1use std::sync::Arc;
2
3use anyhow::Result;
4use assistant_tool::ToolWorkingSet;
5use chrono::{DateTime, Utc};
6use collections::{BTreeMap, HashMap, HashSet};
7use futures::future::Shared;
8use futures::{FutureExt as _, StreamExt as _};
9use gpui::{AppContext, EventEmitter, ModelContext, SharedString, Task};
10use language_model::{
11 LanguageModel, LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
12 LanguageModelRequestMessage, LanguageModelToolResult, LanguageModelToolUse,
13 LanguageModelToolUseId, MessageContent, Role, StopReason,
14};
15use language_models::provider::cloud::{MaxMonthlySpendReachedError, PaymentRequiredError};
16use serde::{Deserialize, Serialize};
17use util::{post_inc, TryFutureExt as _};
18use uuid::Uuid;
19
20use crate::context::{attach_context_to_message, ContextId, ContextSnapshot};
21
22#[derive(Debug, Clone, Copy)]
23pub enum RequestKind {
24 Chat,
25}
26
27#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Serialize, Deserialize)]
28pub struct ThreadId(Arc<str>);
29
30impl ThreadId {
31 pub fn new() -> Self {
32 Self(Uuid::new_v4().to_string().into())
33 }
34}
35
36impl std::fmt::Display for ThreadId {
37 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
38 write!(f, "{}", self.0)
39 }
40}
41
42#[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Clone, Copy, Serialize, Deserialize)]
43pub struct MessageId(usize);
44
45impl MessageId {
46 fn post_inc(&mut self) -> Self {
47 Self(post_inc(&mut self.0))
48 }
49}
50
51/// A message in a [`Thread`].
52#[derive(Debug, Clone)]
53pub struct Message {
54 pub id: MessageId,
55 pub role: Role,
56 pub text: String,
57}
58
59/// A thread of conversation with the LLM.
60pub struct Thread {
61 id: ThreadId,
62 updated_at: DateTime<Utc>,
63 summary: Option<SharedString>,
64 pending_summary: Task<Option<()>>,
65 messages: Vec<Message>,
66 next_message_id: MessageId,
67 context: BTreeMap<ContextId, ContextSnapshot>,
68 context_by_message: HashMap<MessageId, Vec<ContextId>>,
69 completion_count: usize,
70 pending_completions: Vec<PendingCompletion>,
71 tools: Arc<ToolWorkingSet>,
72 tool_uses_by_message: HashMap<MessageId, Vec<LanguageModelToolUse>>,
73 tool_results_by_message: HashMap<MessageId, Vec<LanguageModelToolResult>>,
74 pending_tool_uses_by_id: HashMap<LanguageModelToolUseId, PendingToolUse>,
75}
76
77impl Thread {
78 pub fn new(tools: Arc<ToolWorkingSet>, _cx: &mut ModelContext<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_uses_by_message: HashMap::default(),
92 tool_results_by_message: HashMap::default(),
93 pending_tool_uses_by_id: HashMap::default(),
94 }
95 }
96
97 pub fn id(&self) -> &ThreadId {
98 &self.id
99 }
100
101 pub fn is_empty(&self) -> bool {
102 self.messages.is_empty()
103 }
104
105 pub fn updated_at(&self) -> DateTime<Utc> {
106 self.updated_at
107 }
108
109 pub fn touch_updated_at(&mut self) {
110 self.updated_at = Utc::now();
111 }
112
113 pub fn summary(&self) -> Option<SharedString> {
114 self.summary.clone()
115 }
116
117 pub fn summary_or_default(&self) -> SharedString {
118 const DEFAULT: SharedString = SharedString::new_static("New Thread");
119 self.summary.clone().unwrap_or(DEFAULT)
120 }
121
122 pub fn set_summary(&mut self, summary: impl Into<SharedString>, cx: &mut ModelContext<Self>) {
123 self.summary = Some(summary.into());
124 cx.emit(ThreadEvent::SummaryChanged);
125 }
126
127 pub fn message(&self, id: MessageId) -> Option<&Message> {
128 self.messages.iter().find(|message| message.id == id)
129 }
130
131 pub fn messages(&self) -> impl Iterator<Item = &Message> {
132 self.messages.iter()
133 }
134
135 pub fn tools(&self) -> &Arc<ToolWorkingSet> {
136 &self.tools
137 }
138
139 pub fn context_for_message(&self, id: MessageId) -> Option<Vec<ContextSnapshot>> {
140 let context = self.context_by_message.get(&id)?;
141 Some(
142 context
143 .into_iter()
144 .filter_map(|context_id| self.context.get(&context_id))
145 .cloned()
146 .collect::<Vec<_>>(),
147 )
148 }
149
150 pub fn pending_tool_uses(&self) -> Vec<&PendingToolUse> {
151 self.pending_tool_uses_by_id.values().collect()
152 }
153
154 pub fn insert_user_message(
155 &mut self,
156 text: impl Into<String>,
157 context: Vec<ContextSnapshot>,
158 cx: &mut ModelContext<Self>,
159 ) {
160 let message_id = self.insert_message(Role::User, text, cx);
161 let context_ids = context.iter().map(|context| context.id).collect::<Vec<_>>();
162 self.context
163 .extend(context.into_iter().map(|context| (context.id, context)));
164 self.context_by_message.insert(message_id, context_ids);
165 }
166
167 pub fn insert_message(
168 &mut self,
169 role: Role,
170 text: impl Into<String>,
171 cx: &mut ModelContext<Self>,
172 ) -> MessageId {
173 let id = self.next_message_id.post_inc();
174 self.messages.push(Message {
175 id,
176 role,
177 text: text.into(),
178 });
179 self.touch_updated_at();
180 cx.emit(ThreadEvent::MessageAdded(id));
181 id
182 }
183
184 /// Returns the representation of this [`Thread`] in a textual form.
185 ///
186 /// This is the representation we use when attaching a thread as context to another thread.
187 pub fn text(&self) -> String {
188 let mut text = String::new();
189
190 for message in &self.messages {
191 text.push_str(match message.role {
192 language_model::Role::User => "User:",
193 language_model::Role::Assistant => "Assistant:",
194 language_model::Role::System => "System:",
195 });
196 text.push('\n');
197
198 text.push_str(&message.text);
199 text.push('\n');
200 }
201
202 text
203 }
204
205 pub fn to_completion_request(
206 &self,
207 _request_kind: RequestKind,
208 _cx: &AppContext,
209 ) -> LanguageModelRequest {
210 let mut request = LanguageModelRequest {
211 messages: vec![],
212 tools: Vec::new(),
213 stop: Vec::new(),
214 temperature: None,
215 };
216
217 let mut referenced_context_ids = HashSet::default();
218
219 for message in &self.messages {
220 if let Some(context_ids) = self.context_by_message.get(&message.id) {
221 referenced_context_ids.extend(context_ids);
222 }
223
224 let mut request_message = LanguageModelRequestMessage {
225 role: message.role,
226 content: Vec::new(),
227 cache: false,
228 };
229
230 if let Some(tool_results) = self.tool_results_by_message.get(&message.id) {
231 for tool_result in tool_results {
232 request_message
233 .content
234 .push(MessageContent::ToolResult(tool_result.clone()));
235 }
236 }
237
238 if !message.text.is_empty() {
239 request_message
240 .content
241 .push(MessageContent::Text(message.text.clone()));
242 }
243
244 if let Some(tool_uses) = self.tool_uses_by_message.get(&message.id) {
245 for tool_use in tool_uses {
246 request_message
247 .content
248 .push(MessageContent::ToolUse(tool_use.clone()));
249 }
250 }
251
252 request.messages.push(request_message);
253 }
254
255 if !referenced_context_ids.is_empty() {
256 let mut context_message = LanguageModelRequestMessage {
257 role: Role::User,
258 content: Vec::new(),
259 cache: false,
260 };
261
262 let referenced_context = referenced_context_ids
263 .into_iter()
264 .filter_map(|context_id| self.context.get(context_id))
265 .cloned();
266 attach_context_to_message(&mut context_message, referenced_context);
267
268 request.messages.push(context_message);
269 }
270
271 request
272 }
273
274 pub fn stream_completion(
275 &mut self,
276 request: LanguageModelRequest,
277 model: Arc<dyn LanguageModel>,
278 cx: &mut ModelContext<Self>,
279 ) {
280 let pending_completion_id = post_inc(&mut self.completion_count);
281
282 let task = cx.spawn(|thread, mut cx| async move {
283 let stream = model.stream_completion(request, &cx);
284 let stream_completion = async {
285 let mut events = stream.await?;
286 let mut stop_reason = StopReason::EndTurn;
287
288 while let Some(event) = events.next().await {
289 let event = event?;
290
291 thread.update(&mut cx, |thread, cx| {
292 match event {
293 LanguageModelCompletionEvent::StartMessage { .. } => {
294 thread.insert_message(Role::Assistant, String::new(), cx);
295 }
296 LanguageModelCompletionEvent::Stop(reason) => {
297 stop_reason = reason;
298 }
299 LanguageModelCompletionEvent::Text(chunk) => {
300 if let Some(last_message) = thread.messages.last_mut() {
301 if last_message.role == Role::Assistant {
302 last_message.text.push_str(&chunk);
303 cx.emit(ThreadEvent::StreamedAssistantText(
304 last_message.id,
305 chunk,
306 ));
307 }
308 }
309 }
310 LanguageModelCompletionEvent::ToolUse(tool_use) => {
311 if let Some(last_assistant_message) = thread
312 .messages
313 .iter()
314 .rfind(|message| message.role == Role::Assistant)
315 {
316 thread
317 .tool_uses_by_message
318 .entry(last_assistant_message.id)
319 .or_default()
320 .push(tool_use.clone());
321
322 thread.pending_tool_uses_by_id.insert(
323 tool_use.id.clone(),
324 PendingToolUse {
325 assistant_message_id: last_assistant_message.id,
326 id: tool_use.id,
327 name: tool_use.name,
328 input: tool_use.input,
329 status: PendingToolUseStatus::Idle,
330 },
331 );
332 }
333 }
334 }
335
336 thread.touch_updated_at();
337 cx.emit(ThreadEvent::StreamedCompletion);
338 cx.notify();
339 })?;
340
341 smol::future::yield_now().await;
342 }
343
344 thread.update(&mut cx, |thread, cx| {
345 thread
346 .pending_completions
347 .retain(|completion| completion.id != pending_completion_id);
348
349 if thread.summary.is_none() && thread.messages.len() >= 2 {
350 thread.summarize(cx);
351 }
352 })?;
353
354 anyhow::Ok(stop_reason)
355 };
356
357 let result = stream_completion.await;
358
359 thread
360 .update(&mut cx, |_thread, cx| match result.as_ref() {
361 Ok(stop_reason) => match stop_reason {
362 StopReason::ToolUse => {
363 cx.emit(ThreadEvent::UsePendingTools);
364 }
365 StopReason::EndTurn => {}
366 StopReason::MaxTokens => {}
367 },
368 Err(error) => {
369 if error.is::<PaymentRequiredError>() {
370 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
371 } else if error.is::<MaxMonthlySpendReachedError>() {
372 cx.emit(ThreadEvent::ShowError(ThreadError::MaxMonthlySpendReached));
373 } else {
374 let error_message = error
375 .chain()
376 .map(|err| err.to_string())
377 .collect::<Vec<_>>()
378 .join("\n");
379 cx.emit(ThreadEvent::ShowError(ThreadError::Message(
380 SharedString::from(error_message.clone()),
381 )));
382 }
383 }
384 })
385 .ok();
386 });
387
388 self.pending_completions.push(PendingCompletion {
389 id: pending_completion_id,
390 _task: task,
391 });
392 }
393
394 pub fn summarize(&mut self, cx: &mut ModelContext<Self>) {
395 let Some(provider) = LanguageModelRegistry::read_global(cx).active_provider() else {
396 return;
397 };
398 let Some(model) = LanguageModelRegistry::read_global(cx).active_model() else {
399 return;
400 };
401
402 if !provider.is_authenticated(cx) {
403 return;
404 }
405
406 let mut request = self.to_completion_request(RequestKind::Chat, cx);
407 request.messages.push(LanguageModelRequestMessage {
408 role: Role::User,
409 content: vec![
410 "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:`"
411 .into(),
412 ],
413 cache: false,
414 });
415
416 self.pending_summary = cx.spawn(|this, mut cx| {
417 async move {
418 let stream = model.stream_completion_text(request, &cx);
419 let mut messages = stream.await?;
420
421 let mut new_summary = String::new();
422 while let Some(message) = messages.stream.next().await {
423 let text = message?;
424 let mut lines = text.lines();
425 new_summary.extend(lines.next());
426
427 // Stop if the LLM generated multiple lines.
428 if lines.next().is_some() {
429 break;
430 }
431 }
432
433 this.update(&mut cx, |this, cx| {
434 if !new_summary.is_empty() {
435 this.summary = Some(new_summary.into());
436 }
437
438 cx.emit(ThreadEvent::SummaryChanged);
439 })?;
440
441 anyhow::Ok(())
442 }
443 .log_err()
444 });
445 }
446
447 pub fn insert_tool_output(
448 &mut self,
449 assistant_message_id: MessageId,
450 tool_use_id: LanguageModelToolUseId,
451 output: Task<Result<String>>,
452 cx: &mut ModelContext<Self>,
453 ) {
454 let insert_output_task = cx.spawn(|thread, mut cx| {
455 let tool_use_id = tool_use_id.clone();
456 async move {
457 let output = output.await;
458 thread
459 .update(&mut cx, |thread, cx| {
460 // The tool use was requested by an Assistant message,
461 // so we want to attach the tool results to the next
462 // user message.
463 let next_user_message = MessageId(assistant_message_id.0 + 1);
464
465 let tool_results = thread
466 .tool_results_by_message
467 .entry(next_user_message)
468 .or_default();
469
470 match output {
471 Ok(output) => {
472 tool_results.push(LanguageModelToolResult {
473 tool_use_id: tool_use_id.to_string(),
474 content: output,
475 is_error: false,
476 });
477
478 cx.emit(ThreadEvent::ToolFinished { tool_use_id });
479 }
480 Err(err) => {
481 tool_results.push(LanguageModelToolResult {
482 tool_use_id: tool_use_id.to_string(),
483 content: err.to_string(),
484 is_error: true,
485 });
486
487 if let Some(tool_use) =
488 thread.pending_tool_uses_by_id.get_mut(&tool_use_id)
489 {
490 tool_use.status = PendingToolUseStatus::Error(err.to_string());
491 }
492 }
493 }
494 })
495 .ok();
496 }
497 });
498
499 if let Some(tool_use) = self.pending_tool_uses_by_id.get_mut(&tool_use_id) {
500 tool_use.status = PendingToolUseStatus::Running {
501 _task: insert_output_task.shared(),
502 };
503 }
504 }
505}
506
507#[derive(Debug, Clone)]
508pub enum ThreadError {
509 PaymentRequired,
510 MaxMonthlySpendReached,
511 Message(SharedString),
512}
513
514#[derive(Debug, Clone)]
515pub enum ThreadEvent {
516 ShowError(ThreadError),
517 StreamedCompletion,
518 StreamedAssistantText(MessageId, String),
519 MessageAdded(MessageId),
520 SummaryChanged,
521 UsePendingTools,
522 ToolFinished {
523 #[allow(unused)]
524 tool_use_id: LanguageModelToolUseId,
525 },
526}
527
528impl EventEmitter<ThreadEvent> for Thread {}
529
530struct PendingCompletion {
531 id: usize,
532 _task: Task<()>,
533}
534
535#[derive(Debug, Clone)]
536pub struct PendingToolUse {
537 pub id: LanguageModelToolUseId,
538 /// The ID of the Assistant message in which the tool use was requested.
539 pub assistant_message_id: MessageId,
540 pub name: String,
541 pub input: serde_json::Value,
542 pub status: PendingToolUseStatus,
543}
544
545#[derive(Debug, Clone)]
546pub enum PendingToolUseStatus {
547 Idle,
548 Running { _task: Shared<Task<()>> },
549 Error(#[allow(unused)] String),
550}
551
552impl PendingToolUseStatus {
553 pub fn is_idle(&self) -> bool {
554 matches!(self, PendingToolUseStatus::Idle)
555 }
556}