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::{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, 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
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
429 && response
430 .headers()
431 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
432 .is_some()
433 {
434 token = llm_api_token.refresh(&client).await?;
435 refreshed_token = true;
436 continue;
437 }
438
439 if status == StatusCode::PAYMENT_REQUIRED {
440 return Err(anyhow!(PaymentRequiredError));
441 }
442
443 let mut body = String::new();
444 let headers = response.headers().clone();
445 response.body_mut().read_to_string(&mut body).await?;
446 return Err(anyhow!(ApiError {
447 status,
448 body,
449 headers
450 }));
451 }
452 }
453}
454
455#[derive(Debug, Error)]
456#[error("cloud language model request failed with status {status}: {body}")]
457struct ApiError {
458 status: StatusCode,
459 body: String,
460 headers: HeaderMap<HeaderValue>,
461}
462
463/// Represents error responses from Zed's cloud API.
464///
465/// Example JSON for an upstream HTTP error:
466/// ```json
467/// {
468/// "code": "upstream_http_error",
469/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
470/// "upstream_status": 503
471/// }
472/// ```
473#[derive(Debug, serde::Deserialize)]
474struct CloudApiError {
475 code: String,
476 message: String,
477 #[serde(default)]
478 #[serde(deserialize_with = "deserialize_optional_status_code")]
479 upstream_status: Option<StatusCode>,
480 #[serde(default)]
481 retry_after: Option<f64>,
482}
483
484fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
485where
486 D: serde::Deserializer<'de>,
487{
488 let opt: Option<u16> = Option::deserialize(deserializer)?;
489 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
490}
491
492impl From<ApiError> for LanguageModelCompletionError {
493 fn from(error: ApiError) -> Self {
494 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
495 if cloud_error.code.starts_with("upstream_http_") {
496 let status = if let Some(status) = cloud_error.upstream_status {
497 status
498 } else if cloud_error.code.ends_with("_error") {
499 error.status
500 } else {
501 // If there's a status code in the code string (e.g. "upstream_http_429")
502 // then use that; otherwise, see if the JSON contains a status code.
503 cloud_error
504 .code
505 .strip_prefix("upstream_http_")
506 .and_then(|code_str| code_str.parse::<u16>().ok())
507 .and_then(|code| StatusCode::from_u16(code).ok())
508 .unwrap_or(error.status)
509 };
510
511 return LanguageModelCompletionError::UpstreamProviderError {
512 message: cloud_error.message,
513 status,
514 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
515 };
516 }
517
518 return LanguageModelCompletionError::from_http_status(
519 PROVIDER_NAME,
520 error.status,
521 cloud_error.message,
522 None,
523 );
524 }
525
526 let retry_after = None;
527 LanguageModelCompletionError::from_http_status(
528 PROVIDER_NAME,
529 error.status,
530 error.body,
531 retry_after,
532 )
533 }
534}
535
536impl LanguageModel for CloudLanguageModel {
537 fn id(&self) -> LanguageModelId {
538 self.id.clone()
539 }
540
541 fn name(&self) -> LanguageModelName {
542 LanguageModelName::from(self.model.display_name.clone())
543 }
544
545 fn provider_id(&self) -> LanguageModelProviderId {
546 PROVIDER_ID
547 }
548
549 fn provider_name(&self) -> LanguageModelProviderName {
550 PROVIDER_NAME
551 }
552
553 fn upstream_provider_id(&self) -> LanguageModelProviderId {
554 use cloud_llm_client::LanguageModelProvider::*;
555 match self.model.provider {
556 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
557 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
558 Google => language_model::GOOGLE_PROVIDER_ID,
559 XAi => language_model::X_AI_PROVIDER_ID,
560 }
561 }
562
563 fn upstream_provider_name(&self) -> LanguageModelProviderName {
564 use cloud_llm_client::LanguageModelProvider::*;
565 match self.model.provider {
566 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
567 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
568 Google => language_model::GOOGLE_PROVIDER_NAME,
569 XAi => language_model::X_AI_PROVIDER_NAME,
570 }
571 }
572
573 fn supports_tools(&self) -> bool {
574 self.model.supports_tools
575 }
576
577 fn supports_images(&self) -> bool {
578 self.model.supports_images
579 }
580
581 fn supports_thinking(&self) -> bool {
582 self.model.supports_thinking
583 }
584
585 fn supports_streaming_tools(&self) -> bool {
586 self.model.supports_streaming_tools
587 }
588
589 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
590 match choice {
591 LanguageModelToolChoice::Auto
592 | LanguageModelToolChoice::Any
593 | LanguageModelToolChoice::None => true,
594 }
595 }
596
597 fn supports_split_token_display(&self) -> bool {
598 use cloud_llm_client::LanguageModelProvider::*;
599 matches!(self.model.provider, OpenAi)
600 }
601
602 fn telemetry_id(&self) -> String {
603 format!("zed.dev/{}", self.model.id)
604 }
605
606 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
607 match self.model.provider {
608 cloud_llm_client::LanguageModelProvider::Anthropic
609 | cloud_llm_client::LanguageModelProvider::OpenAi
610 | cloud_llm_client::LanguageModelProvider::XAi => {
611 LanguageModelToolSchemaFormat::JsonSchema
612 }
613 cloud_llm_client::LanguageModelProvider::Google => {
614 LanguageModelToolSchemaFormat::JsonSchemaSubset
615 }
616 }
617 }
618
619 fn max_token_count(&self) -> u64 {
620 self.model.max_token_count as u64
621 }
622
623 fn max_output_tokens(&self) -> Option<u64> {
624 Some(self.model.max_output_tokens as u64)
625 }
626
627 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
628 match &self.model.provider {
629 cloud_llm_client::LanguageModelProvider::Anthropic => {
630 Some(LanguageModelCacheConfiguration {
631 min_total_token: 2_048,
632 should_speculate: true,
633 max_cache_anchors: 4,
634 })
635 }
636 cloud_llm_client::LanguageModelProvider::OpenAi
637 | cloud_llm_client::LanguageModelProvider::XAi
638 | cloud_llm_client::LanguageModelProvider::Google => None,
639 }
640 }
641
642 fn count_tokens(
643 &self,
644 request: LanguageModelRequest,
645 cx: &App,
646 ) -> BoxFuture<'static, Result<u64>> {
647 match self.model.provider {
648 cloud_llm_client::LanguageModelProvider::Anthropic => cx
649 .background_spawn(async move { count_anthropic_tokens_with_tiktoken(request) })
650 .boxed(),
651 cloud_llm_client::LanguageModelProvider::OpenAi => {
652 let model = match open_ai::Model::from_id(&self.model.id.0) {
653 Ok(model) => model,
654 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
655 };
656 count_open_ai_tokens(request, model, cx)
657 }
658 cloud_llm_client::LanguageModelProvider::XAi => {
659 let model = match x_ai::Model::from_id(&self.model.id.0) {
660 Ok(model) => model,
661 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
662 };
663 count_xai_tokens(request, model, cx)
664 }
665 cloud_llm_client::LanguageModelProvider::Google => {
666 let client = self.client.clone();
667 let llm_api_token = self.llm_api_token.clone();
668 let model_id = self.model.id.to_string();
669 let generate_content_request =
670 into_google(request, model_id.clone(), GoogleModelMode::Default);
671 async move {
672 let http_client = &client.http_client();
673 let token = llm_api_token.acquire(&client).await?;
674
675 let request_body = CountTokensBody {
676 provider: cloud_llm_client::LanguageModelProvider::Google,
677 model: model_id,
678 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
679 generate_content_request,
680 })?,
681 };
682 let request = http_client::Request::builder()
683 .method(Method::POST)
684 .uri(
685 http_client
686 .build_zed_llm_url("/count_tokens", &[])?
687 .as_ref(),
688 )
689 .header("Content-Type", "application/json")
690 .header("Authorization", format!("Bearer {token}"))
691 .body(serde_json::to_string(&request_body)?.into())?;
692 let mut response = http_client.send(request).await?;
693 let status = response.status();
694 let headers = response.headers().clone();
695 let mut response_body = String::new();
696 response
697 .body_mut()
698 .read_to_string(&mut response_body)
699 .await?;
700
701 if status.is_success() {
702 let response_body: CountTokensResponse =
703 serde_json::from_str(&response_body)?;
704
705 Ok(response_body.tokens as u64)
706 } else {
707 Err(anyhow!(ApiError {
708 status,
709 body: response_body,
710 headers
711 }))
712 }
713 }
714 .boxed()
715 }
716 }
717 }
718
719 fn stream_completion(
720 &self,
721 request: LanguageModelRequest,
722 cx: &AsyncApp,
723 ) -> BoxFuture<
724 'static,
725 Result<
726 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
727 LanguageModelCompletionError,
728 >,
729 > {
730 let thread_id = request.thread_id.clone();
731 let prompt_id = request.prompt_id.clone();
732 let intent = request.intent;
733 let app_version = Some(cx.update(|cx| AppVersion::global(cx)));
734 let thinking_allowed = request.thinking_allowed;
735 let is_thinking_toggle_enabled =
736 cx.update(|cx| cx.has_flag::<CloudThinkingToggleFeatureFlag>());
737 let enable_thinking = if is_thinking_toggle_enabled {
738 thinking_allowed && self.model.supports_thinking
739 } else {
740 thinking_allowed && self.model.id.0.ends_with("-thinking")
741 };
742 let provider_name = provider_name(&self.model.provider);
743 match self.model.provider {
744 cloud_llm_client::LanguageModelProvider::Anthropic => {
745 let request = into_anthropic(
746 request,
747 self.model.id.to_string(),
748 1.0,
749 self.model.max_output_tokens as u64,
750 if enable_thinking {
751 AnthropicModelMode::Thinking {
752 budget_tokens: Some(4_096),
753 }
754 } else {
755 AnthropicModelMode::Default
756 },
757 );
758 let client = self.client.clone();
759 let llm_api_token = self.llm_api_token.clone();
760 let future = self.request_limiter.stream(async move {
761 let PerformLlmCompletionResponse {
762 response,
763 includes_status_messages,
764 } = Self::perform_llm_completion(
765 client.clone(),
766 llm_api_token,
767 app_version,
768 CompletionBody {
769 thread_id,
770 prompt_id,
771 intent,
772 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
773 model: request.model.clone(),
774 provider_request: serde_json::to_value(&request)
775 .map_err(|e| anyhow!(e))?,
776 },
777 )
778 .await
779 .map_err(|err| match err.downcast::<ApiError>() {
780 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
781 Err(err) => anyhow!(err),
782 })?;
783
784 let mut mapper = AnthropicEventMapper::new();
785 Ok(map_cloud_completion_events(
786 Box::pin(response_lines(response, includes_status_messages)),
787 &provider_name,
788 move |event| mapper.map_event(event),
789 ))
790 });
791 async move { Ok(future.await?.boxed()) }.boxed()
792 }
793 cloud_llm_client::LanguageModelProvider::OpenAi => {
794 let client = self.client.clone();
795 let llm_api_token = self.llm_api_token.clone();
796
797 let request = into_open_ai_response(
798 request,
799 &self.model.id.0,
800 self.model.supports_parallel_tool_calls,
801 true,
802 None,
803 None,
804 );
805 let future = self.request_limiter.stream(async move {
806 let PerformLlmCompletionResponse {
807 response,
808 includes_status_messages,
809 } = Self::perform_llm_completion(
810 client.clone(),
811 llm_api_token,
812 app_version,
813 CompletionBody {
814 thread_id,
815 prompt_id,
816 intent,
817 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
818 model: request.model.clone(),
819 provider_request: serde_json::to_value(&request)
820 .map_err(|e| anyhow!(e))?,
821 },
822 )
823 .await?;
824
825 let mut mapper = OpenAiResponseEventMapper::new();
826 Ok(map_cloud_completion_events(
827 Box::pin(response_lines(response, includes_status_messages)),
828 &provider_name,
829 move |event| mapper.map_event(event),
830 ))
831 });
832 async move { Ok(future.await?.boxed()) }.boxed()
833 }
834 cloud_llm_client::LanguageModelProvider::XAi => {
835 let client = self.client.clone();
836 let request = into_open_ai(
837 request,
838 &self.model.id.0,
839 self.model.supports_parallel_tool_calls,
840 false,
841 None,
842 None,
843 );
844 let llm_api_token = self.llm_api_token.clone();
845 let future = self.request_limiter.stream(async move {
846 let PerformLlmCompletionResponse {
847 response,
848 includes_status_messages,
849 } = Self::perform_llm_completion(
850 client.clone(),
851 llm_api_token,
852 app_version,
853 CompletionBody {
854 thread_id,
855 prompt_id,
856 intent,
857 provider: cloud_llm_client::LanguageModelProvider::XAi,
858 model: request.model.clone(),
859 provider_request: serde_json::to_value(&request)
860 .map_err(|e| anyhow!(e))?,
861 },
862 )
863 .await?;
864
865 let mut mapper = OpenAiEventMapper::new();
866 Ok(map_cloud_completion_events(
867 Box::pin(response_lines(response, includes_status_messages)),
868 &provider_name,
869 move |event| mapper.map_event(event),
870 ))
871 });
872 async move { Ok(future.await?.boxed()) }.boxed()
873 }
874 cloud_llm_client::LanguageModelProvider::Google => {
875 let client = self.client.clone();
876 let request =
877 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
878 let llm_api_token = self.llm_api_token.clone();
879 let future = self.request_limiter.stream(async move {
880 let PerformLlmCompletionResponse {
881 response,
882 includes_status_messages,
883 } = Self::perform_llm_completion(
884 client.clone(),
885 llm_api_token,
886 app_version,
887 CompletionBody {
888 thread_id,
889 prompt_id,
890 intent,
891 provider: cloud_llm_client::LanguageModelProvider::Google,
892 model: request.model.model_id.clone(),
893 provider_request: serde_json::to_value(&request)
894 .map_err(|e| anyhow!(e))?,
895 },
896 )
897 .await?;
898
899 let mut mapper = GoogleEventMapper::new();
900 Ok(map_cloud_completion_events(
901 Box::pin(response_lines(response, includes_status_messages)),
902 &provider_name,
903 move |event| mapper.map_event(event),
904 ))
905 });
906 async move { Ok(future.await?.boxed()) }.boxed()
907 }
908 }
909 }
910}
911
912fn map_cloud_completion_events<T, F>(
913 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
914 provider: &LanguageModelProviderName,
915 mut map_callback: F,
916) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
917where
918 T: DeserializeOwned + 'static,
919 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
920 + Send
921 + 'static,
922{
923 let provider = provider.clone();
924 stream
925 .flat_map(move |event| {
926 futures::stream::iter(match event {
927 Err(error) => {
928 vec![Err(LanguageModelCompletionError::from(error))]
929 }
930 Ok(CompletionEvent::Status(event)) => {
931 vec![
932 LanguageModelCompletionEvent::from_completion_request_status(
933 event,
934 provider.clone(),
935 ),
936 ]
937 }
938 Ok(CompletionEvent::Event(event)) => map_callback(event),
939 })
940 })
941 .boxed()
942}
943
944fn provider_name(provider: &cloud_llm_client::LanguageModelProvider) -> LanguageModelProviderName {
945 match provider {
946 cloud_llm_client::LanguageModelProvider::Anthropic => {
947 language_model::ANTHROPIC_PROVIDER_NAME
948 }
949 cloud_llm_client::LanguageModelProvider::OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
950 cloud_llm_client::LanguageModelProvider::Google => language_model::GOOGLE_PROVIDER_NAME,
951 cloud_llm_client::LanguageModelProvider::XAi => language_model::X_AI_PROVIDER_NAME,
952 }
953}
954
955fn response_lines<T: DeserializeOwned>(
956 response: Response<AsyncBody>,
957 includes_status_messages: bool,
958) -> impl Stream<Item = Result<CompletionEvent<T>>> {
959 futures::stream::try_unfold(
960 (String::new(), BufReader::new(response.into_body())),
961 move |(mut line, mut body)| async move {
962 match body.read_line(&mut line).await {
963 Ok(0) => Ok(None),
964 Ok(_) => {
965 let event = if includes_status_messages {
966 serde_json::from_str::<CompletionEvent<T>>(&line)?
967 } else {
968 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
969 };
970
971 line.clear();
972 Ok(Some((event, (line, body))))
973 }
974 Err(e) => Err(e.into()),
975 }
976 },
977 )
978}
979
980#[derive(IntoElement, RegisterComponent)]
981struct ZedAiConfiguration {
982 is_connected: bool,
983 plan: Option<Plan>,
984 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
985 eligible_for_trial: bool,
986 account_too_young: bool,
987 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
988}
989
990impl RenderOnce for ZedAiConfiguration {
991 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
992 let is_pro = self
993 .plan
994 .is_some_and(|plan| plan == Plan::V2(PlanV2::ZedPro));
995 let subscription_text = match (self.plan, self.subscription_period) {
996 (Some(Plan::V2(PlanV2::ZedPro)), Some(_)) => {
997 "You have access to Zed's hosted models through your Pro subscription."
998 }
999 (Some(Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1000 "You have access to Zed's hosted models through your Pro trial."
1001 }
1002 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1003 if self.eligible_for_trial {
1004 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1005 } else {
1006 "Subscribe for access to Zed's hosted models."
1007 }
1008 }
1009 _ => {
1010 if self.eligible_for_trial {
1011 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1012 } else {
1013 "Subscribe for access to Zed's hosted models."
1014 }
1015 }
1016 };
1017
1018 let manage_subscription_buttons = if is_pro {
1019 Button::new("manage_settings", "Manage Subscription")
1020 .full_width()
1021 .style(ButtonStyle::Tinted(TintColor::Accent))
1022 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1023 .into_any_element()
1024 } else if self.plan.is_none() || self.eligible_for_trial {
1025 Button::new("start_trial", "Start 14-day Free Pro Trial")
1026 .full_width()
1027 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1028 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1029 .into_any_element()
1030 } else {
1031 Button::new("upgrade", "Upgrade to Pro")
1032 .full_width()
1033 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1034 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1035 .into_any_element()
1036 };
1037
1038 if !self.is_connected {
1039 return v_flex()
1040 .gap_2()
1041 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1042 .child(
1043 Button::new("sign_in", "Sign In to use Zed AI")
1044 .icon_color(Color::Muted)
1045 .icon(IconName::Github)
1046 .icon_size(IconSize::Small)
1047 .icon_position(IconPosition::Start)
1048 .full_width()
1049 .on_click({
1050 let callback = self.sign_in_callback.clone();
1051 move |_, window, cx| (callback)(window, cx)
1052 }),
1053 );
1054 }
1055
1056 v_flex().gap_2().w_full().map(|this| {
1057 if self.account_too_young {
1058 this.child(YoungAccountBanner).child(
1059 Button::new("upgrade", "Upgrade to Pro")
1060 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1061 .full_width()
1062 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1063 )
1064 } else {
1065 this.text_sm()
1066 .child(subscription_text)
1067 .child(manage_subscription_buttons)
1068 }
1069 })
1070 }
1071}
1072
1073struct ConfigurationView {
1074 state: Entity<State>,
1075 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1076}
1077
1078impl ConfigurationView {
1079 fn new(state: Entity<State>) -> Self {
1080 let sign_in_callback = Arc::new({
1081 let state = state.clone();
1082 move |_window: &mut Window, cx: &mut App| {
1083 state.update(cx, |state, cx| {
1084 state.authenticate(cx).detach_and_log_err(cx);
1085 });
1086 }
1087 });
1088
1089 Self {
1090 state,
1091 sign_in_callback,
1092 }
1093 }
1094}
1095
1096impl Render for ConfigurationView {
1097 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1098 let state = self.state.read(cx);
1099 let user_store = state.user_store.read(cx);
1100
1101 ZedAiConfiguration {
1102 is_connected: !state.is_signed_out(cx),
1103 plan: user_store.plan(),
1104 subscription_period: user_store.subscription_period(),
1105 eligible_for_trial: user_store.trial_started_at().is_none(),
1106 account_too_young: user_store.account_too_young(),
1107 sign_in_callback: self.sign_in_callback.clone(),
1108 }
1109 }
1110}
1111
1112impl Component for ZedAiConfiguration {
1113 fn name() -> &'static str {
1114 "AI Configuration Content"
1115 }
1116
1117 fn sort_name() -> &'static str {
1118 "AI Configuration Content"
1119 }
1120
1121 fn scope() -> ComponentScope {
1122 ComponentScope::Onboarding
1123 }
1124
1125 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1126 fn configuration(
1127 is_connected: bool,
1128 plan: Option<Plan>,
1129 eligible_for_trial: bool,
1130 account_too_young: bool,
1131 ) -> AnyElement {
1132 ZedAiConfiguration {
1133 is_connected,
1134 plan,
1135 subscription_period: plan
1136 .is_some()
1137 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1138 eligible_for_trial,
1139 account_too_young,
1140 sign_in_callback: Arc::new(|_, _| {}),
1141 }
1142 .into_any_element()
1143 }
1144
1145 Some(
1146 v_flex()
1147 .p_4()
1148 .gap_4()
1149 .children(vec![
1150 single_example("Not connected", configuration(false, None, false, false)),
1151 single_example(
1152 "Accept Terms of Service",
1153 configuration(true, None, true, false),
1154 ),
1155 single_example(
1156 "No Plan - Not eligible for trial",
1157 configuration(true, None, false, false),
1158 ),
1159 single_example(
1160 "No Plan - Eligible for trial",
1161 configuration(true, None, true, false),
1162 ),
1163 single_example(
1164 "Free Plan",
1165 configuration(true, Some(Plan::V2(PlanV2::ZedFree)), true, false),
1166 ),
1167 single_example(
1168 "Zed Pro Trial Plan",
1169 configuration(true, Some(Plan::V2(PlanV2::ZedProTrial)), true, false),
1170 ),
1171 single_example(
1172 "Zed Pro Plan",
1173 configuration(true, Some(Plan::V2(PlanV2::ZedPro)), true, false),
1174 ),
1175 ])
1176 .into_any_element(),
1177 )
1178 }
1179}
1180
1181#[cfg(test)]
1182mod tests {
1183 use super::*;
1184 use http_client::http::{HeaderMap, StatusCode};
1185 use language_model::LanguageModelCompletionError;
1186
1187 #[test]
1188 fn test_api_error_conversion_with_upstream_http_error() {
1189 // upstream_http_error with 503 status should become ServerOverloaded
1190 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}"#;
1191
1192 let api_error = ApiError {
1193 status: StatusCode::INTERNAL_SERVER_ERROR,
1194 body: error_body.to_string(),
1195 headers: HeaderMap::new(),
1196 };
1197
1198 let completion_error: LanguageModelCompletionError = api_error.into();
1199
1200 match completion_error {
1201 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1202 assert_eq!(
1203 message,
1204 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1205 );
1206 }
1207 _ => panic!(
1208 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1209 completion_error
1210 ),
1211 }
1212
1213 // upstream_http_error with 500 status should become ApiInternalServerError
1214 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1215
1216 let api_error = ApiError {
1217 status: StatusCode::INTERNAL_SERVER_ERROR,
1218 body: error_body.to_string(),
1219 headers: HeaderMap::new(),
1220 };
1221
1222 let completion_error: LanguageModelCompletionError = api_error.into();
1223
1224 match completion_error {
1225 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1226 assert_eq!(
1227 message,
1228 "Received an error from the OpenAI API: internal server error"
1229 );
1230 }
1231 _ => panic!(
1232 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1233 completion_error
1234 ),
1235 }
1236
1237 // upstream_http_error with 429 status should become RateLimitExceeded
1238 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1239
1240 let api_error = ApiError {
1241 status: StatusCode::INTERNAL_SERVER_ERROR,
1242 body: error_body.to_string(),
1243 headers: HeaderMap::new(),
1244 };
1245
1246 let completion_error: LanguageModelCompletionError = api_error.into();
1247
1248 match completion_error {
1249 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1250 assert_eq!(
1251 message,
1252 "Received an error from the Google API: rate limit exceeded"
1253 );
1254 }
1255 _ => panic!(
1256 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1257 completion_error
1258 ),
1259 }
1260
1261 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1262 let error_body = "Regular internal server error";
1263
1264 let api_error = ApiError {
1265 status: StatusCode::INTERNAL_SERVER_ERROR,
1266 body: error_body.to_string(),
1267 headers: HeaderMap::new(),
1268 };
1269
1270 let completion_error: LanguageModelCompletionError = api_error.into();
1271
1272 match completion_error {
1273 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1274 assert_eq!(provider, PROVIDER_NAME);
1275 assert_eq!(message, "Regular internal server error");
1276 }
1277 _ => panic!(
1278 "Expected ApiInternalServerError for regular 500, got: {:?}",
1279 completion_error
1280 ),
1281 }
1282
1283 // upstream_http_429 format should be converted to UpstreamProviderError
1284 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1285
1286 let api_error = ApiError {
1287 status: StatusCode::INTERNAL_SERVER_ERROR,
1288 body: error_body.to_string(),
1289 headers: HeaderMap::new(),
1290 };
1291
1292 let completion_error: LanguageModelCompletionError = api_error.into();
1293
1294 match completion_error {
1295 LanguageModelCompletionError::UpstreamProviderError {
1296 message,
1297 status,
1298 retry_after,
1299 } => {
1300 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1301 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1302 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1303 }
1304 _ => panic!(
1305 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1306 completion_error
1307 ),
1308 }
1309
1310 // Invalid JSON in error body should fall back to regular error handling
1311 let error_body = "Not JSON at all";
1312
1313 let api_error = ApiError {
1314 status: StatusCode::INTERNAL_SERVER_ERROR,
1315 body: error_body.to_string(),
1316 headers: HeaderMap::new(),
1317 };
1318
1319 let completion_error: LanguageModelCompletionError = api_error.into();
1320
1321 match completion_error {
1322 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1323 assert_eq!(provider, PROVIDER_NAME);
1324 }
1325 _ => panic!(
1326 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1327 completion_error
1328 ),
1329 }
1330 }
1331}