1use crate::AllLanguageModelSettings;
2use crate::ui::InstructionListItem;
3use anthropic::{
4 AnthropicError, AnthropicModelMode, ContentDelta, Event, ResponseContent, ToolResultContent,
5 ToolResultPart, Usage,
6};
7use anyhow::{Context as _, Result, anyhow};
8use collections::{BTreeMap, HashMap};
9use credentials_provider::CredentialsProvider;
10use editor::{Editor, EditorElement, EditorStyle};
11use futures::Stream;
12use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};
13use gpui::{
14 AnyView, App, AsyncApp, Context, Entity, FontStyle, Subscription, Task, TextStyle, WhiteSpace,
15};
16use http_client::HttpClient;
17use language_model::{
18 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
19 LanguageModelCompletionError, LanguageModelId, LanguageModelKnownError, LanguageModelName,
20 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
21 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
22 LanguageModelToolResultContent, MessageContent, RateLimiter, Role,
23};
24use language_model::{LanguageModelCompletionEvent, LanguageModelToolUse, StopReason};
25use schemars::JsonSchema;
26use serde::{Deserialize, Serialize};
27use settings::{Settings, SettingsStore};
28use std::pin::Pin;
29use std::str::FromStr;
30use std::sync::Arc;
31use strum::IntoEnumIterator;
32use theme::ThemeSettings;
33use ui::{Icon, IconName, List, Tooltip, prelude::*};
34use util::ResultExt;
35
36const PROVIDER_ID: &str = language_model::ANTHROPIC_PROVIDER_ID;
37const PROVIDER_NAME: &str = "Anthropic";
38
39#[derive(Default, Clone, Debug, PartialEq)]
40pub struct AnthropicSettings {
41 pub api_url: String,
42 /// Extend Zed's list of Anthropic models.
43 pub available_models: Vec<AvailableModel>,
44 pub needs_setting_migration: bool,
45}
46
47#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
48pub struct AvailableModel {
49 /// The model's name in the Anthropic API. e.g. claude-3-5-sonnet-latest, claude-3-opus-20240229, etc
50 pub name: String,
51 /// The model's name in Zed's UI, such as in the model selector dropdown menu in the assistant panel.
52 pub display_name: Option<String>,
53 /// The model's context window size.
54 pub max_tokens: u64,
55 /// A model `name` to substitute when calling tools, in case the primary model doesn't support tool calling.
56 pub tool_override: Option<String>,
57 /// Configuration of Anthropic's caching API.
58 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
59 pub max_output_tokens: Option<u64>,
60 pub default_temperature: Option<f32>,
61 #[serde(default)]
62 pub extra_beta_headers: Vec<String>,
63 /// The model's mode (e.g. thinking)
64 pub mode: Option<ModelMode>,
65}
66
67#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
68#[serde(tag = "type", rename_all = "lowercase")]
69pub enum ModelMode {
70 #[default]
71 Default,
72 Thinking {
73 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
74 budget_tokens: Option<u32>,
75 },
76}
77
78impl From<ModelMode> for AnthropicModelMode {
79 fn from(value: ModelMode) -> Self {
80 match value {
81 ModelMode::Default => AnthropicModelMode::Default,
82 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
83 }
84 }
85}
86
87impl From<AnthropicModelMode> for ModelMode {
88 fn from(value: AnthropicModelMode) -> Self {
89 match value {
90 AnthropicModelMode::Default => ModelMode::Default,
91 AnthropicModelMode::Thinking { budget_tokens } => ModelMode::Thinking { budget_tokens },
92 }
93 }
94}
95
96pub struct AnthropicLanguageModelProvider {
97 http_client: Arc<dyn HttpClient>,
98 state: gpui::Entity<State>,
99}
100
101const ANTHROPIC_API_KEY_VAR: &str = "ANTHROPIC_API_KEY";
102
103pub struct State {
104 api_key: Option<String>,
105 api_key_from_env: bool,
106 _subscription: Subscription,
107}
108
109impl State {
110 fn reset_api_key(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
111 let credentials_provider = <dyn CredentialsProvider>::global(cx);
112 let api_url = AllLanguageModelSettings::get_global(cx)
113 .anthropic
114 .api_url
115 .clone();
116 cx.spawn(async move |this, cx| {
117 credentials_provider
118 .delete_credentials(&api_url, &cx)
119 .await
120 .ok();
121 this.update(cx, |this, cx| {
122 this.api_key = None;
123 this.api_key_from_env = false;
124 cx.notify();
125 })
126 })
127 }
128
129 fn set_api_key(&mut self, api_key: String, cx: &mut Context<Self>) -> Task<Result<()>> {
130 let credentials_provider = <dyn CredentialsProvider>::global(cx);
131 let api_url = AllLanguageModelSettings::get_global(cx)
132 .anthropic
133 .api_url
134 .clone();
135 cx.spawn(async move |this, cx| {
136 credentials_provider
137 .write_credentials(&api_url, "Bearer", api_key.as_bytes(), &cx)
138 .await
139 .ok();
140
141 this.update(cx, |this, cx| {
142 this.api_key = Some(api_key);
143 cx.notify();
144 })
145 })
146 }
147
148 fn is_authenticated(&self) -> bool {
149 self.api_key.is_some()
150 }
151
152 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
153 if self.is_authenticated() {
154 return Task::ready(Ok(()));
155 }
156
157 let credentials_provider = <dyn CredentialsProvider>::global(cx);
158 let api_url = AllLanguageModelSettings::get_global(cx)
159 .anthropic
160 .api_url
161 .clone();
162
163 cx.spawn(async move |this, cx| {
164 let (api_key, from_env) = if let Ok(api_key) = std::env::var(ANTHROPIC_API_KEY_VAR) {
165 (api_key, true)
166 } else {
167 let (_, api_key) = credentials_provider
168 .read_credentials(&api_url, &cx)
169 .await?
170 .ok_or(AuthenticateError::CredentialsNotFound)?;
171 (
172 String::from_utf8(api_key).context("invalid {PROVIDER_NAME} API key")?,
173 false,
174 )
175 };
176
177 this.update(cx, |this, cx| {
178 this.api_key = Some(api_key);
179 this.api_key_from_env = from_env;
180 cx.notify();
181 })?;
182
183 Ok(())
184 })
185 }
186}
187
188impl AnthropicLanguageModelProvider {
189 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
190 let state = cx.new(|cx| State {
191 api_key: None,
192 api_key_from_env: false,
193 _subscription: cx.observe_global::<SettingsStore>(|_, cx| {
194 cx.notify();
195 }),
196 });
197
198 Self { http_client, state }
199 }
200
201 fn create_language_model(&self, model: anthropic::Model) -> Arc<dyn LanguageModel> {
202 Arc::new(AnthropicModel {
203 id: LanguageModelId::from(model.id().to_string()),
204 model,
205 state: self.state.clone(),
206 http_client: self.http_client.clone(),
207 request_limiter: RateLimiter::new(4),
208 })
209 }
210}
211
212impl LanguageModelProviderState for AnthropicLanguageModelProvider {
213 type ObservableEntity = State;
214
215 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
216 Some(self.state.clone())
217 }
218}
219
220impl LanguageModelProvider for AnthropicLanguageModelProvider {
221 fn id(&self) -> LanguageModelProviderId {
222 LanguageModelProviderId(PROVIDER_ID.into())
223 }
224
225 fn name(&self) -> LanguageModelProviderName {
226 LanguageModelProviderName(PROVIDER_NAME.into())
227 }
228
229 fn icon(&self) -> IconName {
230 IconName::AiAnthropic
231 }
232
233 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
234 Some(self.create_language_model(anthropic::Model::default()))
235 }
236
237 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
238 Some(self.create_language_model(anthropic::Model::default_fast()))
239 }
240
241 fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {
242 [
243 anthropic::Model::ClaudeSonnet4,
244 anthropic::Model::ClaudeSonnet4Thinking,
245 ]
246 .into_iter()
247 .map(|model| self.create_language_model(model))
248 .collect()
249 }
250
251 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
252 let mut models = BTreeMap::default();
253
254 // Add base models from anthropic::Model::iter()
255 for model in anthropic::Model::iter() {
256 if !matches!(model, anthropic::Model::Custom { .. }) {
257 models.insert(model.id().to_string(), model);
258 }
259 }
260
261 // Override with available models from settings
262 for model in AllLanguageModelSettings::get_global(cx)
263 .anthropic
264 .available_models
265 .iter()
266 {
267 models.insert(
268 model.name.clone(),
269 anthropic::Model::Custom {
270 name: model.name.clone(),
271 display_name: model.display_name.clone(),
272 max_tokens: model.max_tokens,
273 tool_override: model.tool_override.clone(),
274 cache_configuration: model.cache_configuration.as_ref().map(|config| {
275 anthropic::AnthropicModelCacheConfiguration {
276 max_cache_anchors: config.max_cache_anchors,
277 should_speculate: config.should_speculate,
278 min_total_token: config.min_total_token,
279 }
280 }),
281 max_output_tokens: model.max_output_tokens,
282 default_temperature: model.default_temperature,
283 extra_beta_headers: model.extra_beta_headers.clone(),
284 mode: model.mode.clone().unwrap_or_default().into(),
285 },
286 );
287 }
288
289 models
290 .into_values()
291 .map(|model| self.create_language_model(model))
292 .collect()
293 }
294
295 fn is_authenticated(&self, cx: &App) -> bool {
296 self.state.read(cx).is_authenticated()
297 }
298
299 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
300 self.state.update(cx, |state, cx| state.authenticate(cx))
301 }
302
303 fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView {
304 cx.new(|cx| ConfigurationView::new(self.state.clone(), window, cx))
305 .into()
306 }
307
308 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
309 self.state.update(cx, |state, cx| state.reset_api_key(cx))
310 }
311}
312
313pub struct AnthropicModel {
314 id: LanguageModelId,
315 model: anthropic::Model,
316 state: gpui::Entity<State>,
317 http_client: Arc<dyn HttpClient>,
318 request_limiter: RateLimiter,
319}
320
321pub fn count_anthropic_tokens(
322 request: LanguageModelRequest,
323 cx: &App,
324) -> BoxFuture<'static, Result<u64>> {
325 cx.background_spawn(async move {
326 let messages = request.messages;
327 let mut tokens_from_images = 0;
328 let mut string_messages = Vec::with_capacity(messages.len());
329
330 for message in messages {
331 use language_model::MessageContent;
332
333 let mut string_contents = String::new();
334
335 for content in message.content {
336 match content {
337 MessageContent::Text(text) => {
338 string_contents.push_str(&text);
339 }
340 MessageContent::Thinking { .. } => {
341 // Thinking blocks are not included in the input token count.
342 }
343 MessageContent::RedactedThinking(_) => {
344 // Thinking blocks are not included in the input token count.
345 }
346 MessageContent::Image(image) => {
347 tokens_from_images += image.estimate_tokens();
348 }
349 MessageContent::ToolUse(_tool_use) => {
350 // TODO: Estimate token usage from tool uses.
351 }
352 MessageContent::ToolResult(tool_result) => match &tool_result.content {
353 LanguageModelToolResultContent::Text(text) => {
354 string_contents.push_str(text);
355 }
356 LanguageModelToolResultContent::Image(image) => {
357 tokens_from_images += image.estimate_tokens();
358 }
359 },
360 }
361 }
362
363 if !string_contents.is_empty() {
364 string_messages.push(tiktoken_rs::ChatCompletionRequestMessage {
365 role: match message.role {
366 Role::User => "user".into(),
367 Role::Assistant => "assistant".into(),
368 Role::System => "system".into(),
369 },
370 content: Some(string_contents),
371 name: None,
372 function_call: None,
373 });
374 }
375 }
376
377 // Tiktoken doesn't yet support these models, so we manually use the
378 // same tokenizer as GPT-4.
379 tiktoken_rs::num_tokens_from_messages("gpt-4", &string_messages)
380 .map(|tokens| (tokens + tokens_from_images) as u64)
381 })
382 .boxed()
383}
384
385impl AnthropicModel {
386 fn stream_completion(
387 &self,
388 request: anthropic::Request,
389 cx: &AsyncApp,
390 ) -> BoxFuture<
391 'static,
392 Result<
393 BoxStream<'static, Result<anthropic::Event, AnthropicError>>,
394 LanguageModelCompletionError,
395 >,
396 > {
397 let http_client = self.http_client.clone();
398
399 let Ok((api_key, api_url)) = cx.read_entity(&self.state, |state, cx| {
400 let settings = &AllLanguageModelSettings::get_global(cx).anthropic;
401 (state.api_key.clone(), settings.api_url.clone())
402 }) else {
403 return futures::future::ready(Err(anyhow!("App state dropped").into())).boxed();
404 };
405
406 async move {
407 let api_key = api_key.context("Missing Anthropic API Key")?;
408 let request =
409 anthropic::stream_completion(http_client.as_ref(), &api_url, &api_key, request);
410 request.await.map_err(|err| match err {
411 AnthropicError::RateLimit(duration) => {
412 LanguageModelCompletionError::RateLimit(duration)
413 }
414 err @ (AnthropicError::ApiError(..) | AnthropicError::Other(..)) => {
415 LanguageModelCompletionError::Other(anthropic_err_to_anyhow(err))
416 }
417 })
418 }
419 .boxed()
420 }
421}
422
423impl LanguageModel for AnthropicModel {
424 fn id(&self) -> LanguageModelId {
425 self.id.clone()
426 }
427
428 fn name(&self) -> LanguageModelName {
429 LanguageModelName::from(self.model.display_name().to_string())
430 }
431
432 fn provider_id(&self) -> LanguageModelProviderId {
433 LanguageModelProviderId(PROVIDER_ID.into())
434 }
435
436 fn provider_name(&self) -> LanguageModelProviderName {
437 LanguageModelProviderName(PROVIDER_NAME.into())
438 }
439
440 fn supports_tools(&self) -> bool {
441 true
442 }
443
444 fn supports_images(&self) -> bool {
445 true
446 }
447
448 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
449 match choice {
450 LanguageModelToolChoice::Auto
451 | LanguageModelToolChoice::Any
452 | LanguageModelToolChoice::None => true,
453 }
454 }
455
456 fn telemetry_id(&self) -> String {
457 format!("anthropic/{}", self.model.id())
458 }
459
460 fn api_key(&self, cx: &App) -> Option<String> {
461 self.state.read(cx).api_key.clone()
462 }
463
464 fn max_token_count(&self) -> u64 {
465 self.model.max_token_count()
466 }
467
468 fn max_output_tokens(&self) -> Option<u64> {
469 Some(self.model.max_output_tokens())
470 }
471
472 fn count_tokens(
473 &self,
474 request: LanguageModelRequest,
475 cx: &App,
476 ) -> BoxFuture<'static, Result<u64>> {
477 count_anthropic_tokens(request, cx)
478 }
479
480 fn stream_completion(
481 &self,
482 request: LanguageModelRequest,
483 cx: &AsyncApp,
484 ) -> BoxFuture<
485 'static,
486 Result<
487 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
488 LanguageModelCompletionError,
489 >,
490 > {
491 let request = into_anthropic(
492 request,
493 self.model.request_id().into(),
494 self.model.default_temperature(),
495 self.model.max_output_tokens(),
496 self.model.mode(),
497 );
498 let request = self.stream_completion(request, cx);
499 let future = self.request_limiter.stream(async move {
500 let response = request.await?;
501 Ok(AnthropicEventMapper::new().map_stream(response))
502 });
503 async move { Ok(future.await?.boxed()) }.boxed()
504 }
505
506 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
507 self.model
508 .cache_configuration()
509 .map(|config| LanguageModelCacheConfiguration {
510 max_cache_anchors: config.max_cache_anchors,
511 should_speculate: config.should_speculate,
512 min_total_token: config.min_total_token,
513 })
514 }
515}
516
517pub fn into_anthropic(
518 request: LanguageModelRequest,
519 model: String,
520 default_temperature: f32,
521 max_output_tokens: u64,
522 mode: AnthropicModelMode,
523) -> anthropic::Request {
524 let mut new_messages: Vec<anthropic::Message> = Vec::new();
525 let mut system_message = String::new();
526
527 for message in request.messages {
528 if message.contents_empty() {
529 continue;
530 }
531
532 match message.role {
533 Role::User | Role::Assistant => {
534 let mut anthropic_message_content: Vec<anthropic::RequestContent> = message
535 .content
536 .into_iter()
537 .filter_map(|content| match content {
538 MessageContent::Text(text) => {
539 if !text.is_empty() {
540 Some(anthropic::RequestContent::Text {
541 text,
542 cache_control: None,
543 })
544 } else {
545 None
546 }
547 }
548 MessageContent::Thinking {
549 text: thinking,
550 signature,
551 } => {
552 if !thinking.is_empty() {
553 Some(anthropic::RequestContent::Thinking {
554 thinking,
555 signature: signature.unwrap_or_default(),
556 cache_control: None,
557 })
558 } else {
559 None
560 }
561 }
562 MessageContent::RedactedThinking(data) => {
563 if !data.is_empty() {
564 Some(anthropic::RequestContent::RedactedThinking { data })
565 } else {
566 None
567 }
568 }
569 MessageContent::Image(image) => Some(anthropic::RequestContent::Image {
570 source: anthropic::ImageSource {
571 source_type: "base64".to_string(),
572 media_type: "image/png".to_string(),
573 data: image.source.to_string(),
574 },
575 cache_control: None,
576 }),
577 MessageContent::ToolUse(tool_use) => {
578 Some(anthropic::RequestContent::ToolUse {
579 id: tool_use.id.to_string(),
580 name: tool_use.name.to_string(),
581 input: tool_use.input,
582 cache_control: None,
583 })
584 }
585 MessageContent::ToolResult(tool_result) => {
586 Some(anthropic::RequestContent::ToolResult {
587 tool_use_id: tool_result.tool_use_id.to_string(),
588 is_error: tool_result.is_error,
589 content: match tool_result.content {
590 LanguageModelToolResultContent::Text(text) => {
591 ToolResultContent::Plain(text.to_string())
592 }
593 LanguageModelToolResultContent::Image(image) => {
594 ToolResultContent::Multipart(vec![ToolResultPart::Image {
595 source: anthropic::ImageSource {
596 source_type: "base64".to_string(),
597 media_type: "image/png".to_string(),
598 data: image.source.to_string(),
599 },
600 }])
601 }
602 },
603 cache_control: None,
604 })
605 }
606 })
607 .collect();
608 let anthropic_role = match message.role {
609 Role::User => anthropic::Role::User,
610 Role::Assistant => anthropic::Role::Assistant,
611 Role::System => unreachable!("System role should never occur here"),
612 };
613 if let Some(last_message) = new_messages.last_mut() {
614 if last_message.role == anthropic_role {
615 last_message.content.extend(anthropic_message_content);
616 continue;
617 }
618 }
619
620 // Mark the last segment of the message as cached
621 if message.cache {
622 let cache_control_value = Some(anthropic::CacheControl {
623 cache_type: anthropic::CacheControlType::Ephemeral,
624 });
625 for message_content in anthropic_message_content.iter_mut().rev() {
626 match message_content {
627 anthropic::RequestContent::RedactedThinking { .. } => {
628 // Caching is not possible, fallback to next message
629 }
630 anthropic::RequestContent::Text { cache_control, .. }
631 | anthropic::RequestContent::Thinking { cache_control, .. }
632 | anthropic::RequestContent::Image { cache_control, .. }
633 | anthropic::RequestContent::ToolUse { cache_control, .. }
634 | anthropic::RequestContent::ToolResult { cache_control, .. } => {
635 *cache_control = cache_control_value;
636 break;
637 }
638 }
639 }
640 }
641
642 new_messages.push(anthropic::Message {
643 role: anthropic_role,
644 content: anthropic_message_content,
645 });
646 }
647 Role::System => {
648 if !system_message.is_empty() {
649 system_message.push_str("\n\n");
650 }
651 system_message.push_str(&message.string_contents());
652 }
653 }
654 }
655
656 anthropic::Request {
657 model,
658 messages: new_messages,
659 max_tokens: max_output_tokens,
660 system: if system_message.is_empty() {
661 None
662 } else {
663 Some(anthropic::StringOrContents::String(system_message))
664 },
665 thinking: if let AnthropicModelMode::Thinking { budget_tokens } = mode {
666 Some(anthropic::Thinking::Enabled { budget_tokens })
667 } else {
668 None
669 },
670 tools: request
671 .tools
672 .into_iter()
673 .map(|tool| anthropic::Tool {
674 name: tool.name,
675 description: tool.description,
676 input_schema: tool.input_schema,
677 })
678 .collect(),
679 tool_choice: request.tool_choice.map(|choice| match choice {
680 LanguageModelToolChoice::Auto => anthropic::ToolChoice::Auto,
681 LanguageModelToolChoice::Any => anthropic::ToolChoice::Any,
682 LanguageModelToolChoice::None => anthropic::ToolChoice::None,
683 }),
684 metadata: None,
685 stop_sequences: Vec::new(),
686 temperature: request.temperature.or(Some(default_temperature)),
687 top_k: None,
688 top_p: None,
689 }
690}
691
692pub struct AnthropicEventMapper {
693 tool_uses_by_index: HashMap<usize, RawToolUse>,
694 usage: Usage,
695 stop_reason: StopReason,
696}
697
698impl AnthropicEventMapper {
699 pub fn new() -> Self {
700 Self {
701 tool_uses_by_index: HashMap::default(),
702 usage: Usage::default(),
703 stop_reason: StopReason::EndTurn,
704 }
705 }
706
707 pub fn map_stream(
708 mut self,
709 events: Pin<Box<dyn Send + Stream<Item = Result<Event, AnthropicError>>>>,
710 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
711 {
712 events.flat_map(move |event| {
713 futures::stream::iter(match event {
714 Ok(event) => self.map_event(event),
715 Err(error) => vec![Err(LanguageModelCompletionError::Other(anyhow!(error)))],
716 })
717 })
718 }
719
720 pub fn map_event(
721 &mut self,
722 event: Event,
723 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
724 match event {
725 Event::ContentBlockStart {
726 index,
727 content_block,
728 } => match content_block {
729 ResponseContent::Text { text } => {
730 vec![Ok(LanguageModelCompletionEvent::Text(text))]
731 }
732 ResponseContent::Thinking { thinking } => {
733 vec![Ok(LanguageModelCompletionEvent::Thinking {
734 text: thinking,
735 signature: None,
736 })]
737 }
738 ResponseContent::RedactedThinking { data } => {
739 vec![Ok(LanguageModelCompletionEvent::RedactedThinking { data })]
740 }
741 ResponseContent::ToolUse { id, name, .. } => {
742 self.tool_uses_by_index.insert(
743 index,
744 RawToolUse {
745 id,
746 name,
747 input_json: String::new(),
748 },
749 );
750 Vec::new()
751 }
752 },
753 Event::ContentBlockDelta { index, delta } => match delta {
754 ContentDelta::TextDelta { text } => {
755 vec![Ok(LanguageModelCompletionEvent::Text(text))]
756 }
757 ContentDelta::ThinkingDelta { thinking } => {
758 vec![Ok(LanguageModelCompletionEvent::Thinking {
759 text: thinking,
760 signature: None,
761 })]
762 }
763 ContentDelta::SignatureDelta { signature } => {
764 vec![Ok(LanguageModelCompletionEvent::Thinking {
765 text: "".to_string(),
766 signature: Some(signature),
767 })]
768 }
769 ContentDelta::InputJsonDelta { partial_json } => {
770 if let Some(tool_use) = self.tool_uses_by_index.get_mut(&index) {
771 tool_use.input_json.push_str(&partial_json);
772
773 // Try to convert invalid (incomplete) JSON into
774 // valid JSON that serde can accept, e.g. by closing
775 // unclosed delimiters. This way, we can update the
776 // UI with whatever has been streamed back so far.
777 if let Ok(input) = serde_json::Value::from_str(
778 &partial_json_fixer::fix_json(&tool_use.input_json),
779 ) {
780 return vec![Ok(LanguageModelCompletionEvent::ToolUse(
781 LanguageModelToolUse {
782 id: tool_use.id.clone().into(),
783 name: tool_use.name.clone().into(),
784 is_input_complete: false,
785 raw_input: tool_use.input_json.clone(),
786 input,
787 },
788 ))];
789 }
790 }
791 return vec![];
792 }
793 },
794 Event::ContentBlockStop { index } => {
795 if let Some(tool_use) = self.tool_uses_by_index.remove(&index) {
796 let input_json = tool_use.input_json.trim();
797 let input_value = if input_json.is_empty() {
798 Ok(serde_json::Value::Object(serde_json::Map::default()))
799 } else {
800 serde_json::Value::from_str(input_json)
801 };
802 let event_result = match input_value {
803 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
804 LanguageModelToolUse {
805 id: tool_use.id.into(),
806 name: tool_use.name.into(),
807 is_input_complete: true,
808 input,
809 raw_input: tool_use.input_json.clone(),
810 },
811 )),
812 Err(json_parse_err) => Err(LanguageModelCompletionError::BadInputJson {
813 id: tool_use.id.into(),
814 tool_name: tool_use.name.into(),
815 raw_input: input_json.into(),
816 json_parse_error: json_parse_err.to_string(),
817 }),
818 };
819
820 vec![event_result]
821 } else {
822 Vec::new()
823 }
824 }
825 Event::MessageStart { message } => {
826 update_usage(&mut self.usage, &message.usage);
827 vec![
828 Ok(LanguageModelCompletionEvent::UsageUpdate(convert_usage(
829 &self.usage,
830 ))),
831 Ok(LanguageModelCompletionEvent::StartMessage {
832 message_id: message.id,
833 }),
834 ]
835 }
836 Event::MessageDelta { delta, usage } => {
837 update_usage(&mut self.usage, &usage);
838 if let Some(stop_reason) = delta.stop_reason.as_deref() {
839 self.stop_reason = match stop_reason {
840 "end_turn" => StopReason::EndTurn,
841 "max_tokens" => StopReason::MaxTokens,
842 "tool_use" => StopReason::ToolUse,
843 "refusal" => StopReason::Refusal,
844 _ => {
845 log::error!("Unexpected anthropic stop_reason: {stop_reason}");
846 StopReason::EndTurn
847 }
848 };
849 }
850 vec![Ok(LanguageModelCompletionEvent::UsageUpdate(
851 convert_usage(&self.usage),
852 ))]
853 }
854 Event::MessageStop => {
855 vec![Ok(LanguageModelCompletionEvent::Stop(self.stop_reason))]
856 }
857 Event::Error { error } => {
858 vec![Err(LanguageModelCompletionError::Other(anyhow!(
859 AnthropicError::ApiError(error)
860 )))]
861 }
862 _ => Vec::new(),
863 }
864 }
865}
866
867struct RawToolUse {
868 id: String,
869 name: String,
870 input_json: String,
871}
872
873pub fn anthropic_err_to_anyhow(err: AnthropicError) -> anyhow::Error {
874 if let AnthropicError::ApiError(api_err) = &err {
875 if let Some(tokens) = api_err.match_window_exceeded() {
876 return anyhow!(LanguageModelKnownError::ContextWindowLimitExceeded { tokens });
877 }
878 }
879
880 anyhow!(err)
881}
882
883/// Updates usage data by preferring counts from `new`.
884fn update_usage(usage: &mut Usage, new: &Usage) {
885 if let Some(input_tokens) = new.input_tokens {
886 usage.input_tokens = Some(input_tokens);
887 }
888 if let Some(output_tokens) = new.output_tokens {
889 usage.output_tokens = Some(output_tokens);
890 }
891 if let Some(cache_creation_input_tokens) = new.cache_creation_input_tokens {
892 usage.cache_creation_input_tokens = Some(cache_creation_input_tokens);
893 }
894 if let Some(cache_read_input_tokens) = new.cache_read_input_tokens {
895 usage.cache_read_input_tokens = Some(cache_read_input_tokens);
896 }
897}
898
899fn convert_usage(usage: &Usage) -> language_model::TokenUsage {
900 language_model::TokenUsage {
901 input_tokens: usage.input_tokens.unwrap_or(0),
902 output_tokens: usage.output_tokens.unwrap_or(0),
903 cache_creation_input_tokens: usage.cache_creation_input_tokens.unwrap_or(0),
904 cache_read_input_tokens: usage.cache_read_input_tokens.unwrap_or(0),
905 }
906}
907
908struct ConfigurationView {
909 api_key_editor: Entity<Editor>,
910 state: gpui::Entity<State>,
911 load_credentials_task: Option<Task<()>>,
912}
913
914impl ConfigurationView {
915 const PLACEHOLDER_TEXT: &'static str = "sk-ant-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
916
917 fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
918 cx.observe(&state, |_, _, cx| {
919 cx.notify();
920 })
921 .detach();
922
923 let load_credentials_task = Some(cx.spawn({
924 let state = state.clone();
925 async move |this, cx| {
926 if let Some(task) = state
927 .update(cx, |state, cx| state.authenticate(cx))
928 .log_err()
929 {
930 // We don't log an error, because "not signed in" is also an error.
931 let _ = task.await;
932 }
933 this.update(cx, |this, cx| {
934 this.load_credentials_task = None;
935 cx.notify();
936 })
937 .log_err();
938 }
939 }));
940
941 Self {
942 api_key_editor: cx.new(|cx| {
943 let mut editor = Editor::single_line(window, cx);
944 editor.set_placeholder_text(Self::PLACEHOLDER_TEXT, cx);
945 editor
946 }),
947 state,
948 load_credentials_task,
949 }
950 }
951
952 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
953 let api_key = self.api_key_editor.read(cx).text(cx);
954 if api_key.is_empty() {
955 return;
956 }
957
958 let state = self.state.clone();
959 cx.spawn_in(window, async move |_, cx| {
960 state
961 .update(cx, |state, cx| state.set_api_key(api_key, cx))?
962 .await
963 })
964 .detach_and_log_err(cx);
965
966 cx.notify();
967 }
968
969 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
970 self.api_key_editor
971 .update(cx, |editor, cx| editor.set_text("", window, cx));
972
973 let state = self.state.clone();
974 cx.spawn_in(window, async move |_, cx| {
975 state.update(cx, |state, cx| state.reset_api_key(cx))?.await
976 })
977 .detach_and_log_err(cx);
978
979 cx.notify();
980 }
981
982 fn render_api_key_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
983 let settings = ThemeSettings::get_global(cx);
984 let text_style = TextStyle {
985 color: cx.theme().colors().text,
986 font_family: settings.ui_font.family.clone(),
987 font_features: settings.ui_font.features.clone(),
988 font_fallbacks: settings.ui_font.fallbacks.clone(),
989 font_size: rems(0.875).into(),
990 font_weight: settings.ui_font.weight,
991 font_style: FontStyle::Normal,
992 line_height: relative(1.3),
993 white_space: WhiteSpace::Normal,
994 ..Default::default()
995 };
996 EditorElement::new(
997 &self.api_key_editor,
998 EditorStyle {
999 background: cx.theme().colors().editor_background,
1000 local_player: cx.theme().players().local(),
1001 text: text_style,
1002 ..Default::default()
1003 },
1004 )
1005 }
1006
1007 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
1008 !self.state.read(cx).is_authenticated()
1009 }
1010}
1011
1012impl Render for ConfigurationView {
1013 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1014 let env_var_set = self.state.read(cx).api_key_from_env;
1015
1016 if self.load_credentials_task.is_some() {
1017 div().child(Label::new("Loading credentials...")).into_any()
1018 } else if self.should_render_editor(cx) {
1019 v_flex()
1020 .size_full()
1021 .on_action(cx.listener(Self::save_api_key))
1022 .child(Label::new("To use Zed's assistant with Anthropic, you need to add an API key. Follow these steps:"))
1023 .child(
1024 List::new()
1025 .child(
1026 InstructionListItem::new(
1027 "Create one by visiting",
1028 Some("Anthropic's settings"),
1029 Some("https://console.anthropic.com/settings/keys")
1030 )
1031 )
1032 .child(
1033 InstructionListItem::text_only("Paste your API key below and hit enter to start using the assistant")
1034 )
1035 )
1036 .child(
1037 h_flex()
1038 .w_full()
1039 .my_2()
1040 .px_2()
1041 .py_1()
1042 .bg(cx.theme().colors().editor_background)
1043 .border_1()
1044 .border_color(cx.theme().colors().border)
1045 .rounded_sm()
1046 .child(self.render_api_key_editor(cx)),
1047 )
1048 .child(
1049 Label::new(
1050 format!("You can also assign the {ANTHROPIC_API_KEY_VAR} environment variable and restart Zed."),
1051 )
1052 .size(LabelSize::Small)
1053 .color(Color::Muted),
1054 )
1055 .into_any()
1056 } else {
1057 h_flex()
1058 .mt_1()
1059 .p_1()
1060 .justify_between()
1061 .rounded_md()
1062 .border_1()
1063 .border_color(cx.theme().colors().border)
1064 .bg(cx.theme().colors().background)
1065 .child(
1066 h_flex()
1067 .gap_1()
1068 .child(Icon::new(IconName::Check).color(Color::Success))
1069 .child(Label::new(if env_var_set {
1070 format!("API key set in {ANTHROPIC_API_KEY_VAR} environment variable.")
1071 } else {
1072 "API key configured.".to_string()
1073 })),
1074 )
1075 .child(
1076 Button::new("reset-key", "Reset Key")
1077 .label_size(LabelSize::Small)
1078 .icon(Some(IconName::Trash))
1079 .icon_size(IconSize::Small)
1080 .icon_position(IconPosition::Start)
1081 .disabled(env_var_set)
1082 .when(env_var_set, |this| {
1083 this.tooltip(Tooltip::text(format!("To reset your API key, unset the {ANTHROPIC_API_KEY_VAR} environment variable.")))
1084 })
1085 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx))),
1086 )
1087 .into_any()
1088 }
1089 }
1090}
1091
1092#[cfg(test)]
1093mod tests {
1094 use super::*;
1095 use anthropic::AnthropicModelMode;
1096 use language_model::{LanguageModelRequestMessage, MessageContent};
1097
1098 #[test]
1099 fn test_cache_control_only_on_last_segment() {
1100 let request = LanguageModelRequest {
1101 messages: vec![LanguageModelRequestMessage {
1102 role: Role::User,
1103 content: vec![
1104 MessageContent::Text("Some prompt".to_string()),
1105 MessageContent::Image(language_model::LanguageModelImage::empty()),
1106 MessageContent::Image(language_model::LanguageModelImage::empty()),
1107 MessageContent::Image(language_model::LanguageModelImage::empty()),
1108 MessageContent::Image(language_model::LanguageModelImage::empty()),
1109 ],
1110 cache: true,
1111 }],
1112 thread_id: None,
1113 prompt_id: None,
1114 intent: None,
1115 mode: None,
1116 stop: vec![],
1117 temperature: None,
1118 tools: vec![],
1119 tool_choice: None,
1120 };
1121
1122 let anthropic_request = into_anthropic(
1123 request,
1124 "claude-3-5-sonnet".to_string(),
1125 0.7,
1126 4096,
1127 AnthropicModelMode::Default,
1128 );
1129
1130 assert_eq!(anthropic_request.messages.len(), 1);
1131
1132 let message = &anthropic_request.messages[0];
1133 assert_eq!(message.content.len(), 5);
1134
1135 assert!(matches!(
1136 message.content[0],
1137 anthropic::RequestContent::Text {
1138 cache_control: None,
1139 ..
1140 }
1141 ));
1142 for i in 1..3 {
1143 assert!(matches!(
1144 message.content[i],
1145 anthropic::RequestContent::Image {
1146 cache_control: None,
1147 ..
1148 }
1149 ));
1150 }
1151
1152 assert!(matches!(
1153 message.content[4],
1154 anthropic::RequestContent::Image {
1155 cache_control: Some(anthropic::CacheControl {
1156 cache_type: anthropic::CacheControlType::Ephemeral,
1157 }),
1158 ..
1159 }
1160 ));
1161 }
1162}