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(
395 &self,
396 _target_agent: language_model::ConfigurationViewTargetAgent,
397 _: &mut Window,
398 cx: &mut App,
399 ) -> AnyView {
400 cx.new(|_| ConfigurationView::new(self.state.clone()))
401 .into()
402 }
403
404 fn must_accept_terms(&self, cx: &App) -> bool {
405 !self.state.read(cx).has_accepted_terms_of_service(cx)
406 }
407
408 fn render_accept_terms(
409 &self,
410 view: LanguageModelProviderTosView,
411 cx: &mut App,
412 ) -> Option<AnyElement> {
413 let state = self.state.read(cx);
414 if state.has_accepted_terms_of_service(cx) {
415 return None;
416 }
417 Some(
418 render_accept_terms(view, state.accept_terms_of_service_task.is_some(), {
419 let state = self.state.clone();
420 move |_window, cx| {
421 state.update(cx, |state, cx| state.accept_terms_of_service(cx));
422 }
423 })
424 .into_any_element(),
425 )
426 }
427
428 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
429 Task::ready(Ok(()))
430 }
431}
432
433fn render_accept_terms(
434 view_kind: LanguageModelProviderTosView,
435 accept_terms_of_service_in_progress: bool,
436 accept_terms_callback: impl Fn(&mut Window, &mut App) + 'static,
437) -> impl IntoElement {
438 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
439 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadEmptyState);
440
441 let terms_button = Button::new("terms_of_service", "Terms of Service")
442 .style(ButtonStyle::Subtle)
443 .icon(IconName::ArrowUpRight)
444 .icon_color(Color::Muted)
445 .icon_size(IconSize::Small)
446 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
447 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
448
449 let button_container = h_flex().child(
450 Button::new("accept_terms", "I accept the Terms of Service")
451 .when(!thread_empty_state, |this| {
452 this.full_width()
453 .style(ButtonStyle::Tinted(TintColor::Accent))
454 .icon(IconName::Check)
455 .icon_position(IconPosition::Start)
456 .icon_size(IconSize::Small)
457 })
458 .when(thread_empty_state, |this| {
459 this.style(ButtonStyle::Tinted(TintColor::Warning))
460 .label_size(LabelSize::Small)
461 })
462 .disabled(accept_terms_of_service_in_progress)
463 .on_click(move |_, window, cx| (accept_terms_callback)(window, cx)),
464 );
465
466 if thread_empty_state {
467 h_flex()
468 .w_full()
469 .flex_wrap()
470 .justify_between()
471 .child(
472 h_flex()
473 .child(
474 Label::new("To start using Zed AI, please read and accept the")
475 .size(LabelSize::Small),
476 )
477 .child(terms_button),
478 )
479 .child(button_container)
480 } else {
481 v_flex()
482 .w_full()
483 .gap_2()
484 .child(
485 h_flex()
486 .flex_wrap()
487 .when(thread_fresh_start, |this| this.justify_center())
488 .child(Label::new(
489 "To start using Zed AI, please read and accept the",
490 ))
491 .child(terms_button),
492 )
493 .child({
494 match view_kind {
495 LanguageModelProviderTosView::TextThreadPopup => {
496 button_container.w_full().justify_end()
497 }
498 LanguageModelProviderTosView::Configuration => {
499 button_container.w_full().justify_start()
500 }
501 LanguageModelProviderTosView::ThreadFreshStart => {
502 button_container.w_full().justify_center()
503 }
504 LanguageModelProviderTosView::ThreadEmptyState => div().w_0(),
505 }
506 })
507 }
508}
509
510pub struct CloudLanguageModel {
511 id: LanguageModelId,
512 model: Arc<cloud_llm_client::LanguageModel>,
513 llm_api_token: LlmApiToken,
514 client: Arc<Client>,
515 request_limiter: RateLimiter,
516}
517
518struct PerformLlmCompletionResponse {
519 response: Response<AsyncBody>,
520 usage: Option<ModelRequestUsage>,
521 tool_use_limit_reached: bool,
522 includes_status_messages: bool,
523}
524
525impl CloudLanguageModel {
526 async fn perform_llm_completion(
527 client: Arc<Client>,
528 llm_api_token: LlmApiToken,
529 app_version: Option<SemanticVersion>,
530 body: CompletionBody,
531 ) -> Result<PerformLlmCompletionResponse> {
532 let http_client = &client.http_client();
533
534 let mut token = llm_api_token.acquire(&client).await?;
535 let mut refreshed_token = false;
536
537 loop {
538 let request_builder = http_client::Request::builder()
539 .method(Method::POST)
540 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
541 let request_builder = if let Some(app_version) = app_version {
542 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
543 } else {
544 request_builder
545 };
546
547 let request = request_builder
548 .header("Content-Type", "application/json")
549 .header("Authorization", format!("Bearer {token}"))
550 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
551 .body(serde_json::to_string(&body)?.into())?;
552 let mut response = http_client.send(request).await?;
553 let status = response.status();
554 if status.is_success() {
555 let includes_status_messages = response
556 .headers()
557 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
558 .is_some();
559
560 let tool_use_limit_reached = response
561 .headers()
562 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
563 .is_some();
564
565 let usage = if includes_status_messages {
566 None
567 } else {
568 ModelRequestUsage::from_headers(response.headers()).ok()
569 };
570
571 return Ok(PerformLlmCompletionResponse {
572 response,
573 usage,
574 includes_status_messages,
575 tool_use_limit_reached,
576 });
577 }
578
579 if !refreshed_token
580 && response
581 .headers()
582 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
583 .is_some()
584 {
585 token = llm_api_token.refresh(&client).await?;
586 refreshed_token = true;
587 continue;
588 }
589
590 if status == StatusCode::FORBIDDEN
591 && response
592 .headers()
593 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
594 .is_some()
595 {
596 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
597 .headers()
598 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
599 .and_then(|resource| resource.to_str().ok())
600 && let Some(plan) = response
601 .headers()
602 .get(CURRENT_PLAN_HEADER_NAME)
603 .and_then(|plan| plan.to_str().ok())
604 .and_then(|plan| cloud_llm_client::Plan::from_str(plan).ok())
605 {
606 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
607 }
608 } else if status == StatusCode::PAYMENT_REQUIRED {
609 return Err(anyhow!(PaymentRequiredError));
610 }
611
612 let mut body = String::new();
613 let headers = response.headers().clone();
614 response.body_mut().read_to_string(&mut body).await?;
615 return Err(anyhow!(ApiError {
616 status,
617 body,
618 headers
619 }));
620 }
621 }
622}
623
624#[derive(Debug, Error)]
625#[error("cloud language model request failed with status {status}: {body}")]
626struct ApiError {
627 status: StatusCode,
628 body: String,
629 headers: HeaderMap<HeaderValue>,
630}
631
632/// Represents error responses from Zed's cloud API.
633///
634/// Example JSON for an upstream HTTP error:
635/// ```json
636/// {
637/// "code": "upstream_http_error",
638/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
639/// "upstream_status": 503
640/// }
641/// ```
642#[derive(Debug, serde::Deserialize)]
643struct CloudApiError {
644 code: String,
645 message: String,
646 #[serde(default)]
647 #[serde(deserialize_with = "deserialize_optional_status_code")]
648 upstream_status: Option<StatusCode>,
649 #[serde(default)]
650 retry_after: Option<f64>,
651}
652
653fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
654where
655 D: serde::Deserializer<'de>,
656{
657 let opt: Option<u16> = Option::deserialize(deserializer)?;
658 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
659}
660
661impl From<ApiError> for LanguageModelCompletionError {
662 fn from(error: ApiError) -> Self {
663 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body)
664 && cloud_error.code.starts_with("upstream_http_") {
665 let status = if let Some(status) = cloud_error.upstream_status {
666 status
667 } else if cloud_error.code.ends_with("_error") {
668 error.status
669 } else {
670 // If there's a status code in the code string (e.g. "upstream_http_429")
671 // then use that; otherwise, see if the JSON contains a status code.
672 cloud_error
673 .code
674 .strip_prefix("upstream_http_")
675 .and_then(|code_str| code_str.parse::<u16>().ok())
676 .and_then(|code| StatusCode::from_u16(code).ok())
677 .unwrap_or(error.status)
678 };
679
680 return LanguageModelCompletionError::UpstreamProviderError {
681 message: cloud_error.message,
682 status,
683 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
684 };
685 }
686
687 let retry_after = None;
688 LanguageModelCompletionError::from_http_status(
689 PROVIDER_NAME,
690 error.status,
691 error.body,
692 retry_after,
693 )
694 }
695}
696
697impl LanguageModel for CloudLanguageModel {
698 fn id(&self) -> LanguageModelId {
699 self.id.clone()
700 }
701
702 fn name(&self) -> LanguageModelName {
703 LanguageModelName::from(self.model.display_name.clone())
704 }
705
706 fn provider_id(&self) -> LanguageModelProviderId {
707 PROVIDER_ID
708 }
709
710 fn provider_name(&self) -> LanguageModelProviderName {
711 PROVIDER_NAME
712 }
713
714 fn upstream_provider_id(&self) -> LanguageModelProviderId {
715 use cloud_llm_client::LanguageModelProvider::*;
716 match self.model.provider {
717 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
718 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
719 Google => language_model::GOOGLE_PROVIDER_ID,
720 }
721 }
722
723 fn upstream_provider_name(&self) -> LanguageModelProviderName {
724 use cloud_llm_client::LanguageModelProvider::*;
725 match self.model.provider {
726 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
727 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
728 Google => language_model::GOOGLE_PROVIDER_NAME,
729 }
730 }
731
732 fn supports_tools(&self) -> bool {
733 self.model.supports_tools
734 }
735
736 fn supports_images(&self) -> bool {
737 self.model.supports_images
738 }
739
740 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
741 match choice {
742 LanguageModelToolChoice::Auto
743 | LanguageModelToolChoice::Any
744 | LanguageModelToolChoice::None => true,
745 }
746 }
747
748 fn supports_burn_mode(&self) -> bool {
749 self.model.supports_max_mode
750 }
751
752 fn telemetry_id(&self) -> String {
753 format!("zed.dev/{}", self.model.id)
754 }
755
756 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
757 match self.model.provider {
758 cloud_llm_client::LanguageModelProvider::Anthropic
759 | cloud_llm_client::LanguageModelProvider::OpenAi => {
760 LanguageModelToolSchemaFormat::JsonSchema
761 }
762 cloud_llm_client::LanguageModelProvider::Google => {
763 LanguageModelToolSchemaFormat::JsonSchemaSubset
764 }
765 }
766 }
767
768 fn max_token_count(&self) -> u64 {
769 self.model.max_token_count as u64
770 }
771
772 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
773 self.model
774 .max_token_count_in_max_mode
775 .filter(|_| self.model.supports_max_mode)
776 .map(|max_token_count| max_token_count as u64)
777 }
778
779 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
780 match &self.model.provider {
781 cloud_llm_client::LanguageModelProvider::Anthropic => {
782 Some(LanguageModelCacheConfiguration {
783 min_total_token: 2_048,
784 should_speculate: true,
785 max_cache_anchors: 4,
786 })
787 }
788 cloud_llm_client::LanguageModelProvider::OpenAi
789 | cloud_llm_client::LanguageModelProvider::Google => None,
790 }
791 }
792
793 fn count_tokens(
794 &self,
795 request: LanguageModelRequest,
796 cx: &App,
797 ) -> BoxFuture<'static, Result<u64>> {
798 match self.model.provider {
799 cloud_llm_client::LanguageModelProvider::Anthropic => {
800 count_anthropic_tokens(request, cx)
801 }
802 cloud_llm_client::LanguageModelProvider::OpenAi => {
803 let model = match open_ai::Model::from_id(&self.model.id.0) {
804 Ok(model) => model,
805 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
806 };
807 count_open_ai_tokens(request, model, cx)
808 }
809 cloud_llm_client::LanguageModelProvider::Google => {
810 let client = self.client.clone();
811 let llm_api_token = self.llm_api_token.clone();
812 let model_id = self.model.id.to_string();
813 let generate_content_request =
814 into_google(request, model_id.clone(), GoogleModelMode::Default);
815 async move {
816 let http_client = &client.http_client();
817 let token = llm_api_token.acquire(&client).await?;
818
819 let request_body = CountTokensBody {
820 provider: cloud_llm_client::LanguageModelProvider::Google,
821 model: model_id,
822 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
823 generate_content_request,
824 })?,
825 };
826 let request = http_client::Request::builder()
827 .method(Method::POST)
828 .uri(
829 http_client
830 .build_zed_llm_url("/count_tokens", &[])?
831 .as_ref(),
832 )
833 .header("Content-Type", "application/json")
834 .header("Authorization", format!("Bearer {token}"))
835 .body(serde_json::to_string(&request_body)?.into())?;
836 let mut response = http_client.send(request).await?;
837 let status = response.status();
838 let headers = response.headers().clone();
839 let mut response_body = String::new();
840 response
841 .body_mut()
842 .read_to_string(&mut response_body)
843 .await?;
844
845 if status.is_success() {
846 let response_body: CountTokensResponse =
847 serde_json::from_str(&response_body)?;
848
849 Ok(response_body.tokens as u64)
850 } else {
851 Err(anyhow!(ApiError {
852 status,
853 body: response_body,
854 headers
855 }))
856 }
857 }
858 .boxed()
859 }
860 }
861 }
862
863 fn stream_completion(
864 &self,
865 request: LanguageModelRequest,
866 cx: &AsyncApp,
867 ) -> BoxFuture<
868 'static,
869 Result<
870 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
871 LanguageModelCompletionError,
872 >,
873 > {
874 let thread_id = request.thread_id.clone();
875 let prompt_id = request.prompt_id.clone();
876 let intent = request.intent;
877 let mode = request.mode;
878 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
879 let thinking_allowed = request.thinking_allowed;
880 match self.model.provider {
881 cloud_llm_client::LanguageModelProvider::Anthropic => {
882 let request = into_anthropic(
883 request,
884 self.model.id.to_string(),
885 1.0,
886 self.model.max_output_tokens as u64,
887 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
888 AnthropicModelMode::Thinking {
889 budget_tokens: Some(4_096),
890 }
891 } else {
892 AnthropicModelMode::Default
893 },
894 );
895 let client = self.client.clone();
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 usage,
901 includes_status_messages,
902 tool_use_limit_reached,
903 } = Self::perform_llm_completion(
904 client.clone(),
905 llm_api_token,
906 app_version,
907 CompletionBody {
908 thread_id,
909 prompt_id,
910 intent,
911 mode,
912 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
913 model: request.model.clone(),
914 provider_request: serde_json::to_value(&request)
915 .map_err(|e| anyhow!(e))?,
916 },
917 )
918 .await
919 .map_err(|err| match err.downcast::<ApiError>() {
920 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
921 Err(err) => anyhow!(err),
922 })?;
923
924 let mut mapper = AnthropicEventMapper::new();
925 Ok(map_cloud_completion_events(
926 Box::pin(
927 response_lines(response, includes_status_messages)
928 .chain(usage_updated_event(usage))
929 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
930 ),
931 move |event| mapper.map_event(event),
932 ))
933 });
934 async move { Ok(future.await?.boxed()) }.boxed()
935 }
936 cloud_llm_client::LanguageModelProvider::OpenAi => {
937 let client = self.client.clone();
938 let model = match open_ai::Model::from_id(&self.model.id.0) {
939 Ok(model) => model,
940 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
941 };
942 let request = into_open_ai(
943 request,
944 model.id(),
945 model.supports_parallel_tool_calls(),
946 model.supports_prompt_cache_key(),
947 None,
948 None,
949 );
950 let llm_api_token = self.llm_api_token.clone();
951 let future = self.request_limiter.stream(async move {
952 let PerformLlmCompletionResponse {
953 response,
954 usage,
955 includes_status_messages,
956 tool_use_limit_reached,
957 } = Self::perform_llm_completion(
958 client.clone(),
959 llm_api_token,
960 app_version,
961 CompletionBody {
962 thread_id,
963 prompt_id,
964 intent,
965 mode,
966 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
967 model: request.model.clone(),
968 provider_request: serde_json::to_value(&request)
969 .map_err(|e| anyhow!(e))?,
970 },
971 )
972 .await?;
973
974 let mut mapper = OpenAiEventMapper::new();
975 Ok(map_cloud_completion_events(
976 Box::pin(
977 response_lines(response, includes_status_messages)
978 .chain(usage_updated_event(usage))
979 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
980 ),
981 move |event| mapper.map_event(event),
982 ))
983 });
984 async move { Ok(future.await?.boxed()) }.boxed()
985 }
986 cloud_llm_client::LanguageModelProvider::Google => {
987 let client = self.client.clone();
988 let request =
989 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
990 let llm_api_token = self.llm_api_token.clone();
991 let future = self.request_limiter.stream(async move {
992 let PerformLlmCompletionResponse {
993 response,
994 usage,
995 includes_status_messages,
996 tool_use_limit_reached,
997 } = Self::perform_llm_completion(
998 client.clone(),
999 llm_api_token,
1000 app_version,
1001 CompletionBody {
1002 thread_id,
1003 prompt_id,
1004 intent,
1005 mode,
1006 provider: cloud_llm_client::LanguageModelProvider::Google,
1007 model: request.model.model_id.clone(),
1008 provider_request: serde_json::to_value(&request)
1009 .map_err(|e| anyhow!(e))?,
1010 },
1011 )
1012 .await?;
1013
1014 let mut mapper = GoogleEventMapper::new();
1015 Ok(map_cloud_completion_events(
1016 Box::pin(
1017 response_lines(response, includes_status_messages)
1018 .chain(usage_updated_event(usage))
1019 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
1020 ),
1021 move |event| mapper.map_event(event),
1022 ))
1023 });
1024 async move { Ok(future.await?.boxed()) }.boxed()
1025 }
1026 }
1027 }
1028}
1029
1030fn map_cloud_completion_events<T, F>(
1031 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
1032 mut map_callback: F,
1033) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1034where
1035 T: DeserializeOwned + 'static,
1036 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1037 + Send
1038 + 'static,
1039{
1040 stream
1041 .flat_map(move |event| {
1042 futures::stream::iter(match event {
1043 Err(error) => {
1044 vec![Err(LanguageModelCompletionError::from(error))]
1045 }
1046 Ok(CompletionEvent::Status(event)) => {
1047 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
1048 }
1049 Ok(CompletionEvent::Event(event)) => map_callback(event),
1050 })
1051 })
1052 .boxed()
1053}
1054
1055fn usage_updated_event<T>(
1056 usage: Option<ModelRequestUsage>,
1057) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1058 futures::stream::iter(usage.map(|usage| {
1059 Ok(CompletionEvent::Status(
1060 CompletionRequestStatus::UsageUpdated {
1061 amount: usage.amount as usize,
1062 limit: usage.limit,
1063 },
1064 ))
1065 }))
1066}
1067
1068fn tool_use_limit_reached_event<T>(
1069 tool_use_limit_reached: bool,
1070) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1071 futures::stream::iter(tool_use_limit_reached.then(|| {
1072 Ok(CompletionEvent::Status(
1073 CompletionRequestStatus::ToolUseLimitReached,
1074 ))
1075 }))
1076}
1077
1078fn response_lines<T: DeserializeOwned>(
1079 response: Response<AsyncBody>,
1080 includes_status_messages: bool,
1081) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1082 futures::stream::try_unfold(
1083 (String::new(), BufReader::new(response.into_body())),
1084 move |(mut line, mut body)| async move {
1085 match body.read_line(&mut line).await {
1086 Ok(0) => Ok(None),
1087 Ok(_) => {
1088 let event = if includes_status_messages {
1089 serde_json::from_str::<CompletionEvent<T>>(&line)?
1090 } else {
1091 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1092 };
1093
1094 line.clear();
1095 Ok(Some((event, (line, body))))
1096 }
1097 Err(e) => Err(e.into()),
1098 }
1099 },
1100 )
1101}
1102
1103#[derive(IntoElement, RegisterComponent)]
1104struct ZedAiConfiguration {
1105 is_connected: bool,
1106 plan: Option<Plan>,
1107 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1108 eligible_for_trial: bool,
1109 has_accepted_terms_of_service: bool,
1110 account_too_young: bool,
1111 accept_terms_of_service_in_progress: bool,
1112 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1113 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1114}
1115
1116impl RenderOnce for ZedAiConfiguration {
1117 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1118 let young_account_banner = YoungAccountBanner;
1119
1120 let is_pro = self.plan == Some(Plan::ZedPro);
1121 let subscription_text = match (self.plan, self.subscription_period) {
1122 (Some(Plan::ZedPro), Some(_)) => {
1123 "You have access to Zed's hosted models through your Pro subscription."
1124 }
1125 (Some(Plan::ZedProTrial), Some(_)) => {
1126 "You have access to Zed's hosted models through your Pro trial."
1127 }
1128 (Some(Plan::ZedFree), Some(_)) => {
1129 "You have basic access to Zed's hosted models through the Free plan."
1130 }
1131 _ => {
1132 if self.eligible_for_trial {
1133 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1134 } else {
1135 "Subscribe for access to Zed's hosted models."
1136 }
1137 }
1138 };
1139
1140 let manage_subscription_buttons = if is_pro {
1141 Button::new("manage_settings", "Manage Subscription")
1142 .full_width()
1143 .style(ButtonStyle::Tinted(TintColor::Accent))
1144 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1145 .into_any_element()
1146 } else if self.plan.is_none() || self.eligible_for_trial {
1147 Button::new("start_trial", "Start 14-day Free Pro Trial")
1148 .full_width()
1149 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1150 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1151 .into_any_element()
1152 } else {
1153 Button::new("upgrade", "Upgrade to Pro")
1154 .full_width()
1155 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1156 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1157 .into_any_element()
1158 };
1159
1160 if !self.is_connected {
1161 return v_flex()
1162 .gap_2()
1163 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1164 .child(
1165 Button::new("sign_in", "Sign In to use Zed AI")
1166 .icon_color(Color::Muted)
1167 .icon(IconName::Github)
1168 .icon_size(IconSize::Small)
1169 .icon_position(IconPosition::Start)
1170 .full_width()
1171 .on_click({
1172 let callback = self.sign_in_callback.clone();
1173 move |_, window, cx| (callback)(window, cx)
1174 }),
1175 );
1176 }
1177
1178 v_flex()
1179 .gap_2()
1180 .w_full()
1181 .when(!self.has_accepted_terms_of_service, |this| {
1182 this.child(render_accept_terms(
1183 LanguageModelProviderTosView::Configuration,
1184 self.accept_terms_of_service_in_progress,
1185 {
1186 let callback = self.accept_terms_of_service_callback.clone();
1187 move |window, cx| (callback)(window, cx)
1188 },
1189 ))
1190 })
1191 .map(|this| {
1192 if self.has_accepted_terms_of_service && self.account_too_young {
1193 this.child(young_account_banner).child(
1194 Button::new("upgrade", "Upgrade to Pro")
1195 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1196 .full_width()
1197 .on_click(|_, _, cx| {
1198 cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))
1199 }),
1200 )
1201 } else if self.has_accepted_terms_of_service {
1202 this.text_sm()
1203 .child(subscription_text)
1204 .child(manage_subscription_buttons)
1205 } else {
1206 this
1207 }
1208 })
1209 .when(self.has_accepted_terms_of_service, |this| this)
1210 }
1211}
1212
1213struct ConfigurationView {
1214 state: Entity<State>,
1215 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1216 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1217}
1218
1219impl ConfigurationView {
1220 fn new(state: Entity<State>) -> Self {
1221 let accept_terms_of_service_callback = Arc::new({
1222 let state = state.clone();
1223 move |_window: &mut Window, cx: &mut App| {
1224 state.update(cx, |state, cx| {
1225 state.accept_terms_of_service(cx);
1226 });
1227 }
1228 });
1229
1230 let sign_in_callback = Arc::new({
1231 let state = state.clone();
1232 move |_window: &mut Window, cx: &mut App| {
1233 state.update(cx, |state, cx| {
1234 state.authenticate(cx).detach_and_log_err(cx);
1235 });
1236 }
1237 });
1238
1239 Self {
1240 state,
1241 accept_terms_of_service_callback,
1242 sign_in_callback,
1243 }
1244 }
1245}
1246
1247impl Render for ConfigurationView {
1248 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1249 let state = self.state.read(cx);
1250 let user_store = state.user_store.read(cx);
1251
1252 ZedAiConfiguration {
1253 is_connected: !state.is_signed_out(cx),
1254 plan: user_store.plan(),
1255 subscription_period: user_store.subscription_period(),
1256 eligible_for_trial: user_store.trial_started_at().is_none(),
1257 has_accepted_terms_of_service: state.has_accepted_terms_of_service(cx),
1258 account_too_young: user_store.account_too_young(),
1259 accept_terms_of_service_in_progress: state.accept_terms_of_service_task.is_some(),
1260 accept_terms_of_service_callback: self.accept_terms_of_service_callback.clone(),
1261 sign_in_callback: self.sign_in_callback.clone(),
1262 }
1263 }
1264}
1265
1266impl Component for ZedAiConfiguration {
1267 fn name() -> &'static str {
1268 "AI Configuration Content"
1269 }
1270
1271 fn sort_name() -> &'static str {
1272 "AI Configuration Content"
1273 }
1274
1275 fn scope() -> ComponentScope {
1276 ComponentScope::Onboarding
1277 }
1278
1279 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1280 fn configuration(
1281 is_connected: bool,
1282 plan: Option<Plan>,
1283 eligible_for_trial: bool,
1284 account_too_young: bool,
1285 has_accepted_terms_of_service: bool,
1286 ) -> AnyElement {
1287 ZedAiConfiguration {
1288 is_connected,
1289 plan,
1290 subscription_period: plan
1291 .is_some()
1292 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1293 eligible_for_trial,
1294 has_accepted_terms_of_service,
1295 account_too_young,
1296 accept_terms_of_service_in_progress: false,
1297 accept_terms_of_service_callback: Arc::new(|_, _| {}),
1298 sign_in_callback: Arc::new(|_, _| {}),
1299 }
1300 .into_any_element()
1301 }
1302
1303 Some(
1304 v_flex()
1305 .p_4()
1306 .gap_4()
1307 .children(vec![
1308 single_example(
1309 "Not connected",
1310 configuration(false, None, false, false, true),
1311 ),
1312 single_example(
1313 "Accept Terms of Service",
1314 configuration(true, None, true, false, false),
1315 ),
1316 single_example(
1317 "No Plan - Not eligible for trial",
1318 configuration(true, None, false, false, true),
1319 ),
1320 single_example(
1321 "No Plan - Eligible for trial",
1322 configuration(true, None, true, false, true),
1323 ),
1324 single_example(
1325 "Free Plan",
1326 configuration(true, Some(Plan::ZedFree), true, false, true),
1327 ),
1328 single_example(
1329 "Zed Pro Trial Plan",
1330 configuration(true, Some(Plan::ZedProTrial), true, false, true),
1331 ),
1332 single_example(
1333 "Zed Pro Plan",
1334 configuration(true, Some(Plan::ZedPro), true, false, true),
1335 ),
1336 ])
1337 .into_any_element(),
1338 )
1339 }
1340}
1341
1342#[cfg(test)]
1343mod tests {
1344 use super::*;
1345 use http_client::http::{HeaderMap, StatusCode};
1346 use language_model::LanguageModelCompletionError;
1347
1348 #[test]
1349 fn test_api_error_conversion_with_upstream_http_error() {
1350 // upstream_http_error with 503 status should become ServerOverloaded
1351 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}"#;
1352
1353 let api_error = ApiError {
1354 status: StatusCode::INTERNAL_SERVER_ERROR,
1355 body: error_body.to_string(),
1356 headers: HeaderMap::new(),
1357 };
1358
1359 let completion_error: LanguageModelCompletionError = api_error.into();
1360
1361 match completion_error {
1362 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1363 assert_eq!(
1364 message,
1365 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1366 );
1367 }
1368 _ => panic!(
1369 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1370 completion_error
1371 ),
1372 }
1373
1374 // upstream_http_error with 500 status should become ApiInternalServerError
1375 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1376
1377 let api_error = ApiError {
1378 status: StatusCode::INTERNAL_SERVER_ERROR,
1379 body: error_body.to_string(),
1380 headers: HeaderMap::new(),
1381 };
1382
1383 let completion_error: LanguageModelCompletionError = api_error.into();
1384
1385 match completion_error {
1386 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1387 assert_eq!(
1388 message,
1389 "Received an error from the OpenAI API: internal server error"
1390 );
1391 }
1392 _ => panic!(
1393 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1394 completion_error
1395 ),
1396 }
1397
1398 // upstream_http_error with 429 status should become RateLimitExceeded
1399 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1400
1401 let api_error = ApiError {
1402 status: StatusCode::INTERNAL_SERVER_ERROR,
1403 body: error_body.to_string(),
1404 headers: HeaderMap::new(),
1405 };
1406
1407 let completion_error: LanguageModelCompletionError = api_error.into();
1408
1409 match completion_error {
1410 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1411 assert_eq!(
1412 message,
1413 "Received an error from the Google API: rate limit exceeded"
1414 );
1415 }
1416 _ => panic!(
1417 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1418 completion_error
1419 ),
1420 }
1421
1422 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1423 let error_body = "Regular internal server error";
1424
1425 let api_error = ApiError {
1426 status: StatusCode::INTERNAL_SERVER_ERROR,
1427 body: error_body.to_string(),
1428 headers: HeaderMap::new(),
1429 };
1430
1431 let completion_error: LanguageModelCompletionError = api_error.into();
1432
1433 match completion_error {
1434 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1435 assert_eq!(provider, PROVIDER_NAME);
1436 assert_eq!(message, "Regular internal server error");
1437 }
1438 _ => panic!(
1439 "Expected ApiInternalServerError for regular 500, got: {:?}",
1440 completion_error
1441 ),
1442 }
1443
1444 // upstream_http_429 format should be converted to UpstreamProviderError
1445 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1446
1447 let api_error = ApiError {
1448 status: StatusCode::INTERNAL_SERVER_ERROR,
1449 body: error_body.to_string(),
1450 headers: HeaderMap::new(),
1451 };
1452
1453 let completion_error: LanguageModelCompletionError = api_error.into();
1454
1455 match completion_error {
1456 LanguageModelCompletionError::UpstreamProviderError {
1457 message,
1458 status,
1459 retry_after,
1460 } => {
1461 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1462 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1463 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1464 }
1465 _ => panic!(
1466 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1467 completion_error
1468 ),
1469 }
1470
1471 // Invalid JSON in error body should fall back to regular error handling
1472 let error_body = "Not JSON at all";
1473
1474 let api_error = ApiError {
1475 status: StatusCode::INTERNAL_SERVER_ERROR,
1476 body: error_body.to_string(),
1477 headers: HeaderMap::new(),
1478 };
1479
1480 let completion_error: LanguageModelCompletionError = api_error.into();
1481
1482 match completion_error {
1483 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1484 assert_eq!(provider, PROVIDER_NAME);
1485 }
1486 _ => panic!(
1487 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1488 completion_error
1489 ),
1490 }
1491 }
1492}