1use ai_onboarding::YoungAccountBanner;
2use anthropic::AnthropicModelMode;
3use anyhow::{Context as _, Result, anyhow};
4use chrono::{DateTime, Utc};
5use client::{Client, ModelRequestUsage, UserStore, zed_urls};
6use cloud_llm_client::{
7 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
8 CompletionEvent, CompletionRequestStatus, CountTokensBody, CountTokensResponse,
9 EXPIRED_LLM_TOKEN_HEADER_NAME, ListModelsResponse, MODEL_REQUESTS_RESOURCE_HEADER_VALUE, Plan,
10 PlanV1, PlanV2, SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME,
11 SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME, TOOL_USE_LIMIT_REACHED_HEADER_NAME,
12 ZED_VERSION_HEADER_NAME,
13};
14use feature_flags::{BillingV2FeatureFlag, FeatureFlagAppExt};
15use futures::{
16 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
17};
18use google_ai::GoogleModelMode;
19use gpui::{
20 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
21};
22use http_client::http::{HeaderMap, HeaderValue};
23use http_client::{AsyncBody, HttpClient, HttpRequestExt, Method, Response, StatusCode};
24use language_model::{
25 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
26 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
27 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
28 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
29 LanguageModelToolSchemaFormat, LlmApiToken, ModelRequestLimitReachedError,
30 PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
31};
32use release_channel::AppVersion;
33use schemars::JsonSchema;
34use serde::{Deserialize, Serialize, de::DeserializeOwned};
35use settings::SettingsStore;
36pub use settings::ZedDotDevAvailableModel as AvailableModel;
37pub use settings::ZedDotDevAvailableProvider as AvailableProvider;
38use smol::io::{AsyncReadExt, BufReader};
39use std::pin::Pin;
40use std::str::FromStr as _;
41use std::sync::Arc;
42use std::time::Duration;
43use thiserror::Error;
44use ui::{TintColor, prelude::*};
45use util::{ResultExt as _, maybe};
46
47use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
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: gpui::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 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
222 .header("Authorization", format!("Bearer {token}"))
223 .body(AsyncBody::empty())?;
224 let mut response = http_client
225 .send(request)
226 .await
227 .context("failed to send list models request")?;
228
229 if response.status().is_success() {
230 let mut body = String::new();
231 response.body_mut().read_to_string(&mut body).await?;
232 Ok(serde_json::from_str(&body)?)
233 } else {
234 let mut body = String::new();
235 response.body_mut().read_to_string(&mut body).await?;
236 anyhow::bail!(
237 "error listing models.\nStatus: {:?}\nBody: {body}",
238 response.status(),
239 );
240 }
241 }
242}
243
244impl CloudLanguageModelProvider {
245 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
246 let mut status_rx = client.status();
247 let status = *status_rx.borrow();
248
249 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
250
251 let state_ref = state.downgrade();
252 let maintain_client_status = cx.spawn(async move |cx| {
253 while let Some(status) = status_rx.next().await {
254 if let Some(this) = state_ref.upgrade() {
255 _ = this.update(cx, |this, cx| {
256 if this.status != status {
257 this.status = status;
258 cx.notify();
259 }
260 });
261 } else {
262 break;
263 }
264 }
265 });
266
267 Self {
268 client,
269 state,
270 _maintain_client_status: maintain_client_status,
271 }
272 }
273
274 fn create_language_model(
275 &self,
276 model: Arc<cloud_llm_client::LanguageModel>,
277 llm_api_token: LlmApiToken,
278 ) -> Arc<dyn LanguageModel> {
279 Arc::new(CloudLanguageModel {
280 id: LanguageModelId(SharedString::from(model.id.0.clone())),
281 model,
282 llm_api_token,
283 client: self.client.clone(),
284 request_limiter: RateLimiter::new(4),
285 })
286 }
287}
288
289impl LanguageModelProviderState for CloudLanguageModelProvider {
290 type ObservableEntity = State;
291
292 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
293 Some(self.state.clone())
294 }
295}
296
297impl LanguageModelProvider for CloudLanguageModelProvider {
298 fn id(&self) -> LanguageModelProviderId {
299 PROVIDER_ID
300 }
301
302 fn name(&self) -> LanguageModelProviderName {
303 PROVIDER_NAME
304 }
305
306 fn icon(&self) -> IconName {
307 IconName::AiZed
308 }
309
310 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
311 let default_model = self.state.read(cx).default_model.clone()?;
312 let llm_api_token = self.state.read(cx).llm_api_token.clone();
313 Some(self.create_language_model(default_model, llm_api_token))
314 }
315
316 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
317 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
318 let llm_api_token = self.state.read(cx).llm_api_token.clone();
319 Some(self.create_language_model(default_fast_model, llm_api_token))
320 }
321
322 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
323 let llm_api_token = self.state.read(cx).llm_api_token.clone();
324 self.state
325 .read(cx)
326 .recommended_models
327 .iter()
328 .cloned()
329 .map(|model| self.create_language_model(model, llm_api_token.clone()))
330 .collect()
331 }
332
333 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
334 let llm_api_token = self.state.read(cx).llm_api_token.clone();
335 self.state
336 .read(cx)
337 .models
338 .iter()
339 .cloned()
340 .map(|model| self.create_language_model(model, llm_api_token.clone()))
341 .collect()
342 }
343
344 fn is_authenticated(&self, cx: &App) -> bool {
345 let state = self.state.read(cx);
346 !state.is_signed_out(cx)
347 }
348
349 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
350 Task::ready(Ok(()))
351 }
352
353 fn configuration_view(
354 &self,
355 _target_agent: language_model::ConfigurationViewTargetAgent,
356 _: &mut Window,
357 cx: &mut App,
358 ) -> AnyView {
359 cx.new(|_| ConfigurationView::new(self.state.clone()))
360 .into()
361 }
362
363 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
364 Task::ready(Ok(()))
365 }
366}
367
368pub struct CloudLanguageModel {
369 id: LanguageModelId,
370 model: Arc<cloud_llm_client::LanguageModel>,
371 llm_api_token: LlmApiToken,
372 client: Arc<Client>,
373 request_limiter: RateLimiter,
374}
375
376struct PerformLlmCompletionResponse {
377 response: Response<AsyncBody>,
378 usage: Option<ModelRequestUsage>,
379 tool_use_limit_reached: bool,
380 includes_status_messages: bool,
381}
382
383impl CloudLanguageModel {
384 async fn perform_llm_completion(
385 client: Arc<Client>,
386 llm_api_token: LlmApiToken,
387 app_version: Option<SemanticVersion>,
388 body: CompletionBody,
389 ) -> Result<PerformLlmCompletionResponse> {
390 let http_client = &client.http_client();
391
392 let mut token = llm_api_token.acquire(&client).await?;
393 let mut refreshed_token = false;
394
395 loop {
396 let request = http_client::Request::builder()
397 .method(Method::POST)
398 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
399 .when_some(app_version, |builder, app_version| {
400 builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
401 })
402 .header("Content-Type", "application/json")
403 .header("Authorization", format!("Bearer {token}"))
404 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
405 .body(serde_json::to_string(&body)?.into())?;
406
407 let mut response = http_client.send(request).await?;
408 let status = response.status();
409 if status.is_success() {
410 let includes_status_messages = response
411 .headers()
412 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
413 .is_some();
414
415 let tool_use_limit_reached = response
416 .headers()
417 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
418 .is_some();
419
420 let usage = if includes_status_messages {
421 None
422 } else {
423 ModelRequestUsage::from_headers(response.headers()).ok()
424 };
425
426 return Ok(PerformLlmCompletionResponse {
427 response,
428 usage,
429 includes_status_messages,
430 tool_use_limit_reached,
431 });
432 }
433
434 if !refreshed_token
435 && response
436 .headers()
437 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
438 .is_some()
439 {
440 token = llm_api_token.refresh(&client).await?;
441 refreshed_token = true;
442 continue;
443 }
444
445 if status == StatusCode::FORBIDDEN
446 && response
447 .headers()
448 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
449 .is_some()
450 {
451 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
452 .headers()
453 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
454 .and_then(|resource| resource.to_str().ok())
455 && let Some(plan) = response
456 .headers()
457 .get(CURRENT_PLAN_HEADER_NAME)
458 .and_then(|plan| plan.to_str().ok())
459 .and_then(|plan| cloud_llm_client::PlanV1::from_str(plan).ok())
460 .map(Plan::V1)
461 {
462 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
463 }
464 } else if status == StatusCode::PAYMENT_REQUIRED {
465 return Err(anyhow!(PaymentRequiredError));
466 }
467
468 let mut body = String::new();
469 let headers = response.headers().clone();
470 response.body_mut().read_to_string(&mut body).await?;
471 return Err(anyhow!(ApiError {
472 status,
473 body,
474 headers
475 }));
476 }
477 }
478}
479
480#[derive(Debug, Error)]
481#[error("cloud language model request failed with status {status}: {body}")]
482struct ApiError {
483 status: StatusCode,
484 body: String,
485 headers: HeaderMap<HeaderValue>,
486}
487
488/// Represents error responses from Zed's cloud API.
489///
490/// Example JSON for an upstream HTTP error:
491/// ```json
492/// {
493/// "code": "upstream_http_error",
494/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
495/// "upstream_status": 503
496/// }
497/// ```
498#[derive(Debug, serde::Deserialize)]
499struct CloudApiError {
500 code: String,
501 message: String,
502 #[serde(default)]
503 #[serde(deserialize_with = "deserialize_optional_status_code")]
504 upstream_status: Option<StatusCode>,
505 #[serde(default)]
506 retry_after: Option<f64>,
507}
508
509fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
510where
511 D: serde::Deserializer<'de>,
512{
513 let opt: Option<u16> = Option::deserialize(deserializer)?;
514 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
515}
516
517impl From<ApiError> for LanguageModelCompletionError {
518 fn from(error: ApiError) -> Self {
519 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
520 if cloud_error.code.starts_with("upstream_http_") {
521 let status = if let Some(status) = cloud_error.upstream_status {
522 status
523 } else if cloud_error.code.ends_with("_error") {
524 error.status
525 } else {
526 // If there's a status code in the code string (e.g. "upstream_http_429")
527 // then use that; otherwise, see if the JSON contains a status code.
528 cloud_error
529 .code
530 .strip_prefix("upstream_http_")
531 .and_then(|code_str| code_str.parse::<u16>().ok())
532 .and_then(|code| StatusCode::from_u16(code).ok())
533 .unwrap_or(error.status)
534 };
535
536 return LanguageModelCompletionError::UpstreamProviderError {
537 message: cloud_error.message,
538 status,
539 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
540 };
541 }
542
543 return LanguageModelCompletionError::from_http_status(
544 PROVIDER_NAME,
545 error.status,
546 cloud_error.message,
547 None,
548 );
549 }
550
551 let retry_after = None;
552 LanguageModelCompletionError::from_http_status(
553 PROVIDER_NAME,
554 error.status,
555 error.body,
556 retry_after,
557 )
558 }
559}
560
561impl LanguageModel for CloudLanguageModel {
562 fn id(&self) -> LanguageModelId {
563 self.id.clone()
564 }
565
566 fn name(&self) -> LanguageModelName {
567 LanguageModelName::from(self.model.display_name.clone())
568 }
569
570 fn provider_id(&self) -> LanguageModelProviderId {
571 PROVIDER_ID
572 }
573
574 fn provider_name(&self) -> LanguageModelProviderName {
575 PROVIDER_NAME
576 }
577
578 fn upstream_provider_id(&self) -> LanguageModelProviderId {
579 use cloud_llm_client::LanguageModelProvider::*;
580 match self.model.provider {
581 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
582 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
583 Google => language_model::GOOGLE_PROVIDER_ID,
584 XAi => language_model::X_AI_PROVIDER_ID,
585 }
586 }
587
588 fn upstream_provider_name(&self) -> LanguageModelProviderName {
589 use cloud_llm_client::LanguageModelProvider::*;
590 match self.model.provider {
591 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
592 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
593 Google => language_model::GOOGLE_PROVIDER_NAME,
594 XAi => language_model::X_AI_PROVIDER_NAME,
595 }
596 }
597
598 fn supports_tools(&self) -> bool {
599 self.model.supports_tools
600 }
601
602 fn supports_images(&self) -> bool {
603 self.model.supports_images
604 }
605
606 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
607 match choice {
608 LanguageModelToolChoice::Auto
609 | LanguageModelToolChoice::Any
610 | LanguageModelToolChoice::None => true,
611 }
612 }
613
614 fn supports_burn_mode(&self) -> bool {
615 self.model.supports_max_mode
616 }
617
618 fn telemetry_id(&self) -> String {
619 format!("zed.dev/{}", self.model.id)
620 }
621
622 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
623 match self.model.provider {
624 cloud_llm_client::LanguageModelProvider::Anthropic
625 | cloud_llm_client::LanguageModelProvider::OpenAi
626 | cloud_llm_client::LanguageModelProvider::XAi => {
627 LanguageModelToolSchemaFormat::JsonSchema
628 }
629 cloud_llm_client::LanguageModelProvider::Google => {
630 LanguageModelToolSchemaFormat::JsonSchemaSubset
631 }
632 }
633 }
634
635 fn max_token_count(&self) -> u64 {
636 self.model.max_token_count as u64
637 }
638
639 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
640 self.model
641 .max_token_count_in_max_mode
642 .filter(|_| self.model.supports_max_mode)
643 .map(|max_token_count| max_token_count as u64)
644 }
645
646 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
647 match &self.model.provider {
648 cloud_llm_client::LanguageModelProvider::Anthropic => {
649 Some(LanguageModelCacheConfiguration {
650 min_total_token: 2_048,
651 should_speculate: true,
652 max_cache_anchors: 4,
653 })
654 }
655 cloud_llm_client::LanguageModelProvider::OpenAi
656 | cloud_llm_client::LanguageModelProvider::XAi
657 | cloud_llm_client::LanguageModelProvider::Google => None,
658 }
659 }
660
661 fn count_tokens(
662 &self,
663 request: LanguageModelRequest,
664 cx: &App,
665 ) -> BoxFuture<'static, Result<u64>> {
666 match self.model.provider {
667 cloud_llm_client::LanguageModelProvider::Anthropic => {
668 count_anthropic_tokens(request, cx)
669 }
670 cloud_llm_client::LanguageModelProvider::OpenAi => {
671 let model = match open_ai::Model::from_id(&self.model.id.0) {
672 Ok(model) => model,
673 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
674 };
675 count_open_ai_tokens(request, model, cx)
676 }
677 cloud_llm_client::LanguageModelProvider::XAi => {
678 let model = match x_ai::Model::from_id(&self.model.id.0) {
679 Ok(model) => model,
680 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
681 };
682 count_xai_tokens(request, model, cx)
683 }
684 cloud_llm_client::LanguageModelProvider::Google => {
685 let client = self.client.clone();
686 let llm_api_token = self.llm_api_token.clone();
687 let model_id = self.model.id.to_string();
688 let generate_content_request =
689 into_google(request, model_id.clone(), GoogleModelMode::Default);
690 async move {
691 let http_client = &client.http_client();
692 let token = llm_api_token.acquire(&client).await?;
693
694 let request_body = CountTokensBody {
695 provider: cloud_llm_client::LanguageModelProvider::Google,
696 model: model_id,
697 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
698 generate_content_request,
699 })?,
700 };
701 let request = http_client::Request::builder()
702 .method(Method::POST)
703 .uri(
704 http_client
705 .build_zed_llm_url("/count_tokens", &[])?
706 .as_ref(),
707 )
708 .header("Content-Type", "application/json")
709 .header("Authorization", format!("Bearer {token}"))
710 .body(serde_json::to_string(&request_body)?.into())?;
711 let mut response = http_client.send(request).await?;
712 let status = response.status();
713 let headers = response.headers().clone();
714 let mut response_body = String::new();
715 response
716 .body_mut()
717 .read_to_string(&mut response_body)
718 .await?;
719
720 if status.is_success() {
721 let response_body: CountTokensResponse =
722 serde_json::from_str(&response_body)?;
723
724 Ok(response_body.tokens as u64)
725 } else {
726 Err(anyhow!(ApiError {
727 status,
728 body: response_body,
729 headers
730 }))
731 }
732 }
733 .boxed()
734 }
735 }
736 }
737
738 fn stream_completion(
739 &self,
740 request: LanguageModelRequest,
741 cx: &AsyncApp,
742 ) -> BoxFuture<
743 'static,
744 Result<
745 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
746 LanguageModelCompletionError,
747 >,
748 > {
749 let thread_id = request.thread_id.clone();
750 let prompt_id = request.prompt_id.clone();
751 let intent = request.intent;
752 let mode = request.mode;
753 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
754 let thinking_allowed = request.thinking_allowed;
755 match self.model.provider {
756 cloud_llm_client::LanguageModelProvider::Anthropic => {
757 let request = into_anthropic(
758 request,
759 self.model.id.to_string(),
760 1.0,
761 self.model.max_output_tokens as u64,
762 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
763 AnthropicModelMode::Thinking {
764 budget_tokens: Some(4_096),
765 }
766 } else {
767 AnthropicModelMode::Default
768 },
769 );
770 let client = self.client.clone();
771 let llm_api_token = self.llm_api_token.clone();
772 let future = self.request_limiter.stream(async move {
773 let PerformLlmCompletionResponse {
774 response,
775 usage,
776 includes_status_messages,
777 tool_use_limit_reached,
778 } = Self::perform_llm_completion(
779 client.clone(),
780 llm_api_token,
781 app_version,
782 CompletionBody {
783 thread_id,
784 prompt_id,
785 intent,
786 mode,
787 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
788 model: request.model.clone(),
789 provider_request: serde_json::to_value(&request)
790 .map_err(|e| anyhow!(e))?,
791 },
792 )
793 .await
794 .map_err(|err| match err.downcast::<ApiError>() {
795 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
796 Err(err) => anyhow!(err),
797 })?;
798
799 let mut mapper = AnthropicEventMapper::new();
800 Ok(map_cloud_completion_events(
801 Box::pin(
802 response_lines(response, includes_status_messages)
803 .chain(usage_updated_event(usage))
804 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
805 ),
806 move |event| mapper.map_event(event),
807 ))
808 });
809 async move { Ok(future.await?.boxed()) }.boxed()
810 }
811 cloud_llm_client::LanguageModelProvider::OpenAi => {
812 let client = self.client.clone();
813 let model = match open_ai::Model::from_id(&self.model.id.0) {
814 Ok(model) => model,
815 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
816 };
817 let request = into_open_ai(
818 request,
819 model.id(),
820 model.supports_parallel_tool_calls(),
821 model.supports_prompt_cache_key(),
822 None,
823 None,
824 );
825 let llm_api_token = self.llm_api_token.clone();
826 let future = self.request_limiter.stream(async move {
827 let PerformLlmCompletionResponse {
828 response,
829 usage,
830 includes_status_messages,
831 tool_use_limit_reached,
832 } = Self::perform_llm_completion(
833 client.clone(),
834 llm_api_token,
835 app_version,
836 CompletionBody {
837 thread_id,
838 prompt_id,
839 intent,
840 mode,
841 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
842 model: request.model.clone(),
843 provider_request: serde_json::to_value(&request)
844 .map_err(|e| anyhow!(e))?,
845 },
846 )
847 .await?;
848
849 let mut mapper = OpenAiEventMapper::new();
850 Ok(map_cloud_completion_events(
851 Box::pin(
852 response_lines(response, includes_status_messages)
853 .chain(usage_updated_event(usage))
854 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
855 ),
856 move |event| mapper.map_event(event),
857 ))
858 });
859 async move { Ok(future.await?.boxed()) }.boxed()
860 }
861 cloud_llm_client::LanguageModelProvider::XAi => {
862 let client = self.client.clone();
863 let model = match x_ai::Model::from_id(&self.model.id.0) {
864 Ok(model) => model,
865 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
866 };
867 let request = into_open_ai(
868 request,
869 model.id(),
870 model.supports_parallel_tool_calls(),
871 model.supports_prompt_cache_key(),
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 move |event| mapper.map_event(event),
907 ))
908 });
909 async move { Ok(future.await?.boxed()) }.boxed()
910 }
911 cloud_llm_client::LanguageModelProvider::Google => {
912 let client = self.client.clone();
913 let request =
914 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
915 let llm_api_token = self.llm_api_token.clone();
916 let future = self.request_limiter.stream(async move {
917 let PerformLlmCompletionResponse {
918 response,
919 usage,
920 includes_status_messages,
921 tool_use_limit_reached,
922 } = Self::perform_llm_completion(
923 client.clone(),
924 llm_api_token,
925 app_version,
926 CompletionBody {
927 thread_id,
928 prompt_id,
929 intent,
930 mode,
931 provider: cloud_llm_client::LanguageModelProvider::Google,
932 model: request.model.model_id.clone(),
933 provider_request: serde_json::to_value(&request)
934 .map_err(|e| anyhow!(e))?,
935 },
936 )
937 .await?;
938
939 let mut mapper = GoogleEventMapper::new();
940 Ok(map_cloud_completion_events(
941 Box::pin(
942 response_lines(response, includes_status_messages)
943 .chain(usage_updated_event(usage))
944 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
945 ),
946 move |event| mapper.map_event(event),
947 ))
948 });
949 async move { Ok(future.await?.boxed()) }.boxed()
950 }
951 }
952 }
953}
954
955fn map_cloud_completion_events<T, F>(
956 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
957 mut map_callback: F,
958) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
959where
960 T: DeserializeOwned + 'static,
961 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
962 + Send
963 + 'static,
964{
965 stream
966 .flat_map(move |event| {
967 futures::stream::iter(match event {
968 Err(error) => {
969 vec![Err(LanguageModelCompletionError::from(error))]
970 }
971 Ok(CompletionEvent::Status(event)) => {
972 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
973 }
974 Ok(CompletionEvent::Event(event)) => map_callback(event),
975 })
976 })
977 .boxed()
978}
979
980fn usage_updated_event<T>(
981 usage: Option<ModelRequestUsage>,
982) -> impl Stream<Item = Result<CompletionEvent<T>>> {
983 futures::stream::iter(usage.map(|usage| {
984 Ok(CompletionEvent::Status(
985 CompletionRequestStatus::UsageUpdated {
986 amount: usage.amount as usize,
987 limit: usage.limit,
988 },
989 ))
990 }))
991}
992
993fn tool_use_limit_reached_event<T>(
994 tool_use_limit_reached: bool,
995) -> impl Stream<Item = Result<CompletionEvent<T>>> {
996 futures::stream::iter(tool_use_limit_reached.then(|| {
997 Ok(CompletionEvent::Status(
998 CompletionRequestStatus::ToolUseLimitReached,
999 ))
1000 }))
1001}
1002
1003fn response_lines<T: DeserializeOwned>(
1004 response: Response<AsyncBody>,
1005 includes_status_messages: bool,
1006) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1007 futures::stream::try_unfold(
1008 (String::new(), BufReader::new(response.into_body())),
1009 move |(mut line, mut body)| async move {
1010 match body.read_line(&mut line).await {
1011 Ok(0) => Ok(None),
1012 Ok(_) => {
1013 let event = if includes_status_messages {
1014 serde_json::from_str::<CompletionEvent<T>>(&line)?
1015 } else {
1016 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1017 };
1018
1019 line.clear();
1020 Ok(Some((event, (line, body))))
1021 }
1022 Err(e) => Err(e.into()),
1023 }
1024 },
1025 )
1026}
1027
1028#[derive(IntoElement, RegisterComponent)]
1029struct ZedAiConfiguration {
1030 is_connected: bool,
1031 plan: Option<Plan>,
1032 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1033 eligible_for_trial: bool,
1034 account_too_young: bool,
1035 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1036}
1037
1038impl RenderOnce for ZedAiConfiguration {
1039 fn render(self, _window: &mut Window, cx: &mut App) -> impl IntoElement {
1040 let is_pro = self.plan.is_some_and(|plan| {
1041 matches!(plan, Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro))
1042 });
1043 let is_free_v2 = self
1044 .plan
1045 .is_some_and(|plan| plan == Plan::V2(PlanV2::ZedFree));
1046 let subscription_text = match (self.plan, self.subscription_period) {
1047 (Some(Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro)), Some(_)) => {
1048 "You have access to Zed's hosted models through your Pro subscription."
1049 }
1050 (Some(Plan::V1(PlanV1::ZedProTrial) | Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1051 "You have access to Zed's hosted models through your Pro trial."
1052 }
1053 (Some(Plan::V1(PlanV1::ZedFree)), Some(_)) => {
1054 "You have basic access to Zed's hosted models through the Free plan."
1055 }
1056 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1057 if self.eligible_for_trial {
1058 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1059 } else {
1060 "Subscribe for access to Zed's hosted models."
1061 }
1062 }
1063 _ => {
1064 if self.eligible_for_trial {
1065 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1066 } else {
1067 "Subscribe for access to Zed's hosted models."
1068 }
1069 }
1070 };
1071
1072 let manage_subscription_buttons = if is_pro {
1073 Button::new("manage_settings", "Manage Subscription")
1074 .full_width()
1075 .style(ButtonStyle::Tinted(TintColor::Accent))
1076 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1077 .into_any_element()
1078 } else if self.plan.is_none() || self.eligible_for_trial {
1079 Button::new("start_trial", "Start 14-day Free Pro Trial")
1080 .full_width()
1081 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1082 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1083 .into_any_element()
1084 } else {
1085 Button::new("upgrade", "Upgrade to Pro")
1086 .full_width()
1087 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1088 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1089 .into_any_element()
1090 };
1091
1092 if !self.is_connected {
1093 return v_flex()
1094 .gap_2()
1095 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1096 .child(
1097 Button::new("sign_in", "Sign In to use Zed AI")
1098 .icon_color(Color::Muted)
1099 .icon(IconName::Github)
1100 .icon_size(IconSize::Small)
1101 .icon_position(IconPosition::Start)
1102 .full_width()
1103 .on_click({
1104 let callback = self.sign_in_callback.clone();
1105 move |_, window, cx| (callback)(window, cx)
1106 }),
1107 );
1108 }
1109
1110 v_flex().gap_2().w_full().map(|this| {
1111 if self.account_too_young {
1112 this.child(YoungAccountBanner::new(
1113 is_free_v2 || cx.has_flag::<BillingV2FeatureFlag>(),
1114 ))
1115 .child(
1116 Button::new("upgrade", "Upgrade to Pro")
1117 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1118 .full_width()
1119 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1120 )
1121 } else {
1122 this.text_sm()
1123 .child(subscription_text)
1124 .child(manage_subscription_buttons)
1125 }
1126 })
1127 }
1128}
1129
1130struct ConfigurationView {
1131 state: Entity<State>,
1132 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1133}
1134
1135impl ConfigurationView {
1136 fn new(state: Entity<State>) -> Self {
1137 let sign_in_callback = Arc::new({
1138 let state = state.clone();
1139 move |_window: &mut Window, cx: &mut App| {
1140 state.update(cx, |state, cx| {
1141 state.authenticate(cx).detach_and_log_err(cx);
1142 });
1143 }
1144 });
1145
1146 Self {
1147 state,
1148 sign_in_callback,
1149 }
1150 }
1151}
1152
1153impl Render for ConfigurationView {
1154 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1155 let state = self.state.read(cx);
1156 let user_store = state.user_store.read(cx);
1157
1158 ZedAiConfiguration {
1159 is_connected: !state.is_signed_out(cx),
1160 plan: user_store.plan(),
1161 subscription_period: user_store.subscription_period(),
1162 eligible_for_trial: user_store.trial_started_at().is_none(),
1163 account_too_young: user_store.account_too_young(),
1164 sign_in_callback: self.sign_in_callback.clone(),
1165 }
1166 }
1167}
1168
1169impl Component for ZedAiConfiguration {
1170 fn name() -> &'static str {
1171 "AI Configuration Content"
1172 }
1173
1174 fn sort_name() -> &'static str {
1175 "AI Configuration Content"
1176 }
1177
1178 fn scope() -> ComponentScope {
1179 ComponentScope::Onboarding
1180 }
1181
1182 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1183 fn configuration(
1184 is_connected: bool,
1185 plan: Option<Plan>,
1186 eligible_for_trial: bool,
1187 account_too_young: bool,
1188 ) -> AnyElement {
1189 ZedAiConfiguration {
1190 is_connected,
1191 plan,
1192 subscription_period: plan
1193 .is_some()
1194 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1195 eligible_for_trial,
1196 account_too_young,
1197 sign_in_callback: Arc::new(|_, _| {}),
1198 }
1199 .into_any_element()
1200 }
1201
1202 Some(
1203 v_flex()
1204 .p_4()
1205 .gap_4()
1206 .children(vec![
1207 single_example("Not connected", configuration(false, None, false, false)),
1208 single_example(
1209 "Accept Terms of Service",
1210 configuration(true, None, true, false),
1211 ),
1212 single_example(
1213 "No Plan - Not eligible for trial",
1214 configuration(true, None, false, false),
1215 ),
1216 single_example(
1217 "No Plan - Eligible for trial",
1218 configuration(true, None, true, false),
1219 ),
1220 single_example(
1221 "Free Plan",
1222 configuration(true, Some(Plan::V1(PlanV1::ZedFree)), true, false),
1223 ),
1224 single_example(
1225 "Zed Pro Trial Plan",
1226 configuration(true, Some(Plan::V1(PlanV1::ZedProTrial)), true, false),
1227 ),
1228 single_example(
1229 "Zed Pro Plan",
1230 configuration(true, Some(Plan::V1(PlanV1::ZedPro)), true, false),
1231 ),
1232 ])
1233 .into_any_element(),
1234 )
1235 }
1236}
1237
1238#[cfg(test)]
1239mod tests {
1240 use super::*;
1241 use http_client::http::{HeaderMap, StatusCode};
1242 use language_model::LanguageModelCompletionError;
1243
1244 #[test]
1245 fn test_api_error_conversion_with_upstream_http_error() {
1246 // upstream_http_error with 503 status should become ServerOverloaded
1247 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}"#;
1248
1249 let api_error = ApiError {
1250 status: StatusCode::INTERNAL_SERVER_ERROR,
1251 body: error_body.to_string(),
1252 headers: HeaderMap::new(),
1253 };
1254
1255 let completion_error: LanguageModelCompletionError = api_error.into();
1256
1257 match completion_error {
1258 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1259 assert_eq!(
1260 message,
1261 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1262 );
1263 }
1264 _ => panic!(
1265 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1266 completion_error
1267 ),
1268 }
1269
1270 // upstream_http_error with 500 status should become ApiInternalServerError
1271 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1272
1273 let api_error = ApiError {
1274 status: StatusCode::INTERNAL_SERVER_ERROR,
1275 body: error_body.to_string(),
1276 headers: HeaderMap::new(),
1277 };
1278
1279 let completion_error: LanguageModelCompletionError = api_error.into();
1280
1281 match completion_error {
1282 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1283 assert_eq!(
1284 message,
1285 "Received an error from the OpenAI API: internal server error"
1286 );
1287 }
1288 _ => panic!(
1289 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1290 completion_error
1291 ),
1292 }
1293
1294 // upstream_http_error with 429 status should become RateLimitExceeded
1295 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1296
1297 let api_error = ApiError {
1298 status: StatusCode::INTERNAL_SERVER_ERROR,
1299 body: error_body.to_string(),
1300 headers: HeaderMap::new(),
1301 };
1302
1303 let completion_error: LanguageModelCompletionError = api_error.into();
1304
1305 match completion_error {
1306 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1307 assert_eq!(
1308 message,
1309 "Received an error from the Google API: rate limit exceeded"
1310 );
1311 }
1312 _ => panic!(
1313 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1314 completion_error
1315 ),
1316 }
1317
1318 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1319 let error_body = "Regular internal server error";
1320
1321 let api_error = ApiError {
1322 status: StatusCode::INTERNAL_SERVER_ERROR,
1323 body: error_body.to_string(),
1324 headers: HeaderMap::new(),
1325 };
1326
1327 let completion_error: LanguageModelCompletionError = api_error.into();
1328
1329 match completion_error {
1330 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1331 assert_eq!(provider, PROVIDER_NAME);
1332 assert_eq!(message, "Regular internal server error");
1333 }
1334 _ => panic!(
1335 "Expected ApiInternalServerError for regular 500, got: {:?}",
1336 completion_error
1337 ),
1338 }
1339
1340 // upstream_http_429 format should be converted to UpstreamProviderError
1341 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1342
1343 let api_error = ApiError {
1344 status: StatusCode::INTERNAL_SERVER_ERROR,
1345 body: error_body.to_string(),
1346 headers: HeaderMap::new(),
1347 };
1348
1349 let completion_error: LanguageModelCompletionError = api_error.into();
1350
1351 match completion_error {
1352 LanguageModelCompletionError::UpstreamProviderError {
1353 message,
1354 status,
1355 retry_after,
1356 } => {
1357 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1358 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1359 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1360 }
1361 _ => panic!(
1362 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1363 completion_error
1364 ),
1365 }
1366
1367 // Invalid JSON in error body should fall back to regular error handling
1368 let error_body = "Not JSON at all";
1369
1370 let api_error = ApiError {
1371 status: StatusCode::INTERNAL_SERVER_ERROR,
1372 body: error_body.to_string(),
1373 headers: HeaderMap::new(),
1374 };
1375
1376 let completion_error: LanguageModelCompletionError = api_error.into();
1377
1378 match completion_error {
1379 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1380 assert_eq!(provider, PROVIDER_NAME);
1381 }
1382 _ => panic!(
1383 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1384 completion_error
1385 ),
1386 }
1387 }
1388}