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 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 .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<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 XAi => language_model::X_AI_PROVIDER_ID,
584 }
585 }
586
587 fn upstream_provider_name(&self) -> LanguageModelProviderName {
588 use cloud_llm_client::LanguageModelProvider::*;
589 match self.model.provider {
590 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
591 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
592 Google => language_model::GOOGLE_PROVIDER_NAME,
593 XAi => language_model::X_AI_PROVIDER_NAME,
594 }
595 }
596
597 fn supports_tools(&self) -> bool {
598 self.model.supports_tools
599 }
600
601 fn supports_images(&self) -> bool {
602 self.model.supports_images
603 }
604
605 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
606 match choice {
607 LanguageModelToolChoice::Auto
608 | LanguageModelToolChoice::Any
609 | LanguageModelToolChoice::None => true,
610 }
611 }
612
613 fn supports_burn_mode(&self) -> bool {
614 self.model.supports_max_mode
615 }
616
617 fn telemetry_id(&self) -> String {
618 format!("zed.dev/{}", self.model.id)
619 }
620
621 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
622 match self.model.provider {
623 cloud_llm_client::LanguageModelProvider::Anthropic
624 | cloud_llm_client::LanguageModelProvider::OpenAi
625 | cloud_llm_client::LanguageModelProvider::XAi => {
626 LanguageModelToolSchemaFormat::JsonSchema
627 }
628 cloud_llm_client::LanguageModelProvider::Google => {
629 LanguageModelToolSchemaFormat::JsonSchemaSubset
630 }
631 }
632 }
633
634 fn max_token_count(&self) -> u64 {
635 self.model.max_token_count as u64
636 }
637
638 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
639 self.model
640 .max_token_count_in_max_mode
641 .filter(|_| self.model.supports_max_mode)
642 .map(|max_token_count| max_token_count as u64)
643 }
644
645 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
646 match &self.model.provider {
647 cloud_llm_client::LanguageModelProvider::Anthropic => {
648 Some(LanguageModelCacheConfiguration {
649 min_total_token: 2_048,
650 should_speculate: true,
651 max_cache_anchors: 4,
652 })
653 }
654 cloud_llm_client::LanguageModelProvider::OpenAi
655 | cloud_llm_client::LanguageModelProvider::XAi
656 | cloud_llm_client::LanguageModelProvider::Google => None,
657 }
658 }
659
660 fn count_tokens(
661 &self,
662 request: LanguageModelRequest,
663 cx: &App,
664 ) -> BoxFuture<'static, Result<u64>> {
665 match self.model.provider {
666 cloud_llm_client::LanguageModelProvider::Anthropic => {
667 count_anthropic_tokens(request, cx)
668 }
669 cloud_llm_client::LanguageModelProvider::OpenAi => {
670 let model = match open_ai::Model::from_id(&self.model.id.0) {
671 Ok(model) => model,
672 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
673 };
674 count_open_ai_tokens(request, model, cx)
675 }
676 cloud_llm_client::LanguageModelProvider::XAi => {
677 let model = match x_ai::Model::from_id(&self.model.id.0) {
678 Ok(model) => model,
679 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
680 };
681 count_xai_tokens(request, model, cx)
682 }
683 cloud_llm_client::LanguageModelProvider::Google => {
684 let client = self.client.clone();
685 let llm_api_token = self.llm_api_token.clone();
686 let model_id = self.model.id.to_string();
687 let generate_content_request =
688 into_google(request, model_id.clone(), GoogleModelMode::Default);
689 async move {
690 let http_client = &client.http_client();
691 let token = llm_api_token.acquire(&client).await?;
692
693 let request_body = CountTokensBody {
694 provider: cloud_llm_client::LanguageModelProvider::Google,
695 model: model_id,
696 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
697 generate_content_request,
698 })?,
699 };
700 let request = http_client::Request::builder()
701 .method(Method::POST)
702 .uri(
703 http_client
704 .build_zed_llm_url("/count_tokens", &[])?
705 .as_ref(),
706 )
707 .header("Content-Type", "application/json")
708 .header("Authorization", format!("Bearer {token}"))
709 .body(serde_json::to_string(&request_body)?.into())?;
710 let mut response = http_client.send(request).await?;
711 let status = response.status();
712 let headers = response.headers().clone();
713 let mut response_body = String::new();
714 response
715 .body_mut()
716 .read_to_string(&mut response_body)
717 .await?;
718
719 if status.is_success() {
720 let response_body: CountTokensResponse =
721 serde_json::from_str(&response_body)?;
722
723 Ok(response_body.tokens as u64)
724 } else {
725 Err(anyhow!(ApiError {
726 status,
727 body: response_body,
728 headers
729 }))
730 }
731 }
732 .boxed()
733 }
734 }
735 }
736
737 fn stream_completion(
738 &self,
739 request: LanguageModelRequest,
740 cx: &AsyncApp,
741 ) -> BoxFuture<
742 'static,
743 Result<
744 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
745 LanguageModelCompletionError,
746 >,
747 > {
748 let thread_id = request.thread_id.clone();
749 let prompt_id = request.prompt_id.clone();
750 let intent = request.intent;
751 let mode = request.mode;
752 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
753 let thinking_allowed = request.thinking_allowed;
754 match self.model.provider {
755 cloud_llm_client::LanguageModelProvider::Anthropic => {
756 let request = into_anthropic(
757 request,
758 self.model.id.to_string(),
759 1.0,
760 self.model.max_output_tokens as u64,
761 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
762 AnthropicModelMode::Thinking {
763 budget_tokens: Some(4_096),
764 }
765 } else {
766 AnthropicModelMode::Default
767 },
768 );
769 let client = self.client.clone();
770 let llm_api_token = self.llm_api_token.clone();
771 let future = self.request_limiter.stream(async move {
772 let PerformLlmCompletionResponse {
773 response,
774 usage,
775 includes_status_messages,
776 tool_use_limit_reached,
777 } = Self::perform_llm_completion(
778 client.clone(),
779 llm_api_token,
780 app_version,
781 CompletionBody {
782 thread_id,
783 prompt_id,
784 intent,
785 mode,
786 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
787 model: request.model.clone(),
788 provider_request: serde_json::to_value(&request)
789 .map_err(|e| anyhow!(e))?,
790 },
791 )
792 .await
793 .map_err(|err| match err.downcast::<ApiError>() {
794 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
795 Err(err) => anyhow!(err),
796 })?;
797
798 let mut mapper = AnthropicEventMapper::new();
799 Ok(map_cloud_completion_events(
800 Box::pin(
801 response_lines(response, includes_status_messages)
802 .chain(usage_updated_event(usage))
803 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
804 ),
805 move |event| mapper.map_event(event),
806 ))
807 });
808 async move { Ok(future.await?.boxed()) }.boxed()
809 }
810 cloud_llm_client::LanguageModelProvider::OpenAi => {
811 let client = self.client.clone();
812 let model = match open_ai::Model::from_id(&self.model.id.0) {
813 Ok(model) => model,
814 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
815 };
816 let request = into_open_ai(
817 request,
818 model.id(),
819 model.supports_parallel_tool_calls(),
820 model.supports_prompt_cache_key(),
821 None,
822 None,
823 );
824 let llm_api_token = self.llm_api_token.clone();
825 let future = self.request_limiter.stream(async move {
826 let PerformLlmCompletionResponse {
827 response,
828 usage,
829 includes_status_messages,
830 tool_use_limit_reached,
831 } = Self::perform_llm_completion(
832 client.clone(),
833 llm_api_token,
834 app_version,
835 CompletionBody {
836 thread_id,
837 prompt_id,
838 intent,
839 mode,
840 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
841 model: request.model.clone(),
842 provider_request: serde_json::to_value(&request)
843 .map_err(|e| anyhow!(e))?,
844 },
845 )
846 .await?;
847
848 let mut mapper = OpenAiEventMapper::new();
849 Ok(map_cloud_completion_events(
850 Box::pin(
851 response_lines(response, includes_status_messages)
852 .chain(usage_updated_event(usage))
853 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
854 ),
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 model = match x_ai::Model::from_id(&self.model.id.0) {
863 Ok(model) => model,
864 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
865 };
866 let request = into_open_ai(
867 request,
868 model.id(),
869 model.supports_parallel_tool_calls(),
870 model.supports_prompt_cache_key(),
871 None,
872 None,
873 );
874 let llm_api_token = self.llm_api_token.clone();
875 let future = self.request_limiter.stream(async move {
876 let PerformLlmCompletionResponse {
877 response,
878 usage,
879 includes_status_messages,
880 tool_use_limit_reached,
881 } = Self::perform_llm_completion(
882 client.clone(),
883 llm_api_token,
884 app_version,
885 CompletionBody {
886 thread_id,
887 prompt_id,
888 intent,
889 mode,
890 provider: cloud_llm_client::LanguageModelProvider::XAi,
891 model: request.model.clone(),
892 provider_request: serde_json::to_value(&request)
893 .map_err(|e| anyhow!(e))?,
894 },
895 )
896 .await?;
897
898 let mut mapper = OpenAiEventMapper::new();
899 Ok(map_cloud_completion_events(
900 Box::pin(
901 response_lines(response, includes_status_messages)
902 .chain(usage_updated_event(usage))
903 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
904 ),
905 move |event| mapper.map_event(event),
906 ))
907 });
908 async move { Ok(future.await?.boxed()) }.boxed()
909 }
910 cloud_llm_client::LanguageModelProvider::Google => {
911 let client = self.client.clone();
912 let request =
913 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
914 let llm_api_token = self.llm_api_token.clone();
915 let future = self.request_limiter.stream(async move {
916 let PerformLlmCompletionResponse {
917 response,
918 usage,
919 includes_status_messages,
920 tool_use_limit_reached,
921 } = Self::perform_llm_completion(
922 client.clone(),
923 llm_api_token,
924 app_version,
925 CompletionBody {
926 thread_id,
927 prompt_id,
928 intent,
929 mode,
930 provider: cloud_llm_client::LanguageModelProvider::Google,
931 model: request.model.model_id.clone(),
932 provider_request: serde_json::to_value(&request)
933 .map_err(|e| anyhow!(e))?,
934 },
935 )
936 .await?;
937
938 let mut mapper = GoogleEventMapper::new();
939 Ok(map_cloud_completion_events(
940 Box::pin(
941 response_lines(response, includes_status_messages)
942 .chain(usage_updated_event(usage))
943 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
944 ),
945 move |event| mapper.map_event(event),
946 ))
947 });
948 async move { Ok(future.await?.boxed()) }.boxed()
949 }
950 }
951 }
952}
953
954fn map_cloud_completion_events<T, F>(
955 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
956 mut map_callback: F,
957) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
958where
959 T: DeserializeOwned + 'static,
960 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
961 + Send
962 + 'static,
963{
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![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
972 }
973 Ok(CompletionEvent::Event(event)) => map_callback(event),
974 })
975 })
976 .boxed()
977}
978
979fn usage_updated_event<T>(
980 usage: Option<ModelRequestUsage>,
981) -> impl Stream<Item = Result<CompletionEvent<T>>> {
982 futures::stream::iter(usage.map(|usage| {
983 Ok(CompletionEvent::Status(
984 CompletionRequestStatus::UsageUpdated {
985 amount: usage.amount as usize,
986 limit: usage.limit,
987 },
988 ))
989 }))
990}
991
992fn tool_use_limit_reached_event<T>(
993 tool_use_limit_reached: bool,
994) -> impl Stream<Item = Result<CompletionEvent<T>>> {
995 futures::stream::iter(tool_use_limit_reached.then(|| {
996 Ok(CompletionEvent::Status(
997 CompletionRequestStatus::ToolUseLimitReached,
998 ))
999 }))
1000}
1001
1002fn response_lines<T: DeserializeOwned>(
1003 response: Response<AsyncBody>,
1004 includes_status_messages: bool,
1005) -> impl Stream<Item = Result<CompletionEvent<T>>> {
1006 futures::stream::try_unfold(
1007 (String::new(), BufReader::new(response.into_body())),
1008 move |(mut line, mut body)| async move {
1009 match body.read_line(&mut line).await {
1010 Ok(0) => Ok(None),
1011 Ok(_) => {
1012 let event = if includes_status_messages {
1013 serde_json::from_str::<CompletionEvent<T>>(&line)?
1014 } else {
1015 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1016 };
1017
1018 line.clear();
1019 Ok(Some((event, (line, body))))
1020 }
1021 Err(e) => Err(e.into()),
1022 }
1023 },
1024 )
1025}
1026
1027#[derive(IntoElement, RegisterComponent)]
1028struct ZedAiConfiguration {
1029 is_connected: bool,
1030 plan: Option<Plan>,
1031 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1032 eligible_for_trial: bool,
1033 account_too_young: bool,
1034 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1035}
1036
1037impl RenderOnce for ZedAiConfiguration {
1038 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1039 let is_pro = self.plan.is_some_and(|plan| {
1040 matches!(plan, Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro))
1041 });
1042 let subscription_text = match (self.plan, self.subscription_period) {
1043 (Some(Plan::V1(PlanV1::ZedPro) | Plan::V2(PlanV2::ZedPro)), Some(_)) => {
1044 "You have access to Zed's hosted models through your Pro subscription."
1045 }
1046 (Some(Plan::V1(PlanV1::ZedProTrial) | Plan::V2(PlanV2::ZedProTrial)), Some(_)) => {
1047 "You have access to Zed's hosted models through your Pro trial."
1048 }
1049 (Some(Plan::V1(PlanV1::ZedFree)), Some(_)) => {
1050 "You have basic access to Zed's hosted models through the Free plan."
1051 }
1052 (Some(Plan::V2(PlanV2::ZedFree)), Some(_)) => {
1053 if self.eligible_for_trial {
1054 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1055 } else {
1056 "Subscribe for access to Zed's hosted models."
1057 }
1058 }
1059 _ => {
1060 if self.eligible_for_trial {
1061 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1062 } else {
1063 "Subscribe for access to Zed's hosted models."
1064 }
1065 }
1066 };
1067
1068 let manage_subscription_buttons = if is_pro {
1069 Button::new("manage_settings", "Manage Subscription")
1070 .full_width()
1071 .style(ButtonStyle::Tinted(TintColor::Accent))
1072 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1073 .into_any_element()
1074 } else if self.plan.is_none() || self.eligible_for_trial {
1075 Button::new("start_trial", "Start 14-day Free Pro Trial")
1076 .full_width()
1077 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1078 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1079 .into_any_element()
1080 } else {
1081 Button::new("upgrade", "Upgrade to Pro")
1082 .full_width()
1083 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1084 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1085 .into_any_element()
1086 };
1087
1088 if !self.is_connected {
1089 return v_flex()
1090 .gap_2()
1091 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1092 .child(
1093 Button::new("sign_in", "Sign In to use Zed AI")
1094 .icon_color(Color::Muted)
1095 .icon(IconName::Github)
1096 .icon_size(IconSize::Small)
1097 .icon_position(IconPosition::Start)
1098 .full_width()
1099 .on_click({
1100 let callback = self.sign_in_callback.clone();
1101 move |_, window, cx| (callback)(window, cx)
1102 }),
1103 );
1104 }
1105
1106 v_flex().gap_2().w_full().map(|this| {
1107 if self.account_too_young {
1108 this.child(YoungAccountBanner).child(
1109 Button::new("upgrade", "Upgrade to Pro")
1110 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1111 .full_width()
1112 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1113 )
1114 } else {
1115 this.text_sm()
1116 .child(subscription_text)
1117 .child(manage_subscription_buttons)
1118 }
1119 })
1120 }
1121}
1122
1123struct ConfigurationView {
1124 state: Entity<State>,
1125 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1126}
1127
1128impl ConfigurationView {
1129 fn new(state: Entity<State>) -> Self {
1130 let sign_in_callback = Arc::new({
1131 let state = state.clone();
1132 move |_window: &mut Window, cx: &mut App| {
1133 state.update(cx, |state, cx| {
1134 state.authenticate(cx).detach_and_log_err(cx);
1135 });
1136 }
1137 });
1138
1139 Self {
1140 state,
1141 sign_in_callback,
1142 }
1143 }
1144}
1145
1146impl Render for ConfigurationView {
1147 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1148 let state = self.state.read(cx);
1149 let user_store = state.user_store.read(cx);
1150
1151 ZedAiConfiguration {
1152 is_connected: !state.is_signed_out(cx),
1153 plan: user_store.plan(),
1154 subscription_period: user_store.subscription_period(),
1155 eligible_for_trial: user_store.trial_started_at().is_none(),
1156 account_too_young: user_store.account_too_young(),
1157 sign_in_callback: self.sign_in_callback.clone(),
1158 }
1159 }
1160}
1161
1162impl Component for ZedAiConfiguration {
1163 fn name() -> &'static str {
1164 "AI Configuration Content"
1165 }
1166
1167 fn sort_name() -> &'static str {
1168 "AI Configuration Content"
1169 }
1170
1171 fn scope() -> ComponentScope {
1172 ComponentScope::Onboarding
1173 }
1174
1175 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1176 fn configuration(
1177 is_connected: bool,
1178 plan: Option<Plan>,
1179 eligible_for_trial: bool,
1180 account_too_young: bool,
1181 ) -> AnyElement {
1182 ZedAiConfiguration {
1183 is_connected,
1184 plan,
1185 subscription_period: plan
1186 .is_some()
1187 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1188 eligible_for_trial,
1189 account_too_young,
1190 sign_in_callback: Arc::new(|_, _| {}),
1191 }
1192 .into_any_element()
1193 }
1194
1195 Some(
1196 v_flex()
1197 .p_4()
1198 .gap_4()
1199 .children(vec![
1200 single_example("Not connected", configuration(false, None, false, false)),
1201 single_example(
1202 "Accept Terms of Service",
1203 configuration(true, None, true, false),
1204 ),
1205 single_example(
1206 "No Plan - Not eligible for trial",
1207 configuration(true, None, false, false),
1208 ),
1209 single_example(
1210 "No Plan - Eligible for trial",
1211 configuration(true, None, true, false),
1212 ),
1213 single_example(
1214 "Free Plan",
1215 configuration(true, Some(Plan::V1(PlanV1::ZedFree)), true, false),
1216 ),
1217 single_example(
1218 "Zed Pro Trial Plan",
1219 configuration(true, Some(Plan::V1(PlanV1::ZedProTrial)), true, false),
1220 ),
1221 single_example(
1222 "Zed Pro Plan",
1223 configuration(true, Some(Plan::V1(PlanV1::ZedPro)), true, false),
1224 ),
1225 ])
1226 .into_any_element(),
1227 )
1228 }
1229}
1230
1231#[cfg(test)]
1232mod tests {
1233 use super::*;
1234 use http_client::http::{HeaderMap, StatusCode};
1235 use language_model::LanguageModelCompletionError;
1236
1237 #[test]
1238 fn test_api_error_conversion_with_upstream_http_error() {
1239 // upstream_http_error with 503 status should become ServerOverloaded
1240 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}"#;
1241
1242 let api_error = ApiError {
1243 status: StatusCode::INTERNAL_SERVER_ERROR,
1244 body: error_body.to_string(),
1245 headers: HeaderMap::new(),
1246 };
1247
1248 let completion_error: LanguageModelCompletionError = api_error.into();
1249
1250 match completion_error {
1251 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1252 assert_eq!(
1253 message,
1254 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1255 );
1256 }
1257 _ => panic!(
1258 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1259 completion_error
1260 ),
1261 }
1262
1263 // upstream_http_error with 500 status should become ApiInternalServerError
1264 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1265
1266 let api_error = ApiError {
1267 status: StatusCode::INTERNAL_SERVER_ERROR,
1268 body: error_body.to_string(),
1269 headers: HeaderMap::new(),
1270 };
1271
1272 let completion_error: LanguageModelCompletionError = api_error.into();
1273
1274 match completion_error {
1275 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1276 assert_eq!(
1277 message,
1278 "Received an error from the OpenAI API: internal server error"
1279 );
1280 }
1281 _ => panic!(
1282 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1283 completion_error
1284 ),
1285 }
1286
1287 // upstream_http_error with 429 status should become RateLimitExceeded
1288 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1289
1290 let api_error = ApiError {
1291 status: StatusCode::INTERNAL_SERVER_ERROR,
1292 body: error_body.to_string(),
1293 headers: HeaderMap::new(),
1294 };
1295
1296 let completion_error: LanguageModelCompletionError = api_error.into();
1297
1298 match completion_error {
1299 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1300 assert_eq!(
1301 message,
1302 "Received an error from the Google API: rate limit exceeded"
1303 );
1304 }
1305 _ => panic!(
1306 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1307 completion_error
1308 ),
1309 }
1310
1311 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1312 let error_body = "Regular internal server error";
1313
1314 let api_error = ApiError {
1315 status: StatusCode::INTERNAL_SERVER_ERROR,
1316 body: error_body.to_string(),
1317 headers: HeaderMap::new(),
1318 };
1319
1320 let completion_error: LanguageModelCompletionError = api_error.into();
1321
1322 match completion_error {
1323 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1324 assert_eq!(provider, PROVIDER_NAME);
1325 assert_eq!(message, "Regular internal server error");
1326 }
1327 _ => panic!(
1328 "Expected ApiInternalServerError for regular 500, got: {:?}",
1329 completion_error
1330 ),
1331 }
1332
1333 // upstream_http_429 format should be converted to UpstreamProviderError
1334 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1335
1336 let api_error = ApiError {
1337 status: StatusCode::INTERNAL_SERVER_ERROR,
1338 body: error_body.to_string(),
1339 headers: HeaderMap::new(),
1340 };
1341
1342 let completion_error: LanguageModelCompletionError = api_error.into();
1343
1344 match completion_error {
1345 LanguageModelCompletionError::UpstreamProviderError {
1346 message,
1347 status,
1348 retry_after,
1349 } => {
1350 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1351 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1352 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1353 }
1354 _ => panic!(
1355 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1356 completion_error
1357 ),
1358 }
1359
1360 // Invalid JSON in error body should fall back to regular error handling
1361 let error_body = "Not JSON at all";
1362
1363 let api_error = ApiError {
1364 status: StatusCode::INTERNAL_SERVER_ERROR,
1365 body: error_body.to_string(),
1366 headers: HeaderMap::new(),
1367 };
1368
1369 let completion_error: LanguageModelCompletionError = api_error.into();
1370
1371 match completion_error {
1372 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1373 assert_eq!(provider, PROVIDER_NAME);
1374 }
1375 _ => panic!(
1376 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1377 completion_error
1378 ),
1379 }
1380 }
1381}