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