1use ai_onboarding::YoungAccountBanner;
2use anthropic::AnthropicModelMode;
3use anyhow::{Context as _, Result, anyhow};
4use chrono::{DateTime, Utc};
5use client::{Client, ModelRequestUsage, UserStore, zed_urls};
6use cloud_llm_client::{
7 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
8 CompletionEvent, CompletionRequestStatus, CountTokensBody, CountTokensResponse,
9 EXPIRED_LLM_TOKEN_HEADER_NAME, ListModelsResponse, MODEL_REQUESTS_RESOURCE_HEADER_VALUE, Plan,
10 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
11 TOOL_USE_LIMIT_REACHED_HEADER_NAME, ZED_VERSION_HEADER_NAME,
12};
13use futures::{
14 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
15};
16use google_ai::GoogleModelMode;
17use gpui::{
18 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
19};
20use http_client::http::{HeaderMap, HeaderValue};
21use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
22use language_model::{
23 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
24 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
25 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
26 LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest,
27 LanguageModelToolChoice, LanguageModelToolSchemaFormat, LlmApiToken,
28 ModelRequestLimitReachedError, PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
29};
30use release_channel::AppVersion;
31use schemars::JsonSchema;
32use serde::{Deserialize, Serialize, de::DeserializeOwned};
33use settings::SettingsStore;
34use smol::io::{AsyncReadExt, BufReader};
35use std::pin::Pin;
36use std::str::FromStr as _;
37use std::sync::Arc;
38use std::time::Duration;
39use thiserror::Error;
40use ui::{TintColor, prelude::*};
41use util::{ResultExt as _, maybe};
42
43use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
44use crate::provider::google::{GoogleEventMapper, into_google};
45use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
46
47const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
48const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
49
50#[derive(Default, Clone, Debug, PartialEq)]
51pub struct ZedDotDevSettings {
52 pub available_models: Vec<AvailableModel>,
53}
54
55#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
56#[serde(rename_all = "lowercase")]
57pub enum AvailableProvider {
58 Anthropic,
59 OpenAi,
60 Google,
61}
62
63#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
64pub struct AvailableModel {
65 /// The provider of the language model.
66 pub provider: AvailableProvider,
67 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
68 pub name: String,
69 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
70 pub display_name: Option<String>,
71 /// The size of the context window, indicating the maximum number of tokens the model can process.
72 pub max_tokens: usize,
73 /// The maximum number of output tokens allowed by the model.
74 pub max_output_tokens: Option<u64>,
75 /// The maximum number of completion tokens allowed by the model (o1-* only)
76 pub max_completion_tokens: Option<u64>,
77 /// Override this model with a different Anthropic model for tool calls.
78 pub tool_override: Option<String>,
79 /// Indicates whether this custom model supports caching.
80 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
81 /// The default temperature to use for this model.
82 pub default_temperature: Option<f32>,
83 /// Any extra beta headers to provide when using the model.
84 #[serde(default)]
85 pub extra_beta_headers: Vec<String>,
86 /// The model's mode (e.g. thinking)
87 pub mode: Option<ModelMode>,
88}
89
90#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
91#[serde(tag = "type", rename_all = "lowercase")]
92pub enum ModelMode {
93 #[default]
94 Default,
95 Thinking {
96 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
97 budget_tokens: Option<u32>,
98 },
99}
100
101impl From<ModelMode> for AnthropicModelMode {
102 fn from(value: ModelMode) -> Self {
103 match value {
104 ModelMode::Default => AnthropicModelMode::Default,
105 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
106 }
107 }
108}
109
110pub struct CloudLanguageModelProvider {
111 client: Arc<Client>,
112 state: gpui::Entity<State>,
113 _maintain_client_status: Task<()>,
114}
115
116pub struct State {
117 client: Arc<Client>,
118 llm_api_token: LlmApiToken,
119 user_store: Entity<UserStore>,
120 status: client::Status,
121 accept_terms_of_service_task: Option<Task<Result<()>>>,
122 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
123 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
124 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
125 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
126 _fetch_models_task: Task<()>,
127 _settings_subscription: Subscription,
128 _llm_token_subscription: Subscription,
129}
130
131impl State {
132 fn new(
133 client: Arc<Client>,
134 user_store: Entity<UserStore>,
135 status: client::Status,
136 cx: &mut Context<Self>,
137 ) -> Self {
138 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
139 let mut current_user = user_store.read(cx).watch_current_user();
140 Self {
141 client: client.clone(),
142 llm_api_token: LlmApiToken::default(),
143 user_store: user_store.clone(),
144 status,
145 accept_terms_of_service_task: None,
146 models: Vec::new(),
147 default_model: None,
148 default_fast_model: None,
149 recommended_models: Vec::new(),
150 _fetch_models_task: cx.spawn(async move |this, cx| {
151 maybe!(async move {
152 let (client, llm_api_token) = this
153 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
154
155 while current_user.borrow().is_none() {
156 current_user.next().await;
157 }
158
159 let response =
160 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
161 this.update(cx, |this, cx| this.update_models(response, cx))?;
162 anyhow::Ok(())
163 })
164 .await
165 .context("failed to fetch Zed models")
166 .log_err();
167 }),
168 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
169 cx.notify();
170 }),
171 _llm_token_subscription: cx.subscribe(
172 &refresh_llm_token_listener,
173 move |this, _listener, _event, cx| {
174 let client = this.client.clone();
175 let llm_api_token = this.llm_api_token.clone();
176 cx.spawn(async move |this, cx| {
177 llm_api_token.refresh(&client).await?;
178 let response = Self::fetch_models(client, llm_api_token).await?;
179 this.update(cx, |this, cx| {
180 this.update_models(response, cx);
181 })
182 })
183 .detach_and_log_err(cx);
184 },
185 ),
186 }
187 }
188
189 fn is_signed_out(&self, cx: &App) -> bool {
190 self.user_store.read(cx).current_user().is_none()
191 }
192
193 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
194 let client = self.client.clone();
195 cx.spawn(async move |state, cx| {
196 client.sign_in_with_optional_connect(true, &cx).await?;
197 state.update(cx, |_, cx| cx.notify())
198 })
199 }
200
201 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
202 self.user_store.read(cx).has_accepted_terms_of_service()
203 }
204
205 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
206 let user_store = self.user_store.clone();
207 self.accept_terms_of_service_task = Some(cx.spawn(async move |this, cx| {
208 let _ = user_store
209 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
210 .await;
211 this.update(cx, |this, cx| {
212 this.accept_terms_of_service_task = None;
213 cx.notify()
214 })
215 }));
216 }
217
218 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
219 let mut models = Vec::new();
220
221 for model in response.models {
222 models.push(Arc::new(model.clone()));
223
224 // Right now we represent thinking variants of models as separate models on the client,
225 // so we need to insert variants for any model that supports thinking.
226 if model.supports_thinking {
227 models.push(Arc::new(cloud_llm_client::LanguageModel {
228 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
229 display_name: format!("{} Thinking", model.display_name),
230 ..model
231 }));
232 }
233 }
234
235 self.default_model = models
236 .iter()
237 .find(|model| model.id == response.default_model)
238 .cloned();
239 self.default_fast_model = models
240 .iter()
241 .find(|model| model.id == response.default_fast_model)
242 .cloned();
243 self.recommended_models = response
244 .recommended_models
245 .iter()
246 .filter_map(|id| models.iter().find(|model| &model.id == id))
247 .cloned()
248 .collect();
249 self.models = models;
250 cx.notify();
251 }
252
253 async fn fetch_models(
254 client: Arc<Client>,
255 llm_api_token: LlmApiToken,
256 ) -> Result<ListModelsResponse> {
257 let http_client = &client.http_client();
258 let token = llm_api_token.acquire(&client).await?;
259
260 let request = http_client::Request::builder()
261 .method(Method::GET)
262 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
263 .header("Authorization", format!("Bearer {token}"))
264 .body(AsyncBody::empty())?;
265 let mut response = http_client
266 .send(request)
267 .await
268 .context("failed to send list models request")?;
269
270 if response.status().is_success() {
271 let mut body = String::new();
272 response.body_mut().read_to_string(&mut body).await?;
273 return Ok(serde_json::from_str(&body)?);
274 } else {
275 let mut body = String::new();
276 response.body_mut().read_to_string(&mut body).await?;
277 anyhow::bail!(
278 "error listing models.\nStatus: {:?}\nBody: {body}",
279 response.status(),
280 );
281 }
282 }
283}
284
285impl CloudLanguageModelProvider {
286 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
287 let mut status_rx = client.status();
288 let status = *status_rx.borrow();
289
290 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
291
292 let state_ref = state.downgrade();
293 let maintain_client_status = cx.spawn(async move |cx| {
294 while let Some(status) = status_rx.next().await {
295 if let Some(this) = state_ref.upgrade() {
296 _ = this.update(cx, |this, cx| {
297 if this.status != status {
298 this.status = status;
299 cx.notify();
300 }
301 });
302 } else {
303 break;
304 }
305 }
306 });
307
308 Self {
309 client,
310 state: state.clone(),
311 _maintain_client_status: maintain_client_status,
312 }
313 }
314
315 fn create_language_model(
316 &self,
317 model: Arc<cloud_llm_client::LanguageModel>,
318 llm_api_token: LlmApiToken,
319 ) -> Arc<dyn LanguageModel> {
320 Arc::new(CloudLanguageModel {
321 id: LanguageModelId(SharedString::from(model.id.0.clone())),
322 model,
323 llm_api_token: llm_api_token.clone(),
324 client: self.client.clone(),
325 request_limiter: RateLimiter::new(4),
326 })
327 }
328}
329
330impl LanguageModelProviderState for CloudLanguageModelProvider {
331 type ObservableEntity = State;
332
333 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
334 Some(self.state.clone())
335 }
336}
337
338impl LanguageModelProvider for CloudLanguageModelProvider {
339 fn id(&self) -> LanguageModelProviderId {
340 PROVIDER_ID
341 }
342
343 fn name(&self) -> LanguageModelProviderName {
344 PROVIDER_NAME
345 }
346
347 fn icon(&self) -> IconName {
348 IconName::AiZed
349 }
350
351 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
352 let default_model = self.state.read(cx).default_model.clone()?;
353 let llm_api_token = self.state.read(cx).llm_api_token.clone();
354 Some(self.create_language_model(default_model, llm_api_token))
355 }
356
357 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
358 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
359 let llm_api_token = self.state.read(cx).llm_api_token.clone();
360 Some(self.create_language_model(default_fast_model, llm_api_token))
361 }
362
363 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
364 let llm_api_token = self.state.read(cx).llm_api_token.clone();
365 self.state
366 .read(cx)
367 .recommended_models
368 .iter()
369 .cloned()
370 .map(|model| self.create_language_model(model, llm_api_token.clone()))
371 .collect()
372 }
373
374 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
375 let llm_api_token = self.state.read(cx).llm_api_token.clone();
376 self.state
377 .read(cx)
378 .models
379 .iter()
380 .cloned()
381 .map(|model| self.create_language_model(model, llm_api_token.clone()))
382 .collect()
383 }
384
385 fn is_authenticated(&self, cx: &App) -> bool {
386 let state = self.state.read(cx);
387 !state.is_signed_out(cx) && state.has_accepted_terms_of_service(cx)
388 }
389
390 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
391 Task::ready(Ok(()))
392 }
393
394 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
395 cx.new(|_| ConfigurationView::new(self.state.clone()))
396 .into()
397 }
398
399 fn must_accept_terms(&self, cx: &App) -> bool {
400 !self.state.read(cx).has_accepted_terms_of_service(cx)
401 }
402
403 fn render_accept_terms(
404 &self,
405 view: LanguageModelProviderTosView,
406 cx: &mut App,
407 ) -> Option<AnyElement> {
408 let state = self.state.read(cx);
409 if state.has_accepted_terms_of_service(cx) {
410 return None;
411 }
412 Some(
413 render_accept_terms(
414 view,
415 state.accept_terms_of_service_task.is_some(),
416 {
417 let state = self.state.clone();
418 move |_window, cx| {
419 state.update(cx, |state, cx| state.accept_terms_of_service(cx));
420 }
421 },
422 cx,
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 app: &App,
438) -> impl IntoElement {
439 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
440 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadEmptyState);
441
442 let terms_button = Button::new("terms_of_service", "Terms of Service", app)
443 .style(ButtonStyle::Subtle)
444 .icon(IconName::ArrowUpRight)
445 .icon_color(Color::Muted)
446 .icon_size(IconSize::Small)
447 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
448 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
449
450 let button_container = h_flex().child(
451 Button::new("accept_terms", "I accept the Terms of Service", app)
452 .when(!thread_empty_state, |this| {
453 this.full_width()
454 .style(ButtonStyle::Tinted(TintColor::Accent))
455 .icon(IconName::Check)
456 .icon_position(IconPosition::Start)
457 .icon_size(IconSize::Small)
458 })
459 .when(thread_empty_state, |this| {
460 this.style(ButtonStyle::Tinted(TintColor::Warning))
461 .label_size(LabelSize::Small)
462 })
463 .disabled(accept_terms_of_service_in_progress)
464 .on_click(move |_, window, cx| (accept_terms_callback)(window, cx)),
465 );
466
467 if thread_empty_state {
468 h_flex()
469 .w_full()
470 .flex_wrap()
471 .justify_between()
472 .child(
473 h_flex()
474 .child(
475 Label::new("To start using Zed AI, please read and accept the")
476 .size(LabelSize::Small),
477 )
478 .child(terms_button),
479 )
480 .child(button_container)
481 } else {
482 v_flex()
483 .w_full()
484 .gap_2()
485 .child(
486 h_flex()
487 .flex_wrap()
488 .when(thread_fresh_start, |this| this.justify_center())
489 .child(Label::new(
490 "To start using Zed AI, please read and accept the",
491 ))
492 .child(terms_button),
493 )
494 .child({
495 match view_kind {
496 LanguageModelProviderTosView::TextThreadPopup => {
497 button_container.w_full().justify_end()
498 }
499 LanguageModelProviderTosView::Configuration => {
500 button_container.w_full().justify_start()
501 }
502 LanguageModelProviderTosView::ThreadFreshStart => {
503 button_container.w_full().justify_center()
504 }
505 LanguageModelProviderTosView::ThreadEmptyState => div().w_0(),
506 }
507 })
508 }
509}
510
511pub struct CloudLanguageModel {
512 id: LanguageModelId,
513 model: Arc<cloud_llm_client::LanguageModel>,
514 llm_api_token: LlmApiToken,
515 client: Arc<Client>,
516 request_limiter: RateLimiter,
517}
518
519struct PerformLlmCompletionResponse {
520 response: Response<AsyncBody>,
521 usage: Option<ModelRequestUsage>,
522 tool_use_limit_reached: bool,
523 includes_status_messages: bool,
524}
525
526impl CloudLanguageModel {
527 async fn perform_llm_completion(
528 client: Arc<Client>,
529 llm_api_token: LlmApiToken,
530 app_version: Option<SemanticVersion>,
531 body: CompletionBody,
532 ) -> Result<PerformLlmCompletionResponse> {
533 let http_client = &client.http_client();
534
535 let mut token = llm_api_token.acquire(&client).await?;
536 let mut refreshed_token = false;
537
538 loop {
539 let request_builder = http_client::Request::builder()
540 .method(Method::POST)
541 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
542 let request_builder = if let Some(app_version) = app_version {
543 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
544 } else {
545 request_builder
546 };
547
548 let request = request_builder
549 .header("Content-Type", "application/json")
550 .header("Authorization", format!("Bearer {token}"))
551 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
552 .body(serde_json::to_string(&body)?.into())?;
553 let mut response = http_client.send(request).await?;
554 let status = response.status();
555 if status.is_success() {
556 let includes_status_messages = response
557 .headers()
558 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
559 .is_some();
560
561 let tool_use_limit_reached = response
562 .headers()
563 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
564 .is_some();
565
566 let usage = if includes_status_messages {
567 None
568 } else {
569 ModelRequestUsage::from_headers(response.headers()).ok()
570 };
571
572 return Ok(PerformLlmCompletionResponse {
573 response,
574 usage,
575 includes_status_messages,
576 tool_use_limit_reached,
577 });
578 }
579
580 if !refreshed_token
581 && response
582 .headers()
583 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
584 .is_some()
585 {
586 token = llm_api_token.refresh(&client).await?;
587 refreshed_token = true;
588 continue;
589 }
590
591 if status == StatusCode::FORBIDDEN
592 && response
593 .headers()
594 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
595 .is_some()
596 {
597 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
598 .headers()
599 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
600 .and_then(|resource| resource.to_str().ok())
601 {
602 if let Some(plan) = response
603 .headers()
604 .get(CURRENT_PLAN_HEADER_NAME)
605 .and_then(|plan| plan.to_str().ok())
606 .and_then(|plan| cloud_llm_client::Plan::from_str(plan).ok())
607 {
608 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
609 }
610 }
611 } else if status == StatusCode::PAYMENT_REQUIRED {
612 return Err(anyhow!(PaymentRequiredError));
613 }
614
615 let mut body = String::new();
616 let headers = response.headers().clone();
617 response.body_mut().read_to_string(&mut body).await?;
618 return Err(anyhow!(ApiError {
619 status,
620 body,
621 headers
622 }));
623 }
624 }
625}
626
627#[derive(Debug, Error)]
628#[error("cloud language model request failed with status {status}: {body}")]
629struct ApiError {
630 status: StatusCode,
631 body: String,
632 headers: HeaderMap<HeaderValue>,
633}
634
635/// Represents error responses from Zed's cloud API.
636///
637/// Example JSON for an upstream HTTP error:
638/// ```json
639/// {
640/// "code": "upstream_http_error",
641/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
642/// "upstream_status": 503
643/// }
644/// ```
645#[derive(Debug, serde::Deserialize)]
646struct CloudApiError {
647 code: String,
648 message: String,
649 #[serde(default)]
650 #[serde(deserialize_with = "deserialize_optional_status_code")]
651 upstream_status: Option<StatusCode>,
652 #[serde(default)]
653 retry_after: Option<f64>,
654}
655
656fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
657where
658 D: serde::Deserializer<'de>,
659{
660 let opt: Option<u16> = Option::deserialize(deserializer)?;
661 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
662}
663
664impl From<ApiError> for LanguageModelCompletionError {
665 fn from(error: ApiError) -> Self {
666 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
667 if cloud_error.code.starts_with("upstream_http_") {
668 let status = if let Some(status) = cloud_error.upstream_status {
669 status
670 } else if cloud_error.code.ends_with("_error") {
671 error.status
672 } else {
673 // If there's a status code in the code string (e.g. "upstream_http_429")
674 // then use that; otherwise, see if the JSON contains a status code.
675 cloud_error
676 .code
677 .strip_prefix("upstream_http_")
678 .and_then(|code_str| code_str.parse::<u16>().ok())
679 .and_then(|code| StatusCode::from_u16(code).ok())
680 .unwrap_or(error.status)
681 };
682
683 return LanguageModelCompletionError::UpstreamProviderError {
684 message: cloud_error.message,
685 status,
686 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
687 };
688 }
689 }
690
691 let retry_after = None;
692 LanguageModelCompletionError::from_http_status(
693 PROVIDER_NAME,
694 error.status,
695 error.body,
696 retry_after,
697 )
698 }
699}
700
701impl LanguageModel for CloudLanguageModel {
702 fn id(&self) -> LanguageModelId {
703 self.id.clone()
704 }
705
706 fn name(&self) -> LanguageModelName {
707 LanguageModelName::from(self.model.display_name.clone())
708 }
709
710 fn provider_id(&self) -> LanguageModelProviderId {
711 PROVIDER_ID
712 }
713
714 fn provider_name(&self) -> LanguageModelProviderName {
715 PROVIDER_NAME
716 }
717
718 fn upstream_provider_id(&self) -> LanguageModelProviderId {
719 use cloud_llm_client::LanguageModelProvider::*;
720 match self.model.provider {
721 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
722 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
723 Google => language_model::GOOGLE_PROVIDER_ID,
724 }
725 }
726
727 fn upstream_provider_name(&self) -> LanguageModelProviderName {
728 use cloud_llm_client::LanguageModelProvider::*;
729 match self.model.provider {
730 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
731 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
732 Google => language_model::GOOGLE_PROVIDER_NAME,
733 }
734 }
735
736 fn supports_tools(&self) -> bool {
737 self.model.supports_tools
738 }
739
740 fn supports_images(&self) -> bool {
741 self.model.supports_images
742 }
743
744 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
745 match choice {
746 LanguageModelToolChoice::Auto
747 | LanguageModelToolChoice::Any
748 | LanguageModelToolChoice::None => true,
749 }
750 }
751
752 fn supports_burn_mode(&self) -> bool {
753 self.model.supports_max_mode
754 }
755
756 fn telemetry_id(&self) -> String {
757 format!("zed.dev/{}", self.model.id)
758 }
759
760 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
761 match self.model.provider {
762 cloud_llm_client::LanguageModelProvider::Anthropic
763 | cloud_llm_client::LanguageModelProvider::OpenAi => {
764 LanguageModelToolSchemaFormat::JsonSchema
765 }
766 cloud_llm_client::LanguageModelProvider::Google => {
767 LanguageModelToolSchemaFormat::JsonSchemaSubset
768 }
769 }
770 }
771
772 fn max_token_count(&self) -> u64 {
773 self.model.max_token_count as u64
774 }
775
776 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
777 self.model
778 .max_token_count_in_max_mode
779 .filter(|_| self.model.supports_max_mode)
780 .map(|max_token_count| max_token_count as u64)
781 }
782
783 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
784 match &self.model.provider {
785 cloud_llm_client::LanguageModelProvider::Anthropic => {
786 Some(LanguageModelCacheConfiguration {
787 min_total_token: 2_048,
788 should_speculate: true,
789 max_cache_anchors: 4,
790 })
791 }
792 cloud_llm_client::LanguageModelProvider::OpenAi
793 | cloud_llm_client::LanguageModelProvider::Google => None,
794 }
795 }
796
797 fn count_tokens(
798 &self,
799 request: LanguageModelRequest,
800 cx: &App,
801 ) -> BoxFuture<'static, Result<u64>> {
802 match self.model.provider {
803 cloud_llm_client::LanguageModelProvider::Anthropic => {
804 count_anthropic_tokens(request, cx)
805 }
806 cloud_llm_client::LanguageModelProvider::OpenAi => {
807 let model = match open_ai::Model::from_id(&self.model.id.0) {
808 Ok(model) => model,
809 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
810 };
811 count_open_ai_tokens(request, model, cx)
812 }
813 cloud_llm_client::LanguageModelProvider::Google => {
814 let client = self.client.clone();
815 let llm_api_token = self.llm_api_token.clone();
816 let model_id = self.model.id.to_string();
817 let generate_content_request =
818 into_google(request, model_id.clone(), GoogleModelMode::Default);
819 async move {
820 let http_client = &client.http_client();
821 let token = llm_api_token.acquire(&client).await?;
822
823 let request_body = CountTokensBody {
824 provider: cloud_llm_client::LanguageModelProvider::Google,
825 model: model_id,
826 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
827 generate_content_request,
828 })?,
829 };
830 let request = http_client::Request::builder()
831 .method(Method::POST)
832 .uri(
833 http_client
834 .build_zed_llm_url("/count_tokens", &[])?
835 .as_ref(),
836 )
837 .header("Content-Type", "application/json")
838 .header("Authorization", format!("Bearer {token}"))
839 .body(serde_json::to_string(&request_body)?.into())?;
840 let mut response = http_client.send(request).await?;
841 let status = response.status();
842 let headers = response.headers().clone();
843 let mut response_body = String::new();
844 response
845 .body_mut()
846 .read_to_string(&mut response_body)
847 .await?;
848
849 if status.is_success() {
850 let response_body: CountTokensResponse =
851 serde_json::from_str(&response_body)?;
852
853 Ok(response_body.tokens as u64)
854 } else {
855 Err(anyhow!(ApiError {
856 status,
857 body: response_body,
858 headers
859 }))
860 }
861 }
862 .boxed()
863 }
864 }
865 }
866
867 fn stream_completion(
868 &self,
869 request: LanguageModelRequest,
870 cx: &AsyncApp,
871 ) -> BoxFuture<
872 'static,
873 Result<
874 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
875 LanguageModelCompletionError,
876 >,
877 > {
878 let thread_id = request.thread_id.clone();
879 let prompt_id = request.prompt_id.clone();
880 let intent = request.intent;
881 let mode = request.mode;
882 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
883 let thinking_allowed = request.thinking_allowed;
884 match self.model.provider {
885 cloud_llm_client::LanguageModelProvider::Anthropic => {
886 let request = into_anthropic(
887 request,
888 self.model.id.to_string(),
889 1.0,
890 self.model.max_output_tokens as u64,
891 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
892 AnthropicModelMode::Thinking {
893 budget_tokens: Some(4_096),
894 }
895 } else {
896 AnthropicModelMode::Default
897 },
898 );
899 let client = self.client.clone();
900 let llm_api_token = self.llm_api_token.clone();
901 let future = self.request_limiter.stream(async move {
902 let PerformLlmCompletionResponse {
903 response,
904 usage,
905 includes_status_messages,
906 tool_use_limit_reached,
907 } = Self::perform_llm_completion(
908 client.clone(),
909 llm_api_token,
910 app_version,
911 CompletionBody {
912 thread_id,
913 prompt_id,
914 intent,
915 mode,
916 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
917 model: request.model.clone(),
918 provider_request: serde_json::to_value(&request)
919 .map_err(|e| anyhow!(e))?,
920 },
921 )
922 .await
923 .map_err(|err| match err.downcast::<ApiError>() {
924 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
925 Err(err) => anyhow!(err),
926 })?;
927
928 let mut mapper = AnthropicEventMapper::new();
929 Ok(map_cloud_completion_events(
930 Box::pin(
931 response_lines(response, includes_status_messages)
932 .chain(usage_updated_event(usage))
933 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
934 ),
935 move |event| mapper.map_event(event),
936 ))
937 });
938 async move { Ok(future.await?.boxed()) }.boxed()
939 }
940 cloud_llm_client::LanguageModelProvider::OpenAi => {
941 let client = self.client.clone();
942 let model = match open_ai::Model::from_id(&self.model.id.0) {
943 Ok(model) => model,
944 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
945 };
946 let request = into_open_ai(
947 request,
948 model.id(),
949 model.supports_parallel_tool_calls(),
950 None,
951 None,
952 );
953 let llm_api_token = self.llm_api_token.clone();
954 let future = self.request_limiter.stream(async move {
955 let PerformLlmCompletionResponse {
956 response,
957 usage,
958 includes_status_messages,
959 tool_use_limit_reached,
960 } = Self::perform_llm_completion(
961 client.clone(),
962 llm_api_token,
963 app_version,
964 CompletionBody {
965 thread_id,
966 prompt_id,
967 intent,
968 mode,
969 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
970 model: request.model.clone(),
971 provider_request: serde_json::to_value(&request)
972 .map_err(|e| anyhow!(e))?,
973 },
974 )
975 .await?;
976
977 let mut mapper = OpenAiEventMapper::new();
978 Ok(map_cloud_completion_events(
979 Box::pin(
980 response_lines(response, includes_status_messages)
981 .chain(usage_updated_event(usage))
982 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
983 ),
984 move |event| mapper.map_event(event),
985 ))
986 });
987 async move { Ok(future.await?.boxed()) }.boxed()
988 }
989 cloud_llm_client::LanguageModelProvider::Google => {
990 let client = self.client.clone();
991 let request =
992 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
993 let llm_api_token = self.llm_api_token.clone();
994 let future = self.request_limiter.stream(async move {
995 let PerformLlmCompletionResponse {
996 response,
997 usage,
998 includes_status_messages,
999 tool_use_limit_reached,
1000 } = Self::perform_llm_completion(
1001 client.clone(),
1002 llm_api_token,
1003 app_version,
1004 CompletionBody {
1005 thread_id,
1006 prompt_id,
1007 intent,
1008 mode,
1009 provider: cloud_llm_client::LanguageModelProvider::Google,
1010 model: request.model.model_id.clone(),
1011 provider_request: serde_json::to_value(&request)
1012 .map_err(|e| anyhow!(e))?,
1013 },
1014 )
1015 .await?;
1016
1017 let mut mapper = GoogleEventMapper::new();
1018 Ok(map_cloud_completion_events(
1019 Box::pin(
1020 response_lines(response, includes_status_messages)
1021 .chain(usage_updated_event(usage))
1022 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
1023 ),
1024 move |event| mapper.map_event(event),
1025 ))
1026 });
1027 async move { Ok(future.await?.boxed()) }.boxed()
1028 }
1029 }
1030 }
1031}
1032
1033fn map_cloud_completion_events<T, F>(
1034 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
1035 mut map_callback: F,
1036) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1037where
1038 T: DeserializeOwned + 'static,
1039 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1040 + Send
1041 + 'static,
1042{
1043 stream
1044 .flat_map(move |event| {
1045 futures::stream::iter(match event {
1046 Err(error) => {
1047 vec![Err(LanguageModelCompletionError::from(error))]
1048 }
1049 Ok(CompletionEvent::Status(event)) => {
1050 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
1051 }
1052 Ok(CompletionEvent::Event(event)) => map_callback(event),
1053 })
1054 })
1055 .boxed()
1056}
1057
1058fn usage_updated_event<T>(
1059 usage: Option<ModelRequestUsage>,
1060) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1061 futures::stream::iter(usage.map(|usage| {
1062 Ok(CompletionEvent::Status(
1063 CompletionRequestStatus::UsageUpdated {
1064 amount: usage.amount as usize,
1065 limit: usage.limit,
1066 },
1067 ))
1068 }))
1069}
1070
1071fn tool_use_limit_reached_event<T>(
1072 tool_use_limit_reached: bool,
1073) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1074 futures::stream::iter(tool_use_limit_reached.then(|| {
1075 Ok(CompletionEvent::Status(
1076 CompletionRequestStatus::ToolUseLimitReached,
1077 ))
1078 }))
1079}
1080
1081fn response_lines<T: DeserializeOwned>(
1082 response: Response<AsyncBody>,
1083 includes_status_messages: bool,
1084) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1085 futures::stream::try_unfold(
1086 (String::new(), BufReader::new(response.into_body())),
1087 move |(mut line, mut body)| async move {
1088 match body.read_line(&mut line).await {
1089 Ok(0) => Ok(None),
1090 Ok(_) => {
1091 let event = if includes_status_messages {
1092 serde_json::from_str::<CompletionEvent<T>>(&line)?
1093 } else {
1094 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1095 };
1096
1097 line.clear();
1098 Ok(Some((event, (line, body))))
1099 }
1100 Err(e) => Err(e.into()),
1101 }
1102 },
1103 )
1104}
1105
1106#[derive(IntoElement, RegisterComponent)]
1107struct ZedAiConfiguration {
1108 is_connected: bool,
1109 plan: Option<Plan>,
1110 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1111 eligible_for_trial: bool,
1112 has_accepted_terms_of_service: bool,
1113 account_too_young: bool,
1114 accept_terms_of_service_in_progress: bool,
1115 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1116 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1117}
1118
1119impl RenderOnce for ZedAiConfiguration {
1120 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1121 let young_account_banner = YoungAccountBanner;
1122
1123 let is_pro = self.plan == Some(Plan::ZedPro);
1124 let subscription_text = match (self.plan, self.subscription_period) {
1125 (Some(Plan::ZedPro), Some(_)) => {
1126 "You have access to Zed's hosted models through your Pro subscription."
1127 }
1128 (Some(Plan::ZedProTrial), Some(_)) => {
1129 "You have access to Zed's hosted models through your Pro trial."
1130 }
1131 (Some(Plan::ZedFree), Some(_)) => {
1132 "You have basic access to Zed's hosted models through the Free plan."
1133 }
1134 _ => {
1135 if self.eligible_for_trial {
1136 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1137 } else {
1138 "Subscribe for access to Zed's hosted models."
1139 }
1140 }
1141 };
1142
1143 let manage_subscription_buttons = if is_pro {
1144 Button::new("manage_settings", "Manage Subscription", _cx)
1145 .full_width()
1146 .style(ButtonStyle::Tinted(TintColor::Accent))
1147 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1148 .into_any_element()
1149 } else if self.plan.is_none() || self.eligible_for_trial {
1150 Button::new("start_trial", "Start 14-day Free Pro Trial", _cx)
1151 .full_width()
1152 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1153 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1154 .into_any_element()
1155 } else {
1156 Button::new("upgrade", "Upgrade to Pro", _cx)
1157 .full_width()
1158 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1159 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1160 .into_any_element()
1161 };
1162
1163 if !self.is_connected {
1164 return v_flex()
1165 .gap_2()
1166 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1167 .child(
1168 Button::new("sign_in", "Sign In to use Zed AI", _cx)
1169 .icon_color(Color::Muted)
1170 .icon(IconName::Github)
1171 .icon_size(IconSize::Small)
1172 .icon_position(IconPosition::Start)
1173 .full_width()
1174 .on_click({
1175 let callback = self.sign_in_callback.clone();
1176 move |_, window, cx| (callback)(window, cx)
1177 }),
1178 );
1179 }
1180
1181 v_flex()
1182 .gap_2()
1183 .w_full()
1184 .when(!self.has_accepted_terms_of_service, |this| {
1185 this.child(render_accept_terms(
1186 LanguageModelProviderTosView::Configuration,
1187 self.accept_terms_of_service_in_progress,
1188 {
1189 let callback = self.accept_terms_of_service_callback.clone();
1190 move |window, cx| (callback)(window, cx)
1191 },
1192 _cx,
1193 ))
1194 })
1195 .map(|this| {
1196 if self.has_accepted_terms_of_service && self.account_too_young {
1197 this.child(young_account_banner).child(
1198 Button::new("upgrade", "Upgrade to Pro", _cx)
1199 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1200 .full_width()
1201 .on_click(|_, _, cx| {
1202 cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))
1203 }),
1204 )
1205 } else if self.has_accepted_terms_of_service {
1206 this.text_sm()
1207 .child(subscription_text)
1208 .child(manage_subscription_buttons)
1209 } else {
1210 this
1211 }
1212 })
1213 .when(self.has_accepted_terms_of_service, |this| this)
1214 }
1215}
1216
1217struct ConfigurationView {
1218 state: Entity<State>,
1219 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1220 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1221}
1222
1223impl ConfigurationView {
1224 fn new(state: Entity<State>) -> Self {
1225 let accept_terms_of_service_callback = Arc::new({
1226 let state = state.clone();
1227 move |_window: &mut Window, cx: &mut App| {
1228 state.update(cx, |state, cx| {
1229 state.accept_terms_of_service(cx);
1230 });
1231 }
1232 });
1233
1234 let sign_in_callback = Arc::new({
1235 let state = state.clone();
1236 move |_window: &mut Window, cx: &mut App| {
1237 state.update(cx, |state, cx| {
1238 state.authenticate(cx).detach_and_log_err(cx);
1239 });
1240 }
1241 });
1242
1243 Self {
1244 state,
1245 accept_terms_of_service_callback,
1246 sign_in_callback,
1247 }
1248 }
1249}
1250
1251impl Render for ConfigurationView {
1252 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1253 let state = self.state.read(cx);
1254 let user_store = state.user_store.read(cx);
1255
1256 ZedAiConfiguration {
1257 is_connected: !state.is_signed_out(cx),
1258 plan: user_store.plan(),
1259 subscription_period: user_store.subscription_period(),
1260 eligible_for_trial: user_store.trial_started_at().is_none(),
1261 has_accepted_terms_of_service: state.has_accepted_terms_of_service(cx),
1262 account_too_young: user_store.account_too_young(),
1263 accept_terms_of_service_in_progress: state.accept_terms_of_service_task.is_some(),
1264 accept_terms_of_service_callback: self.accept_terms_of_service_callback.clone(),
1265 sign_in_callback: self.sign_in_callback.clone(),
1266 }
1267 }
1268}
1269
1270impl Component for ZedAiConfiguration {
1271 fn name() -> &'static str {
1272 "AI Configuration Content"
1273 }
1274
1275 fn sort_name() -> &'static str {
1276 "AI Configuration Content"
1277 }
1278
1279 fn scope() -> ComponentScope {
1280 ComponentScope::Onboarding
1281 }
1282
1283 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1284 fn configuration(
1285 is_connected: bool,
1286 plan: Option<Plan>,
1287 eligible_for_trial: bool,
1288 account_too_young: bool,
1289 has_accepted_terms_of_service: bool,
1290 ) -> AnyElement {
1291 ZedAiConfiguration {
1292 is_connected,
1293 plan,
1294 subscription_period: plan
1295 .is_some()
1296 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1297 eligible_for_trial,
1298 has_accepted_terms_of_service,
1299 account_too_young,
1300 accept_terms_of_service_in_progress: false,
1301 accept_terms_of_service_callback: Arc::new(|_, _| {}),
1302 sign_in_callback: Arc::new(|_, _| {}),
1303 }
1304 .into_any_element()
1305 }
1306
1307 Some(
1308 v_flex()
1309 .p_4()
1310 .gap_4()
1311 .children(vec![
1312 single_example(
1313 "Not connected",
1314 configuration(false, None, false, false, true),
1315 ),
1316 single_example(
1317 "Accept Terms of Service",
1318 configuration(true, None, true, false, false),
1319 ),
1320 single_example(
1321 "No Plan - Not eligible for trial",
1322 configuration(true, None, false, false, true),
1323 ),
1324 single_example(
1325 "No Plan - Eligible for trial",
1326 configuration(true, None, true, false, true),
1327 ),
1328 single_example(
1329 "Free Plan",
1330 configuration(true, Some(Plan::ZedFree), true, false, true),
1331 ),
1332 single_example(
1333 "Zed Pro Trial Plan",
1334 configuration(true, Some(Plan::ZedProTrial), true, false, true),
1335 ),
1336 single_example(
1337 "Zed Pro Plan",
1338 configuration(true, Some(Plan::ZedPro), true, false, true),
1339 ),
1340 ])
1341 .into_any_element(),
1342 )
1343 }
1344}
1345
1346#[cfg(test)]
1347mod tests {
1348 use super::*;
1349 use http_client::http::{HeaderMap, StatusCode};
1350 use language_model::LanguageModelCompletionError;
1351
1352 #[test]
1353 fn test_api_error_conversion_with_upstream_http_error() {
1354 // upstream_http_error with 503 status should become ServerOverloaded
1355 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}"#;
1356
1357 let api_error = ApiError {
1358 status: StatusCode::INTERNAL_SERVER_ERROR,
1359 body: error_body.to_string(),
1360 headers: HeaderMap::new(),
1361 };
1362
1363 let completion_error: LanguageModelCompletionError = api_error.into();
1364
1365 match completion_error {
1366 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1367 assert_eq!(
1368 message,
1369 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1370 );
1371 }
1372 _ => panic!(
1373 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1374 completion_error
1375 ),
1376 }
1377
1378 // upstream_http_error with 500 status should become ApiInternalServerError
1379 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1380
1381 let api_error = ApiError {
1382 status: StatusCode::INTERNAL_SERVER_ERROR,
1383 body: error_body.to_string(),
1384 headers: HeaderMap::new(),
1385 };
1386
1387 let completion_error: LanguageModelCompletionError = api_error.into();
1388
1389 match completion_error {
1390 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1391 assert_eq!(
1392 message,
1393 "Received an error from the OpenAI API: internal server error"
1394 );
1395 }
1396 _ => panic!(
1397 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1398 completion_error
1399 ),
1400 }
1401
1402 // upstream_http_error with 429 status should become RateLimitExceeded
1403 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1404
1405 let api_error = ApiError {
1406 status: StatusCode::INTERNAL_SERVER_ERROR,
1407 body: error_body.to_string(),
1408 headers: HeaderMap::new(),
1409 };
1410
1411 let completion_error: LanguageModelCompletionError = api_error.into();
1412
1413 match completion_error {
1414 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1415 assert_eq!(
1416 message,
1417 "Received an error from the Google API: rate limit exceeded"
1418 );
1419 }
1420 _ => panic!(
1421 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1422 completion_error
1423 ),
1424 }
1425
1426 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1427 let error_body = "Regular internal server error";
1428
1429 let api_error = ApiError {
1430 status: StatusCode::INTERNAL_SERVER_ERROR,
1431 body: error_body.to_string(),
1432 headers: HeaderMap::new(),
1433 };
1434
1435 let completion_error: LanguageModelCompletionError = api_error.into();
1436
1437 match completion_error {
1438 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1439 assert_eq!(provider, PROVIDER_NAME);
1440 assert_eq!(message, "Regular internal server error");
1441 }
1442 _ => panic!(
1443 "Expected ApiInternalServerError for regular 500, got: {:?}",
1444 completion_error
1445 ),
1446 }
1447
1448 // upstream_http_429 format should be converted to UpstreamProviderError
1449 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1450
1451 let api_error = ApiError {
1452 status: StatusCode::INTERNAL_SERVER_ERROR,
1453 body: error_body.to_string(),
1454 headers: HeaderMap::new(),
1455 };
1456
1457 let completion_error: LanguageModelCompletionError = api_error.into();
1458
1459 match completion_error {
1460 LanguageModelCompletionError::UpstreamProviderError {
1461 message,
1462 status,
1463 retry_after,
1464 } => {
1465 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1466 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1467 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1468 }
1469 _ => panic!(
1470 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1471 completion_error
1472 ),
1473 }
1474
1475 // Invalid JSON in error body should fall back to regular error handling
1476 let error_body = "Not JSON at all";
1477
1478 let api_error = ApiError {
1479 status: StatusCode::INTERNAL_SERVER_ERROR,
1480 body: error_body.to_string(),
1481 headers: HeaderMap::new(),
1482 };
1483
1484 let completion_error: LanguageModelCompletionError = api_error.into();
1485
1486 match completion_error {
1487 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1488 assert_eq!(provider, PROVIDER_NAME);
1489 }
1490 _ => panic!(
1491 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1492 completion_error
1493 ),
1494 }
1495 }
1496}