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, EventEmitter, SharedString, Task};
9use language_model::{
10 LanguageModel, LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
11 LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelToolUseId,
12 MaxMonthlySpendReachedError, MessageContent, PaymentRequiredError, Role, StopReason,
13};
14use serde::{Deserialize, Serialize};
15use util::{post_inc, TryFutureExt as _};
16use uuid::Uuid;
17
18use crate::context::{attach_context_to_message, ContextId, ContextSnapshot};
19use crate::thread_store::SavedThread;
20use crate::tool_use::{PendingToolUse, ToolUse, ToolUseState};
21
22#[derive(Debug, Clone, Copy)]
23pub enum RequestKind {
24 Chat,
25 /// Used when summarizing a thread.
26 Summarize,
27}
28
29#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
30pub struct ThreadId(Arc<str>);
31
32impl ThreadId {
33 pub fn new() -> Self {
34 Self(Uuid::new_v4().to_string().into())
35 }
36}
37
38impl std::fmt::Display for ThreadId {
39 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
40 write!(f, "{}", self.0)
41 }
42}
43
44#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Copy, Serialize, Deserialize)]
45pub struct MessageId(pub(crate) usize);
46
47impl MessageId {
48 fn post_inc(&mut self) -> Self {
49 Self(post_inc(&mut self.0))
50 }
51}
52
53/// A message in a [`Thread`].
54#[derive(Debug, Clone)]
55pub struct Message {
56 pub id: MessageId,
57 pub role: Role,
58 pub text: String,
59}
60
61/// A thread of conversation with the LLM.
62pub struct Thread {
63 id: ThreadId,
64 updated_at: DateTime<Utc>,
65 summary: Option<SharedString>,
66 pending_summary: Task<Option<()>>,
67 messages: Vec<Message>,
68 next_message_id: MessageId,
69 context: BTreeMap<ContextId, ContextSnapshot>,
70 context_by_message: HashMap<MessageId, Vec<ContextId>>,
71 completion_count: usize,
72 pending_completions: Vec<PendingCompletion>,
73 tools: Arc<ToolWorkingSet>,
74 tool_use: ToolUseState,
75}
76
77impl Thread {
78 pub fn new(tools: Arc<ToolWorkingSet>, _cx: &mut Context<Self>) -> Self {
79 Self {
80 id: ThreadId::new(),
81 updated_at: Utc::now(),
82 summary: None,
83 pending_summary: Task::ready(None),
84 messages: Vec::new(),
85 next_message_id: MessageId(0),
86 context: BTreeMap::default(),
87 context_by_message: HashMap::default(),
88 completion_count: 0,
89 pending_completions: Vec::new(),
90 tools,
91 tool_use: ToolUseState::default(),
92 }
93 }
94
95 pub fn from_saved(
96 id: ThreadId,
97 saved: SavedThread,
98 tools: Arc<ToolWorkingSet>,
99 _cx: &mut Context<Self>,
100 ) -> Self {
101 let next_message_id = MessageId(saved.messages.len());
102
103 Self {
104 id,
105 updated_at: saved.updated_at,
106 summary: Some(saved.summary),
107 pending_summary: Task::ready(None),
108 messages: saved
109 .messages
110 .into_iter()
111 .map(|message| Message {
112 id: message.id,
113 role: message.role,
114 text: message.text,
115 })
116 .collect(),
117 next_message_id,
118 context: BTreeMap::default(),
119 context_by_message: HashMap::default(),
120 completion_count: 0,
121 pending_completions: Vec::new(),
122 tools,
123 tool_use: ToolUseState::default(),
124 }
125 }
126
127 pub fn id(&self) -> &ThreadId {
128 &self.id
129 }
130
131 pub fn is_empty(&self) -> bool {
132 self.messages.is_empty()
133 }
134
135 pub fn updated_at(&self) -> DateTime<Utc> {
136 self.updated_at
137 }
138
139 pub fn touch_updated_at(&mut self) {
140 self.updated_at = Utc::now();
141 }
142
143 pub fn summary(&self) -> Option<SharedString> {
144 self.summary.clone()
145 }
146
147 pub fn summary_or_default(&self) -> SharedString {
148 const DEFAULT: SharedString = SharedString::new_static("New Thread");
149 self.summary.clone().unwrap_or(DEFAULT)
150 }
151
152 pub fn set_summary(&mut self, summary: impl Into<SharedString>, cx: &mut Context<Self>) {
153 self.summary = Some(summary.into());
154 cx.emit(ThreadEvent::SummaryChanged);
155 }
156
157 pub fn message(&self, id: MessageId) -> Option<&Message> {
158 self.messages.iter().find(|message| message.id == id)
159 }
160
161 pub fn messages(&self) -> impl Iterator<Item = &Message> {
162 self.messages.iter()
163 }
164
165 pub fn is_streaming(&self) -> bool {
166 !self.pending_completions.is_empty()
167 }
168
169 pub fn tools(&self) -> &Arc<ToolWorkingSet> {
170 &self.tools
171 }
172
173 pub fn context_for_message(&self, id: MessageId) -> Option<Vec<ContextSnapshot>> {
174 let context = self.context_by_message.get(&id)?;
175 Some(
176 context
177 .into_iter()
178 .filter_map(|context_id| self.context.get(&context_id))
179 .cloned()
180 .collect::<Vec<_>>(),
181 )
182 }
183
184 pub fn pending_tool_uses(&self) -> Vec<&PendingToolUse> {
185 self.tool_use.pending_tool_uses()
186 }
187
188 pub fn tool_uses_for_message(&self, id: MessageId) -> Vec<ToolUse> {
189 self.tool_use.tool_uses_for_message(id)
190 }
191
192 pub fn message_has_tool_results(&self, message_id: MessageId) -> bool {
193 self.tool_use.message_has_tool_results(message_id)
194 }
195
196 pub fn insert_user_message(
197 &mut self,
198 text: impl Into<String>,
199 context: Vec<ContextSnapshot>,
200 cx: &mut Context<Self>,
201 ) {
202 let message_id = self.insert_message(Role::User, text, cx);
203 let context_ids = context.iter().map(|context| context.id).collect::<Vec<_>>();
204 self.context
205 .extend(context.into_iter().map(|context| (context.id, context)));
206 self.context_by_message.insert(message_id, context_ids);
207 }
208
209 pub fn insert_message(
210 &mut self,
211 role: Role,
212 text: impl Into<String>,
213 cx: &mut Context<Self>,
214 ) -> MessageId {
215 let id = self.next_message_id.post_inc();
216 self.messages.push(Message {
217 id,
218 role,
219 text: text.into(),
220 });
221 self.touch_updated_at();
222 cx.emit(ThreadEvent::MessageAdded(id));
223 id
224 }
225
226 /// Returns the representation of this [`Thread`] in a textual form.
227 ///
228 /// This is the representation we use when attaching a thread as context to another thread.
229 pub fn text(&self) -> String {
230 let mut text = String::new();
231
232 for message in &self.messages {
233 text.push_str(match message.role {
234 language_model::Role::User => "User:",
235 language_model::Role::Assistant => "Assistant:",
236 language_model::Role::System => "System:",
237 });
238 text.push('\n');
239
240 text.push_str(&message.text);
241 text.push('\n');
242 }
243
244 text
245 }
246
247 pub fn send_to_model(
248 &mut self,
249 model: Arc<dyn LanguageModel>,
250 request_kind: RequestKind,
251 use_tools: bool,
252 cx: &mut Context<Self>,
253 ) {
254 let mut request = self.to_completion_request(request_kind, cx);
255
256 if use_tools {
257 request.tools = self
258 .tools()
259 .tools(cx)
260 .into_iter()
261 .map(|tool| LanguageModelRequestTool {
262 name: tool.name(),
263 description: tool.description(),
264 input_schema: tool.input_schema(),
265 })
266 .collect();
267 }
268
269 self.stream_completion(request, model, cx);
270 }
271
272 pub fn to_completion_request(
273 &self,
274 request_kind: RequestKind,
275 _cx: &App,
276 ) -> LanguageModelRequest {
277 let mut request = LanguageModelRequest {
278 messages: vec![],
279 tools: Vec::new(),
280 stop: Vec::new(),
281 temperature: None,
282 };
283
284 let mut referenced_context_ids = HashSet::default();
285
286 for message in &self.messages {
287 if let Some(context_ids) = self.context_by_message.get(&message.id) {
288 referenced_context_ids.extend(context_ids);
289 }
290
291 let mut request_message = LanguageModelRequestMessage {
292 role: message.role,
293 content: Vec::new(),
294 cache: false,
295 };
296 match request_kind {
297 RequestKind::Chat => {
298 self.tool_use
299 .attach_tool_results(message.id, &mut request_message);
300 }
301 RequestKind::Summarize => {
302 // We don't care about tool use during summarization.
303 }
304 }
305
306 if !message.text.is_empty() {
307 request_message
308 .content
309 .push(MessageContent::Text(message.text.clone()));
310 }
311
312 match request_kind {
313 RequestKind::Chat => {
314 self.tool_use
315 .attach_tool_uses(message.id, &mut request_message);
316 }
317 RequestKind::Summarize => {
318 // We don't care about tool use during summarization.
319 }
320 }
321
322 request.messages.push(request_message);
323 }
324
325 if !referenced_context_ids.is_empty() {
326 let mut context_message = LanguageModelRequestMessage {
327 role: Role::User,
328 content: Vec::new(),
329 cache: false,
330 };
331
332 let referenced_context = referenced_context_ids
333 .into_iter()
334 .filter_map(|context_id| self.context.get(context_id))
335 .cloned();
336 attach_context_to_message(&mut context_message, referenced_context);
337
338 request.messages.push(context_message);
339 }
340
341 request
342 }
343
344 pub fn stream_completion(
345 &mut self,
346 request: LanguageModelRequest,
347 model: Arc<dyn LanguageModel>,
348 cx: &mut Context<Self>,
349 ) {
350 let pending_completion_id = post_inc(&mut self.completion_count);
351
352 let task = cx.spawn(|thread, mut cx| async move {
353 let stream = model.stream_completion(request, &cx);
354 let stream_completion = async {
355 let mut events = stream.await?;
356 let mut stop_reason = StopReason::EndTurn;
357
358 while let Some(event) = events.next().await {
359 let event = event?;
360
361 thread.update(&mut cx, |thread, cx| {
362 match event {
363 LanguageModelCompletionEvent::StartMessage { .. } => {
364 thread.insert_message(Role::Assistant, String::new(), cx);
365 }
366 LanguageModelCompletionEvent::Stop(reason) => {
367 stop_reason = reason;
368 }
369 LanguageModelCompletionEvent::Text(chunk) => {
370 if let Some(last_message) = thread.messages.last_mut() {
371 if last_message.role == Role::Assistant {
372 last_message.text.push_str(&chunk);
373 cx.emit(ThreadEvent::StreamedAssistantText(
374 last_message.id,
375 chunk,
376 ));
377 } else {
378 // If we won't have an Assistant message yet, assume this chunk marks the beginning
379 // of a new Assistant response.
380 //
381 // Importantly: We do *not* want to emit a `StreamedAssistantText` event here, as it
382 // will result in duplicating the text of the chunk in the rendered Markdown.
383 thread.insert_message(Role::Assistant, chunk, cx);
384 }
385 }
386 }
387 LanguageModelCompletionEvent::ToolUse(tool_use) => {
388 if let Some(last_assistant_message) = thread
389 .messages
390 .iter()
391 .rfind(|message| message.role == Role::Assistant)
392 {
393 thread
394 .tool_use
395 .request_tool_use(last_assistant_message.id, tool_use);
396 }
397 }
398 }
399
400 thread.touch_updated_at();
401 cx.emit(ThreadEvent::StreamedCompletion);
402 cx.notify();
403 })?;
404
405 smol::future::yield_now().await;
406 }
407
408 thread.update(&mut cx, |thread, cx| {
409 thread
410 .pending_completions
411 .retain(|completion| completion.id != pending_completion_id);
412
413 if thread.summary.is_none() && thread.messages.len() >= 2 {
414 thread.summarize(cx);
415 }
416 })?;
417
418 anyhow::Ok(stop_reason)
419 };
420
421 let result = stream_completion.await;
422
423 thread
424 .update(&mut cx, |thread, cx| match result.as_ref() {
425 Ok(stop_reason) => match stop_reason {
426 StopReason::ToolUse => {
427 cx.emit(ThreadEvent::UsePendingTools);
428 }
429 StopReason::EndTurn => {}
430 StopReason::MaxTokens => {}
431 },
432 Err(error) => {
433 if error.is::<PaymentRequiredError>() {
434 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
435 } else if error.is::<MaxMonthlySpendReachedError>() {
436 cx.emit(ThreadEvent::ShowError(ThreadError::MaxMonthlySpendReached));
437 } else {
438 let error_message = error
439 .chain()
440 .map(|err| err.to_string())
441 .collect::<Vec<_>>()
442 .join("\n");
443 cx.emit(ThreadEvent::ShowError(ThreadError::Message(
444 SharedString::from(error_message.clone()),
445 )));
446 }
447
448 thread.cancel_last_completion();
449 }
450 })
451 .ok();
452 });
453
454 self.pending_completions.push(PendingCompletion {
455 id: pending_completion_id,
456 _task: task,
457 });
458 }
459
460 pub fn summarize(&mut self, cx: &mut Context<Self>) {
461 let Some(provider) = LanguageModelRegistry::read_global(cx).active_provider() else {
462 return;
463 };
464 let Some(model) = LanguageModelRegistry::read_global(cx).active_model() else {
465 return;
466 };
467
468 if !provider.is_authenticated(cx) {
469 return;
470 }
471
472 let mut request = self.to_completion_request(RequestKind::Summarize, cx);
473 request.messages.push(LanguageModelRequestMessage {
474 role: Role::User,
475 content: vec![
476 "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:`"
477 .into(),
478 ],
479 cache: false,
480 });
481
482 self.pending_summary = cx.spawn(|this, mut cx| {
483 async move {
484 let stream = model.stream_completion_text(request, &cx);
485 let mut messages = stream.await?;
486
487 let mut new_summary = String::new();
488 while let Some(message) = messages.stream.next().await {
489 let text = message?;
490 let mut lines = text.lines();
491 new_summary.extend(lines.next());
492
493 // Stop if the LLM generated multiple lines.
494 if lines.next().is_some() {
495 break;
496 }
497 }
498
499 this.update(&mut cx, |this, cx| {
500 if !new_summary.is_empty() {
501 this.summary = Some(new_summary.into());
502 }
503
504 cx.emit(ThreadEvent::SummaryChanged);
505 })?;
506
507 anyhow::Ok(())
508 }
509 .log_err()
510 });
511 }
512
513 pub fn insert_tool_output(
514 &mut self,
515 tool_use_id: LanguageModelToolUseId,
516 output: Task<Result<String>>,
517 cx: &mut Context<Self>,
518 ) {
519 let insert_output_task = cx.spawn(|thread, mut cx| {
520 let tool_use_id = tool_use_id.clone();
521 async move {
522 let output = output.await;
523 thread
524 .update(&mut cx, |thread, cx| {
525 thread
526 .tool_use
527 .insert_tool_output(tool_use_id.clone(), output);
528
529 cx.emit(ThreadEvent::ToolFinished { tool_use_id });
530 })
531 .ok();
532 }
533 });
534
535 self.tool_use
536 .run_pending_tool(tool_use_id, insert_output_task);
537 }
538
539 /// Cancels the last pending completion, if there are any pending.
540 ///
541 /// Returns whether a completion was canceled.
542 pub fn cancel_last_completion(&mut self) -> bool {
543 if let Some(_last_completion) = self.pending_completions.pop() {
544 true
545 } else {
546 false
547 }
548 }
549}
550
551#[derive(Debug, Clone)]
552pub enum ThreadError {
553 PaymentRequired,
554 MaxMonthlySpendReached,
555 Message(SharedString),
556}
557
558#[derive(Debug, Clone)]
559pub enum ThreadEvent {
560 ShowError(ThreadError),
561 StreamedCompletion,
562 StreamedAssistantText(MessageId, String),
563 MessageAdded(MessageId),
564 SummaryChanged,
565 UsePendingTools,
566 ToolFinished {
567 #[allow(unused)]
568 tool_use_id: LanguageModelToolUseId,
569 },
570}
571
572impl EventEmitter<ThreadEvent> for Thread {}
573
574struct PendingCompletion {
575 id: usize,
576 _task: Task<()>,
577}