1use ai_onboarding::YoungAccountBanner;
2use anthropic::AnthropicModelMode;
3use anyhow::{Context as _, Result, anyhow};
4use chrono::{DateTime, Utc};
5use client::{Client, ModelRequestUsage, UserStore, zed_urls};
6use cloud_llm_client::{
7 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CLIENT_SUPPORTS_X_AI_HEADER_NAME,
8 CURRENT_PLAN_HEADER_NAME, CompletionBody, CompletionEvent, CompletionRequestStatus,
9 CountTokensBody, CountTokensResponse, EXPIRED_LLM_TOKEN_HEADER_NAME, ListModelsResponse,
10 MODEL_REQUESTS_RESOURCE_HEADER_VALUE, Plan, PlanV1, PlanV2,
11 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
12 TOOL_USE_LIMIT_REACHED_HEADER_NAME, ZED_VERSION_HEADER_NAME,
13};
14use futures::{
15 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
16};
17use google_ai::GoogleModelMode;
18use gpui::{AnyElement, AnyView, App, AsyncApp, Context, Entity, Subscription, Task};
19use http_client::http::{HeaderMap, HeaderValue};
20use http_client::{AsyncBody, HttpClient, HttpRequestExt, Method, Response, StatusCode};
21use language_model::{
22 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
23 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
24 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
25 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
26 LanguageModelToolSchemaFormat, LlmApiToken, ModelRequestLimitReachedError,
27 PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
28};
29use release_channel::AppVersion;
30use schemars::JsonSchema;
31use semver::Version;
32use serde::{Deserialize, Serialize, de::DeserializeOwned};
33use settings::SettingsStore;
34pub use settings::ZedDotDevAvailableModel as AvailableModel;
35pub use settings::ZedDotDevAvailableProvider as AvailableProvider;
36use smol::io::{AsyncReadExt, BufReader};
37use std::pin::Pin;
38use std::str::FromStr as _;
39use std::sync::Arc;
40use std::time::Duration;
41use thiserror::Error;
42use ui::{TintColor, prelude::*};
43use util::{ResultExt as _, maybe};
44
45use crate::provider::anthropic::{
46 AnthropicEventMapper, count_anthropic_tokens_with_tiktoken, into_anthropic,
47};
48use crate::provider::google::{GoogleEventMapper, into_google};
49use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
50use crate::provider::x_ai::count_xai_tokens;
51
52const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
53const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
54
55#[derive(Default, Clone, Debug, PartialEq)]
56pub struct ZedDotDevSettings {
57 pub available_models: Vec<AvailableModel>,
58}
59#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
60#[serde(tag = "type", rename_all = "lowercase")]
61pub enum ModelMode {
62 #[default]
63 Default,
64 Thinking {
65 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
66 budget_tokens: Option<u32>,
67 },
68}
69
70impl From<ModelMode> for AnthropicModelMode {
71 fn from(value: ModelMode) -> Self {
72 match value {
73 ModelMode::Default => AnthropicModelMode::Default,
74 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
75 }
76 }
77}
78
79pub struct CloudLanguageModelProvider {
80 client: Arc<Client>,
81 state: Entity<State>,
82 _maintain_client_status: Task<()>,
83}
84
85pub struct State {
86 client: Arc<Client>,
87 llm_api_token: LlmApiToken,
88 user_store: Entity<UserStore>,
89 status: client::Status,
90 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
91 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
92 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
93 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
94 _fetch_models_task: Task<()>,
95 _settings_subscription: Subscription,
96 _llm_token_subscription: Subscription,
97}
98
99impl State {
100 fn new(
101 client: Arc<Client>,
102 user_store: Entity<UserStore>,
103 status: client::Status,
104 cx: &mut Context<Self>,
105 ) -> Self {
106 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
107 let mut current_user = user_store.read(cx).watch_current_user();
108 Self {
109 client: client.clone(),
110 llm_api_token: LlmApiToken::default(),
111 user_store,
112 status,
113 models: Vec::new(),
114 default_model: None,
115 default_fast_model: None,
116 recommended_models: Vec::new(),
117 _fetch_models_task: cx.spawn(async move |this, cx| {
118 maybe!(async move {
119 let (client, llm_api_token) = this
120 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
121
122 while current_user.borrow().is_none() {
123 current_user.next().await;
124 }
125
126 let response =
127 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
128 this.update(cx, |this, cx| this.update_models(response, cx))?;
129 anyhow::Ok(())
130 })
131 .await
132 .context("failed to fetch Zed models")
133 .log_err();
134 }),
135 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
136 cx.notify();
137 }),
138 _llm_token_subscription: cx.subscribe(
139 &refresh_llm_token_listener,
140 move |this, _listener, _event, cx| {
141 let client = this.client.clone();
142 let llm_api_token = this.llm_api_token.clone();
143 cx.spawn(async move |this, cx| {
144 llm_api_token.refresh(&client).await?;
145 let response = Self::fetch_models(client, llm_api_token).await?;
146 this.update(cx, |this, cx| {
147 this.update_models(response, cx);
148 })
149 })
150 .detach_and_log_err(cx);
151 },
152 ),
153 }
154 }
155
156 fn is_signed_out(&self, cx: &App) -> bool {
157 self.user_store.read(cx).current_user().is_none()
158 }
159
160 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
161 let client = self.client.clone();
162 cx.spawn(async move |state, cx| {
163 client.sign_in_with_optional_connect(true, cx).await?;
164 state.update(cx, |_, cx| cx.notify())
165 })
166 }
167 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
168 let mut models = Vec::new();
169
170 for model in response.models {
171 models.push(Arc::new(model.clone()));
172
173 // Right now we represent thinking variants of models as separate models on the client,
174 // so we need to insert variants for any model that supports thinking.
175 if model.supports_thinking {
176 models.push(Arc::new(cloud_llm_client::LanguageModel {
177 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
178 display_name: format!("{} Thinking", model.display_name),
179 ..model
180 }));
181 }
182 }
183
184 self.default_model = models
185 .iter()
186 .find(|model| {
187 response
188 .default_model
189 .as_ref()
190 .is_some_and(|default_model_id| &model.id == default_model_id)
191 })
192 .cloned();
193 self.default_fast_model = models
194 .iter()
195 .find(|model| {
196 response
197 .default_fast_model
198 .as_ref()
199 .is_some_and(|default_fast_model_id| &model.id == default_fast_model_id)
200 })
201 .cloned();
202 self.recommended_models = response
203 .recommended_models
204 .iter()
205 .filter_map(|id| models.iter().find(|model| &model.id == id))
206 .cloned()
207 .collect();
208 self.models = models;
209 cx.notify();
210 }
211
212 async fn fetch_models(
213 client: Arc<Client>,
214 llm_api_token: LlmApiToken,
215 ) -> Result<ListModelsResponse> {
216 let http_client = &client.http_client();
217 let token = llm_api_token.acquire(&client).await?;
218
219 let request = http_client::Request::builder()
220 .method(Method::GET)
221 .header(CLIENT_SUPPORTS_X_AI_HEADER_NAME, "true")
222 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
223 .header("Authorization", format!("Bearer {token}"))
224 .body(AsyncBody::empty())?;
225 let mut response = http_client
226 .send(request)
227 .await
228 .context("failed to send list models request")?;
229
230 if response.status().is_success() {
231 let mut body = String::new();
232 response.body_mut().read_to_string(&mut body).await?;
233 Ok(serde_json::from_str(&body)?)
234 } else {
235 let mut body = String::new();
236 response.body_mut().read_to_string(&mut body).await?;
237 anyhow::bail!(
238 "error listing models.\nStatus: {:?}\nBody: {body}",
239 response.status(),
240 );
241 }
242 }
243}
244
245impl CloudLanguageModelProvider {
246 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
247 let mut status_rx = client.status();
248 let status = *status_rx.borrow();
249
250 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
251
252 let state_ref = state.downgrade();
253 let maintain_client_status = cx.spawn(async move |cx| {
254 while let Some(status) = status_rx.next().await {
255 if let Some(this) = state_ref.upgrade() {
256 _ = this.update(cx, |this, cx| {
257 if this.status != status {
258 this.status = status;
259 cx.notify();
260 }
261 });
262 } else {
263 break;
264 }
265 }
266 });
267
268 Self {
269 client,
270 state,
271 _maintain_client_status: maintain_client_status,
272 }
273 }
274
275 fn create_language_model(
276 &self,
277 model: Arc<cloud_llm_client::LanguageModel>,
278 llm_api_token: LlmApiToken,
279 ) -> Arc<dyn LanguageModel> {
280 Arc::new(CloudLanguageModel {
281 id: LanguageModelId(SharedString::from(model.id.0.clone())),
282 model,
283 llm_api_token,
284 client: self.client.clone(),
285 request_limiter: RateLimiter::new(4),
286 })
287 }
288}
289
290impl LanguageModelProviderState for CloudLanguageModelProvider {
291 type ObservableEntity = State;
292
293 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
294 Some(self.state.clone())
295 }
296}
297
298impl LanguageModelProvider for CloudLanguageModelProvider {
299 fn id(&self) -> LanguageModelProviderId {
300 PROVIDER_ID
301 }
302
303 fn name(&self) -> LanguageModelProviderName {
304 PROVIDER_NAME
305 }
306
307 fn icon(&self) -> IconName {
308 IconName::AiZed
309 }
310
311 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
312 let default_model = self.state.read(cx).default_model.clone()?;
313 let llm_api_token = self.state.read(cx).llm_api_token.clone();
314 Some(self.create_language_model(default_model, llm_api_token))
315 }
316
317 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
318 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
319 let llm_api_token = self.state.read(cx).llm_api_token.clone();
320 Some(self.create_language_model(default_fast_model, llm_api_token))
321 }
322
323 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
324 let llm_api_token = self.state.read(cx).llm_api_token.clone();
325 self.state
326 .read(cx)
327 .recommended_models
328 .iter()
329 .cloned()
330 .map(|model| self.create_language_model(model, llm_api_token.clone()))
331 .collect()
332 }
333
334 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
335 let llm_api_token = self.state.read(cx).llm_api_token.clone();
336 self.state
337 .read(cx)
338 .models
339 .iter()
340 .cloned()
341 .map(|model| self.create_language_model(model, llm_api_token.clone()))
342 .collect()
343 }
344
345 fn is_authenticated(&self, cx: &App) -> bool {
346 let state = self.state.read(cx);
347 !state.is_signed_out(cx)
348 }
349
350 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
351 Task::ready(Ok(()))
352 }
353
354 fn configuration_view(
355 &self,
356 _target_agent: language_model::ConfigurationViewTargetAgent,
357 _: &mut Window,
358 cx: &mut App,
359 ) -> AnyView {
360 cx.new(|_| ConfigurationView::new(self.state.clone()))
361 .into()
362 }
363
364 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
365 Task::ready(Ok(()))
366 }
367}
368
369pub struct CloudLanguageModel {
370 id: LanguageModelId,
371 model: Arc<cloud_llm_client::LanguageModel>,
372 llm_api_token: LlmApiToken,
373 client: Arc<Client>,
374 request_limiter: RateLimiter,
375}
376
377struct PerformLlmCompletionResponse {
378 response: Response<AsyncBody>,
379 usage: Option<ModelRequestUsage>,
380 tool_use_limit_reached: bool,
381 includes_status_messages: bool,
382}
383
384impl CloudLanguageModel {
385 async fn perform_llm_completion(
386 client: Arc<Client>,
387 llm_api_token: LlmApiToken,
388 app_version: Option<Version>,
389 body: CompletionBody,
390 ) -> Result<PerformLlmCompletionResponse> {
391 let http_client = &client.http_client();
392
393 let mut token = llm_api_token.acquire(&client).await?;
394 let mut refreshed_token = false;
395
396 loop {
397 let request = http_client::Request::builder()
398 .method(Method::POST)
399 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
400 .when_some(app_version.as_ref(), |builder, app_version| {
401 builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
402 })
403 .header("Content-Type", "application/json")
404 .header("Authorization", format!("Bearer {token}"))
405 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
406 .body(serde_json::to_string(&body)?.into())?;
407
408 let mut response = http_client.send(request).await?;
409 let status = response.status();
410 if status.is_success() {
411 let includes_status_messages = response
412 .headers()
413 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
414 .is_some();
415
416 let tool_use_limit_reached = response
417 .headers()
418 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
419 .is_some();
420
421 let usage = if includes_status_messages {
422 None
423 } else {
424 ModelRequestUsage::from_headers(response.headers()).ok()
425 };
426
427 return Ok(PerformLlmCompletionResponse {
428 response,
429 usage,
430 includes_status_messages,
431 tool_use_limit_reached,
432 });
433 }
434
435 if !refreshed_token
436 && response
437 .headers()
438 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
439 .is_some()
440 {
441 token = llm_api_token.refresh(&client).await?;
442 refreshed_token = true;
443 continue;
444 }
445
446 if status == StatusCode::FORBIDDEN
447 && response
448 .headers()
449 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
450 .is_some()
451 {
452 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
453 .headers()
454 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
455 .and_then(|resource| resource.to_str().ok())
456 && let Some(plan) = response
457 .headers()
458 .get(CURRENT_PLAN_HEADER_NAME)
459 .and_then(|plan| plan.to_str().ok())
460 .and_then(|plan| cloud_llm_client::PlanV1::from_str(plan).ok())
461 .map(Plan::V1)
462 {
463 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
464 }
465 } else if status == StatusCode::PAYMENT_REQUIRED {
466 return Err(anyhow!(PaymentRequiredError));
467 }
468
469 let mut body = String::new();
470 let headers = response.headers().clone();
471 response.body_mut().read_to_string(&mut body).await?;
472 return Err(anyhow!(ApiError {
473 status,
474 body,
475 headers
476 }));
477 }
478 }
479}
480
481#[derive(Debug, Error)]
482#[error("cloud language model request failed with status {status}: {body}")]
483struct ApiError {
484 status: StatusCode,
485 body: String,
486 headers: HeaderMap<HeaderValue>,
487}
488
489/// Represents error responses from Zed's cloud API.
490///
491/// Example JSON for an upstream HTTP error:
492/// ```json
493/// {
494/// "code": "upstream_http_error",
495/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
496/// "upstream_status": 503
497/// }
498/// ```
499#[derive(Debug, serde::Deserialize)]
500struct CloudApiError {
501 code: String,
502 message: String,
503 #[serde(default)]
504 #[serde(deserialize_with = "deserialize_optional_status_code")]
505 upstream_status: Option<StatusCode>,
506 #[serde(default)]
507 retry_after: Option<f64>,
508}
509
510fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
511where
512 D: serde::Deserializer<'de>,
513{
514 let opt: Option<u16> = Option::deserialize(deserializer)?;
515 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
516}
517
518impl From<ApiError> for LanguageModelCompletionError {
519 fn from(error: ApiError) -> Self {
520 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
521 if cloud_error.code.starts_with("upstream_http_") {
522 let status = if let Some(status) = cloud_error.upstream_status {
523 status
524 } else if cloud_error.code.ends_with("_error") {
525 error.status
526 } else {
527 // If there's a status code in the code string (e.g. "upstream_http_429")
528 // then use that; otherwise, see if the JSON contains a status code.
529 cloud_error
530 .code
531 .strip_prefix("upstream_http_")
532 .and_then(|code_str| code_str.parse::<u16>().ok())
533 .and_then(|code| StatusCode::from_u16(code).ok())
534 .unwrap_or(error.status)
535 };
536
537 return LanguageModelCompletionError::UpstreamProviderError {
538 message: cloud_error.message,
539 status,
540 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
541 };
542 }
543
544 return LanguageModelCompletionError::from_http_status(
545 PROVIDER_NAME,
546 error.status,
547 cloud_error.message,
548 None,
549 );
550 }
551
552 let retry_after = None;
553 LanguageModelCompletionError::from_http_status(
554 PROVIDER_NAME,
555 error.status,
556 error.body,
557 retry_after,
558 )
559 }
560}
561
562impl LanguageModel for CloudLanguageModel {
563 fn id(&self) -> LanguageModelId {
564 self.id.clone()
565 }
566
567 fn name(&self) -> LanguageModelName {
568 LanguageModelName::from(self.model.display_name.clone())
569 }
570
571 fn provider_id(&self) -> LanguageModelProviderId {
572 PROVIDER_ID
573 }
574
575 fn provider_name(&self) -> LanguageModelProviderName {
576 PROVIDER_NAME
577 }
578
579 fn upstream_provider_id(&self) -> LanguageModelProviderId {
580 use cloud_llm_client::LanguageModelProvider::*;
581 match self.model.provider {
582 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
583 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
584 Google => language_model::GOOGLE_PROVIDER_ID,
585 XAi => language_model::X_AI_PROVIDER_ID,
586 }
587 }
588
589 fn upstream_provider_name(&self) -> LanguageModelProviderName {
590 use cloud_llm_client::LanguageModelProvider::*;
591 match self.model.provider {
592 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
593 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
594 Google => language_model::GOOGLE_PROVIDER_NAME,
595 XAi => language_model::X_AI_PROVIDER_NAME,
596 }
597 }
598
599 fn supports_tools(&self) -> bool {
600 self.model.supports_tools
601 }
602
603 fn supports_images(&self) -> bool {
604 self.model.supports_images
605 }
606
607 fn supports_streaming_tools(&self) -> bool {
608 self.model.supports_streaming_tools
609 }
610
611 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
612 match choice {
613 LanguageModelToolChoice::Auto
614 | LanguageModelToolChoice::Any
615 | LanguageModelToolChoice::None => true,
616 }
617 }
618
619 fn supports_burn_mode(&self) -> bool {
620 self.model.supports_max_mode
621 }
622
623 fn telemetry_id(&self) -> String {
624 format!("zed.dev/{}", self.model.id)
625 }
626
627 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
628 match self.model.provider {
629 cloud_llm_client::LanguageModelProvider::Anthropic
630 | cloud_llm_client::LanguageModelProvider::OpenAi
631 | cloud_llm_client::LanguageModelProvider::XAi => {
632 LanguageModelToolSchemaFormat::JsonSchema
633 }
634 cloud_llm_client::LanguageModelProvider::Google => {
635 LanguageModelToolSchemaFormat::JsonSchemaSubset
636 }
637 }
638 }
639
640 fn max_token_count(&self) -> u64 {
641 self.model.max_token_count as u64
642 }
643
644 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
645 self.model
646 .max_token_count_in_max_mode
647 .filter(|_| self.model.supports_max_mode)
648 .map(|max_token_count| max_token_count as u64)
649 }
650
651 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
652 match &self.model.provider {
653 cloud_llm_client::LanguageModelProvider::Anthropic => {
654 Some(LanguageModelCacheConfiguration {
655 min_total_token: 2_048,
656 should_speculate: true,
657 max_cache_anchors: 4,
658 })
659 }
660 cloud_llm_client::LanguageModelProvider::OpenAi
661 | cloud_llm_client::LanguageModelProvider::XAi
662 | cloud_llm_client::LanguageModelProvider::Google => None,
663 }
664 }
665
666 fn count_tokens(
667 &self,
668 request: LanguageModelRequest,
669 cx: &App,
670 ) -> BoxFuture<'static, Result<u64>> {
671 match self.model.provider {
672 cloud_llm_client::LanguageModelProvider::Anthropic => cx
673 .background_spawn(async move { count_anthropic_tokens_with_tiktoken(request) })
674 .boxed(),
675 cloud_llm_client::LanguageModelProvider::OpenAi => {
676 let model = match open_ai::Model::from_id(&self.model.id.0) {
677 Ok(model) => model,
678 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
679 };
680 count_open_ai_tokens(request, model, cx)
681 }
682 cloud_llm_client::LanguageModelProvider::XAi => {
683 let model = match x_ai::Model::from_id(&self.model.id.0) {
684 Ok(model) => model,
685 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
686 };
687 count_xai_tokens(request, model, cx)
688 }
689 cloud_llm_client::LanguageModelProvider::Google => {
690 let client = self.client.clone();
691 let llm_api_token = self.llm_api_token.clone();
692 let model_id = self.model.id.to_string();
693 let generate_content_request =
694 into_google(request, model_id.clone(), GoogleModelMode::Default);
695 async move {
696 let http_client = &client.http_client();
697 let token = llm_api_token.acquire(&client).await?;
698
699 let request_body = CountTokensBody {
700 provider: cloud_llm_client::LanguageModelProvider::Google,
701 model: model_id,
702 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
703 generate_content_request,
704 })?,
705 };
706 let request = http_client::Request::builder()
707 .method(Method::POST)
708 .uri(
709 http_client
710 .build_zed_llm_url("/count_tokens", &[])?
711 .as_ref(),
712 )
713 .header("Content-Type", "application/json")
714 .header("Authorization", format!("Bearer {token}"))
715 .body(serde_json::to_string(&request_body)?.into())?;
716 let mut response = http_client.send(request).await?;
717 let status = response.status();
718 let headers = response.headers().clone();
719 let mut response_body = String::new();
720 response
721 .body_mut()
722 .read_to_string(&mut response_body)
723 .await?;
724
725 if status.is_success() {
726 let response_body: CountTokensResponse =
727 serde_json::from_str(&response_body)?;
728
729 Ok(response_body.tokens as u64)
730 } else {
731 Err(anyhow!(ApiError {
732 status,
733 body: response_body,
734 headers
735 }))
736 }
737 }
738 .boxed()
739 }
740 }
741 }
742
743 fn stream_completion(
744 &self,
745 request: LanguageModelRequest,
746 cx: &AsyncApp,
747 ) -> BoxFuture<
748 'static,
749 Result<
750 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
751 LanguageModelCompletionError,
752 >,
753 > {
754 let thread_id = request.thread_id.clone();
755 let prompt_id = request.prompt_id.clone();
756 let intent = request.intent;
757 let mode = request.mode;
758 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
759 let thinking_allowed = request.thinking_allowed;
760 let provider_name = provider_name(&self.model.provider);
761 match self.model.provider {
762 cloud_llm_client::LanguageModelProvider::Anthropic => {
763 let request = into_anthropic(
764 request,
765 self.model.id.to_string(),
766 1.0,
767 self.model.max_output_tokens as u64,
768 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
769 AnthropicModelMode::Thinking {
770 budget_tokens: Some(4_096),
771 }
772 } else {
773 AnthropicModelMode::Default
774 },
775 );
776 let client = self.client.clone();
777 let llm_api_token = self.llm_api_token.clone();
778 let future = self.request_limiter.stream(async move {
779 let PerformLlmCompletionResponse {
780 response,
781 usage,
782 includes_status_messages,
783 tool_use_limit_reached,
784 } = Self::perform_llm_completion(
785 client.clone(),
786 llm_api_token,
787 app_version,
788 CompletionBody {
789 thread_id,
790 prompt_id,
791 intent,
792 mode,
793 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
794 model: request.model.clone(),
795 provider_request: serde_json::to_value(&request)
796 .map_err(|e| anyhow!(e))?,
797 },
798 )
799 .await
800 .map_err(|err| match err.downcast::<ApiError>() {
801 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
802 Err(err) => anyhow!(err),
803 })?;
804
805 let mut mapper = AnthropicEventMapper::new();
806 Ok(map_cloud_completion_events(
807 Box::pin(
808 response_lines(response, includes_status_messages)
809 .chain(usage_updated_event(usage))
810 .chain(tool_use_limit_reached_event(tool_use_limit_reached)), // .map(|_| {}),
811 ),
812 &provider_name,
813 move |event| mapper.map_event(event),
814 ))
815 });
816 async move { Ok(future.await?.boxed()) }.boxed()
817 }
818 cloud_llm_client::LanguageModelProvider::OpenAi => {
819 let client = self.client.clone();
820 let request = into_open_ai(
821 request,
822 &self.model.id.0,
823 self.model.supports_parallel_tool_calls,
824 true,
825 None,
826 None,
827 );
828 let llm_api_token = self.llm_api_token.clone();
829 let future = self.request_limiter.stream(async move {
830 let PerformLlmCompletionResponse {
831 response,
832 usage,
833 includes_status_messages,
834 tool_use_limit_reached,
835 } = Self::perform_llm_completion(
836 client.clone(),
837 llm_api_token,
838 app_version,
839 CompletionBody {
840 thread_id,
841 prompt_id,
842 intent,
843 mode,
844 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
845 model: request.model.clone(),
846 provider_request: serde_json::to_value(&request)
847 .map_err(|e| anyhow!(e))?,
848 },
849 )
850 .await?;
851
852 let mut mapper = OpenAiEventMapper::new();
853 Ok(map_cloud_completion_events(
854 Box::pin(
855 response_lines(response, includes_status_messages)
856 .chain(usage_updated_event(usage))
857 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
858 ),
859 &provider_name,
860 move |event| mapper.map_event(event),
861 ))
862 });
863 async move { Ok(future.await?.boxed()) }.boxed()
864 }
865 cloud_llm_client::LanguageModelProvider::XAi => {
866 let client = self.client.clone();
867 let request = into_open_ai(
868 request,
869 &self.model.id.0,
870 self.model.supports_parallel_tool_calls,
871 false,
872 None,
873 None,
874 );
875 let llm_api_token = self.llm_api_token.clone();
876 let future = self.request_limiter.stream(async move {
877 let PerformLlmCompletionResponse {
878 response,
879 usage,
880 includes_status_messages,
881 tool_use_limit_reached,
882 } = Self::perform_llm_completion(
883 client.clone(),
884 llm_api_token,
885 app_version,
886 CompletionBody {
887 thread_id,
888 prompt_id,
889 intent,
890 mode,
891 provider: cloud_llm_client::LanguageModelProvider::XAi,
892 model: request.model.clone(),
893 provider_request: serde_json::to_value(&request)
894 .map_err(|e| anyhow!(e))?,
895 },
896 )
897 .await?;
898
899 let mut mapper = OpenAiEventMapper::new();
900 Ok(map_cloud_completion_events(
901 Box::pin(
902 response_lines(response, includes_status_messages)
903 .chain(usage_updated_event(usage))
904 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
905 ),
906 &provider_name,
907 move |event| mapper.map_event(event),
908 ))
909 });
910 async move { Ok(future.await?.boxed()) }.boxed()
911 }
912 cloud_llm_client::LanguageModelProvider::Google => {
913 let client = self.client.clone();
914 let request =
915 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
916 let llm_api_token = self.llm_api_token.clone();
917 let future = self.request_limiter.stream(async move {
918 let PerformLlmCompletionResponse {
919 response,
920 usage,
921 includes_status_messages,
922 tool_use_limit_reached,
923 } = Self::perform_llm_completion(
924 client.clone(),
925 llm_api_token,
926 app_version,
927 CompletionBody {
928 thread_id,
929 prompt_id,
930 intent,
931 mode,
932 provider: cloud_llm_client::LanguageModelProvider::Google,
933 model: request.model.model_id.clone(),
934 provider_request: serde_json::to_value(&request)
935 .map_err(|e| anyhow!(e))?,
936 },
937 )
938 .await?;
939
940 let mut mapper = GoogleEventMapper::new();
941 Ok(map_cloud_completion_events(
942 Box::pin(
943 response_lines(response, includes_status_messages)
944 .chain(usage_updated_event(usage))
945 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
946 ),
947 &provider_name,
948 move |event| mapper.map_event(event),
949 ))
950 });
951 async move { Ok(future.await?.boxed()) }.boxed()
952 }
953 }
954 }
955}
956
957fn map_cloud_completion_events<T, F>(
958 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
959 provider: &LanguageModelProviderName,
960 mut map_callback: F,
961) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
962where
963 T: DeserializeOwned + 'static,
964 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
965 + Send
966 + 'static,
967{
968 let provider = provider.clone();
969 stream
970 .flat_map(move |event| {
971 futures::stream::iter(match event {
972 Err(error) => {
973 vec![Err(LanguageModelCompletionError::from(error))]
974 }
975 Ok(CompletionEvent::Status(event)) => {
976 vec![
977 LanguageModelCompletionEvent::from_completion_request_status(
978 event,
979 provider.clone(),
980 ),
981 ]
982 }
983 Ok(CompletionEvent::Event(event)) => map_callback(event),
984 })
985 })
986 .boxed()
987}
988
989fn provider_name(provider: &cloud_llm_client::LanguageModelProvider) -> LanguageModelProviderName {
990 match provider {
991 cloud_llm_client::LanguageModelProvider::Anthropic => {
992 language_model::ANTHROPIC_PROVIDER_NAME
993 }
994 cloud_llm_client::LanguageModelProvider::OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
995 cloud_llm_client::LanguageModelProvider::Google => language_model::GOOGLE_PROVIDER_NAME,
996 cloud_llm_client::LanguageModelProvider::XAi => language_model::X_AI_PROVIDER_NAME,
997 }
998}
999
1000fn usage_updated_event<T>(
1001 usage: Option<ModelRequestUsage>,
1002) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1003 futures::stream::iter(usage.map(|usage| {
1004 Ok(CompletionEvent::Status(
1005 CompletionRequestStatus::UsageUpdated {
1006 amount: usage.amount as usize,
1007 limit: usage.limit,
1008 },
1009 ))
1010 }))
1011}
1012
1013fn tool_use_limit_reached_event<T>(
1014 tool_use_limit_reached: bool,
1015) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1016 futures::stream::iter(tool_use_limit_reached.then(|| {
1017 Ok(CompletionEvent::Status(
1018 CompletionRequestStatus::ToolUseLimitReached,
1019 ))
1020 }))
1021}
1022
1023fn response_lines<T: DeserializeOwned>(
1024 response: Response<AsyncBody>,
1025 includes_status_messages: bool,
1026) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1027 futures::stream::try_unfold(
1028 (String::new(), BufReader::new(response.into_body())),
1029 move |(mut line, mut body)| async move {
1030 match body.read_line(&mut line).await {
1031 Ok(0) => Ok(None),
1032 Ok(_) => {
1033 let event = if includes_status_messages {
1034 serde_json::from_str::<CompletionEvent<T>>(&line)?
1035 } else {
1036 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1037 };
1038
1039 line.clear();
1040 Ok(Some((event, (line, body))))
1041 }
1042 Err(e) => Err(e.into()),
1043 }
1044 },
1045 )
1046}
1047
1048#[derive(IntoElement, RegisterComponent)]
1049struct ZedAiConfiguration {
1050 is_connected: bool,
1051 plan: Option<Plan>,
1052 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1053 eligible_for_trial: bool,
1054 account_too_young: bool,
1055 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1056}
1057
1058impl RenderOnce for ZedAiConfiguration {
1059 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1060 let is_pro = self.plan.is_some_and(|plan| {
1061 matches!(plan, Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro))
1062 });
1063 let subscription_text = match (self.plan, self.subscription_period) {
1064 (Some(Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro)), Some(_)) => {
1065 "You have access to Zed's hosted models through your Pro subscription."
1066 }
1067 (Some(Plan::V1(PlanV1::ZedProTrial) | Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1068 "You have access to Zed's hosted models through your Pro trial."
1069 }
1070 (Some(Plan::V1(PlanV1::ZedFree)), Some(_)) => {
1071 "You have basic access to Zed's hosted models through the Free plan."
1072 }
1073 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1074 if self.eligible_for_trial {
1075 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1076 } else {
1077 "Subscribe for access to Zed's hosted models."
1078 }
1079 }
1080 _ => {
1081 if self.eligible_for_trial {
1082 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1083 } else {
1084 "Subscribe for access to Zed's hosted models."
1085 }
1086 }
1087 };
1088
1089 let manage_subscription_buttons = if is_pro {
1090 Button::new("manage_settings", "Manage Subscription")
1091 .full_width()
1092 .style(ButtonStyle::Tinted(TintColor::Accent))
1093 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1094 .into_any_element()
1095 } else if self.plan.is_none() || self.eligible_for_trial {
1096 Button::new("start_trial", "Start 14-day Free Pro Trial")
1097 .full_width()
1098 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1099 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1100 .into_any_element()
1101 } else {
1102 Button::new("upgrade", "Upgrade to Pro")
1103 .full_width()
1104 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1105 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1106 .into_any_element()
1107 };
1108
1109 if !self.is_connected {
1110 return v_flex()
1111 .gap_2()
1112 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1113 .child(
1114 Button::new("sign_in", "Sign In to use Zed AI")
1115 .icon_color(Color::Muted)
1116 .icon(IconName::Github)
1117 .icon_size(IconSize::Small)
1118 .icon_position(IconPosition::Start)
1119 .full_width()
1120 .on_click({
1121 let callback = self.sign_in_callback.clone();
1122 move |_, window, cx| (callback)(window, cx)
1123 }),
1124 );
1125 }
1126
1127 v_flex().gap_2().w_full().map(|this| {
1128 if self.account_too_young {
1129 this.child(YoungAccountBanner).child(
1130 Button::new("upgrade", "Upgrade to Pro")
1131 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1132 .full_width()
1133 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1134 )
1135 } else {
1136 this.text_sm()
1137 .child(subscription_text)
1138 .child(manage_subscription_buttons)
1139 }
1140 })
1141 }
1142}
1143
1144struct ConfigurationView {
1145 state: Entity<State>,
1146 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1147}
1148
1149impl ConfigurationView {
1150 fn new(state: Entity<State>) -> Self {
1151 let sign_in_callback = Arc::new({
1152 let state = state.clone();
1153 move |_window: &mut Window, cx: &mut App| {
1154 state.update(cx, |state, cx| {
1155 state.authenticate(cx).detach_and_log_err(cx);
1156 });
1157 }
1158 });
1159
1160 Self {
1161 state,
1162 sign_in_callback,
1163 }
1164 }
1165}
1166
1167impl Render for ConfigurationView {
1168 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1169 let state = self.state.read(cx);
1170 let user_store = state.user_store.read(cx);
1171
1172 ZedAiConfiguration {
1173 is_connected: !state.is_signed_out(cx),
1174 plan: user_store.plan(),
1175 subscription_period: user_store.subscription_period(),
1176 eligible_for_trial: user_store.trial_started_at().is_none(),
1177 account_too_young: user_store.account_too_young(),
1178 sign_in_callback: self.sign_in_callback.clone(),
1179 }
1180 }
1181}
1182
1183impl Component for ZedAiConfiguration {
1184 fn name() -> &'static str {
1185 "AI Configuration Content"
1186 }
1187
1188 fn sort_name() -> &'static str {
1189 "AI Configuration Content"
1190 }
1191
1192 fn scope() -> ComponentScope {
1193 ComponentScope::Onboarding
1194 }
1195
1196 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1197 fn configuration(
1198 is_connected: bool,
1199 plan: Option<Plan>,
1200 eligible_for_trial: bool,
1201 account_too_young: bool,
1202 ) -> AnyElement {
1203 ZedAiConfiguration {
1204 is_connected,
1205 plan,
1206 subscription_period: plan
1207 .is_some()
1208 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1209 eligible_for_trial,
1210 account_too_young,
1211 sign_in_callback: Arc::new(|_, _| {}),
1212 }
1213 .into_any_element()
1214 }
1215
1216 Some(
1217 v_flex()
1218 .p_4()
1219 .gap_4()
1220 .children(vec![
1221 single_example("Not connected", configuration(false, None, false, false)),
1222 single_example(
1223 "Accept Terms of Service",
1224 configuration(true, None, true, false),
1225 ),
1226 single_example(
1227 "No Plan - Not eligible for trial",
1228 configuration(true, None, false, false),
1229 ),
1230 single_example(
1231 "No Plan - Eligible for trial",
1232 configuration(true, None, true, false),
1233 ),
1234 single_example(
1235 "Free Plan",
1236 configuration(true, Some(Plan::V1(PlanV1::ZedFree)), true, false),
1237 ),
1238 single_example(
1239 "Zed Pro Trial Plan",
1240 configuration(true, Some(Plan::V1(PlanV1::ZedProTrial)), true, false),
1241 ),
1242 single_example(
1243 "Zed Pro Plan",
1244 configuration(true, Some(Plan::V1(PlanV1::ZedPro)), true, false),
1245 ),
1246 ])
1247 .into_any_element(),
1248 )
1249 }
1250}
1251
1252#[cfg(test)]
1253mod tests {
1254 use super::*;
1255 use http_client::http::{HeaderMap, StatusCode};
1256 use language_model::LanguageModelCompletionError;
1257
1258 #[test]
1259 fn test_api_error_conversion_with_upstream_http_error() {
1260 // upstream_http_error with 503 status should become ServerOverloaded
1261 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}"#;
1262
1263 let api_error = ApiError {
1264 status: StatusCode::INTERNAL_SERVER_ERROR,
1265 body: error_body.to_string(),
1266 headers: HeaderMap::new(),
1267 };
1268
1269 let completion_error: LanguageModelCompletionError = api_error.into();
1270
1271 match completion_error {
1272 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1273 assert_eq!(
1274 message,
1275 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1276 );
1277 }
1278 _ => panic!(
1279 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1280 completion_error
1281 ),
1282 }
1283
1284 // upstream_http_error with 500 status should become ApiInternalServerError
1285 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1286
1287 let api_error = ApiError {
1288 status: StatusCode::INTERNAL_SERVER_ERROR,
1289 body: error_body.to_string(),
1290 headers: HeaderMap::new(),
1291 };
1292
1293 let completion_error: LanguageModelCompletionError = api_error.into();
1294
1295 match completion_error {
1296 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1297 assert_eq!(
1298 message,
1299 "Received an error from the OpenAI API: internal server error"
1300 );
1301 }
1302 _ => panic!(
1303 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1304 completion_error
1305 ),
1306 }
1307
1308 // upstream_http_error with 429 status should become RateLimitExceeded
1309 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1310
1311 let api_error = ApiError {
1312 status: StatusCode::INTERNAL_SERVER_ERROR,
1313 body: error_body.to_string(),
1314 headers: HeaderMap::new(),
1315 };
1316
1317 let completion_error: LanguageModelCompletionError = api_error.into();
1318
1319 match completion_error {
1320 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1321 assert_eq!(
1322 message,
1323 "Received an error from the Google API: rate limit exceeded"
1324 );
1325 }
1326 _ => panic!(
1327 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1328 completion_error
1329 ),
1330 }
1331
1332 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1333 let error_body = "Regular internal server error";
1334
1335 let api_error = ApiError {
1336 status: StatusCode::INTERNAL_SERVER_ERROR,
1337 body: error_body.to_string(),
1338 headers: HeaderMap::new(),
1339 };
1340
1341 let completion_error: LanguageModelCompletionError = api_error.into();
1342
1343 match completion_error {
1344 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1345 assert_eq!(provider, PROVIDER_NAME);
1346 assert_eq!(message, "Regular internal server error");
1347 }
1348 _ => panic!(
1349 "Expected ApiInternalServerError for regular 500, got: {:?}",
1350 completion_error
1351 ),
1352 }
1353
1354 // upstream_http_429 format should be converted to UpstreamProviderError
1355 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1356
1357 let api_error = ApiError {
1358 status: StatusCode::INTERNAL_SERVER_ERROR,
1359 body: error_body.to_string(),
1360 headers: HeaderMap::new(),
1361 };
1362
1363 let completion_error: LanguageModelCompletionError = api_error.into();
1364
1365 match completion_error {
1366 LanguageModelCompletionError::UpstreamProviderError {
1367 message,
1368 status,
1369 retry_after,
1370 } => {
1371 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1372 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1373 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1374 }
1375 _ => panic!(
1376 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1377 completion_error
1378 ),
1379 }
1380
1381 // Invalid JSON in error body should fall back to regular error handling
1382 let error_body = "Not JSON at all";
1383
1384 let api_error = ApiError {
1385 status: StatusCode::INTERNAL_SERVER_ERROR,
1386 body: error_body.to_string(),
1387 headers: HeaderMap::new(),
1388 };
1389
1390 let completion_error: LanguageModelCompletionError = api_error.into();
1391
1392 match completion_error {
1393 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1394 assert_eq!(provider, PROVIDER_NAME);
1395 }
1396 _ => panic!(
1397 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1398 completion_error
1399 ),
1400 }
1401 }
1402}