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