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;
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, LanguageModelEffortLevel,
23 LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
24 LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,
25 LanguageModelToolChoice, LanguageModelToolSchemaFormat, LlmApiToken, NeedsLlmTokenRefresh,
26 PaymentRequiredError, 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 supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {
581 self.model
582 .supported_effort_levels
583 .iter()
584 .map(|effort_level| LanguageModelEffortLevel {
585 name: effort_level.name.clone().into(),
586 value: effort_level.value.clone().into(),
587 })
588 .collect()
589 }
590
591 fn supports_streaming_tools(&self) -> bool {
592 self.model.supports_streaming_tools
593 }
594
595 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
596 match choice {
597 LanguageModelToolChoice::Auto
598 | LanguageModelToolChoice::Any
599 | LanguageModelToolChoice::None => true,
600 }
601 }
602
603 fn supports_split_token_display(&self) -> bool {
604 use cloud_llm_client::LanguageModelProvider::*;
605 matches!(self.model.provider, OpenAi)
606 }
607
608 fn telemetry_id(&self) -> String {
609 format!("zed.dev/{}", self.model.id)
610 }
611
612 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
613 match self.model.provider {
614 cloud_llm_client::LanguageModelProvider::Anthropic
615 | cloud_llm_client::LanguageModelProvider::OpenAi
616 | cloud_llm_client::LanguageModelProvider::XAi => {
617 LanguageModelToolSchemaFormat::JsonSchema
618 }
619 cloud_llm_client::LanguageModelProvider::Google => {
620 LanguageModelToolSchemaFormat::JsonSchemaSubset
621 }
622 }
623 }
624
625 fn max_token_count(&self) -> u64 {
626 self.model.max_token_count as u64
627 }
628
629 fn max_output_tokens(&self) -> Option<u64> {
630 Some(self.model.max_output_tokens as u64)
631 }
632
633 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
634 match &self.model.provider {
635 cloud_llm_client::LanguageModelProvider::Anthropic => {
636 Some(LanguageModelCacheConfiguration {
637 min_total_token: 2_048,
638 should_speculate: true,
639 max_cache_anchors: 4,
640 })
641 }
642 cloud_llm_client::LanguageModelProvider::OpenAi
643 | cloud_llm_client::LanguageModelProvider::XAi
644 | cloud_llm_client::LanguageModelProvider::Google => None,
645 }
646 }
647
648 fn count_tokens(
649 &self,
650 request: LanguageModelRequest,
651 cx: &App,
652 ) -> BoxFuture<'static, Result<u64>> {
653 match self.model.provider {
654 cloud_llm_client::LanguageModelProvider::Anthropic => cx
655 .background_spawn(async move { count_anthropic_tokens_with_tiktoken(request) })
656 .boxed(),
657 cloud_llm_client::LanguageModelProvider::OpenAi => {
658 let model = match open_ai::Model::from_id(&self.model.id.0) {
659 Ok(model) => model,
660 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
661 };
662 count_open_ai_tokens(request, model, cx)
663 }
664 cloud_llm_client::LanguageModelProvider::XAi => {
665 let model = match x_ai::Model::from_id(&self.model.id.0) {
666 Ok(model) => model,
667 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
668 };
669 count_xai_tokens(request, model, cx)
670 }
671 cloud_llm_client::LanguageModelProvider::Google => {
672 let client = self.client.clone();
673 let llm_api_token = self.llm_api_token.clone();
674 let model_id = self.model.id.to_string();
675 let generate_content_request =
676 into_google(request, model_id.clone(), GoogleModelMode::Default);
677 async move {
678 let http_client = &client.http_client();
679 let token = llm_api_token.acquire(&client).await?;
680
681 let request_body = CountTokensBody {
682 provider: cloud_llm_client::LanguageModelProvider::Google,
683 model: model_id,
684 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
685 generate_content_request,
686 })?,
687 };
688 let request = http_client::Request::builder()
689 .method(Method::POST)
690 .uri(
691 http_client
692 .build_zed_llm_url("/count_tokens", &[])?
693 .as_ref(),
694 )
695 .header("Content-Type", "application/json")
696 .header("Authorization", format!("Bearer {token}"))
697 .body(serde_json::to_string(&request_body)?.into())?;
698 let mut response = http_client.send(request).await?;
699 let status = response.status();
700 let headers = response.headers().clone();
701 let mut response_body = String::new();
702 response
703 .body_mut()
704 .read_to_string(&mut response_body)
705 .await?;
706
707 if status.is_success() {
708 let response_body: CountTokensResponse =
709 serde_json::from_str(&response_body)?;
710
711 Ok(response_body.tokens as u64)
712 } else {
713 Err(anyhow!(ApiError {
714 status,
715 body: response_body,
716 headers
717 }))
718 }
719 }
720 .boxed()
721 }
722 }
723 }
724
725 fn stream_completion(
726 &self,
727 request: LanguageModelRequest,
728 cx: &AsyncApp,
729 ) -> BoxFuture<
730 'static,
731 Result<
732 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
733 LanguageModelCompletionError,
734 >,
735 > {
736 let thread_id = request.thread_id.clone();
737 let prompt_id = request.prompt_id.clone();
738 let intent = request.intent;
739 let app_version = Some(cx.update(|cx| AppVersion::global(cx)));
740 let thinking_allowed = request.thinking_allowed;
741 let is_thinking_toggle_enabled =
742 cx.update(|cx| cx.has_flag::<CloudThinkingToggleFeatureFlag>());
743 let enable_thinking = if is_thinking_toggle_enabled {
744 thinking_allowed && self.model.supports_thinking
745 } else {
746 thinking_allowed && self.model.id.0.ends_with("-thinking")
747 };
748 let provider_name = provider_name(&self.model.provider);
749 match self.model.provider {
750 cloud_llm_client::LanguageModelProvider::Anthropic => {
751 let request = into_anthropic(
752 request,
753 self.model.id.to_string(),
754 1.0,
755 self.model.max_output_tokens as u64,
756 if enable_thinking {
757 AnthropicModelMode::Thinking {
758 budget_tokens: Some(4_096),
759 }
760 } else {
761 AnthropicModelMode::Default
762 },
763 );
764 let client = self.client.clone();
765 let llm_api_token = self.llm_api_token.clone();
766 let future = self.request_limiter.stream(async move {
767 let PerformLlmCompletionResponse {
768 response,
769 includes_status_messages,
770 } = Self::perform_llm_completion(
771 client.clone(),
772 llm_api_token,
773 app_version,
774 CompletionBody {
775 thread_id,
776 prompt_id,
777 intent,
778 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
779 model: request.model.clone(),
780 provider_request: serde_json::to_value(&request)
781 .map_err(|e| anyhow!(e))?,
782 },
783 )
784 .await
785 .map_err(|err| match err.downcast::<ApiError>() {
786 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
787 Err(err) => anyhow!(err),
788 })?;
789
790 let mut mapper = AnthropicEventMapper::new();
791 Ok(map_cloud_completion_events(
792 Box::pin(response_lines(response, includes_status_messages)),
793 &provider_name,
794 move |event| mapper.map_event(event),
795 ))
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(async move {
812 let PerformLlmCompletionResponse {
813 response,
814 includes_status_messages,
815 } = Self::perform_llm_completion(
816 client.clone(),
817 llm_api_token,
818 app_version,
819 CompletionBody {
820 thread_id,
821 prompt_id,
822 intent,
823 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
824 model: request.model.clone(),
825 provider_request: serde_json::to_value(&request)
826 .map_err(|e| anyhow!(e))?,
827 },
828 )
829 .await?;
830
831 let mut mapper = OpenAiResponseEventMapper::new();
832 Ok(map_cloud_completion_events(
833 Box::pin(response_lines(response, includes_status_messages)),
834 &provider_name,
835 move |event| mapper.map_event(event),
836 ))
837 });
838 async move { Ok(future.await?.boxed()) }.boxed()
839 }
840 cloud_llm_client::LanguageModelProvider::XAi => {
841 let client = self.client.clone();
842 let request = into_open_ai(
843 request,
844 &self.model.id.0,
845 self.model.supports_parallel_tool_calls,
846 false,
847 None,
848 None,
849 );
850 let llm_api_token = self.llm_api_token.clone();
851 let future = self.request_limiter.stream(async move {
852 let PerformLlmCompletionResponse {
853 response,
854 includes_status_messages,
855 } = Self::perform_llm_completion(
856 client.clone(),
857 llm_api_token,
858 app_version,
859 CompletionBody {
860 thread_id,
861 prompt_id,
862 intent,
863 provider: cloud_llm_client::LanguageModelProvider::XAi,
864 model: request.model.clone(),
865 provider_request: serde_json::to_value(&request)
866 .map_err(|e| anyhow!(e))?,
867 },
868 )
869 .await?;
870
871 let mut mapper = OpenAiEventMapper::new();
872 Ok(map_cloud_completion_events(
873 Box::pin(response_lines(response, includes_status_messages)),
874 &provider_name,
875 move |event| mapper.map_event(event),
876 ))
877 });
878 async move { Ok(future.await?.boxed()) }.boxed()
879 }
880 cloud_llm_client::LanguageModelProvider::Google => {
881 let client = self.client.clone();
882 let request =
883 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
884 let llm_api_token = self.llm_api_token.clone();
885 let future = self.request_limiter.stream(async move {
886 let PerformLlmCompletionResponse {
887 response,
888 includes_status_messages,
889 } = Self::perform_llm_completion(
890 client.clone(),
891 llm_api_token,
892 app_version,
893 CompletionBody {
894 thread_id,
895 prompt_id,
896 intent,
897 provider: cloud_llm_client::LanguageModelProvider::Google,
898 model: request.model.model_id.clone(),
899 provider_request: serde_json::to_value(&request)
900 .map_err(|e| anyhow!(e))?,
901 },
902 )
903 .await?;
904
905 let mut mapper = GoogleEventMapper::new();
906 Ok(map_cloud_completion_events(
907 Box::pin(response_lines(response, includes_status_messages)),
908 &provider_name,
909 move |event| mapper.map_event(event),
910 ))
911 });
912 async move { Ok(future.await?.boxed()) }.boxed()
913 }
914 }
915 }
916}
917
918fn map_cloud_completion_events<T, F>(
919 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
920 provider: &LanguageModelProviderName,
921 mut map_callback: F,
922) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
923where
924 T: DeserializeOwned + 'static,
925 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
926 + Send
927 + 'static,
928{
929 let provider = provider.clone();
930 stream
931 .flat_map(move |event| {
932 futures::stream::iter(match event {
933 Err(error) => {
934 vec![Err(LanguageModelCompletionError::from(error))]
935 }
936 Ok(CompletionEvent::Status(event)) => {
937 vec![
938 LanguageModelCompletionEvent::from_completion_request_status(
939 event,
940 provider.clone(),
941 ),
942 ]
943 }
944 Ok(CompletionEvent::Event(event)) => map_callback(event),
945 })
946 })
947 .boxed()
948}
949
950fn provider_name(provider: &cloud_llm_client::LanguageModelProvider) -> LanguageModelProviderName {
951 match provider {
952 cloud_llm_client::LanguageModelProvider::Anthropic => {
953 language_model::ANTHROPIC_PROVIDER_NAME
954 }
955 cloud_llm_client::LanguageModelProvider::OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
956 cloud_llm_client::LanguageModelProvider::Google => language_model::GOOGLE_PROVIDER_NAME,
957 cloud_llm_client::LanguageModelProvider::XAi => language_model::X_AI_PROVIDER_NAME,
958 }
959}
960
961fn response_lines<T: DeserializeOwned>(
962 response: Response<AsyncBody>,
963 includes_status_messages: bool,
964) -> impl Stream<Item = Result<CompletionEvent<T>>> {
965 futures::stream::try_unfold(
966 (String::new(), BufReader::new(response.into_body())),
967 move |(mut line, mut body)| async move {
968 match body.read_line(&mut line).await {
969 Ok(0) => Ok(None),
970 Ok(_) => {
971 let event = if includes_status_messages {
972 serde_json::from_str::<CompletionEvent<T>>(&line)?
973 } else {
974 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
975 };
976
977 line.clear();
978 Ok(Some((event, (line, body))))
979 }
980 Err(e) => Err(e.into()),
981 }
982 },
983 )
984}
985
986#[derive(IntoElement, RegisterComponent)]
987struct ZedAiConfiguration {
988 is_connected: bool,
989 plan: Option<Plan>,
990 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
991 eligible_for_trial: bool,
992 account_too_young: bool,
993 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
994}
995
996impl RenderOnce for ZedAiConfiguration {
997 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
998 let is_pro = self.plan.is_some_and(|plan| plan == Plan::ZedPro);
999 let subscription_text = match (self.plan, self.subscription_period) {
1000 (Some(Plan::ZedPro), Some(_)) => {
1001 "You have access to Zed's hosted models through your Pro subscription."
1002 }
1003 (Some(Plan::ZedProTrial), Some(_)) => {
1004 "You have access to Zed's hosted models through your Pro trial."
1005 }
1006 (Some(Plan::ZedFree), Some(_)) => {
1007 if self.eligible_for_trial {
1008 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1009 } else {
1010 "Subscribe for access to Zed's hosted models."
1011 }
1012 }
1013 _ => {
1014 if self.eligible_for_trial {
1015 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1016 } else {
1017 "Subscribe for access to Zed's hosted models."
1018 }
1019 }
1020 };
1021
1022 let manage_subscription_buttons = if is_pro {
1023 Button::new("manage_settings", "Manage Subscription")
1024 .full_width()
1025 .style(ButtonStyle::Tinted(TintColor::Accent))
1026 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1027 .into_any_element()
1028 } else if self.plan.is_none() || self.eligible_for_trial {
1029 Button::new("start_trial", "Start 14-day Free Pro Trial")
1030 .full_width()
1031 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1032 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1033 .into_any_element()
1034 } else {
1035 Button::new("upgrade", "Upgrade to Pro")
1036 .full_width()
1037 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1038 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1039 .into_any_element()
1040 };
1041
1042 if !self.is_connected {
1043 return v_flex()
1044 .gap_2()
1045 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1046 .child(
1047 Button::new("sign_in", "Sign In to use Zed AI")
1048 .icon_color(Color::Muted)
1049 .icon(IconName::Github)
1050 .icon_size(IconSize::Small)
1051 .icon_position(IconPosition::Start)
1052 .full_width()
1053 .on_click({
1054 let callback = self.sign_in_callback.clone();
1055 move |_, window, cx| (callback)(window, cx)
1056 }),
1057 );
1058 }
1059
1060 v_flex().gap_2().w_full().map(|this| {
1061 if self.account_too_young {
1062 this.child(YoungAccountBanner).child(
1063 Button::new("upgrade", "Upgrade to Pro")
1064 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1065 .full_width()
1066 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1067 )
1068 } else {
1069 this.text_sm()
1070 .child(subscription_text)
1071 .child(manage_subscription_buttons)
1072 }
1073 })
1074 }
1075}
1076
1077struct ConfigurationView {
1078 state: Entity<State>,
1079 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1080}
1081
1082impl ConfigurationView {
1083 fn new(state: Entity<State>) -> Self {
1084 let sign_in_callback = Arc::new({
1085 let state = state.clone();
1086 move |_window: &mut Window, cx: &mut App| {
1087 state.update(cx, |state, cx| {
1088 state.authenticate(cx).detach_and_log_err(cx);
1089 });
1090 }
1091 });
1092
1093 Self {
1094 state,
1095 sign_in_callback,
1096 }
1097 }
1098}
1099
1100impl Render for ConfigurationView {
1101 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1102 let state = self.state.read(cx);
1103 let user_store = state.user_store.read(cx);
1104
1105 ZedAiConfiguration {
1106 is_connected: !state.is_signed_out(cx),
1107 plan: user_store.plan(),
1108 subscription_period: user_store.subscription_period(),
1109 eligible_for_trial: user_store.trial_started_at().is_none(),
1110 account_too_young: user_store.account_too_young(),
1111 sign_in_callback: self.sign_in_callback.clone(),
1112 }
1113 }
1114}
1115
1116impl Component for ZedAiConfiguration {
1117 fn name() -> &'static str {
1118 "AI Configuration Content"
1119 }
1120
1121 fn sort_name() -> &'static str {
1122 "AI Configuration Content"
1123 }
1124
1125 fn scope() -> ComponentScope {
1126 ComponentScope::Onboarding
1127 }
1128
1129 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1130 fn configuration(
1131 is_connected: bool,
1132 plan: Option<Plan>,
1133 eligible_for_trial: bool,
1134 account_too_young: bool,
1135 ) -> AnyElement {
1136 ZedAiConfiguration {
1137 is_connected,
1138 plan,
1139 subscription_period: plan
1140 .is_some()
1141 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1142 eligible_for_trial,
1143 account_too_young,
1144 sign_in_callback: Arc::new(|_, _| {}),
1145 }
1146 .into_any_element()
1147 }
1148
1149 Some(
1150 v_flex()
1151 .p_4()
1152 .gap_4()
1153 .children(vec![
1154 single_example("Not connected", configuration(false, None, false, false)),
1155 single_example(
1156 "Accept Terms of Service",
1157 configuration(true, None, true, false),
1158 ),
1159 single_example(
1160 "No Plan - Not eligible for trial",
1161 configuration(true, None, false, false),
1162 ),
1163 single_example(
1164 "No Plan - Eligible for trial",
1165 configuration(true, None, true, false),
1166 ),
1167 single_example(
1168 "Free Plan",
1169 configuration(true, Some(Plan::ZedFree), true, false),
1170 ),
1171 single_example(
1172 "Zed Pro Trial Plan",
1173 configuration(true, Some(Plan::ZedProTrial), true, false),
1174 ),
1175 single_example(
1176 "Zed Pro Plan",
1177 configuration(true, Some(Plan::ZedPro), true, false),
1178 ),
1179 ])
1180 .into_any_element(),
1181 )
1182 }
1183}
1184
1185#[cfg(test)]
1186mod tests {
1187 use super::*;
1188 use http_client::http::{HeaderMap, StatusCode};
1189 use language_model::LanguageModelCompletionError;
1190
1191 #[test]
1192 fn test_api_error_conversion_with_upstream_http_error() {
1193 // upstream_http_error with 503 status should become ServerOverloaded
1194 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}"#;
1195
1196 let api_error = ApiError {
1197 status: StatusCode::INTERNAL_SERVER_ERROR,
1198 body: error_body.to_string(),
1199 headers: HeaderMap::new(),
1200 };
1201
1202 let completion_error: LanguageModelCompletionError = api_error.into();
1203
1204 match completion_error {
1205 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1206 assert_eq!(
1207 message,
1208 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1209 );
1210 }
1211 _ => panic!(
1212 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1213 completion_error
1214 ),
1215 }
1216
1217 // upstream_http_error with 500 status should become ApiInternalServerError
1218 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1219
1220 let api_error = ApiError {
1221 status: StatusCode::INTERNAL_SERVER_ERROR,
1222 body: error_body.to_string(),
1223 headers: HeaderMap::new(),
1224 };
1225
1226 let completion_error: LanguageModelCompletionError = api_error.into();
1227
1228 match completion_error {
1229 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1230 assert_eq!(
1231 message,
1232 "Received an error from the OpenAI API: internal server error"
1233 );
1234 }
1235 _ => panic!(
1236 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1237 completion_error
1238 ),
1239 }
1240
1241 // upstream_http_error with 429 status should become RateLimitExceeded
1242 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1243
1244 let api_error = ApiError {
1245 status: StatusCode::INTERNAL_SERVER_ERROR,
1246 body: error_body.to_string(),
1247 headers: HeaderMap::new(),
1248 };
1249
1250 let completion_error: LanguageModelCompletionError = api_error.into();
1251
1252 match completion_error {
1253 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1254 assert_eq!(
1255 message,
1256 "Received an error from the Google API: rate limit exceeded"
1257 );
1258 }
1259 _ => panic!(
1260 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1261 completion_error
1262 ),
1263 }
1264
1265 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1266 let error_body = "Regular internal server error";
1267
1268 let api_error = ApiError {
1269 status: StatusCode::INTERNAL_SERVER_ERROR,
1270 body: error_body.to_string(),
1271 headers: HeaderMap::new(),
1272 };
1273
1274 let completion_error: LanguageModelCompletionError = api_error.into();
1275
1276 match completion_error {
1277 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1278 assert_eq!(provider, PROVIDER_NAME);
1279 assert_eq!(message, "Regular internal server error");
1280 }
1281 _ => panic!(
1282 "Expected ApiInternalServerError for regular 500, got: {:?}",
1283 completion_error
1284 ),
1285 }
1286
1287 // upstream_http_429 format should be converted to UpstreamProviderError
1288 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1289
1290 let api_error = ApiError {
1291 status: StatusCode::INTERNAL_SERVER_ERROR,
1292 body: error_body.to_string(),
1293 headers: HeaderMap::new(),
1294 };
1295
1296 let completion_error: LanguageModelCompletionError = api_error.into();
1297
1298 match completion_error {
1299 LanguageModelCompletionError::UpstreamProviderError {
1300 message,
1301 status,
1302 retry_after,
1303 } => {
1304 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1305 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1306 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1307 }
1308 _ => panic!(
1309 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1310 completion_error
1311 ),
1312 }
1313
1314 // Invalid JSON in error body should fall back to regular error handling
1315 let error_body = "Not JSON at all";
1316
1317 let api_error = ApiError {
1318 status: StatusCode::INTERNAL_SERVER_ERROR,
1319 body: error_body.to_string(),
1320 headers: HeaderMap::new(),
1321 };
1322
1323 let completion_error: LanguageModelCompletionError = api_error.into();
1324
1325 match completion_error {
1326 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1327 assert_eq!(provider, PROVIDER_NAME);
1328 }
1329 _ => panic!(
1330 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1331 completion_error
1332 ),
1333 }
1334 }
1335}