1use std::sync::Arc;
2
3use anyhow::Result;
4use assistant_tool::ToolWorkingSet;
5use chrono::{DateTime, Utc};
6use collections::{BTreeMap, HashMap, HashSet};
7use futures::StreamExt as _;
8use gpui::{App, Context, Entity, EventEmitter, SharedString, Task};
9use language_model::{
10 LanguageModel, LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
11 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolResult,
12 LanguageModelToolUseId, MaxMonthlySpendReachedError, MessageContent, PaymentRequiredError,
13 Role, StopReason,
14};
15use project::Project;
16use scripting_tool::ScriptingTool;
17use serde::{Deserialize, Serialize};
18use util::{post_inc, TryFutureExt as _};
19use uuid::Uuid;
20
21use crate::context::{attach_context_to_message, ContextId, ContextSnapshot};
22use crate::thread_store::SavedThread;
23use crate::tool_use::{ToolUse, ToolUseState};
24
25#[derive(Debug, Clone, Copy)]
26pub enum RequestKind {
27 Chat,
28 /// Used when summarizing a thread.
29 Summarize,
30}
31
32#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
33pub struct ThreadId(Arc<str>);
34
35impl ThreadId {
36 pub fn new() -> Self {
37 Self(Uuid::new_v4().to_string().into())
38 }
39}
40
41impl std::fmt::Display for ThreadId {
42 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
43 write!(f, "{}", self.0)
44 }
45}
46
47#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Copy, Serialize, Deserialize)]
48pub struct MessageId(pub(crate) usize);
49
50impl MessageId {
51 fn post_inc(&mut self) -> Self {
52 Self(post_inc(&mut self.0))
53 }
54}
55
56/// A message in a [`Thread`].
57#[derive(Debug, Clone)]
58pub struct Message {
59 pub id: MessageId,
60 pub role: Role,
61 pub text: String,
62}
63
64/// A thread of conversation with the LLM.
65pub struct Thread {
66 id: ThreadId,
67 updated_at: DateTime<Utc>,
68 summary: Option<SharedString>,
69 pending_summary: Task<Option<()>>,
70 messages: Vec<Message>,
71 next_message_id: MessageId,
72 context: BTreeMap<ContextId, ContextSnapshot>,
73 context_by_message: HashMap<MessageId, Vec<ContextId>>,
74 completion_count: usize,
75 pending_completions: Vec<PendingCompletion>,
76 project: Entity<Project>,
77 tools: Arc<ToolWorkingSet>,
78 tool_use: ToolUseState,
79 scripting_tool_use: ToolUseState,
80}
81
82impl Thread {
83 pub fn new(
84 project: Entity<Project>,
85 tools: Arc<ToolWorkingSet>,
86 _cx: &mut Context<Self>,
87 ) -> Self {
88 Self {
89 id: ThreadId::new(),
90 updated_at: Utc::now(),
91 summary: None,
92 pending_summary: Task::ready(None),
93 messages: Vec::new(),
94 next_message_id: MessageId(0),
95 context: BTreeMap::default(),
96 context_by_message: HashMap::default(),
97 completion_count: 0,
98 pending_completions: Vec::new(),
99 project,
100 tools,
101 tool_use: ToolUseState::new(),
102 scripting_tool_use: ToolUseState::new(),
103 }
104 }
105
106 pub fn from_saved(
107 id: ThreadId,
108 saved: SavedThread,
109 project: Entity<Project>,
110 tools: Arc<ToolWorkingSet>,
111 _cx: &mut Context<Self>,
112 ) -> Self {
113 let next_message_id = MessageId(
114 saved
115 .messages
116 .last()
117 .map(|message| message.id.0 + 1)
118 .unwrap_or(0),
119 );
120 let tool_use = ToolUseState::from_saved_messages(&saved.messages);
121 let scripting_tool_use = ToolUseState::new();
122
123 Self {
124 id,
125 updated_at: saved.updated_at,
126 summary: Some(saved.summary),
127 pending_summary: Task::ready(None),
128 messages: saved
129 .messages
130 .into_iter()
131 .map(|message| Message {
132 id: message.id,
133 role: message.role,
134 text: message.text,
135 })
136 .collect(),
137 next_message_id,
138 context: BTreeMap::default(),
139 context_by_message: HashMap::default(),
140 completion_count: 0,
141 pending_completions: Vec::new(),
142 project,
143 tools,
144 tool_use,
145 scripting_tool_use,
146 }
147 }
148
149 pub fn id(&self) -> &ThreadId {
150 &self.id
151 }
152
153 pub fn is_empty(&self) -> bool {
154 self.messages.is_empty()
155 }
156
157 pub fn updated_at(&self) -> DateTime<Utc> {
158 self.updated_at
159 }
160
161 pub fn touch_updated_at(&mut self) {
162 self.updated_at = Utc::now();
163 }
164
165 pub fn summary(&self) -> Option<SharedString> {
166 self.summary.clone()
167 }
168
169 pub fn summary_or_default(&self) -> SharedString {
170 const DEFAULT: SharedString = SharedString::new_static("New Thread");
171 self.summary.clone().unwrap_or(DEFAULT)
172 }
173
174 pub fn set_summary(&mut self, summary: impl Into<SharedString>, cx: &mut Context<Self>) {
175 self.summary = Some(summary.into());
176 cx.emit(ThreadEvent::SummaryChanged);
177 }
178
179 pub fn message(&self, id: MessageId) -> Option<&Message> {
180 self.messages.iter().find(|message| message.id == id)
181 }
182
183 pub fn messages(&self) -> impl Iterator<Item = &Message> {
184 self.messages.iter()
185 }
186
187 pub fn is_streaming(&self) -> bool {
188 !self.pending_completions.is_empty()
189 }
190
191 pub fn tools(&self) -> &Arc<ToolWorkingSet> {
192 &self.tools
193 }
194
195 pub fn context_for_message(&self, id: MessageId) -> Option<Vec<ContextSnapshot>> {
196 let context = self.context_by_message.get(&id)?;
197 Some(
198 context
199 .into_iter()
200 .filter_map(|context_id| self.context.get(&context_id))
201 .cloned()
202 .collect::<Vec<_>>(),
203 )
204 }
205
206 /// Returns whether all of the tool uses have finished running.
207 pub fn all_tools_finished(&self) -> bool {
208 let mut all_pending_tool_uses = self
209 .tool_use
210 .pending_tool_uses()
211 .into_iter()
212 .chain(self.scripting_tool_use.pending_tool_uses());
213
214 // If the only pending tool uses left are the ones with errors, then that means that we've finished running all
215 // of the pending tools.
216 all_pending_tool_uses.all(|tool_use| tool_use.status.is_error())
217 }
218
219 pub fn tool_uses_for_message(&self, id: MessageId) -> Vec<ToolUse> {
220 self.tool_use.tool_uses_for_message(id)
221 }
222
223 pub fn scripting_tool_uses_for_message(&self, id: MessageId) -> Vec<ToolUse> {
224 self.scripting_tool_use.tool_uses_for_message(id)
225 }
226
227 pub fn tool_results_for_message(&self, id: MessageId) -> Vec<&LanguageModelToolResult> {
228 self.tool_use.tool_results_for_message(id)
229 }
230
231 pub fn scripting_tool_results_for_message(
232 &self,
233 id: MessageId,
234 ) -> Vec<&LanguageModelToolResult> {
235 self.scripting_tool_use.tool_results_for_message(id)
236 }
237
238 pub fn message_has_tool_results(&self, message_id: MessageId) -> bool {
239 self.tool_use.message_has_tool_results(message_id)
240 }
241
242 pub fn message_has_scripting_tool_results(&self, message_id: MessageId) -> bool {
243 self.scripting_tool_use.message_has_tool_results(message_id)
244 }
245
246 pub fn insert_user_message(
247 &mut self,
248 text: impl Into<String>,
249 context: Vec<ContextSnapshot>,
250 cx: &mut Context<Self>,
251 ) -> MessageId {
252 let message_id = self.insert_message(Role::User, text, cx);
253 let context_ids = context.iter().map(|context| context.id).collect::<Vec<_>>();
254 self.context
255 .extend(context.into_iter().map(|context| (context.id, context)));
256 self.context_by_message.insert(message_id, context_ids);
257 message_id
258 }
259
260 pub fn insert_message(
261 &mut self,
262 role: Role,
263 text: impl Into<String>,
264 cx: &mut Context<Self>,
265 ) -> MessageId {
266 let id = self.next_message_id.post_inc();
267 self.messages.push(Message {
268 id,
269 role,
270 text: text.into(),
271 });
272 self.touch_updated_at();
273 cx.emit(ThreadEvent::MessageAdded(id));
274 id
275 }
276
277 pub fn edit_message(
278 &mut self,
279 id: MessageId,
280 new_role: Role,
281 new_text: String,
282 cx: &mut Context<Self>,
283 ) -> bool {
284 let Some(message) = self.messages.iter_mut().find(|message| message.id == id) else {
285 return false;
286 };
287 message.role = new_role;
288 message.text = new_text;
289 self.touch_updated_at();
290 cx.emit(ThreadEvent::MessageEdited(id));
291 true
292 }
293
294 pub fn delete_message(&mut self, id: MessageId, cx: &mut Context<Self>) -> bool {
295 let Some(index) = self.messages.iter().position(|message| message.id == id) else {
296 return false;
297 };
298 self.messages.remove(index);
299 self.context_by_message.remove(&id);
300 self.touch_updated_at();
301 cx.emit(ThreadEvent::MessageDeleted(id));
302 true
303 }
304
305 /// Returns the representation of this [`Thread`] in a textual form.
306 ///
307 /// This is the representation we use when attaching a thread as context to another thread.
308 pub fn text(&self) -> String {
309 let mut text = String::new();
310
311 for message in &self.messages {
312 text.push_str(match message.role {
313 language_model::Role::User => "User:",
314 language_model::Role::Assistant => "Assistant:",
315 language_model::Role::System => "System:",
316 });
317 text.push('\n');
318
319 text.push_str(&message.text);
320 text.push('\n');
321 }
322
323 text
324 }
325
326 pub fn send_to_model(
327 &mut self,
328 model: Arc<dyn LanguageModel>,
329 request_kind: RequestKind,
330 use_tools: bool,
331 cx: &mut Context<Self>,
332 ) {
333 let mut request = self.to_completion_request(request_kind, cx);
334
335 if use_tools {
336 let mut tools = Vec::new();
337 tools.push(LanguageModelRequestTool {
338 name: ScriptingTool::NAME.into(),
339 description: ScriptingTool::DESCRIPTION.into(),
340 input_schema: ScriptingTool::input_schema(),
341 });
342
343 tools.extend(
344 self.tools()
345 .tools(cx)
346 .into_iter()
347 .map(|tool| LanguageModelRequestTool {
348 name: tool.name(),
349 description: tool.description(),
350 input_schema: tool.input_schema(),
351 }),
352 );
353
354 request.tools = tools;
355 }
356
357 self.stream_completion(request, model, cx);
358 }
359
360 pub fn to_completion_request(
361 &self,
362 request_kind: RequestKind,
363 _cx: &App,
364 ) -> LanguageModelRequest {
365 let mut request = LanguageModelRequest {
366 messages: vec![],
367 tools: Vec::new(),
368 stop: Vec::new(),
369 temperature: None,
370 };
371
372 let mut referenced_context_ids = HashSet::default();
373
374 for message in &self.messages {
375 if let Some(context_ids) = self.context_by_message.get(&message.id) {
376 referenced_context_ids.extend(context_ids);
377 }
378
379 let mut request_message = LanguageModelRequestMessage {
380 role: message.role,
381 content: Vec::new(),
382 cache: false,
383 };
384
385 match request_kind {
386 RequestKind::Chat => {
387 self.tool_use
388 .attach_tool_results(message.id, &mut request_message);
389 self.scripting_tool_use
390 .attach_tool_results(message.id, &mut request_message);
391 }
392 RequestKind::Summarize => {
393 // We don't care about tool use during summarization.
394 }
395 }
396
397 if !message.text.is_empty() {
398 request_message
399 .content
400 .push(MessageContent::Text(message.text.clone()));
401 }
402
403 match request_kind {
404 RequestKind::Chat => {
405 self.tool_use
406 .attach_tool_uses(message.id, &mut request_message);
407 self.scripting_tool_use
408 .attach_tool_uses(message.id, &mut request_message);
409 }
410 RequestKind::Summarize => {
411 // We don't care about tool use during summarization.
412 }
413 };
414
415 request.messages.push(request_message);
416 }
417
418 if !referenced_context_ids.is_empty() {
419 let mut context_message = LanguageModelRequestMessage {
420 role: Role::User,
421 content: Vec::new(),
422 cache: false,
423 };
424
425 let referenced_context = referenced_context_ids
426 .into_iter()
427 .filter_map(|context_id| self.context.get(context_id))
428 .cloned();
429 attach_context_to_message(&mut context_message, referenced_context);
430
431 request.messages.push(context_message);
432 }
433
434 request
435 }
436
437 pub fn stream_completion(
438 &mut self,
439 request: LanguageModelRequest,
440 model: Arc<dyn LanguageModel>,
441 cx: &mut Context<Self>,
442 ) {
443 let pending_completion_id = post_inc(&mut self.completion_count);
444
445 let task = cx.spawn(|thread, mut cx| async move {
446 let stream = model.stream_completion(request, &cx);
447 let stream_completion = async {
448 let mut events = stream.await?;
449 let mut stop_reason = StopReason::EndTurn;
450
451 while let Some(event) = events.next().await {
452 let event = event?;
453
454 thread.update(&mut cx, |thread, cx| {
455 match event {
456 LanguageModelCompletionEvent::StartMessage { .. } => {
457 thread.insert_message(Role::Assistant, String::new(), cx);
458 }
459 LanguageModelCompletionEvent::Stop(reason) => {
460 stop_reason = reason;
461 }
462 LanguageModelCompletionEvent::Text(chunk) => {
463 if let Some(last_message) = thread.messages.last_mut() {
464 if last_message.role == Role::Assistant {
465 last_message.text.push_str(&chunk);
466 cx.emit(ThreadEvent::StreamedAssistantText(
467 last_message.id,
468 chunk,
469 ));
470 } else {
471 // If we won't have an Assistant message yet, assume this chunk marks the beginning
472 // of a new Assistant response.
473 //
474 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
475 // will result in duplicating the text of the chunk in the rendered Markdown.
476 thread.insert_message(Role::Assistant, chunk, cx);
477 };
478 }
479 }
480 LanguageModelCompletionEvent::ToolUse(tool_use) => {
481 if let Some(last_assistant_message) = thread
482 .messages
483 .iter()
484 .rfind(|message| message.role == Role::Assistant)
485 {
486 if tool_use.name.as_ref() == ScriptingTool::NAME {
487 thread
488 .scripting_tool_use
489 .request_tool_use(last_assistant_message.id, tool_use);
490 } else {
491 thread
492 .tool_use
493 .request_tool_use(last_assistant_message.id, tool_use);
494 }
495 }
496 }
497 }
498
499 thread.touch_updated_at();
500 cx.emit(ThreadEvent::StreamedCompletion);
501 cx.notify();
502 })?;
503
504 smol::future::yield_now().await;
505 }
506
507 thread.update(&mut cx, |thread, cx| {
508 thread
509 .pending_completions
510 .retain(|completion| completion.id != pending_completion_id);
511
512 if thread.summary.is_none() && thread.messages.len() >= 2 {
513 thread.summarize(cx);
514 }
515 })?;
516
517 anyhow::Ok(stop_reason)
518 };
519
520 let result = stream_completion.await;
521
522 thread
523 .update(&mut cx, |thread, cx| match result.as_ref() {
524 Ok(stop_reason) => match stop_reason {
525 StopReason::ToolUse => {
526 cx.emit(ThreadEvent::UsePendingTools);
527 }
528 StopReason::EndTurn => {}
529 StopReason::MaxTokens => {}
530 },
531 Err(error) => {
532 if error.is::<PaymentRequiredError>() {
533 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
534 } else if error.is::<MaxMonthlySpendReachedError>() {
535 cx.emit(ThreadEvent::ShowError(ThreadError::MaxMonthlySpendReached));
536 } else {
537 let error_message = error
538 .chain()
539 .map(|err| err.to_string())
540 .collect::<Vec<_>>()
541 .join("\n");
542 cx.emit(ThreadEvent::ShowError(ThreadError::Message(
543 SharedString::from(error_message.clone()),
544 )));
545 }
546
547 thread.cancel_last_completion();
548 }
549 })
550 .ok();
551 });
552
553 self.pending_completions.push(PendingCompletion {
554 id: pending_completion_id,
555 _task: task,
556 });
557 }
558
559 pub fn summarize(&mut self, cx: &mut Context<Self>) {
560 let Some(provider) = LanguageModelRegistry::read_global(cx).active_provider() else {
561 return;
562 };
563 let Some(model) = LanguageModelRegistry::read_global(cx).active_model() else {
564 return;
565 };
566
567 if !provider.is_authenticated(cx) {
568 return;
569 }
570
571 let mut request = self.to_completion_request(RequestKind::Summarize, cx);
572 request.messages.push(LanguageModelRequestMessage {
573 role: Role::User,
574 content: vec![
575 "Generate a concise 3-7 word title for this conversation, omitting punctuation. Go straight to the title, without any preamble and prefix like `Here's a concise suggestion:...` or `Title:`"
576 .into(),
577 ],
578 cache: false,
579 });
580
581 self.pending_summary = cx.spawn(|this, mut cx| {
582 async move {
583 let stream = model.stream_completion_text(request, &cx);
584 let mut messages = stream.await?;
585
586 let mut new_summary = String::new();
587 while let Some(message) = messages.stream.next().await {
588 let text = message?;
589 let mut lines = text.lines();
590 new_summary.extend(lines.next());
591
592 // Stop if the LLM generated multiple lines.
593 if lines.next().is_some() {
594 break;
595 }
596 }
597
598 this.update(&mut cx, |this, cx| {
599 if !new_summary.is_empty() {
600 this.summary = Some(new_summary.into());
601 }
602
603 cx.emit(ThreadEvent::SummaryChanged);
604 })?;
605
606 anyhow::Ok(())
607 }
608 .log_err()
609 });
610 }
611
612 pub fn use_pending_tools(&mut self, cx: &mut Context<Self>) {
613 let pending_tool_uses = self
614 .tool_use
615 .pending_tool_uses()
616 .into_iter()
617 .filter(|tool_use| tool_use.status.is_idle())
618 .cloned()
619 .collect::<Vec<_>>();
620
621 for tool_use in pending_tool_uses {
622 if let Some(tool) = self.tools.tool(&tool_use.name, cx) {
623 let task = tool.run(tool_use.input, self.project.clone(), cx);
624
625 self.insert_tool_output(tool_use.id.clone(), task, cx);
626 }
627 }
628
629 let pending_scripting_tool_uses = self
630 .scripting_tool_use
631 .pending_tool_uses()
632 .into_iter()
633 .filter(|tool_use| tool_use.status.is_idle())
634 .cloned()
635 .collect::<Vec<_>>();
636
637 for scripting_tool_use in pending_scripting_tool_uses {
638 let task = ScriptingTool.run(scripting_tool_use.input, self.project.clone(), cx);
639
640 self.insert_scripting_tool_output(scripting_tool_use.id.clone(), task, cx);
641 }
642 }
643
644 pub fn insert_tool_output(
645 &mut self,
646 tool_use_id: LanguageModelToolUseId,
647 output: Task<Result<String>>,
648 cx: &mut Context<Self>,
649 ) {
650 let insert_output_task = cx.spawn(|thread, mut cx| {
651 let tool_use_id = tool_use_id.clone();
652 async move {
653 let output = output.await;
654 thread
655 .update(&mut cx, |thread, cx| {
656 thread
657 .tool_use
658 .insert_tool_output(tool_use_id.clone(), output);
659
660 cx.emit(ThreadEvent::ToolFinished { tool_use_id });
661 })
662 .ok();
663 }
664 });
665
666 self.tool_use
667 .run_pending_tool(tool_use_id, insert_output_task);
668 }
669
670 pub fn insert_scripting_tool_output(
671 &mut self,
672 tool_use_id: LanguageModelToolUseId,
673 output: Task<Result<String>>,
674 cx: &mut Context<Self>,
675 ) {
676 let insert_output_task = cx.spawn(|thread, mut cx| {
677 let tool_use_id = tool_use_id.clone();
678 async move {
679 let output = output.await;
680 thread
681 .update(&mut cx, |thread, cx| {
682 thread
683 .scripting_tool_use
684 .insert_tool_output(tool_use_id.clone(), output);
685
686 cx.emit(ThreadEvent::ToolFinished { tool_use_id });
687 })
688 .ok();
689 }
690 });
691
692 self.scripting_tool_use
693 .run_pending_tool(tool_use_id, insert_output_task);
694 }
695
696 pub fn send_tool_results_to_model(
697 &mut self,
698 model: Arc<dyn LanguageModel>,
699 cx: &mut Context<Self>,
700 ) {
701 // Insert a user message to contain the tool results.
702 self.insert_user_message(
703 // TODO: Sending up a user message without any content results in the model sending back
704 // responses that also don't have any content. We currently don't handle this case well,
705 // so for now we provide some text to keep the model on track.
706 "Here are the tool results.",
707 Vec::new(),
708 cx,
709 );
710 self.send_to_model(model, RequestKind::Chat, true, cx);
711 }
712
713 /// Cancels the last pending completion, if there are any pending.
714 ///
715 /// Returns whether a completion was canceled.
716 pub fn cancel_last_completion(&mut self) -> bool {
717 if let Some(_last_completion) = self.pending_completions.pop() {
718 true
719 } else {
720 false
721 }
722 }
723}
724
725#[derive(Debug, Clone)]
726pub enum ThreadError {
727 PaymentRequired,
728 MaxMonthlySpendReached,
729 Message(SharedString),
730}
731
732#[derive(Debug, Clone)]
733pub enum ThreadEvent {
734 ShowError(ThreadError),
735 StreamedCompletion,
736 StreamedAssistantText(MessageId, String),
737 MessageAdded(MessageId),
738 MessageEdited(MessageId),
739 MessageDeleted(MessageId),
740 SummaryChanged,
741 UsePendingTools,
742 ToolFinished {
743 #[allow(unused)]
744 tool_use_id: LanguageModelToolUseId,
745 },
746 ScriptFinished,
747}
748
749impl EventEmitter<ThreadEvent> for Thread {}
750
751struct PendingCompletion {
752 id: usize,
753 _task: Task<()>,
754}