1use std::path::PathBuf;
2use std::sync::Arc;
3use std::sync::OnceLock;
4
5use anyhow::Context as _;
6use anyhow::{Result, anyhow};
7use chrono::DateTime;
8use collections::HashSet;
9use fs::Fs;
10use futures::{AsyncBufReadExt, AsyncReadExt, StreamExt, io::BufReader, stream::BoxStream};
11use gpui::{App, AsyncApp, Global, prelude::*};
12use http_client::{AsyncBody, HttpClient, Method, Request as HttpRequest};
13use itertools::Itertools;
14use paths::home_dir;
15use serde::{Deserialize, Serialize};
16use settings::watch_config_dir;
17
18pub const COPILOT_CHAT_COMPLETION_URL: &str = "https://api.githubcopilot.com/chat/completions";
19pub const COPILOT_CHAT_AUTH_URL: &str = "https://api.github.com/copilot_internal/v2/token";
20pub const COPILOT_CHAT_MODELS_URL: &str = "https://api.githubcopilot.com/models";
21
22// Copilot's base model; defined by Microsoft in premium requests table
23// This will be moved to the front of the Copilot model list, and will be used for
24// 'fast' requests (e.g. title generation)
25// https://docs.github.com/en/copilot/managing-copilot/monitoring-usage-and-entitlements/about-premium-requests
26const DEFAULT_MODEL_ID: &str = "gpt-4.1";
27
28#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
29#[serde(rename_all = "lowercase")]
30pub enum Role {
31 User,
32 Assistant,
33 System,
34}
35
36#[derive(Deserialize)]
37struct ModelSchema {
38 #[serde(deserialize_with = "deserialize_models_skip_errors")]
39 data: Vec<Model>,
40}
41
42fn deserialize_models_skip_errors<'de, D>(deserializer: D) -> Result<Vec<Model>, D::Error>
43where
44 D: serde::Deserializer<'de>,
45{
46 let raw_values = Vec::<serde_json::Value>::deserialize(deserializer)?;
47 let models = raw_values
48 .into_iter()
49 .filter_map(|value| match serde_json::from_value::<Model>(value) {
50 Ok(model) => Some(model),
51 Err(err) => {
52 log::warn!("GitHub Copilot Chat model failed to deserialize: {:?}", err);
53 None
54 }
55 })
56 .collect();
57
58 Ok(models)
59}
60
61#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
62pub struct Model {
63 capabilities: ModelCapabilities,
64 id: String,
65 name: String,
66 policy: Option<ModelPolicy>,
67 vendor: ModelVendor,
68 model_picker_enabled: bool,
69}
70
71#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
72struct ModelCapabilities {
73 family: String,
74 #[serde(default)]
75 limits: ModelLimits,
76 supports: ModelSupportedFeatures,
77}
78
79#[derive(Default, Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
80struct ModelLimits {
81 #[serde(default)]
82 max_context_window_tokens: usize,
83 #[serde(default)]
84 max_output_tokens: usize,
85 #[serde(default)]
86 max_prompt_tokens: usize,
87}
88
89#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
90struct ModelPolicy {
91 state: String,
92}
93
94#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
95struct ModelSupportedFeatures {
96 #[serde(default)]
97 streaming: bool,
98 #[serde(default)]
99 tool_calls: bool,
100 #[serde(default)]
101 parallel_tool_calls: bool,
102 #[serde(default)]
103 vision: bool,
104}
105
106#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
107pub enum ModelVendor {
108 // Azure OpenAI should have no functional difference from OpenAI in Copilot Chat
109 #[serde(alias = "Azure OpenAI")]
110 OpenAI,
111 Google,
112 Anthropic,
113}
114
115#[derive(Serialize, Deserialize, Debug, Eq, PartialEq, Clone)]
116#[serde(tag = "type")]
117pub enum ChatMessagePart {
118 #[serde(rename = "text")]
119 Text { text: String },
120 #[serde(rename = "image_url")]
121 Image { image_url: ImageUrl },
122}
123
124#[derive(Serialize, Deserialize, Debug, Eq, PartialEq, Clone)]
125pub struct ImageUrl {
126 pub url: String,
127}
128
129impl Model {
130 pub fn uses_streaming(&self) -> bool {
131 self.capabilities.supports.streaming
132 }
133
134 pub fn id(&self) -> &str {
135 self.id.as_str()
136 }
137
138 pub fn display_name(&self) -> &str {
139 self.name.as_str()
140 }
141
142 pub fn max_token_count(&self) -> usize {
143 self.capabilities.limits.max_prompt_tokens
144 }
145
146 pub fn supports_tools(&self) -> bool {
147 self.capabilities.supports.tool_calls
148 }
149
150 pub fn vendor(&self) -> ModelVendor {
151 self.vendor
152 }
153
154 pub fn supports_vision(&self) -> bool {
155 self.capabilities.supports.vision
156 }
157
158 pub fn supports_parallel_tool_calls(&self) -> bool {
159 self.capabilities.supports.parallel_tool_calls
160 }
161}
162
163#[derive(Serialize, Deserialize)]
164pub struct Request {
165 pub intent: bool,
166 pub n: usize,
167 pub stream: bool,
168 pub temperature: f32,
169 pub model: String,
170 pub messages: Vec<ChatMessage>,
171 #[serde(default, skip_serializing_if = "Vec::is_empty")]
172 pub tools: Vec<Tool>,
173 #[serde(default, skip_serializing_if = "Option::is_none")]
174 pub tool_choice: Option<ToolChoice>,
175}
176
177#[derive(Serialize, Deserialize)]
178pub struct Function {
179 pub name: String,
180 pub description: String,
181 pub parameters: serde_json::Value,
182}
183
184#[derive(Serialize, Deserialize)]
185#[serde(tag = "type", rename_all = "snake_case")]
186pub enum Tool {
187 Function { function: Function },
188}
189
190#[derive(Serialize, Deserialize)]
191#[serde(rename_all = "lowercase")]
192pub enum ToolChoice {
193 Auto,
194 Any,
195 None,
196}
197
198#[derive(Serialize, Deserialize, Debug)]
199#[serde(tag = "role", rename_all = "lowercase")]
200pub enum ChatMessage {
201 Assistant {
202 content: ChatMessageContent,
203 #[serde(default, skip_serializing_if = "Vec::is_empty")]
204 tool_calls: Vec<ToolCall>,
205 },
206 User {
207 content: ChatMessageContent,
208 },
209 System {
210 content: String,
211 },
212 Tool {
213 content: ChatMessageContent,
214 tool_call_id: String,
215 },
216}
217
218#[derive(Debug, Serialize, Deserialize)]
219#[serde(untagged)]
220pub enum ChatMessageContent {
221 Plain(String),
222 Multipart(Vec<ChatMessagePart>),
223}
224
225impl ChatMessageContent {
226 pub fn empty() -> Self {
227 ChatMessageContent::Multipart(vec![])
228 }
229}
230
231impl From<Vec<ChatMessagePart>> for ChatMessageContent {
232 fn from(mut parts: Vec<ChatMessagePart>) -> Self {
233 if let [ChatMessagePart::Text { text }] = parts.as_mut_slice() {
234 ChatMessageContent::Plain(std::mem::take(text))
235 } else {
236 ChatMessageContent::Multipart(parts)
237 }
238 }
239}
240
241impl From<String> for ChatMessageContent {
242 fn from(text: String) -> Self {
243 ChatMessageContent::Plain(text)
244 }
245}
246
247#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
248pub struct ToolCall {
249 pub id: String,
250 #[serde(flatten)]
251 pub content: ToolCallContent,
252}
253
254#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
255#[serde(tag = "type", rename_all = "lowercase")]
256pub enum ToolCallContent {
257 Function { function: FunctionContent },
258}
259
260#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
261pub struct FunctionContent {
262 pub name: String,
263 pub arguments: String,
264}
265
266#[derive(Deserialize, Debug)]
267#[serde(tag = "type", rename_all = "snake_case")]
268pub struct ResponseEvent {
269 pub choices: Vec<ResponseChoice>,
270 pub id: String,
271}
272
273#[derive(Debug, Deserialize)]
274pub struct ResponseChoice {
275 pub index: usize,
276 pub finish_reason: Option<String>,
277 pub delta: Option<ResponseDelta>,
278 pub message: Option<ResponseDelta>,
279}
280
281#[derive(Debug, Deserialize)]
282pub struct ResponseDelta {
283 pub content: Option<String>,
284 pub role: Option<Role>,
285 #[serde(default)]
286 pub tool_calls: Vec<ToolCallChunk>,
287}
288
289#[derive(Deserialize, Debug, Eq, PartialEq)]
290pub struct ToolCallChunk {
291 pub index: usize,
292 pub id: Option<String>,
293 pub function: Option<FunctionChunk>,
294}
295
296#[derive(Deserialize, Debug, Eq, PartialEq)]
297pub struct FunctionChunk {
298 pub name: Option<String>,
299 pub arguments: Option<String>,
300}
301
302#[derive(Deserialize)]
303struct ApiTokenResponse {
304 token: String,
305 expires_at: i64,
306}
307
308#[derive(Clone)]
309struct ApiToken {
310 api_key: String,
311 expires_at: DateTime<chrono::Utc>,
312}
313
314impl ApiToken {
315 pub fn remaining_seconds(&self) -> i64 {
316 self.expires_at
317 .timestamp()
318 .saturating_sub(chrono::Utc::now().timestamp())
319 }
320}
321
322impl TryFrom<ApiTokenResponse> for ApiToken {
323 type Error = anyhow::Error;
324
325 fn try_from(response: ApiTokenResponse) -> Result<Self, Self::Error> {
326 let expires_at =
327 DateTime::from_timestamp(response.expires_at, 0).context("invalid expires_at")?;
328
329 Ok(Self {
330 api_key: response.token,
331 expires_at,
332 })
333 }
334}
335
336struct GlobalCopilotChat(gpui::Entity<CopilotChat>);
337
338impl Global for GlobalCopilotChat {}
339
340pub struct CopilotChat {
341 oauth_token: Option<String>,
342 api_token: Option<ApiToken>,
343 models: Option<Vec<Model>>,
344 client: Arc<dyn HttpClient>,
345}
346
347pub fn init(fs: Arc<dyn Fs>, client: Arc<dyn HttpClient>, cx: &mut App) {
348 let copilot_chat = cx.new(|cx| CopilotChat::new(fs, client, cx));
349 cx.set_global(GlobalCopilotChat(copilot_chat));
350}
351
352pub fn copilot_chat_config_dir() -> &'static PathBuf {
353 static COPILOT_CHAT_CONFIG_DIR: OnceLock<PathBuf> = OnceLock::new();
354
355 COPILOT_CHAT_CONFIG_DIR.get_or_init(|| {
356 if cfg!(target_os = "windows") {
357 home_dir().join("AppData").join("Local")
358 } else {
359 home_dir().join(".config")
360 }
361 .join("github-copilot")
362 })
363}
364
365fn copilot_chat_config_paths() -> [PathBuf; 2] {
366 let base_dir = copilot_chat_config_dir();
367 [base_dir.join("hosts.json"), base_dir.join("apps.json")]
368}
369
370impl CopilotChat {
371 pub fn global(cx: &App) -> Option<gpui::Entity<Self>> {
372 cx.try_global::<GlobalCopilotChat>()
373 .map(|model| model.0.clone())
374 }
375
376 pub fn new(fs: Arc<dyn Fs>, client: Arc<dyn HttpClient>, cx: &App) -> Self {
377 let config_paths: HashSet<PathBuf> = copilot_chat_config_paths().into_iter().collect();
378 let dir_path = copilot_chat_config_dir();
379
380 cx.spawn({
381 let client = client.clone();
382 async move |cx| {
383 let mut parent_watch_rx = watch_config_dir(
384 cx.background_executor(),
385 fs.clone(),
386 dir_path.clone(),
387 config_paths,
388 );
389 while let Some(contents) = parent_watch_rx.next().await {
390 let oauth_token = extract_oauth_token(contents);
391 cx.update(|cx| {
392 if let Some(this) = Self::global(cx).as_ref() {
393 this.update(cx, |this, cx| {
394 this.oauth_token = oauth_token.clone();
395 cx.notify();
396 });
397 }
398 })?;
399
400 if let Some(ref oauth_token) = oauth_token {
401 let api_token = request_api_token(oauth_token, client.clone()).await?;
402 cx.update(|cx| {
403 if let Some(this) = Self::global(cx).as_ref() {
404 this.update(cx, |this, cx| {
405 this.api_token = Some(api_token.clone());
406 cx.notify();
407 });
408 }
409 })?;
410 let models = get_models(api_token.api_key, client.clone()).await?;
411 cx.update(|cx| {
412 if let Some(this) = Self::global(cx).as_ref() {
413 this.update(cx, |this, cx| {
414 this.models = Some(models);
415 cx.notify();
416 });
417 }
418 })?;
419 }
420 }
421 anyhow::Ok(())
422 }
423 })
424 .detach_and_log_err(cx);
425
426 Self {
427 oauth_token: None,
428 api_token: None,
429 models: None,
430 client,
431 }
432 }
433
434 pub fn is_authenticated(&self) -> bool {
435 self.oauth_token.is_some()
436 }
437
438 pub fn models(&self) -> Option<&[Model]> {
439 self.models.as_deref()
440 }
441
442 pub async fn stream_completion(
443 request: Request,
444 mut cx: AsyncApp,
445 ) -> Result<BoxStream<'static, Result<ResponseEvent>>> {
446 let this = cx
447 .update(|cx| Self::global(cx))
448 .ok()
449 .flatten()
450 .context("Copilot chat is not enabled")?;
451
452 let (oauth_token, api_token, client) = this.read_with(&cx, |this, _| {
453 (
454 this.oauth_token.clone(),
455 this.api_token.clone(),
456 this.client.clone(),
457 )
458 })?;
459
460 let oauth_token = oauth_token.context("No OAuth token available")?;
461
462 let token = match api_token {
463 Some(api_token) if api_token.remaining_seconds() > 5 * 60 => api_token.clone(),
464 _ => {
465 let token = request_api_token(&oauth_token, client.clone()).await?;
466 this.update(&mut cx, |this, cx| {
467 this.api_token = Some(token.clone());
468 cx.notify();
469 })?;
470 token
471 }
472 };
473
474 stream_completion(client.clone(), token.api_key, request).await
475 }
476}
477
478async fn get_models(api_token: String, client: Arc<dyn HttpClient>) -> Result<Vec<Model>> {
479 let all_models = request_models(api_token, client).await?;
480
481 let mut models: Vec<Model> = all_models
482 .into_iter()
483 .filter(|model| {
484 // Ensure user has access to the model; Policy is present only for models that must be
485 // enabled in the GitHub dashboard
486 model.model_picker_enabled
487 && model
488 .policy
489 .as_ref()
490 .is_none_or(|policy| policy.state == "enabled")
491 })
492 // The first model from the API response, in any given family, appear to be the non-tagged
493 // models, which are likely the best choice (e.g. gpt-4o rather than gpt-4o-2024-11-20)
494 .dedup_by(|a, b| a.capabilities.family == b.capabilities.family)
495 .collect();
496
497 if let Some(default_model_position) =
498 models.iter().position(|model| model.id == DEFAULT_MODEL_ID)
499 {
500 let default_model = models.remove(default_model_position);
501 models.insert(0, default_model);
502 }
503
504 Ok(models)
505}
506
507async fn request_models(api_token: String, client: Arc<dyn HttpClient>) -> Result<Vec<Model>> {
508 let request_builder = HttpRequest::builder()
509 .method(Method::GET)
510 .uri(COPILOT_CHAT_MODELS_URL)
511 .header("Authorization", format!("Bearer {}", api_token))
512 .header("Content-Type", "application/json")
513 .header("Copilot-Integration-Id", "vscode-chat");
514
515 let request = request_builder.body(AsyncBody::empty())?;
516
517 let mut response = client.send(request).await?;
518
519 anyhow::ensure!(
520 response.status().is_success(),
521 "Failed to request models: {}",
522 response.status()
523 );
524 let mut body = Vec::new();
525 response.body_mut().read_to_end(&mut body).await?;
526
527 let body_str = std::str::from_utf8(&body)?;
528
529 let models = serde_json::from_str::<ModelSchema>(body_str)?.data;
530
531 Ok(models)
532}
533
534async fn request_api_token(oauth_token: &str, client: Arc<dyn HttpClient>) -> Result<ApiToken> {
535 let request_builder = HttpRequest::builder()
536 .method(Method::GET)
537 .uri(COPILOT_CHAT_AUTH_URL)
538 .header("Authorization", format!("token {}", oauth_token))
539 .header("Accept", "application/json");
540
541 let request = request_builder.body(AsyncBody::empty())?;
542
543 let mut response = client.send(request).await?;
544
545 if response.status().is_success() {
546 let mut body = Vec::new();
547 response.body_mut().read_to_end(&mut body).await?;
548
549 let body_str = std::str::from_utf8(&body)?;
550
551 let parsed: ApiTokenResponse = serde_json::from_str(body_str)?;
552 ApiToken::try_from(parsed)
553 } else {
554 let mut body = Vec::new();
555 response.body_mut().read_to_end(&mut body).await?;
556
557 let body_str = std::str::from_utf8(&body)?;
558 anyhow::bail!("Failed to request API token: {body_str}");
559 }
560}
561
562fn extract_oauth_token(contents: String) -> Option<String> {
563 serde_json::from_str::<serde_json::Value>(&contents)
564 .map(|v| {
565 v.as_object().and_then(|obj| {
566 obj.iter().find_map(|(key, value)| {
567 if key.starts_with("github.com") {
568 value["oauth_token"].as_str().map(|v| v.to_string())
569 } else {
570 None
571 }
572 })
573 })
574 })
575 .ok()
576 .flatten()
577}
578
579async fn stream_completion(
580 client: Arc<dyn HttpClient>,
581 api_key: String,
582 request: Request,
583) -> Result<BoxStream<'static, Result<ResponseEvent>>> {
584 let request_builder = HttpRequest::builder()
585 .method(Method::POST)
586 .uri(COPILOT_CHAT_COMPLETION_URL)
587 .header(
588 "Editor-Version",
589 format!(
590 "Zed/{}",
591 option_env!("CARGO_PKG_VERSION").unwrap_or("unknown")
592 ),
593 )
594 .header("Authorization", format!("Bearer {}", api_key))
595 .header("Content-Type", "application/json")
596 .header("Copilot-Integration-Id", "vscode-chat")
597 .header("Copilot-Vision-Request", "true");
598
599 let is_streaming = request.stream;
600
601 let json = serde_json::to_string(&request)?;
602 let request = request_builder.body(AsyncBody::from(json))?;
603 let mut response = client.send(request).await?;
604
605 if !response.status().is_success() {
606 let mut body = Vec::new();
607 response.body_mut().read_to_end(&mut body).await?;
608 let body_str = std::str::from_utf8(&body)?;
609 anyhow::bail!(
610 "Failed to connect to API: {} {}",
611 response.status(),
612 body_str
613 );
614 }
615
616 if is_streaming {
617 let reader = BufReader::new(response.into_body());
618 Ok(reader
619 .lines()
620 .filter_map(|line| async move {
621 match line {
622 Ok(line) => {
623 let line = line.strip_prefix("data: ")?;
624 if line.starts_with("[DONE]") {
625 return None;
626 }
627
628 match serde_json::from_str::<ResponseEvent>(line) {
629 Ok(response) => {
630 if response.choices.is_empty() {
631 None
632 } else {
633 Some(Ok(response))
634 }
635 }
636 Err(error) => Some(Err(anyhow!(error))),
637 }
638 }
639 Err(error) => Some(Err(anyhow!(error))),
640 }
641 })
642 .boxed())
643 } else {
644 let mut body = Vec::new();
645 response.body_mut().read_to_end(&mut body).await?;
646 let body_str = std::str::from_utf8(&body)?;
647 let response: ResponseEvent = serde_json::from_str(body_str)?;
648
649 Ok(futures::stream::once(async move { Ok(response) }).boxed())
650 }
651}
652
653#[cfg(test)]
654mod tests {
655 use super::*;
656
657 #[test]
658 fn test_resilient_model_schema_deserialize() {
659 let json = r#"{
660 "data": [
661 {
662 "capabilities": {
663 "family": "gpt-4",
664 "limits": {
665 "max_context_window_tokens": 32768,
666 "max_output_tokens": 4096,
667 "max_prompt_tokens": 32768
668 },
669 "object": "model_capabilities",
670 "supports": { "streaming": true, "tool_calls": true },
671 "tokenizer": "cl100k_base",
672 "type": "chat"
673 },
674 "id": "gpt-4",
675 "model_picker_enabled": false,
676 "name": "GPT 4",
677 "object": "model",
678 "preview": false,
679 "vendor": "Azure OpenAI",
680 "version": "gpt-4-0613"
681 },
682 {
683 "some-unknown-field": 123
684 },
685 {
686 "capabilities": {
687 "family": "claude-3.7-sonnet",
688 "limits": {
689 "max_context_window_tokens": 200000,
690 "max_output_tokens": 16384,
691 "max_prompt_tokens": 90000,
692 "vision": {
693 "max_prompt_image_size": 3145728,
694 "max_prompt_images": 1,
695 "supported_media_types": ["image/jpeg", "image/png", "image/webp"]
696 }
697 },
698 "object": "model_capabilities",
699 "supports": {
700 "parallel_tool_calls": true,
701 "streaming": true,
702 "tool_calls": true,
703 "vision": true
704 },
705 "tokenizer": "o200k_base",
706 "type": "chat"
707 },
708 "id": "claude-3.7-sonnet",
709 "model_picker_enabled": true,
710 "name": "Claude 3.7 Sonnet",
711 "object": "model",
712 "policy": {
713 "state": "enabled",
714 "terms": "Enable access to the latest Claude 3.7 Sonnet model from Anthropic. [Learn more about how GitHub Copilot serves Claude 3.7 Sonnet](https://docs.github.com/copilot/using-github-copilot/using-claude-sonnet-in-github-copilot)."
715 },
716 "preview": false,
717 "vendor": "Anthropic",
718 "version": "claude-3.7-sonnet"
719 }
720 ],
721 "object": "list"
722 }"#;
723
724 let schema: ModelSchema = serde_json::from_str(&json).unwrap();
725
726 assert_eq!(schema.data.len(), 2);
727 assert_eq!(schema.data[0].id, "gpt-4");
728 assert_eq!(schema.data[1].id, "claude-3.7-sonnet");
729 }
730}