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 Self {
140 client: client.clone(),
141 llm_api_token: LlmApiToken::default(),
142 user_store: user_store.clone(),
143 status,
144 accept_terms_of_service_task: None,
145 models: Vec::new(),
146 default_model: None,
147 default_fast_model: None,
148 recommended_models: Vec::new(),
149 _fetch_models_task: cx.spawn(async move |this, cx| {
150 maybe!(async move {
151 let (client, llm_api_token) = this
152 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
153
154 loop {
155 let is_authenticated = user_store
156 .read_with(cx, |user_store, _cx| user_store.current_user().is_some())?;
157 if is_authenticated {
158 break;
159 }
160
161 cx.background_executor()
162 .timer(Duration::from_millis(100))
163 .await;
164 }
165
166 let response = Self::fetch_models(client, llm_api_token).await?;
167 this.update(cx, |this, cx| {
168 this.update_models(response, cx);
169 })
170 })
171 .await
172 .context("failed to fetch Zed models")
173 .log_err();
174 }),
175 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
176 cx.notify();
177 }),
178 _llm_token_subscription: cx.subscribe(
179 &refresh_llm_token_listener,
180 move |this, _listener, _event, cx| {
181 let client = this.client.clone();
182 let llm_api_token = this.llm_api_token.clone();
183 cx.spawn(async move |this, cx| {
184 llm_api_token.refresh(&client).await?;
185 let response = Self::fetch_models(client, llm_api_token).await?;
186 this.update(cx, |this, cx| {
187 this.update_models(response, cx);
188 })
189 })
190 .detach_and_log_err(cx);
191 },
192 ),
193 }
194 }
195
196 fn is_signed_out(&self, cx: &App) -> bool {
197 self.user_store.read(cx).current_user().is_none()
198 }
199
200 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
201 let client = self.client.clone();
202 cx.spawn(async move |state, cx| {
203 client.sign_in_with_optional_connect(true, &cx).await?;
204 state.update(cx, |_, cx| cx.notify())
205 })
206 }
207
208 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
209 self.user_store.read(cx).has_accepted_terms_of_service()
210 }
211
212 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
213 let user_store = self.user_store.clone();
214 self.accept_terms_of_service_task = Some(cx.spawn(async move |this, cx| {
215 let _ = user_store
216 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
217 .await;
218 this.update(cx, |this, cx| {
219 this.accept_terms_of_service_task = None;
220 cx.notify()
221 })
222 }));
223 }
224
225 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
226 let mut models = Vec::new();
227
228 for model in response.models {
229 models.push(Arc::new(model.clone()));
230
231 // Right now we represent thinking variants of models as separate models on the client,
232 // so we need to insert variants for any model that supports thinking.
233 if model.supports_thinking {
234 models.push(Arc::new(cloud_llm_client::LanguageModel {
235 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
236 display_name: format!("{} Thinking", model.display_name),
237 ..model
238 }));
239 }
240 }
241
242 self.default_model = models
243 .iter()
244 .find(|model| model.id == response.default_model)
245 .cloned();
246 self.default_fast_model = models
247 .iter()
248 .find(|model| model.id == response.default_fast_model)
249 .cloned();
250 self.recommended_models = response
251 .recommended_models
252 .iter()
253 .filter_map(|id| models.iter().find(|model| &model.id == id))
254 .cloned()
255 .collect();
256 self.models = models;
257 cx.notify();
258 }
259
260 async fn fetch_models(
261 client: Arc<Client>,
262 llm_api_token: LlmApiToken,
263 ) -> Result<ListModelsResponse> {
264 let http_client = &client.http_client();
265 let token = llm_api_token.acquire(&client).await?;
266
267 let request = http_client::Request::builder()
268 .method(Method::GET)
269 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
270 .header("Authorization", format!("Bearer {token}"))
271 .body(AsyncBody::empty())?;
272 let mut response = http_client
273 .send(request)
274 .await
275 .context("failed to send list models request")?;
276
277 if response.status().is_success() {
278 let mut body = String::new();
279 response.body_mut().read_to_string(&mut body).await?;
280 return Ok(serde_json::from_str(&body)?);
281 } else {
282 let mut body = String::new();
283 response.body_mut().read_to_string(&mut body).await?;
284 anyhow::bail!(
285 "error listing models.\nStatus: {:?}\nBody: {body}",
286 response.status(),
287 );
288 }
289 }
290}
291
292impl CloudLanguageModelProvider {
293 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
294 let mut status_rx = client.status();
295 let status = *status_rx.borrow();
296
297 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
298
299 let state_ref = state.downgrade();
300 let maintain_client_status = cx.spawn(async move |cx| {
301 while let Some(status) = status_rx.next().await {
302 if let Some(this) = state_ref.upgrade() {
303 _ = this.update(cx, |this, cx| {
304 if this.status != status {
305 this.status = status;
306 cx.notify();
307 }
308 });
309 } else {
310 break;
311 }
312 }
313 });
314
315 Self {
316 client,
317 state: state.clone(),
318 _maintain_client_status: maintain_client_status,
319 }
320 }
321
322 fn create_language_model(
323 &self,
324 model: Arc<cloud_llm_client::LanguageModel>,
325 llm_api_token: LlmApiToken,
326 ) -> Arc<dyn LanguageModel> {
327 Arc::new(CloudLanguageModel {
328 id: LanguageModelId(SharedString::from(model.id.0.clone())),
329 model,
330 llm_api_token: llm_api_token.clone(),
331 client: self.client.clone(),
332 request_limiter: RateLimiter::new(4),
333 })
334 }
335}
336
337impl LanguageModelProviderState for CloudLanguageModelProvider {
338 type ObservableEntity = State;
339
340 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
341 Some(self.state.clone())
342 }
343}
344
345impl LanguageModelProvider for CloudLanguageModelProvider {
346 fn id(&self) -> LanguageModelProviderId {
347 PROVIDER_ID
348 }
349
350 fn name(&self) -> LanguageModelProviderName {
351 PROVIDER_NAME
352 }
353
354 fn icon(&self) -> IconName {
355 IconName::AiZed
356 }
357
358 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
359 let default_model = self.state.read(cx).default_model.clone()?;
360 let llm_api_token = self.state.read(cx).llm_api_token.clone();
361 Some(self.create_language_model(default_model, llm_api_token))
362 }
363
364 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
365 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
366 let llm_api_token = self.state.read(cx).llm_api_token.clone();
367 Some(self.create_language_model(default_fast_model, llm_api_token))
368 }
369
370 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
371 let llm_api_token = self.state.read(cx).llm_api_token.clone();
372 self.state
373 .read(cx)
374 .recommended_models
375 .iter()
376 .cloned()
377 .map(|model| self.create_language_model(model, llm_api_token.clone()))
378 .collect()
379 }
380
381 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
382 let llm_api_token = self.state.read(cx).llm_api_token.clone();
383 self.state
384 .read(cx)
385 .models
386 .iter()
387 .cloned()
388 .map(|model| self.create_language_model(model, llm_api_token.clone()))
389 .collect()
390 }
391
392 fn is_authenticated(&self, cx: &App) -> bool {
393 let state = self.state.read(cx);
394 !state.is_signed_out(cx) && state.has_accepted_terms_of_service(cx)
395 }
396
397 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
398 Task::ready(Ok(()))
399 }
400
401 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
402 cx.new(|_| ConfigurationView::new(self.state.clone()))
403 .into()
404 }
405
406 fn must_accept_terms(&self, cx: &App) -> bool {
407 !self.state.read(cx).has_accepted_terms_of_service(cx)
408 }
409
410 fn render_accept_terms(
411 &self,
412 view: LanguageModelProviderTosView,
413 cx: &mut App,
414 ) -> Option<AnyElement> {
415 let state = self.state.read(cx);
416 if state.has_accepted_terms_of_service(cx) {
417 return None;
418 }
419 Some(
420 render_accept_terms(view, state.accept_terms_of_service_task.is_some(), {
421 let state = self.state.clone();
422 move |_window, cx| {
423 state.update(cx, |state, cx| state.accept_terms_of_service(cx));
424 }
425 })
426 .into_any_element(),
427 )
428 }
429
430 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
431 Task::ready(Ok(()))
432 }
433}
434
435fn render_accept_terms(
436 view_kind: LanguageModelProviderTosView,
437 accept_terms_of_service_in_progress: bool,
438 accept_terms_callback: impl Fn(&mut Window, &mut App) + 'static,
439) -> impl IntoElement {
440 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
441 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadEmptyState);
442
443 let terms_button = Button::new("terms_of_service", "Terms of Service")
444 .style(ButtonStyle::Subtle)
445 .icon(IconName::ArrowUpRight)
446 .icon_color(Color::Muted)
447 .icon_size(IconSize::XSmall)
448 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
449 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
450
451 let button_container = h_flex().child(
452 Button::new("accept_terms", "I accept the Terms of Service")
453 .when(!thread_empty_state, |this| {
454 this.full_width()
455 .style(ButtonStyle::Tinted(TintColor::Accent))
456 .icon(IconName::Check)
457 .icon_position(IconPosition::Start)
458 .icon_size(IconSize::Small)
459 })
460 .when(thread_empty_state, |this| {
461 this.style(ButtonStyle::Tinted(TintColor::Warning))
462 .label_size(LabelSize::Small)
463 })
464 .disabled(accept_terms_of_service_in_progress)
465 .on_click(move |_, window, cx| (accept_terms_callback)(window, cx)),
466 );
467
468 if thread_empty_state {
469 h_flex()
470 .w_full()
471 .flex_wrap()
472 .justify_between()
473 .child(
474 h_flex()
475 .child(
476 Label::new("To start using Zed AI, please read and accept the")
477 .size(LabelSize::Small),
478 )
479 .child(terms_button),
480 )
481 .child(button_container)
482 } else {
483 v_flex()
484 .w_full()
485 .gap_2()
486 .child(
487 h_flex()
488 .flex_wrap()
489 .when(thread_fresh_start, |this| this.justify_center())
490 .child(Label::new(
491 "To start using Zed AI, please read and accept the",
492 ))
493 .child(terms_button),
494 )
495 .child({
496 match view_kind {
497 LanguageModelProviderTosView::TextThreadPopup => {
498 button_container.w_full().justify_end()
499 }
500 LanguageModelProviderTosView::Configuration => {
501 button_container.w_full().justify_start()
502 }
503 LanguageModelProviderTosView::ThreadFreshStart => {
504 button_container.w_full().justify_center()
505 }
506 LanguageModelProviderTosView::ThreadEmptyState => div().w_0(),
507 }
508 })
509 }
510}
511
512pub struct CloudLanguageModel {
513 id: LanguageModelId,
514 model: Arc<cloud_llm_client::LanguageModel>,
515 llm_api_token: LlmApiToken,
516 client: Arc<Client>,
517 request_limiter: RateLimiter,
518}
519
520struct PerformLlmCompletionResponse {
521 response: Response<AsyncBody>,
522 usage: Option<ModelRequestUsage>,
523 tool_use_limit_reached: bool,
524 includes_status_messages: bool,
525}
526
527impl CloudLanguageModel {
528 async fn perform_llm_completion(
529 client: Arc<Client>,
530 llm_api_token: LlmApiToken,
531 app_version: Option<SemanticVersion>,
532 body: CompletionBody,
533 ) -> Result<PerformLlmCompletionResponse> {
534 let http_client = &client.http_client();
535
536 let mut token = llm_api_token.acquire(&client).await?;
537 let mut refreshed_token = false;
538
539 loop {
540 let request_builder = http_client::Request::builder()
541 .method(Method::POST)
542 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
543 let request_builder = if let Some(app_version) = app_version {
544 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
545 } else {
546 request_builder
547 };
548
549 let request = request_builder
550 .header("Content-Type", "application/json")
551 .header("Authorization", format!("Bearer {token}"))
552 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
553 .body(serde_json::to_string(&body)?.into())?;
554 let mut response = http_client.send(request).await?;
555 let status = response.status();
556 if status.is_success() {
557 let includes_status_messages = response
558 .headers()
559 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
560 .is_some();
561
562 let tool_use_limit_reached = response
563 .headers()
564 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
565 .is_some();
566
567 let usage = if includes_status_messages {
568 None
569 } else {
570 ModelRequestUsage::from_headers(response.headers()).ok()
571 };
572
573 return Ok(PerformLlmCompletionResponse {
574 response,
575 usage,
576 includes_status_messages,
577 tool_use_limit_reached,
578 });
579 }
580
581 if !refreshed_token
582 && response
583 .headers()
584 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
585 .is_some()
586 {
587 token = llm_api_token.refresh(&client).await?;
588 refreshed_token = true;
589 continue;
590 }
591
592 if status == StatusCode::FORBIDDEN
593 && response
594 .headers()
595 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
596 .is_some()
597 {
598 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
599 .headers()
600 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
601 .and_then(|resource| resource.to_str().ok())
602 {
603 if let Some(plan) = response
604 .headers()
605 .get(CURRENT_PLAN_HEADER_NAME)
606 .and_then(|plan| plan.to_str().ok())
607 .and_then(|plan| cloud_llm_client::Plan::from_str(plan).ok())
608 {
609 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
610 }
611 }
612 } else if status == StatusCode::PAYMENT_REQUIRED {
613 return Err(anyhow!(PaymentRequiredError));
614 }
615
616 let mut body = String::new();
617 let headers = response.headers().clone();
618 response.body_mut().read_to_string(&mut body).await?;
619 return Err(anyhow!(ApiError {
620 status,
621 body,
622 headers
623 }));
624 }
625 }
626}
627
628#[derive(Debug, Error)]
629#[error("cloud language model request failed with status {status}: {body}")]
630struct ApiError {
631 status: StatusCode,
632 body: String,
633 headers: HeaderMap<HeaderValue>,
634}
635
636/// Represents error responses from Zed's cloud API.
637///
638/// Example JSON for an upstream HTTP error:
639/// ```json
640/// {
641/// "code": "upstream_http_error",
642/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
643/// "upstream_status": 503
644/// }
645/// ```
646#[derive(Debug, serde::Deserialize)]
647struct CloudApiError {
648 code: String,
649 message: String,
650 #[serde(default)]
651 #[serde(deserialize_with = "deserialize_optional_status_code")]
652 upstream_status: Option<StatusCode>,
653 #[serde(default)]
654 retry_after: Option<f64>,
655}
656
657fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
658where
659 D: serde::Deserializer<'de>,
660{
661 let opt: Option<u16> = Option::deserialize(deserializer)?;
662 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
663}
664
665impl From<ApiError> for LanguageModelCompletionError {
666 fn from(error: ApiError) -> Self {
667 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
668 if cloud_error.code.starts_with("upstream_http_") {
669 let status = if let Some(status) = cloud_error.upstream_status {
670 status
671 } else if cloud_error.code.ends_with("_error") {
672 error.status
673 } else {
674 // If there's a status code in the code string (e.g. "upstream_http_429")
675 // then use that; otherwise, see if the JSON contains a status code.
676 cloud_error
677 .code
678 .strip_prefix("upstream_http_")
679 .and_then(|code_str| code_str.parse::<u16>().ok())
680 .and_then(|code| StatusCode::from_u16(code).ok())
681 .unwrap_or(error.status)
682 };
683
684 return LanguageModelCompletionError::UpstreamProviderError {
685 message: cloud_error.message,
686 status,
687 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
688 };
689 }
690 }
691
692 let retry_after = None;
693 LanguageModelCompletionError::from_http_status(
694 PROVIDER_NAME,
695 error.status,
696 error.body,
697 retry_after,
698 )
699 }
700}
701
702impl LanguageModel for CloudLanguageModel {
703 fn id(&self) -> LanguageModelId {
704 self.id.clone()
705 }
706
707 fn name(&self) -> LanguageModelName {
708 LanguageModelName::from(self.model.display_name.clone())
709 }
710
711 fn provider_id(&self) -> LanguageModelProviderId {
712 PROVIDER_ID
713 }
714
715 fn provider_name(&self) -> LanguageModelProviderName {
716 PROVIDER_NAME
717 }
718
719 fn upstream_provider_id(&self) -> LanguageModelProviderId {
720 use cloud_llm_client::LanguageModelProvider::*;
721 match self.model.provider {
722 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
723 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
724 Google => language_model::GOOGLE_PROVIDER_ID,
725 }
726 }
727
728 fn upstream_provider_name(&self) -> LanguageModelProviderName {
729 use cloud_llm_client::LanguageModelProvider::*;
730 match self.model.provider {
731 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
732 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
733 Google => language_model::GOOGLE_PROVIDER_NAME,
734 }
735 }
736
737 fn supports_tools(&self) -> bool {
738 self.model.supports_tools
739 }
740
741 fn supports_images(&self) -> bool {
742 self.model.supports_images
743 }
744
745 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
746 match choice {
747 LanguageModelToolChoice::Auto
748 | LanguageModelToolChoice::Any
749 | LanguageModelToolChoice::None => true,
750 }
751 }
752
753 fn supports_burn_mode(&self) -> bool {
754 self.model.supports_max_mode
755 }
756
757 fn telemetry_id(&self) -> String {
758 format!("zed.dev/{}", self.model.id)
759 }
760
761 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
762 match self.model.provider {
763 cloud_llm_client::LanguageModelProvider::Anthropic
764 | cloud_llm_client::LanguageModelProvider::OpenAi => {
765 LanguageModelToolSchemaFormat::JsonSchema
766 }
767 cloud_llm_client::LanguageModelProvider::Google => {
768 LanguageModelToolSchemaFormat::JsonSchemaSubset
769 }
770 }
771 }
772
773 fn max_token_count(&self) -> u64 {
774 self.model.max_token_count as u64
775 }
776
777 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
778 self.model
779 .max_token_count_in_max_mode
780 .filter(|_| self.model.supports_max_mode)
781 .map(|max_token_count| max_token_count as u64)
782 }
783
784 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
785 match &self.model.provider {
786 cloud_llm_client::LanguageModelProvider::Anthropic => {
787 Some(LanguageModelCacheConfiguration {
788 min_total_token: 2_048,
789 should_speculate: true,
790 max_cache_anchors: 4,
791 })
792 }
793 cloud_llm_client::LanguageModelProvider::OpenAi
794 | cloud_llm_client::LanguageModelProvider::Google => None,
795 }
796 }
797
798 fn count_tokens(
799 &self,
800 request: LanguageModelRequest,
801 cx: &App,
802 ) -> BoxFuture<'static, Result<u64>> {
803 match self.model.provider {
804 cloud_llm_client::LanguageModelProvider::Anthropic => {
805 count_anthropic_tokens(request, cx)
806 }
807 cloud_llm_client::LanguageModelProvider::OpenAi => {
808 let model = match open_ai::Model::from_id(&self.model.id.0) {
809 Ok(model) => model,
810 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
811 };
812 count_open_ai_tokens(request, model, cx)
813 }
814 cloud_llm_client::LanguageModelProvider::Google => {
815 let client = self.client.clone();
816 let llm_api_token = self.llm_api_token.clone();
817 let model_id = self.model.id.to_string();
818 let generate_content_request =
819 into_google(request, model_id.clone(), GoogleModelMode::Default);
820 async move {
821 let http_client = &client.http_client();
822 let token = llm_api_token.acquire(&client).await?;
823
824 let request_body = CountTokensBody {
825 provider: cloud_llm_client::LanguageModelProvider::Google,
826 model: model_id,
827 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
828 generate_content_request,
829 })?,
830 };
831 let request = http_client::Request::builder()
832 .method(Method::POST)
833 .uri(
834 http_client
835 .build_zed_llm_url("/count_tokens", &[])?
836 .as_ref(),
837 )
838 .header("Content-Type", "application/json")
839 .header("Authorization", format!("Bearer {token}"))
840 .body(serde_json::to_string(&request_body)?.into())?;
841 let mut response = http_client.send(request).await?;
842 let status = response.status();
843 let headers = response.headers().clone();
844 let mut response_body = String::new();
845 response
846 .body_mut()
847 .read_to_string(&mut response_body)
848 .await?;
849
850 if status.is_success() {
851 let response_body: CountTokensResponse =
852 serde_json::from_str(&response_body)?;
853
854 Ok(response_body.tokens as u64)
855 } else {
856 Err(anyhow!(ApiError {
857 status,
858 body: response_body,
859 headers
860 }))
861 }
862 }
863 .boxed()
864 }
865 }
866 }
867
868 fn stream_completion(
869 &self,
870 request: LanguageModelRequest,
871 cx: &AsyncApp,
872 ) -> BoxFuture<
873 'static,
874 Result<
875 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
876 LanguageModelCompletionError,
877 >,
878 > {
879 let thread_id = request.thread_id.clone();
880 let prompt_id = request.prompt_id.clone();
881 let intent = request.intent;
882 let mode = request.mode;
883 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
884 let thinking_allowed = request.thinking_allowed;
885 match self.model.provider {
886 cloud_llm_client::LanguageModelProvider::Anthropic => {
887 let request = into_anthropic(
888 request,
889 self.model.id.to_string(),
890 1.0,
891 self.model.max_output_tokens as u64,
892 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
893 AnthropicModelMode::Thinking {
894 budget_tokens: Some(4_096),
895 }
896 } else {
897 AnthropicModelMode::Default
898 },
899 );
900 let client = self.client.clone();
901 let llm_api_token = self.llm_api_token.clone();
902 let future = self.request_limiter.stream(async move {
903 let PerformLlmCompletionResponse {
904 response,
905 usage,
906 includes_status_messages,
907 tool_use_limit_reached,
908 } = Self::perform_llm_completion(
909 client.clone(),
910 llm_api_token,
911 app_version,
912 CompletionBody {
913 thread_id,
914 prompt_id,
915 intent,
916 mode,
917 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
918 model: request.model.clone(),
919 provider_request: serde_json::to_value(&request)
920 .map_err(|e| anyhow!(e))?,
921 },
922 )
923 .await
924 .map_err(|err| match err.downcast::<ApiError>() {
925 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
926 Err(err) => anyhow!(err),
927 })?;
928
929 let mut mapper = AnthropicEventMapper::new();
930 Ok(map_cloud_completion_events(
931 Box::pin(
932 response_lines(response, includes_status_messages)
933 .chain(usage_updated_event(usage))
934 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
935 ),
936 move |event| mapper.map_event(event),
937 ))
938 });
939 async move { Ok(future.await?.boxed()) }.boxed()
940 }
941 cloud_llm_client::LanguageModelProvider::OpenAi => {
942 let client = self.client.clone();
943 let model = match open_ai::Model::from_id(&self.model.id.0) {
944 Ok(model) => model,
945 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
946 };
947 let request = into_open_ai(
948 request,
949 model.id(),
950 model.supports_parallel_tool_calls(),
951 None,
952 );
953 let llm_api_token = self.llm_api_token.clone();
954 let future = self.request_limiter.stream(async move {
955 let PerformLlmCompletionResponse {
956 response,
957 usage,
958 includes_status_messages,
959 tool_use_limit_reached,
960 } = Self::perform_llm_completion(
961 client.clone(),
962 llm_api_token,
963 app_version,
964 CompletionBody {
965 thread_id,
966 prompt_id,
967 intent,
968 mode,
969 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
970 model: request.model.clone(),
971 provider_request: serde_json::to_value(&request)
972 .map_err(|e| anyhow!(e))?,
973 },
974 )
975 .await?;
976
977 let mut mapper = OpenAiEventMapper::new();
978 Ok(map_cloud_completion_events(
979 Box::pin(
980 response_lines(response, includes_status_messages)
981 .chain(usage_updated_event(usage))
982 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
983 ),
984 move |event| mapper.map_event(event),
985 ))
986 });
987 async move { Ok(future.await?.boxed()) }.boxed()
988 }
989 cloud_llm_client::LanguageModelProvider::Google => {
990 let client = self.client.clone();
991 let request =
992 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
993 let llm_api_token = self.llm_api_token.clone();
994 let future = self.request_limiter.stream(async move {
995 let PerformLlmCompletionResponse {
996 response,
997 usage,
998 includes_status_messages,
999 tool_use_limit_reached,
1000 } = Self::perform_llm_completion(
1001 client.clone(),
1002 llm_api_token,
1003 app_version,
1004 CompletionBody {
1005 thread_id,
1006 prompt_id,
1007 intent,
1008 mode,
1009 provider: cloud_llm_client::LanguageModelProvider::Google,
1010 model: request.model.model_id.clone(),
1011 provider_request: serde_json::to_value(&request)
1012 .map_err(|e| anyhow!(e))?,
1013 },
1014 )
1015 .await?;
1016
1017 let mut mapper = GoogleEventMapper::new();
1018 Ok(map_cloud_completion_events(
1019 Box::pin(
1020 response_lines(response, includes_status_messages)
1021 .chain(usage_updated_event(usage))
1022 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
1023 ),
1024 move |event| mapper.map_event(event),
1025 ))
1026 });
1027 async move { Ok(future.await?.boxed()) }.boxed()
1028 }
1029 }
1030 }
1031}
1032
1033fn map_cloud_completion_events<T, F>(
1034 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
1035 mut map_callback: F,
1036) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1037where
1038 T: DeserializeOwned + 'static,
1039 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1040 + Send
1041 + 'static,
1042{
1043 stream
1044 .flat_map(move |event| {
1045 futures::stream::iter(match event {
1046 Err(error) => {
1047 vec![Err(LanguageModelCompletionError::from(error))]
1048 }
1049 Ok(CompletionEvent::Status(event)) => {
1050 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
1051 }
1052 Ok(CompletionEvent::Event(event)) => map_callback(event),
1053 })
1054 })
1055 .boxed()
1056}
1057
1058fn usage_updated_event<T>(
1059 usage: Option<ModelRequestUsage>,
1060) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1061 futures::stream::iter(usage.map(|usage| {
1062 Ok(CompletionEvent::Status(
1063 CompletionRequestStatus::UsageUpdated {
1064 amount: usage.amount as usize,
1065 limit: usage.limit,
1066 },
1067 ))
1068 }))
1069}
1070
1071fn tool_use_limit_reached_event<T>(
1072 tool_use_limit_reached: bool,
1073) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1074 futures::stream::iter(tool_use_limit_reached.then(|| {
1075 Ok(CompletionEvent::Status(
1076 CompletionRequestStatus::ToolUseLimitReached,
1077 ))
1078 }))
1079}
1080
1081fn response_lines<T: DeserializeOwned>(
1082 response: Response<AsyncBody>,
1083 includes_status_messages: bool,
1084) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1085 futures::stream::try_unfold(
1086 (String::new(), BufReader::new(response.into_body())),
1087 move |(mut line, mut body)| async move {
1088 match body.read_line(&mut line).await {
1089 Ok(0) => Ok(None),
1090 Ok(_) => {
1091 let event = if includes_status_messages {
1092 serde_json::from_str::<CompletionEvent<T>>(&line)?
1093 } else {
1094 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1095 };
1096
1097 line.clear();
1098 Ok(Some((event, (line, body))))
1099 }
1100 Err(e) => Err(e.into()),
1101 }
1102 },
1103 )
1104}
1105
1106#[derive(IntoElement, RegisterComponent)]
1107struct ZedAiConfiguration {
1108 is_connected: bool,
1109 plan: Option<Plan>,
1110 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1111 eligible_for_trial: bool,
1112 has_accepted_terms_of_service: bool,
1113 account_too_young: bool,
1114 accept_terms_of_service_in_progress: bool,
1115 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1116 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1117}
1118
1119impl RenderOnce for ZedAiConfiguration {
1120 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1121 let young_account_banner = YoungAccountBanner;
1122
1123 let is_pro = self.plan == Some(Plan::ZedPro);
1124 let subscription_text = match (self.plan, self.subscription_period) {
1125 (Some(Plan::ZedPro), Some(_)) => {
1126 "You have access to Zed's hosted models through your Pro subscription."
1127 }
1128 (Some(Plan::ZedProTrial), Some(_)) => {
1129 "You have access to Zed's hosted models through your Pro trial."
1130 }
1131 (Some(Plan::ZedFree), Some(_)) => {
1132 "You have basic access to Zed's hosted models through the Free plan."
1133 }
1134 _ => {
1135 if self.eligible_for_trial {
1136 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1137 } else {
1138 "Subscribe for access to Zed's hosted models."
1139 }
1140 }
1141 };
1142
1143 let manage_subscription_buttons = if is_pro {
1144 Button::new("manage_settings", "Manage Subscription")
1145 .full_width()
1146 .style(ButtonStyle::Tinted(TintColor::Accent))
1147 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1148 .into_any_element()
1149 } else if self.plan.is_none() || self.eligible_for_trial {
1150 Button::new("start_trial", "Start 14-day Free Pro Trial")
1151 .full_width()
1152 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1153 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1154 .into_any_element()
1155 } else {
1156 Button::new("upgrade", "Upgrade to Pro")
1157 .full_width()
1158 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1159 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1160 .into_any_element()
1161 };
1162
1163 if !self.is_connected {
1164 return v_flex()
1165 .gap_2()
1166 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1167 .child(
1168 Button::new("sign_in", "Sign In to use Zed AI")
1169 .icon_color(Color::Muted)
1170 .icon(IconName::Github)
1171 .icon_size(IconSize::Small)
1172 .icon_position(IconPosition::Start)
1173 .full_width()
1174 .on_click({
1175 let callback = self.sign_in_callback.clone();
1176 move |_, window, cx| (callback)(window, cx)
1177 }),
1178 );
1179 }
1180
1181 v_flex()
1182 .gap_2()
1183 .w_full()
1184 .when(!self.has_accepted_terms_of_service, |this| {
1185 this.child(render_accept_terms(
1186 LanguageModelProviderTosView::Configuration,
1187 self.accept_terms_of_service_in_progress,
1188 {
1189 let callback = self.accept_terms_of_service_callback.clone();
1190 move |window, cx| (callback)(window, cx)
1191 },
1192 ))
1193 })
1194 .map(|this| {
1195 if self.has_accepted_terms_of_service && self.account_too_young {
1196 this.child(young_account_banner).child(
1197 Button::new("upgrade", "Upgrade to Pro")
1198 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1199 .full_width()
1200 .on_click(|_, _, cx| {
1201 cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))
1202 }),
1203 )
1204 } else if self.has_accepted_terms_of_service {
1205 this.text_sm()
1206 .child(subscription_text)
1207 .child(manage_subscription_buttons)
1208 } else {
1209 this
1210 }
1211 })
1212 .when(self.has_accepted_terms_of_service, |this| this)
1213 }
1214}
1215
1216struct ConfigurationView {
1217 state: Entity<State>,
1218 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1219 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1220}
1221
1222impl ConfigurationView {
1223 fn new(state: Entity<State>) -> Self {
1224 let accept_terms_of_service_callback = Arc::new({
1225 let state = state.clone();
1226 move |_window: &mut Window, cx: &mut App| {
1227 state.update(cx, |state, cx| {
1228 state.accept_terms_of_service(cx);
1229 });
1230 }
1231 });
1232
1233 let sign_in_callback = Arc::new({
1234 let state = state.clone();
1235 move |_window: &mut Window, cx: &mut App| {
1236 state.update(cx, |state, cx| {
1237 state.authenticate(cx).detach_and_log_err(cx);
1238 });
1239 }
1240 });
1241
1242 Self {
1243 state,
1244 accept_terms_of_service_callback,
1245 sign_in_callback,
1246 }
1247 }
1248}
1249
1250impl Render for ConfigurationView {
1251 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1252 let state = self.state.read(cx);
1253 let user_store = state.user_store.read(cx);
1254
1255 ZedAiConfiguration {
1256 is_connected: !state.is_signed_out(cx),
1257 plan: user_store.plan(),
1258 subscription_period: user_store.subscription_period(),
1259 eligible_for_trial: user_store.trial_started_at().is_none(),
1260 has_accepted_terms_of_service: state.has_accepted_terms_of_service(cx),
1261 account_too_young: user_store.account_too_young(),
1262 accept_terms_of_service_in_progress: state.accept_terms_of_service_task.is_some(),
1263 accept_terms_of_service_callback: self.accept_terms_of_service_callback.clone(),
1264 sign_in_callback: self.sign_in_callback.clone(),
1265 }
1266 }
1267}
1268
1269impl Component for ZedAiConfiguration {
1270 fn name() -> &'static str {
1271 "AI Configuration Content"
1272 }
1273
1274 fn sort_name() -> &'static str {
1275 "AI Configuration Content"
1276 }
1277
1278 fn scope() -> ComponentScope {
1279 ComponentScope::Onboarding
1280 }
1281
1282 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1283 fn configuration(
1284 is_connected: bool,
1285 plan: Option<Plan>,
1286 eligible_for_trial: bool,
1287 account_too_young: bool,
1288 has_accepted_terms_of_service: bool,
1289 ) -> AnyElement {
1290 ZedAiConfiguration {
1291 is_connected,
1292 plan,
1293 subscription_period: plan
1294 .is_some()
1295 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1296 eligible_for_trial,
1297 has_accepted_terms_of_service,
1298 account_too_young,
1299 accept_terms_of_service_in_progress: false,
1300 accept_terms_of_service_callback: Arc::new(|_, _| {}),
1301 sign_in_callback: Arc::new(|_, _| {}),
1302 }
1303 .into_any_element()
1304 }
1305
1306 Some(
1307 v_flex()
1308 .p_4()
1309 .gap_4()
1310 .children(vec![
1311 single_example(
1312 "Not connected",
1313 configuration(false, None, false, false, true),
1314 ),
1315 single_example(
1316 "Accept Terms of Service",
1317 configuration(true, None, true, false, false),
1318 ),
1319 single_example(
1320 "No Plan - Not eligible for trial",
1321 configuration(true, None, false, false, true),
1322 ),
1323 single_example(
1324 "No Plan - Eligible for trial",
1325 configuration(true, None, true, false, true),
1326 ),
1327 single_example(
1328 "Free Plan",
1329 configuration(true, Some(Plan::ZedFree), true, false, true),
1330 ),
1331 single_example(
1332 "Zed Pro Trial Plan",
1333 configuration(true, Some(Plan::ZedProTrial), true, false, true),
1334 ),
1335 single_example(
1336 "Zed Pro Plan",
1337 configuration(true, Some(Plan::ZedPro), true, false, true),
1338 ),
1339 ])
1340 .into_any_element(),
1341 )
1342 }
1343}
1344
1345#[cfg(test)]
1346mod tests {
1347 use super::*;
1348 use http_client::http::{HeaderMap, StatusCode};
1349 use language_model::LanguageModelCompletionError;
1350
1351 #[test]
1352 fn test_api_error_conversion_with_upstream_http_error() {
1353 // upstream_http_error with 503 status should become ServerOverloaded
1354 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}"#;
1355
1356 let api_error = ApiError {
1357 status: StatusCode::INTERNAL_SERVER_ERROR,
1358 body: error_body.to_string(),
1359 headers: HeaderMap::new(),
1360 };
1361
1362 let completion_error: LanguageModelCompletionError = api_error.into();
1363
1364 match completion_error {
1365 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1366 assert_eq!(
1367 message,
1368 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1369 );
1370 }
1371 _ => panic!(
1372 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1373 completion_error
1374 ),
1375 }
1376
1377 // upstream_http_error with 500 status should become ApiInternalServerError
1378 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1379
1380 let api_error = ApiError {
1381 status: StatusCode::INTERNAL_SERVER_ERROR,
1382 body: error_body.to_string(),
1383 headers: HeaderMap::new(),
1384 };
1385
1386 let completion_error: LanguageModelCompletionError = api_error.into();
1387
1388 match completion_error {
1389 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1390 assert_eq!(
1391 message,
1392 "Received an error from the OpenAI API: internal server error"
1393 );
1394 }
1395 _ => panic!(
1396 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1397 completion_error
1398 ),
1399 }
1400
1401 // upstream_http_error with 429 status should become RateLimitExceeded
1402 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1403
1404 let api_error = ApiError {
1405 status: StatusCode::INTERNAL_SERVER_ERROR,
1406 body: error_body.to_string(),
1407 headers: HeaderMap::new(),
1408 };
1409
1410 let completion_error: LanguageModelCompletionError = api_error.into();
1411
1412 match completion_error {
1413 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1414 assert_eq!(
1415 message,
1416 "Received an error from the Google API: rate limit exceeded"
1417 );
1418 }
1419 _ => panic!(
1420 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1421 completion_error
1422 ),
1423 }
1424
1425 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1426 let error_body = "Regular internal server error";
1427
1428 let api_error = ApiError {
1429 status: StatusCode::INTERNAL_SERVER_ERROR,
1430 body: error_body.to_string(),
1431 headers: HeaderMap::new(),
1432 };
1433
1434 let completion_error: LanguageModelCompletionError = api_error.into();
1435
1436 match completion_error {
1437 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1438 assert_eq!(provider, PROVIDER_NAME);
1439 assert_eq!(message, "Regular internal server error");
1440 }
1441 _ => panic!(
1442 "Expected ApiInternalServerError for regular 500, got: {:?}",
1443 completion_error
1444 ),
1445 }
1446
1447 // upstream_http_429 format should be converted to UpstreamProviderError
1448 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1449
1450 let api_error = ApiError {
1451 status: StatusCode::INTERNAL_SERVER_ERROR,
1452 body: error_body.to_string(),
1453 headers: HeaderMap::new(),
1454 };
1455
1456 let completion_error: LanguageModelCompletionError = api_error.into();
1457
1458 match completion_error {
1459 LanguageModelCompletionError::UpstreamProviderError {
1460 message,
1461 status,
1462 retry_after,
1463 } => {
1464 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1465 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1466 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1467 }
1468 _ => panic!(
1469 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1470 completion_error
1471 ),
1472 }
1473
1474 // Invalid JSON in error body should fall back to regular error handling
1475 let error_body = "Not JSON at all";
1476
1477 let api_error = ApiError {
1478 status: StatusCode::INTERNAL_SERVER_ERROR,
1479 body: error_body.to_string(),
1480 headers: HeaderMap::new(),
1481 };
1482
1483 let completion_error: LanguageModelCompletionError = api_error.into();
1484
1485 match completion_error {
1486 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1487 assert_eq!(provider, PROVIDER_NAME);
1488 }
1489 _ => panic!(
1490 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1491 completion_error
1492 ),
1493 }
1494 }
1495}