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};
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_by_message: HashMap<MessageId, Vec<Context>>,
68 completion_count: usize,
69 pending_completions: Vec<PendingCompletion>,
70 tools: Arc<ToolWorkingSet>,
71 tool_uses_by_message: HashMap<MessageId, Vec<LanguageModelToolUse>>,
72 tool_results_by_message: HashMap<MessageId, Vec<LanguageModelToolResult>>,
73 pending_tool_uses_by_id: HashMap<LanguageModelToolUseId, PendingToolUse>,
74}
75
76impl Thread {
77 pub fn new(tools: Arc<ToolWorkingSet>, _cx: &mut ModelContext<Self>) -> Self {
78 Self {
79 id: ThreadId::new(),
80 updated_at: Utc::now(),
81 summary: None,
82 pending_summary: Task::ready(None),
83 messages: Vec::new(),
84 next_message_id: MessageId(0),
85 context_by_message: HashMap::default(),
86 completion_count: 0,
87 pending_completions: Vec::new(),
88 tools,
89 tool_uses_by_message: HashMap::default(),
90 tool_results_by_message: HashMap::default(),
91 pending_tool_uses_by_id: HashMap::default(),
92 }
93 }
94
95 pub fn id(&self) -> &ThreadId {
96 &self.id
97 }
98
99 pub fn is_empty(&self) -> bool {
100 self.messages.is_empty()
101 }
102
103 pub fn updated_at(&self) -> DateTime<Utc> {
104 self.updated_at
105 }
106
107 pub fn touch_updated_at(&mut self) {
108 self.updated_at = Utc::now();
109 }
110
111 pub fn summary(&self) -> Option<SharedString> {
112 self.summary.clone()
113 }
114
115 pub fn set_summary(&mut self, summary: impl Into<SharedString>, cx: &mut ModelContext<Self>) {
116 self.summary = Some(summary.into());
117 cx.emit(ThreadEvent::SummaryChanged);
118 }
119
120 pub fn message(&self, id: MessageId) -> Option<&Message> {
121 self.messages.iter().find(|message| message.id == id)
122 }
123
124 pub fn messages(&self) -> impl Iterator<Item = &Message> {
125 self.messages.iter()
126 }
127
128 pub fn tools(&self) -> &Arc<ToolWorkingSet> {
129 &self.tools
130 }
131
132 pub fn context_for_message(&self, id: MessageId) -> Option<&Vec<Context>> {
133 self.context_by_message.get(&id)
134 }
135
136 pub fn pending_tool_uses(&self) -> Vec<&PendingToolUse> {
137 self.pending_tool_uses_by_id.values().collect()
138 }
139
140 pub fn insert_user_message(
141 &mut self,
142 text: impl Into<String>,
143 context: Vec<Context>,
144 cx: &mut ModelContext<Self>,
145 ) {
146 let message_id = self.insert_message(Role::User, text, cx);
147 self.context_by_message.insert(message_id, context);
148 }
149
150 pub fn insert_message(
151 &mut self,
152 role: Role,
153 text: impl Into<String>,
154 cx: &mut ModelContext<Self>,
155 ) -> MessageId {
156 let id = self.next_message_id.post_inc();
157 self.messages.push(Message {
158 id,
159 role,
160 text: text.into(),
161 });
162 self.touch_updated_at();
163 cx.emit(ThreadEvent::MessageAdded(id));
164 id
165 }
166
167 pub fn to_completion_request(
168 &self,
169 _request_kind: RequestKind,
170 _cx: &AppContext,
171 ) -> LanguageModelRequest {
172 let mut request = LanguageModelRequest {
173 messages: vec![],
174 tools: Vec::new(),
175 stop: Vec::new(),
176 temperature: None,
177 };
178
179 for message in &self.messages {
180 let mut request_message = LanguageModelRequestMessage {
181 role: message.role,
182 content: Vec::new(),
183 cache: false,
184 };
185
186 if let Some(tool_results) = self.tool_results_by_message.get(&message.id) {
187 for tool_result in tool_results {
188 request_message
189 .content
190 .push(MessageContent::ToolResult(tool_result.clone()));
191 }
192 }
193
194 if let Some(context) = self.context_for_message(message.id) {
195 attach_context_to_message(&mut request_message, context.clone());
196 }
197
198 if !message.text.is_empty() {
199 request_message
200 .content
201 .push(MessageContent::Text(message.text.clone()));
202 }
203
204 if let Some(tool_uses) = self.tool_uses_by_message.get(&message.id) {
205 for tool_use in tool_uses {
206 request_message
207 .content
208 .push(MessageContent::ToolUse(tool_use.clone()));
209 }
210 }
211
212 request.messages.push(request_message);
213 }
214
215 request
216 }
217
218 pub fn stream_completion(
219 &mut self,
220 request: LanguageModelRequest,
221 model: Arc<dyn LanguageModel>,
222 cx: &mut ModelContext<Self>,
223 ) {
224 let pending_completion_id = post_inc(&mut self.completion_count);
225
226 let task = cx.spawn(|thread, mut cx| async move {
227 let stream = model.stream_completion(request, &cx);
228 let stream_completion = async {
229 let mut events = stream.await?;
230 let mut stop_reason = StopReason::EndTurn;
231
232 while let Some(event) = events.next().await {
233 let event = event?;
234
235 thread.update(&mut cx, |thread, cx| {
236 match event {
237 LanguageModelCompletionEvent::StartMessage { .. } => {
238 thread.insert_message(Role::Assistant, String::new(), cx);
239 }
240 LanguageModelCompletionEvent::Stop(reason) => {
241 stop_reason = reason;
242 }
243 LanguageModelCompletionEvent::Text(chunk) => {
244 if let Some(last_message) = thread.messages.last_mut() {
245 if last_message.role == Role::Assistant {
246 last_message.text.push_str(&chunk);
247 cx.emit(ThreadEvent::StreamedAssistantText(
248 last_message.id,
249 chunk,
250 ));
251 }
252 }
253 }
254 LanguageModelCompletionEvent::ToolUse(tool_use) => {
255 if let Some(last_assistant_message) = thread
256 .messages
257 .iter()
258 .rfind(|message| message.role == Role::Assistant)
259 {
260 thread
261 .tool_uses_by_message
262 .entry(last_assistant_message.id)
263 .or_default()
264 .push(tool_use.clone());
265
266 thread.pending_tool_uses_by_id.insert(
267 tool_use.id.clone(),
268 PendingToolUse {
269 assistant_message_id: last_assistant_message.id,
270 id: tool_use.id,
271 name: tool_use.name,
272 input: tool_use.input,
273 status: PendingToolUseStatus::Idle,
274 },
275 );
276 }
277 }
278 }
279
280 thread.touch_updated_at();
281 cx.emit(ThreadEvent::StreamedCompletion);
282 cx.notify();
283 })?;
284
285 smol::future::yield_now().await;
286 }
287
288 thread.update(&mut cx, |thread, cx| {
289 thread
290 .pending_completions
291 .retain(|completion| completion.id != pending_completion_id);
292
293 if thread.summary.is_none() && thread.messages.len() >= 2 {
294 thread.summarize(cx);
295 }
296 })?;
297
298 anyhow::Ok(stop_reason)
299 };
300
301 let result = stream_completion.await;
302
303 thread
304 .update(&mut cx, |_thread, cx| match result.as_ref() {
305 Ok(stop_reason) => match stop_reason {
306 StopReason::ToolUse => {
307 cx.emit(ThreadEvent::UsePendingTools);
308 }
309 StopReason::EndTurn => {}
310 StopReason::MaxTokens => {}
311 },
312 Err(error) => {
313 if error.is::<PaymentRequiredError>() {
314 cx.emit(ThreadEvent::ShowError(ThreadError::PaymentRequired));
315 } else if error.is::<MaxMonthlySpendReachedError>() {
316 cx.emit(ThreadEvent::ShowError(ThreadError::MaxMonthlySpendReached));
317 } else {
318 let error_message = error
319 .chain()
320 .map(|err| err.to_string())
321 .collect::<Vec<_>>()
322 .join("\n");
323 cx.emit(ThreadEvent::ShowError(ThreadError::Message(
324 SharedString::from(error_message.clone()),
325 )));
326 }
327 }
328 })
329 .ok();
330 });
331
332 self.pending_completions.push(PendingCompletion {
333 id: pending_completion_id,
334 _task: task,
335 });
336 }
337
338 pub fn summarize(&mut self, cx: &mut ModelContext<Self>) {
339 let Some(provider) = LanguageModelRegistry::read_global(cx).active_provider() else {
340 return;
341 };
342 let Some(model) = LanguageModelRegistry::read_global(cx).active_model() else {
343 return;
344 };
345
346 if !provider.is_authenticated(cx) {
347 return;
348 }
349
350 let mut request = self.to_completion_request(RequestKind::Chat, cx);
351 request.messages.push(LanguageModelRequestMessage {
352 role: Role::User,
353 content: vec![
354 "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:`"
355 .into(),
356 ],
357 cache: false,
358 });
359
360 self.pending_summary = cx.spawn(|this, mut cx| {
361 async move {
362 let stream = model.stream_completion_text(request, &cx);
363 let mut messages = stream.await?;
364
365 let mut new_summary = String::new();
366 while let Some(message) = messages.stream.next().await {
367 let text = message?;
368 let mut lines = text.lines();
369 new_summary.extend(lines.next());
370
371 // Stop if the LLM generated multiple lines.
372 if lines.next().is_some() {
373 break;
374 }
375 }
376
377 this.update(&mut cx, |this, cx| {
378 if !new_summary.is_empty() {
379 this.summary = Some(new_summary.into());
380 }
381
382 cx.emit(ThreadEvent::SummaryChanged);
383 })?;
384
385 anyhow::Ok(())
386 }
387 .log_err()
388 });
389 }
390
391 pub fn insert_tool_output(
392 &mut self,
393 assistant_message_id: MessageId,
394 tool_use_id: LanguageModelToolUseId,
395 output: Task<Result<String>>,
396 cx: &mut ModelContext<Self>,
397 ) {
398 let insert_output_task = cx.spawn(|thread, mut cx| {
399 let tool_use_id = tool_use_id.clone();
400 async move {
401 let output = output.await;
402 thread
403 .update(&mut cx, |thread, cx| {
404 // The tool use was requested by an Assistant message,
405 // so we want to attach the tool results to the next
406 // user message.
407 let next_user_message = MessageId(assistant_message_id.0 + 1);
408
409 let tool_results = thread
410 .tool_results_by_message
411 .entry(next_user_message)
412 .or_default();
413
414 match output {
415 Ok(output) => {
416 tool_results.push(LanguageModelToolResult {
417 tool_use_id: tool_use_id.to_string(),
418 content: output,
419 is_error: false,
420 });
421
422 cx.emit(ThreadEvent::ToolFinished { tool_use_id });
423 }
424 Err(err) => {
425 tool_results.push(LanguageModelToolResult {
426 tool_use_id: tool_use_id.to_string(),
427 content: err.to_string(),
428 is_error: true,
429 });
430
431 if let Some(tool_use) =
432 thread.pending_tool_uses_by_id.get_mut(&tool_use_id)
433 {
434 tool_use.status = PendingToolUseStatus::Error(err.to_string());
435 }
436 }
437 }
438 })
439 .ok();
440 }
441 });
442
443 if let Some(tool_use) = self.pending_tool_uses_by_id.get_mut(&tool_use_id) {
444 tool_use.status = PendingToolUseStatus::Running {
445 _task: insert_output_task.shared(),
446 };
447 }
448 }
449}
450
451#[derive(Debug, Clone)]
452pub enum ThreadError {
453 PaymentRequired,
454 MaxMonthlySpendReached,
455 Message(SharedString),
456}
457
458#[derive(Debug, Clone)]
459pub enum ThreadEvent {
460 ShowError(ThreadError),
461 StreamedCompletion,
462 StreamedAssistantText(MessageId, String),
463 MessageAdded(MessageId),
464 SummaryChanged,
465 UsePendingTools,
466 ToolFinished {
467 #[allow(unused)]
468 tool_use_id: LanguageModelToolUseId,
469 },
470}
471
472impl EventEmitter<ThreadEvent> for Thread {}
473
474struct PendingCompletion {
475 id: usize,
476 _task: Task<()>,
477}
478
479#[derive(Debug, Clone)]
480pub struct PendingToolUse {
481 pub id: LanguageModelToolUseId,
482 /// The ID of the Assistant message in which the tool use was requested.
483 pub assistant_message_id: MessageId,
484 pub name: String,
485 pub input: serde_json::Value,
486 pub status: PendingToolUseStatus,
487}
488
489#[derive(Debug, Clone)]
490pub enum PendingToolUseStatus {
491 Idle,
492 Running { _task: Shared<Task<()>> },
493 Error(#[allow(unused)] String),
494}
495
496impl PendingToolUseStatus {
497 pub fn is_idle(&self) -> bool {
498 matches!(self, PendingToolUseStatus::Idle)
499 }
500}