1use anthropic::AnthropicModelMode;
2use anyhow::{Context as _, Result, anyhow};
3use chrono::{DateTime, Utc};
4use client::{Client, ModelRequestUsage, UserStore, zed_urls};
5use feature_flags::{FeatureFlagAppExt as _, ZedCloudFeatureFlag};
6use futures::{
7 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
8};
9use google_ai::GoogleModelMode;
10use gpui::{
11 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
12};
13use http_client::http::{HeaderMap, HeaderValue};
14use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
15use language_model::{
16 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
17 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
18 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
19 LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest,
20 LanguageModelToolChoice, LanguageModelToolSchemaFormat, LlmApiToken,
21 ModelRequestLimitReachedError, PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
22};
23use proto::Plan;
24use release_channel::AppVersion;
25use schemars::JsonSchema;
26use serde::{Deserialize, Serialize, de::DeserializeOwned};
27use settings::SettingsStore;
28use smol::io::{AsyncReadExt, BufReader};
29use std::pin::Pin;
30use std::str::FromStr as _;
31use std::sync::Arc;
32use std::time::Duration;
33use thiserror::Error;
34use ui::{TintColor, prelude::*};
35use util::{ResultExt as _, maybe};
36use zed_llm_client::{
37 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
38 CompletionRequestStatus, CountTokensBody, CountTokensResponse, EXPIRED_LLM_TOKEN_HEADER_NAME,
39 ListModelsResponse, MODEL_REQUESTS_RESOURCE_HEADER_VALUE,
40 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
41 TOOL_USE_LIMIT_REACHED_HEADER_NAME, ZED_VERSION_HEADER_NAME,
42};
43
44use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
45use crate::provider::google::{GoogleEventMapper, into_google};
46use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
47
48const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
49const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
50
51#[derive(Default, Clone, Debug, PartialEq)]
52pub struct ZedDotDevSettings {
53 pub available_models: Vec<AvailableModel>,
54}
55
56#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
57#[serde(rename_all = "lowercase")]
58pub enum AvailableProvider {
59 Anthropic,
60 OpenAi,
61 Google,
62}
63
64#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
65pub struct AvailableModel {
66 /// The provider of the language model.
67 pub provider: AvailableProvider,
68 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
69 pub name: String,
70 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
71 pub display_name: Option<String>,
72 /// The size of the context window, indicating the maximum number of tokens the model can process.
73 pub max_tokens: usize,
74 /// The maximum number of output tokens allowed by the model.
75 pub max_output_tokens: Option<u64>,
76 /// The maximum number of completion tokens allowed by the model (o1-* only)
77 pub max_completion_tokens: Option<u64>,
78 /// Override this model with a different Anthropic model for tool calls.
79 pub tool_override: Option<String>,
80 /// Indicates whether this custom model supports caching.
81 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
82 /// The default temperature to use for this model.
83 pub default_temperature: Option<f32>,
84 /// Any extra beta headers to provide when using the model.
85 #[serde(default)]
86 pub extra_beta_headers: Vec<String>,
87 /// The model's mode (e.g. thinking)
88 pub mode: Option<ModelMode>,
89}
90
91#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
92#[serde(tag = "type", rename_all = "lowercase")]
93pub enum ModelMode {
94 #[default]
95 Default,
96 Thinking {
97 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
98 budget_tokens: Option<u32>,
99 },
100}
101
102impl From<ModelMode> for AnthropicModelMode {
103 fn from(value: ModelMode) -> Self {
104 match value {
105 ModelMode::Default => AnthropicModelMode::Default,
106 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
107 }
108 }
109}
110
111pub struct CloudLanguageModelProvider {
112 client: Arc<Client>,
113 state: gpui::Entity<State>,
114 _maintain_client_status: Task<()>,
115}
116
117pub struct State {
118 client: Arc<Client>,
119 llm_api_token: LlmApiToken,
120 user_store: Entity<UserStore>,
121 status: client::Status,
122 accept_terms_of_service_task: Option<Task<Result<()>>>,
123 models: Vec<Arc<zed_llm_client::LanguageModel>>,
124 default_model: Option<Arc<zed_llm_client::LanguageModel>>,
125 default_fast_model: Option<Arc<zed_llm_client::LanguageModel>>,
126 recommended_models: Vec<Arc<zed_llm_client::LanguageModel>>,
127 _fetch_models_task: Task<()>,
128 _settings_subscription: Subscription,
129 _llm_token_subscription: Subscription,
130}
131
132impl State {
133 fn new(
134 client: Arc<Client>,
135 user_store: Entity<UserStore>,
136 status: client::Status,
137 cx: &mut Context<Self>,
138 ) -> Self {
139 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
140 let use_cloud = cx.has_flag::<ZedCloudFeatureFlag>();
141
142 Self {
143 client: client.clone(),
144 llm_api_token: LlmApiToken::default(),
145 user_store,
146 status,
147 accept_terms_of_service_task: None,
148 models: Vec::new(),
149 default_model: None,
150 default_fast_model: None,
151 recommended_models: Vec::new(),
152 _fetch_models_task: cx.spawn(async move |this, cx| {
153 maybe!(async move {
154 let (client, llm_api_token) = this
155 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
156
157 loop {
158 let status = this.read_with(cx, |this, _cx| this.status)?;
159 if matches!(status, client::Status::Connected { .. }) {
160 break;
161 }
162
163 cx.background_executor()
164 .timer(Duration::from_millis(100))
165 .await;
166 }
167
168 let response = Self::fetch_models(client, llm_api_token, use_cloud).await?;
169 this.update(cx, |this, cx| {
170 this.update_models(response, cx);
171 })
172 })
173 .await
174 .context("failed to fetch Zed models")
175 .log_err();
176 }),
177 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
178 cx.notify();
179 }),
180 _llm_token_subscription: cx.subscribe(
181 &refresh_llm_token_listener,
182 move |this, _listener, _event, cx| {
183 let client = this.client.clone();
184 let llm_api_token = this.llm_api_token.clone();
185 cx.spawn(async move |this, cx| {
186 llm_api_token.refresh(&client).await?;
187 let response = Self::fetch_models(client, llm_api_token, use_cloud).await?;
188 this.update(cx, |this, cx| {
189 this.update_models(response, cx);
190 })
191 })
192 .detach_and_log_err(cx);
193 },
194 ),
195 }
196 }
197
198 fn is_signed_out(&self) -> bool {
199 self.status.is_signed_out()
200 }
201
202 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
203 let client = self.client.clone();
204 cx.spawn(async move |state, cx| {
205 client
206 .authenticate_and_connect(true, &cx)
207 .await
208 .into_response()?;
209 state.update(cx, |_, cx| cx.notify())
210 })
211 }
212
213 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
214 self.user_store
215 .read(cx)
216 .current_user_has_accepted_terms()
217 .unwrap_or(false)
218 }
219
220 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
221 let user_store = self.user_store.clone();
222 self.accept_terms_of_service_task = Some(cx.spawn(async move |this, cx| {
223 let _ = user_store
224 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
225 .await;
226 this.update(cx, |this, cx| {
227 this.accept_terms_of_service_task = None;
228 cx.notify()
229 })
230 }));
231 }
232
233 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
234 let mut models = Vec::new();
235
236 for model in response.models {
237 models.push(Arc::new(model.clone()));
238
239 // Right now we represent thinking variants of models as separate models on the client,
240 // so we need to insert variants for any model that supports thinking.
241 if model.supports_thinking {
242 models.push(Arc::new(zed_llm_client::LanguageModel {
243 id: zed_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
244 display_name: format!("{} Thinking", model.display_name),
245 ..model
246 }));
247 }
248 }
249
250 self.default_model = models
251 .iter()
252 .find(|model| model.id == response.default_model)
253 .cloned();
254 self.default_fast_model = models
255 .iter()
256 .find(|model| model.id == response.default_fast_model)
257 .cloned();
258 self.recommended_models = response
259 .recommended_models
260 .iter()
261 .filter_map(|id| models.iter().find(|model| &model.id == id))
262 .cloned()
263 .collect();
264 self.models = models;
265 cx.notify();
266 }
267
268 async fn fetch_models(
269 client: Arc<Client>,
270 llm_api_token: LlmApiToken,
271 use_cloud: bool,
272 ) -> Result<ListModelsResponse> {
273 let http_client = &client.http_client();
274 let token = llm_api_token.acquire(&client).await?;
275
276 let request = http_client::Request::builder()
277 .method(Method::GET)
278 .uri(
279 http_client
280 .build_zed_llm_url("/models", &[], use_cloud)?
281 .as_ref(),
282 )
283 .header("Authorization", format!("Bearer {token}"))
284 .body(AsyncBody::empty())?;
285 let mut response = http_client
286 .send(request)
287 .await
288 .context("failed to send list models request")?;
289
290 if response.status().is_success() {
291 let mut body = String::new();
292 response.body_mut().read_to_string(&mut body).await?;
293 return Ok(serde_json::from_str(&body)?);
294 } else {
295 let mut body = String::new();
296 response.body_mut().read_to_string(&mut body).await?;
297 anyhow::bail!(
298 "error listing models.\nStatus: {:?}\nBody: {body}",
299 response.status(),
300 );
301 }
302 }
303}
304
305impl CloudLanguageModelProvider {
306 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
307 let mut status_rx = client.status();
308 let status = *status_rx.borrow();
309
310 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
311
312 let state_ref = state.downgrade();
313 let maintain_client_status = cx.spawn(async move |cx| {
314 while let Some(status) = status_rx.next().await {
315 if let Some(this) = state_ref.upgrade() {
316 _ = this.update(cx, |this, cx| {
317 if this.status != status {
318 this.status = status;
319 cx.notify();
320 }
321 });
322 } else {
323 break;
324 }
325 }
326 });
327
328 Self {
329 client,
330 state: state.clone(),
331 _maintain_client_status: maintain_client_status,
332 }
333 }
334
335 fn create_language_model(
336 &self,
337 model: Arc<zed_llm_client::LanguageModel>,
338 llm_api_token: LlmApiToken,
339 ) -> Arc<dyn LanguageModel> {
340 Arc::new(CloudLanguageModel {
341 id: LanguageModelId(SharedString::from(model.id.0.clone())),
342 model,
343 llm_api_token: llm_api_token.clone(),
344 client: self.client.clone(),
345 request_limiter: RateLimiter::new(4),
346 })
347 }
348}
349
350impl LanguageModelProviderState for CloudLanguageModelProvider {
351 type ObservableEntity = State;
352
353 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
354 Some(self.state.clone())
355 }
356}
357
358impl LanguageModelProvider for CloudLanguageModelProvider {
359 fn id(&self) -> LanguageModelProviderId {
360 PROVIDER_ID
361 }
362
363 fn name(&self) -> LanguageModelProviderName {
364 PROVIDER_NAME
365 }
366
367 fn icon(&self) -> IconName {
368 IconName::AiZed
369 }
370
371 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
372 let default_model = self.state.read(cx).default_model.clone()?;
373 let llm_api_token = self.state.read(cx).llm_api_token.clone();
374 Some(self.create_language_model(default_model, llm_api_token))
375 }
376
377 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
378 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
379 let llm_api_token = self.state.read(cx).llm_api_token.clone();
380 Some(self.create_language_model(default_fast_model, llm_api_token))
381 }
382
383 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
384 let llm_api_token = self.state.read(cx).llm_api_token.clone();
385 self.state
386 .read(cx)
387 .recommended_models
388 .iter()
389 .cloned()
390 .map(|model| self.create_language_model(model, llm_api_token.clone()))
391 .collect()
392 }
393
394 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
395 let llm_api_token = self.state.read(cx).llm_api_token.clone();
396 self.state
397 .read(cx)
398 .models
399 .iter()
400 .cloned()
401 .map(|model| self.create_language_model(model, llm_api_token.clone()))
402 .collect()
403 }
404
405 fn is_authenticated(&self, cx: &App) -> bool {
406 let state = self.state.read(cx);
407 !state.is_signed_out() && state.has_accepted_terms_of_service(cx)
408 }
409
410 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
411 Task::ready(Ok(()))
412 }
413
414 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
415 cx.new(|_| ConfigurationView::new(self.state.clone()))
416 .into()
417 }
418
419 fn must_accept_terms(&self, cx: &App) -> bool {
420 !self.state.read(cx).has_accepted_terms_of_service(cx)
421 }
422
423 fn render_accept_terms(
424 &self,
425 view: LanguageModelProviderTosView,
426 cx: &mut App,
427 ) -> Option<AnyElement> {
428 let state = self.state.read(cx);
429 if state.has_accepted_terms_of_service(cx) {
430 return None;
431 }
432 Some(
433 render_accept_terms(view, state.accept_terms_of_service_task.is_some(), {
434 let state = self.state.clone();
435 move |_window, cx| {
436 state.update(cx, |state, cx| state.accept_terms_of_service(cx));
437 }
438 })
439 .into_any_element(),
440 )
441 }
442
443 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
444 Task::ready(Ok(()))
445 }
446}
447
448fn render_accept_terms(
449 view_kind: LanguageModelProviderTosView,
450 accept_terms_of_service_in_progress: bool,
451 accept_terms_callback: impl Fn(&mut Window, &mut App) + 'static,
452) -> impl IntoElement {
453 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
454 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadEmptyState);
455
456 let terms_button = Button::new("terms_of_service", "Terms of Service")
457 .style(ButtonStyle::Subtle)
458 .icon(IconName::ArrowUpRight)
459 .icon_color(Color::Muted)
460 .icon_size(IconSize::XSmall)
461 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
462 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
463
464 let button_container = h_flex().child(
465 Button::new("accept_terms", "I accept the Terms of Service")
466 .when(!thread_empty_state, |this| {
467 this.full_width()
468 .style(ButtonStyle::Tinted(TintColor::Accent))
469 .icon(IconName::Check)
470 .icon_position(IconPosition::Start)
471 .icon_size(IconSize::Small)
472 })
473 .when(thread_empty_state, |this| {
474 this.style(ButtonStyle::Tinted(TintColor::Warning))
475 .label_size(LabelSize::Small)
476 })
477 .disabled(accept_terms_of_service_in_progress)
478 .on_click(move |_, window, cx| (accept_terms_callback)(window, cx)),
479 );
480
481 if thread_empty_state {
482 h_flex()
483 .w_full()
484 .flex_wrap()
485 .justify_between()
486 .child(
487 h_flex()
488 .child(
489 Label::new("To start using Zed AI, please read and accept the")
490 .size(LabelSize::Small),
491 )
492 .child(terms_button),
493 )
494 .child(button_container)
495 } else {
496 v_flex()
497 .w_full()
498 .gap_2()
499 .child(
500 h_flex()
501 .flex_wrap()
502 .when(thread_fresh_start, |this| this.justify_center())
503 .child(Label::new(
504 "To start using Zed AI, please read and accept the",
505 ))
506 .child(terms_button),
507 )
508 .child({
509 match view_kind {
510 LanguageModelProviderTosView::PromptEditorPopup => {
511 button_container.w_full().justify_end()
512 }
513 LanguageModelProviderTosView::Configuration => {
514 button_container.w_full().justify_start()
515 }
516 LanguageModelProviderTosView::ThreadFreshStart => {
517 button_container.w_full().justify_center()
518 }
519 LanguageModelProviderTosView::ThreadEmptyState => div().w_0(),
520 }
521 })
522 }
523}
524
525pub struct CloudLanguageModel {
526 id: LanguageModelId,
527 model: Arc<zed_llm_client::LanguageModel>,
528 llm_api_token: LlmApiToken,
529 client: Arc<Client>,
530 request_limiter: RateLimiter,
531}
532
533struct PerformLlmCompletionResponse {
534 response: Response<AsyncBody>,
535 usage: Option<ModelRequestUsage>,
536 tool_use_limit_reached: bool,
537 includes_status_messages: bool,
538}
539
540impl CloudLanguageModel {
541 async fn perform_llm_completion(
542 client: Arc<Client>,
543 llm_api_token: LlmApiToken,
544 app_version: Option<SemanticVersion>,
545 body: CompletionBody,
546 use_cloud: bool,
547 ) -> Result<PerformLlmCompletionResponse> {
548 let http_client = &client.http_client();
549
550 let mut token = llm_api_token.acquire(&client).await?;
551 let mut refreshed_token = false;
552
553 loop {
554 let request_builder = http_client::Request::builder().method(Method::POST).uri(
555 http_client
556 .build_zed_llm_url("/completions", &[], use_cloud)?
557 .as_ref(),
558 );
559 let request_builder = if let Some(app_version) = app_version {
560 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
561 } else {
562 request_builder
563 };
564
565 let request = request_builder
566 .header("Content-Type", "application/json")
567 .header("Authorization", format!("Bearer {token}"))
568 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
569 .body(serde_json::to_string(&body)?.into())?;
570 let mut response = http_client.send(request).await?;
571 let status = response.status();
572 if status.is_success() {
573 let includes_status_messages = response
574 .headers()
575 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
576 .is_some();
577
578 let tool_use_limit_reached = response
579 .headers()
580 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
581 .is_some();
582
583 let usage = if includes_status_messages {
584 None
585 } else {
586 ModelRequestUsage::from_headers(response.headers()).ok()
587 };
588
589 return Ok(PerformLlmCompletionResponse {
590 response,
591 usage,
592 includes_status_messages,
593 tool_use_limit_reached,
594 });
595 }
596
597 if !refreshed_token
598 && response
599 .headers()
600 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
601 .is_some()
602 {
603 token = llm_api_token.refresh(&client).await?;
604 refreshed_token = true;
605 continue;
606 }
607
608 if status == StatusCode::FORBIDDEN
609 && response
610 .headers()
611 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
612 .is_some()
613 {
614 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
615 .headers()
616 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
617 .and_then(|resource| resource.to_str().ok())
618 {
619 if let Some(plan) = response
620 .headers()
621 .get(CURRENT_PLAN_HEADER_NAME)
622 .and_then(|plan| plan.to_str().ok())
623 .and_then(|plan| zed_llm_client::Plan::from_str(plan).ok())
624 {
625 let plan = match plan {
626 zed_llm_client::Plan::ZedFree => Plan::Free,
627 zed_llm_client::Plan::ZedPro => Plan::ZedPro,
628 zed_llm_client::Plan::ZedProTrial => Plan::ZedProTrial,
629 };
630 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
631 }
632 }
633 } else if status == StatusCode::PAYMENT_REQUIRED {
634 return Err(anyhow!(PaymentRequiredError));
635 }
636
637 let mut body = String::new();
638 let headers = response.headers().clone();
639 response.body_mut().read_to_string(&mut body).await?;
640 return Err(anyhow!(ApiError {
641 status,
642 body,
643 headers
644 }));
645 }
646 }
647}
648
649#[derive(Debug, Error)]
650#[error("cloud language model request failed with status {status}: {body}")]
651struct ApiError {
652 status: StatusCode,
653 body: String,
654 headers: HeaderMap<HeaderValue>,
655}
656
657/// Represents error responses from Zed's cloud API.
658///
659/// Example JSON for an upstream HTTP error:
660/// ```json
661/// {
662/// "code": "upstream_http_error",
663/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
664/// "upstream_status": 503
665/// }
666/// ```
667#[derive(Debug, serde::Deserialize)]
668struct CloudApiError {
669 code: String,
670 message: String,
671 #[serde(default)]
672 #[serde(deserialize_with = "deserialize_optional_status_code")]
673 upstream_status: Option<StatusCode>,
674 #[serde(default)]
675 retry_after: Option<f64>,
676}
677
678fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
679where
680 D: serde::Deserializer<'de>,
681{
682 let opt: Option<u16> = Option::deserialize(deserializer)?;
683 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
684}
685
686impl From<ApiError> for LanguageModelCompletionError {
687 fn from(error: ApiError) -> Self {
688 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
689 if cloud_error.code.starts_with("upstream_http_") {
690 let status = if let Some(status) = cloud_error.upstream_status {
691 status
692 } else if cloud_error.code.ends_with("_error") {
693 error.status
694 } else {
695 // If there's a status code in the code string (e.g. "upstream_http_429")
696 // then use that; otherwise, see if the JSON contains a status code.
697 cloud_error
698 .code
699 .strip_prefix("upstream_http_")
700 .and_then(|code_str| code_str.parse::<u16>().ok())
701 .and_then(|code| StatusCode::from_u16(code).ok())
702 .unwrap_or(error.status)
703 };
704
705 return LanguageModelCompletionError::UpstreamProviderError {
706 message: cloud_error.message,
707 status,
708 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
709 };
710 }
711 }
712
713 let retry_after = None;
714 LanguageModelCompletionError::from_http_status(
715 PROVIDER_NAME,
716 error.status,
717 error.body,
718 retry_after,
719 )
720 }
721}
722
723impl LanguageModel for CloudLanguageModel {
724 fn id(&self) -> LanguageModelId {
725 self.id.clone()
726 }
727
728 fn name(&self) -> LanguageModelName {
729 LanguageModelName::from(self.model.display_name.clone())
730 }
731
732 fn provider_id(&self) -> LanguageModelProviderId {
733 PROVIDER_ID
734 }
735
736 fn provider_name(&self) -> LanguageModelProviderName {
737 PROVIDER_NAME
738 }
739
740 fn upstream_provider_id(&self) -> LanguageModelProviderId {
741 use zed_llm_client::LanguageModelProvider::*;
742 match self.model.provider {
743 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
744 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
745 Google => language_model::GOOGLE_PROVIDER_ID,
746 }
747 }
748
749 fn upstream_provider_name(&self) -> LanguageModelProviderName {
750 use zed_llm_client::LanguageModelProvider::*;
751 match self.model.provider {
752 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
753 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
754 Google => language_model::GOOGLE_PROVIDER_NAME,
755 }
756 }
757
758 fn supports_tools(&self) -> bool {
759 self.model.supports_tools
760 }
761
762 fn supports_images(&self) -> bool {
763 self.model.supports_images
764 }
765
766 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
767 match choice {
768 LanguageModelToolChoice::Auto
769 | LanguageModelToolChoice::Any
770 | LanguageModelToolChoice::None => true,
771 }
772 }
773
774 fn supports_burn_mode(&self) -> bool {
775 self.model.supports_max_mode
776 }
777
778 fn telemetry_id(&self) -> String {
779 format!("zed.dev/{}", self.model.id)
780 }
781
782 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
783 match self.model.provider {
784 zed_llm_client::LanguageModelProvider::Anthropic
785 | zed_llm_client::LanguageModelProvider::OpenAi => {
786 LanguageModelToolSchemaFormat::JsonSchema
787 }
788 zed_llm_client::LanguageModelProvider::Google => {
789 LanguageModelToolSchemaFormat::JsonSchemaSubset
790 }
791 }
792 }
793
794 fn max_token_count(&self) -> u64 {
795 self.model.max_token_count as u64
796 }
797
798 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
799 self.model
800 .max_token_count_in_max_mode
801 .filter(|_| self.model.supports_max_mode)
802 .map(|max_token_count| max_token_count as u64)
803 }
804
805 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
806 match &self.model.provider {
807 zed_llm_client::LanguageModelProvider::Anthropic => {
808 Some(LanguageModelCacheConfiguration {
809 min_total_token: 2_048,
810 should_speculate: true,
811 max_cache_anchors: 4,
812 })
813 }
814 zed_llm_client::LanguageModelProvider::OpenAi
815 | zed_llm_client::LanguageModelProvider::Google => None,
816 }
817 }
818
819 fn count_tokens(
820 &self,
821 request: LanguageModelRequest,
822 cx: &App,
823 ) -> BoxFuture<'static, Result<u64>> {
824 match self.model.provider {
825 zed_llm_client::LanguageModelProvider::Anthropic => count_anthropic_tokens(request, cx),
826 zed_llm_client::LanguageModelProvider::OpenAi => {
827 let model = match open_ai::Model::from_id(&self.model.id.0) {
828 Ok(model) => model,
829 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
830 };
831 count_open_ai_tokens(request, model, cx)
832 }
833 zed_llm_client::LanguageModelProvider::Google => {
834 let client = self.client.clone();
835 let llm_api_token = self.llm_api_token.clone();
836 let model_id = self.model.id.to_string();
837 let generate_content_request =
838 into_google(request, model_id.clone(), GoogleModelMode::Default);
839 let use_cloud = cx.has_flag::<ZedCloudFeatureFlag>();
840 async move {
841 let http_client = &client.http_client();
842 let token = llm_api_token.acquire(&client).await?;
843
844 let request_body = CountTokensBody {
845 provider: zed_llm_client::LanguageModelProvider::Google,
846 model: model_id,
847 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
848 generate_content_request,
849 })?,
850 };
851 let request = http_client::Request::builder()
852 .method(Method::POST)
853 .uri(
854 http_client
855 .build_zed_llm_url("/count_tokens", &[], use_cloud)?
856 .as_ref(),
857 )
858 .header("Content-Type", "application/json")
859 .header("Authorization", format!("Bearer {token}"))
860 .body(serde_json::to_string(&request_body)?.into())?;
861 let mut response = http_client.send(request).await?;
862 let status = response.status();
863 let headers = response.headers().clone();
864 let mut response_body = String::new();
865 response
866 .body_mut()
867 .read_to_string(&mut response_body)
868 .await?;
869
870 if status.is_success() {
871 let response_body: CountTokensResponse =
872 serde_json::from_str(&response_body)?;
873
874 Ok(response_body.tokens as u64)
875 } else {
876 Err(anyhow!(ApiError {
877 status,
878 body: response_body,
879 headers
880 }))
881 }
882 }
883 .boxed()
884 }
885 }
886 }
887
888 fn stream_completion(
889 &self,
890 request: LanguageModelRequest,
891 cx: &AsyncApp,
892 ) -> BoxFuture<
893 'static,
894 Result<
895 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
896 LanguageModelCompletionError,
897 >,
898 > {
899 let thread_id = request.thread_id.clone();
900 let prompt_id = request.prompt_id.clone();
901 let intent = request.intent;
902 let mode = request.mode;
903 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
904 let use_cloud = cx
905 .update(|cx| cx.has_flag::<ZedCloudFeatureFlag>())
906 .unwrap_or(false);
907 match self.model.provider {
908 zed_llm_client::LanguageModelProvider::Anthropic => {
909 let request = into_anthropic(
910 request,
911 self.model.id.to_string(),
912 1.0,
913 self.model.max_output_tokens as u64,
914 if self.model.id.0.ends_with("-thinking") {
915 AnthropicModelMode::Thinking {
916 budget_tokens: Some(4_096),
917 }
918 } else {
919 AnthropicModelMode::Default
920 },
921 );
922 let client = self.client.clone();
923 let llm_api_token = self.llm_api_token.clone();
924 let future = self.request_limiter.stream(async move {
925 let PerformLlmCompletionResponse {
926 response,
927 usage,
928 includes_status_messages,
929 tool_use_limit_reached,
930 } = Self::perform_llm_completion(
931 client.clone(),
932 llm_api_token,
933 app_version,
934 CompletionBody {
935 thread_id,
936 prompt_id,
937 intent,
938 mode,
939 provider: zed_llm_client::LanguageModelProvider::Anthropic,
940 model: request.model.clone(),
941 provider_request: serde_json::to_value(&request)
942 .map_err(|e| anyhow!(e))?,
943 },
944 use_cloud,
945 )
946 .await
947 .map_err(|err| match err.downcast::<ApiError>() {
948 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
949 Err(err) => anyhow!(err),
950 })?;
951
952 let mut mapper = AnthropicEventMapper::new();
953 Ok(map_cloud_completion_events(
954 Box::pin(
955 response_lines(response, includes_status_messages)
956 .chain(usage_updated_event(usage))
957 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
958 ),
959 move |event| mapper.map_event(event),
960 ))
961 });
962 async move { Ok(future.await?.boxed()) }.boxed()
963 }
964 zed_llm_client::LanguageModelProvider::OpenAi => {
965 let client = self.client.clone();
966 let model = match open_ai::Model::from_id(&self.model.id.0) {
967 Ok(model) => model,
968 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
969 };
970 let request = into_open_ai(
971 request,
972 model.id(),
973 model.supports_parallel_tool_calls(),
974 None,
975 );
976 let llm_api_token = self.llm_api_token.clone();
977 let future = self.request_limiter.stream(async move {
978 let PerformLlmCompletionResponse {
979 response,
980 usage,
981 includes_status_messages,
982 tool_use_limit_reached,
983 } = Self::perform_llm_completion(
984 client.clone(),
985 llm_api_token,
986 app_version,
987 CompletionBody {
988 thread_id,
989 prompt_id,
990 intent,
991 mode,
992 provider: zed_llm_client::LanguageModelProvider::OpenAi,
993 model: request.model.clone(),
994 provider_request: serde_json::to_value(&request)
995 .map_err(|e| anyhow!(e))?,
996 },
997 use_cloud,
998 )
999 .await?;
1000
1001 let mut mapper = OpenAiEventMapper::new();
1002 Ok(map_cloud_completion_events(
1003 Box::pin(
1004 response_lines(response, includes_status_messages)
1005 .chain(usage_updated_event(usage))
1006 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
1007 ),
1008 move |event| mapper.map_event(event),
1009 ))
1010 });
1011 async move { Ok(future.await?.boxed()) }.boxed()
1012 }
1013 zed_llm_client::LanguageModelProvider::Google => {
1014 let client = self.client.clone();
1015 let request =
1016 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
1017 let llm_api_token = self.llm_api_token.clone();
1018 let future = self.request_limiter.stream(async move {
1019 let PerformLlmCompletionResponse {
1020 response,
1021 usage,
1022 includes_status_messages,
1023 tool_use_limit_reached,
1024 } = Self::perform_llm_completion(
1025 client.clone(),
1026 llm_api_token,
1027 app_version,
1028 CompletionBody {
1029 thread_id,
1030 prompt_id,
1031 intent,
1032 mode,
1033 provider: zed_llm_client::LanguageModelProvider::Google,
1034 model: request.model.model_id.clone(),
1035 provider_request: serde_json::to_value(&request)
1036 .map_err(|e| anyhow!(e))?,
1037 },
1038 use_cloud,
1039 )
1040 .await?;
1041
1042 let mut mapper = GoogleEventMapper::new();
1043 Ok(map_cloud_completion_events(
1044 Box::pin(
1045 response_lines(response, includes_status_messages)
1046 .chain(usage_updated_event(usage))
1047 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
1048 ),
1049 move |event| mapper.map_event(event),
1050 ))
1051 });
1052 async move { Ok(future.await?.boxed()) }.boxed()
1053 }
1054 }
1055 }
1056}
1057
1058#[derive(Serialize, Deserialize)]
1059#[serde(rename_all = "snake_case")]
1060pub enum CloudCompletionEvent<T> {
1061 Status(CompletionRequestStatus),
1062 Event(T),
1063}
1064
1065fn map_cloud_completion_events<T, F>(
1066 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
1067 mut map_callback: F,
1068) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1069where
1070 T: DeserializeOwned + 'static,
1071 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1072 + Send
1073 + 'static,
1074{
1075 stream
1076 .flat_map(move |event| {
1077 futures::stream::iter(match event {
1078 Err(error) => {
1079 vec![Err(LanguageModelCompletionError::from(error))]
1080 }
1081 Ok(CloudCompletionEvent::Status(event)) => {
1082 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
1083 }
1084 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
1085 })
1086 })
1087 .boxed()
1088}
1089
1090fn usage_updated_event<T>(
1091 usage: Option<ModelRequestUsage>,
1092) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1093 futures::stream::iter(usage.map(|usage| {
1094 Ok(CloudCompletionEvent::Status(
1095 CompletionRequestStatus::UsageUpdated {
1096 amount: usage.amount as usize,
1097 limit: usage.limit,
1098 },
1099 ))
1100 }))
1101}
1102
1103fn tool_use_limit_reached_event<T>(
1104 tool_use_limit_reached: bool,
1105) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1106 futures::stream::iter(tool_use_limit_reached.then(|| {
1107 Ok(CloudCompletionEvent::Status(
1108 CompletionRequestStatus::ToolUseLimitReached,
1109 ))
1110 }))
1111}
1112
1113fn response_lines<T: DeserializeOwned>(
1114 response: Response<AsyncBody>,
1115 includes_status_messages: bool,
1116) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1117 futures::stream::try_unfold(
1118 (String::new(), BufReader::new(response.into_body())),
1119 move |(mut line, mut body)| async move {
1120 match body.read_line(&mut line).await {
1121 Ok(0) => Ok(None),
1122 Ok(_) => {
1123 let event = if includes_status_messages {
1124 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
1125 } else {
1126 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1127 };
1128
1129 line.clear();
1130 Ok(Some((event, (line, body))))
1131 }
1132 Err(e) => Err(e.into()),
1133 }
1134 },
1135 )
1136}
1137
1138#[derive(IntoElement, RegisterComponent)]
1139struct ZedAiConfiguration {
1140 is_connected: bool,
1141 plan: Option<proto::Plan>,
1142 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1143 eligible_for_trial: bool,
1144 has_accepted_terms_of_service: bool,
1145 accept_terms_of_service_in_progress: bool,
1146 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1147 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1148}
1149
1150impl RenderOnce for ZedAiConfiguration {
1151 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1152 const ZED_PRICING_URL: &str = "https://zed.dev/pricing";
1153
1154 let is_pro = self.plan == Some(proto::Plan::ZedPro);
1155 let subscription_text = match (self.plan, self.subscription_period) {
1156 (Some(proto::Plan::ZedPro), Some(_)) => {
1157 "You have access to Zed's hosted LLMs through your Zed Pro subscription."
1158 }
1159 (Some(proto::Plan::ZedProTrial), Some(_)) => {
1160 "You have access to Zed's hosted LLMs through your Zed Pro trial."
1161 }
1162 (Some(proto::Plan::Free), Some(_)) => {
1163 "You have basic access to Zed's hosted LLMs through your Zed Free subscription."
1164 }
1165 _ => {
1166 if self.eligible_for_trial {
1167 "Subscribe for access to Zed's hosted LLMs. Start with a 14 day free trial."
1168 } else {
1169 "Subscribe for access to Zed's hosted LLMs."
1170 }
1171 }
1172 };
1173 let manage_subscription_buttons = if is_pro {
1174 h_flex().child(
1175 Button::new("manage_settings", "Manage Subscription")
1176 .style(ButtonStyle::Tinted(TintColor::Accent))
1177 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1178 )
1179 } else {
1180 h_flex()
1181 .gap_2()
1182 .child(
1183 Button::new("learn_more", "Learn more")
1184 .style(ButtonStyle::Subtle)
1185 .on_click(|_, _, cx| cx.open_url(ZED_PRICING_URL)),
1186 )
1187 .child(
1188 Button::new(
1189 "upgrade",
1190 if self.plan.is_none() && self.eligible_for_trial {
1191 "Start Trial"
1192 } else {
1193 "Upgrade"
1194 },
1195 )
1196 .style(ButtonStyle::Subtle)
1197 .color(Color::Accent)
1198 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1199 )
1200 };
1201
1202 if self.is_connected {
1203 v_flex()
1204 .gap_3()
1205 .w_full()
1206 .when(!self.has_accepted_terms_of_service, |this| {
1207 this.child(render_accept_terms(
1208 LanguageModelProviderTosView::Configuration,
1209 self.accept_terms_of_service_in_progress,
1210 {
1211 let callback = self.accept_terms_of_service_callback.clone();
1212 move |window, cx| (callback)(window, cx)
1213 },
1214 ))
1215 })
1216 .when(self.has_accepted_terms_of_service, |this| {
1217 this.child(subscription_text)
1218 .child(manage_subscription_buttons)
1219 })
1220 } else {
1221 v_flex()
1222 .gap_2()
1223 .child(Label::new("Use Zed AI to access hosted language models."))
1224 .child(
1225 Button::new("sign_in", "Sign In")
1226 .icon_color(Color::Muted)
1227 .icon(IconName::Github)
1228 .icon_position(IconPosition::Start)
1229 .on_click({
1230 let callback = self.sign_in_callback.clone();
1231 move |_, window, cx| (callback)(window, cx)
1232 }),
1233 )
1234 }
1235 }
1236}
1237
1238struct ConfigurationView {
1239 state: Entity<State>,
1240 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1241 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1242}
1243
1244impl ConfigurationView {
1245 fn new(state: Entity<State>) -> Self {
1246 let accept_terms_of_service_callback = Arc::new({
1247 let state = state.clone();
1248 move |_window: &mut Window, cx: &mut App| {
1249 state.update(cx, |state, cx| {
1250 state.accept_terms_of_service(cx);
1251 });
1252 }
1253 });
1254
1255 let sign_in_callback = Arc::new({
1256 let state = state.clone();
1257 move |_window: &mut Window, cx: &mut App| {
1258 state.update(cx, |state, cx| {
1259 state.authenticate(cx).detach_and_log_err(cx);
1260 });
1261 }
1262 });
1263
1264 Self {
1265 state,
1266 accept_terms_of_service_callback,
1267 sign_in_callback,
1268 }
1269 }
1270}
1271
1272impl Render for ConfigurationView {
1273 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1274 let state = self.state.read(cx);
1275 let user_store = state.user_store.read(cx);
1276
1277 ZedAiConfiguration {
1278 is_connected: !state.is_signed_out(),
1279 plan: user_store.current_plan(),
1280 subscription_period: user_store.subscription_period(),
1281 eligible_for_trial: user_store.trial_started_at().is_none(),
1282 has_accepted_terms_of_service: state.has_accepted_terms_of_service(cx),
1283 accept_terms_of_service_in_progress: state.accept_terms_of_service_task.is_some(),
1284 accept_terms_of_service_callback: self.accept_terms_of_service_callback.clone(),
1285 sign_in_callback: self.sign_in_callback.clone(),
1286 }
1287 }
1288}
1289
1290impl Component for ZedAiConfiguration {
1291 fn scope() -> ComponentScope {
1292 ComponentScope::Agent
1293 }
1294
1295 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1296 fn configuration(
1297 is_connected: bool,
1298 plan: Option<proto::Plan>,
1299 eligible_for_trial: bool,
1300 has_accepted_terms_of_service: bool,
1301 ) -> AnyElement {
1302 ZedAiConfiguration {
1303 is_connected,
1304 plan,
1305 subscription_period: plan
1306 .is_some()
1307 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1308 eligible_for_trial,
1309 has_accepted_terms_of_service,
1310 accept_terms_of_service_in_progress: false,
1311 accept_terms_of_service_callback: Arc::new(|_, _| {}),
1312 sign_in_callback: Arc::new(|_, _| {}),
1313 }
1314 .into_any_element()
1315 }
1316
1317 Some(
1318 v_flex()
1319 .p_4()
1320 .gap_4()
1321 .children(vec![
1322 single_example("Not connected", configuration(false, None, false, true)),
1323 single_example(
1324 "Accept Terms of Service",
1325 configuration(true, None, true, false),
1326 ),
1327 single_example(
1328 "No Plan - Not eligible for trial",
1329 configuration(true, None, false, true),
1330 ),
1331 single_example(
1332 "No Plan - Eligible for trial",
1333 configuration(true, None, true, true),
1334 ),
1335 single_example(
1336 "Free Plan",
1337 configuration(true, Some(proto::Plan::Free), true, true),
1338 ),
1339 single_example(
1340 "Zed Pro Trial Plan",
1341 configuration(true, Some(proto::Plan::ZedProTrial), true, true),
1342 ),
1343 single_example(
1344 "Zed Pro Plan",
1345 configuration(true, Some(proto::Plan::ZedPro), true, true),
1346 ),
1347 ])
1348 .into_any_element(),
1349 )
1350 }
1351}
1352
1353#[cfg(test)]
1354mod tests {
1355 use super::*;
1356 use http_client::http::{HeaderMap, StatusCode};
1357 use language_model::LanguageModelCompletionError;
1358
1359 #[test]
1360 fn test_api_error_conversion_with_upstream_http_error() {
1361 // upstream_http_error with 503 status should become ServerOverloaded
1362 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}"#;
1363
1364 let api_error = ApiError {
1365 status: StatusCode::INTERNAL_SERVER_ERROR,
1366 body: error_body.to_string(),
1367 headers: HeaderMap::new(),
1368 };
1369
1370 let completion_error: LanguageModelCompletionError = api_error.into();
1371
1372 match completion_error {
1373 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1374 assert_eq!(
1375 message,
1376 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1377 );
1378 }
1379 _ => panic!(
1380 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1381 completion_error
1382 ),
1383 }
1384
1385 // upstream_http_error with 500 status should become ApiInternalServerError
1386 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1387
1388 let api_error = ApiError {
1389 status: StatusCode::INTERNAL_SERVER_ERROR,
1390 body: error_body.to_string(),
1391 headers: HeaderMap::new(),
1392 };
1393
1394 let completion_error: LanguageModelCompletionError = api_error.into();
1395
1396 match completion_error {
1397 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1398 assert_eq!(
1399 message,
1400 "Received an error from the OpenAI API: internal server error"
1401 );
1402 }
1403 _ => panic!(
1404 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1405 completion_error
1406 ),
1407 }
1408
1409 // upstream_http_error with 429 status should become RateLimitExceeded
1410 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1411
1412 let api_error = ApiError {
1413 status: StatusCode::INTERNAL_SERVER_ERROR,
1414 body: error_body.to_string(),
1415 headers: HeaderMap::new(),
1416 };
1417
1418 let completion_error: LanguageModelCompletionError = api_error.into();
1419
1420 match completion_error {
1421 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1422 assert_eq!(
1423 message,
1424 "Received an error from the Google API: rate limit exceeded"
1425 );
1426 }
1427 _ => panic!(
1428 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1429 completion_error
1430 ),
1431 }
1432
1433 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1434 let error_body = "Regular internal server error";
1435
1436 let api_error = ApiError {
1437 status: StatusCode::INTERNAL_SERVER_ERROR,
1438 body: error_body.to_string(),
1439 headers: HeaderMap::new(),
1440 };
1441
1442 let completion_error: LanguageModelCompletionError = api_error.into();
1443
1444 match completion_error {
1445 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1446 assert_eq!(provider, PROVIDER_NAME);
1447 assert_eq!(message, "Regular internal server error");
1448 }
1449 _ => panic!(
1450 "Expected ApiInternalServerError for regular 500, got: {:?}",
1451 completion_error
1452 ),
1453 }
1454
1455 // upstream_http_429 format should be converted to UpstreamProviderError
1456 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1457
1458 let api_error = ApiError {
1459 status: StatusCode::INTERNAL_SERVER_ERROR,
1460 body: error_body.to_string(),
1461 headers: HeaderMap::new(),
1462 };
1463
1464 let completion_error: LanguageModelCompletionError = api_error.into();
1465
1466 match completion_error {
1467 LanguageModelCompletionError::UpstreamProviderError {
1468 message,
1469 status,
1470 retry_after,
1471 } => {
1472 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1473 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1474 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1475 }
1476 _ => panic!(
1477 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1478 completion_error
1479 ),
1480 }
1481
1482 // Invalid JSON in error body should fall back to regular error handling
1483 let error_body = "Not JSON at all";
1484
1485 let api_error = ApiError {
1486 status: StatusCode::INTERNAL_SERVER_ERROR,
1487 body: error_body.to_string(),
1488 headers: HeaderMap::new(),
1489 };
1490
1491 let completion_error: LanguageModelCompletionError = api_error.into();
1492
1493 match completion_error {
1494 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1495 assert_eq!(provider, PROVIDER_NAME);
1496 }
1497 _ => panic!(
1498 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1499 completion_error
1500 ),
1501 }
1502 }
1503}