insert_user_message(
1495 // TODO: Sending up a user message without any content results in the model sending back
1496 // responses that also don't have any content. We currently don't handle this case well,
1497 // so for now we provide some text to keep the model on track.
1498 "Here are the tool results.",
1499 Vec::new(),
1500 None,
1501 cx,
1502 );
1503 }
1504
1505 /// Cancels the last pending completion, if there are any pending.
1506 ///
1507 /// Returns whether a completion was canceled.
1508 pub fn cancel_last_completion(&mut self, cx: &mut Context<Self>) -> bool {
1509 let canceled = if self.pending_completions.pop().is_some() {
1510 true
1511 } else {
1512 let mut canceled = false;
1513 for pending_tool_use in self.tool_use.cancel_pending() {
1514 canceled = true;
1515 cx.emit(ThreadEvent::ToolFinished {
1516 tool_use_id: pending_tool_use.id.clone(),
1517 pending_tool_use: Some(pending_tool_use),
1518 canceled: true,
1519 });
1520 }
1521 canceled
1522 };
1523 self.finalize_pending_checkpoint(cx);
1524 canceled
1525 }
1526
1527 /// Returns the feedback given to the thread, if any.
1528 pub fn feedback(&self) -> Option<ThreadFeedback> {
1529 self.feedback
1530 }
1531
1532 /// Reports feedback about the thread and stores it in our telemetry backend.
1533 pub fn report_feedback(
1534 &mut self,
1535 feedback: ThreadFeedback,
1536 cx: &mut Context<Self>,
1537 ) -> Task<Result<()>> {
1538 let final_project_snapshot = Self::project_snapshot(self.project.clone(), cx);
1539 let serialized_thread = self.serialize(cx);
1540 let thread_id = self.id().clone();
1541 let client = self.project.read(cx).client();
1542 self.feedback = Some(feedback);
1543 cx.notify();
1544
1545 cx.background_spawn(async move {
1546 let final_project_snapshot = final_project_snapshot.await;
1547 let serialized_thread = serialized_thread.await?;
1548 let thread_data =
1549 serde_json::to_value(serialized_thread).unwrap_or_else(|_| serde_json::Value::Null);
1550
1551 let rating = match feedback {
1552 ThreadFeedback::Positive => "positive",
1553 ThreadFeedback::Negative => "negative",
1554 };
1555 telemetry::event!(
1556 "Assistant Thread Rated",
1557 rating,
1558 thread_id,
1559 thread_data,
1560 final_project_snapshot
1561 );
1562 client.telemetry().flush_events();
1563
1564 Ok(())
1565 })
1566 }
1567
1568 /// Create a snapshot of the current project state including git information and unsaved buffers.
1569 fn project_snapshot(
1570 project: Entity<Project>,
1571 cx: &mut Context<Self>,
1572 ) -> Task<Arc<ProjectSnapshot>> {
1573 let git_store = project.read(cx).git_store().clone();
1574 let worktree_snapshots: Vec<_> = project
1575 .read(cx)
1576 .visible_worktrees(cx)
1577 .map(|worktree| Self::worktree_snapshot(worktree, git_store.clone(), cx))
1578 .collect();
1579
1580 cx.spawn(async move |_, cx| {
1581 let worktree_snapshots = futures::future::join_all(worktree_snapshots).await;
1582
1583 let mut unsaved_buffers = Vec::new();
1584 cx.update(|app_cx| {
1585 let buffer_store = project.read(app_cx).buffer_store();
1586 for buffer_handle in buffer_store.read(app_cx).buffers() {
1587 let buffer = buffer_handle.read(app_cx);
1588 if buffer.is_dirty() {
1589 if let Some(file) = buffer.file() {
1590 let path = file.path().to_string_lossy().to_string();
1591 unsaved_buffers.push(path);
1592 }
1593 }
1594 }
1595 })
1596 .ok();
1597
1598 Arc::new(ProjectSnapshot {
1599 worktree_snapshots,
1600 unsaved_buffer_paths: unsaved_buffers,
1601 timestamp: Utc::now(),
1602 })
1603 })
1604 }
1605
1606 fn worktree_snapshot(
1607 worktree: Entity<project::Worktree>,
1608 git_store: Entity<GitStore>,
1609 cx: &App,
1610 ) -> Task<WorktreeSnapshot> {
1611 cx.spawn(async move |cx| {
1612 // Get worktree path and snapshot
1613 let worktree_info = cx.update(|app_cx| {
1614 let worktree = worktree.read(app_cx);
1615 let path = worktree.abs_path().to_string_lossy().to_string();
1616 let snapshot = worktree.snapshot();
1617 (path, snapshot)
1618 });
1619
1620 let Ok((worktree_path, _snapshot)) = worktree_info else {
1621 return WorktreeSnapshot {
1622 worktree_path: String::new(),
1623 git_state: None,
1624 };
1625 };
1626
1627 let git_state = git_store
1628 .update(cx, |git_store, cx| {
1629 git_store
1630 .repositories()
1631 .values()
1632 .find(|repo| {
1633 repo.read(cx)
1634 .abs_path_to_repo_path(&worktree.read(cx).abs_path())
1635 .is_some()
1636 })
1637 .cloned()
1638 })
1639 .ok()
1640 .flatten()
1641 .map(|repo| {
1642 repo.read_with(cx, |repo, _| {
1643 let current_branch =
1644 repo.branch.as_ref().map(|branch| branch.name.to_string());
1645 repo.send_job(|state, _| async move {
1646 let RepositoryState::Local { backend, .. } = state else {
1647 return GitState {
1648 remote_url: None,
1649 head_sha: None,
1650 current_branch,
1651 diff: None,
1652 };
1653 };
1654
1655 let remote_url = backend.remote_url("origin");
1656 let head_sha = backend.head_sha();
1657 let diff = backend.diff(DiffType::HeadToWorktree).await.ok();
1658
1659 GitState {
1660 remote_url,
1661 head_sha,
1662 current_branch,
1663 diff,
1664 }
1665 })
1666 })
1667 });
1668
1669 let git_state = match git_state {
1670 Some(git_state) => match git_state.ok() {
1671 Some(git_state) => git_state.await.ok(),
1672 None => None,
1673 },
1674 None => None,
1675 };
1676
1677 WorktreeSnapshot {
1678 worktree_path,
1679 git_state,
1680 }
1681 })
1682 }
1683
1684 pub fn to_markdown(&self, cx: &App) -> Result<String> {
1685 let mut markdown = Vec::new();
1686
1687 if let Some(summary) = self.summary() {
1688 writeln!(markdown, "# {summary}\n")?;
1689 };
1690
1691 for message in self.messages() {
1692 writeln!(
1693 markdown,
1694 "## {role}\n",
1695 role = match message.role {
1696 Role::User => "User",
1697 Role::Assistant => "Assistant",
1698 Role::System => "System",
1699 }
1700 )?;
1701
1702 if !message.context.is_empty() {
1703 writeln!(markdown, "{}", message.context)?;
1704 }
1705
1706 for segment in &message.segments {
1707 match segment {
1708 MessageSegment::Text(text) => writeln!(markdown, "{}\n", text)?,
1709 MessageSegment::Thinking(text) => {
1710 writeln!(markdown, "<think>{}</think>\n", text)?
1711 }
1712 }
1713 }
1714
1715 for tool_use in self.tool_uses_for_message(message.id, cx) {
1716 writeln!(
1717 markdown,
1718 "**Use Tool: {} ({})**",
1719 tool_use.name, tool_use.id
1720 )?;
1721 writeln!(markdown, "```json")?;
1722 writeln!(
1723 markdown,
1724 "{}",
1725 serde_json::to_string_pretty(&tool_use.input)?
1726 )?;
1727 writeln!(markdown, "```")?;
1728 }
1729
1730 for tool_result in self.tool_results_for_message(message.id) {
1731 write!(markdown, "**Tool Results: {}", tool_result.tool_use_id)?;
1732 if tool_result.is_error {
1733 write!(markdown, " (Error)")?;
1734 }
1735
1736 writeln!(markdown, "**\n")?;
1737 writeln!(markdown, "{}", tool_result.content)?;
1738 }
1739 }
1740
1741 Ok(String::from_utf8_lossy(&markdown).to_string())
1742 }
1743
1744 pub fn keep_edits_in_range(
1745 &mut self,
1746 buffer: Entity<language::Buffer>,
1747 buffer_range: Range<language::Anchor>,
1748 cx: &mut Context<Self>,
1749 ) {
1750 self.action_log.update(cx, |action_log, cx| {
1751 action_log.keep_edits_in_range(buffer, buffer_range, cx)
1752 });
1753 }
1754
1755 pub fn keep_all_edits(&mut self, cx: &mut Context<Self>) {
1756 self.action_log
1757 .update(cx, |action_log, cx| action_log.keep_all_edits(cx));
1758 }
1759
1760 pub fn action_log(&self) -> &Entity<ActionLog> {
1761 &self.action_log
1762 }
1763
1764 pub fn project(&self) -> &Entity<Project> {
1765 &self.project
1766 }
1767
1768 pub fn cumulative_token_usage(&self) -> TokenUsage {
1769 self.cumulative_token_usage.clone()
1770 }
1771
1772 pub fn total_token_usage(&self, cx: &App) -> TotalTokenUsage {
1773 let model_registry = LanguageModelRegistry::read_global(cx);
1774 let Some(model) = model_registry.active_model() else {
1775 return TotalTokenUsage::default();
1776 };
1777
1778 let max = model.max_token_count();
1779
1780 #[cfg(debug_assertions)]
1781 let warning_threshold: f32 = std::env::var("ZED_THREAD_WARNING_THRESHOLD")
1782 .unwrap_or("0.8".to_string())
1783 .parse()
1784 .unwrap();
1785 #[cfg(not(debug_assertions))]
1786 let warning_threshold: f32 = 0.8;
1787
1788 let total = self.cumulative_token_usage.total_tokens() as usize;
1789
1790 let ratio = if total >= max {
1791 TokenUsageRatio::Exceeded
1792 } else if total as f32 / max as f32 >= warning_threshold {
1793 TokenUsageRatio::Warning
1794 } else {
1795 TokenUsageRatio::Normal
1796 };
1797
1798 TotalTokenUsage { total, max, ratio }
1799 }
1800
1801 pub fn deny_tool_use(
1802 &mut self,
1803 tool_use_id: LanguageModelToolUseId,
1804 tool_name: Arc<str>,
1805 cx: &mut Context<Self>,
1806 ) {
1807 let err = Err(anyhow::anyhow!(
1808 "Permission to run tool action denied by user"
1809 ));
1810
1811 self.tool_use
1812 .insert_tool_output(tool_use_id.clone(), tool_name, err);
1813
1814 cx.emit(ThreadEvent::ToolFinished {
1815 tool_use_id,
1816 pending_tool_use: None,
1817 canceled: true,
1818 });
1819 }
1820}
1821
1822#[derive(Debug, Clone)]
1823pub enum ThreadError {
1824 PaymentRequired,
1825 MaxMonthlySpendReached,
1826 Message {
1827 header: SharedString,
1828 message: SharedString,
1829 },
1830}
1831
1832#[derive(Debug, Clone)]
1833pub enum ThreadEvent {
1834 ShowError(ThreadError),
1835 StreamedCompletion,
1836 StreamedAssistantText(MessageId, String),
1837 StreamedAssistantThinking(MessageId, String),
1838 DoneStreaming,
1839 MessageAdded(MessageId),
1840 MessageEdited(MessageId),
1841 MessageDeleted(MessageId),
1842 SummaryChanged,
1843 UsePendingTools,
1844 ToolFinished {
1845 #[allow(unused)]
1846 tool_use_id: LanguageModelToolUseId,
1847 /// The pending tool use that corresponds to this tool.
1848 pending_tool_use: Option<PendingToolUse>,
1849 /// Whether the tool was canceled by the user.
1850 canceled: bool,
1851 },
1852 CheckpointChanged,
1853 ToolConfirmationNeeded,
1854}
1855
1856impl EventEmitter<ThreadEvent> for Thread {}
1857
1858struct PendingCompletion {
1859 id: usize,
1860 _task: Task<()>,
1861}
1862
1863#[cfg(test)]
1864mod tests {
1865 use super::*;
1866 use crate::{ThreadStore, context_store::ContextStore, thread_store};
1867 use assistant_settings::AssistantSettings;
1868 use context_server::ContextServerSettings;
1869 use editor::EditorSettings;
1870 use gpui::TestAppContext;
1871 use project::{FakeFs, Project};
1872 use prompt_store::PromptBuilder;
1873 use serde_json::json;
1874 use settings::{Settings, SettingsStore};
1875 use std::sync::Arc;
1876 use theme::ThemeSettings;
1877 use util::path;
1878 use workspace::Workspace;
1879
1880 #[gpui::test]
1881 async fn test_message_with_context(cx: &mut TestAppContext) {
1882 init_test_settings(cx);
1883
1884 let project = create_test_project(
1885 cx,
1886 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
1887 )
1888 .await;
1889
1890 let (_workspace, _thread_store, thread, context_store) =
1891 setup_test_environment(cx, project.clone()).await;
1892
1893 add_file_to_context(&project, &context_store, "test/code.rs", cx)
1894 .await
1895 .unwrap();
1896
1897 let context =
1898 context_store.update(cx, |store, _| store.context().first().cloned().unwrap());
1899
1900 // Insert user message with context
1901 let message_id = thread.update(cx, |thread, cx| {
1902 thread.insert_user_message("Please explain this code", vec![context], None, cx)
1903 });
1904
1905 // Check content and context in message object
1906 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
1907
1908 // Use different path format strings based on platform for the test
1909 #[cfg(windows)]
1910 let path_part = r"test\code.rs";
1911 #[cfg(not(windows))]
1912 let path_part = "test/code.rs";
1913
1914 let expected_context = format!(
1915 r#"
1916<context>
1917The following items were attached by the user. You don't need to use other tools to read them.
1918
1919<files>
1920```rs {path_part}
1921fn main() {{
1922 println!("Hello, world!");
1923}}
1924```
1925</files>
1926</context>
1927"#
1928 );
1929
1930 assert_eq!(message.role, Role::User);
1931 assert_eq!(message.segments.len(), 1);
1932 assert_eq!(
1933 message.segments[0],
1934 MessageSegment::Text("Please explain this code".to_string())
1935 );
1936 assert_eq!(message.context, expected_context);
1937
1938 // Check message in request
1939 let request = thread.read_with(cx, |thread, cx| {
1940 thread.to_completion_request(RequestKind::Chat, cx)
1941 });
1942
1943 assert_eq!(request.messages.len(), 1);
1944 let expected_full_message = format!("{}Please explain this code", expected_context);
1945 assert_eq!(request.messages[0].string_contents(), expected_full_message);
1946 }
1947
1948 #[gpui::test]
1949 async fn test_only_include_new_contexts(cx: &mut TestAppContext) {
1950 init_test_settings(cx);
1951
1952 let project = create_test_project(
1953 cx,
1954 json!({
1955 "file1.rs": "fn function1() {}\n",
1956 "file2.rs": "fn function2() {}\n",
1957 "file3.rs": "fn function3() {}\n",
1958 }),
1959 )
1960 .await;
1961
1962 let (_, _thread_store, thread, context_store) =
1963 setup_test_environment(cx, project.clone()).await;
1964
1965 // Open files individually
1966 add_file_to_context(&project, &context_store, "test/file1.rs", cx)
1967 .await
1968 .unwrap();
1969 add_file_to_context(&project, &context_store, "test/file2.rs", cx)
1970 .await
1971 .unwrap();
1972 add_file_to_context(&project, &context_store, "test/file3.rs", cx)
1973 .await
1974 .unwrap();
1975
1976 // Get the context objects
1977 let contexts = context_store.update(cx, |store, _| store.context().clone());
1978 assert_eq!(contexts.len(), 3);
1979
1980 // First message with context 1
1981 let message1_id = thread.update(cx, |thread, cx| {
1982 thread.insert_user_message("Message 1", vec![contexts[0].clone()], None, cx)
1983 });
1984
1985 // Second message with contexts 1 and 2 (context 1 should be skipped as it's already included)
1986 let message2_id = thread.update(cx, |thread, cx| {
1987 thread.insert_user_message(
1988 "Message 2",
1989 vec![contexts[0].clone(), contexts[1].clone()],
1990 None,
1991 cx,
1992 )
1993 });
1994
1995 // Third message with all three contexts (contexts 1 and 2 should be skipped)
1996 let message3_id = thread.update(cx, |thread, cx| {
1997 thread.insert_user_message(
1998 "Message 3",
1999 vec![
2000 contexts[0].clone(),
2001 contexts[1].clone(),
2002 contexts[2].clone(),
2003 ],
2004 None,
2005 cx,
2006 )
2007 });
2008
2009 // Check what contexts are included in each message
2010 let (message1, message2, message3) = thread.read_with(cx, |thread, _| {
2011 (
2012 thread.message(message1_id).unwrap().clone(),
2013 thread.message(message2_id).unwrap().clone(),
2014 thread.message(message3_id).unwrap().clone(),
2015 )
2016 });
2017
2018 // First message should include context 1
2019 assert!(message1.context.contains("file1.rs"));
2020
2021 // Second message should include only context 2 (not 1)
2022 assert!(!message2.context.contains("file1.rs"));
2023 assert!(message2.context.contains("file2.rs"));
2024
2025 // Third message should include only context 3 (not 1 or 2)
2026 assert!(!message3.context.contains("file1.rs"));
2027 assert!(!message3.context.contains("file2.rs"));
2028 assert!(message3.context.contains("file3.rs"));
2029
2030 // Check entire request to make sure all contexts are properly included
2031 let request = thread.read_with(cx, |thread, cx| {
2032 thread.to_completion_request(RequestKind::Chat, cx)
2033 });
2034
2035 // The request should contain all 3 messages
2036 assert_eq!(request.messages.len(), 3);
2037
2038 // Check that the contexts are properly formatted in each message
2039 assert!(request.messages[0].string_contents().contains("file1.rs"));
2040 assert!(!request.messages[0].string_contents().contains("file2.rs"));
2041 assert!(!request.messages[0].string_contents().contains("file3.rs"));
2042
2043 assert!(!request.messages[1].string_contents().contains("file1.rs"));
2044 assert!(request.messages[1].string_contents().contains("file2.rs"));
2045 assert!(!request.messages[1].string_contents().contains("file3.rs"));
2046
2047 assert!(!request.messages[2].string_contents().contains("file1.rs"));
2048 assert!(!request.messages[2].string_contents().contains("file2.rs"));
2049 assert!(request.messages[2].string_contents().contains("file3.rs"));
2050 }
2051
2052 #[gpui::test]
2053 async fn test_message_without_files(cx: &mut TestAppContext) {
2054 init_test_settings(cx);
2055
2056 let project = create_test_project(
2057 cx,
2058 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
2059 )
2060 .await;
2061
2062 let (_, _thread_store, thread, _context_store) =
2063 setup_test_environment(cx, project.clone()).await;
2064
2065 // Insert user message without any context (empty context vector)
2066 let message_id = thread.update(cx, |thread, cx| {
2067 thread.insert_user_message("What is the best way to learn Rust?", vec![], None, cx)
2068 });
2069
2070 // Check content and context in message object
2071 let message = thread.read_with(cx, |thread, _| thread.message(message_id).unwrap().clone());
2072
2073 // Context should be empty when no files are included
2074 assert_eq!(message.role, Role::User);
2075 assert_eq!(message.segments.len(), 1);
2076 assert_eq!(
2077 message.segments[0],
2078 MessageSegment::Text("What is the best way to learn Rust?".to_string())
2079 );
2080 assert_eq!(message.context, "");
2081
2082 // Check message in request
2083 let request = thread.read_with(cx, |thread, cx| {
2084 thread.to_completion_request(RequestKind::Chat, cx)
2085 });
2086
2087 assert_eq!(request.messages.len(), 1);
2088 assert_eq!(
2089 request.messages[0].string_contents(),
2090 "What is the best way to learn Rust?"
2091 );
2092
2093 // Add second message, also without context
2094 let message2_id = thread.update(cx, |thread, cx| {
2095 thread.insert_user_message("Are there any good books?", vec![], None, cx)
2096 });
2097
2098 let message2 =
2099 thread.read_with(cx, |thread, _| thread.message(message2_id).unwrap().clone());
2100 assert_eq!(message2.context, "");
2101
2102 // Check that both messages appear in the request
2103 let request = thread.read_with(cx, |thread, cx| {
2104 thread.to_completion_request(RequestKind::Chat, cx)
2105 });
2106
2107 assert_eq!(request.messages.len(), 2);
2108 assert_eq!(
2109 request.messages[0].string_contents(),
2110 "What is the best way to learn Rust?"
2111 );
2112 assert_eq!(
2113 request.messages[1].string_contents(),
2114 "Are there any good books?"
2115 );
2116 }
2117
2118 #[gpui::test]
2119 async fn test_stale_buffer_notification(cx: &mut TestAppContext) {
2120 init_test_settings(cx);
2121
2122 let project = create_test_project(
2123 cx,
2124 json!({"code.rs": "fn main() {\n println!(\"Hello, world!\");\n}"}),
2125 )
2126 .await;
2127
2128 let (_workspace, _thread_store, thread, context_store) =
2129 setup_test_environment(cx, project.clone()).await;
2130
2131 // Open buffer and add it to context
2132 let buffer = add_file_to_context(&project, &context_store, "test/code.rs", cx)
2133 .await
2134 .unwrap();
2135
2136 let context =
2137 context_store.update(cx, |store, _| store.context().first().cloned().unwrap());
2138
2139 // Insert user message with the buffer as context
2140 thread.update(cx, |thread, cx| {
2141 thread.insert_user_message("Explain this code", vec![context], None, cx)
2142 });
2143
2144 // Create a request and check that it doesn't have a stale buffer warning yet
2145 let initial_request = thread.read_with(cx, |thread, cx| {
2146 thread.to_completion_request(RequestKind::Chat, cx)
2147 });
2148
2149 // Make sure we don't have a stale file warning yet
2150 let has_stale_warning = initial_request.messages.iter().any(|msg| {
2151 msg.string_contents()
2152 .contains("These files changed since last read:")
2153 });
2154 assert!(
2155 !has_stale_warning,
2156 "Should not have stale buffer warning before buffer is modified"
2157 );
2158
2159 // Modify the buffer
2160 buffer.update(cx, |buffer, cx| {
2161 // Find a position at the end of line 1
2162 buffer.edit(
2163 [(1..1, "\n println!(\"Added a new line\");\n")],
2164 None,
2165 cx,
2166 );
2167 });
2168
2169 // Insert another user message without context
2170 thread.update(cx, |thread, cx| {
2171 thread.insert_user_message("What does the code do now?", vec![], None, cx)
2172 });
2173
2174 // Create a new request and check for the stale buffer warning
2175 let new_request = thread.read_with(cx, |thread, cx| {
2176 thread.to_completion_request(RequestKind::Chat, cx)
2177 });
2178
2179 // We should have a stale file warning as the last message
2180 let last_message = new_request
2181 .messages
2182 .last()
2183 .expect("Request should have messages");
2184
2185 // The last message should be the stale buffer notification
2186 assert_eq!(last_message.role, Role::User);
2187
2188 // Check the exact content of the message
2189 let expected_content = "These files changed since last read:\n- code.rs\n";
2190 assert_eq!(
2191 last_message.string_contents(),
2192 expected_content,
2193 "Last message should be exactly the stale buffer notification"
2194 );
2195 }
2196
2197 fn init_test_settings(cx: &mut TestAppContext) {
2198 cx.update(|cx| {
2199 let settings_store = SettingsStore::test(cx);
2200 cx.set_global(settings_store);
2201 language::init(cx);
2202 Project::init_settings(cx);
2203 AssistantSettings::register(cx);
2204 thread_store::init(cx);
2205 workspace::init_settings(cx);
2206 ThemeSettings::register(cx);
2207 ContextServerSettings::register(cx);
2208 EditorSettings::register(cx);
2209 });
2210 }
2211
2212 // Helper to create a test project with test files
2213 async fn create_test_project(
2214 cx: &mut TestAppContext,
2215 files: serde_json::Value,
2216 ) -> Entity<Project> {
2217 let fs = FakeFs::new(cx.executor());
2218 fs.insert_tree(path!("/test"), files).await;
2219 Project::test(fs, [path!("/test").as_ref()], cx).await
2220 }
2221
2222 async fn setup_test_environment(
2223 cx: &mut TestAppContext,
2224 project: Entity<Project>,
2225 ) -> (
2226 Entity<Workspace>,
2227 Entity<ThreadStore>,
2228 Entity<Thread>,
2229 Entity<ContextStore>,
2230 ) {
2231 let (workspace, cx) =
2232 cx.add_window_view(|window, cx| Workspace::test_new(project.clone(), window, cx));
2233
2234 let thread_store = cx.update(|_, cx| {
2235 ThreadStore::new(
2236 project.clone(),
2237 Arc::default(),
2238 Arc::new(PromptBuilder::new(None).unwrap()),
2239 cx,
2240 )
2241 .unwrap()
2242 });
2243
2244 let thread = thread_store.update(cx, |store, cx| store.create_thread(cx));
2245 let context_store = cx.new(|_cx| ContextStore::new(workspace.downgrade(), None));
2246
2247 (workspace, thread_store, thread, context_store)
2248 }
2249
2250 async fn add_file_to_context(
2251 project: &Entity<Project>,
2252 context_store: &Entity<ContextStore>,
2253 path: &str,
2254 cx: &mut TestAppContext,
2255 ) -> Result<Entity<language::Buffer>> {
2256 let buffer_path = project
2257 .read_with(cx, |project, cx| project.find_project_path(path, cx))
2258 .unwrap();
2259
2260 let buffer = project
2261 .update(cx, |project, cx| project.open_buffer(buffer_path, cx))
2262 .await
2263 .unwrap();
2264
2265 context_store
2266 .update(cx, |store, cx| {
2267 store.add_file_from_buffer(buffer.clone(), cx)
2268 })
2269 .await?;
2270
2271 Ok(buffer)
2272 }
2273}