cloud.rs

  1use super::open_ai::count_open_ai_tokens;
  2use anthropic::AnthropicError;
  3use anyhow::{anyhow, Result};
  4use client::{
  5    zed_urls, Client, PerformCompletionParams, UserStore, EXPIRED_LLM_TOKEN_HEADER_NAME,
  6    MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
  7};
  8use collections::BTreeMap;
  9use feature_flags::{FeatureFlagAppExt, LlmClosedBeta, ZedPro};
 10use futures::{
 11    future::BoxFuture, stream::BoxStream, AsyncBufReadExt, FutureExt, Stream, StreamExt,
 12    TryStreamExt as _,
 13};
 14use gpui::{
 15    AnyElement, AnyView, AppContext, AsyncAppContext, EventEmitter, FontWeight, Global, Model,
 16    ModelContext, ReadGlobal, Subscription, Task,
 17};
 18use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
 19use language_model::{
 20    CloudModel, LanguageModel, LanguageModelCacheConfiguration, LanguageModelId, LanguageModelName,
 21    LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
 22    LanguageModelRequest, RateLimiter, ZED_CLOUD_PROVIDER_ID,
 23};
 24use language_model::{
 25    LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider,
 26};
 27use proto::TypedEnvelope;
 28use schemars::JsonSchema;
 29use serde::{de::DeserializeOwned, Deserialize, Serialize};
 30use serde_json::value::RawValue;
 31use settings::{Settings, SettingsStore};
 32use smol::{
 33    io::{AsyncReadExt, BufReader},
 34    lock::{RwLock, RwLockUpgradableReadGuard, RwLockWriteGuard},
 35};
 36use std::fmt;
 37use std::{
 38    future,
 39    sync::{Arc, LazyLock},
 40};
 41use strum::IntoEnumIterator;
 42use thiserror::Error;
 43use ui::{prelude::*, TintColor};
 44
 45use crate::provider::anthropic::map_to_language_model_completion_events;
 46use crate::AllLanguageModelSettings;
 47
 48use super::anthropic::count_anthropic_tokens;
 49
 50pub const PROVIDER_NAME: &str = "Zed";
 51
 52const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
 53    option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
 54
 55fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
 56    static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
 57        ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
 58            .map(|json| serde_json::from_str(json).unwrap())
 59            .unwrap_or_default()
 60    });
 61    ADDITIONAL_MODELS.as_slice()
 62}
 63
 64#[derive(Default, Clone, Debug, PartialEq)]
 65pub struct ZedDotDevSettings {
 66    pub available_models: Vec<AvailableModel>,
 67}
 68
 69#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
 70#[serde(rename_all = "lowercase")]
 71pub enum AvailableProvider {
 72    Anthropic,
 73    OpenAi,
 74    Google,
 75}
 76
 77#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
 78pub struct AvailableModel {
 79    /// The provider of the language model.
 80    pub provider: AvailableProvider,
 81    /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
 82    pub name: String,
 83    /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
 84    pub display_name: Option<String>,
 85    /// The size of the context window, indicating the maximum number of tokens the model can process.
 86    pub max_tokens: usize,
 87    /// The maximum number of output tokens allowed by the model.
 88    pub max_output_tokens: Option<u32>,
 89    /// The maximum number of completion tokens allowed by the model (o1-* only)
 90    pub max_completion_tokens: Option<u32>,
 91    /// Override this model with a different Anthropic model for tool calls.
 92    pub tool_override: Option<String>,
 93    /// Indicates whether this custom model supports caching.
 94    pub cache_configuration: Option<LanguageModelCacheConfiguration>,
 95    /// The default temperature to use for this model.
 96    pub default_temperature: Option<f32>,
 97}
 98
 99struct GlobalRefreshLlmTokenListener(Model<RefreshLlmTokenListener>);
100
101impl Global for GlobalRefreshLlmTokenListener {}
102
103pub struct RefreshLlmTokenEvent;
104
105pub struct RefreshLlmTokenListener {
106    _llm_token_subscription: client::Subscription,
107}
108
109impl EventEmitter<RefreshLlmTokenEvent> for RefreshLlmTokenListener {}
110
111impl RefreshLlmTokenListener {
112    pub fn register(client: Arc<Client>, cx: &mut AppContext) {
113        let listener = cx.new_model(|cx| RefreshLlmTokenListener::new(client, cx));
114        cx.set_global(GlobalRefreshLlmTokenListener(listener));
115    }
116
117    pub fn global(cx: &AppContext) -> Model<Self> {
118        GlobalRefreshLlmTokenListener::global(cx).0.clone()
119    }
120
121    fn new(client: Arc<Client>, cx: &mut ModelContext<Self>) -> Self {
122        Self {
123            _llm_token_subscription: client
124                .add_message_handler(cx.weak_model(), Self::handle_refresh_llm_token),
125        }
126    }
127
128    async fn handle_refresh_llm_token(
129        this: Model<Self>,
130        _: TypedEnvelope<proto::RefreshLlmToken>,
131        mut cx: AsyncAppContext,
132    ) -> Result<()> {
133        this.update(&mut cx, |_this, cx| cx.emit(RefreshLlmTokenEvent))
134    }
135}
136
137pub struct CloudLanguageModelProvider {
138    client: Arc<Client>,
139    state: gpui::Model<State>,
140    _maintain_client_status: Task<()>,
141}
142
143pub struct State {
144    client: Arc<Client>,
145    llm_api_token: LlmApiToken,
146    user_store: Model<UserStore>,
147    status: client::Status,
148    accept_terms: Option<Task<Result<()>>>,
149    _settings_subscription: Subscription,
150    _llm_token_subscription: Subscription,
151}
152
153impl State {
154    fn new(
155        client: Arc<Client>,
156        user_store: Model<UserStore>,
157        status: client::Status,
158        cx: &mut ModelContext<Self>,
159    ) -> Self {
160        let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
161
162        Self {
163            client: client.clone(),
164            llm_api_token: LlmApiToken::default(),
165            user_store,
166            status,
167            accept_terms: None,
168            _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
169                cx.notify();
170            }),
171            _llm_token_subscription: cx.subscribe(
172                &refresh_llm_token_listener,
173                |this, _listener, _event, cx| {
174                    let client = this.client.clone();
175                    let llm_api_token = this.llm_api_token.clone();
176                    cx.spawn(|_this, _cx| async move {
177                        llm_api_token.refresh(&client).await?;
178                        anyhow::Ok(())
179                    })
180                    .detach_and_log_err(cx);
181                },
182            ),
183        }
184    }
185
186    fn is_signed_out(&self) -> bool {
187        self.status.is_signed_out()
188    }
189
190    fn authenticate(&self, cx: &mut ModelContext<Self>) -> Task<Result<()>> {
191        let client = self.client.clone();
192        cx.spawn(move |this, mut cx| async move {
193            client.authenticate_and_connect(true, &cx).await?;
194            this.update(&mut cx, |_, cx| cx.notify())
195        })
196    }
197
198    fn has_accepted_terms_of_service(&self, cx: &AppContext) -> bool {
199        self.user_store
200            .read(cx)
201            .current_user_has_accepted_terms()
202            .unwrap_or(false)
203    }
204
205    fn accept_terms_of_service(&mut self, cx: &mut ModelContext<Self>) {
206        let user_store = self.user_store.clone();
207        self.accept_terms = Some(cx.spawn(move |this, mut cx| async move {
208            let _ = user_store
209                .update(&mut cx, |store, cx| store.accept_terms_of_service(cx))?
210                .await;
211            this.update(&mut cx, |this, cx| {
212                this.accept_terms = None;
213                cx.notify()
214            })
215        }));
216    }
217}
218
219impl CloudLanguageModelProvider {
220    pub fn new(user_store: Model<UserStore>, client: Arc<Client>, cx: &mut AppContext) -> Self {
221        let mut status_rx = client.status();
222        let status = *status_rx.borrow();
223
224        let state = cx.new_model(|cx| State::new(client.clone(), user_store.clone(), status, cx));
225
226        let state_ref = state.downgrade();
227        let maintain_client_status = cx.spawn(|mut cx| async move {
228            while let Some(status) = status_rx.next().await {
229                if let Some(this) = state_ref.upgrade() {
230                    _ = this.update(&mut cx, |this, cx| {
231                        if this.status != status {
232                            this.status = status;
233                            cx.notify();
234                        }
235                    });
236                } else {
237                    break;
238                }
239            }
240        });
241
242        Self {
243            client,
244            state: state.clone(),
245            _maintain_client_status: maintain_client_status,
246        }
247    }
248}
249
250impl LanguageModelProviderState for CloudLanguageModelProvider {
251    type ObservableEntity = State;
252
253    fn observable_entity(&self) -> Option<gpui::Model<Self::ObservableEntity>> {
254        Some(self.state.clone())
255    }
256}
257
258impl LanguageModelProvider for CloudLanguageModelProvider {
259    fn id(&self) -> LanguageModelProviderId {
260        LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
261    }
262
263    fn name(&self) -> LanguageModelProviderName {
264        LanguageModelProviderName(PROVIDER_NAME.into())
265    }
266
267    fn icon(&self) -> IconName {
268        IconName::AiZed
269    }
270
271    fn provided_models(&self, cx: &AppContext) -> Vec<Arc<dyn LanguageModel>> {
272        let mut models = BTreeMap::default();
273
274        if cx.is_staff() {
275            for model in anthropic::Model::iter() {
276                if !matches!(model, anthropic::Model::Custom { .. }) {
277                    models.insert(model.id().to_string(), CloudModel::Anthropic(model));
278                }
279            }
280            for model in open_ai::Model::iter() {
281                if !matches!(model, open_ai::Model::Custom { .. }) {
282                    models.insert(model.id().to_string(), CloudModel::OpenAi(model));
283                }
284            }
285            for model in google_ai::Model::iter() {
286                if !matches!(model, google_ai::Model::Custom { .. }) {
287                    models.insert(model.id().to_string(), CloudModel::Google(model));
288                }
289            }
290        } else {
291            models.insert(
292                anthropic::Model::Claude3_5Sonnet.id().to_string(),
293                CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
294            );
295        }
296
297        let llm_closed_beta_models = if cx.has_flag::<LlmClosedBeta>() {
298            zed_cloud_provider_additional_models()
299        } else {
300            &[]
301        };
302
303        // Override with available models from settings
304        for model in AllLanguageModelSettings::get_global(cx)
305            .zed_dot_dev
306            .available_models
307            .iter()
308            .chain(llm_closed_beta_models)
309            .cloned()
310        {
311            let model = match model.provider {
312                AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
313                    name: model.name.clone(),
314                    display_name: model.display_name.clone(),
315                    max_tokens: model.max_tokens,
316                    tool_override: model.tool_override.clone(),
317                    cache_configuration: model.cache_configuration.as_ref().map(|config| {
318                        anthropic::AnthropicModelCacheConfiguration {
319                            max_cache_anchors: config.max_cache_anchors,
320                            should_speculate: config.should_speculate,
321                            min_total_token: config.min_total_token,
322                        }
323                    }),
324                    default_temperature: model.default_temperature,
325                    max_output_tokens: model.max_output_tokens,
326                }),
327                AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
328                    name: model.name.clone(),
329                    display_name: model.display_name.clone(),
330                    max_tokens: model.max_tokens,
331                    max_output_tokens: model.max_output_tokens,
332                    max_completion_tokens: model.max_completion_tokens,
333                }),
334                AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
335                    name: model.name.clone(),
336                    display_name: model.display_name.clone(),
337                    max_tokens: model.max_tokens,
338                }),
339            };
340            models.insert(model.id().to_string(), model.clone());
341        }
342
343        let llm_api_token = self.state.read(cx).llm_api_token.clone();
344        models
345            .into_values()
346            .map(|model| {
347                Arc::new(CloudLanguageModel {
348                    id: LanguageModelId::from(model.id().to_string()),
349                    model,
350                    llm_api_token: llm_api_token.clone(),
351                    client: self.client.clone(),
352                    request_limiter: RateLimiter::new(4),
353                }) as Arc<dyn LanguageModel>
354            })
355            .collect()
356    }
357
358    fn is_authenticated(&self, cx: &AppContext) -> bool {
359        !self.state.read(cx).is_signed_out()
360    }
361
362    fn authenticate(&self, _cx: &mut AppContext) -> Task<Result<()>> {
363        Task::ready(Ok(()))
364    }
365
366    fn configuration_view(&self, cx: &mut WindowContext) -> AnyView {
367        cx.new_view(|_cx| ConfigurationView {
368            state: self.state.clone(),
369        })
370        .into()
371    }
372
373    fn must_accept_terms(&self, cx: &AppContext) -> bool {
374        !self.state.read(cx).has_accepted_terms_of_service(cx)
375    }
376
377    fn render_accept_terms(&self, cx: &mut WindowContext) -> Option<AnyElement> {
378        let state = self.state.read(cx);
379
380        let terms = [(
381            "terms_of_service",
382            "Terms of Service",
383            "https://zed.dev/terms-of-service",
384        )]
385        .map(|(id, label, url)| {
386            Button::new(id, label)
387                .style(ButtonStyle::Subtle)
388                .icon(IconName::ExternalLink)
389                .icon_size(IconSize::XSmall)
390                .icon_color(Color::Muted)
391                .on_click(move |_, cx| cx.open_url(url))
392        });
393
394        if state.has_accepted_terms_of_service(cx) {
395            None
396        } else {
397            let disabled = state.accept_terms.is_some();
398            Some(
399                v_flex()
400                    .gap_2()
401                    .child(
402                        v_flex()
403                            .child(Label::new("Terms and Conditions").weight(FontWeight::MEDIUM))
404                            .child(
405                                Label::new(
406                                    "Please read and accept our terms and conditions to continue.",
407                                )
408                                .size(LabelSize::Small),
409                            ),
410                    )
411                    .child(v_flex().gap_1().children(terms))
412                    .child(
413                        h_flex().justify_end().child(
414                            Button::new("accept_terms", "I've read it and accept it")
415                                .disabled(disabled)
416                                .on_click({
417                                    let state = self.state.downgrade();
418                                    move |_, cx| {
419                                        state
420                                            .update(cx, |state, cx| {
421                                                state.accept_terms_of_service(cx)
422                                            })
423                                            .ok();
424                                    }
425                                }),
426                        ),
427                    )
428                    .into_any(),
429            )
430        }
431    }
432
433    fn reset_credentials(&self, _cx: &mut AppContext) -> Task<Result<()>> {
434        Task::ready(Ok(()))
435    }
436}
437
438pub struct CloudLanguageModel {
439    id: LanguageModelId,
440    model: CloudModel,
441    llm_api_token: LlmApiToken,
442    client: Arc<Client>,
443    request_limiter: RateLimiter,
444}
445
446#[derive(Clone, Default)]
447struct LlmApiToken(Arc<RwLock<Option<String>>>);
448
449#[derive(Error, Debug)]
450pub struct PaymentRequiredError;
451
452impl fmt::Display for PaymentRequiredError {
453    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
454        write!(
455            f,
456            "Payment required to use this language model. Please upgrade your account."
457        )
458    }
459}
460
461#[derive(Error, Debug)]
462pub struct MaxMonthlySpendReachedError;
463
464impl fmt::Display for MaxMonthlySpendReachedError {
465    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
466        write!(
467            f,
468            "Maximum spending limit reached for this month. For more usage, increase your spending limit."
469        )
470    }
471}
472
473impl CloudLanguageModel {
474    async fn perform_llm_completion(
475        client: Arc<Client>,
476        llm_api_token: LlmApiToken,
477        body: PerformCompletionParams,
478    ) -> Result<Response<AsyncBody>> {
479        let http_client = &client.http_client();
480
481        let mut token = llm_api_token.acquire(&client).await?;
482        let mut did_retry = false;
483
484        let response = loop {
485            let request_builder = http_client::Request::builder();
486            let request = request_builder
487                .method(Method::POST)
488                .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
489                .header("Content-Type", "application/json")
490                .header("Authorization", format!("Bearer {token}"))
491                .body(serde_json::to_string(&body)?.into())?;
492            let mut response = http_client.send(request).await?;
493            if response.status().is_success() {
494                break response;
495            } else if !did_retry
496                && response
497                    .headers()
498                    .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
499                    .is_some()
500            {
501                did_retry = true;
502                token = llm_api_token.refresh(&client).await?;
503            } else if response.status() == StatusCode::FORBIDDEN
504                && response
505                    .headers()
506                    .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
507                    .is_some()
508            {
509                break Err(anyhow!(MaxMonthlySpendReachedError))?;
510            } else if response.status() == StatusCode::PAYMENT_REQUIRED {
511                break Err(anyhow!(PaymentRequiredError))?;
512            } else {
513                let mut body = String::new();
514                response.body_mut().read_to_string(&mut body).await?;
515                break Err(anyhow!(
516                    "cloud language model completion failed with status {}: {body}",
517                    response.status()
518                ))?;
519            }
520        };
521
522        Ok(response)
523    }
524}
525
526impl LanguageModel for CloudLanguageModel {
527    fn id(&self) -> LanguageModelId {
528        self.id.clone()
529    }
530
531    fn name(&self) -> LanguageModelName {
532        LanguageModelName::from(self.model.display_name().to_string())
533    }
534
535    fn icon(&self) -> Option<IconName> {
536        self.model.icon()
537    }
538
539    fn provider_id(&self) -> LanguageModelProviderId {
540        LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
541    }
542
543    fn provider_name(&self) -> LanguageModelProviderName {
544        LanguageModelProviderName(PROVIDER_NAME.into())
545    }
546
547    fn telemetry_id(&self) -> String {
548        format!("zed.dev/{}", self.model.id())
549    }
550
551    fn availability(&self) -> LanguageModelAvailability {
552        self.model.availability()
553    }
554
555    fn max_token_count(&self) -> usize {
556        self.model.max_token_count()
557    }
558
559    fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
560        match &self.model {
561            CloudModel::Anthropic(model) => {
562                model
563                    .cache_configuration()
564                    .map(|cache| LanguageModelCacheConfiguration {
565                        max_cache_anchors: cache.max_cache_anchors,
566                        should_speculate: cache.should_speculate,
567                        min_total_token: cache.min_total_token,
568                    })
569            }
570            CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
571        }
572    }
573
574    fn count_tokens(
575        &self,
576        request: LanguageModelRequest,
577        cx: &AppContext,
578    ) -> BoxFuture<'static, Result<usize>> {
579        match self.model.clone() {
580            CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
581            CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
582            CloudModel::Google(model) => {
583                let client = self.client.clone();
584                let request = request.into_google(model.id().into());
585                let request = google_ai::CountTokensRequest {
586                    contents: request.contents,
587                };
588                async move {
589                    let request = serde_json::to_string(&request)?;
590                    let response = client
591                        .request(proto::CountLanguageModelTokens {
592                            provider: proto::LanguageModelProvider::Google as i32,
593                            request,
594                        })
595                        .await?;
596                    Ok(response.token_count as usize)
597                }
598                .boxed()
599            }
600        }
601    }
602
603    fn stream_completion(
604        &self,
605        request: LanguageModelRequest,
606        _cx: &AsyncAppContext,
607    ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
608        match &self.model {
609            CloudModel::Anthropic(model) => {
610                let request = request.into_anthropic(
611                    model.id().into(),
612                    model.default_temperature(),
613                    model.max_output_tokens(),
614                );
615                let client = self.client.clone();
616                let llm_api_token = self.llm_api_token.clone();
617                let future = self.request_limiter.stream(async move {
618                    let response = Self::perform_llm_completion(
619                        client.clone(),
620                        llm_api_token,
621                        PerformCompletionParams {
622                            provider: client::LanguageModelProvider::Anthropic,
623                            model: request.model.clone(),
624                            provider_request: RawValue::from_string(serde_json::to_string(
625                                &request,
626                            )?)?,
627                        },
628                    )
629                    .await?;
630                    Ok(map_to_language_model_completion_events(Box::pin(
631                        response_lines(response).map_err(AnthropicError::Other),
632                    )))
633                });
634                async move { Ok(future.await?.boxed()) }.boxed()
635            }
636            CloudModel::OpenAi(model) => {
637                let client = self.client.clone();
638                let request = request.into_open_ai(model.id().into(), model.max_output_tokens());
639                let llm_api_token = self.llm_api_token.clone();
640                let future = self.request_limiter.stream(async move {
641                    let response = Self::perform_llm_completion(
642                        client.clone(),
643                        llm_api_token,
644                        PerformCompletionParams {
645                            provider: client::LanguageModelProvider::OpenAi,
646                            model: request.model.clone(),
647                            provider_request: RawValue::from_string(serde_json::to_string(
648                                &request,
649                            )?)?,
650                        },
651                    )
652                    .await?;
653                    Ok(open_ai::extract_text_from_events(response_lines(response)))
654                });
655                async move {
656                    Ok(future
657                        .await?
658                        .map(|result| result.map(LanguageModelCompletionEvent::Text))
659                        .boxed())
660                }
661                .boxed()
662            }
663            CloudModel::Google(model) => {
664                let client = self.client.clone();
665                let request = request.into_google(model.id().into());
666                let llm_api_token = self.llm_api_token.clone();
667                let future = self.request_limiter.stream(async move {
668                    let response = Self::perform_llm_completion(
669                        client.clone(),
670                        llm_api_token,
671                        PerformCompletionParams {
672                            provider: client::LanguageModelProvider::Google,
673                            model: request.model.clone(),
674                            provider_request: RawValue::from_string(serde_json::to_string(
675                                &request,
676                            )?)?,
677                        },
678                    )
679                    .await?;
680                    Ok(google_ai::extract_text_from_events(response_lines(
681                        response,
682                    )))
683                });
684                async move {
685                    Ok(future
686                        .await?
687                        .map(|result| result.map(LanguageModelCompletionEvent::Text))
688                        .boxed())
689                }
690                .boxed()
691            }
692        }
693    }
694
695    fn use_any_tool(
696        &self,
697        request: LanguageModelRequest,
698        tool_name: String,
699        tool_description: String,
700        input_schema: serde_json::Value,
701        _cx: &AsyncAppContext,
702    ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
703        let client = self.client.clone();
704        let llm_api_token = self.llm_api_token.clone();
705
706        match &self.model {
707            CloudModel::Anthropic(model) => {
708                let mut request = request.into_anthropic(
709                    model.tool_model_id().into(),
710                    model.default_temperature(),
711                    model.max_output_tokens(),
712                );
713                request.tool_choice = Some(anthropic::ToolChoice::Tool {
714                    name: tool_name.clone(),
715                });
716                request.tools = vec![anthropic::Tool {
717                    name: tool_name.clone(),
718                    description: tool_description,
719                    input_schema,
720                }];
721
722                self.request_limiter
723                    .run(async move {
724                        let response = Self::perform_llm_completion(
725                            client.clone(),
726                            llm_api_token,
727                            PerformCompletionParams {
728                                provider: client::LanguageModelProvider::Anthropic,
729                                model: request.model.clone(),
730                                provider_request: RawValue::from_string(serde_json::to_string(
731                                    &request,
732                                )?)?,
733                            },
734                        )
735                        .await?;
736
737                        Ok(anthropic::extract_tool_args_from_events(
738                            tool_name,
739                            Box::pin(response_lines(response)),
740                        )
741                        .await?
742                        .boxed())
743                    })
744                    .boxed()
745            }
746            CloudModel::OpenAi(model) => {
747                let mut request =
748                    request.into_open_ai(model.id().into(), model.max_output_tokens());
749                request.tool_choice = Some(open_ai::ToolChoice::Other(
750                    open_ai::ToolDefinition::Function {
751                        function: open_ai::FunctionDefinition {
752                            name: tool_name.clone(),
753                            description: None,
754                            parameters: None,
755                        },
756                    },
757                ));
758                request.tools = vec![open_ai::ToolDefinition::Function {
759                    function: open_ai::FunctionDefinition {
760                        name: tool_name.clone(),
761                        description: Some(tool_description),
762                        parameters: Some(input_schema),
763                    },
764                }];
765
766                self.request_limiter
767                    .run(async move {
768                        let response = Self::perform_llm_completion(
769                            client.clone(),
770                            llm_api_token,
771                            PerformCompletionParams {
772                                provider: client::LanguageModelProvider::OpenAi,
773                                model: request.model.clone(),
774                                provider_request: RawValue::from_string(serde_json::to_string(
775                                    &request,
776                                )?)?,
777                            },
778                        )
779                        .await?;
780
781                        Ok(open_ai::extract_tool_args_from_events(
782                            tool_name,
783                            Box::pin(response_lines(response)),
784                        )
785                        .await?
786                        .boxed())
787                    })
788                    .boxed()
789            }
790            CloudModel::Google(_) => {
791                future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
792            }
793        }
794    }
795}
796
797fn response_lines<T: DeserializeOwned>(
798    response: Response<AsyncBody>,
799) -> impl Stream<Item = Result<T>> {
800    futures::stream::try_unfold(
801        (String::new(), BufReader::new(response.into_body())),
802        move |(mut line, mut body)| async {
803            match body.read_line(&mut line).await {
804                Ok(0) => Ok(None),
805                Ok(_) => {
806                    let event: T = serde_json::from_str(&line)?;
807                    line.clear();
808                    Ok(Some((event, (line, body))))
809                }
810                Err(e) => Err(e.into()),
811            }
812        },
813    )
814}
815
816impl LlmApiToken {
817    async fn acquire(&self, client: &Arc<Client>) -> Result<String> {
818        let lock = self.0.upgradable_read().await;
819        if let Some(token) = lock.as_ref() {
820            Ok(token.to_string())
821        } else {
822            Self::fetch(RwLockUpgradableReadGuard::upgrade(lock).await, client).await
823        }
824    }
825
826    async fn refresh(&self, client: &Arc<Client>) -> Result<String> {
827        Self::fetch(self.0.write().await, client).await
828    }
829
830    async fn fetch<'a>(
831        mut lock: RwLockWriteGuard<'a, Option<String>>,
832        client: &Arc<Client>,
833    ) -> Result<String> {
834        let response = client.request(proto::GetLlmToken {}).await?;
835        *lock = Some(response.token.clone());
836        Ok(response.token.clone())
837    }
838}
839
840struct ConfigurationView {
841    state: gpui::Model<State>,
842}
843
844impl ConfigurationView {
845    fn authenticate(&mut self, cx: &mut ViewContext<Self>) {
846        self.state.update(cx, |state, cx| {
847            state.authenticate(cx).detach_and_log_err(cx);
848        });
849        cx.notify();
850    }
851
852    fn render_accept_terms(&mut self, cx: &mut ViewContext<Self>) -> Option<AnyElement> {
853        if self.state.read(cx).has_accepted_terms_of_service(cx) {
854            return None;
855        }
856
857        let accept_terms_disabled = self.state.read(cx).accept_terms.is_some();
858
859        let terms_button = Button::new("terms_of_service", "Terms of Service")
860            .style(ButtonStyle::Subtle)
861            .icon(IconName::ExternalLink)
862            .icon_color(Color::Muted)
863            .on_click(move |_, cx| cx.open_url("https://zed.dev/terms-of-service"));
864
865        let text =
866            "In order to use Zed AI, please read and accept our terms and conditions to continue:";
867
868        let form = v_flex()
869            .gap_2()
870            .child(Label::new("Terms and Conditions"))
871            .child(Label::new(text))
872            .child(h_flex().justify_center().child(terms_button))
873            .child(
874                h_flex().justify_center().child(
875                    Button::new("accept_terms", "I've read and accept the terms of service")
876                        .style(ButtonStyle::Tinted(TintColor::Accent))
877                        .disabled(accept_terms_disabled)
878                        .on_click({
879                            let state = self.state.downgrade();
880                            move |_, cx| {
881                                state
882                                    .update(cx, |state, cx| state.accept_terms_of_service(cx))
883                                    .ok();
884                            }
885                        }),
886                ),
887            );
888
889        Some(form.into_any())
890    }
891}
892
893impl Render for ConfigurationView {
894    fn render(&mut self, cx: &mut ViewContext<Self>) -> impl IntoElement {
895        const ZED_AI_URL: &str = "https://zed.dev/ai";
896
897        let is_connected = !self.state.read(cx).is_signed_out();
898        let plan = self.state.read(cx).user_store.read(cx).current_plan();
899        let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
900
901        let is_pro = plan == Some(proto::Plan::ZedPro);
902        let subscription_text = Label::new(if is_pro {
903            "You have full access to Zed's hosted LLMs, which include models from Anthropic, OpenAI, and Google. They come with faster speeds and higher limits through Zed Pro."
904        } else {
905            "You have basic access to models from Anthropic through the Zed AI Free plan."
906        });
907        let manage_subscription_button = if is_pro {
908            Some(
909                h_flex().child(
910                    Button::new("manage_settings", "Manage Subscription")
911                        .style(ButtonStyle::Tinted(TintColor::Accent))
912                        .on_click(cx.listener(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))),
913                ),
914            )
915        } else if cx.has_flag::<ZedPro>() {
916            Some(
917                h_flex()
918                    .gap_2()
919                    .child(
920                        Button::new("learn_more", "Learn more")
921                            .style(ButtonStyle::Subtle)
922                            .on_click(cx.listener(|_, _, cx| cx.open_url(ZED_AI_URL))),
923                    )
924                    .child(
925                        Button::new("upgrade", "Upgrade")
926                            .style(ButtonStyle::Subtle)
927                            .color(Color::Accent)
928                            .on_click(
929                                cx.listener(|_, _, cx| cx.open_url(&zed_urls::account_url(cx))),
930                            ),
931                    ),
932            )
933        } else {
934            None
935        };
936
937        if is_connected {
938            v_flex()
939                .gap_3()
940                .max_w_4_5()
941                .children(self.render_accept_terms(cx))
942                .when(has_accepted_terms, |this| {
943                    this.child(subscription_text)
944                        .children(manage_subscription_button)
945                })
946        } else {
947            v_flex()
948                .gap_2()
949                .child(Label::new("Use Zed AI to access hosted language models."))
950                .child(
951                    Button::new("sign_in", "Sign In")
952                        .icon_color(Color::Muted)
953                        .icon(IconName::Github)
954                        .icon_position(IconPosition::Start)
955                        .on_click(cx.listener(move |this, _, cx| this.authenticate(cx))),
956                )
957        }
958    }
959}