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