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