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,
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 proto::Plan;
31use release_channel::AppVersion;
32use schemars::JsonSchema;
33use serde::{Deserialize, Serialize, de::DeserializeOwned};
34use settings::SettingsStore;
35use smol::io::{AsyncReadExt, BufReader};
36use std::pin::Pin;
37use std::str::FromStr as _;
38use std::sync::Arc;
39use std::time::Duration;
40use thiserror::Error;
41use ui::{TintColor, prelude::*};
42use util::{ResultExt as _, maybe};
43
44use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
45use crate::provider::google::{GoogleEventMapper, into_google};
46use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
47
48const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
49const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
50
51#[derive(Default, Clone, Debug, PartialEq)]
52pub struct ZedDotDevSettings {
53 pub available_models: Vec<AvailableModel>,
54}
55
56#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
57#[serde(rename_all = "lowercase")]
58pub enum AvailableProvider {
59 Anthropic,
60 OpenAi,
61 Google,
62}
63
64#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
65pub struct AvailableModel {
66 /// The provider of the language model.
67 pub provider: AvailableProvider,
68 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
69 pub name: String,
70 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
71 pub display_name: Option<String>,
72 /// The size of the context window, indicating the maximum number of tokens the model can process.
73 pub max_tokens: usize,
74 /// The maximum number of output tokens allowed by the model.
75 pub max_output_tokens: Option<u64>,
76 /// The maximum number of completion tokens allowed by the model (o1-* only)
77 pub max_completion_tokens: Option<u64>,
78 /// Override this model with a different Anthropic model for tool calls.
79 pub tool_override: Option<String>,
80 /// Indicates whether this custom model supports caching.
81 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
82 /// The default temperature to use for this model.
83 pub default_temperature: Option<f32>,
84 /// Any extra beta headers to provide when using the model.
85 #[serde(default)]
86 pub extra_beta_headers: Vec<String>,
87 /// The model's mode (e.g. thinking)
88 pub mode: Option<ModelMode>,
89}
90
91#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
92#[serde(tag = "type", rename_all = "lowercase")]
93pub enum ModelMode {
94 #[default]
95 Default,
96 Thinking {
97 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
98 budget_tokens: Option<u32>,
99 },
100}
101
102impl From<ModelMode> for AnthropicModelMode {
103 fn from(value: ModelMode) -> Self {
104 match value {
105 ModelMode::Default => AnthropicModelMode::Default,
106 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
107 }
108 }
109}
110
111pub struct CloudLanguageModelProvider {
112 client: Arc<Client>,
113 state: gpui::Entity<State>,
114 _maintain_client_status: Task<()>,
115}
116
117pub struct State {
118 client: Arc<Client>,
119 llm_api_token: LlmApiToken,
120 user_store: Entity<UserStore>,
121 status: client::Status,
122 accept_terms_of_service_task: Option<Task<Result<()>>>,
123 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
124 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
125 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
126 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
127 _fetch_models_task: Task<()>,
128 _settings_subscription: Subscription,
129 _llm_token_subscription: Subscription,
130}
131
132impl State {
133 fn new(
134 client: Arc<Client>,
135 user_store: Entity<UserStore>,
136 status: client::Status,
137 cx: &mut Context<Self>,
138 ) -> Self {
139 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
140
141 Self {
142 client: client.clone(),
143 llm_api_token: LlmApiToken::default(),
144 user_store,
145 status,
146 accept_terms_of_service_task: None,
147 models: Vec::new(),
148 default_model: None,
149 default_fast_model: None,
150 recommended_models: Vec::new(),
151 _fetch_models_task: cx.spawn(async move |this, cx| {
152 maybe!(async move {
153 let (client, llm_api_token) = this
154 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
155
156 loop {
157 let status = this.read_with(cx, |this, _cx| this.status)?;
158 if matches!(status, client::Status::Connected { .. }) {
159 break;
160 }
161
162 cx.background_executor()
163 .timer(Duration::from_millis(100))
164 .await;
165 }
166
167 let response = Self::fetch_models(client, llm_api_token).await?;
168 this.update(cx, |this, cx| {
169 this.update_models(response, cx);
170 })
171 })
172 .await
173 .context("failed to fetch Zed models")
174 .log_err();
175 }),
176 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
177 cx.notify();
178 }),
179 _llm_token_subscription: cx.subscribe(
180 &refresh_llm_token_listener,
181 move |this, _listener, _event, cx| {
182 let client = this.client.clone();
183 let llm_api_token = this.llm_api_token.clone();
184 cx.spawn(async move |this, cx| {
185 llm_api_token.refresh(&client).await?;
186 let response = Self::fetch_models(client, llm_api_token).await?;
187 this.update(cx, |this, cx| {
188 this.update_models(response, cx);
189 })
190 })
191 .detach_and_log_err(cx);
192 },
193 ),
194 }
195 }
196
197 fn is_signed_out(&self) -> bool {
198 self.status.is_signed_out()
199 }
200
201 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
202 let client = self.client.clone();
203 cx.spawn(async move |state, cx| {
204 client
205 .authenticate_and_connect(true, &cx)
206 .await
207 .into_response()?;
208 state.update(cx, |_, cx| cx.notify())
209 })
210 }
211
212 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
213 self.user_store
214 .read(cx)
215 .current_user_has_accepted_terms()
216 .unwrap_or(false)
217 }
218
219 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
220 let user_store = self.user_store.clone();
221 self.accept_terms_of_service_task = Some(cx.spawn(async move |this, cx| {
222 let _ = user_store
223 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
224 .await;
225 this.update(cx, |this, cx| {
226 this.accept_terms_of_service_task = None;
227 cx.notify()
228 })
229 }));
230 }
231
232 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
233 let mut models = Vec::new();
234
235 for model in response.models {
236 models.push(Arc::new(model.clone()));
237
238 // Right now we represent thinking variants of models as separate models on the client,
239 // so we need to insert variants for any model that supports thinking.
240 if model.supports_thinking {
241 models.push(Arc::new(cloud_llm_client::LanguageModel {
242 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
243 display_name: format!("{} Thinking", model.display_name),
244 ..model
245 }));
246 }
247 }
248
249 self.default_model = models
250 .iter()
251 .find(|model| model.id == response.default_model)
252 .cloned();
253 self.default_fast_model = models
254 .iter()
255 .find(|model| model.id == response.default_fast_model)
256 .cloned();
257 self.recommended_models = response
258 .recommended_models
259 .iter()
260 .filter_map(|id| models.iter().find(|model| &model.id == id))
261 .cloned()
262 .collect();
263 self.models = models;
264 cx.notify();
265 }
266
267 async fn fetch_models(
268 client: Arc<Client>,
269 llm_api_token: LlmApiToken,
270 ) -> Result<ListModelsResponse> {
271 let http_client = &client.http_client();
272 let token = llm_api_token.acquire(&client).await?;
273
274 let request = http_client::Request::builder()
275 .method(Method::GET)
276 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
277 .header("Authorization", format!("Bearer {token}"))
278 .body(AsyncBody::empty())?;
279 let mut response = http_client
280 .send(request)
281 .await
282 .context("failed to send list models request")?;
283
284 if response.status().is_success() {
285 let mut body = String::new();
286 response.body_mut().read_to_string(&mut body).await?;
287 return Ok(serde_json::from_str(&body)?);
288 } else {
289 let mut body = String::new();
290 response.body_mut().read_to_string(&mut body).await?;
291 anyhow::bail!(
292 "error listing models.\nStatus: {:?}\nBody: {body}",
293 response.status(),
294 );
295 }
296 }
297}
298
299impl CloudLanguageModelProvider {
300 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
301 let mut status_rx = client.status();
302 let status = *status_rx.borrow();
303
304 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
305
306 let state_ref = state.downgrade();
307 let maintain_client_status = cx.spawn(async move |cx| {
308 while let Some(status) = status_rx.next().await {
309 if let Some(this) = state_ref.upgrade() {
310 _ = this.update(cx, |this, cx| {
311 if this.status != status {
312 this.status = status;
313 cx.notify();
314 }
315 });
316 } else {
317 break;
318 }
319 }
320 });
321
322 Self {
323 client,
324 state: state.clone(),
325 _maintain_client_status: maintain_client_status,
326 }
327 }
328
329 fn create_language_model(
330 &self,
331 model: Arc<cloud_llm_client::LanguageModel>,
332 llm_api_token: LlmApiToken,
333 ) -> Arc<dyn LanguageModel> {
334 Arc::new(CloudLanguageModel {
335 id: LanguageModelId(SharedString::from(model.id.0.clone())),
336 model,
337 llm_api_token: llm_api_token.clone(),
338 client: self.client.clone(),
339 request_limiter: RateLimiter::new(4),
340 })
341 }
342}
343
344impl LanguageModelProviderState for CloudLanguageModelProvider {
345 type ObservableEntity = State;
346
347 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
348 Some(self.state.clone())
349 }
350}
351
352impl LanguageModelProvider for CloudLanguageModelProvider {
353 fn id(&self) -> LanguageModelProviderId {
354 PROVIDER_ID
355 }
356
357 fn name(&self) -> LanguageModelProviderName {
358 PROVIDER_NAME
359 }
360
361 fn icon(&self) -> IconName {
362 IconName::AiZed
363 }
364
365 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
366 let default_model = self.state.read(cx).default_model.clone()?;
367 let llm_api_token = self.state.read(cx).llm_api_token.clone();
368 Some(self.create_language_model(default_model, llm_api_token))
369 }
370
371 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
372 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
373 let llm_api_token = self.state.read(cx).llm_api_token.clone();
374 Some(self.create_language_model(default_fast_model, llm_api_token))
375 }
376
377 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
378 let llm_api_token = self.state.read(cx).llm_api_token.clone();
379 self.state
380 .read(cx)
381 .recommended_models
382 .iter()
383 .cloned()
384 .map(|model| self.create_language_model(model, llm_api_token.clone()))
385 .collect()
386 }
387
388 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
389 let llm_api_token = self.state.read(cx).llm_api_token.clone();
390 self.state
391 .read(cx)
392 .models
393 .iter()
394 .cloned()
395 .map(|model| self.create_language_model(model, llm_api_token.clone()))
396 .collect()
397 }
398
399 fn is_authenticated(&self, cx: &App) -> bool {
400 let state = self.state.read(cx);
401 !state.is_signed_out() && state.has_accepted_terms_of_service(cx)
402 }
403
404 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
405 Task::ready(Ok(()))
406 }
407
408 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
409 cx.new(|_| ConfigurationView::new(self.state.clone()))
410 .into()
411 }
412
413 fn must_accept_terms(&self, cx: &App) -> bool {
414 !self.state.read(cx).has_accepted_terms_of_service(cx)
415 }
416
417 fn render_accept_terms(
418 &self,
419 view: LanguageModelProviderTosView,
420 cx: &mut App,
421 ) -> Option<AnyElement> {
422 let state = self.state.read(cx);
423 if state.has_accepted_terms_of_service(cx) {
424 return None;
425 }
426 Some(
427 render_accept_terms(view, state.accept_terms_of_service_task.is_some(), {
428 let state = self.state.clone();
429 move |_window, cx| {
430 state.update(cx, |state, cx| state.accept_terms_of_service(cx));
431 }
432 })
433 .into_any_element(),
434 )
435 }
436
437 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
438 Task::ready(Ok(()))
439 }
440}
441
442fn render_accept_terms(
443 view_kind: LanguageModelProviderTosView,
444 accept_terms_of_service_in_progress: bool,
445 accept_terms_callback: impl Fn(&mut Window, &mut App) + 'static,
446) -> impl IntoElement {
447 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
448 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadEmptyState);
449
450 let terms_button = Button::new("terms_of_service", "Terms of Service")
451 .style(ButtonStyle::Subtle)
452 .icon(IconName::ArrowUpRight)
453 .icon_color(Color::Muted)
454 .icon_size(IconSize::XSmall)
455 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
456 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
457
458 let button_container = h_flex().child(
459 Button::new("accept_terms", "I accept the Terms of Service")
460 .when(!thread_empty_state, |this| {
461 this.full_width()
462 .style(ButtonStyle::Tinted(TintColor::Accent))
463 .icon(IconName::Check)
464 .icon_position(IconPosition::Start)
465 .icon_size(IconSize::Small)
466 })
467 .when(thread_empty_state, |this| {
468 this.style(ButtonStyle::Tinted(TintColor::Warning))
469 .label_size(LabelSize::Small)
470 })
471 .disabled(accept_terms_of_service_in_progress)
472 .on_click(move |_, window, cx| (accept_terms_callback)(window, cx)),
473 );
474
475 if thread_empty_state {
476 h_flex()
477 .w_full()
478 .flex_wrap()
479 .justify_between()
480 .child(
481 h_flex()
482 .child(
483 Label::new("To start using Zed AI, please read and accept the")
484 .size(LabelSize::Small),
485 )
486 .child(terms_button),
487 )
488 .child(button_container)
489 } else {
490 v_flex()
491 .w_full()
492 .gap_2()
493 .child(
494 h_flex()
495 .flex_wrap()
496 .when(thread_fresh_start, |this| this.justify_center())
497 .child(Label::new(
498 "To start using Zed AI, please read and accept the",
499 ))
500 .child(terms_button),
501 )
502 .child({
503 match view_kind {
504 LanguageModelProviderTosView::TextThreadPopup => {
505 button_container.w_full().justify_end()
506 }
507 LanguageModelProviderTosView::Configuration => {
508 button_container.w_full().justify_start()
509 }
510 LanguageModelProviderTosView::ThreadFreshStart => {
511 button_container.w_full().justify_center()
512 }
513 LanguageModelProviderTosView::ThreadEmptyState => div().w_0(),
514 }
515 })
516 }
517}
518
519pub struct CloudLanguageModel {
520 id: LanguageModelId,
521 model: Arc<cloud_llm_client::LanguageModel>,
522 llm_api_token: LlmApiToken,
523 client: Arc<Client>,
524 request_limiter: RateLimiter,
525}
526
527struct PerformLlmCompletionResponse {
528 response: Response<AsyncBody>,
529 usage: Option<ModelRequestUsage>,
530 tool_use_limit_reached: bool,
531 includes_status_messages: bool,
532}
533
534impl CloudLanguageModel {
535 async fn perform_llm_completion(
536 client: Arc<Client>,
537 llm_api_token: LlmApiToken,
538 app_version: Option<SemanticVersion>,
539 body: CompletionBody,
540 ) -> Result<PerformLlmCompletionResponse> {
541 let http_client = &client.http_client();
542
543 let mut token = llm_api_token.acquire(&client).await?;
544 let mut refreshed_token = false;
545
546 loop {
547 let request_builder = http_client::Request::builder()
548 .method(Method::POST)
549 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
550 let request_builder = if let Some(app_version) = app_version {
551 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
552 } else {
553 request_builder
554 };
555
556 let request = request_builder
557 .header("Content-Type", "application/json")
558 .header("Authorization", format!("Bearer {token}"))
559 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
560 .body(serde_json::to_string(&body)?.into())?;
561 let mut response = http_client.send(request).await?;
562 let status = response.status();
563 if status.is_success() {
564 let includes_status_messages = response
565 .headers()
566 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
567 .is_some();
568
569 let tool_use_limit_reached = response
570 .headers()
571 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
572 .is_some();
573
574 let usage = if includes_status_messages {
575 None
576 } else {
577 ModelRequestUsage::from_headers(response.headers()).ok()
578 };
579
580 return Ok(PerformLlmCompletionResponse {
581 response,
582 usage,
583 includes_status_messages,
584 tool_use_limit_reached,
585 });
586 }
587
588 if !refreshed_token
589 && response
590 .headers()
591 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
592 .is_some()
593 {
594 token = llm_api_token.refresh(&client).await?;
595 refreshed_token = true;
596 continue;
597 }
598
599 if status == StatusCode::FORBIDDEN
600 && response
601 .headers()
602 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
603 .is_some()
604 {
605 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
606 .headers()
607 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
608 .and_then(|resource| resource.to_str().ok())
609 {
610 if let Some(plan) = response
611 .headers()
612 .get(CURRENT_PLAN_HEADER_NAME)
613 .and_then(|plan| plan.to_str().ok())
614 .and_then(|plan| cloud_llm_client::Plan::from_str(plan).ok())
615 {
616 let plan = match plan {
617 cloud_llm_client::Plan::ZedFree => Plan::Free,
618 cloud_llm_client::Plan::ZedPro => Plan::ZedPro,
619 cloud_llm_client::Plan::ZedProTrial => Plan::ZedProTrial,
620 };
621 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
622 }
623 }
624 } else if status == StatusCode::PAYMENT_REQUIRED {
625 return Err(anyhow!(PaymentRequiredError));
626 }
627
628 let mut body = String::new();
629 let headers = response.headers().clone();
630 response.body_mut().read_to_string(&mut body).await?;
631 return Err(anyhow!(ApiError {
632 status,
633 body,
634 headers
635 }));
636 }
637 }
638}
639
640#[derive(Debug, Error)]
641#[error("cloud language model request failed with status {status}: {body}")]
642struct ApiError {
643 status: StatusCode,
644 body: String,
645 headers: HeaderMap<HeaderValue>,
646}
647
648/// Represents error responses from Zed's cloud API.
649///
650/// Example JSON for an upstream HTTP error:
651/// ```json
652/// {
653/// "code": "upstream_http_error",
654/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
655/// "upstream_status": 503
656/// }
657/// ```
658#[derive(Debug, serde::Deserialize)]
659struct CloudApiError {
660 code: String,
661 message: String,
662 #[serde(default)]
663 #[serde(deserialize_with = "deserialize_optional_status_code")]
664 upstream_status: Option<StatusCode>,
665 #[serde(default)]
666 retry_after: Option<f64>,
667}
668
669fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
670where
671 D: serde::Deserializer<'de>,
672{
673 let opt: Option<u16> = Option::deserialize(deserializer)?;
674 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
675}
676
677impl From<ApiError> for LanguageModelCompletionError {
678 fn from(error: ApiError) -> Self {
679 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
680 if cloud_error.code.starts_with("upstream_http_") {
681 let status = if let Some(status) = cloud_error.upstream_status {
682 status
683 } else if cloud_error.code.ends_with("_error") {
684 error.status
685 } else {
686 // If there's a status code in the code string (e.g. "upstream_http_429")
687 // then use that; otherwise, see if the JSON contains a status code.
688 cloud_error
689 .code
690 .strip_prefix("upstream_http_")
691 .and_then(|code_str| code_str.parse::<u16>().ok())
692 .and_then(|code| StatusCode::from_u16(code).ok())
693 .unwrap_or(error.status)
694 };
695
696 return LanguageModelCompletionError::UpstreamProviderError {
697 message: cloud_error.message,
698 status,
699 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
700 };
701 }
702 }
703
704 let retry_after = None;
705 LanguageModelCompletionError::from_http_status(
706 PROVIDER_NAME,
707 error.status,
708 error.body,
709 retry_after,
710 )
711 }
712}
713
714impl LanguageModel for CloudLanguageModel {
715 fn id(&self) -> LanguageModelId {
716 self.id.clone()
717 }
718
719 fn name(&self) -> LanguageModelName {
720 LanguageModelName::from(self.model.display_name.clone())
721 }
722
723 fn provider_id(&self) -> LanguageModelProviderId {
724 PROVIDER_ID
725 }
726
727 fn provider_name(&self) -> LanguageModelProviderName {
728 PROVIDER_NAME
729 }
730
731 fn upstream_provider_id(&self) -> LanguageModelProviderId {
732 use cloud_llm_client::LanguageModelProvider::*;
733 match self.model.provider {
734 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
735 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
736 Google => language_model::GOOGLE_PROVIDER_ID,
737 }
738 }
739
740 fn upstream_provider_name(&self) -> LanguageModelProviderName {
741 use cloud_llm_client::LanguageModelProvider::*;
742 match self.model.provider {
743 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
744 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
745 Google => language_model::GOOGLE_PROVIDER_NAME,
746 }
747 }
748
749 fn supports_tools(&self) -> bool {
750 self.model.supports_tools
751 }
752
753 fn supports_images(&self) -> bool {
754 self.model.supports_images
755 }
756
757 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
758 match choice {
759 LanguageModelToolChoice::Auto
760 | LanguageModelToolChoice::Any
761 | LanguageModelToolChoice::None => true,
762 }
763 }
764
765 fn supports_burn_mode(&self) -> bool {
766 self.model.supports_max_mode
767 }
768
769 fn telemetry_id(&self) -> String {
770 format!("zed.dev/{}", self.model.id)
771 }
772
773 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
774 match self.model.provider {
775 cloud_llm_client::LanguageModelProvider::Anthropic
776 | cloud_llm_client::LanguageModelProvider::OpenAi => {
777 LanguageModelToolSchemaFormat::JsonSchema
778 }
779 cloud_llm_client::LanguageModelProvider::Google => {
780 LanguageModelToolSchemaFormat::JsonSchemaSubset
781 }
782 }
783 }
784
785 fn max_token_count(&self) -> u64 {
786 self.model.max_token_count as u64
787 }
788
789 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
790 self.model
791 .max_token_count_in_max_mode
792 .filter(|_| self.model.supports_max_mode)
793 .map(|max_token_count| max_token_count as u64)
794 }
795
796 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
797 match &self.model.provider {
798 cloud_llm_client::LanguageModelProvider::Anthropic => {
799 Some(LanguageModelCacheConfiguration {
800 min_total_token: 2_048,
801 should_speculate: true,
802 max_cache_anchors: 4,
803 })
804 }
805 cloud_llm_client::LanguageModelProvider::OpenAi
806 | cloud_llm_client::LanguageModelProvider::Google => None,
807 }
808 }
809
810 fn count_tokens(
811 &self,
812 request: LanguageModelRequest,
813 cx: &App,
814 ) -> BoxFuture<'static, Result<u64>> {
815 match self.model.provider {
816 cloud_llm_client::LanguageModelProvider::Anthropic => {
817 count_anthropic_tokens(request, cx)
818 }
819 cloud_llm_client::LanguageModelProvider::OpenAi => {
820 let model = match open_ai::Model::from_id(&self.model.id.0) {
821 Ok(model) => model,
822 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
823 };
824 count_open_ai_tokens(request, model, cx)
825 }
826 cloud_llm_client::LanguageModelProvider::Google => {
827 let client = self.client.clone();
828 let llm_api_token = self.llm_api_token.clone();
829 let model_id = self.model.id.to_string();
830 let generate_content_request =
831 into_google(request, model_id.clone(), GoogleModelMode::Default);
832 async move {
833 let http_client = &client.http_client();
834 let token = llm_api_token.acquire(&client).await?;
835
836 let request_body = CountTokensBody {
837 provider: cloud_llm_client::LanguageModelProvider::Google,
838 model: model_id,
839 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
840 generate_content_request,
841 })?,
842 };
843 let request = http_client::Request::builder()
844 .method(Method::POST)
845 .uri(
846 http_client
847 .build_zed_llm_url("/count_tokens", &[])?
848 .as_ref(),
849 )
850 .header("Content-Type", "application/json")
851 .header("Authorization", format!("Bearer {token}"))
852 .body(serde_json::to_string(&request_body)?.into())?;
853 let mut response = http_client.send(request).await?;
854 let status = response.status();
855 let headers = response.headers().clone();
856 let mut response_body = String::new();
857 response
858 .body_mut()
859 .read_to_string(&mut response_body)
860 .await?;
861
862 if status.is_success() {
863 let response_body: CountTokensResponse =
864 serde_json::from_str(&response_body)?;
865
866 Ok(response_body.tokens as u64)
867 } else {
868 Err(anyhow!(ApiError {
869 status,
870 body: response_body,
871 headers
872 }))
873 }
874 }
875 .boxed()
876 }
877 }
878 }
879
880 fn stream_completion(
881 &self,
882 request: LanguageModelRequest,
883 cx: &AsyncApp,
884 ) -> BoxFuture<
885 'static,
886 Result<
887 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
888 LanguageModelCompletionError,
889 >,
890 > {
891 let thread_id = request.thread_id.clone();
892 let prompt_id = request.prompt_id.clone();
893 let intent = request.intent;
894 let mode = request.mode;
895 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
896 let thinking_allowed = request.thinking_allowed;
897 match self.model.provider {
898 cloud_llm_client::LanguageModelProvider::Anthropic => {
899 let request = into_anthropic(
900 request,
901 self.model.id.to_string(),
902 1.0,
903 self.model.max_output_tokens as u64,
904 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
905 AnthropicModelMode::Thinking {
906 budget_tokens: Some(4_096),
907 }
908 } else {
909 AnthropicModelMode::Default
910 },
911 );
912 let client = self.client.clone();
913 let llm_api_token = self.llm_api_token.clone();
914 let future = self.request_limiter.stream(async move {
915 let PerformLlmCompletionResponse {
916 response,
917 usage,
918 includes_status_messages,
919 tool_use_limit_reached,
920 } = Self::perform_llm_completion(
921 client.clone(),
922 llm_api_token,
923 app_version,
924 CompletionBody {
925 thread_id,
926 prompt_id,
927 intent,
928 mode,
929 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
930 model: request.model.clone(),
931 provider_request: serde_json::to_value(&request)
932 .map_err(|e| anyhow!(e))?,
933 },
934 )
935 .await
936 .map_err(|err| match err.downcast::<ApiError>() {
937 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
938 Err(err) => anyhow!(err),
939 })?;
940
941 let mut mapper = AnthropicEventMapper::new();
942 Ok(map_cloud_completion_events(
943 Box::pin(
944 response_lines(response, includes_status_messages)
945 .chain(usage_updated_event(usage))
946 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
947 ),
948 move |event| mapper.map_event(event),
949 ))
950 });
951 async move { Ok(future.await?.boxed()) }.boxed()
952 }
953 cloud_llm_client::LanguageModelProvider::OpenAi => {
954 let client = self.client.clone();
955 let model = match open_ai::Model::from_id(&self.model.id.0) {
956 Ok(model) => model,
957 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
958 };
959 let request = into_open_ai(
960 request,
961 model.id(),
962 model.supports_parallel_tool_calls(),
963 None,
964 );
965 let llm_api_token = self.llm_api_token.clone();
966 let future = self.request_limiter.stream(async move {
967 let PerformLlmCompletionResponse {
968 response,
969 usage,
970 includes_status_messages,
971 tool_use_limit_reached,
972 } = Self::perform_llm_completion(
973 client.clone(),
974 llm_api_token,
975 app_version,
976 CompletionBody {
977 thread_id,
978 prompt_id,
979 intent,
980 mode,
981 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
982 model: request.model.clone(),
983 provider_request: serde_json::to_value(&request)
984 .map_err(|e| anyhow!(e))?,
985 },
986 )
987 .await?;
988
989 let mut mapper = OpenAiEventMapper::new();
990 Ok(map_cloud_completion_events(
991 Box::pin(
992 response_lines(response, includes_status_messages)
993 .chain(usage_updated_event(usage))
994 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
995 ),
996 move |event| mapper.map_event(event),
997 ))
998 });
999 async move { Ok(future.await?.boxed()) }.boxed()
1000 }
1001 cloud_llm_client::LanguageModelProvider::Google => {
1002 let client = self.client.clone();
1003 let request =
1004 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
1005 let llm_api_token = self.llm_api_token.clone();
1006 let future = self.request_limiter.stream(async move {
1007 let PerformLlmCompletionResponse {
1008 response,
1009 usage,
1010 includes_status_messages,
1011 tool_use_limit_reached,
1012 } = Self::perform_llm_completion(
1013 client.clone(),
1014 llm_api_token,
1015 app_version,
1016 CompletionBody {
1017 thread_id,
1018 prompt_id,
1019 intent,
1020 mode,
1021 provider: cloud_llm_client::LanguageModelProvider::Google,
1022 model: request.model.model_id.clone(),
1023 provider_request: serde_json::to_value(&request)
1024 .map_err(|e| anyhow!(e))?,
1025 },
1026 )
1027 .await?;
1028
1029 let mut mapper = GoogleEventMapper::new();
1030 Ok(map_cloud_completion_events(
1031 Box::pin(
1032 response_lines(response, includes_status_messages)
1033 .chain(usage_updated_event(usage))
1034 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
1035 ),
1036 move |event| mapper.map_event(event),
1037 ))
1038 });
1039 async move { Ok(future.await?.boxed()) }.boxed()
1040 }
1041 }
1042 }
1043}
1044
1045fn map_cloud_completion_events<T, F>(
1046 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
1047 mut map_callback: F,
1048) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1049where
1050 T: DeserializeOwned + 'static,
1051 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1052 + Send
1053 + 'static,
1054{
1055 stream
1056 .flat_map(move |event| {
1057 futures::stream::iter(match event {
1058 Err(error) => {
1059 vec![Err(LanguageModelCompletionError::from(error))]
1060 }
1061 Ok(CompletionEvent::Status(event)) => {
1062 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
1063 }
1064 Ok(CompletionEvent::Event(event)) => map_callback(event),
1065 })
1066 })
1067 .boxed()
1068}
1069
1070fn usage_updated_event<T>(
1071 usage: Option<ModelRequestUsage>,
1072) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1073 futures::stream::iter(usage.map(|usage| {
1074 Ok(CompletionEvent::Status(
1075 CompletionRequestStatus::UsageUpdated {
1076 amount: usage.amount as usize,
1077 limit: usage.limit,
1078 },
1079 ))
1080 }))
1081}
1082
1083fn tool_use_limit_reached_event<T>(
1084 tool_use_limit_reached: bool,
1085) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1086 futures::stream::iter(tool_use_limit_reached.then(|| {
1087 Ok(CompletionEvent::Status(
1088 CompletionRequestStatus::ToolUseLimitReached,
1089 ))
1090 }))
1091}
1092
1093fn response_lines<T: DeserializeOwned>(
1094 response: Response<AsyncBody>,
1095 includes_status_messages: bool,
1096) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1097 futures::stream::try_unfold(
1098 (String::new(), BufReader::new(response.into_body())),
1099 move |(mut line, mut body)| async move {
1100 match body.read_line(&mut line).await {
1101 Ok(0) => Ok(None),
1102 Ok(_) => {
1103 let event = if includes_status_messages {
1104 serde_json::from_str::<CompletionEvent<T>>(&line)?
1105 } else {
1106 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1107 };
1108
1109 line.clear();
1110 Ok(Some((event, (line, body))))
1111 }
1112 Err(e) => Err(e.into()),
1113 }
1114 },
1115 )
1116}
1117
1118#[derive(IntoElement, RegisterComponent)]
1119struct ZedAiConfiguration {
1120 is_connected: bool,
1121 plan: Option<proto::Plan>,
1122 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1123 eligible_for_trial: bool,
1124 has_accepted_terms_of_service: bool,
1125 account_too_young: bool,
1126 accept_terms_of_service_in_progress: bool,
1127 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1128 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1129}
1130
1131impl RenderOnce for ZedAiConfiguration {
1132 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1133 let young_account_banner = YoungAccountBanner;
1134
1135 let is_pro = self.plan == Some(proto::Plan::ZedPro);
1136 let subscription_text = match (self.plan, self.subscription_period) {
1137 (Some(proto::Plan::ZedPro), Some(_)) => {
1138 "You have access to Zed's hosted models through your Pro subscription."
1139 }
1140 (Some(proto::Plan::ZedProTrial), Some(_)) => {
1141 "You have access to Zed's hosted models through your Pro trial."
1142 }
1143 (Some(proto::Plan::Free), Some(_)) => {
1144 "You have basic access to Zed's hosted models through the Free plan."
1145 }
1146 _ => {
1147 if self.eligible_for_trial {
1148 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1149 } else {
1150 "Subscribe for access to Zed's hosted models."
1151 }
1152 }
1153 };
1154
1155 let manage_subscription_buttons = if is_pro {
1156 Button::new("manage_settings", "Manage Subscription")
1157 .full_width()
1158 .style(ButtonStyle::Tinted(TintColor::Accent))
1159 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1160 .into_any_element()
1161 } else if self.plan.is_none() || self.eligible_for_trial {
1162 Button::new("start_trial", "Start 14-day Free Pro Trial")
1163 .full_width()
1164 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1165 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1166 .into_any_element()
1167 } else {
1168 Button::new("upgrade", "Upgrade to Pro")
1169 .full_width()
1170 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1171 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1172 .into_any_element()
1173 };
1174
1175 if !self.is_connected {
1176 return v_flex()
1177 .gap_2()
1178 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1179 .child(
1180 Button::new("sign_in", "Sign In to use Zed AI")
1181 .icon_color(Color::Muted)
1182 .icon(IconName::Github)
1183 .icon_size(IconSize::Small)
1184 .icon_position(IconPosition::Start)
1185 .full_width()
1186 .on_click({
1187 let callback = self.sign_in_callback.clone();
1188 move |_, window, cx| (callback)(window, cx)
1189 }),
1190 );
1191 }
1192
1193 v_flex()
1194 .gap_2()
1195 .w_full()
1196 .when(!self.has_accepted_terms_of_service, |this| {
1197 this.child(render_accept_terms(
1198 LanguageModelProviderTosView::Configuration,
1199 self.accept_terms_of_service_in_progress,
1200 {
1201 let callback = self.accept_terms_of_service_callback.clone();
1202 move |window, cx| (callback)(window, cx)
1203 },
1204 ))
1205 })
1206 .map(|this| {
1207 if self.has_accepted_terms_of_service && self.account_too_young {
1208 this.child(young_account_banner).child(
1209 Button::new("upgrade", "Upgrade to Pro")
1210 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1211 .full_width()
1212 .on_click(|_, _, cx| {
1213 cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))
1214 }),
1215 )
1216 } else if self.has_accepted_terms_of_service {
1217 this.text_sm()
1218 .child(subscription_text)
1219 .child(manage_subscription_buttons)
1220 } else {
1221 this
1222 }
1223 })
1224 .when(self.has_accepted_terms_of_service, |this| this)
1225 }
1226}
1227
1228struct ConfigurationView {
1229 state: Entity<State>,
1230 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1231 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1232}
1233
1234impl ConfigurationView {
1235 fn new(state: Entity<State>) -> Self {
1236 let accept_terms_of_service_callback = Arc::new({
1237 let state = state.clone();
1238 move |_window: &mut Window, cx: &mut App| {
1239 state.update(cx, |state, cx| {
1240 state.accept_terms_of_service(cx);
1241 });
1242 }
1243 });
1244
1245 let sign_in_callback = Arc::new({
1246 let state = state.clone();
1247 move |_window: &mut Window, cx: &mut App| {
1248 state.update(cx, |state, cx| {
1249 state.authenticate(cx).detach_and_log_err(cx);
1250 });
1251 }
1252 });
1253
1254 Self {
1255 state,
1256 accept_terms_of_service_callback,
1257 sign_in_callback,
1258 }
1259 }
1260}
1261
1262impl Render for ConfigurationView {
1263 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1264 let state = self.state.read(cx);
1265 let user_store = state.user_store.read(cx);
1266
1267 ZedAiConfiguration {
1268 is_connected: !state.is_signed_out(),
1269 plan: user_store.current_plan(),
1270 subscription_period: user_store.subscription_period(),
1271 eligible_for_trial: user_store.trial_started_at().is_none(),
1272 has_accepted_terms_of_service: state.has_accepted_terms_of_service(cx),
1273 account_too_young: user_store.account_too_young(),
1274 accept_terms_of_service_in_progress: state.accept_terms_of_service_task.is_some(),
1275 accept_terms_of_service_callback: self.accept_terms_of_service_callback.clone(),
1276 sign_in_callback: self.sign_in_callback.clone(),
1277 }
1278 }
1279}
1280
1281impl Component for ZedAiConfiguration {
1282 fn scope() -> ComponentScope {
1283 ComponentScope::Agent
1284 }
1285
1286 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1287 fn configuration(
1288 is_connected: bool,
1289 plan: Option<proto::Plan>,
1290 eligible_for_trial: bool,
1291 account_too_young: bool,
1292 has_accepted_terms_of_service: bool,
1293 ) -> AnyElement {
1294 ZedAiConfiguration {
1295 is_connected,
1296 plan,
1297 subscription_period: plan
1298 .is_some()
1299 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1300 eligible_for_trial,
1301 has_accepted_terms_of_service,
1302 account_too_young,
1303 accept_terms_of_service_in_progress: false,
1304 accept_terms_of_service_callback: Arc::new(|_, _| {}),
1305 sign_in_callback: Arc::new(|_, _| {}),
1306 }
1307 .into_any_element()
1308 }
1309
1310 Some(
1311 v_flex()
1312 .p_4()
1313 .gap_4()
1314 .children(vec![
1315 single_example(
1316 "Not connected",
1317 configuration(false, None, false, false, true),
1318 ),
1319 single_example(
1320 "Accept Terms of Service",
1321 configuration(true, None, true, false, false),
1322 ),
1323 single_example(
1324 "No Plan - Not eligible for trial",
1325 configuration(true, None, false, false, true),
1326 ),
1327 single_example(
1328 "No Plan - Eligible for trial",
1329 configuration(true, None, true, false, true),
1330 ),
1331 single_example(
1332 "Free Plan",
1333 configuration(true, Some(proto::Plan::Free), true, false, true),
1334 ),
1335 single_example(
1336 "Zed Pro Trial Plan",
1337 configuration(true, Some(proto::Plan::ZedProTrial), true, false, true),
1338 ),
1339 single_example(
1340 "Zed Pro Plan",
1341 configuration(true, Some(proto::Plan::ZedPro), true, false, true),
1342 ),
1343 ])
1344 .into_any_element(),
1345 )
1346 }
1347}
1348
1349#[cfg(test)]
1350mod tests {
1351 use super::*;
1352 use http_client::http::{HeaderMap, StatusCode};
1353 use language_model::LanguageModelCompletionError;
1354
1355 #[test]
1356 fn test_api_error_conversion_with_upstream_http_error() {
1357 // upstream_http_error with 503 status should become ServerOverloaded
1358 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}"#;
1359
1360 let api_error = ApiError {
1361 status: StatusCode::INTERNAL_SERVER_ERROR,
1362 body: error_body.to_string(),
1363 headers: HeaderMap::new(),
1364 };
1365
1366 let completion_error: LanguageModelCompletionError = api_error.into();
1367
1368 match completion_error {
1369 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1370 assert_eq!(
1371 message,
1372 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1373 );
1374 }
1375 _ => panic!(
1376 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1377 completion_error
1378 ),
1379 }
1380
1381 // upstream_http_error with 500 status should become ApiInternalServerError
1382 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1383
1384 let api_error = ApiError {
1385 status: StatusCode::INTERNAL_SERVER_ERROR,
1386 body: error_body.to_string(),
1387 headers: HeaderMap::new(),
1388 };
1389
1390 let completion_error: LanguageModelCompletionError = api_error.into();
1391
1392 match completion_error {
1393 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1394 assert_eq!(
1395 message,
1396 "Received an error from the OpenAI API: internal server error"
1397 );
1398 }
1399 _ => panic!(
1400 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1401 completion_error
1402 ),
1403 }
1404
1405 // upstream_http_error with 429 status should become RateLimitExceeded
1406 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1407
1408 let api_error = ApiError {
1409 status: StatusCode::INTERNAL_SERVER_ERROR,
1410 body: error_body.to_string(),
1411 headers: HeaderMap::new(),
1412 };
1413
1414 let completion_error: LanguageModelCompletionError = api_error.into();
1415
1416 match completion_error {
1417 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1418 assert_eq!(
1419 message,
1420 "Received an error from the Google API: rate limit exceeded"
1421 );
1422 }
1423 _ => panic!(
1424 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1425 completion_error
1426 ),
1427 }
1428
1429 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1430 let error_body = "Regular internal server error";
1431
1432 let api_error = ApiError {
1433 status: StatusCode::INTERNAL_SERVER_ERROR,
1434 body: error_body.to_string(),
1435 headers: HeaderMap::new(),
1436 };
1437
1438 let completion_error: LanguageModelCompletionError = api_error.into();
1439
1440 match completion_error {
1441 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1442 assert_eq!(provider, PROVIDER_NAME);
1443 assert_eq!(message, "Regular internal server error");
1444 }
1445 _ => panic!(
1446 "Expected ApiInternalServerError for regular 500, got: {:?}",
1447 completion_error
1448 ),
1449 }
1450
1451 // upstream_http_429 format should be converted to UpstreamProviderError
1452 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1453
1454 let api_error = ApiError {
1455 status: StatusCode::INTERNAL_SERVER_ERROR,
1456 body: error_body.to_string(),
1457 headers: HeaderMap::new(),
1458 };
1459
1460 let completion_error: LanguageModelCompletionError = api_error.into();
1461
1462 match completion_error {
1463 LanguageModelCompletionError::UpstreamProviderError {
1464 message,
1465 status,
1466 retry_after,
1467 } => {
1468 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1469 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1470 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1471 }
1472 _ => panic!(
1473 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1474 completion_error
1475 ),
1476 }
1477
1478 // Invalid JSON in error body should fall back to regular error handling
1479 let error_body = "Not JSON at all";
1480
1481 let api_error = ApiError {
1482 status: StatusCode::INTERNAL_SERVER_ERROR,
1483 body: error_body.to_string(),
1484 headers: HeaderMap::new(),
1485 };
1486
1487 let completion_error: LanguageModelCompletionError = api_error.into();
1488
1489 match completion_error {
1490 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1491 assert_eq!(provider, PROVIDER_NAME);
1492 }
1493 _ => panic!(
1494 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1495 completion_error
1496 ),
1497 }
1498 }
1499}