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