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