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};
50
51const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
52const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
53
54#[derive(Default, Clone, Debug, PartialEq)]
55pub struct ZedDotDevSettings {
56 pub available_models: Vec<AvailableModel>,
57}
58#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
59#[serde(tag = "type", rename_all = "lowercase")]
60pub enum ModelMode {
61 #[default]
62 Default,
63 Thinking {
64 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
65 budget_tokens: Option<u32>,
66 },
67}
68
69impl From<ModelMode> for AnthropicModelMode {
70 fn from(value: ModelMode) -> Self {
71 match value {
72 ModelMode::Default => AnthropicModelMode::Default,
73 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
74 }
75 }
76}
77
78pub struct CloudLanguageModelProvider {
79 client: Arc<Client>,
80 state: gpui::Entity<State>,
81 _maintain_client_status: Task<()>,
82}
83
84pub struct State {
85 client: Arc<Client>,
86 llm_api_token: LlmApiToken,
87 user_store: Entity<UserStore>,
88 status: client::Status,
89 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
90 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
91 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
92 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
93 _fetch_models_task: Task<()>,
94 _settings_subscription: Subscription,
95 _llm_token_subscription: Subscription,
96}
97
98impl State {
99 fn new(
100 client: Arc<Client>,
101 user_store: Entity<UserStore>,
102 status: client::Status,
103 cx: &mut Context<Self>,
104 ) -> Self {
105 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
106 let mut current_user = user_store.read(cx).watch_current_user();
107 Self {
108 client: client.clone(),
109 llm_api_token: LlmApiToken::default(),
110 user_store,
111 status,
112 models: Vec::new(),
113 default_model: None,
114 default_fast_model: None,
115 recommended_models: Vec::new(),
116 _fetch_models_task: cx.spawn(async move |this, cx| {
117 maybe!(async move {
118 let (client, llm_api_token) = this
119 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
120
121 while current_user.borrow().is_none() {
122 current_user.next().await;
123 }
124
125 let response =
126 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
127 this.update(cx, |this, cx| this.update_models(response, cx))?;
128 anyhow::Ok(())
129 })
130 .await
131 .context("failed to fetch Zed models")
132 .log_err();
133 }),
134 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
135 cx.notify();
136 }),
137 _llm_token_subscription: cx.subscribe(
138 &refresh_llm_token_listener,
139 move |this, _listener, _event, cx| {
140 let client = this.client.clone();
141 let llm_api_token = this.llm_api_token.clone();
142 cx.spawn(async move |this, cx| {
143 llm_api_token.refresh(&client).await?;
144 let response = Self::fetch_models(client, llm_api_token).await?;
145 this.update(cx, |this, cx| {
146 this.update_models(response, cx);
147 })
148 })
149 .detach_and_log_err(cx);
150 },
151 ),
152 }
153 }
154
155 fn is_signed_out(&self, cx: &App) -> bool {
156 self.user_store.read(cx).current_user().is_none()
157 }
158
159 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
160 let client = self.client.clone();
161 cx.spawn(async move |state, cx| {
162 client.sign_in_with_optional_connect(true, cx).await?;
163 state.update(cx, |_, cx| cx.notify())
164 })
165 }
166 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
167 let mut models = Vec::new();
168
169 for model in response.models {
170 models.push(Arc::new(model.clone()));
171
172 // Right now we represent thinking variants of models as separate models on the client,
173 // so we need to insert variants for any model that supports thinking.
174 if model.supports_thinking {
175 models.push(Arc::new(cloud_llm_client::LanguageModel {
176 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
177 display_name: format!("{} Thinking", model.display_name),
178 ..model
179 }));
180 }
181 }
182
183 self.default_model = models
184 .iter()
185 .find(|model| {
186 response
187 .default_model
188 .as_ref()
189 .is_some_and(|default_model_id| &model.id == default_model_id)
190 })
191 .cloned();
192 self.default_fast_model = models
193 .iter()
194 .find(|model| {
195 response
196 .default_fast_model
197 .as_ref()
198 .is_some_and(|default_fast_model_id| &model.id == default_fast_model_id)
199 })
200 .cloned();
201 self.recommended_models = response
202 .recommended_models
203 .iter()
204 .filter_map(|id| models.iter().find(|model| &model.id == id))
205 .cloned()
206 .collect();
207 self.models = models;
208 cx.notify();
209 }
210
211 async fn fetch_models(
212 client: Arc<Client>,
213 llm_api_token: LlmApiToken,
214 ) -> Result<ListModelsResponse> {
215 let http_client = &client.http_client();
216 let token = llm_api_token.acquire(&client).await?;
217
218 let request = http_client::Request::builder()
219 .method(Method::GET)
220 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
221 .header("Authorization", format!("Bearer {token}"))
222 .body(AsyncBody::empty())?;
223 let mut response = http_client
224 .send(request)
225 .await
226 .context("failed to send list models request")?;
227
228 if response.status().is_success() {
229 let mut body = String::new();
230 response.body_mut().read_to_string(&mut body).await?;
231 Ok(serde_json::from_str(&body)?)
232 } else {
233 let mut body = String::new();
234 response.body_mut().read_to_string(&mut body).await?;
235 anyhow::bail!(
236 "error listing models.\nStatus: {:?}\nBody: {body}",
237 response.status(),
238 );
239 }
240 }
241}
242
243impl CloudLanguageModelProvider {
244 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
245 let mut status_rx = client.status();
246 let status = *status_rx.borrow();
247
248 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
249
250 let state_ref = state.downgrade();
251 let maintain_client_status = cx.spawn(async move |cx| {
252 while let Some(status) = status_rx.next().await {
253 if let Some(this) = state_ref.upgrade() {
254 _ = this.update(cx, |this, cx| {
255 if this.status != status {
256 this.status = status;
257 cx.notify();
258 }
259 });
260 } else {
261 break;
262 }
263 }
264 });
265
266 Self {
267 client,
268 state,
269 _maintain_client_status: maintain_client_status,
270 }
271 }
272
273 fn create_language_model(
274 &self,
275 model: Arc<cloud_llm_client::LanguageModel>,
276 llm_api_token: LlmApiToken,
277 ) -> Arc<dyn LanguageModel> {
278 Arc::new(CloudLanguageModel {
279 id: LanguageModelId(SharedString::from(model.id.0.clone())),
280 model,
281 llm_api_token,
282 client: self.client.clone(),
283 request_limiter: RateLimiter::new(4),
284 })
285 }
286}
287
288impl LanguageModelProviderState for CloudLanguageModelProvider {
289 type ObservableEntity = State;
290
291 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
292 Some(self.state.clone())
293 }
294}
295
296impl LanguageModelProvider for CloudLanguageModelProvider {
297 fn id(&self) -> LanguageModelProviderId {
298 PROVIDER_ID
299 }
300
301 fn name(&self) -> LanguageModelProviderName {
302 PROVIDER_NAME
303 }
304
305 fn icon(&self) -> IconName {
306 IconName::AiZed
307 }
308
309 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
310 let default_model = self.state.read(cx).default_model.clone()?;
311 let llm_api_token = self.state.read(cx).llm_api_token.clone();
312 Some(self.create_language_model(default_model, llm_api_token))
313 }
314
315 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
316 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
317 let llm_api_token = self.state.read(cx).llm_api_token.clone();
318 Some(self.create_language_model(default_fast_model, llm_api_token))
319 }
320
321 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
322 let llm_api_token = self.state.read(cx).llm_api_token.clone();
323 self.state
324 .read(cx)
325 .recommended_models
326 .iter()
327 .cloned()
328 .map(|model| self.create_language_model(model, llm_api_token.clone()))
329 .collect()
330 }
331
332 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
333 let llm_api_token = self.state.read(cx).llm_api_token.clone();
334 self.state
335 .read(cx)
336 .models
337 .iter()
338 .cloned()
339 .map(|model| self.create_language_model(model, llm_api_token.clone()))
340 .collect()
341 }
342
343 fn is_authenticated(&self, cx: &App) -> bool {
344 let state = self.state.read(cx);
345 !state.is_signed_out(cx)
346 }
347
348 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
349 Task::ready(Ok(()))
350 }
351
352 fn configuration_view(
353 &self,
354 _target_agent: language_model::ConfigurationViewTargetAgent,
355 _: &mut Window,
356 cx: &mut App,
357 ) -> AnyView {
358 cx.new(|_| ConfigurationView::new(self.state.clone()))
359 .into()
360 }
361
362 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
363 Task::ready(Ok(()))
364 }
365}
366
367pub struct CloudLanguageModel {
368 id: LanguageModelId,
369 model: Arc<cloud_llm_client::LanguageModel>,
370 llm_api_token: LlmApiToken,
371 client: Arc<Client>,
372 request_limiter: RateLimiter,
373}
374
375struct PerformLlmCompletionResponse {
376 response: Response<AsyncBody>,
377 usage: Option<ModelRequestUsage>,
378 tool_use_limit_reached: bool,
379 includes_status_messages: bool,
380}
381
382impl CloudLanguageModel {
383 async fn perform_llm_completion(
384 client: Arc<Client>,
385 llm_api_token: LlmApiToken,
386 app_version: Option<SemanticVersion>,
387 body: CompletionBody,
388 ) -> Result<PerformLlmCompletionResponse> {
389 let http_client = &client.http_client();
390
391 let mut token = llm_api_token.acquire(&client).await?;
392 let mut refreshed_token = false;
393
394 loop {
395 let request = http_client::Request::builder()
396 .method(Method::POST)
397 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
398 .when_some(app_version, |builder, app_version| {
399 builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
400 })
401 .header("Content-Type", "application/json")
402 .header("Authorization", format!("Bearer {token}"))
403 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
404 .body(serde_json::to_string(&body)?.into())?;
405
406 let mut response = http_client.send(request).await?;
407 let status = response.status();
408 if status.is_success() {
409 let includes_status_messages = response
410 .headers()
411 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
412 .is_some();
413
414 let tool_use_limit_reached = response
415 .headers()
416 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
417 .is_some();
418
419 let usage = if includes_status_messages {
420 None
421 } else {
422 ModelRequestUsage::from_headers(response.headers()).ok()
423 };
424
425 return Ok(PerformLlmCompletionResponse {
426 response,
427 usage,
428 includes_status_messages,
429 tool_use_limit_reached,
430 });
431 }
432
433 if !refreshed_token
434 && response
435 .headers()
436 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
437 .is_some()
438 {
439 token = llm_api_token.refresh(&client).await?;
440 refreshed_token = true;
441 continue;
442 }
443
444 if status == StatusCode::FORBIDDEN
445 && response
446 .headers()
447 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
448 .is_some()
449 {
450 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
451 .headers()
452 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
453 .and_then(|resource| resource.to_str().ok())
454 && let Some(plan) = response
455 .headers()
456 .get(CURRENT_PLAN_HEADER_NAME)
457 .and_then(|plan| plan.to_str().ok())
458 .and_then(|plan| cloud_llm_client::PlanV1::from_str(plan).ok())
459 .map(Plan::V1)
460 {
461 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
462 }
463 } else if status == StatusCode::PAYMENT_REQUIRED {
464 return Err(anyhow!(PaymentRequiredError));
465 }
466
467 let mut body = String::new();
468 let headers = response.headers().clone();
469 response.body_mut().read_to_string(&mut body).await?;
470 return Err(anyhow!(ApiError {
471 status,
472 body,
473 headers
474 }));
475 }
476 }
477}
478
479#[derive(Debug, Error)]
480#[error("cloud language model request failed with status {status}: {body}")]
481struct ApiError {
482 status: StatusCode,
483 body: String,
484 headers: HeaderMap<HeaderValue>,
485}
486
487/// Represents error responses from Zed's cloud API.
488///
489/// Example JSON for an upstream HTTP error:
490/// ```json
491/// {
492/// "code": "upstream_http_error",
493/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
494/// "upstream_status": 503
495/// }
496/// ```
497#[derive(Debug, serde::Deserialize)]
498struct CloudApiError {
499 code: String,
500 message: String,
501 #[serde(default)]
502 #[serde(deserialize_with = "deserialize_optional_status_code")]
503 upstream_status: Option<StatusCode>,
504 #[serde(default)]
505 retry_after: Option<f64>,
506}
507
508fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
509where
510 D: serde::Deserializer<'de>,
511{
512 let opt: Option<u16> = Option::deserialize(deserializer)?;
513 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
514}
515
516impl From<ApiError> for LanguageModelCompletionError {
517 fn from(error: ApiError) -> Self {
518 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body) {
519 if cloud_error.code.starts_with("upstream_http_") {
520 let status = if let Some(status) = cloud_error.upstream_status {
521 status
522 } else if cloud_error.code.ends_with("_error") {
523 error.status
524 } else {
525 // If there's a status code in the code string (e.g. "upstream_http_429")
526 // then use that; otherwise, see if the JSON contains a status code.
527 cloud_error
528 .code
529 .strip_prefix("upstream_http_")
530 .and_then(|code_str| code_str.parse::<u16>().ok())
531 .and_then(|code| StatusCode::from_u16(code).ok())
532 .unwrap_or(error.status)
533 };
534
535 return LanguageModelCompletionError::UpstreamProviderError {
536 message: cloud_error.message,
537 status,
538 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
539 };
540 }
541
542 return LanguageModelCompletionError::from_http_status(
543 PROVIDER_NAME,
544 error.status,
545 cloud_error.message,
546 None,
547 );
548 }
549
550 let retry_after = None;
551 LanguageModelCompletionError::from_http_status(
552 PROVIDER_NAME,
553 error.status,
554 error.body,
555 retry_after,
556 )
557 }
558}
559
560impl LanguageModel for CloudLanguageModel {
561 fn id(&self) -> LanguageModelId {
562 self.id.clone()
563 }
564
565 fn name(&self) -> LanguageModelName {
566 LanguageModelName::from(self.model.display_name.clone())
567 }
568
569 fn provider_id(&self) -> LanguageModelProviderId {
570 PROVIDER_ID
571 }
572
573 fn provider_name(&self) -> LanguageModelProviderName {
574 PROVIDER_NAME
575 }
576
577 fn upstream_provider_id(&self) -> LanguageModelProviderId {
578 use cloud_llm_client::LanguageModelProvider::*;
579 match self.model.provider {
580 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
581 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
582 Google => language_model::GOOGLE_PROVIDER_ID,
583 }
584 }
585
586 fn upstream_provider_name(&self) -> LanguageModelProviderName {
587 use cloud_llm_client::LanguageModelProvider::*;
588 match self.model.provider {
589 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
590 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
591 Google => language_model::GOOGLE_PROVIDER_NAME,
592 }
593 }
594
595 fn supports_tools(&self) -> bool {
596 self.model.supports_tools
597 }
598
599 fn supports_images(&self) -> bool {
600 self.model.supports_images
601 }
602
603 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
604 match choice {
605 LanguageModelToolChoice::Auto
606 | LanguageModelToolChoice::Any
607 | LanguageModelToolChoice::None => true,
608 }
609 }
610
611 fn supports_burn_mode(&self) -> bool {
612 self.model.supports_max_mode
613 }
614
615 fn telemetry_id(&self) -> String {
616 format!("zed.dev/{}", self.model.id)
617 }
618
619 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
620 match self.model.provider {
621 cloud_llm_client::LanguageModelProvider::Anthropic
622 | cloud_llm_client::LanguageModelProvider::OpenAi => {
623 LanguageModelToolSchemaFormat::JsonSchema
624 }
625 cloud_llm_client::LanguageModelProvider::Google => {
626 LanguageModelToolSchemaFormat::JsonSchemaSubset
627 }
628 }
629 }
630
631 fn max_token_count(&self) -> u64 {
632 self.model.max_token_count as u64
633 }
634
635 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
636 self.model
637 .max_token_count_in_max_mode
638 .filter(|_| self.model.supports_max_mode)
639 .map(|max_token_count| max_token_count as u64)
640 }
641
642 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
643 match &self.model.provider {
644 cloud_llm_client::LanguageModelProvider::Anthropic => {
645 Some(LanguageModelCacheConfiguration {
646 min_total_token: 2_048,
647 should_speculate: true,
648 max_cache_anchors: 4,
649 })
650 }
651 cloud_llm_client::LanguageModelProvider::OpenAi
652 | cloud_llm_client::LanguageModelProvider::Google => None,
653 }
654 }
655
656 fn count_tokens(
657 &self,
658 request: LanguageModelRequest,
659 cx: &App,
660 ) -> BoxFuture<'static, Result<u64>> {
661 match self.model.provider {
662 cloud_llm_client::LanguageModelProvider::Anthropic => {
663 count_anthropic_tokens(request, cx)
664 }
665 cloud_llm_client::LanguageModelProvider::OpenAi => {
666 let model = match open_ai::Model::from_id(&self.model.id.0) {
667 Ok(model) => model,
668 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
669 };
670 count_open_ai_tokens(request, model, cx)
671 }
672 cloud_llm_client::LanguageModelProvider::Google => {
673 let client = self.client.clone();
674 let llm_api_token = self.llm_api_token.clone();
675 let model_id = self.model.id.to_string();
676 let generate_content_request =
677 into_google(request, model_id.clone(), GoogleModelMode::Default);
678 async move {
679 let http_client = &client.http_client();
680 let token = llm_api_token.acquire(&client).await?;
681
682 let request_body = CountTokensBody {
683 provider: cloud_llm_client::LanguageModelProvider::Google,
684 model: model_id,
685 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
686 generate_content_request,
687 })?,
688 };
689 let request = http_client::Request::builder()
690 .method(Method::POST)
691 .uri(
692 http_client
693 .build_zed_llm_url("/count_tokens", &[])?
694 .as_ref(),
695 )
696 .header("Content-Type", "application/json")
697 .header("Authorization", format!("Bearer {token}"))
698 .body(serde_json::to_string(&request_body)?.into())?;
699 let mut response = http_client.send(request).await?;
700 let status = response.status();
701 let headers = response.headers().clone();
702 let mut response_body = String::new();
703 response
704 .body_mut()
705 .read_to_string(&mut response_body)
706 .await?;
707
708 if status.is_success() {
709 let response_body: CountTokensResponse =
710 serde_json::from_str(&response_body)?;
711
712 Ok(response_body.tokens as u64)
713 } else {
714 Err(anyhow!(ApiError {
715 status,
716 body: response_body,
717 headers
718 }))
719 }
720 }
721 .boxed()
722 }
723 }
724 }
725
726 fn stream_completion(
727 &self,
728 request: LanguageModelRequest,
729 cx: &AsyncApp,
730 ) -> BoxFuture<
731 'static,
732 Result<
733 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
734 LanguageModelCompletionError,
735 >,
736 > {
737 let thread_id = request.thread_id.clone();
738 let prompt_id = request.prompt_id.clone();
739 let intent = request.intent;
740 let mode = request.mode;
741 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
742 let thinking_allowed = request.thinking_allowed;
743 match self.model.provider {
744 cloud_llm_client::LanguageModelProvider::Anthropic => {
745 let request = into_anthropic(
746 request,
747 self.model.id.to_string(),
748 1.0,
749 self.model.max_output_tokens as u64,
750 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
751 AnthropicModelMode::Thinking {
752 budget_tokens: Some(4_096),
753 }
754 } else {
755 AnthropicModelMode::Default
756 },
757 );
758 let client = self.client.clone();
759 let llm_api_token = self.llm_api_token.clone();
760 let future = self.request_limiter.stream(async move {
761 let PerformLlmCompletionResponse {
762 response,
763 usage,
764 includes_status_messages,
765 tool_use_limit_reached,
766 } = Self::perform_llm_completion(
767 client.clone(),
768 llm_api_token,
769 app_version,
770 CompletionBody {
771 thread_id,
772 prompt_id,
773 intent,
774 mode,
775 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
776 model: request.model.clone(),
777 provider_request: serde_json::to_value(&request)
778 .map_err(|e| anyhow!(e))?,
779 },
780 )
781 .await
782 .map_err(|err| match err.downcast::<ApiError>() {
783 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
784 Err(err) => anyhow!(err),
785 })?;
786
787 let mut mapper = AnthropicEventMapper::new();
788 Ok(map_cloud_completion_events(
789 Box::pin(
790 response_lines(response, includes_status_messages)
791 .chain(usage_updated_event(usage))
792 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
793 ),
794 move |event| mapper.map_event(event),
795 ))
796 });
797 async move { Ok(future.await?.boxed()) }.boxed()
798 }
799 cloud_llm_client::LanguageModelProvider::OpenAi => {
800 let client = self.client.clone();
801 let model = match open_ai::Model::from_id(&self.model.id.0) {
802 Ok(model) => model,
803 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
804 };
805 let request = into_open_ai(
806 request,
807 model.id(),
808 model.supports_parallel_tool_calls(),
809 model.supports_prompt_cache_key(),
810 None,
811 None,
812 );
813 let llm_api_token = self.llm_api_token.clone();
814 let future = self.request_limiter.stream(async move {
815 let PerformLlmCompletionResponse {
816 response,
817 usage,
818 includes_status_messages,
819 tool_use_limit_reached,
820 } = Self::perform_llm_completion(
821 client.clone(),
822 llm_api_token,
823 app_version,
824 CompletionBody {
825 thread_id,
826 prompt_id,
827 intent,
828 mode,
829 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
830 model: request.model.clone(),
831 provider_request: serde_json::to_value(&request)
832 .map_err(|e| anyhow!(e))?,
833 },
834 )
835 .await?;
836
837 let mut mapper = OpenAiEventMapper::new();
838 Ok(map_cloud_completion_events(
839 Box::pin(
840 response_lines(response, includes_status_messages)
841 .chain(usage_updated_event(usage))
842 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
843 ),
844 move |event| mapper.map_event(event),
845 ))
846 });
847 async move { Ok(future.await?.boxed()) }.boxed()
848 }
849 cloud_llm_client::LanguageModelProvider::Google => {
850 let client = self.client.clone();
851 let request =
852 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
853 let llm_api_token = self.llm_api_token.clone();
854 let future = self.request_limiter.stream(async move {
855 let PerformLlmCompletionResponse {
856 response,
857 usage,
858 includes_status_messages,
859 tool_use_limit_reached,
860 } = Self::perform_llm_completion(
861 client.clone(),
862 llm_api_token,
863 app_version,
864 CompletionBody {
865 thread_id,
866 prompt_id,
867 intent,
868 mode,
869 provider: cloud_llm_client::LanguageModelProvider::Google,
870 model: request.model.model_id.clone(),
871 provider_request: serde_json::to_value(&request)
872 .map_err(|e| anyhow!(e))?,
873 },
874 )
875 .await?;
876
877 let mut mapper = GoogleEventMapper::new();
878 Ok(map_cloud_completion_events(
879 Box::pin(
880 response_lines(response, includes_status_messages)
881 .chain(usage_updated_event(usage))
882 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
883 ),
884 move |event| mapper.map_event(event),
885 ))
886 });
887 async move { Ok(future.await?.boxed()) }.boxed()
888 }
889 }
890 }
891}
892
893fn map_cloud_completion_events<T, F>(
894 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
895 mut map_callback: F,
896) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
897where
898 T: DeserializeOwned + 'static,
899 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
900 + Send
901 + 'static,
902{
903 stream
904 .flat_map(move |event| {
905 futures::stream::iter(match event {
906 Err(error) => {
907 vec![Err(LanguageModelCompletionError::from(error))]
908 }
909 Ok(CompletionEvent::Status(event)) => {
910 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
911 }
912 Ok(CompletionEvent::Event(event)) => map_callback(event),
913 })
914 })
915 .boxed()
916}
917
918fn usage_updated_event<T>(
919 usage: Option<ModelRequestUsage>,
920) -> impl Stream<Item = Result<CompletionEvent<T>>> {
921 futures::stream::iter(usage.map(|usage| {
922 Ok(CompletionEvent::Status(
923 CompletionRequestStatus::UsageUpdated {
924 amount: usage.amount as usize,
925 limit: usage.limit,
926 },
927 ))
928 }))
929}
930
931fn tool_use_limit_reached_event<T>(
932 tool_use_limit_reached: bool,
933) -> impl Stream<Item = Result<CompletionEvent<T>>> {
934 futures::stream::iter(tool_use_limit_reached.then(|| {
935 Ok(CompletionEvent::Status(
936 CompletionRequestStatus::ToolUseLimitReached,
937 ))
938 }))
939}
940
941fn response_lines<T: DeserializeOwned>(
942 response: Response<AsyncBody>,
943 includes_status_messages: bool,
944) -> impl Stream<Item = Result<CompletionEvent<T>>> {
945 futures::stream::try_unfold(
946 (String::new(), BufReader::new(response.into_body())),
947 move |(mut line, mut body)| async move {
948 match body.read_line(&mut line).await {
949 Ok(0) => Ok(None),
950 Ok(_) => {
951 let event = if includes_status_messages {
952 serde_json::from_str::<CompletionEvent<T>>(&line)?
953 } else {
954 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
955 };
956
957 line.clear();
958 Ok(Some((event, (line, body))))
959 }
960 Err(e) => Err(e.into()),
961 }
962 },
963 )
964}
965
966#[derive(IntoElement, RegisterComponent)]
967struct ZedAiConfiguration {
968 is_connected: bool,
969 plan: Option<Plan>,
970 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
971 eligible_for_trial: bool,
972 account_too_young: bool,
973 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
974}
975
976impl RenderOnce for ZedAiConfiguration {
977 fn render(self, _window: &mut Window, cx: &mut App) -> impl IntoElement {
978 let is_pro = self.plan.is_some_and(|plan| {
979 matches!(plan, Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro))
980 });
981 let is_free_v2 = self
982 .plan
983 .is_some_and(|plan| plan == Plan::V2(PlanV2::ZedFree));
984 let subscription_text = match (self.plan, self.subscription_period) {
985 (Some(Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro)), Some(_)) => {
986 "You have access to Zed's hosted models through your Pro subscription."
987 }
988 (Some(Plan::V1(PlanV1::ZedProTrial) | Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
989 "You have access to Zed's hosted models through your Pro trial."
990 }
991 (Some(Plan::V1(PlanV1::ZedFree)), Some(_)) => {
992 "You have basic access to Zed's hosted models through the Free plan."
993 }
994 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
995 if self.eligible_for_trial {
996 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
997 } else {
998 "Subscribe for access to Zed's hosted models."
999 }
1000 }
1001 _ => {
1002 if self.eligible_for_trial {
1003 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1004 } else {
1005 "Subscribe for access to Zed's hosted models."
1006 }
1007 }
1008 };
1009
1010 let manage_subscription_buttons = if is_pro {
1011 Button::new("manage_settings", "Manage Subscription")
1012 .full_width()
1013 .style(ButtonStyle::Tinted(TintColor::Accent))
1014 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1015 .into_any_element()
1016 } else if self.plan.is_none() || self.eligible_for_trial {
1017 Button::new("start_trial", "Start 14-day Free Pro Trial")
1018 .full_width()
1019 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1020 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1021 .into_any_element()
1022 } else {
1023 Button::new("upgrade", "Upgrade to Pro")
1024 .full_width()
1025 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1026 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1027 .into_any_element()
1028 };
1029
1030 if !self.is_connected {
1031 return v_flex()
1032 .gap_2()
1033 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1034 .child(
1035 Button::new("sign_in", "Sign In to use Zed AI")
1036 .icon_color(Color::Muted)
1037 .icon(IconName::Github)
1038 .icon_size(IconSize::Small)
1039 .icon_position(IconPosition::Start)
1040 .full_width()
1041 .on_click({
1042 let callback = self.sign_in_callback.clone();
1043 move |_, window, cx| (callback)(window, cx)
1044 }),
1045 );
1046 }
1047
1048 v_flex().gap_2().w_full().map(|this| {
1049 if self.account_too_young {
1050 this.child(YoungAccountBanner::new(
1051 is_free_v2 || cx.has_flag::<BillingV2FeatureFlag>(),
1052 ))
1053 .child(
1054 Button::new("upgrade", "Upgrade to Pro")
1055 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1056 .full_width()
1057 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1058 )
1059 } else {
1060 this.text_sm()
1061 .child(subscription_text)
1062 .child(manage_subscription_buttons)
1063 }
1064 })
1065 }
1066}
1067
1068struct ConfigurationView {
1069 state: Entity<State>,
1070 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1071}
1072
1073impl ConfigurationView {
1074 fn new(state: Entity<State>) -> Self {
1075 let sign_in_callback = Arc::new({
1076 let state = state.clone();
1077 move |_window: &mut Window, cx: &mut App| {
1078 state.update(cx, |state, cx| {
1079 state.authenticate(cx).detach_and_log_err(cx);
1080 });
1081 }
1082 });
1083
1084 Self {
1085 state,
1086 sign_in_callback,
1087 }
1088 }
1089}
1090
1091impl Render for ConfigurationView {
1092 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1093 let state = self.state.read(cx);
1094 let user_store = state.user_store.read(cx);
1095
1096 ZedAiConfiguration {
1097 is_connected: !state.is_signed_out(cx),
1098 plan: user_store.plan(),
1099 subscription_period: user_store.subscription_period(),
1100 eligible_for_trial: user_store.trial_started_at().is_none(),
1101 account_too_young: user_store.account_too_young(),
1102 sign_in_callback: self.sign_in_callback.clone(),
1103 }
1104 }
1105}
1106
1107impl Component for ZedAiConfiguration {
1108 fn name() -> &'static str {
1109 "AI Configuration Content"
1110 }
1111
1112 fn sort_name() -> &'static str {
1113 "AI Configuration Content"
1114 }
1115
1116 fn scope() -> ComponentScope {
1117 ComponentScope::Onboarding
1118 }
1119
1120 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1121 fn configuration(
1122 is_connected: bool,
1123 plan: Option<Plan>,
1124 eligible_for_trial: bool,
1125 account_too_young: bool,
1126 ) -> AnyElement {
1127 ZedAiConfiguration {
1128 is_connected,
1129 plan,
1130 subscription_period: plan
1131 .is_some()
1132 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1133 eligible_for_trial,
1134 account_too_young,
1135 sign_in_callback: Arc::new(|_, _| {}),
1136 }
1137 .into_any_element()
1138 }
1139
1140 Some(
1141 v_flex()
1142 .p_4()
1143 .gap_4()
1144 .children(vec![
1145 single_example("Not connected", configuration(false, None, false, false)),
1146 single_example(
1147 "Accept Terms of Service",
1148 configuration(true, None, true, false),
1149 ),
1150 single_example(
1151 "No Plan - Not eligible for trial",
1152 configuration(true, None, false, false),
1153 ),
1154 single_example(
1155 "No Plan - Eligible for trial",
1156 configuration(true, None, true, false),
1157 ),
1158 single_example(
1159 "Free Plan",
1160 configuration(true, Some(Plan::V1(PlanV1::ZedFree)), true, false),
1161 ),
1162 single_example(
1163 "Zed Pro Trial Plan",
1164 configuration(true, Some(Plan::V1(PlanV1::ZedProTrial)), true, false),
1165 ),
1166 single_example(
1167 "Zed Pro Plan",
1168 configuration(true, Some(Plan::V1(PlanV1::ZedPro)), true, false),
1169 ),
1170 ])
1171 .into_any_element(),
1172 )
1173 }
1174}
1175
1176#[cfg(test)]
1177mod tests {
1178 use super::*;
1179 use http_client::http::{HeaderMap, StatusCode};
1180 use language_model::LanguageModelCompletionError;
1181
1182 #[test]
1183 fn test_api_error_conversion_with_upstream_http_error() {
1184 // upstream_http_error with 503 status should become ServerOverloaded
1185 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}"#;
1186
1187 let api_error = ApiError {
1188 status: StatusCode::INTERNAL_SERVER_ERROR,
1189 body: error_body.to_string(),
1190 headers: HeaderMap::new(),
1191 };
1192
1193 let completion_error: LanguageModelCompletionError = api_error.into();
1194
1195 match completion_error {
1196 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1197 assert_eq!(
1198 message,
1199 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1200 );
1201 }
1202 _ => panic!(
1203 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1204 completion_error
1205 ),
1206 }
1207
1208 // upstream_http_error with 500 status should become ApiInternalServerError
1209 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1210
1211 let api_error = ApiError {
1212 status: StatusCode::INTERNAL_SERVER_ERROR,
1213 body: error_body.to_string(),
1214 headers: HeaderMap::new(),
1215 };
1216
1217 let completion_error: LanguageModelCompletionError = api_error.into();
1218
1219 match completion_error {
1220 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1221 assert_eq!(
1222 message,
1223 "Received an error from the OpenAI API: internal server error"
1224 );
1225 }
1226 _ => panic!(
1227 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1228 completion_error
1229 ),
1230 }
1231
1232 // upstream_http_error with 429 status should become RateLimitExceeded
1233 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1234
1235 let api_error = ApiError {
1236 status: StatusCode::INTERNAL_SERVER_ERROR,
1237 body: error_body.to_string(),
1238 headers: HeaderMap::new(),
1239 };
1240
1241 let completion_error: LanguageModelCompletionError = api_error.into();
1242
1243 match completion_error {
1244 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1245 assert_eq!(
1246 message,
1247 "Received an error from the Google API: rate limit exceeded"
1248 );
1249 }
1250 _ => panic!(
1251 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1252 completion_error
1253 ),
1254 }
1255
1256 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1257 let error_body = "Regular internal server error";
1258
1259 let api_error = ApiError {
1260 status: StatusCode::INTERNAL_SERVER_ERROR,
1261 body: error_body.to_string(),
1262 headers: HeaderMap::new(),
1263 };
1264
1265 let completion_error: LanguageModelCompletionError = api_error.into();
1266
1267 match completion_error {
1268 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1269 assert_eq!(provider, PROVIDER_NAME);
1270 assert_eq!(message, "Regular internal server error");
1271 }
1272 _ => panic!(
1273 "Expected ApiInternalServerError for regular 500, got: {:?}",
1274 completion_error
1275 ),
1276 }
1277
1278 // upstream_http_429 format should be converted to UpstreamProviderError
1279 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1280
1281 let api_error = ApiError {
1282 status: StatusCode::INTERNAL_SERVER_ERROR,
1283 body: error_body.to_string(),
1284 headers: HeaderMap::new(),
1285 };
1286
1287 let completion_error: LanguageModelCompletionError = api_error.into();
1288
1289 match completion_error {
1290 LanguageModelCompletionError::UpstreamProviderError {
1291 message,
1292 status,
1293 retry_after,
1294 } => {
1295 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1296 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1297 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1298 }
1299 _ => panic!(
1300 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1301 completion_error
1302 ),
1303 }
1304
1305 // Invalid JSON in error body should fall back to regular error handling
1306 let error_body = "Not JSON at all";
1307
1308 let api_error = ApiError {
1309 status: StatusCode::INTERNAL_SERVER_ERROR,
1310 body: error_body.to_string(),
1311 headers: HeaderMap::new(),
1312 };
1313
1314 let completion_error: LanguageModelCompletionError = api_error.into();
1315
1316 match completion_error {
1317 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1318 assert_eq!(provider, PROVIDER_NAME);
1319 }
1320 _ => panic!(
1321 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1322 completion_error
1323 ),
1324 }
1325 }
1326}