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