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, ConfigurationViewTargetAgent, LanguageModel,
19 LanguageModelCacheConfiguration, LanguageModelCompletionError, LanguageModelId,
20 LanguageModelName, 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: LanguageModelProviderId = language_model::ANTHROPIC_PROVIDER_ID;
37const PROVIDER_NAME: LanguageModelProviderName = language_model::ANTHROPIC_PROVIDER_NAME;
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}
45
46#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
47pub struct AvailableModel {
48 /// The model's name in the Anthropic API. e.g. claude-3-5-sonnet-latest, claude-3-opus-20240229, etc
49 pub name: String,
50 /// The model's name in Zed's UI, such as in the model selector dropdown menu in the assistant panel.
51 pub display_name: Option<String>,
52 /// The model's context window size.
53 pub max_tokens: u64,
54 /// A model `name` to substitute when calling tools, in case the primary model doesn't support tool calling.
55 pub tool_override: Option<String>,
56 /// Configuration of Anthropic's caching API.
57 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
58 pub max_output_tokens: Option<u64>,
59 pub default_temperature: Option<f32>,
60 #[serde(default)]
61 pub extra_beta_headers: Vec<String>,
62 /// The model's mode (e.g. thinking)
63 pub mode: Option<ModelMode>,
64}
65
66#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
67#[serde(tag = "type", rename_all = "lowercase")]
68pub enum ModelMode {
69 #[default]
70 Default,
71 Thinking {
72 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
73 budget_tokens: Option<u32>,
74 },
75}
76
77impl From<ModelMode> for AnthropicModelMode {
78 fn from(value: ModelMode) -> Self {
79 match value {
80 ModelMode::Default => AnthropicModelMode::Default,
81 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
82 }
83 }
84}
85
86impl From<AnthropicModelMode> for ModelMode {
87 fn from(value: AnthropicModelMode) -> Self {
88 match value {
89 AnthropicModelMode::Default => ModelMode::Default,
90 AnthropicModelMode::Thinking { budget_tokens } => ModelMode::Thinking { budget_tokens },
91 }
92 }
93}
94
95pub struct AnthropicLanguageModelProvider {
96 http_client: Arc<dyn HttpClient>,
97 state: gpui::Entity<State>,
98}
99
100const ANTHROPIC_API_KEY_VAR: &str = "ANTHROPIC_API_KEY";
101
102pub struct State {
103 api_key: Option<String>,
104 api_key_from_env: bool,
105 _subscription: Subscription,
106}
107
108impl State {
109 fn reset_api_key(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
110 let credentials_provider = <dyn CredentialsProvider>::global(cx);
111 let api_url = AllLanguageModelSettings::get_global(cx)
112 .anthropic
113 .api_url
114 .clone();
115 cx.spawn(async move |this, cx| {
116 credentials_provider
117 .delete_credentials(&api_url, cx)
118 .await
119 .ok();
120 this.update(cx, |this, cx| {
121 this.api_key = None;
122 this.api_key_from_env = false;
123 cx.notify();
124 })
125 })
126 }
127
128 fn set_api_key(&mut self, api_key: String, cx: &mut Context<Self>) -> Task<Result<()>> {
129 let credentials_provider = <dyn CredentialsProvider>::global(cx);
130 let api_url = AllLanguageModelSettings::get_global(cx)
131 .anthropic
132 .api_url
133 .clone();
134 cx.spawn(async move |this, cx| {
135 credentials_provider
136 .write_credentials(&api_url, "Bearer", api_key.as_bytes(), cx)
137 .await
138 .ok();
139
140 this.update(cx, |this, cx| {
141 this.api_key = Some(api_key);
142 cx.notify();
143 })
144 })
145 }
146
147 fn is_authenticated(&self) -> bool {
148 self.api_key.is_some()
149 }
150
151 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
152 if self.is_authenticated() {
153 return Task::ready(Ok(()));
154 }
155
156 let key = AnthropicLanguageModelProvider::api_key(cx);
157
158 cx.spawn(async move |this, cx| {
159 let key = key.await?;
160
161 this.update(cx, |this, cx| {
162 this.api_key = Some(key.key);
163 this.api_key_from_env = key.from_env;
164 cx.notify();
165 })?;
166
167 Ok(())
168 })
169 }
170}
171
172pub struct ApiKey {
173 pub key: String,
174 pub from_env: bool,
175}
176
177impl AnthropicLanguageModelProvider {
178 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
179 let state = cx.new(|cx| State {
180 api_key: None,
181 api_key_from_env: false,
182 _subscription: cx.observe_global::<SettingsStore>(|_, cx| {
183 cx.notify();
184 }),
185 });
186
187 Self { http_client, state }
188 }
189
190 fn create_language_model(&self, model: anthropic::Model) -> Arc<dyn LanguageModel> {
191 Arc::new(AnthropicModel {
192 id: LanguageModelId::from(model.id().to_string()),
193 model,
194 state: self.state.clone(),
195 http_client: self.http_client.clone(),
196 request_limiter: RateLimiter::new(4),
197 })
198 }
199
200 pub fn api_key(cx: &mut App) -> Task<Result<ApiKey, AuthenticateError>> {
201 let credentials_provider = <dyn CredentialsProvider>::global(cx);
202 let api_url = AllLanguageModelSettings::get_global(cx)
203 .anthropic
204 .api_url
205 .clone();
206
207 if let Ok(key) = std::env::var(ANTHROPIC_API_KEY_VAR) {
208 Task::ready(Ok(ApiKey {
209 key,
210 from_env: true,
211 }))
212 } else {
213 cx.spawn(async move |cx| {
214 let (_, api_key) = credentials_provider
215 .read_credentials(&api_url, cx)
216 .await?
217 .ok_or(AuthenticateError::CredentialsNotFound)?;
218
219 Ok(ApiKey {
220 key: String::from_utf8(api_key).context("invalid {PROVIDER_NAME} API key")?,
221 from_env: false,
222 })
223 })
224 }
225 }
226}
227
228impl LanguageModelProviderState for AnthropicLanguageModelProvider {
229 type ObservableEntity = State;
230
231 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
232 Some(self.state.clone())
233 }
234}
235
236impl LanguageModelProvider for AnthropicLanguageModelProvider {
237 fn id(&self) -> LanguageModelProviderId {
238 PROVIDER_ID
239 }
240
241 fn name(&self) -> LanguageModelProviderName {
242 PROVIDER_NAME
243 }
244
245 fn icon(&self) -> IconName {
246 IconName::AiAnthropic
247 }
248
249 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
250 Some(self.create_language_model(anthropic::Model::default()))
251 }
252
253 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
254 Some(self.create_language_model(anthropic::Model::default_fast()))
255 }
256
257 fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {
258 [
259 anthropic::Model::ClaudeSonnet4,
260 anthropic::Model::ClaudeSonnet4Thinking,
261 ]
262 .into_iter()
263 .map(|model| self.create_language_model(model))
264 .collect()
265 }
266
267 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
268 let mut models = BTreeMap::default();
269
270 // Add base models from anthropic::Model::iter()
271 for model in anthropic::Model::iter() {
272 if !matches!(model, anthropic::Model::Custom { .. }) {
273 models.insert(model.id().to_string(), model);
274 }
275 }
276
277 // Override with available models from settings
278 for model in AllLanguageModelSettings::get_global(cx)
279 .anthropic
280 .available_models
281 .iter()
282 {
283 models.insert(
284 model.name.clone(),
285 anthropic::Model::Custom {
286 name: model.name.clone(),
287 display_name: model.display_name.clone(),
288 max_tokens: model.max_tokens,
289 tool_override: model.tool_override.clone(),
290 cache_configuration: model.cache_configuration.as_ref().map(|config| {
291 anthropic::AnthropicModelCacheConfiguration {
292 max_cache_anchors: config.max_cache_anchors,
293 should_speculate: config.should_speculate,
294 min_total_token: config.min_total_token,
295 }
296 }),
297 max_output_tokens: model.max_output_tokens,
298 default_temperature: model.default_temperature,
299 extra_beta_headers: model.extra_beta_headers.clone(),
300 mode: model.mode.clone().unwrap_or_default().into(),
301 },
302 );
303 }
304
305 models
306 .into_values()
307 .map(|model| self.create_language_model(model))
308 .collect()
309 }
310
311 fn is_authenticated(&self, cx: &App) -> bool {
312 self.state.read(cx).is_authenticated()
313 }
314
315 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
316 self.state.update(cx, |state, cx| state.authenticate(cx))
317 }
318
319 fn configuration_view(
320 &self,
321 target_agent: ConfigurationViewTargetAgent,
322 window: &mut Window,
323 cx: &mut App,
324 ) -> AnyView {
325 cx.new(|cx| ConfigurationView::new(self.state.clone(), target_agent, window, cx))
326 .into()
327 }
328
329 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
330 self.state.update(cx, |state, cx| state.reset_api_key(cx))
331 }
332}
333
334pub struct AnthropicModel {
335 id: LanguageModelId,
336 model: anthropic::Model,
337 state: gpui::Entity<State>,
338 http_client: Arc<dyn HttpClient>,
339 request_limiter: RateLimiter,
340}
341
342pub fn count_anthropic_tokens(
343 request: LanguageModelRequest,
344 cx: &App,
345) -> BoxFuture<'static, Result<u64>> {
346 cx.background_spawn(async move {
347 let messages = request.messages;
348 let mut tokens_from_images = 0;
349 let mut string_messages = Vec::with_capacity(messages.len());
350
351 for message in messages {
352 use language_model::MessageContent;
353
354 let mut string_contents = String::new();
355
356 for content in message.content {
357 match content {
358 MessageContent::Text(text) => {
359 string_contents.push_str(&text);
360 }
361 MessageContent::Thinking { .. } => {
362 // Thinking blocks are not included in the input token count.
363 }
364 MessageContent::RedactedThinking(_) => {
365 // Thinking blocks are not included in the input token count.
366 }
367 MessageContent::Image(image) => {
368 tokens_from_images += image.estimate_tokens();
369 }
370 MessageContent::ToolUse(_tool_use) => {
371 // TODO: Estimate token usage from tool uses.
372 }
373 MessageContent::ToolResult(tool_result) => match &tool_result.content {
374 LanguageModelToolResultContent::Text(text) => {
375 string_contents.push_str(text);
376 }
377 LanguageModelToolResultContent::Image(image) => {
378 tokens_from_images += image.estimate_tokens();
379 }
380 },
381 }
382 }
383
384 if !string_contents.is_empty() {
385 string_messages.push(tiktoken_rs::ChatCompletionRequestMessage {
386 role: match message.role {
387 Role::User => "user".into(),
388 Role::Assistant => "assistant".into(),
389 Role::System => "system".into(),
390 },
391 content: Some(string_contents),
392 name: None,
393 function_call: None,
394 });
395 }
396 }
397
398 // Tiktoken doesn't yet support these models, so we manually use the
399 // same tokenizer as GPT-4.
400 tiktoken_rs::num_tokens_from_messages("gpt-4", &string_messages)
401 .map(|tokens| (tokens + tokens_from_images) as u64)
402 })
403 .boxed()
404}
405
406impl AnthropicModel {
407 fn stream_completion(
408 &self,
409 request: anthropic::Request,
410 cx: &AsyncApp,
411 ) -> BoxFuture<
412 'static,
413 Result<
414 BoxStream<'static, Result<anthropic::Event, AnthropicError>>,
415 LanguageModelCompletionError,
416 >,
417 > {
418 let http_client = self.http_client.clone();
419
420 let Ok((api_key, api_url)) = cx.read_entity(&self.state, |state, cx| {
421 let settings = &AllLanguageModelSettings::get_global(cx).anthropic;
422 (state.api_key.clone(), settings.api_url.clone())
423 }) else {
424 return futures::future::ready(Err(anyhow!("App state dropped").into())).boxed();
425 };
426
427 let beta_headers = self.model.beta_headers();
428
429 async move {
430 let Some(api_key) = api_key else {
431 return Err(LanguageModelCompletionError::NoApiKey {
432 provider: PROVIDER_NAME,
433 });
434 };
435 let request = anthropic::stream_completion(
436 http_client.as_ref(),
437 &api_url,
438 &api_key,
439 request,
440 beta_headers,
441 );
442 request.await.map_err(Into::into)
443 }
444 .boxed()
445 }
446}
447
448impl LanguageModel for AnthropicModel {
449 fn id(&self) -> LanguageModelId {
450 self.id.clone()
451 }
452
453 fn name(&self) -> LanguageModelName {
454 LanguageModelName::from(self.model.display_name().to_string())
455 }
456
457 fn provider_id(&self) -> LanguageModelProviderId {
458 PROVIDER_ID
459 }
460
461 fn provider_name(&self) -> LanguageModelProviderName {
462 PROVIDER_NAME
463 }
464
465 fn supports_tools(&self) -> bool {
466 true
467 }
468
469 fn supports_images(&self) -> bool {
470 true
471 }
472
473 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
474 match choice {
475 LanguageModelToolChoice::Auto
476 | LanguageModelToolChoice::Any
477 | LanguageModelToolChoice::None => true,
478 }
479 }
480
481 fn telemetry_id(&self) -> String {
482 format!("anthropic/{}", self.model.id())
483 }
484
485 fn api_key(&self, cx: &App) -> Option<String> {
486 self.state.read(cx).api_key.clone()
487 }
488
489 fn max_token_count(&self) -> u64 {
490 self.model.max_token_count()
491 }
492
493 fn max_output_tokens(&self) -> Option<u64> {
494 Some(self.model.max_output_tokens())
495 }
496
497 fn count_tokens(
498 &self,
499 request: LanguageModelRequest,
500 cx: &App,
501 ) -> BoxFuture<'static, Result<u64>> {
502 count_anthropic_tokens(request, cx)
503 }
504
505 fn stream_completion(
506 &self,
507 request: LanguageModelRequest,
508 cx: &AsyncApp,
509 ) -> BoxFuture<
510 'static,
511 Result<
512 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
513 LanguageModelCompletionError,
514 >,
515 > {
516 let request = into_anthropic(
517 request,
518 self.model.request_id().into(),
519 self.model.default_temperature(),
520 self.model.max_output_tokens(),
521 self.model.mode(),
522 );
523 let request = self.stream_completion(request, cx);
524 let future = self.request_limiter.stream(async move {
525 let response = request.await?;
526 Ok(AnthropicEventMapper::new().map_stream(response))
527 });
528 async move { Ok(future.await?.boxed()) }.boxed()
529 }
530
531 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
532 self.model
533 .cache_configuration()
534 .map(|config| LanguageModelCacheConfiguration {
535 max_cache_anchors: config.max_cache_anchors,
536 should_speculate: config.should_speculate,
537 min_total_token: config.min_total_token,
538 })
539 }
540}
541
542pub fn into_anthropic(
543 request: LanguageModelRequest,
544 model: String,
545 default_temperature: f32,
546 max_output_tokens: u64,
547 mode: AnthropicModelMode,
548) -> anthropic::Request {
549 let mut new_messages: Vec<anthropic::Message> = Vec::new();
550 let mut system_message = String::new();
551
552 for message in request.messages {
553 if message.contents_empty() {
554 continue;
555 }
556
557 match message.role {
558 Role::User | Role::Assistant => {
559 let mut anthropic_message_content: Vec<anthropic::RequestContent> = message
560 .content
561 .into_iter()
562 .filter_map(|content| match content {
563 MessageContent::Text(text) => {
564 let text = if text.chars().last().is_some_and(|c| c.is_whitespace()) {
565 text.trim_end().to_string()
566 } else {
567 text
568 };
569 if !text.is_empty() {
570 Some(anthropic::RequestContent::Text {
571 text,
572 cache_control: None,
573 })
574 } else {
575 None
576 }
577 }
578 MessageContent::Thinking {
579 text: thinking,
580 signature,
581 } => {
582 if !thinking.is_empty() {
583 Some(anthropic::RequestContent::Thinking {
584 thinking,
585 signature: signature.unwrap_or_default(),
586 cache_control: None,
587 })
588 } else {
589 None
590 }
591 }
592 MessageContent::RedactedThinking(data) => {
593 if !data.is_empty() {
594 Some(anthropic::RequestContent::RedactedThinking { data })
595 } else {
596 None
597 }
598 }
599 MessageContent::Image(image) => Some(anthropic::RequestContent::Image {
600 source: anthropic::ImageSource {
601 source_type: "base64".to_string(),
602 media_type: "image/png".to_string(),
603 data: image.source.to_string(),
604 },
605 cache_control: None,
606 }),
607 MessageContent::ToolUse(tool_use) => {
608 Some(anthropic::RequestContent::ToolUse {
609 id: tool_use.id.to_string(),
610 name: tool_use.name.to_string(),
611 input: tool_use.input,
612 cache_control: None,
613 })
614 }
615 MessageContent::ToolResult(tool_result) => {
616 Some(anthropic::RequestContent::ToolResult {
617 tool_use_id: tool_result.tool_use_id.to_string(),
618 is_error: tool_result.is_error,
619 content: match tool_result.content {
620 LanguageModelToolResultContent::Text(text) => {
621 ToolResultContent::Plain(text.to_string())
622 }
623 LanguageModelToolResultContent::Image(image) => {
624 ToolResultContent::Multipart(vec![ToolResultPart::Image {
625 source: anthropic::ImageSource {
626 source_type: "base64".to_string(),
627 media_type: "image/png".to_string(),
628 data: image.source.to_string(),
629 },
630 }])
631 }
632 },
633 cache_control: None,
634 })
635 }
636 })
637 .collect();
638 let anthropic_role = match message.role {
639 Role::User => anthropic::Role::User,
640 Role::Assistant => anthropic::Role::Assistant,
641 Role::System => unreachable!("System role should never occur here"),
642 };
643 if let Some(last_message) = new_messages.last_mut()
644 && last_message.role == anthropic_role
645 {
646 last_message.content.extend(anthropic_message_content);
647 continue;
648 }
649
650 // Mark the last segment of the message as cached
651 if message.cache {
652 let cache_control_value = Some(anthropic::CacheControl {
653 cache_type: anthropic::CacheControlType::Ephemeral,
654 });
655 for message_content in anthropic_message_content.iter_mut().rev() {
656 match message_content {
657 anthropic::RequestContent::RedactedThinking { .. } => {
658 // Caching is not possible, fallback to next message
659 }
660 anthropic::RequestContent::Text { cache_control, .. }
661 | anthropic::RequestContent::Thinking { cache_control, .. }
662 | anthropic::RequestContent::Image { cache_control, .. }
663 | anthropic::RequestContent::ToolUse { cache_control, .. }
664 | anthropic::RequestContent::ToolResult { cache_control, .. } => {
665 *cache_control = cache_control_value;
666 break;
667 }
668 }
669 }
670 }
671
672 new_messages.push(anthropic::Message {
673 role: anthropic_role,
674 content: anthropic_message_content,
675 });
676 }
677 Role::System => {
678 if !system_message.is_empty() {
679 system_message.push_str("\n\n");
680 }
681 system_message.push_str(&message.string_contents());
682 }
683 }
684 }
685
686 anthropic::Request {
687 model,
688 messages: new_messages,
689 max_tokens: max_output_tokens,
690 system: if system_message.is_empty() {
691 None
692 } else {
693 Some(anthropic::StringOrContents::String(system_message))
694 },
695 thinking: if request.thinking_allowed
696 && let AnthropicModelMode::Thinking { budget_tokens } = mode
697 {
698 Some(anthropic::Thinking::Enabled { budget_tokens })
699 } else {
700 None
701 },
702 tools: request
703 .tools
704 .into_iter()
705 .map(|tool| anthropic::Tool {
706 name: tool.name,
707 description: tool.description,
708 input_schema: tool.input_schema,
709 })
710 .collect(),
711 tool_choice: request.tool_choice.map(|choice| match choice {
712 LanguageModelToolChoice::Auto => anthropic::ToolChoice::Auto,
713 LanguageModelToolChoice::Any => anthropic::ToolChoice::Any,
714 LanguageModelToolChoice::None => anthropic::ToolChoice::None,
715 }),
716 metadata: None,
717 stop_sequences: Vec::new(),
718 temperature: request.temperature.or(Some(default_temperature)),
719 top_k: None,
720 top_p: None,
721 }
722}
723
724pub struct AnthropicEventMapper {
725 tool_uses_by_index: HashMap<usize, RawToolUse>,
726 usage: Usage,
727 stop_reason: StopReason,
728}
729
730impl AnthropicEventMapper {
731 pub fn new() -> Self {
732 Self {
733 tool_uses_by_index: HashMap::default(),
734 usage: Usage::default(),
735 stop_reason: StopReason::EndTurn,
736 }
737 }
738
739 pub fn map_stream(
740 mut self,
741 events: Pin<Box<dyn Send + Stream<Item = Result<Event, AnthropicError>>>>,
742 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
743 {
744 events.flat_map(move |event| {
745 futures::stream::iter(match event {
746 Ok(event) => self.map_event(event),
747 Err(error) => vec![Err(error.into())],
748 })
749 })
750 }
751
752 pub fn map_event(
753 &mut self,
754 event: Event,
755 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
756 match event {
757 Event::ContentBlockStart {
758 index,
759 content_block,
760 } => match content_block {
761 ResponseContent::Text { text } => {
762 vec![Ok(LanguageModelCompletionEvent::Text(text))]
763 }
764 ResponseContent::Thinking { thinking } => {
765 vec![Ok(LanguageModelCompletionEvent::Thinking {
766 text: thinking,
767 signature: None,
768 })]
769 }
770 ResponseContent::RedactedThinking { data } => {
771 vec![Ok(LanguageModelCompletionEvent::RedactedThinking { data })]
772 }
773 ResponseContent::ToolUse { id, name, .. } => {
774 self.tool_uses_by_index.insert(
775 index,
776 RawToolUse {
777 id,
778 name,
779 input_json: String::new(),
780 },
781 );
782 Vec::new()
783 }
784 },
785 Event::ContentBlockDelta { index, delta } => match delta {
786 ContentDelta::TextDelta { text } => {
787 vec![Ok(LanguageModelCompletionEvent::Text(text))]
788 }
789 ContentDelta::ThinkingDelta { thinking } => {
790 vec![Ok(LanguageModelCompletionEvent::Thinking {
791 text: thinking,
792 signature: None,
793 })]
794 }
795 ContentDelta::SignatureDelta { signature } => {
796 vec![Ok(LanguageModelCompletionEvent::Thinking {
797 text: "".to_string(),
798 signature: Some(signature),
799 })]
800 }
801 ContentDelta::InputJsonDelta { partial_json } => {
802 if let Some(tool_use) = self.tool_uses_by_index.get_mut(&index) {
803 tool_use.input_json.push_str(&partial_json);
804
805 // Try to convert invalid (incomplete) JSON into
806 // valid JSON that serde can accept, e.g. by closing
807 // unclosed delimiters. This way, we can update the
808 // UI with whatever has been streamed back so far.
809 if let Ok(input) = serde_json::Value::from_str(
810 &partial_json_fixer::fix_json(&tool_use.input_json),
811 ) {
812 return vec![Ok(LanguageModelCompletionEvent::ToolUse(
813 LanguageModelToolUse {
814 id: tool_use.id.clone().into(),
815 name: tool_use.name.clone().into(),
816 is_input_complete: false,
817 raw_input: tool_use.input_json.clone(),
818 input,
819 },
820 ))];
821 }
822 }
823 vec![]
824 }
825 },
826 Event::ContentBlockStop { index } => {
827 if let Some(tool_use) = self.tool_uses_by_index.remove(&index) {
828 let input_json = tool_use.input_json.trim();
829 let input_value = if input_json.is_empty() {
830 Ok(serde_json::Value::Object(serde_json::Map::default()))
831 } else {
832 serde_json::Value::from_str(input_json)
833 };
834 let event_result = match input_value {
835 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
836 LanguageModelToolUse {
837 id: tool_use.id.into(),
838 name: tool_use.name.into(),
839 is_input_complete: true,
840 input,
841 raw_input: tool_use.input_json.clone(),
842 },
843 )),
844 Err(json_parse_err) => {
845 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
846 id: tool_use.id.into(),
847 tool_name: tool_use.name.into(),
848 raw_input: input_json.into(),
849 json_parse_error: json_parse_err.to_string(),
850 })
851 }
852 };
853
854 vec![event_result]
855 } else {
856 Vec::new()
857 }
858 }
859 Event::MessageStart { message } => {
860 update_usage(&mut self.usage, &message.usage);
861 vec![
862 Ok(LanguageModelCompletionEvent::UsageUpdate(convert_usage(
863 &self.usage,
864 ))),
865 Ok(LanguageModelCompletionEvent::StartMessage {
866 message_id: message.id,
867 }),
868 ]
869 }
870 Event::MessageDelta { delta, usage } => {
871 update_usage(&mut self.usage, &usage);
872 if let Some(stop_reason) = delta.stop_reason.as_deref() {
873 self.stop_reason = match stop_reason {
874 "end_turn" => StopReason::EndTurn,
875 "max_tokens" => StopReason::MaxTokens,
876 "tool_use" => StopReason::ToolUse,
877 "refusal" => StopReason::Refusal,
878 _ => {
879 log::error!("Unexpected anthropic stop_reason: {stop_reason}");
880 StopReason::EndTurn
881 }
882 };
883 }
884 vec![Ok(LanguageModelCompletionEvent::UsageUpdate(
885 convert_usage(&self.usage),
886 ))]
887 }
888 Event::MessageStop => {
889 vec![Ok(LanguageModelCompletionEvent::Stop(self.stop_reason))]
890 }
891 Event::Error { error } => {
892 vec![Err(error.into())]
893 }
894 _ => Vec::new(),
895 }
896 }
897}
898
899struct RawToolUse {
900 id: String,
901 name: String,
902 input_json: String,
903}
904
905/// Updates usage data by preferring counts from `new`.
906fn update_usage(usage: &mut Usage, new: &Usage) {
907 if let Some(input_tokens) = new.input_tokens {
908 usage.input_tokens = Some(input_tokens);
909 }
910 if let Some(output_tokens) = new.output_tokens {
911 usage.output_tokens = Some(output_tokens);
912 }
913 if let Some(cache_creation_input_tokens) = new.cache_creation_input_tokens {
914 usage.cache_creation_input_tokens = Some(cache_creation_input_tokens);
915 }
916 if let Some(cache_read_input_tokens) = new.cache_read_input_tokens {
917 usage.cache_read_input_tokens = Some(cache_read_input_tokens);
918 }
919}
920
921fn convert_usage(usage: &Usage) -> language_model::TokenUsage {
922 language_model::TokenUsage {
923 input_tokens: usage.input_tokens.unwrap_or(0),
924 output_tokens: usage.output_tokens.unwrap_or(0),
925 cache_creation_input_tokens: usage.cache_creation_input_tokens.unwrap_or(0),
926 cache_read_input_tokens: usage.cache_read_input_tokens.unwrap_or(0),
927 }
928}
929
930struct ConfigurationView {
931 api_key_editor: Entity<Editor>,
932 state: gpui::Entity<State>,
933 load_credentials_task: Option<Task<()>>,
934 target_agent: ConfigurationViewTargetAgent,
935}
936
937impl ConfigurationView {
938 const PLACEHOLDER_TEXT: &'static str = "sk-ant-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
939
940 fn new(
941 state: gpui::Entity<State>,
942 target_agent: ConfigurationViewTargetAgent,
943 window: &mut Window,
944 cx: &mut Context<Self>,
945 ) -> Self {
946 cx.observe(&state, |_, _, cx| {
947 cx.notify();
948 })
949 .detach();
950
951 let load_credentials_task = Some(cx.spawn({
952 let state = state.clone();
953 async move |this, cx| {
954 if let Some(task) = state
955 .update(cx, |state, cx| state.authenticate(cx))
956 .log_err()
957 {
958 // We don't log an error, because "not signed in" is also an error.
959 let _ = task.await;
960 }
961 this.update(cx, |this, cx| {
962 this.load_credentials_task = None;
963 cx.notify();
964 })
965 .log_err();
966 }
967 }));
968
969 Self {
970 api_key_editor: cx.new(|cx| {
971 let mut editor = Editor::single_line(window, cx);
972 editor.set_placeholder_text(Self::PLACEHOLDER_TEXT, window, cx);
973 editor
974 }),
975 state,
976 load_credentials_task,
977 target_agent,
978 }
979 }
980
981 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
982 let api_key = self.api_key_editor.read(cx).text(cx);
983 if api_key.is_empty() {
984 return;
985 }
986
987 let state = self.state.clone();
988 cx.spawn_in(window, async move |_, cx| {
989 state
990 .update(cx, |state, cx| state.set_api_key(api_key, cx))?
991 .await
992 })
993 .detach_and_log_err(cx);
994
995 cx.notify();
996 }
997
998 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
999 self.api_key_editor
1000 .update(cx, |editor, cx| editor.set_text("", window, cx));
1001
1002 let state = self.state.clone();
1003 cx.spawn_in(window, async move |_, cx| {
1004 state.update(cx, |state, cx| state.reset_api_key(cx))?.await
1005 })
1006 .detach_and_log_err(cx);
1007
1008 cx.notify();
1009 }
1010
1011 fn render_api_key_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1012 let settings = ThemeSettings::get_global(cx);
1013 let text_style = TextStyle {
1014 color: cx.theme().colors().text,
1015 font_family: settings.ui_font.family.clone(),
1016 font_features: settings.ui_font.features.clone(),
1017 font_fallbacks: settings.ui_font.fallbacks.clone(),
1018 font_size: rems(0.875).into(),
1019 font_weight: settings.ui_font.weight,
1020 font_style: FontStyle::Normal,
1021 line_height: relative(1.3),
1022 white_space: WhiteSpace::Normal,
1023 ..Default::default()
1024 };
1025 EditorElement::new(
1026 &self.api_key_editor,
1027 EditorStyle {
1028 background: cx.theme().colors().editor_background,
1029 local_player: cx.theme().players().local(),
1030 text: text_style,
1031 ..Default::default()
1032 },
1033 )
1034 }
1035
1036 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
1037 !self.state.read(cx).is_authenticated()
1038 }
1039}
1040
1041impl Render for ConfigurationView {
1042 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1043 let env_var_set = self.state.read(cx).api_key_from_env;
1044
1045 if self.load_credentials_task.is_some() {
1046 div().child(Label::new("Loading credentials...")).into_any()
1047 } else if self.should_render_editor(cx) {
1048 v_flex()
1049 .size_full()
1050 .on_action(cx.listener(Self::save_api_key))
1051 .child(Label::new(format!("To use {}, you need to add an API key. Follow these steps:", match &self.target_agent {
1052 ConfigurationViewTargetAgent::ZedAgent => "Zed's agent with Anthropic".into(),
1053 ConfigurationViewTargetAgent::Other(agent) => agent.clone(),
1054 })))
1055 .child(
1056 List::new()
1057 .child(
1058 InstructionListItem::new(
1059 "Create one by visiting",
1060 Some("Anthropic's settings"),
1061 Some("https://console.anthropic.com/settings/keys")
1062 )
1063 )
1064 .child(
1065 InstructionListItem::text_only("Paste your API key below and hit enter to start using the agent")
1066 )
1067 )
1068 .child(
1069 h_flex()
1070 .w_full()
1071 .my_2()
1072 .px_2()
1073 .py_1()
1074 .bg(cx.theme().colors().editor_background)
1075 .border_1()
1076 .border_color(cx.theme().colors().border)
1077 .rounded_sm()
1078 .child(self.render_api_key_editor(cx)),
1079 )
1080 .child(
1081 Label::new(
1082 format!("You can also assign the {ANTHROPIC_API_KEY_VAR} environment variable and restart Zed."),
1083 )
1084 .size(LabelSize::Small)
1085 .color(Color::Muted),
1086 )
1087 .into_any()
1088 } else {
1089 h_flex()
1090 .mt_1()
1091 .p_1()
1092 .justify_between()
1093 .rounded_md()
1094 .border_1()
1095 .border_color(cx.theme().colors().border)
1096 .bg(cx.theme().colors().background)
1097 .child(
1098 h_flex()
1099 .gap_1()
1100 .child(Icon::new(IconName::Check).color(Color::Success))
1101 .child(Label::new(if env_var_set {
1102 format!("API key set in {ANTHROPIC_API_KEY_VAR} environment variable.")
1103 } else {
1104 "API key configured.".to_string()
1105 })),
1106 )
1107 .child(
1108 Button::new("reset-key", "Reset Key")
1109 .label_size(LabelSize::Small)
1110 .icon(Some(IconName::Trash))
1111 .icon_size(IconSize::Small)
1112 .icon_position(IconPosition::Start)
1113 .disabled(env_var_set)
1114 .when(env_var_set, |this| {
1115 this.tooltip(Tooltip::text(format!("To reset your API key, unset the {ANTHROPIC_API_KEY_VAR} environment variable.")))
1116 })
1117 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx))),
1118 )
1119 .into_any()
1120 }
1121 }
1122}
1123
1124#[cfg(test)]
1125mod tests {
1126 use super::*;
1127 use anthropic::AnthropicModelMode;
1128 use language_model::{LanguageModelRequestMessage, MessageContent};
1129
1130 #[test]
1131 fn test_cache_control_only_on_last_segment() {
1132 let request = LanguageModelRequest {
1133 messages: vec![LanguageModelRequestMessage {
1134 role: Role::User,
1135 content: vec![
1136 MessageContent::Text("Some prompt".to_string()),
1137 MessageContent::Image(language_model::LanguageModelImage::empty()),
1138 MessageContent::Image(language_model::LanguageModelImage::empty()),
1139 MessageContent::Image(language_model::LanguageModelImage::empty()),
1140 MessageContent::Image(language_model::LanguageModelImage::empty()),
1141 ],
1142 cache: true,
1143 }],
1144 thread_id: None,
1145 prompt_id: None,
1146 intent: None,
1147 mode: None,
1148 stop: vec![],
1149 temperature: None,
1150 tools: vec![],
1151 tool_choice: None,
1152 thinking_allowed: true,
1153 };
1154
1155 let anthropic_request = into_anthropic(
1156 request,
1157 "claude-3-5-sonnet".to_string(),
1158 0.7,
1159 4096,
1160 AnthropicModelMode::Default,
1161 );
1162
1163 assert_eq!(anthropic_request.messages.len(), 1);
1164
1165 let message = &anthropic_request.messages[0];
1166 assert_eq!(message.content.len(), 5);
1167
1168 assert!(matches!(
1169 message.content[0],
1170 anthropic::RequestContent::Text {
1171 cache_control: None,
1172 ..
1173 }
1174 ));
1175 for i in 1..3 {
1176 assert!(matches!(
1177 message.content[i],
1178 anthropic::RequestContent::Image {
1179 cache_control: None,
1180 ..
1181 }
1182 ));
1183 }
1184
1185 assert!(matches!(
1186 message.content[4],
1187 anthropic::RequestContent::Image {
1188 cache_control: Some(anthropic::CacheControl {
1189 cache_type: anthropic::CacheControlType::Ephemeral,
1190 }),
1191 ..
1192 }
1193 ));
1194 }
1195}