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::{
19 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
20};
21use http_client::http::{HeaderMap, HeaderValue};
22use http_client::{AsyncBody, HttpClient, HttpRequestExt, Method, Response, StatusCode};
23use language_model::{
24 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
25 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
26 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
27 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
28 LanguageModelToolSchemaFormat, LlmApiToken, ModelRequestLimitReachedError,
29 PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
30};
31use release_channel::AppVersion;
32use schemars::JsonSchema;
33use serde::{Deserialize, Serialize, de::DeserializeOwned};
34use settings::SettingsStore;
35pub use settings::ZedDotDevAvailableModel as AvailableModel;
36pub use settings::ZedDotDevAvailableProvider as AvailableProvider;
37use smol::io::{AsyncReadExt, BufReader};
38use std::pin::Pin;
39use std::str::FromStr as _;
40use std::sync::Arc;
41use std::time::Duration;
42use thiserror::Error;
43use ui::{TintColor, prelude::*};
44use util::{ResultExt as _, maybe};
45
46use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
47use crate::provider::google::{GoogleEventMapper, into_google};
48use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
49use crate::provider::x_ai::count_xai_tokens;
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: 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 .header(CLIENT_SUPPORTS_X_AI_HEADER_NAME, "true")
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<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 let provider_name = provider_name(&self.model.provider);
756 match self.model.provider {
757 cloud_llm_client::LanguageModelProvider::Anthropic => {
758 let request = into_anthropic(
759 request,
760 self.model.id.to_string(),
761 1.0,
762 self.model.max_output_tokens as u64,
763 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
764 AnthropicModelMode::Thinking {
765 budget_tokens: Some(4_096),
766 }
767 } else {
768 AnthropicModelMode::Default
769 },
770 );
771 let client = self.client.clone();
772 let llm_api_token = self.llm_api_token.clone();
773 let future = self.request_limiter.stream(async move {
774 let PerformLlmCompletionResponse {
775 response,
776 usage,
777 includes_status_messages,
778 tool_use_limit_reached,
779 } = Self::perform_llm_completion(
780 client.clone(),
781 llm_api_token,
782 app_version,
783 CompletionBody {
784 thread_id,
785 prompt_id,
786 intent,
787 mode,
788 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
789 model: request.model.clone(),
790 provider_request: serde_json::to_value(&request)
791 .map_err(|e| anyhow!(e))?,
792 },
793 )
794 .await
795 .map_err(|err| match err.downcast::<ApiError>() {
796 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
797 Err(err) => anyhow!(err),
798 })?;
799
800 let mut mapper = AnthropicEventMapper::new();
801 Ok(map_cloud_completion_events(
802 Box::pin(
803 response_lines(response, includes_status_messages)
804 .chain(usage_updated_event(usage))
805 .chain(tool_use_limit_reached_event(tool_use_limit_reached)), // .map(|_| {}),
806 ),
807 &provider_name,
808 move |event| mapper.map_event(event),
809 ))
810 });
811 async move { Ok(future.await?.boxed()) }.boxed()
812 }
813 cloud_llm_client::LanguageModelProvider::OpenAi => {
814 let client = self.client.clone();
815 let request = into_open_ai(
816 request,
817 &self.model.id.0,
818 self.model.supports_parallel_tool_calls,
819 true,
820 None,
821 None,
822 );
823 let llm_api_token = self.llm_api_token.clone();
824 let future = self.request_limiter.stream(async move {
825 let PerformLlmCompletionResponse {
826 response,
827 usage,
828 includes_status_messages,
829 tool_use_limit_reached,
830 } = Self::perform_llm_completion(
831 client.clone(),
832 llm_api_token,
833 app_version,
834 CompletionBody {
835 thread_id,
836 prompt_id,
837 intent,
838 mode,
839 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
840 model: request.model.clone(),
841 provider_request: serde_json::to_value(&request)
842 .map_err(|e| anyhow!(e))?,
843 },
844 )
845 .await?;
846
847 let mut mapper = OpenAiEventMapper::new();
848 Ok(map_cloud_completion_events(
849 Box::pin(
850 response_lines(response, includes_status_messages)
851 .chain(usage_updated_event(usage))
852 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
853 ),
854 &provider_name,
855 move |event| mapper.map_event(event),
856 ))
857 });
858 async move { Ok(future.await?.boxed()) }.boxed()
859 }
860 cloud_llm_client::LanguageModelProvider::XAi => {
861 let client = self.client.clone();
862 let request = into_open_ai(
863 request,
864 &self.model.id.0,
865 self.model.supports_parallel_tool_calls,
866 false,
867 None,
868 None,
869 );
870 let llm_api_token = self.llm_api_token.clone();
871 let future = self.request_limiter.stream(async move {
872 let PerformLlmCompletionResponse {
873 response,
874 usage,
875 includes_status_messages,
876 tool_use_limit_reached,
877 } = Self::perform_llm_completion(
878 client.clone(),
879 llm_api_token,
880 app_version,
881 CompletionBody {
882 thread_id,
883 prompt_id,
884 intent,
885 mode,
886 provider: cloud_llm_client::LanguageModelProvider::XAi,
887 model: request.model.clone(),
888 provider_request: serde_json::to_value(&request)
889 .map_err(|e| anyhow!(e))?,
890 },
891 )
892 .await?;
893
894 let mut mapper = OpenAiEventMapper::new();
895 Ok(map_cloud_completion_events(
896 Box::pin(
897 response_lines(response, includes_status_messages)
898 .chain(usage_updated_event(usage))
899 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
900 ),
901 &provider_name,
902 move |event| mapper.map_event(event),
903 ))
904 });
905 async move { Ok(future.await?.boxed()) }.boxed()
906 }
907 cloud_llm_client::LanguageModelProvider::Google => {
908 let client = self.client.clone();
909 let request =
910 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
911 let llm_api_token = self.llm_api_token.clone();
912 let future = self.request_limiter.stream(async move {
913 let PerformLlmCompletionResponse {
914 response,
915 usage,
916 includes_status_messages,
917 tool_use_limit_reached,
918 } = Self::perform_llm_completion(
919 client.clone(),
920 llm_api_token,
921 app_version,
922 CompletionBody {
923 thread_id,
924 prompt_id,
925 intent,
926 mode,
927 provider: cloud_llm_client::LanguageModelProvider::Google,
928 model: request.model.model_id.clone(),
929 provider_request: serde_json::to_value(&request)
930 .map_err(|e| anyhow!(e))?,
931 },
932 )
933 .await?;
934
935 let mut mapper = GoogleEventMapper::new();
936 Ok(map_cloud_completion_events(
937 Box::pin(
938 response_lines(response, includes_status_messages)
939 .chain(usage_updated_event(usage))
940 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
941 ),
942 &provider_name,
943 move |event| mapper.map_event(event),
944 ))
945 });
946 async move { Ok(future.await?.boxed()) }.boxed()
947 }
948 }
949 }
950}
951
952fn map_cloud_completion_events<T, F>(
953 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
954 provider: &LanguageModelProviderName,
955 mut map_callback: F,
956) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
957where
958 T: DeserializeOwned + 'static,
959 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
960 + Send
961 + 'static,
962{
963 let provider = provider.clone();
964 stream
965 .flat_map(move |event| {
966 futures::stream::iter(match event {
967 Err(error) => {
968 vec![Err(LanguageModelCompletionError::from(error))]
969 }
970 Ok(CompletionEvent::Status(event)) => {
971 vec![
972 LanguageModelCompletionEvent::from_completion_request_status(
973 event,
974 provider.clone(),
975 ),
976 ]
977 }
978 Ok(CompletionEvent::Event(event)) => map_callback(event),
979 })
980 })
981 .boxed()
982}
983
984fn provider_name(provider: &cloud_llm_client::LanguageModelProvider) -> LanguageModelProviderName {
985 match provider {
986 cloud_llm_client::LanguageModelProvider::Anthropic => {
987 language_model::ANTHROPIC_PROVIDER_NAME
988 }
989 cloud_llm_client::LanguageModelProvider::OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
990 cloud_llm_client::LanguageModelProvider::Google => language_model::GOOGLE_PROVIDER_NAME,
991 cloud_llm_client::LanguageModelProvider::XAi => language_model::X_AI_PROVIDER_NAME,
992 }
993}
994
995fn usage_updated_event<T>(
996 usage: Option<ModelRequestUsage>,
997) -> impl Stream<Item = Result<CompletionEvent<T>>> {
998 futures::stream::iter(usage.map(|usage| {
999 Ok(CompletionEvent::Status(
1000 CompletionRequestStatus::UsageUpdated {
1001 amount: usage.amount as usize,
1002 limit: usage.limit,
1003 },
1004 ))
1005 }))
1006}
1007
1008fn tool_use_limit_reached_event<T>(
1009 tool_use_limit_reached: bool,
1010) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1011 futures::stream::iter(tool_use_limit_reached.then(|| {
1012 Ok(CompletionEvent::Status(
1013 CompletionRequestStatus::ToolUseLimitReached,
1014 ))
1015 }))
1016}
1017
1018fn response_lines<T: DeserializeOwned>(
1019 response: Response<AsyncBody>,
1020 includes_status_messages: bool,
1021) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1022 futures::stream::try_unfold(
1023 (String::new(), BufReader::new(response.into_body())),
1024 move |(mut line, mut body)| async move {
1025 match body.read_line(&mut line).await {
1026 Ok(0) => Ok(None),
1027 Ok(_) => {
1028 let event = if includes_status_messages {
1029 serde_json::from_str::<CompletionEvent<T>>(&line)?
1030 } else {
1031 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1032 };
1033
1034 line.clear();
1035 Ok(Some((event, (line, body))))
1036 }
1037 Err(e) => Err(e.into()),
1038 }
1039 },
1040 )
1041}
1042
1043#[derive(IntoElement, RegisterComponent)]
1044struct ZedAiConfiguration {
1045 is_connected: bool,
1046 plan: Option<Plan>,
1047 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1048 eligible_for_trial: bool,
1049 account_too_young: bool,
1050 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1051}
1052
1053impl RenderOnce for ZedAiConfiguration {
1054 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1055 let is_pro = self.plan.is_some_and(|plan| {
1056 matches!(plan, Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro))
1057 });
1058 let subscription_text = match (self.plan, self.subscription_period) {
1059 (Some(Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro)), Some(_)) => {
1060 "You have access to Zed's hosted models through your Pro subscription."
1061 }
1062 (Some(Plan::V1(PlanV1::ZedProTrial) | Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1063 "You have access to Zed's hosted models through your Pro trial."
1064 }
1065 (Some(Plan::V1(PlanV1::ZedFree)), Some(_)) => {
1066 "You have basic access to Zed's hosted models through the Free plan."
1067 }
1068 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1069 if self.eligible_for_trial {
1070 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1071 } else {
1072 "Subscribe for access to Zed's hosted models."
1073 }
1074 }
1075 _ => {
1076 if self.eligible_for_trial {
1077 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1078 } else {
1079 "Subscribe for access to Zed's hosted models."
1080 }
1081 }
1082 };
1083
1084 let manage_subscription_buttons = if is_pro {
1085 Button::new("manage_settings", "Manage Subscription")
1086 .full_width()
1087 .style(ButtonStyle::Tinted(TintColor::Accent))
1088 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1089 .into_any_element()
1090 } else if self.plan.is_none() || self.eligible_for_trial {
1091 Button::new("start_trial", "Start 14-day Free Pro Trial")
1092 .full_width()
1093 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1094 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1095 .into_any_element()
1096 } else {
1097 Button::new("upgrade", "Upgrade to Pro")
1098 .full_width()
1099 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1100 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1101 .into_any_element()
1102 };
1103
1104 if !self.is_connected {
1105 return v_flex()
1106 .gap_2()
1107 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1108 .child(
1109 Button::new("sign_in", "Sign In to use Zed AI")
1110 .icon_color(Color::Muted)
1111 .icon(IconName::Github)
1112 .icon_size(IconSize::Small)
1113 .icon_position(IconPosition::Start)
1114 .full_width()
1115 .on_click({
1116 let callback = self.sign_in_callback.clone();
1117 move |_, window, cx| (callback)(window, cx)
1118 }),
1119 );
1120 }
1121
1122 v_flex().gap_2().w_full().map(|this| {
1123 if self.account_too_young {
1124 this.child(YoungAccountBanner).child(
1125 Button::new("upgrade", "Upgrade to Pro")
1126 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1127 .full_width()
1128 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1129 )
1130 } else {
1131 this.text_sm()
1132 .child(subscription_text)
1133 .child(manage_subscription_buttons)
1134 }
1135 })
1136 }
1137}
1138
1139struct ConfigurationView {
1140 state: Entity<State>,
1141 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1142}
1143
1144impl ConfigurationView {
1145 fn new(state: Entity<State>) -> Self {
1146 let sign_in_callback = Arc::new({
1147 let state = state.clone();
1148 move |_window: &mut Window, cx: &mut App| {
1149 state.update(cx, |state, cx| {
1150 state.authenticate(cx).detach_and_log_err(cx);
1151 });
1152 }
1153 });
1154
1155 Self {
1156 state,
1157 sign_in_callback,
1158 }
1159 }
1160}
1161
1162impl Render for ConfigurationView {
1163 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1164 let state = self.state.read(cx);
1165 let user_store = state.user_store.read(cx);
1166
1167 ZedAiConfiguration {
1168 is_connected: !state.is_signed_out(cx),
1169 plan: user_store.plan(),
1170 subscription_period: user_store.subscription_period(),
1171 eligible_for_trial: user_store.trial_started_at().is_none(),
1172 account_too_young: user_store.account_too_young(),
1173 sign_in_callback: self.sign_in_callback.clone(),
1174 }
1175 }
1176}
1177
1178impl Component for ZedAiConfiguration {
1179 fn name() -> &'static str {
1180 "AI Configuration Content"
1181 }
1182
1183 fn sort_name() -> &'static str {
1184 "AI Configuration Content"
1185 }
1186
1187 fn scope() -> ComponentScope {
1188 ComponentScope::Onboarding
1189 }
1190
1191 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1192 fn configuration(
1193 is_connected: bool,
1194 plan: Option<Plan>,
1195 eligible_for_trial: bool,
1196 account_too_young: bool,
1197 ) -> AnyElement {
1198 ZedAiConfiguration {
1199 is_connected,
1200 plan,
1201 subscription_period: plan
1202 .is_some()
1203 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1204 eligible_for_trial,
1205 account_too_young,
1206 sign_in_callback: Arc::new(|_, _| {}),
1207 }
1208 .into_any_element()
1209 }
1210
1211 Some(
1212 v_flex()
1213 .p_4()
1214 .gap_4()
1215 .children(vec![
1216 single_example("Not connected", configuration(false, None, false, false)),
1217 single_example(
1218 "Accept Terms of Service",
1219 configuration(true, None, true, false),
1220 ),
1221 single_example(
1222 "No Plan - Not eligible for trial",
1223 configuration(true, None, false, false),
1224 ),
1225 single_example(
1226 "No Plan - Eligible for trial",
1227 configuration(true, None, true, false),
1228 ),
1229 single_example(
1230 "Free Plan",
1231 configuration(true, Some(Plan::V1(PlanV1::ZedFree)), true, false),
1232 ),
1233 single_example(
1234 "Zed Pro Trial Plan",
1235 configuration(true, Some(Plan::V1(PlanV1::ZedProTrial)), true, false),
1236 ),
1237 single_example(
1238 "Zed Pro Plan",
1239 configuration(true, Some(Plan::V1(PlanV1::ZedPro)), true, false),
1240 ),
1241 ])
1242 .into_any_element(),
1243 )
1244 }
1245}
1246
1247#[cfg(test)]
1248mod tests {
1249 use super::*;
1250 use http_client::http::{HeaderMap, StatusCode};
1251 use language_model::LanguageModelCompletionError;
1252
1253 #[test]
1254 fn test_api_error_conversion_with_upstream_http_error() {
1255 // upstream_http_error with 503 status should become ServerOverloaded
1256 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}"#;
1257
1258 let api_error = ApiError {
1259 status: StatusCode::INTERNAL_SERVER_ERROR,
1260 body: error_body.to_string(),
1261 headers: HeaderMap::new(),
1262 };
1263
1264 let completion_error: LanguageModelCompletionError = api_error.into();
1265
1266 match completion_error {
1267 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1268 assert_eq!(
1269 message,
1270 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1271 );
1272 }
1273 _ => panic!(
1274 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1275 completion_error
1276 ),
1277 }
1278
1279 // upstream_http_error with 500 status should become ApiInternalServerError
1280 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1281
1282 let api_error = ApiError {
1283 status: StatusCode::INTERNAL_SERVER_ERROR,
1284 body: error_body.to_string(),
1285 headers: HeaderMap::new(),
1286 };
1287
1288 let completion_error: LanguageModelCompletionError = api_error.into();
1289
1290 match completion_error {
1291 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1292 assert_eq!(
1293 message,
1294 "Received an error from the OpenAI API: internal server error"
1295 );
1296 }
1297 _ => panic!(
1298 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1299 completion_error
1300 ),
1301 }
1302
1303 // upstream_http_error with 429 status should become RateLimitExceeded
1304 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1305
1306 let api_error = ApiError {
1307 status: StatusCode::INTERNAL_SERVER_ERROR,
1308 body: error_body.to_string(),
1309 headers: HeaderMap::new(),
1310 };
1311
1312 let completion_error: LanguageModelCompletionError = api_error.into();
1313
1314 match completion_error {
1315 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1316 assert_eq!(
1317 message,
1318 "Received an error from the Google API: rate limit exceeded"
1319 );
1320 }
1321 _ => panic!(
1322 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1323 completion_error
1324 ),
1325 }
1326
1327 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1328 let error_body = "Regular internal server error";
1329
1330 let api_error = ApiError {
1331 status: StatusCode::INTERNAL_SERVER_ERROR,
1332 body: error_body.to_string(),
1333 headers: HeaderMap::new(),
1334 };
1335
1336 let completion_error: LanguageModelCompletionError = api_error.into();
1337
1338 match completion_error {
1339 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1340 assert_eq!(provider, PROVIDER_NAME);
1341 assert_eq!(message, "Regular internal server error");
1342 }
1343 _ => panic!(
1344 "Expected ApiInternalServerError for regular 500, got: {:?}",
1345 completion_error
1346 ),
1347 }
1348
1349 // upstream_http_429 format should be converted to UpstreamProviderError
1350 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1351
1352 let api_error = ApiError {
1353 status: StatusCode::INTERNAL_SERVER_ERROR,
1354 body: error_body.to_string(),
1355 headers: HeaderMap::new(),
1356 };
1357
1358 let completion_error: LanguageModelCompletionError = api_error.into();
1359
1360 match completion_error {
1361 LanguageModelCompletionError::UpstreamProviderError {
1362 message,
1363 status,
1364 retry_after,
1365 } => {
1366 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1367 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1368 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1369 }
1370 _ => panic!(
1371 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1372 completion_error
1373 ),
1374 }
1375
1376 // Invalid JSON in error body should fall back to regular error handling
1377 let error_body = "Not JSON at all";
1378
1379 let api_error = ApiError {
1380 status: StatusCode::INTERNAL_SERVER_ERROR,
1381 body: error_body.to_string(),
1382 headers: HeaderMap::new(),
1383 };
1384
1385 let completion_error: LanguageModelCompletionError = api_error.into();
1386
1387 match completion_error {
1388 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1389 assert_eq!(provider, PROVIDER_NAME);
1390 }
1391 _ => panic!(
1392 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1393 completion_error
1394 ),
1395 }
1396 }
1397}