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::str::FromStr;
38use std::sync::Arc;
39use std::time::Duration;
40use thiserror::Error;
41use ui::{TintColor, prelude::*};
42use util::{ResultExt as _, maybe};
43
44use crate::provider::anthropic::{
45 AnthropicEventMapper, count_anthropic_tokens_with_tiktoken, into_anthropic,
46};
47use crate::provider::google::{GoogleEventMapper, into_google};
48use crate::provider::open_ai::{
49 OpenAiEventMapper, OpenAiResponseEventMapper, count_open_ai_tokens, into_open_ai,
50 into_open_ai_response,
51};
52use crate::provider::x_ai::count_xai_tokens;
53
54const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
55const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
56
57#[derive(Default, Clone, Debug, PartialEq)]
58pub struct ZedDotDevSettings {
59 pub available_models: Vec<AvailableModel>,
60}
61#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
62#[serde(tag = "type", rename_all = "lowercase")]
63pub enum ModelMode {
64 #[default]
65 Default,
66 Thinking {
67 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
68 budget_tokens: Option<u32>,
69 },
70}
71
72impl From<ModelMode> for AnthropicModelMode {
73 fn from(value: ModelMode) -> Self {
74 match value {
75 ModelMode::Default => AnthropicModelMode::Default,
76 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
77 }
78 }
79}
80
81pub struct CloudLanguageModelProvider {
82 client: Arc<Client>,
83 state: Entity<State>,
84 _maintain_client_status: Task<()>,
85}
86
87pub struct State {
88 client: Arc<Client>,
89 llm_api_token: LlmApiToken,
90 user_store: Entity<UserStore>,
91 status: client::Status,
92 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
93 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
94 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
95 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
96 _fetch_models_task: Task<()>,
97 _settings_subscription: Subscription,
98 _llm_token_subscription: Subscription,
99}
100
101impl State {
102 fn new(
103 client: Arc<Client>,
104 user_store: Entity<UserStore>,
105 status: client::Status,
106 cx: &mut Context<Self>,
107 ) -> Self {
108 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
109 let mut current_user = user_store.read(cx).watch_current_user();
110 Self {
111 client: client.clone(),
112 llm_api_token: LlmApiToken::default(),
113 user_store,
114 status,
115 models: Vec::new(),
116 default_model: None,
117 default_fast_model: None,
118 recommended_models: Vec::new(),
119 _fetch_models_task: cx.spawn(async move |this, cx| {
120 maybe!(async move {
121 let (client, llm_api_token) = this
122 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
123
124 while current_user.borrow().is_none() {
125 current_user.next().await;
126 }
127
128 let response =
129 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
130 this.update(cx, |this, cx| this.update_models(response, cx))?;
131 anyhow::Ok(())
132 })
133 .await
134 .context("failed to fetch Zed models")
135 .log_err();
136 }),
137 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
138 cx.notify();
139 }),
140 _llm_token_subscription: cx.subscribe(
141 &refresh_llm_token_listener,
142 move |this, _listener, _event, cx| {
143 let client = this.client.clone();
144 let llm_api_token = this.llm_api_token.clone();
145 cx.spawn(async move |this, cx| {
146 llm_api_token.refresh(&client).await?;
147 let response = Self::fetch_models(client, llm_api_token).await?;
148 this.update(cx, |this, cx| {
149 this.update_models(response, cx);
150 })
151 })
152 .detach_and_log_err(cx);
153 },
154 ),
155 }
156 }
157
158 fn is_signed_out(&self, cx: &App) -> bool {
159 self.user_store.read(cx).current_user().is_none()
160 }
161
162 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
163 let client = self.client.clone();
164 cx.spawn(async move |state, cx| {
165 client.sign_in_with_optional_connect(true, cx).await?;
166 state.update(cx, |_, cx| cx.notify())
167 })
168 }
169
170 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
171 let is_thinking_toggle_enabled = cx.has_flag::<CloudThinkingToggleFeatureFlag>();
172
173 let mut models = Vec::new();
174
175 for model in response.models {
176 models.push(Arc::new(model.clone()));
177
178 if !is_thinking_toggle_enabled {
179 // Right now we represent thinking variants of models as separate models on the client,
180 // so we need to insert variants for any model that supports thinking.
181 if model.supports_thinking {
182 models.push(Arc::new(cloud_llm_client::LanguageModel {
183 id: cloud_llm_client::LanguageModelId(
184 format!("{}-thinking", model.id).into(),
185 ),
186 display_name: format!("{} Thinking", model.display_name),
187 ..model
188 }));
189 }
190 }
191 }
192
193 self.default_model = models
194 .iter()
195 .find(|model| {
196 response
197 .default_model
198 .as_ref()
199 .is_some_and(|default_model_id| &model.id == default_model_id)
200 })
201 .cloned();
202 self.default_fast_model = models
203 .iter()
204 .find(|model| {
205 response
206 .default_fast_model
207 .as_ref()
208 .is_some_and(|default_fast_model_id| &model.id == default_fast_model_id)
209 })
210 .cloned();
211 self.recommended_models = response
212 .recommended_models
213 .iter()
214 .filter_map(|id| models.iter().find(|model| &model.id == id))
215 .cloned()
216 .collect();
217 self.models = models;
218 cx.notify();
219 }
220
221 async fn fetch_models(
222 client: Arc<Client>,
223 llm_api_token: LlmApiToken,
224 ) -> Result<ListModelsResponse> {
225 let http_client = &client.http_client();
226 let token = llm_api_token.acquire(&client).await?;
227
228 let request = http_client::Request::builder()
229 .method(Method::GET)
230 .header(CLIENT_SUPPORTS_X_AI_HEADER_NAME, "true")
231 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
232 .header("Authorization", format!("Bearer {token}"))
233 .body(AsyncBody::empty())?;
234 let mut response = http_client
235 .send(request)
236 .await
237 .context("failed to send list models request")?;
238
239 if response.status().is_success() {
240 let mut body = String::new();
241 response.body_mut().read_to_string(&mut body).await?;
242 Ok(serde_json::from_str(&body)?)
243 } else {
244 let mut body = String::new();
245 response.body_mut().read_to_string(&mut body).await?;
246 anyhow::bail!(
247 "error listing models.\nStatus: {:?}\nBody: {body}",
248 response.status(),
249 );
250 }
251 }
252}
253
254impl CloudLanguageModelProvider {
255 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
256 let mut status_rx = client.status();
257 let status = *status_rx.borrow();
258
259 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
260
261 let state_ref = state.downgrade();
262 let maintain_client_status = cx.spawn(async move |cx| {
263 while let Some(status) = status_rx.next().await {
264 if let Some(this) = state_ref.upgrade() {
265 _ = this.update(cx, |this, cx| {
266 if this.status != status {
267 this.status = status;
268 cx.notify();
269 }
270 });
271 } else {
272 break;
273 }
274 }
275 });
276
277 Self {
278 client,
279 state,
280 _maintain_client_status: maintain_client_status,
281 }
282 }
283
284 fn create_language_model(
285 &self,
286 model: Arc<cloud_llm_client::LanguageModel>,
287 llm_api_token: LlmApiToken,
288 ) -> Arc<dyn LanguageModel> {
289 Arc::new(CloudLanguageModel {
290 id: LanguageModelId(SharedString::from(model.id.0.clone())),
291 model,
292 llm_api_token,
293 client: self.client.clone(),
294 request_limiter: RateLimiter::new(4),
295 })
296 }
297}
298
299impl LanguageModelProviderState for CloudLanguageModelProvider {
300 type ObservableEntity = State;
301
302 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
303 Some(self.state.clone())
304 }
305}
306
307impl LanguageModelProvider for CloudLanguageModelProvider {
308 fn id(&self) -> LanguageModelProviderId {
309 PROVIDER_ID
310 }
311
312 fn name(&self) -> LanguageModelProviderName {
313 PROVIDER_NAME
314 }
315
316 fn icon(&self) -> IconOrSvg {
317 IconOrSvg::Icon(IconName::AiZed)
318 }
319
320 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
321 let default_model = self.state.read(cx).default_model.clone()?;
322 let llm_api_token = self.state.read(cx).llm_api_token.clone();
323 Some(self.create_language_model(default_model, llm_api_token))
324 }
325
326 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
327 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
328 let llm_api_token = self.state.read(cx).llm_api_token.clone();
329 Some(self.create_language_model(default_fast_model, llm_api_token))
330 }
331
332 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
333 let llm_api_token = self.state.read(cx).llm_api_token.clone();
334 self.state
335 .read(cx)
336 .recommended_models
337 .iter()
338 .cloned()
339 .map(|model| self.create_language_model(model, llm_api_token.clone()))
340 .collect()
341 }
342
343 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
344 let llm_api_token = self.state.read(cx).llm_api_token.clone();
345 self.state
346 .read(cx)
347 .models
348 .iter()
349 .cloned()
350 .map(|model| self.create_language_model(model, llm_api_token.clone()))
351 .collect()
352 }
353
354 fn is_authenticated(&self, cx: &App) -> bool {
355 let state = self.state.read(cx);
356 !state.is_signed_out(cx)
357 }
358
359 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
360 Task::ready(Ok(()))
361 }
362
363 fn configuration_view(
364 &self,
365 _target_agent: language_model::ConfigurationViewTargetAgent,
366 _: &mut Window,
367 cx: &mut App,
368 ) -> AnyView {
369 cx.new(|_| ConfigurationView::new(self.state.clone()))
370 .into()
371 }
372
373 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
374 Task::ready(Ok(()))
375 }
376}
377
378pub struct CloudLanguageModel {
379 id: LanguageModelId,
380 model: Arc<cloud_llm_client::LanguageModel>,
381 llm_api_token: LlmApiToken,
382 client: Arc<Client>,
383 request_limiter: RateLimiter,
384}
385
386struct PerformLlmCompletionResponse {
387 response: Response<AsyncBody>,
388 includes_status_messages: bool,
389}
390
391impl CloudLanguageModel {
392 async fn perform_llm_completion(
393 client: Arc<Client>,
394 llm_api_token: LlmApiToken,
395 app_version: Option<Version>,
396 body: CompletionBody,
397 ) -> Result<PerformLlmCompletionResponse> {
398 let http_client = &client.http_client();
399
400 let mut token = llm_api_token.acquire(&client).await?;
401 let mut refreshed_token = false;
402
403 loop {
404 let request = http_client::Request::builder()
405 .method(Method::POST)
406 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
407 .when_some(app_version.as_ref(), |builder, app_version| {
408 builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
409 })
410 .header("Content-Type", "application/json")
411 .header("Authorization", format!("Bearer {token}"))
412 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
413 .body(serde_json::to_string(&body)?.into())?;
414
415 let mut response = http_client.send(request).await?;
416 let status = response.status();
417 if status.is_success() {
418 let includes_status_messages = response
419 .headers()
420 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
421 .is_some();
422
423 return Ok(PerformLlmCompletionResponse {
424 response,
425 includes_status_messages,
426 });
427 }
428
429 if !refreshed_token && response.needs_llm_token_refresh() {
430 token = llm_api_token.refresh(&client).await?;
431 refreshed_token = true;
432 continue;
433 }
434
435 if status == StatusCode::PAYMENT_REQUIRED {
436 return Err(anyhow!(PaymentRequiredError));
437 }
438
439 let mut body = String::new();
440 let headers = response.headers().clone();
441 response.body_mut().read_to_string(&mut body).await?;
442 return Err(anyhow!(ApiError {
443 status,
444 body,
445 headers
446 }));
447 }
448 }
449}
450
451#[derive(Debug, Error)]
452#[error("cloud language model request failed with status {status}: {body}")]
453struct ApiError {
454 status: StatusCode,
455 body: String,
456 headers: HeaderMap<HeaderValue>,
457}
458
459/// Represents error responses from Zed's cloud API.
460///
461/// Example JSON for an upstream HTTP error:
462/// ```json
463/// {
464/// "code": "upstream_http_error",
465/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
466/// "upstream_status": 503
467/// }
468/// ```
469#[derive(Debug, serde::Deserialize)]
470struct CloudApiError {
471 code: String,
472 message: String,
473 #[serde(default)]
474 #[serde(deserialize_with = "deserialize_optional_status_code")]
475 upstream_status: Option<StatusCode>,
476 #[serde(default)]
477 retry_after: Option<f64>,
478}
479
480fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
481where
482 D: serde::Deserializer<'de>,
483{
484 let opt: Option<u16> = Option::deserialize(deserializer)?;
485 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
486}
487
488impl From<ApiError> for LanguageModelCompletionError {
489 fn from(error: ApiError) -> Self {
490 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
491 if cloud_error.code.starts_with("upstream_http_") {
492 let status = if let Some(status) = cloud_error.upstream_status {
493 status
494 } else if cloud_error.code.ends_with("_error") {
495 error.status
496 } else {
497 // If there's a status code in the code string (e.g. "upstream_http_429")
498 // then use that; otherwise, see if the JSON contains a status code.
499 cloud_error
500 .code
501 .strip_prefix("upstream_http_")
502 .and_then(|code_str| code_str.parse::<u16>().ok())
503 .and_then(|code| StatusCode::from_u16(code).ok())
504 .unwrap_or(error.status)
505 };
506
507 return LanguageModelCompletionError::UpstreamProviderError {
508 message: cloud_error.message,
509 status,
510 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
511 };
512 }
513
514 return LanguageModelCompletionError::from_http_status(
515 PROVIDER_NAME,
516 error.status,
517 cloud_error.message,
518 None,
519 );
520 }
521
522 let retry_after = None;
523 LanguageModelCompletionError::from_http_status(
524 PROVIDER_NAME,
525 error.status,
526 error.body,
527 retry_after,
528 )
529 }
530}
531
532impl LanguageModel for CloudLanguageModel {
533 fn id(&self) -> LanguageModelId {
534 self.id.clone()
535 }
536
537 fn name(&self) -> LanguageModelName {
538 LanguageModelName::from(self.model.display_name.clone())
539 }
540
541 fn provider_id(&self) -> LanguageModelProviderId {
542 PROVIDER_ID
543 }
544
545 fn provider_name(&self) -> LanguageModelProviderName {
546 PROVIDER_NAME
547 }
548
549 fn upstream_provider_id(&self) -> LanguageModelProviderId {
550 use cloud_llm_client::LanguageModelProvider::*;
551 match self.model.provider {
552 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
553 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
554 Google => language_model::GOOGLE_PROVIDER_ID,
555 XAi => language_model::X_AI_PROVIDER_ID,
556 }
557 }
558
559 fn upstream_provider_name(&self) -> LanguageModelProviderName {
560 use cloud_llm_client::LanguageModelProvider::*;
561 match self.model.provider {
562 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
563 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
564 Google => language_model::GOOGLE_PROVIDER_NAME,
565 XAi => language_model::X_AI_PROVIDER_NAME,
566 }
567 }
568
569 fn supports_tools(&self) -> bool {
570 self.model.supports_tools
571 }
572
573 fn supports_images(&self) -> bool {
574 self.model.supports_images
575 }
576
577 fn supports_thinking(&self) -> bool {
578 self.model.supports_thinking
579 }
580
581 fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {
582 self.model
583 .supported_effort_levels
584 .iter()
585 .map(|effort_level| LanguageModelEffortLevel {
586 name: effort_level.name.clone().into(),
587 value: effort_level.value.clone().into(),
588 is_default: effort_level.is_default.unwrap_or(false),
589 })
590 .collect()
591 }
592
593 fn supports_streaming_tools(&self) -> bool {
594 self.model.supports_streaming_tools
595 }
596
597 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
598 match choice {
599 LanguageModelToolChoice::Auto
600 | LanguageModelToolChoice::Any
601 | LanguageModelToolChoice::None => true,
602 }
603 }
604
605 fn supports_split_token_display(&self) -> bool {
606 use cloud_llm_client::LanguageModelProvider::*;
607 matches!(self.model.provider, OpenAi)
608 }
609
610 fn telemetry_id(&self) -> String {
611 format!("zed.dev/{}", self.model.id)
612 }
613
614 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
615 match self.model.provider {
616 cloud_llm_client::LanguageModelProvider::Anthropic
617 | cloud_llm_client::LanguageModelProvider::OpenAi
618 | cloud_llm_client::LanguageModelProvider::XAi => {
619 LanguageModelToolSchemaFormat::JsonSchema
620 }
621 cloud_llm_client::LanguageModelProvider::Google => {
622 LanguageModelToolSchemaFormat::JsonSchemaSubset
623 }
624 }
625 }
626
627 fn max_token_count(&self) -> u64 {
628 self.model.max_token_count as u64
629 }
630
631 fn max_output_tokens(&self) -> Option<u64> {
632 Some(self.model.max_output_tokens as u64)
633 }
634
635 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
636 match &self.model.provider {
637 cloud_llm_client::LanguageModelProvider::Anthropic => {
638 Some(LanguageModelCacheConfiguration {
639 min_total_token: 2_048,
640 should_speculate: true,
641 max_cache_anchors: 4,
642 })
643 }
644 cloud_llm_client::LanguageModelProvider::OpenAi
645 | cloud_llm_client::LanguageModelProvider::XAi
646 | cloud_llm_client::LanguageModelProvider::Google => None,
647 }
648 }
649
650 fn count_tokens(
651 &self,
652 request: LanguageModelRequest,
653 cx: &App,
654 ) -> BoxFuture<'static, Result<u64>> {
655 match self.model.provider {
656 cloud_llm_client::LanguageModelProvider::Anthropic => cx
657 .background_spawn(async move { count_anthropic_tokens_with_tiktoken(request) })
658 .boxed(),
659 cloud_llm_client::LanguageModelProvider::OpenAi => {
660 let model = match open_ai::Model::from_id(&self.model.id.0) {
661 Ok(model) => model,
662 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
663 };
664 count_open_ai_tokens(request, model, cx)
665 }
666 cloud_llm_client::LanguageModelProvider::XAi => {
667 let model = match x_ai::Model::from_id(&self.model.id.0) {
668 Ok(model) => model,
669 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
670 };
671 count_xai_tokens(request, model, cx)
672 }
673 cloud_llm_client::LanguageModelProvider::Google => {
674 let client = self.client.clone();
675 let llm_api_token = self.llm_api_token.clone();
676 let model_id = self.model.id.to_string();
677 let generate_content_request =
678 into_google(request, model_id.clone(), GoogleModelMode::Default);
679 async move {
680 let http_client = &client.http_client();
681 let token = llm_api_token.acquire(&client).await?;
682
683 let request_body = CountTokensBody {
684 provider: cloud_llm_client::LanguageModelProvider::Google,
685 model: model_id,
686 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
687 generate_content_request,
688 })?,
689 };
690 let request = http_client::Request::builder()
691 .method(Method::POST)
692 .uri(
693 http_client
694 .build_zed_llm_url("/count_tokens", &[])?
695 .as_ref(),
696 )
697 .header("Content-Type", "application/json")
698 .header("Authorization", format!("Bearer {token}"))
699 .body(serde_json::to_string(&request_body)?.into())?;
700 let mut response = http_client.send(request).await?;
701 let status = response.status();
702 let headers = response.headers().clone();
703 let mut response_body = String::new();
704 response
705 .body_mut()
706 .read_to_string(&mut response_body)
707 .await?;
708
709 if status.is_success() {
710 let response_body: CountTokensResponse =
711 serde_json::from_str(&response_body)?;
712
713 Ok(response_body.tokens as u64)
714 } else {
715 Err(anyhow!(ApiError {
716 status,
717 body: response_body,
718 headers
719 }))
720 }
721 }
722 .boxed()
723 }
724 }
725 }
726
727 fn stream_completion(
728 &self,
729 request: LanguageModelRequest,
730 cx: &AsyncApp,
731 ) -> BoxFuture<
732 'static,
733 Result<
734 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
735 LanguageModelCompletionError,
736 >,
737 > {
738 let thread_id = request.thread_id.clone();
739 let prompt_id = request.prompt_id.clone();
740 let intent = request.intent;
741 let app_version = Some(cx.update(|cx| AppVersion::global(cx)));
742 let thinking_allowed = request.thinking_allowed;
743 let is_thinking_toggle_enabled =
744 cx.update(|cx| cx.has_flag::<CloudThinkingToggleFeatureFlag>());
745 let enable_thinking = if is_thinking_toggle_enabled {
746 thinking_allowed && self.model.supports_thinking
747 } else {
748 thinking_allowed && self.model.id.0.ends_with("-thinking")
749 };
750 let effort = request
751 .thinking_effort
752 .as_ref()
753 .and_then(|effort| anthropic::Effort::from_str(effort).ok());
754 let provider_name = provider_name(&self.model.provider);
755 match self.model.provider {
756 cloud_llm_client::LanguageModelProvider::Anthropic => {
757 let mut request = into_anthropic(
758 request,
759 self.model.id.to_string(),
760 1.0,
761 self.model.max_output_tokens as u64,
762 if enable_thinking {
763 AnthropicModelMode::Thinking {
764 budget_tokens: Some(4_096),
765 }
766 } else {
767 AnthropicModelMode::Default
768 },
769 );
770
771 if enable_thinking && effort.is_some() {
772 request.thinking = Some(anthropic::Thinking::Adaptive);
773 request.output_config = Some(anthropic::OutputConfig { effort });
774 }
775
776 let client = self.client.clone();
777 let llm_api_token = self.llm_api_token.clone();
778 let future = self.request_limiter.stream(async move {
779 let PerformLlmCompletionResponse {
780 response,
781 includes_status_messages,
782 } = Self::perform_llm_completion(
783 client.clone(),
784 llm_api_token,
785 app_version,
786 CompletionBody {
787 thread_id,
788 prompt_id,
789 intent,
790 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
791 model: request.model.clone(),
792 provider_request: serde_json::to_value(&request)
793 .map_err(|e| anyhow!(e))?,
794 },
795 )
796 .await
797 .map_err(|err| match err.downcast::<ApiError>() {
798 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
799 Err(err) => anyhow!(err),
800 })?;
801
802 let mut mapper = AnthropicEventMapper::new();
803 Ok(map_cloud_completion_events(
804 Box::pin(response_lines(response, includes_status_messages)),
805 &provider_name,
806 move |event| mapper.map_event(event),
807 ))
808 });
809 async move { Ok(future.await?.boxed()) }.boxed()
810 }
811 cloud_llm_client::LanguageModelProvider::OpenAi => {
812 let client = self.client.clone();
813 let llm_api_token = self.llm_api_token.clone();
814
815 let request = into_open_ai_response(
816 request,
817 &self.model.id.0,
818 self.model.supports_parallel_tool_calls,
819 true,
820 None,
821 None,
822 );
823 let future = self.request_limiter.stream(async move {
824 let PerformLlmCompletionResponse {
825 response,
826 includes_status_messages,
827 } = Self::perform_llm_completion(
828 client.clone(),
829 llm_api_token,
830 app_version,
831 CompletionBody {
832 thread_id,
833 prompt_id,
834 intent,
835 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
836 model: request.model.clone(),
837 provider_request: serde_json::to_value(&request)
838 .map_err(|e| anyhow!(e))?,
839 },
840 )
841 .await?;
842
843 let mut mapper = OpenAiResponseEventMapper::new();
844 Ok(map_cloud_completion_events(
845 Box::pin(response_lines(response, includes_status_messages)),
846 &provider_name,
847 move |event| mapper.map_event(event),
848 ))
849 });
850 async move { Ok(future.await?.boxed()) }.boxed()
851 }
852 cloud_llm_client::LanguageModelProvider::XAi => {
853 let client = self.client.clone();
854 let request = into_open_ai(
855 request,
856 &self.model.id.0,
857 self.model.supports_parallel_tool_calls,
858 false,
859 None,
860 None,
861 );
862 let llm_api_token = self.llm_api_token.clone();
863 let future = self.request_limiter.stream(async move {
864 let PerformLlmCompletionResponse {
865 response,
866 includes_status_messages,
867 } = Self::perform_llm_completion(
868 client.clone(),
869 llm_api_token,
870 app_version,
871 CompletionBody {
872 thread_id,
873 prompt_id,
874 intent,
875 provider: cloud_llm_client::LanguageModelProvider::XAi,
876 model: request.model.clone(),
877 provider_request: serde_json::to_value(&request)
878 .map_err(|e| anyhow!(e))?,
879 },
880 )
881 .await?;
882
883 let mut mapper = OpenAiEventMapper::new();
884 Ok(map_cloud_completion_events(
885 Box::pin(response_lines(response, includes_status_messages)),
886 &provider_name,
887 move |event| mapper.map_event(event),
888 ))
889 });
890 async move { Ok(future.await?.boxed()) }.boxed()
891 }
892 cloud_llm_client::LanguageModelProvider::Google => {
893 let client = self.client.clone();
894 let request =
895 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
896 let llm_api_token = self.llm_api_token.clone();
897 let future = self.request_limiter.stream(async move {
898 let PerformLlmCompletionResponse {
899 response,
900 includes_status_messages,
901 } = Self::perform_llm_completion(
902 client.clone(),
903 llm_api_token,
904 app_version,
905 CompletionBody {
906 thread_id,
907 prompt_id,
908 intent,
909 provider: cloud_llm_client::LanguageModelProvider::Google,
910 model: request.model.model_id.clone(),
911 provider_request: serde_json::to_value(&request)
912 .map_err(|e| anyhow!(e))?,
913 },
914 )
915 .await?;
916
917 let mut mapper = GoogleEventMapper::new();
918 Ok(map_cloud_completion_events(
919 Box::pin(response_lines(response, includes_status_messages)),
920 &provider_name,
921 move |event| mapper.map_event(event),
922 ))
923 });
924 async move { Ok(future.await?.boxed()) }.boxed()
925 }
926 }
927 }
928}
929
930fn map_cloud_completion_events<T, F>(
931 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
932 provider: &LanguageModelProviderName,
933 mut map_callback: F,
934) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
935where
936 T: DeserializeOwned + 'static,
937 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
938 + Send
939 + 'static,
940{
941 let provider = provider.clone();
942 stream
943 .flat_map(move |event| {
944 futures::stream::iter(match event {
945 Err(error) => {
946 vec![Err(LanguageModelCompletionError::from(error))]
947 }
948 Ok(CompletionEvent::Status(event)) => {
949 vec![
950 LanguageModelCompletionEvent::from_completion_request_status(
951 event,
952 provider.clone(),
953 ),
954 ]
955 }
956 Ok(CompletionEvent::Event(event)) => map_callback(event),
957 })
958 })
959 .boxed()
960}
961
962fn provider_name(provider: &cloud_llm_client::LanguageModelProvider) -> LanguageModelProviderName {
963 match provider {
964 cloud_llm_client::LanguageModelProvider::Anthropic => {
965 language_model::ANTHROPIC_PROVIDER_NAME
966 }
967 cloud_llm_client::LanguageModelProvider::OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
968 cloud_llm_client::LanguageModelProvider::Google => language_model::GOOGLE_PROVIDER_NAME,
969 cloud_llm_client::LanguageModelProvider::XAi => language_model::X_AI_PROVIDER_NAME,
970 }
971}
972
973fn response_lines<T: DeserializeOwned>(
974 response: Response<AsyncBody>,
975 includes_status_messages: bool,
976) -> impl Stream<Item = Result<CompletionEvent<T>>> {
977 futures::stream::try_unfold(
978 (String::new(), BufReader::new(response.into_body())),
979 move |(mut line, mut body)| async move {
980 match body.read_line(&mut line).await {
981 Ok(0) => Ok(None),
982 Ok(_) => {
983 let event = if includes_status_messages {
984 serde_json::from_str::<CompletionEvent<T>>(&line)?
985 } else {
986 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
987 };
988
989 line.clear();
990 Ok(Some((event, (line, body))))
991 }
992 Err(e) => Err(e.into()),
993 }
994 },
995 )
996}
997
998#[derive(IntoElement, RegisterComponent)]
999struct ZedAiConfiguration {
1000 is_connected: bool,
1001 plan: Option<Plan>,
1002 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1003 eligible_for_trial: bool,
1004 account_too_young: bool,
1005 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1006}
1007
1008impl RenderOnce for ZedAiConfiguration {
1009 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1010 let is_pro = self.plan.is_some_and(|plan| plan == Plan::ZedPro);
1011 let subscription_text = match (self.plan, self.subscription_period) {
1012 (Some(Plan::ZedPro), Some(_)) => {
1013 "You have access to Zed's hosted models through your Pro subscription."
1014 }
1015 (Some(Plan::ZedProTrial), Some(_)) => {
1016 "You have access to Zed's hosted models through your Pro trial."
1017 }
1018 (Some(Plan::ZedFree), Some(_)) => {
1019 if self.eligible_for_trial {
1020 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1021 } else {
1022 "Subscribe for access to Zed's hosted models."
1023 }
1024 }
1025 _ => {
1026 if self.eligible_for_trial {
1027 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1028 } else {
1029 "Subscribe for access to Zed's hosted models."
1030 }
1031 }
1032 };
1033
1034 let manage_subscription_buttons = if is_pro {
1035 Button::new("manage_settings", "Manage Subscription")
1036 .full_width()
1037 .style(ButtonStyle::Tinted(TintColor::Accent))
1038 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1039 .into_any_element()
1040 } else if self.plan.is_none() || self.eligible_for_trial {
1041 Button::new("start_trial", "Start 14-day Free Pro Trial")
1042 .full_width()
1043 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1044 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1045 .into_any_element()
1046 } else {
1047 Button::new("upgrade", "Upgrade to Pro")
1048 .full_width()
1049 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1050 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1051 .into_any_element()
1052 };
1053
1054 if !self.is_connected {
1055 return v_flex()
1056 .gap_2()
1057 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1058 .child(
1059 Button::new("sign_in", "Sign In to use Zed AI")
1060 .icon_color(Color::Muted)
1061 .icon(IconName::Github)
1062 .icon_size(IconSize::Small)
1063 .icon_position(IconPosition::Start)
1064 .full_width()
1065 .on_click({
1066 let callback = self.sign_in_callback.clone();
1067 move |_, window, cx| (callback)(window, cx)
1068 }),
1069 );
1070 }
1071
1072 v_flex().gap_2().w_full().map(|this| {
1073 if self.account_too_young {
1074 this.child(YoungAccountBanner).child(
1075 Button::new("upgrade", "Upgrade to Pro")
1076 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1077 .full_width()
1078 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1079 )
1080 } else {
1081 this.text_sm()
1082 .child(subscription_text)
1083 .child(manage_subscription_buttons)
1084 }
1085 })
1086 }
1087}
1088
1089struct ConfigurationView {
1090 state: Entity<State>,
1091 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1092}
1093
1094impl ConfigurationView {
1095 fn new(state: Entity<State>) -> Self {
1096 let sign_in_callback = Arc::new({
1097 let state = state.clone();
1098 move |_window: &mut Window, cx: &mut App| {
1099 state.update(cx, |state, cx| {
1100 state.authenticate(cx).detach_and_log_err(cx);
1101 });
1102 }
1103 });
1104
1105 Self {
1106 state,
1107 sign_in_callback,
1108 }
1109 }
1110}
1111
1112impl Render for ConfigurationView {
1113 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1114 let state = self.state.read(cx);
1115 let user_store = state.user_store.read(cx);
1116
1117 ZedAiConfiguration {
1118 is_connected: !state.is_signed_out(cx),
1119 plan: user_store.plan(),
1120 subscription_period: user_store.subscription_period(),
1121 eligible_for_trial: user_store.trial_started_at().is_none(),
1122 account_too_young: user_store.account_too_young(),
1123 sign_in_callback: self.sign_in_callback.clone(),
1124 }
1125 }
1126}
1127
1128impl Component for ZedAiConfiguration {
1129 fn name() -> &'static str {
1130 "AI Configuration Content"
1131 }
1132
1133 fn sort_name() -> &'static str {
1134 "AI Configuration Content"
1135 }
1136
1137 fn scope() -> ComponentScope {
1138 ComponentScope::Onboarding
1139 }
1140
1141 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1142 fn configuration(
1143 is_connected: bool,
1144 plan: Option<Plan>,
1145 eligible_for_trial: bool,
1146 account_too_young: bool,
1147 ) -> AnyElement {
1148 ZedAiConfiguration {
1149 is_connected,
1150 plan,
1151 subscription_period: plan
1152 .is_some()
1153 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1154 eligible_for_trial,
1155 account_too_young,
1156 sign_in_callback: Arc::new(|_, _| {}),
1157 }
1158 .into_any_element()
1159 }
1160
1161 Some(
1162 v_flex()
1163 .p_4()
1164 .gap_4()
1165 .children(vec![
1166 single_example("Not connected", configuration(false, None, false, false)),
1167 single_example(
1168 "Accept Terms of Service",
1169 configuration(true, None, true, false),
1170 ),
1171 single_example(
1172 "No Plan - Not eligible for trial",
1173 configuration(true, None, false, false),
1174 ),
1175 single_example(
1176 "No Plan - Eligible for trial",
1177 configuration(true, None, true, false),
1178 ),
1179 single_example(
1180 "Free Plan",
1181 configuration(true, Some(Plan::ZedFree), true, false),
1182 ),
1183 single_example(
1184 "Zed Pro Trial Plan",
1185 configuration(true, Some(Plan::ZedProTrial), true, false),
1186 ),
1187 single_example(
1188 "Zed Pro Plan",
1189 configuration(true, Some(Plan::ZedPro), true, false),
1190 ),
1191 ])
1192 .into_any_element(),
1193 )
1194 }
1195}
1196
1197#[cfg(test)]
1198mod tests {
1199 use super::*;
1200 use http_client::http::{HeaderMap, StatusCode};
1201 use language_model::LanguageModelCompletionError;
1202
1203 #[test]
1204 fn test_api_error_conversion_with_upstream_http_error() {
1205 // upstream_http_error with 503 status should become ServerOverloaded
1206 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}"#;
1207
1208 let api_error = ApiError {
1209 status: StatusCode::INTERNAL_SERVER_ERROR,
1210 body: error_body.to_string(),
1211 headers: HeaderMap::new(),
1212 };
1213
1214 let completion_error: LanguageModelCompletionError = api_error.into();
1215
1216 match completion_error {
1217 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1218 assert_eq!(
1219 message,
1220 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1221 );
1222 }
1223 _ => panic!(
1224 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1225 completion_error
1226 ),
1227 }
1228
1229 // upstream_http_error with 500 status should become ApiInternalServerError
1230 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1231
1232 let api_error = ApiError {
1233 status: StatusCode::INTERNAL_SERVER_ERROR,
1234 body: error_body.to_string(),
1235 headers: HeaderMap::new(),
1236 };
1237
1238 let completion_error: LanguageModelCompletionError = api_error.into();
1239
1240 match completion_error {
1241 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1242 assert_eq!(
1243 message,
1244 "Received an error from the OpenAI API: internal server error"
1245 );
1246 }
1247 _ => panic!(
1248 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1249 completion_error
1250 ),
1251 }
1252
1253 // upstream_http_error with 429 status should become RateLimitExceeded
1254 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1255
1256 let api_error = ApiError {
1257 status: StatusCode::INTERNAL_SERVER_ERROR,
1258 body: error_body.to_string(),
1259 headers: HeaderMap::new(),
1260 };
1261
1262 let completion_error: LanguageModelCompletionError = api_error.into();
1263
1264 match completion_error {
1265 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1266 assert_eq!(
1267 message,
1268 "Received an error from the Google API: rate limit exceeded"
1269 );
1270 }
1271 _ => panic!(
1272 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1273 completion_error
1274 ),
1275 }
1276
1277 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1278 let error_body = "Regular internal server error";
1279
1280 let api_error = ApiError {
1281 status: StatusCode::INTERNAL_SERVER_ERROR,
1282 body: error_body.to_string(),
1283 headers: HeaderMap::new(),
1284 };
1285
1286 let completion_error: LanguageModelCompletionError = api_error.into();
1287
1288 match completion_error {
1289 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1290 assert_eq!(provider, PROVIDER_NAME);
1291 assert_eq!(message, "Regular internal server error");
1292 }
1293 _ => panic!(
1294 "Expected ApiInternalServerError for regular 500, got: {:?}",
1295 completion_error
1296 ),
1297 }
1298
1299 // upstream_http_429 format should be converted to UpstreamProviderError
1300 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1301
1302 let api_error = ApiError {
1303 status: StatusCode::INTERNAL_SERVER_ERROR,
1304 body: error_body.to_string(),
1305 headers: HeaderMap::new(),
1306 };
1307
1308 let completion_error: LanguageModelCompletionError = api_error.into();
1309
1310 match completion_error {
1311 LanguageModelCompletionError::UpstreamProviderError {
1312 message,
1313 status,
1314 retry_after,
1315 } => {
1316 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1317 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1318 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1319 }
1320 _ => panic!(
1321 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1322 completion_error
1323 ),
1324 }
1325
1326 // Invalid JSON in error body should fall back to regular error handling
1327 let error_body = "Not JSON at all";
1328
1329 let api_error = ApiError {
1330 status: StatusCode::INTERNAL_SERVER_ERROR,
1331 body: error_body.to_string(),
1332 headers: HeaderMap::new(),
1333 };
1334
1335 let completion_error: LanguageModelCompletionError = api_error.into();
1336
1337 match completion_error {
1338 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1339 assert_eq!(provider, PROVIDER_NAME);
1340 }
1341 _ => panic!(
1342 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1343 completion_error
1344 ),
1345 }
1346 }
1347}