1use ai_onboarding::YoungAccountBanner;
2use anthropic::AnthropicModelMode;
3use anyhow::{Context as _, Result, anyhow};
4use chrono::{DateTime, Utc};
5use client::{Client, ModelRequestUsage, UserStore, zed_urls};
6use cloud_llm_client::{
7 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
8 CompletionEvent, CompletionRequestStatus, CountTokensBody, CountTokensResponse,
9 EXPIRED_LLM_TOKEN_HEADER_NAME, ListModelsResponse, MODEL_REQUESTS_RESOURCE_HEADER_VALUE, Plan,
10 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
11 TOOL_USE_LIMIT_REACHED_HEADER_NAME, ZED_VERSION_HEADER_NAME,
12};
13use futures::{
14 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
15};
16use google_ai::GoogleModelMode;
17use gpui::{
18 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
19};
20use http_client::http::{HeaderMap, HeaderValue};
21use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
22use language_model::{
23 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
24 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
25 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
26 LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest,
27 LanguageModelToolChoice, LanguageModelToolSchemaFormat, LlmApiToken,
28 ModelRequestLimitReachedError, PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
29};
30use release_channel::AppVersion;
31use schemars::JsonSchema;
32use serde::{Deserialize, Serialize, de::DeserializeOwned};
33use settings::SettingsStore;
34use smol::io::{AsyncReadExt, BufReader};
35use std::pin::Pin;
36use std::str::FromStr as _;
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::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
44use crate::provider::google::{GoogleEventMapper, into_google};
45use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
46
47const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
48const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
49
50#[derive(Default, Clone, Debug, PartialEq)]
51pub struct ZedDotDevSettings {
52 pub available_models: Vec<AvailableModel>,
53}
54
55#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
56#[serde(rename_all = "lowercase")]
57pub enum AvailableProvider {
58 Anthropic,
59 OpenAi,
60 Google,
61}
62
63#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
64pub struct AvailableModel {
65 /// The provider of the language model.
66 pub provider: AvailableProvider,
67 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
68 pub name: String,
69 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
70 pub display_name: Option<String>,
71 /// The size of the context window, indicating the maximum number of tokens the model can process.
72 pub max_tokens: usize,
73 /// The maximum number of output tokens allowed by the model.
74 pub max_output_tokens: Option<u64>,
75 /// The maximum number of completion tokens allowed by the model (o1-* only)
76 pub max_completion_tokens: Option<u64>,
77 /// Override this model with a different Anthropic model for tool calls.
78 pub tool_override: Option<String>,
79 /// Indicates whether this custom model supports caching.
80 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
81 /// The default temperature to use for this model.
82 pub default_temperature: Option<f32>,
83 /// Any extra beta headers to provide when using the model.
84 #[serde(default)]
85 pub extra_beta_headers: Vec<String>,
86 /// The model's mode (e.g. thinking)
87 pub mode: Option<ModelMode>,
88}
89
90#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
91#[serde(tag = "type", rename_all = "lowercase")]
92pub enum ModelMode {
93 #[default]
94 Default,
95 Thinking {
96 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
97 budget_tokens: Option<u32>,
98 },
99}
100
101impl From<ModelMode> for AnthropicModelMode {
102 fn from(value: ModelMode) -> Self {
103 match value {
104 ModelMode::Default => AnthropicModelMode::Default,
105 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
106 }
107 }
108}
109
110pub struct CloudLanguageModelProvider {
111 client: Arc<Client>,
112 state: gpui::Entity<State>,
113 _maintain_client_status: Task<()>,
114}
115
116pub struct State {
117 client: Arc<Client>,
118 llm_api_token: LlmApiToken,
119 user_store: Entity<UserStore>,
120 status: client::Status,
121 accept_terms_of_service_task: Option<Task<Result<()>>>,
122 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
123 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
124 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
125 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
126 _fetch_models_task: Task<()>,
127 _settings_subscription: Subscription,
128 _llm_token_subscription: Subscription,
129}
130
131impl State {
132 fn new(
133 client: Arc<Client>,
134 user_store: Entity<UserStore>,
135 status: client::Status,
136 cx: &mut Context<Self>,
137 ) -> Self {
138 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
139 let mut current_user = user_store.read(cx).watch_current_user();
140 Self {
141 client: client.clone(),
142 llm_api_token: LlmApiToken::default(),
143 user_store: user_store.clone(),
144 status,
145 accept_terms_of_service_task: None,
146 models: Vec::new(),
147 default_model: None,
148 default_fast_model: None,
149 recommended_models: Vec::new(),
150 _fetch_models_task: cx.spawn(async move |this, cx| {
151 maybe!(async move {
152 let (client, llm_api_token) = this
153 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
154
155 while current_user.borrow().is_none() {
156 current_user.next().await;
157 }
158
159 let response =
160 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
161 this.update(cx, |this, cx| this.update_models(response, cx))?;
162 anyhow::Ok(())
163 })
164 .await
165 .context("failed to fetch Zed models")
166 .log_err();
167 }),
168 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
169 cx.notify();
170 }),
171 _llm_token_subscription: cx.subscribe(
172 &refresh_llm_token_listener,
173 move |this, _listener, _event, cx| {
174 let client = this.client.clone();
175 let llm_api_token = this.llm_api_token.clone();
176 cx.spawn(async move |this, cx| {
177 llm_api_token.refresh(&client).await?;
178 let response = Self::fetch_models(client, llm_api_token).await?;
179 this.update(cx, |this, cx| {
180 this.update_models(response, cx);
181 })
182 })
183 .detach_and_log_err(cx);
184 },
185 ),
186 }
187 }
188
189 fn is_signed_out(&self, cx: &App) -> bool {
190 self.user_store.read(cx).current_user().is_none()
191 }
192
193 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
194 let client = self.client.clone();
195 cx.spawn(async move |state, cx| {
196 client.sign_in_with_optional_connect(true, &cx).await?;
197 state.update(cx, |_, cx| cx.notify())
198 })
199 }
200
201 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
202 self.user_store.read(cx).has_accepted_terms_of_service()
203 }
204
205 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
206 let user_store = self.user_store.clone();
207 self.accept_terms_of_service_task = Some(cx.spawn(async move |this, cx| {
208 let _ = user_store
209 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
210 .await;
211 this.update(cx, |this, cx| {
212 this.accept_terms_of_service_task = None;
213 cx.notify()
214 })
215 }));
216 }
217
218 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
219 let mut models = Vec::new();
220
221 for model in response.models {
222 models.push(Arc::new(model.clone()));
223
224 // Right now we represent thinking variants of models as separate models on the client,
225 // so we need to insert variants for any model that supports thinking.
226 if model.supports_thinking {
227 models.push(Arc::new(cloud_llm_client::LanguageModel {
228 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
229 display_name: format!("{} Thinking", model.display_name),
230 ..model
231 }));
232 }
233 }
234
235 self.default_model = models
236 .iter()
237 .find(|model| model.id == response.default_model)
238 .cloned();
239 self.default_fast_model = models
240 .iter()
241 .find(|model| model.id == response.default_fast_model)
242 .cloned();
243 self.recommended_models = response
244 .recommended_models
245 .iter()
246 .filter_map(|id| models.iter().find(|model| &model.id == id))
247 .cloned()
248 .collect();
249 self.models = models;
250 cx.notify();
251 }
252
253 async fn fetch_models(
254 client: Arc<Client>,
255 llm_api_token: LlmApiToken,
256 ) -> Result<ListModelsResponse> {
257 let http_client = &client.http_client();
258 let token = llm_api_token.acquire(&client).await?;
259
260 let request = http_client::Request::builder()
261 .method(Method::GET)
262 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
263 .header("Authorization", format!("Bearer {token}"))
264 .body(AsyncBody::empty())?;
265 let mut response = http_client
266 .send(request)
267 .await
268 .context("failed to send list models request")?;
269
270 if response.status().is_success() {
271 let mut body = String::new();
272 response.body_mut().read_to_string(&mut body).await?;
273 return Ok(serde_json::from_str(&body)?);
274 } else {
275 let mut body = String::new();
276 response.body_mut().read_to_string(&mut body).await?;
277 anyhow::bail!(
278 "error listing models.\nStatus: {:?}\nBody: {body}",
279 response.status(),
280 );
281 }
282 }
283}
284
285impl CloudLanguageModelProvider {
286 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
287 let mut status_rx = client.status();
288 let status = *status_rx.borrow();
289
290 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
291
292 let state_ref = state.downgrade();
293 let maintain_client_status = cx.spawn(async move |cx| {
294 while let Some(status) = status_rx.next().await {
295 if let Some(this) = state_ref.upgrade() {
296 _ = this.update(cx, |this, cx| {
297 if this.status != status {
298 this.status = status;
299 cx.notify();
300 }
301 });
302 } else {
303 break;
304 }
305 }
306 });
307
308 Self {
309 client,
310 state: state.clone(),
311 _maintain_client_status: maintain_client_status,
312 }
313 }
314
315 fn create_language_model(
316 &self,
317 model: Arc<cloud_llm_client::LanguageModel>,
318 llm_api_token: LlmApiToken,
319 ) -> Arc<dyn LanguageModel> {
320 Arc::new(CloudLanguageModel {
321 id: LanguageModelId(SharedString::from(model.id.0.clone())),
322 model,
323 llm_api_token: llm_api_token.clone(),
324 client: self.client.clone(),
325 request_limiter: RateLimiter::new(4),
326 })
327 }
328}
329
330impl LanguageModelProviderState for CloudLanguageModelProvider {
331 type ObservableEntity = State;
332
333 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
334 Some(self.state.clone())
335 }
336}
337
338impl LanguageModelProvider for CloudLanguageModelProvider {
339 fn id(&self) -> LanguageModelProviderId {
340 PROVIDER_ID
341 }
342
343 fn name(&self) -> LanguageModelProviderName {
344 PROVIDER_NAME
345 }
346
347 fn icon(&self) -> IconName {
348 IconName::AiZed
349 }
350
351 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
352 let default_model = self.state.read(cx).default_model.clone()?;
353 let llm_api_token = self.state.read(cx).llm_api_token.clone();
354 Some(self.create_language_model(default_model, llm_api_token))
355 }
356
357 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
358 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
359 let llm_api_token = self.state.read(cx).llm_api_token.clone();
360 Some(self.create_language_model(default_fast_model, llm_api_token))
361 }
362
363 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
364 let llm_api_token = self.state.read(cx).llm_api_token.clone();
365 self.state
366 .read(cx)
367 .recommended_models
368 .iter()
369 .cloned()
370 .map(|model| self.create_language_model(model, llm_api_token.clone()))
371 .collect()
372 }
373
374 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
375 let llm_api_token = self.state.read(cx).llm_api_token.clone();
376 self.state
377 .read(cx)
378 .models
379 .iter()
380 .cloned()
381 .map(|model| self.create_language_model(model, llm_api_token.clone()))
382 .collect()
383 }
384
385 fn is_authenticated(&self, cx: &App) -> bool {
386 let state = self.state.read(cx);
387 !state.is_signed_out(cx) && state.has_accepted_terms_of_service(cx)
388 }
389
390 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
391 Task::ready(Ok(()))
392 }
393
394 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
395 cx.new(|_| ConfigurationView::new(self.state.clone()))
396 .into()
397 }
398
399 fn must_accept_terms(&self, cx: &App) -> bool {
400 !self.state.read(cx).has_accepted_terms_of_service(cx)
401 }
402
403 fn render_accept_terms(
404 &self,
405 view: LanguageModelProviderTosView,
406 cx: &mut App,
407 ) -> Option<AnyElement> {
408 let state = self.state.read(cx);
409 if state.has_accepted_terms_of_service(cx) {
410 return None;
411 }
412 Some(
413 render_accept_terms(view, state.accept_terms_of_service_task.is_some(), {
414 let state = self.state.clone();
415 move |_window, cx| {
416 state.update(cx, |state, cx| state.accept_terms_of_service(cx));
417 }
418 })
419 .into_any_element(),
420 )
421 }
422
423 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
424 Task::ready(Ok(()))
425 }
426}
427
428fn render_accept_terms(
429 view_kind: LanguageModelProviderTosView,
430 accept_terms_of_service_in_progress: bool,
431 accept_terms_callback: impl Fn(&mut Window, &mut App) + 'static,
432) -> impl IntoElement {
433 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
434 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadEmptyState);
435
436 let terms_button = Button::new("terms_of_service", "Terms of Service")
437 .style(ButtonStyle::Subtle)
438 .icon(IconName::ArrowUpRight)
439 .icon_color(Color::Muted)
440 .icon_size(IconSize::Small)
441 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
442 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
443
444 let button_container = h_flex().child(
445 Button::new("accept_terms", "I accept the Terms of Service")
446 .when(!thread_empty_state, |this| {
447 this.full_width()
448 .style(ButtonStyle::Tinted(TintColor::Accent))
449 .icon(IconName::Check)
450 .icon_position(IconPosition::Start)
451 .icon_size(IconSize::Small)
452 })
453 .when(thread_empty_state, |this| {
454 this.style(ButtonStyle::Tinted(TintColor::Warning))
455 .label_size(LabelSize::Small)
456 })
457 .disabled(accept_terms_of_service_in_progress)
458 .on_click(move |_, window, cx| (accept_terms_callback)(window, cx)),
459 );
460
461 if thread_empty_state {
462 h_flex()
463 .w_full()
464 .flex_wrap()
465 .justify_between()
466 .child(
467 h_flex()
468 .child(
469 Label::new("To start using Zed AI, please read and accept the")
470 .size(LabelSize::Small),
471 )
472 .child(terms_button),
473 )
474 .child(button_container)
475 } else {
476 v_flex()
477 .w_full()
478 .gap_2()
479 .child(
480 h_flex()
481 .flex_wrap()
482 .when(thread_fresh_start, |this| this.justify_center())
483 .child(Label::new(
484 "To start using Zed AI, please read and accept the",
485 ))
486 .child(terms_button),
487 )
488 .child({
489 match view_kind {
490 LanguageModelProviderTosView::TextThreadPopup => {
491 button_container.w_full().justify_end()
492 }
493 LanguageModelProviderTosView::Configuration => {
494 button_container.w_full().justify_start()
495 }
496 LanguageModelProviderTosView::ThreadFreshStart => {
497 button_container.w_full().justify_center()
498 }
499 LanguageModelProviderTosView::ThreadEmptyState => div().w_0(),
500 }
501 })
502 }
503}
504
505pub struct CloudLanguageModel {
506 id: LanguageModelId,
507 model: Arc<cloud_llm_client::LanguageModel>,
508 llm_api_token: LlmApiToken,
509 client: Arc<Client>,
510 request_limiter: RateLimiter,
511}
512
513struct PerformLlmCompletionResponse {
514 response: Response<AsyncBody>,
515 usage: Option<ModelRequestUsage>,
516 tool_use_limit_reached: bool,
517 includes_status_messages: bool,
518}
519
520impl CloudLanguageModel {
521 async fn perform_llm_completion(
522 client: Arc<Client>,
523 llm_api_token: LlmApiToken,
524 app_version: Option<SemanticVersion>,
525 body: CompletionBody,
526 ) -> Result<PerformLlmCompletionResponse> {
527 let http_client = &client.http_client();
528
529 let mut token = llm_api_token.acquire(&client).await?;
530 let mut refreshed_token = false;
531
532 loop {
533 let request_builder = http_client::Request::builder()
534 .method(Method::POST)
535 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
536 let request_builder = if let Some(app_version) = app_version {
537 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
538 } else {
539 request_builder
540 };
541
542 let request = request_builder
543 .header("Content-Type", "application/json")
544 .header("Authorization", format!("Bearer {token}"))
545 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
546 .body(serde_json::to_string(&body)?.into())?;
547 let mut response = http_client.send(request).await?;
548 let status = response.status();
549 if status.is_success() {
550 let includes_status_messages = response
551 .headers()
552 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
553 .is_some();
554
555 let tool_use_limit_reached = response
556 .headers()
557 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
558 .is_some();
559
560 let usage = if includes_status_messages {
561 None
562 } else {
563 ModelRequestUsage::from_headers(response.headers()).ok()
564 };
565
566 return Ok(PerformLlmCompletionResponse {
567 response,
568 usage,
569 includes_status_messages,
570 tool_use_limit_reached,
571 });
572 }
573
574 if !refreshed_token
575 && response
576 .headers()
577 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
578 .is_some()
579 {
580 token = llm_api_token.refresh(&client).await?;
581 refreshed_token = true;
582 continue;
583 }
584
585 if status == StatusCode::FORBIDDEN
586 && response
587 .headers()
588 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
589 .is_some()
590 {
591 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
592 .headers()
593 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
594 .and_then(|resource| resource.to_str().ok())
595 {
596 if let Some(plan) = response
597 .headers()
598 .get(CURRENT_PLAN_HEADER_NAME)
599 .and_then(|plan| plan.to_str().ok())
600 .and_then(|plan| cloud_llm_client::Plan::from_str(plan).ok())
601 {
602 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
603 }
604 }
605 } else if status == StatusCode::PAYMENT_REQUIRED {
606 return Err(anyhow!(PaymentRequiredError));
607 }
608
609 let mut body = String::new();
610 let headers = response.headers().clone();
611 response.body_mut().read_to_string(&mut body).await?;
612 return Err(anyhow!(ApiError {
613 status,
614 body,
615 headers
616 }));
617 }
618 }
619}
620
621#[derive(Debug, Error)]
622#[error("cloud language model request failed with status {status}: {body}")]
623struct ApiError {
624 status: StatusCode,
625 body: String,
626 headers: HeaderMap<HeaderValue>,
627}
628
629/// Represents error responses from Zed's cloud API.
630///
631/// Example JSON for an upstream HTTP error:
632/// ```json
633/// {
634/// "code": "upstream_http_error",
635/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
636/// "upstream_status": 503
637/// }
638/// ```
639#[derive(Debug, serde::Deserialize)]
640struct CloudApiError {
641 code: String,
642 message: String,
643 #[serde(default)]
644 #[serde(deserialize_with = "deserialize_optional_status_code")]
645 upstream_status: Option<StatusCode>,
646 #[serde(default)]
647 retry_after: Option<f64>,
648}
649
650fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
651where
652 D: serde::Deserializer<'de>,
653{
654 let opt: Option<u16> = Option::deserialize(deserializer)?;
655 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
656}
657
658impl From<ApiError> for LanguageModelCompletionError {
659 fn from(error: ApiError) -> Self {
660 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
661 if cloud_error.code.starts_with("upstream_http_") {
662 let status = if let Some(status) = cloud_error.upstream_status {
663 status
664 } else if cloud_error.code.ends_with("_error") {
665 error.status
666 } else {
667 // If there's a status code in the code string (e.g. "upstream_http_429")
668 // then use that; otherwise, see if the JSON contains a status code.
669 cloud_error
670 .code
671 .strip_prefix("upstream_http_")
672 .and_then(|code_str| code_str.parse::<u16>().ok())
673 .and_then(|code| StatusCode::from_u16(code).ok())
674 .unwrap_or(error.status)
675 };
676
677 return LanguageModelCompletionError::UpstreamProviderError {
678 message: cloud_error.message,
679 status,
680 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
681 };
682 }
683 }
684
685 let retry_after = None;
686 LanguageModelCompletionError::from_http_status(
687 PROVIDER_NAME,
688 error.status,
689 error.body,
690 retry_after,
691 )
692 }
693}
694
695impl LanguageModel for CloudLanguageModel {
696 fn id(&self) -> LanguageModelId {
697 self.id.clone()
698 }
699
700 fn name(&self) -> LanguageModelName {
701 LanguageModelName::from(self.model.display_name.clone())
702 }
703
704 fn provider_id(&self) -> LanguageModelProviderId {
705 PROVIDER_ID
706 }
707
708 fn provider_name(&self) -> LanguageModelProviderName {
709 PROVIDER_NAME
710 }
711
712 fn upstream_provider_id(&self) -> LanguageModelProviderId {
713 use cloud_llm_client::LanguageModelProvider::*;
714 match self.model.provider {
715 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
716 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
717 Google => language_model::GOOGLE_PROVIDER_ID,
718 }
719 }
720
721 fn upstream_provider_name(&self) -> LanguageModelProviderName {
722 use cloud_llm_client::LanguageModelProvider::*;
723 match self.model.provider {
724 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
725 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
726 Google => language_model::GOOGLE_PROVIDER_NAME,
727 }
728 }
729
730 fn supports_tools(&self) -> bool {
731 self.model.supports_tools
732 }
733
734 fn supports_images(&self) -> bool {
735 self.model.supports_images
736 }
737
738 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
739 match choice {
740 LanguageModelToolChoice::Auto
741 | LanguageModelToolChoice::Any
742 | LanguageModelToolChoice::None => true,
743 }
744 }
745
746 fn supports_burn_mode(&self) -> bool {
747 self.model.supports_max_mode
748 }
749
750 fn telemetry_id(&self) -> String {
751 format!("zed.dev/{}", self.model.id)
752 }
753
754 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
755 match self.model.provider {
756 cloud_llm_client::LanguageModelProvider::Anthropic
757 | cloud_llm_client::LanguageModelProvider::OpenAi => {
758 LanguageModelToolSchemaFormat::JsonSchema
759 }
760 cloud_llm_client::LanguageModelProvider::Google => {
761 LanguageModelToolSchemaFormat::JsonSchemaSubset
762 }
763 }
764 }
765
766 fn max_token_count(&self) -> u64 {
767 self.model.max_token_count as u64
768 }
769
770 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
771 self.model
772 .max_token_count_in_max_mode
773 .filter(|_| self.model.supports_max_mode)
774 .map(|max_token_count| max_token_count as u64)
775 }
776
777 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
778 match &self.model.provider {
779 cloud_llm_client::LanguageModelProvider::Anthropic => {
780 Some(LanguageModelCacheConfiguration {
781 min_total_token: 2_048,
782 should_speculate: true,
783 max_cache_anchors: 4,
784 })
785 }
786 cloud_llm_client::LanguageModelProvider::OpenAi
787 | cloud_llm_client::LanguageModelProvider::Google => None,
788 }
789 }
790
791 fn count_tokens(
792 &self,
793 request: LanguageModelRequest,
794 cx: &App,
795 ) -> BoxFuture<'static, Result<u64>> {
796 match self.model.provider {
797 cloud_llm_client::LanguageModelProvider::Anthropic => {
798 count_anthropic_tokens(request, cx)
799 }
800 cloud_llm_client::LanguageModelProvider::OpenAi => {
801 let model = match open_ai::Model::from_id(&self.model.id.0) {
802 Ok(model) => model,
803 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
804 };
805 count_open_ai_tokens(request, model, cx)
806 }
807 cloud_llm_client::LanguageModelProvider::Google => {
808 let client = self.client.clone();
809 let llm_api_token = self.llm_api_token.clone();
810 let model_id = self.model.id.to_string();
811 let generate_content_request =
812 into_google(request, model_id.clone(), GoogleModelMode::Default);
813 async move {
814 let http_client = &client.http_client();
815 let token = llm_api_token.acquire(&client).await?;
816
817 let request_body = CountTokensBody {
818 provider: cloud_llm_client::LanguageModelProvider::Google,
819 model: model_id,
820 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
821 generate_content_request,
822 })?,
823 };
824 let request = http_client::Request::builder()
825 .method(Method::POST)
826 .uri(
827 http_client
828 .build_zed_llm_url("/count_tokens", &[])?
829 .as_ref(),
830 )
831 .header("Content-Type", "application/json")
832 .header("Authorization", format!("Bearer {token}"))
833 .body(serde_json::to_string(&request_body)?.into())?;
834 let mut response = http_client.send(request).await?;
835 let status = response.status();
836 let headers = response.headers().clone();
837 let mut response_body = String::new();
838 response
839 .body_mut()
840 .read_to_string(&mut response_body)
841 .await?;
842
843 if status.is_success() {
844 let response_body: CountTokensResponse =
845 serde_json::from_str(&response_body)?;
846
847 Ok(response_body.tokens as u64)
848 } else {
849 Err(anyhow!(ApiError {
850 status,
851 body: response_body,
852 headers
853 }))
854 }
855 }
856 .boxed()
857 }
858 }
859 }
860
861 fn stream_completion(
862 &self,
863 request: LanguageModelRequest,
864 cx: &AsyncApp,
865 ) -> BoxFuture<
866 'static,
867 Result<
868 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
869 LanguageModelCompletionError,
870 >,
871 > {
872 let thread_id = request.thread_id.clone();
873 let prompt_id = request.prompt_id.clone();
874 let intent = request.intent;
875 let mode = request.mode;
876 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
877 let thinking_allowed = request.thinking_allowed;
878 match self.model.provider {
879 cloud_llm_client::LanguageModelProvider::Anthropic => {
880 let request = into_anthropic(
881 request,
882 self.model.id.to_string(),
883 1.0,
884 self.model.max_output_tokens as u64,
885 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
886 AnthropicModelMode::Thinking {
887 budget_tokens: Some(4_096),
888 }
889 } else {
890 AnthropicModelMode::Default
891 },
892 );
893 let client = self.client.clone();
894 let llm_api_token = self.llm_api_token.clone();
895 let future = self.request_limiter.stream(async move {
896 let PerformLlmCompletionResponse {
897 response,
898 usage,
899 includes_status_messages,
900 tool_use_limit_reached,
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 mode,
910 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
911 model: request.model.clone(),
912 provider_request: serde_json::to_value(&request)
913 .map_err(|e| anyhow!(e))?,
914 },
915 )
916 .await
917 .map_err(|err| match err.downcast::<ApiError>() {
918 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
919 Err(err) => anyhow!(err),
920 })?;
921
922 let mut mapper = AnthropicEventMapper::new();
923 Ok(map_cloud_completion_events(
924 Box::pin(
925 response_lines(response, includes_status_messages)
926 .chain(usage_updated_event(usage))
927 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
928 ),
929 move |event| mapper.map_event(event),
930 ))
931 });
932 async move { Ok(future.await?.boxed()) }.boxed()
933 }
934 cloud_llm_client::LanguageModelProvider::OpenAi => {
935 let client = self.client.clone();
936 let model = match open_ai::Model::from_id(&self.model.id.0) {
937 Ok(model) => model,
938 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
939 };
940 let request = into_open_ai(
941 request,
942 model.id(),
943 model.supports_parallel_tool_calls(),
944 None,
945 None,
946 );
947 let llm_api_token = self.llm_api_token.clone();
948 let future = self.request_limiter.stream(async move {
949 let PerformLlmCompletionResponse {
950 response,
951 usage,
952 includes_status_messages,
953 tool_use_limit_reached,
954 } = Self::perform_llm_completion(
955 client.clone(),
956 llm_api_token,
957 app_version,
958 CompletionBody {
959 thread_id,
960 prompt_id,
961 intent,
962 mode,
963 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
964 model: request.model.clone(),
965 provider_request: serde_json::to_value(&request)
966 .map_err(|e| anyhow!(e))?,
967 },
968 )
969 .await?;
970
971 let mut mapper = OpenAiEventMapper::new();
972 Ok(map_cloud_completion_events(
973 Box::pin(
974 response_lines(response, includes_status_messages)
975 .chain(usage_updated_event(usage))
976 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
977 ),
978 move |event| mapper.map_event(event),
979 ))
980 });
981 async move { Ok(future.await?.boxed()) }.boxed()
982 }
983 cloud_llm_client::LanguageModelProvider::Google => {
984 let client = self.client.clone();
985 let request =
986 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
987 let llm_api_token = self.llm_api_token.clone();
988 let future = self.request_limiter.stream(async move {
989 let PerformLlmCompletionResponse {
990 response,
991 usage,
992 includes_status_messages,
993 tool_use_limit_reached,
994 } = Self::perform_llm_completion(
995 client.clone(),
996 llm_api_token,
997 app_version,
998 CompletionBody {
999 thread_id,
1000 prompt_id,
1001 intent,
1002 mode,
1003 provider: cloud_llm_client::LanguageModelProvider::Google,
1004 model: request.model.model_id.clone(),
1005 provider_request: serde_json::to_value(&request)
1006 .map_err(|e| anyhow!(e))?,
1007 },
1008 )
1009 .await?;
1010
1011 let mut mapper = GoogleEventMapper::new();
1012 Ok(map_cloud_completion_events(
1013 Box::pin(
1014 response_lines(response, includes_status_messages)
1015 .chain(usage_updated_event(usage))
1016 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
1017 ),
1018 move |event| mapper.map_event(event),
1019 ))
1020 });
1021 async move { Ok(future.await?.boxed()) }.boxed()
1022 }
1023 }
1024 }
1025}
1026
1027fn map_cloud_completion_events<T, F>(
1028 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
1029 mut map_callback: F,
1030) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1031where
1032 T: DeserializeOwned + 'static,
1033 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1034 + Send
1035 + 'static,
1036{
1037 stream
1038 .flat_map(move |event| {
1039 futures::stream::iter(match event {
1040 Err(error) => {
1041 vec![Err(LanguageModelCompletionError::from(error))]
1042 }
1043 Ok(CompletionEvent::Status(event)) => {
1044 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
1045 }
1046 Ok(CompletionEvent::Event(event)) => map_callback(event),
1047 })
1048 })
1049 .boxed()
1050}
1051
1052fn usage_updated_event<T>(
1053 usage: Option<ModelRequestUsage>,
1054) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1055 futures::stream::iter(usage.map(|usage| {
1056 Ok(CompletionEvent::Status(
1057 CompletionRequestStatus::UsageUpdated {
1058 amount: usage.amount as usize,
1059 limit: usage.limit,
1060 },
1061 ))
1062 }))
1063}
1064
1065fn tool_use_limit_reached_event<T>(
1066 tool_use_limit_reached: bool,
1067) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1068 futures::stream::iter(tool_use_limit_reached.then(|| {
1069 Ok(CompletionEvent::Status(
1070 CompletionRequestStatus::ToolUseLimitReached,
1071 ))
1072 }))
1073}
1074
1075fn response_lines<T: DeserializeOwned>(
1076 response: Response<AsyncBody>,
1077 includes_status_messages: bool,
1078) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1079 futures::stream::try_unfold(
1080 (String::new(), BufReader::new(response.into_body())),
1081 move |(mut line, mut body)| async move {
1082 match body.read_line(&mut line).await {
1083 Ok(0) => Ok(None),
1084 Ok(_) => {
1085 let event = if includes_status_messages {
1086 serde_json::from_str::<CompletionEvent<T>>(&line)?
1087 } else {
1088 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1089 };
1090
1091 line.clear();
1092 Ok(Some((event, (line, body))))
1093 }
1094 Err(e) => Err(e.into()),
1095 }
1096 },
1097 )
1098}
1099
1100#[derive(IntoElement, RegisterComponent)]
1101struct ZedAiConfiguration {
1102 is_connected: bool,
1103 plan: Option<Plan>,
1104 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1105 eligible_for_trial: bool,
1106 has_accepted_terms_of_service: bool,
1107 account_too_young: bool,
1108 accept_terms_of_service_in_progress: bool,
1109 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1110 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1111}
1112
1113impl RenderOnce for ZedAiConfiguration {
1114 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1115 let young_account_banner = YoungAccountBanner;
1116
1117 let is_pro = self.plan == Some(Plan::ZedPro);
1118 let subscription_text = match (self.plan, self.subscription_period) {
1119 (Some(Plan::ZedPro), Some(_)) => {
1120 "You have access to Zed's hosted models through your Pro subscription."
1121 }
1122 (Some(Plan::ZedProTrial), Some(_)) => {
1123 "You have access to Zed's hosted models through your Pro trial."
1124 }
1125 (Some(Plan::ZedFree), Some(_)) => {
1126 "You have basic access to Zed's hosted models through the Free plan."
1127 }
1128 _ => {
1129 if self.eligible_for_trial {
1130 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1131 } else {
1132 "Subscribe for access to Zed's hosted models."
1133 }
1134 }
1135 };
1136
1137 let manage_subscription_buttons = if is_pro {
1138 Button::new("manage_settings", "Manage Subscription")
1139 .full_width()
1140 .style(ButtonStyle::Tinted(TintColor::Accent))
1141 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1142 .into_any_element()
1143 } else if self.plan.is_none() || self.eligible_for_trial {
1144 Button::new("start_trial", "Start 14-day Free Pro Trial")
1145 .full_width()
1146 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1147 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1148 .into_any_element()
1149 } else {
1150 Button::new("upgrade", "Upgrade to Pro")
1151 .full_width()
1152 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1153 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1154 .into_any_element()
1155 };
1156
1157 if !self.is_connected {
1158 return v_flex()
1159 .gap_2()
1160 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1161 .child(
1162 Button::new("sign_in", "Sign In to use Zed AI")
1163 .icon_color(Color::Muted)
1164 .icon(IconName::Github)
1165 .icon_size(IconSize::Small)
1166 .icon_position(IconPosition::Start)
1167 .full_width()
1168 .on_click({
1169 let callback = self.sign_in_callback.clone();
1170 move |_, window, cx| (callback)(window, cx)
1171 }),
1172 );
1173 }
1174
1175 v_flex()
1176 .gap_2()
1177 .w_full()
1178 .when(!self.has_accepted_terms_of_service, |this| {
1179 this.child(render_accept_terms(
1180 LanguageModelProviderTosView::Configuration,
1181 self.accept_terms_of_service_in_progress,
1182 {
1183 let callback = self.accept_terms_of_service_callback.clone();
1184 move |window, cx| (callback)(window, cx)
1185 },
1186 ))
1187 })
1188 .map(|this| {
1189 if self.has_accepted_terms_of_service && self.account_too_young {
1190 this.child(young_account_banner).child(
1191 Button::new("upgrade", "Upgrade to Pro")
1192 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1193 .full_width()
1194 .on_click(|_, _, cx| {
1195 cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))
1196 }),
1197 )
1198 } else if self.has_accepted_terms_of_service {
1199 this.text_sm()
1200 .child(subscription_text)
1201 .child(manage_subscription_buttons)
1202 } else {
1203 this
1204 }
1205 })
1206 .when(self.has_accepted_terms_of_service, |this| this)
1207 }
1208}
1209
1210struct ConfigurationView {
1211 state: Entity<State>,
1212 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1213 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1214}
1215
1216impl ConfigurationView {
1217 fn new(state: Entity<State>) -> Self {
1218 let accept_terms_of_service_callback = Arc::new({
1219 let state = state.clone();
1220 move |_window: &mut Window, cx: &mut App| {
1221 state.update(cx, |state, cx| {
1222 state.accept_terms_of_service(cx);
1223 });
1224 }
1225 });
1226
1227 let sign_in_callback = Arc::new({
1228 let state = state.clone();
1229 move |_window: &mut Window, cx: &mut App| {
1230 state.update(cx, |state, cx| {
1231 state.authenticate(cx).detach_and_log_err(cx);
1232 });
1233 }
1234 });
1235
1236 Self {
1237 state,
1238 accept_terms_of_service_callback,
1239 sign_in_callback,
1240 }
1241 }
1242}
1243
1244impl Render for ConfigurationView {
1245 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1246 let state = self.state.read(cx);
1247 let user_store = state.user_store.read(cx);
1248
1249 ZedAiConfiguration {
1250 is_connected: !state.is_signed_out(cx),
1251 plan: user_store.plan(),
1252 subscription_period: user_store.subscription_period(),
1253 eligible_for_trial: user_store.trial_started_at().is_none(),
1254 has_accepted_terms_of_service: state.has_accepted_terms_of_service(cx),
1255 account_too_young: user_store.account_too_young(),
1256 accept_terms_of_service_in_progress: state.accept_terms_of_service_task.is_some(),
1257 accept_terms_of_service_callback: self.accept_terms_of_service_callback.clone(),
1258 sign_in_callback: self.sign_in_callback.clone(),
1259 }
1260 }
1261}
1262
1263impl Component for ZedAiConfiguration {
1264 fn name() -> &'static str {
1265 "AI Configuration Content"
1266 }
1267
1268 fn sort_name() -> &'static str {
1269 "AI Configuration Content"
1270 }
1271
1272 fn scope() -> ComponentScope {
1273 ComponentScope::Onboarding
1274 }
1275
1276 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1277 fn configuration(
1278 is_connected: bool,
1279 plan: Option<Plan>,
1280 eligible_for_trial: bool,
1281 account_too_young: bool,
1282 has_accepted_terms_of_service: bool,
1283 ) -> AnyElement {
1284 ZedAiConfiguration {
1285 is_connected,
1286 plan,
1287 subscription_period: plan
1288 .is_some()
1289 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1290 eligible_for_trial,
1291 has_accepted_terms_of_service,
1292 account_too_young,
1293 accept_terms_of_service_in_progress: false,
1294 accept_terms_of_service_callback: Arc::new(|_, _| {}),
1295 sign_in_callback: Arc::new(|_, _| {}),
1296 }
1297 .into_any_element()
1298 }
1299
1300 Some(
1301 v_flex()
1302 .p_4()
1303 .gap_4()
1304 .children(vec![
1305 single_example(
1306 "Not connected",
1307 configuration(false, None, false, false, true),
1308 ),
1309 single_example(
1310 "Accept Terms of Service",
1311 configuration(true, None, true, false, false),
1312 ),
1313 single_example(
1314 "No Plan - Not eligible for trial",
1315 configuration(true, None, false, false, true),
1316 ),
1317 single_example(
1318 "No Plan - Eligible for trial",
1319 configuration(true, None, true, false, true),
1320 ),
1321 single_example(
1322 "Free Plan",
1323 configuration(true, Some(Plan::ZedFree), true, false, true),
1324 ),
1325 single_example(
1326 "Zed Pro Trial Plan",
1327 configuration(true, Some(Plan::ZedProTrial), true, false, true),
1328 ),
1329 single_example(
1330 "Zed Pro Plan",
1331 configuration(true, Some(Plan::ZedPro), true, false, true),
1332 ),
1333 ])
1334 .into_any_element(),
1335 )
1336 }
1337}
1338
1339#[cfg(test)]
1340mod tests {
1341 use super::*;
1342 use http_client::http::{HeaderMap, StatusCode};
1343 use language_model::LanguageModelCompletionError;
1344
1345 #[test]
1346 fn test_api_error_conversion_with_upstream_http_error() {
1347 // upstream_http_error with 503 status should become ServerOverloaded
1348 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}"#;
1349
1350 let api_error = ApiError {
1351 status: StatusCode::INTERNAL_SERVER_ERROR,
1352 body: error_body.to_string(),
1353 headers: HeaderMap::new(),
1354 };
1355
1356 let completion_error: LanguageModelCompletionError = api_error.into();
1357
1358 match completion_error {
1359 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1360 assert_eq!(
1361 message,
1362 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1363 );
1364 }
1365 _ => panic!(
1366 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1367 completion_error
1368 ),
1369 }
1370
1371 // upstream_http_error with 500 status should become ApiInternalServerError
1372 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1373
1374 let api_error = ApiError {
1375 status: StatusCode::INTERNAL_SERVER_ERROR,
1376 body: error_body.to_string(),
1377 headers: HeaderMap::new(),
1378 };
1379
1380 let completion_error: LanguageModelCompletionError = api_error.into();
1381
1382 match completion_error {
1383 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1384 assert_eq!(
1385 message,
1386 "Received an error from the OpenAI API: internal server error"
1387 );
1388 }
1389 _ => panic!(
1390 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1391 completion_error
1392 ),
1393 }
1394
1395 // upstream_http_error with 429 status should become RateLimitExceeded
1396 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1397
1398 let api_error = ApiError {
1399 status: StatusCode::INTERNAL_SERVER_ERROR,
1400 body: error_body.to_string(),
1401 headers: HeaderMap::new(),
1402 };
1403
1404 let completion_error: LanguageModelCompletionError = api_error.into();
1405
1406 match completion_error {
1407 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1408 assert_eq!(
1409 message,
1410 "Received an error from the Google API: rate limit exceeded"
1411 );
1412 }
1413 _ => panic!(
1414 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1415 completion_error
1416 ),
1417 }
1418
1419 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1420 let error_body = "Regular internal server error";
1421
1422 let api_error = ApiError {
1423 status: StatusCode::INTERNAL_SERVER_ERROR,
1424 body: error_body.to_string(),
1425 headers: HeaderMap::new(),
1426 };
1427
1428 let completion_error: LanguageModelCompletionError = api_error.into();
1429
1430 match completion_error {
1431 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1432 assert_eq!(provider, PROVIDER_NAME);
1433 assert_eq!(message, "Regular internal server error");
1434 }
1435 _ => panic!(
1436 "Expected ApiInternalServerError for regular 500, got: {:?}",
1437 completion_error
1438 ),
1439 }
1440
1441 // upstream_http_429 format should be converted to UpstreamProviderError
1442 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1443
1444 let api_error = ApiError {
1445 status: StatusCode::INTERNAL_SERVER_ERROR,
1446 body: error_body.to_string(),
1447 headers: HeaderMap::new(),
1448 };
1449
1450 let completion_error: LanguageModelCompletionError = api_error.into();
1451
1452 match completion_error {
1453 LanguageModelCompletionError::UpstreamProviderError {
1454 message,
1455 status,
1456 retry_after,
1457 } => {
1458 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1459 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1460 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1461 }
1462 _ => panic!(
1463 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1464 completion_error
1465 ),
1466 }
1467
1468 // Invalid JSON in error body should fall back to regular error handling
1469 let error_body = "Not JSON at all";
1470
1471 let api_error = ApiError {
1472 status: StatusCode::INTERNAL_SERVER_ERROR,
1473 body: error_body.to_string(),
1474 headers: HeaderMap::new(),
1475 };
1476
1477 let completion_error: LanguageModelCompletionError = api_error.into();
1478
1479 match completion_error {
1480 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1481 assert_eq!(provider, PROVIDER_NAME);
1482 }
1483 _ => panic!(
1484 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1485 completion_error
1486 ),
1487 }
1488 }
1489}