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