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 bypass_rate_limit = request.bypass_rate_limit;
734 let app_version = Some(cx.update(|cx| AppVersion::global(cx)));
735 let thinking_allowed = request.thinking_allowed;
736 let is_thinking_toggle_enabled =
737 cx.update(|cx| cx.has_flag::<CloudThinkingToggleFeatureFlag>());
738 let enable_thinking = if is_thinking_toggle_enabled {
739 thinking_allowed && self.model.supports_thinking
740 } else {
741 thinking_allowed && self.model.id.0.ends_with("-thinking")
742 };
743 let provider_name = provider_name(&self.model.provider);
744 match self.model.provider {
745 cloud_llm_client::LanguageModelProvider::Anthropic => {
746 let request = into_anthropic(
747 request,
748 self.model.id.to_string(),
749 1.0,
750 self.model.max_output_tokens as u64,
751 if enable_thinking {
752 AnthropicModelMode::Thinking {
753 budget_tokens: Some(4_096),
754 }
755 } else {
756 AnthropicModelMode::Default
757 },
758 );
759 let client = self.client.clone();
760 let llm_api_token = self.llm_api_token.clone();
761 let future = self.request_limiter.stream_with_bypass(
762 async move {
763 let PerformLlmCompletionResponse {
764 response,
765 includes_status_messages,
766 } = Self::perform_llm_completion(
767 client.clone(),
768 llm_api_token,
769 app_version,
770 CompletionBody {
771 thread_id,
772 prompt_id,
773 intent,
774 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
775 model: request.model.clone(),
776 provider_request: serde_json::to_value(&request)
777 .map_err(|e| anyhow!(e))?,
778 },
779 )
780 .await
781 .map_err(|err| {
782 match err.downcast::<ApiError>() {
783 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
784 Err(err) => anyhow!(err),
785 }
786 })?;
787
788 let mut mapper = AnthropicEventMapper::new();
789 Ok(map_cloud_completion_events(
790 Box::pin(response_lines(response, includes_status_messages)),
791 &provider_name,
792 move |event| mapper.map_event(event),
793 ))
794 },
795 bypass_rate_limit,
796 );
797 async move { Ok(future.await?.boxed()) }.boxed()
798 }
799 cloud_llm_client::LanguageModelProvider::OpenAi => {
800 let client = self.client.clone();
801 let llm_api_token = self.llm_api_token.clone();
802
803 let request = into_open_ai_response(
804 request,
805 &self.model.id.0,
806 self.model.supports_parallel_tool_calls,
807 true,
808 None,
809 None,
810 );
811 let future = self.request_limiter.stream_with_bypass(
812 async move {
813 let PerformLlmCompletionResponse {
814 response,
815 includes_status_messages,
816 } = Self::perform_llm_completion(
817 client.clone(),
818 llm_api_token,
819 app_version,
820 CompletionBody {
821 thread_id,
822 prompt_id,
823 intent,
824 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
825 model: request.model.clone(),
826 provider_request: serde_json::to_value(&request)
827 .map_err(|e| anyhow!(e))?,
828 },
829 )
830 .await?;
831
832 let mut mapper = OpenAiResponseEventMapper::new();
833 Ok(map_cloud_completion_events(
834 Box::pin(response_lines(response, includes_status_messages)),
835 &provider_name,
836 move |event| mapper.map_event(event),
837 ))
838 },
839 bypass_rate_limit,
840 );
841 async move { Ok(future.await?.boxed()) }.boxed()
842 }
843 cloud_llm_client::LanguageModelProvider::XAi => {
844 let client = self.client.clone();
845 let request = into_open_ai(
846 request,
847 &self.model.id.0,
848 self.model.supports_parallel_tool_calls,
849 false,
850 None,
851 None,
852 );
853 let llm_api_token = self.llm_api_token.clone();
854 let future = self.request_limiter.stream_with_bypass(
855 async move {
856 let PerformLlmCompletionResponse {
857 response,
858 includes_status_messages,
859 } = Self::perform_llm_completion(
860 client.clone(),
861 llm_api_token,
862 app_version,
863 CompletionBody {
864 thread_id,
865 prompt_id,
866 intent,
867 provider: cloud_llm_client::LanguageModelProvider::XAi,
868 model: request.model.clone(),
869 provider_request: serde_json::to_value(&request)
870 .map_err(|e| anyhow!(e))?,
871 },
872 )
873 .await?;
874
875 let mut mapper = OpenAiEventMapper::new();
876 Ok(map_cloud_completion_events(
877 Box::pin(response_lines(response, includes_status_messages)),
878 &provider_name,
879 move |event| mapper.map_event(event),
880 ))
881 },
882 bypass_rate_limit,
883 );
884 async move { Ok(future.await?.boxed()) }.boxed()
885 }
886 cloud_llm_client::LanguageModelProvider::Google => {
887 let client = self.client.clone();
888 let request =
889 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
890 let llm_api_token = self.llm_api_token.clone();
891 let future = self.request_limiter.stream_with_bypass(
892 async move {
893 let PerformLlmCompletionResponse {
894 response,
895 includes_status_messages,
896 } = Self::perform_llm_completion(
897 client.clone(),
898 llm_api_token,
899 app_version,
900 CompletionBody {
901 thread_id,
902 prompt_id,
903 intent,
904 provider: cloud_llm_client::LanguageModelProvider::Google,
905 model: request.model.model_id.clone(),
906 provider_request: serde_json::to_value(&request)
907 .map_err(|e| anyhow!(e))?,
908 },
909 )
910 .await?;
911
912 let mut mapper = GoogleEventMapper::new();
913 Ok(map_cloud_completion_events(
914 Box::pin(response_lines(response, includes_status_messages)),
915 &provider_name,
916 move |event| mapper.map_event(event),
917 ))
918 },
919 bypass_rate_limit,
920 );
921 async move { Ok(future.await?.boxed()) }.boxed()
922 }
923 }
924 }
925}
926
927fn map_cloud_completion_events<T, F>(
928 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
929 provider: &LanguageModelProviderName,
930 mut map_callback: F,
931) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
932where
933 T: DeserializeOwned + 'static,
934 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
935 + Send
936 + 'static,
937{
938 let provider = provider.clone();
939 stream
940 .flat_map(move |event| {
941 futures::stream::iter(match event {
942 Err(error) => {
943 vec![Err(LanguageModelCompletionError::from(error))]
944 }
945 Ok(CompletionEvent::Status(event)) => {
946 vec![
947 LanguageModelCompletionEvent::from_completion_request_status(
948 event,
949 provider.clone(),
950 ),
951 ]
952 }
953 Ok(CompletionEvent::Event(event)) => map_callback(event),
954 })
955 })
956 .boxed()
957}
958
959fn provider_name(provider: &cloud_llm_client::LanguageModelProvider) -> LanguageModelProviderName {
960 match provider {
961 cloud_llm_client::LanguageModelProvider::Anthropic => {
962 language_model::ANTHROPIC_PROVIDER_NAME
963 }
964 cloud_llm_client::LanguageModelProvider::OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
965 cloud_llm_client::LanguageModelProvider::Google => language_model::GOOGLE_PROVIDER_NAME,
966 cloud_llm_client::LanguageModelProvider::XAi => language_model::X_AI_PROVIDER_NAME,
967 }
968}
969
970fn response_lines<T: DeserializeOwned>(
971 response: Response<AsyncBody>,
972 includes_status_messages: bool,
973) -> impl Stream<Item = Result<CompletionEvent<T>>> {
974 futures::stream::try_unfold(
975 (String::new(), BufReader::new(response.into_body())),
976 move |(mut line, mut body)| async move {
977 match body.read_line(&mut line).await {
978 Ok(0) => Ok(None),
979 Ok(_) => {
980 let event = if includes_status_messages {
981 serde_json::from_str::<CompletionEvent<T>>(&line)?
982 } else {
983 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
984 };
985
986 line.clear();
987 Ok(Some((event, (line, body))))
988 }
989 Err(e) => Err(e.into()),
990 }
991 },
992 )
993}
994
995#[derive(IntoElement, RegisterComponent)]
996struct ZedAiConfiguration {
997 is_connected: bool,
998 plan: Option<Plan>,
999 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1000 eligible_for_trial: bool,
1001 account_too_young: bool,
1002 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1003}
1004
1005impl RenderOnce for ZedAiConfiguration {
1006 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1007 let is_pro = self
1008 .plan
1009 .is_some_and(|plan| plan == Plan::V2(PlanV2::ZedPro));
1010 let subscription_text = match (self.plan, self.subscription_period) {
1011 (Some(Plan::V2(PlanV2::ZedPro)), Some(_)) => {
1012 "You have access to Zed's hosted models through your Pro subscription."
1013 }
1014 (Some(Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1015 "You have access to Zed's hosted models through your Pro trial."
1016 }
1017 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1018 if self.eligible_for_trial {
1019 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1020 } else {
1021 "Subscribe for access to Zed's hosted models."
1022 }
1023 }
1024 _ => {
1025 if self.eligible_for_trial {
1026 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1027 } else {
1028 "Subscribe for access to Zed's hosted models."
1029 }
1030 }
1031 };
1032
1033 let manage_subscription_buttons = if is_pro {
1034 Button::new("manage_settings", "Manage Subscription")
1035 .full_width()
1036 .style(ButtonStyle::Tinted(TintColor::Accent))
1037 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1038 .into_any_element()
1039 } else if self.plan.is_none() || self.eligible_for_trial {
1040 Button::new("start_trial", "Start 14-day Free Pro Trial")
1041 .full_width()
1042 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1043 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1044 .into_any_element()
1045 } else {
1046 Button::new("upgrade", "Upgrade to Pro")
1047 .full_width()
1048 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1049 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1050 .into_any_element()
1051 };
1052
1053 if !self.is_connected {
1054 return v_flex()
1055 .gap_2()
1056 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1057 .child(
1058 Button::new("sign_in", "Sign In to use Zed AI")
1059 .icon_color(Color::Muted)
1060 .icon(IconName::Github)
1061 .icon_size(IconSize::Small)
1062 .icon_position(IconPosition::Start)
1063 .full_width()
1064 .on_click({
1065 let callback = self.sign_in_callback.clone();
1066 move |_, window, cx| (callback)(window, cx)
1067 }),
1068 );
1069 }
1070
1071 v_flex().gap_2().w_full().map(|this| {
1072 if self.account_too_young {
1073 this.child(YoungAccountBanner).child(
1074 Button::new("upgrade", "Upgrade to Pro")
1075 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1076 .full_width()
1077 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1078 )
1079 } else {
1080 this.text_sm()
1081 .child(subscription_text)
1082 .child(manage_subscription_buttons)
1083 }
1084 })
1085 }
1086}
1087
1088struct ConfigurationView {
1089 state: Entity<State>,
1090 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1091}
1092
1093impl ConfigurationView {
1094 fn new(state: Entity<State>) -> Self {
1095 let sign_in_callback = Arc::new({
1096 let state = state.clone();
1097 move |_window: &mut Window, cx: &mut App| {
1098 state.update(cx, |state, cx| {
1099 state.authenticate(cx).detach_and_log_err(cx);
1100 });
1101 }
1102 });
1103
1104 Self {
1105 state,
1106 sign_in_callback,
1107 }
1108 }
1109}
1110
1111impl Render for ConfigurationView {
1112 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1113 let state = self.state.read(cx);
1114 let user_store = state.user_store.read(cx);
1115
1116 ZedAiConfiguration {
1117 is_connected: !state.is_signed_out(cx),
1118 plan: user_store.plan(),
1119 subscription_period: user_store.subscription_period(),
1120 eligible_for_trial: user_store.trial_started_at().is_none(),
1121 account_too_young: user_store.account_too_young(),
1122 sign_in_callback: self.sign_in_callback.clone(),
1123 }
1124 }
1125}
1126
1127impl Component for ZedAiConfiguration {
1128 fn name() -> &'static str {
1129 "AI Configuration Content"
1130 }
1131
1132 fn sort_name() -> &'static str {
1133 "AI Configuration Content"
1134 }
1135
1136 fn scope() -> ComponentScope {
1137 ComponentScope::Onboarding
1138 }
1139
1140 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1141 fn configuration(
1142 is_connected: bool,
1143 plan: Option<Plan>,
1144 eligible_for_trial: bool,
1145 account_too_young: bool,
1146 ) -> AnyElement {
1147 ZedAiConfiguration {
1148 is_connected,
1149 plan,
1150 subscription_period: plan
1151 .is_some()
1152 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1153 eligible_for_trial,
1154 account_too_young,
1155 sign_in_callback: Arc::new(|_, _| {}),
1156 }
1157 .into_any_element()
1158 }
1159
1160 Some(
1161 v_flex()
1162 .p_4()
1163 .gap_4()
1164 .children(vec![
1165 single_example("Not connected", configuration(false, None, false, false)),
1166 single_example(
1167 "Accept Terms of Service",
1168 configuration(true, None, true, false),
1169 ),
1170 single_example(
1171 "No Plan - Not eligible for trial",
1172 configuration(true, None, false, false),
1173 ),
1174 single_example(
1175 "No Plan - Eligible for trial",
1176 configuration(true, None, true, false),
1177 ),
1178 single_example(
1179 "Free Plan",
1180 configuration(true, Some(Plan::V2(PlanV2::ZedFree)), true, false),
1181 ),
1182 single_example(
1183 "Zed Pro Trial Plan",
1184 configuration(true, Some(Plan::V2(PlanV2::ZedProTrial)), true, false),
1185 ),
1186 single_example(
1187 "Zed Pro Plan",
1188 configuration(true, Some(Plan::V2(PlanV2::ZedPro)), true, false),
1189 ),
1190 ])
1191 .into_any_element(),
1192 )
1193 }
1194}
1195
1196#[cfg(test)]
1197mod tests {
1198 use super::*;
1199 use http_client::http::{HeaderMap, StatusCode};
1200 use language_model::LanguageModelCompletionError;
1201
1202 #[test]
1203 fn test_api_error_conversion_with_upstream_http_error() {
1204 // upstream_http_error with 503 status should become ServerOverloaded
1205 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}"#;
1206
1207 let api_error = ApiError {
1208 status: StatusCode::INTERNAL_SERVER_ERROR,
1209 body: error_body.to_string(),
1210 headers: HeaderMap::new(),
1211 };
1212
1213 let completion_error: LanguageModelCompletionError = api_error.into();
1214
1215 match completion_error {
1216 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1217 assert_eq!(
1218 message,
1219 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1220 );
1221 }
1222 _ => panic!(
1223 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1224 completion_error
1225 ),
1226 }
1227
1228 // upstream_http_error with 500 status should become ApiInternalServerError
1229 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1230
1231 let api_error = ApiError {
1232 status: StatusCode::INTERNAL_SERVER_ERROR,
1233 body: error_body.to_string(),
1234 headers: HeaderMap::new(),
1235 };
1236
1237 let completion_error: LanguageModelCompletionError = api_error.into();
1238
1239 match completion_error {
1240 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1241 assert_eq!(
1242 message,
1243 "Received an error from the OpenAI API: internal server error"
1244 );
1245 }
1246 _ => panic!(
1247 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1248 completion_error
1249 ),
1250 }
1251
1252 // upstream_http_error with 429 status should become RateLimitExceeded
1253 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1254
1255 let api_error = ApiError {
1256 status: StatusCode::INTERNAL_SERVER_ERROR,
1257 body: error_body.to_string(),
1258 headers: HeaderMap::new(),
1259 };
1260
1261 let completion_error: LanguageModelCompletionError = api_error.into();
1262
1263 match completion_error {
1264 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1265 assert_eq!(
1266 message,
1267 "Received an error from the Google API: rate limit exceeded"
1268 );
1269 }
1270 _ => panic!(
1271 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1272 completion_error
1273 ),
1274 }
1275
1276 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1277 let error_body = "Regular internal server error";
1278
1279 let api_error = ApiError {
1280 status: StatusCode::INTERNAL_SERVER_ERROR,
1281 body: error_body.to_string(),
1282 headers: HeaderMap::new(),
1283 };
1284
1285 let completion_error: LanguageModelCompletionError = api_error.into();
1286
1287 match completion_error {
1288 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1289 assert_eq!(provider, PROVIDER_NAME);
1290 assert_eq!(message, "Regular internal server error");
1291 }
1292 _ => panic!(
1293 "Expected ApiInternalServerError for regular 500, got: {:?}",
1294 completion_error
1295 ),
1296 }
1297
1298 // upstream_http_429 format should be converted to UpstreamProviderError
1299 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1300
1301 let api_error = ApiError {
1302 status: StatusCode::INTERNAL_SERVER_ERROR,
1303 body: error_body.to_string(),
1304 headers: HeaderMap::new(),
1305 };
1306
1307 let completion_error: LanguageModelCompletionError = api_error.into();
1308
1309 match completion_error {
1310 LanguageModelCompletionError::UpstreamProviderError {
1311 message,
1312 status,
1313 retry_after,
1314 } => {
1315 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1316 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1317 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1318 }
1319 _ => panic!(
1320 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1321 completion_error
1322 ),
1323 }
1324
1325 // Invalid JSON in error body should fall back to regular error handling
1326 let error_body = "Not JSON at all";
1327
1328 let api_error = ApiError {
1329 status: StatusCode::INTERNAL_SERVER_ERROR,
1330 body: error_body.to_string(),
1331 headers: HeaderMap::new(),
1332 };
1333
1334 let completion_error: LanguageModelCompletionError = api_error.into();
1335
1336 match completion_error {
1337 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1338 assert_eq!(provider, PROVIDER_NAME);
1339 }
1340 _ => panic!(
1341 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1342 completion_error
1343 ),
1344 }
1345 }
1346}