copilot_chat.rs

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