1use std::sync::Arc;
2
3use anyhow::Result;
4use assistant_tool::ToolWorkingSet;
5use chrono::{DateTime, Utc};
6use collections::HashMap;
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 for message in &self.messages {
213 let mut request_message = LanguageModelRequestMessage {
214 role: message.role,
215 content: Vec::new(),
216 cache: false,
217 };
218
219 if let Some(tool_results) = self.tool_results_by_message.get(&message.id) {
220 for tool_result in tool_results {
221 request_message
222 .content
223 .push(MessageContent::ToolResult(tool_result.clone()));
224 }
225 }
226
227 if let Some(context) = self.context_for_message(message.id) {
228 attach_context_to_message(&mut request_message, context.clone());
229 }
230
231 if !message.text.is_empty() {
232 request_message
233 .content
234 .push(MessageContent::Text(message.text.clone()));
235 }
236
237 if let Some(tool_uses) = self.tool_uses_by_message.get(&message.id) {
238 for tool_use in tool_uses {
239 request_message
240 .content
241 .push(MessageContent::ToolUse(tool_use.clone()));
242 }
243 }
244
245 request.messages.push(request_message);
246 }
247
248 request
249 }
250
251 pub fn stream_completion(
252 &mut self,
253 request: LanguageModelRequest,
254 model: Arc<dyn LanguageModel>,
255 cx: &mut ModelContext<Self>,
256 ) {
257 let pending_completion_id = post_inc(&mut self.completion_count);
258
259 let task = cx.spawn(|thread, mut cx| async move {
260 let stream = model.stream_completion(request, &cx);
261 let stream_completion = async {
262 let mut events = stream.await?;
263 let mut stop_reason = StopReason::EndTurn;
264
265 while let Some(event) = events.next().await {
266 let event = event?;
267
268 thread.update(&mut cx, |thread, cx| {
269 match event {
270 LanguageModelCompletionEvent::StartMessage { .. } => {
271 thread.insert_message(Role::Assistant, String::new(), cx);
272 }
273 LanguageModelCompletionEvent::Stop(reason) => {
274 stop_reason = reason;
275 }
276 LanguageModelCompletionEvent::Text(chunk) => {
277 if let Some(last_message) = thread.messages.last_mut() {
278 if last_message.role == Role::Assistant {
279 last_message.text.push_str(&chunk);
280 cx.emit(ThreadEvent::StreamedAssistantText(
281 last_message.id,
282 chunk,
283 ));
284 }
285 }
286 }
287 LanguageModelCompletionEvent::ToolUse(tool_use) => {
288 if let Some(last_assistant_message) = thread
289 .messages
290 .iter()
291 .rfind(|message| message.role == Role::Assistant)
292 {
293 thread
294 .tool_uses_by_message
295 .entry(last_assistant_message.id)
296 .or_default()
297 .push(tool_use.clone());
298
299 thread.pending_tool_uses_by_id.insert(
300 tool_use.id.clone(),
301 PendingToolUse {
302 assistant_message_id: last_assistant_message.id,
303 id: tool_use.id,
304 name: tool_use.name,
305 input: tool_use.input,
306 status: PendingToolUseStatus::Idle,
307 },
308 );
309 }
310 }
311 }
312
313 thread.touch_updated_at();
314 cx.emit(ThreadEvent::StreamedCompletion);
315 cx.notify();
316 })?;
317
318 smol::future::yield_now().await;
319 }
320
321 thread.update(&mut cx, |thread, cx| {
322 thread
323 .pending_completions
324 .retain(|completion| completion.id != pending_completion_id);
325
326 if thread.summary.is_none() && thread.messages.len() >= 2 {
327 thread.summarize(cx);
328 }
329 })?;
330
331 anyhow::Ok(stop_reason)
332 };
333
334 let result = stream_completion.await;
335
336 thread
337 .update(&mut cx, |_thread, cx| match result.as_ref() {
338 Ok(stop_reason) => match stop_reason {
339 StopReason::ToolUse => {
340 cx.emit(ThreadEvent::UsePendingTools);
341 }
342 StopReason::EndTurn => {}
343 StopReason::MaxTokens => {}
344 },
345 Err(error) => {
346 if error.is::<PaymentRequiredError>() {
347 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
348 } else if error.is::<MaxMonthlySpendReachedError>() {
349 cx.emit(ThreadEvent::ShowError(ThreadError::MaxMonthlySpendReached));
350 } else {
351 let error_message = error
352 .chain()
353 .map(|err| err.to_string())
354 .collect::<Vec<_>>()
355 .join("\n");
356 cx.emit(ThreadEvent::ShowError(ThreadError::Message(
357 SharedString::from(error_message.clone()),
358 )));
359 }
360 }
361 })
362 .ok();
363 });
364
365 self.pending_completions.push(PendingCompletion {
366 id: pending_completion_id,
367 _task: task,
368 });
369 }
370
371 pub fn summarize(&mut self, cx: &mut ModelContext<Self>) {
372 let Some(provider) = LanguageModelRegistry::read_global(cx).active_provider() else {
373 return;
374 };
375 let Some(model) = LanguageModelRegistry::read_global(cx).active_model() else {
376 return;
377 };
378
379 if !provider.is_authenticated(cx) {
380 return;
381 }
382
383 let mut request = self.to_completion_request(RequestKind::Chat, cx);
384 request.messages.push(LanguageModelRequestMessage {
385 role: Role::User,
386 content: vec![
387 "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:`"
388 .into(),
389 ],
390 cache: false,
391 });
392
393 self.pending_summary = cx.spawn(|this, mut cx| {
394 async move {
395 let stream = model.stream_completion_text(request, &cx);
396 let mut messages = stream.await?;
397
398 let mut new_summary = String::new();
399 while let Some(message) = messages.stream.next().await {
400 let text = message?;
401 let mut lines = text.lines();
402 new_summary.extend(lines.next());
403
404 // Stop if the LLM generated multiple lines.
405 if lines.next().is_some() {
406 break;
407 }
408 }
409
410 this.update(&mut cx, |this, cx| {
411 if !new_summary.is_empty() {
412 this.summary = Some(new_summary.into());
413 }
414
415 cx.emit(ThreadEvent::SummaryChanged);
416 })?;
417
418 anyhow::Ok(())
419 }
420 .log_err()
421 });
422 }
423
424 pub fn insert_tool_output(
425 &mut self,
426 assistant_message_id: MessageId,
427 tool_use_id: LanguageModelToolUseId,
428 output: Task<Result<String>>,
429 cx: &mut ModelContext<Self>,
430 ) {
431 let insert_output_task = cx.spawn(|thread, mut cx| {
432 let tool_use_id = tool_use_id.clone();
433 async move {
434 let output = output.await;
435 thread
436 .update(&mut cx, |thread, cx| {
437 // The tool use was requested by an Assistant message,
438 // so we want to attach the tool results to the next
439 // user message.
440 let next_user_message = MessageId(assistant_message_id.0 + 1);
441
442 let tool_results = thread
443 .tool_results_by_message
444 .entry(next_user_message)
445 .or_default();
446
447 match output {
448 Ok(output) => {
449 tool_results.push(LanguageModelToolResult {
450 tool_use_id: tool_use_id.to_string(),
451 content: output,
452 is_error: false,
453 });
454
455 cx.emit(ThreadEvent::ToolFinished { tool_use_id });
456 }
457 Err(err) => {
458 tool_results.push(LanguageModelToolResult {
459 tool_use_id: tool_use_id.to_string(),
460 content: err.to_string(),
461 is_error: true,
462 });
463
464 if let Some(tool_use) =
465 thread.pending_tool_uses_by_id.get_mut(&tool_use_id)
466 {
467 tool_use.status = PendingToolUseStatus::Error(err.to_string());
468 }
469 }
470 }
471 })
472 .ok();
473 }
474 });
475
476 if let Some(tool_use) = self.pending_tool_uses_by_id.get_mut(&tool_use_id) {
477 tool_use.status = PendingToolUseStatus::Running {
478 _task: insert_output_task.shared(),
479 };
480 }
481 }
482}
483
484#[derive(Debug, Clone)]
485pub enum ThreadError {
486 PaymentRequired,
487 MaxMonthlySpendReached,
488 Message(SharedString),
489}
490
491#[derive(Debug, Clone)]
492pub enum ThreadEvent {
493 ShowError(ThreadError),
494 StreamedCompletion,
495 StreamedAssistantText(MessageId, String),
496 MessageAdded(MessageId),
497 SummaryChanged,
498 UsePendingTools,
499 ToolFinished {
500 #[allow(unused)]
501 tool_use_id: LanguageModelToolUseId,
502 },
503}
504
505impl EventEmitter<ThreadEvent> for Thread {}
506
507struct PendingCompletion {
508 id: usize,
509 _task: Task<()>,
510}
511
512#[derive(Debug, Clone)]
513pub struct PendingToolUse {
514 pub id: LanguageModelToolUseId,
515 /// The ID of the Assistant message in which the tool use was requested.
516 pub assistant_message_id: MessageId,
517 pub name: String,
518 pub input: serde_json::Value,
519 pub status: PendingToolUseStatus,
520}
521
522#[derive(Debug, Clone)]
523pub enum PendingToolUseStatus {
524 Idle,
525 Running { _task: Shared<Task<()>> },
526 Error(#[allow(unused)] String),
527}
528
529impl PendingToolUseStatus {
530 pub fn is_idle(&self) -> bool {
531 matches!(self, PendingToolUseStatus::Idle)
532 }
533}