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 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
11 TOOL_USE_LIMIT_REACHED_HEADER_NAME, ZED_VERSION_HEADER_NAME,
12};
13use futures::{
14 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
15};
16use google_ai::GoogleModelMode;
17use gpui::{
18 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
19};
20use http_client::http::{HeaderMap, HeaderValue};
21use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
22use language_model::{
23 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
24 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
25 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
26 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
27 LanguageModelToolSchemaFormat, LlmApiToken, ModelRequestLimitReachedError,
28 PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
29};
30use release_channel::AppVersion;
31use schemars::JsonSchema;
32use serde::{Deserialize, Serialize, de::DeserializeOwned};
33use settings::SettingsStore;
34use smol::io::{AsyncReadExt, BufReader};
35use std::pin::Pin;
36use std::str::FromStr as _;
37use std::sync::Arc;
38use std::time::Duration;
39use thiserror::Error;
40use ui::{TintColor, prelude::*};
41use util::{ResultExt as _, maybe};
42
43use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
44use crate::provider::google::{GoogleEventMapper, into_google};
45use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
46
47const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
48const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
49
50#[derive(Default, Clone, Debug, PartialEq)]
51pub struct ZedDotDevSettings {
52 pub available_models: Vec<AvailableModel>,
53}
54
55#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
56#[serde(rename_all = "lowercase")]
57pub enum AvailableProvider {
58 Anthropic,
59 OpenAi,
60 Google,
61}
62
63#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
64pub struct AvailableModel {
65 /// The provider of the language model.
66 pub provider: AvailableProvider,
67 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
68 pub name: String,
69 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
70 pub display_name: Option<String>,
71 /// The size of the context window, indicating the maximum number of tokens the model can process.
72 pub max_tokens: usize,
73 /// The maximum number of output tokens allowed by the model.
74 pub max_output_tokens: Option<u64>,
75 /// The maximum number of completion tokens allowed by the model (o1-* only)
76 pub max_completion_tokens: Option<u64>,
77 /// Override this model with a different Anthropic model for tool calls.
78 pub tool_override: Option<String>,
79 /// Indicates whether this custom model supports caching.
80 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
81 /// The default temperature to use for this model.
82 pub default_temperature: Option<f32>,
83 /// Any extra beta headers to provide when using the model.
84 #[serde(default)]
85 pub extra_beta_headers: Vec<String>,
86 /// The model's mode (e.g. thinking)
87 pub mode: Option<ModelMode>,
88}
89
90#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
91#[serde(tag = "type", rename_all = "lowercase")]
92pub enum ModelMode {
93 #[default]
94 Default,
95 Thinking {
96 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
97 budget_tokens: Option<u32>,
98 },
99}
100
101impl From<ModelMode> for AnthropicModelMode {
102 fn from(value: ModelMode) -> Self {
103 match value {
104 ModelMode::Default => AnthropicModelMode::Default,
105 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
106 }
107 }
108}
109
110pub struct CloudLanguageModelProvider {
111 client: Arc<Client>,
112 state: gpui::Entity<State>,
113 _maintain_client_status: Task<()>,
114}
115
116pub struct State {
117 client: Arc<Client>,
118 llm_api_token: LlmApiToken,
119 user_store: Entity<UserStore>,
120 status: client::Status,
121 models: Vec<Arc<cloud_llm_client::LanguageModel>>,
122 default_model: Option<Arc<cloud_llm_client::LanguageModel>>,
123 default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,
124 recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,
125 _fetch_models_task: Task<()>,
126 _settings_subscription: Subscription,
127 _llm_token_subscription: Subscription,
128}
129
130impl State {
131 fn new(
132 client: Arc<Client>,
133 user_store: Entity<UserStore>,
134 status: client::Status,
135 cx: &mut Context<Self>,
136 ) -> Self {
137 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
138 let mut current_user = user_store.read(cx).watch_current_user();
139 Self {
140 client: client.clone(),
141 llm_api_token: LlmApiToken::default(),
142 user_store,
143 status,
144 models: Vec::new(),
145 default_model: None,
146 default_fast_model: None,
147 recommended_models: Vec::new(),
148 _fetch_models_task: cx.spawn(async move |this, cx| {
149 maybe!(async move {
150 let (client, llm_api_token) = this
151 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
152
153 while current_user.borrow().is_none() {
154 current_user.next().await;
155 }
156
157 let response =
158 Self::fetch_models(client.clone(), llm_api_token.clone()).await?;
159 this.update(cx, |this, cx| this.update_models(response, cx))?;
160 anyhow::Ok(())
161 })
162 .await
163 .context("failed to fetch Zed models")
164 .log_err();
165 }),
166 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
167 cx.notify();
168 }),
169 _llm_token_subscription: cx.subscribe(
170 &refresh_llm_token_listener,
171 move |this, _listener, _event, cx| {
172 let client = this.client.clone();
173 let llm_api_token = this.llm_api_token.clone();
174 cx.spawn(async move |this, cx| {
175 llm_api_token.refresh(&client).await?;
176 let response = Self::fetch_models(client, llm_api_token).await?;
177 this.update(cx, |this, cx| {
178 this.update_models(response, cx);
179 })
180 })
181 .detach_and_log_err(cx);
182 },
183 ),
184 }
185 }
186
187 fn is_signed_out(&self, cx: &App) -> bool {
188 self.user_store.read(cx).current_user().is_none()
189 }
190
191 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
192 let client = self.client.clone();
193 cx.spawn(async move |state, cx| {
194 client.sign_in_with_optional_connect(true, cx).await?;
195 state.update(cx, |_, cx| cx.notify())
196 })
197 }
198 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
199 let mut models = Vec::new();
200
201 for model in response.models {
202 models.push(Arc::new(model.clone()));
203
204 // Right now we represent thinking variants of models as separate models on the client,
205 // so we need to insert variants for any model that supports thinking.
206 if model.supports_thinking {
207 models.push(Arc::new(cloud_llm_client::LanguageModel {
208 id: cloud_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
209 display_name: format!("{} Thinking", model.display_name),
210 ..model
211 }));
212 }
213 }
214
215 self.default_model = models
216 .iter()
217 .find(|model| model.id == response.default_model)
218 .cloned();
219 self.default_fast_model = models
220 .iter()
221 .find(|model| model.id == response.default_fast_model)
222 .cloned();
223 self.recommended_models = response
224 .recommended_models
225 .iter()
226 .filter_map(|id| models.iter().find(|model| &model.id == id))
227 .cloned()
228 .collect();
229 self.models = models;
230 cx.notify();
231 }
232
233 async fn fetch_models(
234 client: Arc<Client>,
235 llm_api_token: LlmApiToken,
236 ) -> Result<ListModelsResponse> {
237 let http_client = &client.http_client();
238 let token = llm_api_token.acquire(&client).await?;
239
240 let request = http_client::Request::builder()
241 .method(Method::GET)
242 .uri(http_client.build_zed_llm_url("/models", &[])?.as_ref())
243 .header("Authorization", format!("Bearer {token}"))
244 .body(AsyncBody::empty())?;
245 let mut response = http_client
246 .send(request)
247 .await
248 .context("failed to send list models request")?;
249
250 if response.status().is_success() {
251 let mut body = String::new();
252 response.body_mut().read_to_string(&mut body).await?;
253 Ok(serde_json::from_str(&body)?)
254 } else {
255 let mut body = String::new();
256 response.body_mut().read_to_string(&mut body).await?;
257 anyhow::bail!(
258 "error listing models.\nStatus: {:?}\nBody: {body}",
259 response.status(),
260 );
261 }
262 }
263}
264
265impl CloudLanguageModelProvider {
266 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
267 let mut status_rx = client.status();
268 let status = *status_rx.borrow();
269
270 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
271
272 let state_ref = state.downgrade();
273 let maintain_client_status = cx.spawn(async move |cx| {
274 while let Some(status) = status_rx.next().await {
275 if let Some(this) = state_ref.upgrade() {
276 _ = this.update(cx, |this, cx| {
277 if this.status != status {
278 this.status = status;
279 cx.notify();
280 }
281 });
282 } else {
283 break;
284 }
285 }
286 });
287
288 Self {
289 client,
290 state,
291 _maintain_client_status: maintain_client_status,
292 }
293 }
294
295 fn create_language_model(
296 &self,
297 model: Arc<cloud_llm_client::LanguageModel>,
298 llm_api_token: LlmApiToken,
299 ) -> Arc<dyn LanguageModel> {
300 Arc::new(CloudLanguageModel {
301 id: LanguageModelId(SharedString::from(model.id.0.clone())),
302 model,
303 llm_api_token,
304 client: self.client.clone(),
305 request_limiter: RateLimiter::new(4),
306 })
307 }
308}
309
310impl LanguageModelProviderState for CloudLanguageModelProvider {
311 type ObservableEntity = State;
312
313 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
314 Some(self.state.clone())
315 }
316}
317
318impl LanguageModelProvider for CloudLanguageModelProvider {
319 fn id(&self) -> LanguageModelProviderId {
320 PROVIDER_ID
321 }
322
323 fn name(&self) -> LanguageModelProviderName {
324 PROVIDER_NAME
325 }
326
327 fn icon(&self) -> IconName {
328 IconName::AiZed
329 }
330
331 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
332 let default_model = self.state.read(cx).default_model.clone()?;
333 let llm_api_token = self.state.read(cx).llm_api_token.clone();
334 Some(self.create_language_model(default_model, llm_api_token))
335 }
336
337 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
338 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
339 let llm_api_token = self.state.read(cx).llm_api_token.clone();
340 Some(self.create_language_model(default_fast_model, llm_api_token))
341 }
342
343 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
344 let llm_api_token = self.state.read(cx).llm_api_token.clone();
345 self.state
346 .read(cx)
347 .recommended_models
348 .iter()
349 .cloned()
350 .map(|model| self.create_language_model(model, llm_api_token.clone()))
351 .collect()
352 }
353
354 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
355 let llm_api_token = self.state.read(cx).llm_api_token.clone();
356 self.state
357 .read(cx)
358 .models
359 .iter()
360 .cloned()
361 .map(|model| self.create_language_model(model, llm_api_token.clone()))
362 .collect()
363 }
364
365 fn is_authenticated(&self, cx: &App) -> bool {
366 let state = self.state.read(cx);
367 !state.is_signed_out(cx)
368 }
369
370 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
371 Task::ready(Ok(()))
372 }
373
374 fn configuration_view(
375 &self,
376 _target_agent: language_model::ConfigurationViewTargetAgent,
377 _: &mut Window,
378 cx: &mut App,
379 ) -> AnyView {
380 cx.new(|_| ConfigurationView::new(self.state.clone()))
381 .into()
382 }
383
384 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
385 Task::ready(Ok(()))
386 }
387}
388
389pub struct CloudLanguageModel {
390 id: LanguageModelId,
391 model: Arc<cloud_llm_client::LanguageModel>,
392 llm_api_token: LlmApiToken,
393 client: Arc<Client>,
394 request_limiter: RateLimiter,
395}
396
397struct PerformLlmCompletionResponse {
398 response: Response<AsyncBody>,
399 usage: Option<ModelRequestUsage>,
400 tool_use_limit_reached: bool,
401 includes_status_messages: bool,
402}
403
404impl CloudLanguageModel {
405 async fn perform_llm_completion(
406 client: Arc<Client>,
407 llm_api_token: LlmApiToken,
408 app_version: Option<SemanticVersion>,
409 body: CompletionBody,
410 ) -> Result<PerformLlmCompletionResponse> {
411 let http_client = &client.http_client();
412
413 let mut token = llm_api_token.acquire(&client).await?;
414 let mut refreshed_token = false;
415
416 loop {
417 let request_builder = http_client::Request::builder()
418 .method(Method::POST)
419 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
420 let request_builder = if let Some(app_version) = app_version {
421 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
422 } else {
423 request_builder
424 };
425
426 let request = request_builder
427 .header("Content-Type", "application/json")
428 .header("Authorization", format!("Bearer {token}"))
429 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
430 .body(serde_json::to_string(&body)?.into())?;
431 let mut response = http_client.send(request).await?;
432 let status = response.status();
433 if status.is_success() {
434 let includes_status_messages = response
435 .headers()
436 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
437 .is_some();
438
439 let tool_use_limit_reached = response
440 .headers()
441 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
442 .is_some();
443
444 let usage = if includes_status_messages {
445 None
446 } else {
447 ModelRequestUsage::from_headers(response.headers()).ok()
448 };
449
450 return Ok(PerformLlmCompletionResponse {
451 response,
452 usage,
453 includes_status_messages,
454 tool_use_limit_reached,
455 });
456 }
457
458 if !refreshed_token
459 && response
460 .headers()
461 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
462 .is_some()
463 {
464 token = llm_api_token.refresh(&client).await?;
465 refreshed_token = true;
466 continue;
467 }
468
469 if status == StatusCode::FORBIDDEN
470 && response
471 .headers()
472 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
473 .is_some()
474 {
475 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
476 .headers()
477 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
478 .and_then(|resource| resource.to_str().ok())
479 && let Some(plan) = response
480 .headers()
481 .get(CURRENT_PLAN_HEADER_NAME)
482 .and_then(|plan| plan.to_str().ok())
483 .and_then(|plan| cloud_llm_client::Plan::from_str(plan).ok())
484 {
485 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
486 }
487 } else if status == StatusCode::PAYMENT_REQUIRED {
488 return Err(anyhow!(PaymentRequiredError));
489 }
490
491 let mut body = String::new();
492 let headers = response.headers().clone();
493 response.body_mut().read_to_string(&mut body).await?;
494 return Err(anyhow!(ApiError {
495 status,
496 body,
497 headers
498 }));
499 }
500 }
501}
502
503#[derive(Debug, Error)]
504#[error("cloud language model request failed with status {status}: {body}")]
505struct ApiError {
506 status: StatusCode,
507 body: String,
508 headers: HeaderMap<HeaderValue>,
509}
510
511/// Represents error responses from Zed's cloud API.
512///
513/// Example JSON for an upstream HTTP error:
514/// ```json
515/// {
516/// "code": "upstream_http_error",
517/// "message": "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout",
518/// "upstream_status": 503
519/// }
520/// ```
521#[derive(Debug, serde::Deserialize)]
522struct CloudApiError {
523 code: String,
524 message: String,
525 #[serde(default)]
526 #[serde(deserialize_with = "deserialize_optional_status_code")]
527 upstream_status: Option<StatusCode>,
528 #[serde(default)]
529 retry_after: Option<f64>,
530}
531
532fn deserialize_optional_status_code<'de, D>(deserializer: D) -> Result<Option<StatusCode>, D::Error>
533where
534 D: serde::Deserializer<'de>,
535{
536 let opt: Option<u16> = Option::deserialize(deserializer)?;
537 Ok(opt.and_then(|code| StatusCode::from_u16(code).ok()))
538}
539
540impl From<ApiError> for LanguageModelCompletionError {
541 fn from(error: ApiError) -> Self {
542 if let Ok(cloud_error) = serde_json::from_str::<CloudApiError>(&error.body)
543 && cloud_error.code.starts_with("upstream_http_")
544 {
545 let status = if let Some(status) = cloud_error.upstream_status {
546 status
547 } else if cloud_error.code.ends_with("_error") {
548 error.status
549 } else {
550 // If there's a status code in the code string (e.g. "upstream_http_429")
551 // then use that; otherwise, see if the JSON contains a status code.
552 cloud_error
553 .code
554 .strip_prefix("upstream_http_")
555 .and_then(|code_str| code_str.parse::<u16>().ok())
556 .and_then(|code| StatusCode::from_u16(code).ok())
557 .unwrap_or(error.status)
558 };
559
560 return LanguageModelCompletionError::UpstreamProviderError {
561 message: cloud_error.message,
562 status,
563 retry_after: cloud_error.retry_after.map(Duration::from_secs_f64),
564 };
565 }
566
567 let retry_after = None;
568 LanguageModelCompletionError::from_http_status(
569 PROVIDER_NAME,
570 error.status,
571 error.body,
572 retry_after,
573 )
574 }
575}
576
577impl LanguageModel for CloudLanguageModel {
578 fn id(&self) -> LanguageModelId {
579 self.id.clone()
580 }
581
582 fn name(&self) -> LanguageModelName {
583 LanguageModelName::from(self.model.display_name.clone())
584 }
585
586 fn provider_id(&self) -> LanguageModelProviderId {
587 PROVIDER_ID
588 }
589
590 fn provider_name(&self) -> LanguageModelProviderName {
591 PROVIDER_NAME
592 }
593
594 fn upstream_provider_id(&self) -> LanguageModelProviderId {
595 use cloud_llm_client::LanguageModelProvider::*;
596 match self.model.provider {
597 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
598 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
599 Google => language_model::GOOGLE_PROVIDER_ID,
600 }
601 }
602
603 fn upstream_provider_name(&self) -> LanguageModelProviderName {
604 use cloud_llm_client::LanguageModelProvider::*;
605 match self.model.provider {
606 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
607 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
608 Google => language_model::GOOGLE_PROVIDER_NAME,
609 }
610 }
611
612 fn supports_tools(&self) -> bool {
613 self.model.supports_tools
614 }
615
616 fn supports_images(&self) -> bool {
617 self.model.supports_images
618 }
619
620 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
621 match choice {
622 LanguageModelToolChoice::Auto
623 | LanguageModelToolChoice::Any
624 | LanguageModelToolChoice::None => true,
625 }
626 }
627
628 fn supports_burn_mode(&self) -> bool {
629 self.model.supports_max_mode
630 }
631
632 fn telemetry_id(&self) -> String {
633 format!("zed.dev/{}", self.model.id)
634 }
635
636 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
637 match self.model.provider {
638 cloud_llm_client::LanguageModelProvider::Anthropic
639 | cloud_llm_client::LanguageModelProvider::OpenAi => {
640 LanguageModelToolSchemaFormat::JsonSchema
641 }
642 cloud_llm_client::LanguageModelProvider::Google => {
643 LanguageModelToolSchemaFormat::JsonSchemaSubset
644 }
645 }
646 }
647
648 fn max_token_count(&self) -> u64 {
649 self.model.max_token_count as u64
650 }
651
652 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
653 self.model
654 .max_token_count_in_max_mode
655 .filter(|_| self.model.supports_max_mode)
656 .map(|max_token_count| max_token_count as u64)
657 }
658
659 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
660 match &self.model.provider {
661 cloud_llm_client::LanguageModelProvider::Anthropic => {
662 Some(LanguageModelCacheConfiguration {
663 min_total_token: 2_048,
664 should_speculate: true,
665 max_cache_anchors: 4,
666 })
667 }
668 cloud_llm_client::LanguageModelProvider::OpenAi
669 | cloud_llm_client::LanguageModelProvider::Google => None,
670 }
671 }
672
673 fn count_tokens(
674 &self,
675 request: LanguageModelRequest,
676 cx: &App,
677 ) -> BoxFuture<'static, Result<u64>> {
678 match self.model.provider {
679 cloud_llm_client::LanguageModelProvider::Anthropic => {
680 count_anthropic_tokens(request, cx)
681 }
682 cloud_llm_client::LanguageModelProvider::OpenAi => {
683 let model = match open_ai::Model::from_id(&self.model.id.0) {
684 Ok(model) => model,
685 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
686 };
687 count_open_ai_tokens(request, model, cx)
688 }
689 cloud_llm_client::LanguageModelProvider::Google => {
690 let client = self.client.clone();
691 let llm_api_token = self.llm_api_token.clone();
692 let model_id = self.model.id.to_string();
693 let generate_content_request =
694 into_google(request, model_id.clone(), GoogleModelMode::Default);
695 async move {
696 let http_client = &client.http_client();
697 let token = llm_api_token.acquire(&client).await?;
698
699 let request_body = CountTokensBody {
700 provider: cloud_llm_client::LanguageModelProvider::Google,
701 model: model_id,
702 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
703 generate_content_request,
704 })?,
705 };
706 let request = http_client::Request::builder()
707 .method(Method::POST)
708 .uri(
709 http_client
710 .build_zed_llm_url("/count_tokens", &[])?
711 .as_ref(),
712 )
713 .header("Content-Type", "application/json")
714 .header("Authorization", format!("Bearer {token}"))
715 .body(serde_json::to_string(&request_body)?.into())?;
716 let mut response = http_client.send(request).await?;
717 let status = response.status();
718 let headers = response.headers().clone();
719 let mut response_body = String::new();
720 response
721 .body_mut()
722 .read_to_string(&mut response_body)
723 .await?;
724
725 if status.is_success() {
726 let response_body: CountTokensResponse =
727 serde_json::from_str(&response_body)?;
728
729 Ok(response_body.tokens as u64)
730 } else {
731 Err(anyhow!(ApiError {
732 status,
733 body: response_body,
734 headers
735 }))
736 }
737 }
738 .boxed()
739 }
740 }
741 }
742
743 fn stream_completion(
744 &self,
745 request: LanguageModelRequest,
746 cx: &AsyncApp,
747 ) -> BoxFuture<
748 'static,
749 Result<
750 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
751 LanguageModelCompletionError,
752 >,
753 > {
754 let thread_id = request.thread_id.clone();
755 let prompt_id = request.prompt_id.clone();
756 let intent = request.intent;
757 let mode = request.mode;
758 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
759 let thinking_allowed = request.thinking_allowed;
760 match self.model.provider {
761 cloud_llm_client::LanguageModelProvider::Anthropic => {
762 let request = into_anthropic(
763 request,
764 self.model.id.to_string(),
765 1.0,
766 self.model.max_output_tokens as u64,
767 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
768 AnthropicModelMode::Thinking {
769 budget_tokens: Some(4_096),
770 }
771 } else {
772 AnthropicModelMode::Default
773 },
774 );
775 let client = self.client.clone();
776 let llm_api_token = self.llm_api_token.clone();
777 let future = self.request_limiter.stream(async move {
778 let PerformLlmCompletionResponse {
779 response,
780 usage,
781 includes_status_messages,
782 tool_use_limit_reached,
783 } = Self::perform_llm_completion(
784 client.clone(),
785 llm_api_token,
786 app_version,
787 CompletionBody {
788 thread_id,
789 prompt_id,
790 intent,
791 mode,
792 provider: cloud_llm_client::LanguageModelProvider::Anthropic,
793 model: request.model.clone(),
794 provider_request: serde_json::to_value(&request)
795 .map_err(|e| anyhow!(e))?,
796 },
797 )
798 .await
799 .map_err(|err| match err.downcast::<ApiError>() {
800 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
801 Err(err) => anyhow!(err),
802 })?;
803
804 let mut mapper = AnthropicEventMapper::new();
805 Ok(map_cloud_completion_events(
806 Box::pin(
807 response_lines(response, includes_status_messages)
808 .chain(usage_updated_event(usage))
809 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
810 ),
811 move |event| mapper.map_event(event),
812 ))
813 });
814 async move { Ok(future.await?.boxed()) }.boxed()
815 }
816 cloud_llm_client::LanguageModelProvider::OpenAi => {
817 let client = self.client.clone();
818 let model = match open_ai::Model::from_id(&self.model.id.0) {
819 Ok(model) => model,
820 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
821 };
822 let request = into_open_ai(
823 request,
824 model.id(),
825 model.supports_parallel_tool_calls(),
826 model.supports_prompt_cache_key(),
827 None,
828 None,
829 );
830 let llm_api_token = self.llm_api_token.clone();
831 let future = self.request_limiter.stream(async move {
832 let PerformLlmCompletionResponse {
833 response,
834 usage,
835 includes_status_messages,
836 tool_use_limit_reached,
837 } = Self::perform_llm_completion(
838 client.clone(),
839 llm_api_token,
840 app_version,
841 CompletionBody {
842 thread_id,
843 prompt_id,
844 intent,
845 mode,
846 provider: cloud_llm_client::LanguageModelProvider::OpenAi,
847 model: request.model.clone(),
848 provider_request: serde_json::to_value(&request)
849 .map_err(|e| anyhow!(e))?,
850 },
851 )
852 .await?;
853
854 let mut mapper = OpenAiEventMapper::new();
855 Ok(map_cloud_completion_events(
856 Box::pin(
857 response_lines(response, includes_status_messages)
858 .chain(usage_updated_event(usage))
859 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
860 ),
861 move |event| mapper.map_event(event),
862 ))
863 });
864 async move { Ok(future.await?.boxed()) }.boxed()
865 }
866 cloud_llm_client::LanguageModelProvider::Google => {
867 let client = self.client.clone();
868 let request =
869 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
870 let llm_api_token = self.llm_api_token.clone();
871 let future = self.request_limiter.stream(async move {
872 let PerformLlmCompletionResponse {
873 response,
874 usage,
875 includes_status_messages,
876 tool_use_limit_reached,
877 } = Self::perform_llm_completion(
878 client.clone(),
879 llm_api_token,
880 app_version,
881 CompletionBody {
882 thread_id,
883 prompt_id,
884 intent,
885 mode,
886 provider: cloud_llm_client::LanguageModelProvider::Google,
887 model: request.model.model_id.clone(),
888 provider_request: serde_json::to_value(&request)
889 .map_err(|e| anyhow!(e))?,
890 },
891 )
892 .await?;
893
894 let mut mapper = GoogleEventMapper::new();
895 Ok(map_cloud_completion_events(
896 Box::pin(
897 response_lines(response, includes_status_messages)
898 .chain(usage_updated_event(usage))
899 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
900 ),
901 move |event| mapper.map_event(event),
902 ))
903 });
904 async move { Ok(future.await?.boxed()) }.boxed()
905 }
906 }
907 }
908}
909
910fn map_cloud_completion_events<T, F>(
911 stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>>> + Send>>,
912 mut map_callback: F,
913) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
914where
915 T: DeserializeOwned + 'static,
916 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
917 + Send
918 + 'static,
919{
920 stream
921 .flat_map(move |event| {
922 futures::stream::iter(match event {
923 Err(error) => {
924 vec![Err(LanguageModelCompletionError::from(error))]
925 }
926 Ok(CompletionEvent::Status(event)) => {
927 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
928 }
929 Ok(CompletionEvent::Event(event)) => map_callback(event),
930 })
931 })
932 .boxed()
933}
934
935fn usage_updated_event<T>(
936 usage: Option<ModelRequestUsage>,
937) -> impl Stream<Item = Result<CompletionEvent<T>>> {
938 futures::stream::iter(usage.map(|usage| {
939 Ok(CompletionEvent::Status(
940 CompletionRequestStatus::UsageUpdated {
941 amount: usage.amount as usize,
942 limit: usage.limit,
943 },
944 ))
945 }))
946}
947
948fn tool_use_limit_reached_event<T>(
949 tool_use_limit_reached: bool,
950) -> impl Stream<Item = Result<CompletionEvent<T>>> {
951 futures::stream::iter(tool_use_limit_reached.then(|| {
952 Ok(CompletionEvent::Status(
953 CompletionRequestStatus::ToolUseLimitReached,
954 ))
955 }))
956}
957
958fn response_lines<T: DeserializeOwned>(
959 response: Response<AsyncBody>,
960 includes_status_messages: bool,
961) -> impl Stream<Item = Result<CompletionEvent<T>>> {
962 futures::stream::try_unfold(
963 (String::new(), BufReader::new(response.into_body())),
964 move |(mut line, mut body)| async move {
965 match body.read_line(&mut line).await {
966 Ok(0) => Ok(None),
967 Ok(_) => {
968 let event = if includes_status_messages {
969 serde_json::from_str::<CompletionEvent<T>>(&line)?
970 } else {
971 CompletionEvent::Event(serde_json::from_str::<T>(&line)?)
972 };
973
974 line.clear();
975 Ok(Some((event, (line, body))))
976 }
977 Err(e) => Err(e.into()),
978 }
979 },
980 )
981}
982
983#[derive(IntoElement, RegisterComponent)]
984struct ZedAiConfiguration {
985 is_connected: bool,
986 plan: Option<Plan>,
987 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
988 eligible_for_trial: bool,
989 account_too_young: bool,
990 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
991}
992
993impl RenderOnce for ZedAiConfiguration {
994 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
995 let young_account_banner = YoungAccountBanner;
996
997 let is_pro = self.plan == Some(Plan::ZedPro);
998 let subscription_text = match (self.plan, self.subscription_period) {
999 (Some(Plan::ZedPro), Some(_)) => {
1000 "You have access to Zed's hosted models through your Pro subscription."
1001 }
1002 (Some(Plan::ZedProTrial), Some(_)) => {
1003 "You have access to Zed's hosted models through your Pro trial."
1004 }
1005 (Some(Plan::ZedFree), Some(_)) => {
1006 "You have basic access to Zed's hosted models through the Free plan."
1007 }
1008 _ => {
1009 if self.eligible_for_trial {
1010 "Subscribe for access to Zed's hosted models. Start with a 14 day free trial."
1011 } else {
1012 "Subscribe for access to Zed's hosted models."
1013 }
1014 }
1015 };
1016
1017 let manage_subscription_buttons = if is_pro {
1018 Button::new("manage_settings", "Manage Subscription")
1019 .full_width()
1020 .style(ButtonStyle::Tinted(TintColor::Accent))
1021 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))
1022 .into_any_element()
1023 } else if self.plan.is_none() || self.eligible_for_trial {
1024 Button::new("start_trial", "Start 14-day Free Pro Trial")
1025 .full_width()
1026 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1027 .on_click(|_, _, cx| cx.open_url(&zed_urls::start_trial_url(cx)))
1028 .into_any_element()
1029 } else {
1030 Button::new("upgrade", "Upgrade to Pro")
1031 .full_width()
1032 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1033 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx)))
1034 .into_any_element()
1035 };
1036
1037 if !self.is_connected {
1038 return v_flex()
1039 .gap_2()
1040 .child(Label::new("Sign in to have access to Zed's complete agentic experience with hosted models."))
1041 .child(
1042 Button::new("sign_in", "Sign In to use Zed AI")
1043 .icon_color(Color::Muted)
1044 .icon(IconName::Github)
1045 .icon_size(IconSize::Small)
1046 .icon_position(IconPosition::Start)
1047 .full_width()
1048 .on_click({
1049 let callback = self.sign_in_callback.clone();
1050 move |_, window, cx| (callback)(window, cx)
1051 }),
1052 );
1053 }
1054
1055 v_flex().gap_2().w_full().map(|this| {
1056 if self.account_too_young {
1057 this.child(young_account_banner).child(
1058 Button::new("upgrade", "Upgrade to Pro")
1059 .style(ui::ButtonStyle::Tinted(ui::TintColor::Accent))
1060 .full_width()
1061 .on_click(|_, _, cx| cx.open_url(&zed_urls::upgrade_to_zed_pro_url(cx))),
1062 )
1063 } else {
1064 this.text_sm()
1065 .child(subscription_text)
1066 .child(manage_subscription_buttons)
1067 }
1068 })
1069 }
1070}
1071
1072struct ConfigurationView {
1073 state: Entity<State>,
1074 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1075}
1076
1077impl ConfigurationView {
1078 fn new(state: Entity<State>) -> Self {
1079 let sign_in_callback = Arc::new({
1080 let state = state.clone();
1081 move |_window: &mut Window, cx: &mut App| {
1082 state.update(cx, |state, cx| {
1083 state.authenticate(cx).detach_and_log_err(cx);
1084 });
1085 }
1086 });
1087
1088 Self {
1089 state,
1090 sign_in_callback,
1091 }
1092 }
1093}
1094
1095impl Render for ConfigurationView {
1096 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1097 let state = self.state.read(cx);
1098 let user_store = state.user_store.read(cx);
1099
1100 ZedAiConfiguration {
1101 is_connected: !state.is_signed_out(cx),
1102 plan: user_store.plan(),
1103 subscription_period: user_store.subscription_period(),
1104 eligible_for_trial: user_store.trial_started_at().is_none(),
1105 account_too_young: user_store.account_too_young(),
1106 sign_in_callback: self.sign_in_callback.clone(),
1107 }
1108 }
1109}
1110
1111impl Component for ZedAiConfiguration {
1112 fn name() -> &'static str {
1113 "AI Configuration Content"
1114 }
1115
1116 fn sort_name() -> &'static str {
1117 "AI Configuration Content"
1118 }
1119
1120 fn scope() -> ComponentScope {
1121 ComponentScope::Onboarding
1122 }
1123
1124 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1125 fn configuration(
1126 is_connected: bool,
1127 plan: Option<Plan>,
1128 eligible_for_trial: bool,
1129 account_too_young: bool,
1130 ) -> AnyElement {
1131 ZedAiConfiguration {
1132 is_connected,
1133 plan,
1134 subscription_period: plan
1135 .is_some()
1136 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1137 eligible_for_trial,
1138 account_too_young,
1139 sign_in_callback: Arc::new(|_, _| {}),
1140 }
1141 .into_any_element()
1142 }
1143
1144 Some(
1145 v_flex()
1146 .p_4()
1147 .gap_4()
1148 .children(vec![
1149 single_example("Not connected", configuration(false, None, false, false)),
1150 single_example(
1151 "Accept Terms of Service",
1152 configuration(true, None, true, false),
1153 ),
1154 single_example(
1155 "No Plan - Not eligible for trial",
1156 configuration(true, None, false, false),
1157 ),
1158 single_example(
1159 "No Plan - Eligible for trial",
1160 configuration(true, None, true, false),
1161 ),
1162 single_example(
1163 "Free Plan",
1164 configuration(true, Some(Plan::ZedFree), true, false),
1165 ),
1166 single_example(
1167 "Zed Pro Trial Plan",
1168 configuration(true, Some(Plan::ZedProTrial), true, false),
1169 ),
1170 single_example(
1171 "Zed Pro Plan",
1172 configuration(true, Some(Plan::ZedPro), true, false),
1173 ),
1174 ])
1175 .into_any_element(),
1176 )
1177 }
1178}
1179
1180#[cfg(test)]
1181mod tests {
1182 use super::*;
1183 use http_client::http::{HeaderMap, StatusCode};
1184 use language_model::LanguageModelCompletionError;
1185
1186 #[test]
1187 fn test_api_error_conversion_with_upstream_http_error() {
1188 // upstream_http_error with 503 status should become ServerOverloaded
1189 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}"#;
1190
1191 let api_error = ApiError {
1192 status: StatusCode::INTERNAL_SERVER_ERROR,
1193 body: error_body.to_string(),
1194 headers: HeaderMap::new(),
1195 };
1196
1197 let completion_error: LanguageModelCompletionError = api_error.into();
1198
1199 match completion_error {
1200 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1201 assert_eq!(
1202 message,
1203 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers, reset reason: connection timeout"
1204 );
1205 }
1206 _ => panic!(
1207 "Expected UpstreamProviderError for upstream 503, got: {:?}",
1208 completion_error
1209 ),
1210 }
1211
1212 // upstream_http_error with 500 status should become ApiInternalServerError
1213 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the OpenAI API: internal server error","upstream_status":500}"#;
1214
1215 let api_error = ApiError {
1216 status: StatusCode::INTERNAL_SERVER_ERROR,
1217 body: error_body.to_string(),
1218 headers: HeaderMap::new(),
1219 };
1220
1221 let completion_error: LanguageModelCompletionError = api_error.into();
1222
1223 match completion_error {
1224 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1225 assert_eq!(
1226 message,
1227 "Received an error from the OpenAI API: internal server error"
1228 );
1229 }
1230 _ => panic!(
1231 "Expected UpstreamProviderError for upstream 500, got: {:?}",
1232 completion_error
1233 ),
1234 }
1235
1236 // upstream_http_error with 429 status should become RateLimitExceeded
1237 let error_body = r#"{"code":"upstream_http_error","message":"Received an error from the Google API: rate limit exceeded","upstream_status":429}"#;
1238
1239 let api_error = ApiError {
1240 status: StatusCode::INTERNAL_SERVER_ERROR,
1241 body: error_body.to_string(),
1242 headers: HeaderMap::new(),
1243 };
1244
1245 let completion_error: LanguageModelCompletionError = api_error.into();
1246
1247 match completion_error {
1248 LanguageModelCompletionError::UpstreamProviderError { message, .. } => {
1249 assert_eq!(
1250 message,
1251 "Received an error from the Google API: rate limit exceeded"
1252 );
1253 }
1254 _ => panic!(
1255 "Expected UpstreamProviderError for upstream 429, got: {:?}",
1256 completion_error
1257 ),
1258 }
1259
1260 // Regular 500 error without upstream_http_error should remain ApiInternalServerError for Zed
1261 let error_body = "Regular internal server error";
1262
1263 let api_error = ApiError {
1264 status: StatusCode::INTERNAL_SERVER_ERROR,
1265 body: error_body.to_string(),
1266 headers: HeaderMap::new(),
1267 };
1268
1269 let completion_error: LanguageModelCompletionError = api_error.into();
1270
1271 match completion_error {
1272 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1273 assert_eq!(provider, PROVIDER_NAME);
1274 assert_eq!(message, "Regular internal server error");
1275 }
1276 _ => panic!(
1277 "Expected ApiInternalServerError for regular 500, got: {:?}",
1278 completion_error
1279 ),
1280 }
1281
1282 // upstream_http_429 format should be converted to UpstreamProviderError
1283 let error_body = r#"{"code":"upstream_http_429","message":"Upstream Anthropic rate limit exceeded.","retry_after":30.5}"#;
1284
1285 let api_error = ApiError {
1286 status: StatusCode::INTERNAL_SERVER_ERROR,
1287 body: error_body.to_string(),
1288 headers: HeaderMap::new(),
1289 };
1290
1291 let completion_error: LanguageModelCompletionError = api_error.into();
1292
1293 match completion_error {
1294 LanguageModelCompletionError::UpstreamProviderError {
1295 message,
1296 status,
1297 retry_after,
1298 } => {
1299 assert_eq!(message, "Upstream Anthropic rate limit exceeded.");
1300 assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
1301 assert_eq!(retry_after, Some(Duration::from_secs_f64(30.5)));
1302 }
1303 _ => panic!(
1304 "Expected UpstreamProviderError for upstream_http_429, got: {:?}",
1305 completion_error
1306 ),
1307 }
1308
1309 // Invalid JSON in error body should fall back to regular error handling
1310 let error_body = "Not JSON at all";
1311
1312 let api_error = ApiError {
1313 status: StatusCode::INTERNAL_SERVER_ERROR,
1314 body: error_body.to_string(),
1315 headers: HeaderMap::new(),
1316 };
1317
1318 let completion_error: LanguageModelCompletionError = api_error.into();
1319
1320 match completion_error {
1321 LanguageModelCompletionError::ApiInternalServerError { provider, .. } => {
1322 assert_eq!(provider, PROVIDER_NAME);
1323 }
1324 _ => panic!(
1325 "Expected ApiInternalServerError for invalid JSON, got: {:?}",
1326 completion_error
1327 ),
1328 }
1329 }
1330}