1use ai_onboarding::YoungAccountBanner;
2use anthropic::AnthropicModelMode;
3use anyhow::{Context as _, Result, anyhow};
4use chrono::{DateTime, Utc};
5use client::{Client, UserStore, zed_urls};
6use cloud_llm_client::{
7 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CLIENT_SUPPORTS_X_AI_HEADER_NAME, CompletionBody,
8 CompletionEvent, CountTokensBody, CountTokensResponse, EXPIRED_LLM_TOKEN_HEADER_NAME,
9 ListModelsResponse, Plan, PlanV2, SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME,
10 ZED_VERSION_HEADER_NAME,
11};
12use feature_flags::{FeatureFlagAppExt as _, OpenAiResponsesApiFeatureFlag};
13use futures::{
14 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
15};
16use google_ai::GoogleModelMode;
17use gpui::{AnyElement, AnyView, App, AsyncApp, Context, Entity, Subscription, Task};
18use http_client::http::{HeaderMap, HeaderValue};
19use http_client::{AsyncBody, HttpClient, HttpRequestExt, Method, Response, StatusCode};
20use language_model::{
21 AuthenticateError, IconOrSvg, LanguageModel, LanguageModelCacheConfiguration,
22 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
23 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
24 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
25 LanguageModelToolSchemaFormat, LlmApiToken, PaymentRequiredError, RateLimiter,
26 RefreshLlmTokenListener,
27};
28use release_channel::AppVersion;
29use schemars::JsonSchema;
30use semver::Version;
31use serde::{Deserialize, Serialize, de::DeserializeOwned};
32use settings::SettingsStore;
33pub use settings::ZedDotDevAvailableModel as AvailableModel;
34pub use settings::ZedDotDevAvailableProvider as AvailableProvider;
35use smol::io::{AsyncReadExt, BufReader};
36use std::pin::Pin;
37use std::sync::Arc;
38use std::time::Duration;
39use thiserror::Error;
40use ui::{TintColor, prelude::*};
41use util::{ResultExt as _, maybe};
42
43use crate::provider::anthropic::{
44 AnthropicEventMapper, count_anthropic_tokens_with_tiktoken, into_anthropic,
45};
46use crate::provider::google::{GoogleEventMapper, into_google};
47use crate::provider::open_ai::{
48 OpenAiEventMapper, OpenAiResponseEventMapper, count_open_ai_tokens, into_open_ai,
49 into_open_ai_response,
50};
51use crate::provider::x_ai::count_xai_tokens;
52
53const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
54const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
55
56#[derive(Default, Clone, Debug, PartialEq)]
57pub struct ZedDotDevSettings {
58 pub available_models: Vec<AvailableModel>,
59}
60#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
61#[serde(tag = "type", rename_all = "lowercase")]
62pub enum ModelMode {
63 #[default]
64 Default,
65 Thinking {
66 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
67 budget_tokens: Option<u32>,
68 },
69}
70
71impl From<ModelMode> for AnthropicModelMode {
72 fn from(value: ModelMode) -> Self {
73 match value {
74 ModelMode::Default => AnthropicModelMode::Default,
75 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
76 }
77 }
78}
79
80pub struct CloudLanguageModelProvider {
81 client: Arc<Client>,
82 state: Entity<State>,
83 _maintain_client_status: Task<()>,
84}
85
86pub struct State {
87 client: Arc<Client>,
88 llm_api_token: LlmApiToken,
89 user_store: Entity<UserStore>,
90 status: client::Status,
91 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
92 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
93 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
94 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
95 _fetch_models_task: Task<()>,
96 _settings_subscription: Subscription,
97 _llm_token_subscription: Subscription,
98}
99
100impl State {
101 fn new(
102 client: Arc<Client>,
103 user_store: Entity<UserStore>,
104 status: client::Status,
105 cx: &mut Context<Self>,
106 ) -> Self {
107 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
108 let mut current_user = user_store.read(cx).watch_current_user();
109 Self {
110 client: client.clone(),
111 llm_api_token: LlmApiToken::default(),
112 user_store,
113 status,
114 models: Vec::new(),
115 default_model: None,
116 default_fast_model: None,
117 recommended_models: Vec::new(),
118 _fetch_models_task: cx.spawn(async move |this, cx| {
119 maybe!(async move {
120 let (client, llm_api_token) = this
121 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
122
123 while current_user.borrow().is_none() {
124 current_user.next().await;
125 }
126
127 let response =
128 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
129 this.update(cx, |this, cx| this.update_models(response, cx))?;
130 anyhow::Ok(())
131 })
132 .await
133 .context("failed to fetch Zed models")
134 .log_err();
135 }),
136 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
137 cx.notify();
138 }),
139 _llm_token_subscription: cx.subscribe(
140 &refresh_llm_token_listener,
141 move |this, _listener, _event, cx| {
142 let client = this.client.clone();
143 let llm_api_token = this.llm_api_token.clone();
144 cx.spawn(async move |this, cx| {
145 llm_api_token.refresh(&client).await?;
146 let response = Self::fetch_models(client, llm_api_token).await?;
147 this.update(cx, |this, cx| {
148 this.update_models(response, cx);
149 })
150 })
151 .detach_and_log_err(cx);
152 },
153 ),
154 }
155 }
156
157 fn is_signed_out(&self, cx: &App) -> bool {
158 self.user_store.read(cx).current_user().is_none()
159 }
160
161 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
162 let client = self.client.clone();
163 cx.spawn(async move |state, cx| {
164 client.sign_in_with_optional_connect(true, cx).await?;
165 state.update(cx, |_, cx| cx.notify())
166 })
167 }
168 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
169 let mut models = Vec::new();
170
171 for model in response.models {
172 models.push(Arc::new(model.clone()));
173
174 // Right now we represent thinking variants of models as separate models on the client,
175 // so we need to insert variants for any model that supports thinking.
176 if model.supports_thinking {
177 models.push(Arc::new(cloud_llm_client::LanguageModel {
178 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
179 display_name: format!("{} Thinking", model.display_name),
180 ..model
181 }));
182 }
183 }
184
185 self.default_model = models
186 .iter()
187 .find(|model| {
188 response
189 .default_model
190 .as_ref()
191 .is_some_and(|default_model_id| &model.id == default_model_id)
192 })
193 .cloned();
194 self.default_fast_model = models
195 .iter()
196 .find(|model| {
197 response
198 .default_fast_model
199 .as_ref()
200 .is_some_and(|default_fast_model_id| &model.id == default_fast_model_id)
201 })
202 .cloned();
203 self.recommended_models = response
204 .recommended_models
205 .iter()
206 .filter_map(|id| models.iter().find(|model| &model.id == id))
207 .cloned()
208 .collect();
209 self.models = models;
210 cx.notify();
211 }
212
213 async fn fetch_models(
214 client: Arc<Client>,
215 llm_api_token: LlmApiToken,
216 ) -> Result<ListModelsResponse> {
217 let http_client = &client.http_client();
218 let token = llm_api_token.acquire(&client).await?;
219
220 let request = http_client::Request::builder()
221 .method(Method::GET)
222 .header(CLIENT_SUPPORTS_X_AI_HEADER_NAME, "true")
223 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
224 .header("Authorization", format!("Bearer {token}"))
225 .body(AsyncBody::empty())?;
226 let mut response = http_client
227 .send(request)
228 .await
229 .context("failed to send list models request")?;
230
231 if response.status().is_success() {
232 let mut body = String::new();
233 response.body_mut().read_to_string(&mut body).await?;
234 Ok(serde_json::from_str(&body)?)
235 } else {
236 let mut body = String::new();
237 response.body_mut().read_to_string(&mut body).await?;
238 anyhow::bail!(
239 "error listing models.\nStatus: {:?}\nBody: {body}",
240 response.status(),
241 );
242 }
243 }
244}
245
246impl CloudLanguageModelProvider {
247 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
248 let mut status_rx = client.status();
249 let status = *status_rx.borrow();
250
251 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
252
253 let state_ref = state.downgrade();
254 let maintain_client_status = cx.spawn(async move |cx| {
255 while let Some(status) = status_rx.next().await {
256 if let Some(this) = state_ref.upgrade() {
257 _ = this.update(cx, |this, cx| {
258 if this.status != status {
259 this.status = status;
260 cx.notify();
261 }
262 });
263 } else {
264 break;
265 }
266 }
267 });
268
269 Self {
270 client,
271 state,
272 _maintain_client_status: maintain_client_status,
273 }
274 }
275
276 fn create_language_model(
277 &self,
278 model: Arc<cloud_llm_client::LanguageModel>,
279 llm_api_token: LlmApiToken,
280 ) -> Arc<dyn LanguageModel> {
281 Arc::new(CloudLanguageModel {
282 id: LanguageModelId(SharedString::from(model.id.0.clone())),
283 model,
284 llm_api_token,
285 client: self.client.clone(),
286 request_limiter: RateLimiter::new(4),
287 })
288 }
289}
290
291impl LanguageModelProviderState for CloudLanguageModelProvider {
292 type ObservableEntity = State;
293
294 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
295 Some(self.state.clone())
296 }
297}
298
299impl LanguageModelProvider for CloudLanguageModelProvider {
300 fn id(&self) -> LanguageModelProviderId {
301 PROVIDER_ID
302 }
303
304 fn name(&self) -> LanguageModelProviderName {
305 PROVIDER_NAME
306 }
307
308 fn icon(&self) -> IconOrSvg {
309 IconOrSvg::Icon(IconName::AiZed)
310 }
311
312 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
313 let default_model = self.state.read(cx).default_model.clone()?;
314 let llm_api_token = self.state.read(cx).llm_api_token.clone();
315 Some(self.create_language_model(default_model, llm_api_token))
316 }
317
318 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
319 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
320 let llm_api_token = self.state.read(cx).llm_api_token.clone();
321 Some(self.create_language_model(default_fast_model, llm_api_token))
322 }
323
324 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
325 let llm_api_token = self.state.read(cx).llm_api_token.clone();
326 self.state
327 .read(cx)
328 .recommended_models
329 .iter()
330 .cloned()
331 .map(|model| self.create_language_model(model, llm_api_token.clone()))
332 .collect()
333 }
334
335 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
336 let llm_api_token = self.state.read(cx).llm_api_token.clone();
337 self.state
338 .read(cx)
339 .models
340 .iter()
341 .cloned()
342 .map(|model| self.create_language_model(model, llm_api_token.clone()))
343 .collect()
344 }
345
346 fn is_authenticated(&self, cx: &App) -> bool {
347 let state = self.state.read(cx);
348 !state.is_signed_out(cx)
349 }
350
351 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
352 Task::ready(Ok(()))
353 }
354
355 fn configuration_view(
356 &self,
357 _target_agent: language_model::ConfigurationViewTargetAgent,
358 _: &mut Window,
359 cx: &mut App,
360 ) -> AnyView {
361 cx.new(|_| ConfigurationView::new(self.state.clone()))
362 .into()
363 }
364
365 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
366 Task::ready(Ok(()))
367 }
368}
369
370pub struct CloudLanguageModel {
371 id: LanguageModelId,
372 model: Arc<cloud_llm_client::LanguageModel>,
373 llm_api_token: LlmApiToken,
374 client: Arc<Client>,
375 request_limiter: RateLimiter,
376}
377
378struct PerformLlmCompletionResponse {
379 response: Response<AsyncBody>,
380 includes_status_messages: bool,
381}
382
383impl CloudLanguageModel {
384 async fn perform_llm_completion(
385 client: Arc<Client>,
386 llm_api_token: LlmApiToken,
387 app_version: Option<Version>,
388 body: CompletionBody,
389 ) -> Result<PerformLlmCompletionResponse> {
390 let http_client = &client.http_client();
391
392 let mut token = llm_api_token.acquire(&client).await?;
393 let mut refreshed_token = false;
394
395 loop {
396 let request = http_client::Request::builder()
397 .method(Method::POST)
398 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
399 .when_some(app_version.as_ref(), |builder, app_version| {
400 builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
401 })
402 .header("Content-Type", "application/json")
403 .header("Authorization", format!("Bearer {token}"))
404 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
405 .body(serde_json::to_string(&body)?.into())?;
406
407 let mut response = http_client.send(request).await?;
408 let status = response.status();
409 if status.is_success() {
410 let includes_status_messages = response
411 .headers()
412 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
413 .is_some();
414
415 return Ok(PerformLlmCompletionResponse {
416 response,
417 includes_status_messages,
418 });
419 }
420
421 if !refreshed_token
422 && response
423 .headers()
424 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
425 .is_some()
426 {
427 token = llm_api_token.refresh(&client).await?;
428 refreshed_token = true;
429 continue;
430 }
431
432 if status == StatusCode::PAYMENT_REQUIRED {
433 return Err(anyhow!(PaymentRequiredError));
434 }
435
436 let mut body = String::new();
437 let headers = response.headers().clone();
438 response.body_mut().read_to_string(&mut body).await?;
439 return Err(anyhow!(ApiError {
440 status,
441 body,
442 headers
443 }));
444 }
445 }
446}
447
448#[derive(Debug, Error)]
449#[error("cloud language model request failed with status {status}: {body}")]
450struct ApiError {
451 status: StatusCode,
452 body: String,
453 headers: HeaderMap<HeaderValue>,
454}
455
456/// Represents error responses from Zed's cloud API.
457///
458/// Example JSON for an upstream HTTP error:
459/// ```json
460/// {
461/// "code": "upstream_http_error",
462/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
463/// "upstream_status": 503
464/// }
465/// ```
466#[derive(Debug, serde::Deserialize)]
467struct CloudApiError {
468 code: String,
469 message: String,
470 #[serde(default)]
471 #[serde(deserialize_with = "deserialize_optional_status_code")]
472 upstream_status: Option<StatusCode>,
473 #[serde(default)]
474 retry_after: Option<f64>,
475}
476
477fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
478where
479 D: serde::Deserializer<'de>,
480{
481 let opt: Option<u16> = Option::deserialize(deserializer)?;
482 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
483}
484
485impl From<ApiError> for LanguageModelCompletionError {
486 fn from(error: ApiError) -> Self {
487 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
488 if cloud_error.code.starts_with("upstream_http_") {
489 let status = if let Some(status) = cloud_error.upstream_status {
490 status
491 } else if cloud_error.code.ends_with("_error") {
492 error.status
493 } else {
494 // If there's a status code in the code string (e.g. "upstream_http_429")
495 // then use that; otherwise, see if the JSON contains a status code.
496 cloud_error
497 .code
498 .strip_prefix("upstream_http_")
499 .and_then(|code_str| code_str.parse::<u16>().ok())
500 .and_then(|code| StatusCode::from_u16(code).ok())
501 .unwrap_or(error.status)
502 };
503
504 return LanguageModelCompletionError::UpstreamProviderError {
505 message: cloud_error.message,
506 status,
507 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
508 };
509 }
510
511 return LanguageModelCompletionError::from_http_status(
512 PROVIDER_NAME,
513 error.status,
514 cloud_error.message,
515 None,
516 );
517 }
518
519 let retry_after = None;
520 LanguageModelCompletionError::from_http_status(
521 PROVIDER_NAME,
522 error.status,
523 error.body,
524 retry_after,
525 )
526 }
527}
528
529impl LanguageModel for CloudLanguageModel {
530 fn id(&self) -> LanguageModelId {
531 self.id.clone()
532 }
533
534 fn name(&self) -> LanguageModelName {
535 LanguageModelName::from(self.model.display_name.clone())
536 }
537
538 fn provider_id(&self) -> LanguageModelProviderId {
539 PROVIDER_ID
540 }
541
542 fn provider_name(&self) -> LanguageModelProviderName {
543 PROVIDER_NAME
544 }
545
546 fn upstream_provider_id(&self) -> LanguageModelProviderId {
547 use cloud_llm_client::LanguageModelProvider::*;
548 match self.model.provider {
549 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
550 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
551 Google => language_model::GOOGLE_PROVIDER_ID,
552 XAi => language_model::X_AI_PROVIDER_ID,
553 }
554 }
555
556 fn upstream_provider_name(&self) -> LanguageModelProviderName {
557 use cloud_llm_client::LanguageModelProvider::*;
558 match self.model.provider {
559 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
560 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
561 Google => language_model::GOOGLE_PROVIDER_NAME,
562 XAi => language_model::X_AI_PROVIDER_NAME,
563 }
564 }
565
566 fn supports_tools(&self) -> bool {
567 self.model.supports_tools
568 }
569
570 fn supports_images(&self) -> bool {
571 self.model.supports_images
572 }
573
574 fn supports_streaming_tools(&self) -> bool {
575 self.model.supports_streaming_tools
576 }
577
578 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
579 match choice {
580 LanguageModelToolChoice::Auto
581 | LanguageModelToolChoice::Any
582 | LanguageModelToolChoice::None => true,
583 }
584 }
585
586 fn supports_split_token_display(&self) -> bool {
587 use cloud_llm_client::LanguageModelProvider::*;
588 matches!(self.model.provider, OpenAi)
589 }
590
591 fn telemetry_id(&self) -> String {
592 format!("zed.dev/{}", self.model.id)
593 }
594
595 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
596 match self.model.provider {
597 cloud_llm_client::LanguageModelProvider::Anthropic
598 | cloud_llm_client::LanguageModelProvider::OpenAi
599 | cloud_llm_client::LanguageModelProvider::XAi => {
600 LanguageModelToolSchemaFormat::JsonSchema
601 }
602 cloud_llm_client::LanguageModelProvider::Google => {
603 LanguageModelToolSchemaFormat::JsonSchemaSubset
604 }
605 }
606 }
607
608 fn max_token_count(&self) -> u64 {
609 self.model.max_token_count as u64
610 }
611
612 fn max_output_tokens(&self) -> Option<u64> {
613 Some(self.model.max_output_tokens as u64)
614 }
615
616 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
617 match &self.model.provider {
618 cloud_llm_client::LanguageModelProvider::Anthropic => {
619 Some(LanguageModelCacheConfiguration {
620 min_total_token: 2_048,
621 should_speculate: true,
622 max_cache_anchors: 4,
623 })
624 }
625 cloud_llm_client::LanguageModelProvider::OpenAi
626 | cloud_llm_client::LanguageModelProvider::XAi
627 | cloud_llm_client::LanguageModelProvider::Google => None,
628 }
629 }
630
631 fn count_tokens(
632 &self,
633 request: LanguageModelRequest,
634 cx: &App,
635 ) -> BoxFuture<'static, Result<u64>> {
636 match self.model.provider {
637 cloud_llm_client::LanguageModelProvider::Anthropic => cx
638 .background_spawn(async move { count_anthropic_tokens_with_tiktoken(request) })
639 .boxed(),
640 cloud_llm_client::LanguageModelProvider::OpenAi => {
641 let model = match open_ai::Model::from_id(&self.model.id.0) {
642 Ok(model) => model,
643 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
644 };
645 count_open_ai_tokens(request, model, cx)
646 }
647 cloud_llm_client::LanguageModelProvider::XAi => {
648 let model = match x_ai::Model::from_id(&self.model.id.0) {
649 Ok(model) => model,
650 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
651 };
652 count_xai_tokens(request, model, cx)
653 }
654 cloud_llm_client::LanguageModelProvider::Google => {
655 let client = self.client.clone();
656 let llm_api_token = self.llm_api_token.clone();
657 let model_id = self.model.id.to_string();
658 let generate_content_request =
659 into_google(request, model_id.clone(), GoogleModelMode::Default);
660 async move {
661 let http_client = &client.http_client();
662 let token = llm_api_token.acquire(&client).await?;
663
664 let request_body = CountTokensBody {
665 provider: cloud_llm_client::LanguageModelProvider::Google,
666 model: model_id,
667 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
668 generate_content_request,
669 })?,
670 };
671 let request = http_client::Request::builder()
672 .method(Method::POST)
673 .uri(
674 http_client
675 .build_zed_llm_url("/count_tokens", &[])?
676 .as_ref(),
677 )
678 .header("Content-Type", "application/json")
679 .header("Authorization", format!("Bearer {token}"))
680 .body(serde_json::to_string(&request_body)?.into())?;
681 let mut response = http_client.send(request).await?;
682 let status = response.status();
683 let headers = response.headers().clone();
684 let mut response_body = String::new();
685 response
686 .body_mut()
687 .read_to_string(&mut response_body)
688 .await?;
689
690 if status.is_success() {
691 let response_body: CountTokensResponse =
692 serde_json::from_str(&response_body)?;
693
694 Ok(response_body.tokens as u64)
695 } else {
696 Err(anyhow!(ApiError {
697 status,
698 body: response_body,
699 headers
700 }))
701 }
702 }
703 .boxed()
704 }
705 }
706 }
707
708 fn stream_completion(
709 &self,
710 request: LanguageModelRequest,
711 cx: &AsyncApp,
712 ) -> BoxFuture<
713 'static,
714 Result<
715 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
716 LanguageModelCompletionError,
717 >,
718 > {
719 let thread_id = request.thread_id.clone();
720 let prompt_id = request.prompt_id.clone();
721 let intent = request.intent;
722 let bypass_rate_limit = request.bypass_rate_limit;
723 let app_version = Some(cx.update(|cx| AppVersion::global(cx)));
724 let use_responses_api = cx.update(|cx| cx.has_flag::<OpenAiResponsesApiFeatureFlag>());
725 let thinking_allowed = request.thinking_allowed;
726 let provider_name = provider_name(&self.model.provider);
727 match self.model.provider {
728 cloud_llm_client::LanguageModelProvider::Anthropic => {
729 let request = into_anthropic(
730 request,
731 self.model.id.to_string(),
732 1.0,
733 self.model.max_output_tokens as u64,
734 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
735 AnthropicModelMode::Thinking {
736 budget_tokens: Some(4_096),
737 }
738 } else {
739 AnthropicModelMode::Default
740 },
741 );
742 let client = self.client.clone();
743 let llm_api_token = self.llm_api_token.clone();
744 let future = self.request_limiter.stream_with_bypass(
745 async move {
746 let PerformLlmCompletionResponse {
747 response,
748 includes_status_messages,
749 } = Self::perform_llm_completion(
750 client.clone(),
751 llm_api_token,
752 app_version,
753 CompletionBody {
754 thread_id,
755 prompt_id,
756 intent,
757 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
758 model: request.model.clone(),
759 provider_request: serde_json::to_value(&request)
760 .map_err(|e| anyhow!(e))?,
761 },
762 )
763 .await
764 .map_err(|err| {
765 match err.downcast::<ApiError>() {
766 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
767 Err(err) => anyhow!(err),
768 }
769 })?;
770
771 let mut mapper = AnthropicEventMapper::new();
772 Ok(map_cloud_completion_events(
773 Box::pin(response_lines(response, includes_status_messages)),
774 &provider_name,
775 move |event| mapper.map_event(event),
776 ))
777 },
778 bypass_rate_limit,
779 );
780 async move { Ok(future.await?.boxed()) }.boxed()
781 }
782 cloud_llm_client::LanguageModelProvider::OpenAi => {
783 let client = self.client.clone();
784 let llm_api_token = self.llm_api_token.clone();
785
786 if use_responses_api {
787 let request = into_open_ai_response(
788 request,
789 &self.model.id.0,
790 self.model.supports_parallel_tool_calls,
791 true,
792 None,
793 None,
794 );
795 let future = self.request_limiter.stream_with_bypass(
796 async move {
797 let PerformLlmCompletionResponse {
798 response,
799 includes_status_messages,
800 } = Self::perform_llm_completion(
801 client.clone(),
802 llm_api_token,
803 app_version,
804 CompletionBody {
805 thread_id,
806 prompt_id,
807 intent,
808 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
809 model: request.model.clone(),
810 provider_request: serde_json::to_value(&request)
811 .map_err(|e| anyhow!(e))?,
812 },
813 )
814 .await?;
815
816 let mut mapper = OpenAiResponseEventMapper::new();
817 Ok(map_cloud_completion_events(
818 Box::pin(response_lines(response, includes_status_messages)),
819 &provider_name,
820 move |event| mapper.map_event(event),
821 ))
822 },
823 bypass_rate_limit,
824 );
825 async move { Ok(future.await?.boxed()) }.boxed()
826 } else {
827 let request = into_open_ai(
828 request,
829 &self.model.id.0,
830 self.model.supports_parallel_tool_calls,
831 true,
832 None,
833 None,
834 );
835 let future = self.request_limiter.stream_with_bypass(
836 async move {
837 let PerformLlmCompletionResponse {
838 response,
839 includes_status_messages,
840 } = Self::perform_llm_completion(
841 client.clone(),
842 llm_api_token,
843 app_version,
844 CompletionBody {
845 thread_id,
846 prompt_id,
847 intent,
848 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
849 model: request.model.clone(),
850 provider_request: serde_json::to_value(&request)
851 .map_err(|e| anyhow!(e))?,
852 },
853 )
854 .await?;
855
856 let mut mapper = OpenAiEventMapper::new();
857 Ok(map_cloud_completion_events(
858 Box::pin(response_lines(response, includes_status_messages)),
859 &provider_name,
860 move |event| mapper.map_event(event),
861 ))
862 },
863 bypass_rate_limit,
864 );
865 async move { Ok(future.await?.boxed()) }.boxed()
866 }
867 }
868 cloud_llm_client::LanguageModelProvider::XAi => {
869 let client = self.client.clone();
870 let request = into_open_ai(
871 request,
872 &self.model.id.0,
873 self.model.supports_parallel_tool_calls,
874 false,
875 None,
876 None,
877 );
878 let llm_api_token = self.llm_api_token.clone();
879 let future = self.request_limiter.stream_with_bypass(
880 async move {
881 let PerformLlmCompletionResponse {
882 response,
883 includes_status_messages,
884 } = Self::perform_llm_completion(
885 client.clone(),
886 llm_api_token,
887 app_version,
888 CompletionBody {
889 thread_id,
890 prompt_id,
891 intent,
892 provider: cloud_llm_client::LanguageModelProvider::XAi,
893 model: request.model.clone(),
894 provider_request: serde_json::to_value(&request)
895 .map_err(|e| anyhow!(e))?,
896 },
897 )
898 .await?;
899
900 let mut mapper = OpenAiEventMapper::new();
901 Ok(map_cloud_completion_events(
902 Box::pin(response_lines(response, includes_status_messages)),
903 &provider_name,
904 move |event| mapper.map_event(event),
905 ))
906 },
907 bypass_rate_limit,
908 );
909 async move { Ok(future.await?.boxed()) }.boxed()
910 }
911 cloud_llm_client::LanguageModelProvider::Google => {
912 let client = self.client.clone();
913 let request =
914 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
915 let llm_api_token = self.llm_api_token.clone();
916 let future = self.request_limiter.stream_with_bypass(
917 async move {
918 let PerformLlmCompletionResponse {
919 response,
920 includes_status_messages,
921 } = Self::perform_llm_completion(
922 client.clone(),
923 llm_api_token,
924 app_version,
925 CompletionBody {
926 thread_id,
927 prompt_id,
928 intent,
929 provider: cloud_llm_client::LanguageModelProvider::Google,
930 model: request.model.model_id.clone(),
931 provider_request: serde_json::to_value(&request)
932 .map_err(|e| anyhow!(e))?,
933 },
934 )
935 .await?;
936
937 let mut mapper = GoogleEventMapper::new();
938 Ok(map_cloud_completion_events(
939 Box::pin(response_lines(response, includes_status_messages)),
940 &provider_name,
941 move |event| mapper.map_event(event),
942 ))
943 },
944 bypass_rate_limit,
945 );
946 async move { Ok(future.await?.boxed()) }.boxed()
947 }
948 }
949 }
950}
951
952fn map_cloud_completion_events<T, F>(
953 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
954 provider: &LanguageModelProviderName,
955 mut map_callback: F,
956) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
957where
958 T: DeserializeOwned + 'static,
959 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
960 + Send
961 + 'static,
962{
963 let provider = provider.clone();
964 stream
965 .flat_map(move |event| {
966 futures::stream::iter(match event {
967 Err(error) => {
968 vec![Err(LanguageModelCompletionError::from(error))]
969 }
970 Ok(CompletionEvent::Status(event)) => {
971 vec![
972 LanguageModelCompletionEvent::from_completion_request_status(
973 event,
974 provider.clone(),
975 ),
976 ]
977 }
978 Ok(CompletionEvent::Event(event)) => map_callback(event),
979 })
980 })
981 .boxed()
982}
983
984fn provider_name(provider: &cloud_llm_client::LanguageModelProvider) -> LanguageModelProviderName {
985 match provider {
986 cloud_llm_client::LanguageModelProvider::Anthropic => {
987 language_model::ANTHROPIC_PROVIDER_NAME
988 }
989 cloud_llm_client::LanguageModelProvider::OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
990 cloud_llm_client::LanguageModelProvider::Google => language_model::GOOGLE_PROVIDER_NAME,
991 cloud_llm_client::LanguageModelProvider::XAi => language_model::X_AI_PROVIDER_NAME,
992 }
993}
994
995fn response_lines<T: DeserializeOwned>(
996 response: Response<AsyncBody>,
997 includes_status_messages: bool,
998) -> impl Stream<Item = Result<CompletionEvent<T>>> {
999 futures::stream::try_unfold(
1000 (String::new(), BufReader::new(response.into_body())),
1001 move |(mut line, mut body)| async move {
1002 match body.read_line(&mut line).await {
1003 Ok(0) => Ok(None),
1004 Ok(_) => {
1005 let event = if includes_status_messages {
1006 serde_json::from_str::<CompletionEvent<T>>(&line)?
1007 } else {
1008 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1009 };
1010
1011 line.clear();
1012 Ok(Some((event, (line, body))))
1013 }
1014 Err(e) => Err(e.into()),
1015 }
1016 },
1017 )
1018}
1019
1020#[derive(IntoElement, RegisterComponent)]
1021struct ZedAiConfiguration {
1022 is_connected: bool,
1023 plan: Option<Plan>,
1024 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1025 eligible_for_trial: bool,
1026 account_too_young: bool,
1027 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1028}
1029
1030impl RenderOnce for ZedAiConfiguration {
1031 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1032 let is_pro = self
1033 .plan
1034 .is_some_and(|plan| plan == Plan::V2(PlanV2::ZedPro));
1035 let subscription_text = match (self.plan, self.subscription_period) {
1036 (Some(Plan::V2(PlanV2::ZedPro)), Some(_)) => {
1037 "You have access to Zed's hosted models through your Pro subscription."
1038 }
1039 (Some(Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1040 "You have access to Zed's hosted models through your Pro trial."
1041 }
1042 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1043 if self.eligible_for_trial {
1044 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1045 } else {
1046 "Subscribe for access to Zed's hosted models."
1047 }
1048 }
1049 _ => {
1050 if self.eligible_for_trial {
1051 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1052 } else {
1053 "Subscribe for access to Zed's hosted models."
1054 }
1055 }
1056 };
1057
1058 let manage_subscription_buttons = if is_pro {
1059 Button::new("manage_settings", "Manage Subscription")
1060 .full_width()
1061 .style(ButtonStyle::Tinted(TintColor::Accent))
1062 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1063 .into_any_element()
1064 } else if self.plan.is_none() || self.eligible_for_trial {
1065 Button::new("start_trial", "Start 14-day Free Pro Trial")
1066 .full_width()
1067 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1068 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1069 .into_any_element()
1070 } else {
1071 Button::new("upgrade", "Upgrade to Pro")
1072 .full_width()
1073 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1074 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1075 .into_any_element()
1076 };
1077
1078 if !self.is_connected {
1079 return v_flex()
1080 .gap_2()
1081 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1082 .child(
1083 Button::new("sign_in", "Sign In to use Zed AI")
1084 .icon_color(Color::Muted)
1085 .icon(IconName::Github)
1086 .icon_size(IconSize::Small)
1087 .icon_position(IconPosition::Start)
1088 .full_width()
1089 .on_click({
1090 let callback = self.sign_in_callback.clone();
1091 move |_, window, cx| (callback)(window, cx)
1092 }),
1093 );
1094 }
1095
1096 v_flex().gap_2().w_full().map(|this| {
1097 if self.account_too_young {
1098 this.child(YoungAccountBanner).child(
1099 Button::new("upgrade", "Upgrade to Pro")
1100 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1101 .full_width()
1102 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1103 )
1104 } else {
1105 this.text_sm()
1106 .child(subscription_text)
1107 .child(manage_subscription_buttons)
1108 }
1109 })
1110 }
1111}
1112
1113struct ConfigurationView {
1114 state: Entity<State>,
1115 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1116}
1117
1118impl ConfigurationView {
1119 fn new(state: Entity<State>) -> Self {
1120 let sign_in_callback = Arc::new({
1121 let state = state.clone();
1122 move |_window: &mut Window, cx: &mut App| {
1123 state.update(cx, |state, cx| {
1124 state.authenticate(cx).detach_and_log_err(cx);
1125 });
1126 }
1127 });
1128
1129 Self {
1130 state,
1131 sign_in_callback,
1132 }
1133 }
1134}
1135
1136impl Render for ConfigurationView {
1137 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1138 let state = self.state.read(cx);
1139 let user_store = state.user_store.read(cx);
1140
1141 ZedAiConfiguration {
1142 is_connected: !state.is_signed_out(cx),
1143 plan: user_store.plan(),
1144 subscription_period: user_store.subscription_period(),
1145 eligible_for_trial: user_store.trial_started_at().is_none(),
1146 account_too_young: user_store.account_too_young(),
1147 sign_in_callback: self.sign_in_callback.clone(),
1148 }
1149 }
1150}
1151
1152impl Component for ZedAiConfiguration {
1153 fn name() -> &'static str {
1154 "AI Configuration Content"
1155 }
1156
1157 fn sort_name() -> &'static str {
1158 "AI Configuration Content"
1159 }
1160
1161 fn scope() -> ComponentScope {
1162 ComponentScope::Onboarding
1163 }
1164
1165 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1166 fn configuration(
1167 is_connected: bool,
1168 plan: Option<Plan>,
1169 eligible_for_trial: bool,
1170 account_too_young: bool,
1171 ) -> AnyElement {
1172 ZedAiConfiguration {
1173 is_connected,
1174 plan,
1175 subscription_period: plan
1176 .is_some()
1177 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1178 eligible_for_trial,
1179 account_too_young,
1180 sign_in_callback: Arc::new(|_, _| {}),
1181 }
1182 .into_any_element()
1183 }
1184
1185 Some(
1186 v_flex()
1187 .p_4()
1188 .gap_4()
1189 .children(vec![
1190 single_example("Not connected", configuration(false, None, false, false)),
1191 single_example(
1192 "Accept Terms of Service",
1193 configuration(true, None, true, false),
1194 ),
1195 single_example(
1196 "No Plan - Not eligible for trial",
1197 configuration(true, None, false, false),
1198 ),
1199 single_example(
1200 "No Plan - Eligible for trial",
1201 configuration(true, None, true, false),
1202 ),
1203 single_example(
1204 "Free Plan",
1205 configuration(true, Some(Plan::V2(PlanV2::ZedFree)), true, false),
1206 ),
1207 single_example(
1208 "Zed Pro Trial Plan",
1209 configuration(true, Some(Plan::V2(PlanV2::ZedProTrial)), true, false),
1210 ),
1211 single_example(
1212 "Zed Pro Plan",
1213 configuration(true, Some(Plan::V2(PlanV2::ZedPro)), true, false),
1214 ),
1215 ])
1216 .into_any_element(),
1217 )
1218 }
1219}
1220
1221#[cfg(test)]
1222mod tests {
1223 use super::*;
1224 use http_client::http::{HeaderMap, StatusCode};
1225 use language_model::LanguageModelCompletionError;
1226
1227 #[test]
1228 fn test_api_error_conversion_with_upstream_http_error() {
1229 // upstream_http_error with 503 status should become ServerOverloaded
1230 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout","upstream_status":503}"#;
1231
1232 let api_error = ApiError {
1233 status: StatusCode::INTERNAL_SERVER_ERROR,
1234 body: error_body.to_string(),
1235 headers: HeaderMap::new(),
1236 };
1237
1238 let completion_error: LanguageModelCompletionError = api_error.into();
1239
1240 match completion_error {
1241 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1242 assert_eq!(
1243 message,
1244 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1245 );
1246 }
1247 _ => panic!(
1248 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1249 completion_error
1250 ),
1251 }
1252
1253 // upstream_http_error with 500 status should become ApiInternalServerError
1254 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1255
1256 let api_error = ApiError {
1257 status: StatusCode::INTERNAL_SERVER_ERROR,
1258 body: error_body.to_string(),
1259 headers: HeaderMap::new(),
1260 };
1261
1262 let completion_error: LanguageModelCompletionError = api_error.into();
1263
1264 match completion_error {
1265 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1266 assert_eq!(
1267 message,
1268 "Received an error from the OpenAI API: internal server error"
1269 );
1270 }
1271 _ => panic!(
1272 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1273 completion_error
1274 ),
1275 }
1276
1277 // upstream_http_error with 429 status should become RateLimitExceeded
1278 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1279
1280 let api_error = ApiError {
1281 status: StatusCode::INTERNAL_SERVER_ERROR,
1282 body: error_body.to_string(),
1283 headers: HeaderMap::new(),
1284 };
1285
1286 let completion_error: LanguageModelCompletionError = api_error.into();
1287
1288 match completion_error {
1289 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1290 assert_eq!(
1291 message,
1292 "Received an error from the Google API: rate limit exceeded"
1293 );
1294 }
1295 _ => panic!(
1296 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1297 completion_error
1298 ),
1299 }
1300
1301 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1302 let error_body = "Regular internal server error";
1303
1304 let api_error = ApiError {
1305 status: StatusCode::INTERNAL_SERVER_ERROR,
1306 body: error_body.to_string(),
1307 headers: HeaderMap::new(),
1308 };
1309
1310 let completion_error: LanguageModelCompletionError = api_error.into();
1311
1312 match completion_error {
1313 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1314 assert_eq!(provider, PROVIDER_NAME);
1315 assert_eq!(message, "Regular internal server error");
1316 }
1317 _ => panic!(
1318 "Expected ApiInternalServerError for regular 500, got: {:?}",
1319 completion_error
1320 ),
1321 }
1322
1323 // upstream_http_429 format should be converted to UpstreamProviderError
1324 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1325
1326 let api_error = ApiError {
1327 status: StatusCode::INTERNAL_SERVER_ERROR,
1328 body: error_body.to_string(),
1329 headers: HeaderMap::new(),
1330 };
1331
1332 let completion_error: LanguageModelCompletionError = api_error.into();
1333
1334 match completion_error {
1335 LanguageModelCompletionError::UpstreamProviderError {
1336 message,
1337 status,
1338 retry_after,
1339 } => {
1340 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1341 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1342 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1343 }
1344 _ => panic!(
1345 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1346 completion_error
1347 ),
1348 }
1349
1350 // Invalid JSON in error body should fall back to regular error handling
1351 let error_body = "Not JSON at all";
1352
1353 let api_error = ApiError {
1354 status: StatusCode::INTERNAL_SERVER_ERROR,
1355 body: error_body.to_string(),
1356 headers: HeaderMap::new(),
1357 };
1358
1359 let completion_error: LanguageModelCompletionError = api_error.into();
1360
1361 match completion_error {
1362 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1363 assert_eq!(provider, PROVIDER_NAME);
1364 }
1365 _ => panic!(
1366 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1367 completion_error
1368 ),
1369 }
1370 }
1371}