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: &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(Into::into)
411 }
412 .boxed()
413 }
414}
415
416impl LanguageModel for AnthropicModel {
417 fn id(&self) -> LanguageModelId {
418 self.id.clone()
419 }
420
421 fn name(&self) -> LanguageModelName {
422 LanguageModelName::from(self.model.display_name().to_string())
423 }
424
425 fn provider_id(&self) -> LanguageModelProviderId {
426 LanguageModelProviderId(PROVIDER_ID.into())
427 }
428
429 fn provider_name(&self) -> LanguageModelProviderName {
430 LanguageModelProviderName(PROVIDER_NAME.into())
431 }
432
433 fn supports_tools(&self) -> bool {
434 true
435 }
436
437 fn supports_images(&self) -> bool {
438 true
439 }
440
441 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
442 match choice {
443 LanguageModelToolChoice::Auto
444 | LanguageModelToolChoice::Any
445 | LanguageModelToolChoice::None => true,
446 }
447 }
448
449 fn telemetry_id(&self) -> String {
450 format!("anthropic/{}", self.model.id())
451 }
452
453 fn api_key(&self, cx: &App) -> Option<String> {
454 self.state.read(cx).api_key.clone()
455 }
456
457 fn max_token_count(&self) -> u64 {
458 self.model.max_token_count()
459 }
460
461 fn max_output_tokens(&self) -> Option<u64> {
462 Some(self.model.max_output_tokens())
463 }
464
465 fn count_tokens(
466 &self,
467 request: LanguageModelRequest,
468 cx: &App,
469 ) -> BoxFuture<'static, Result<u64>> {
470 count_anthropic_tokens(request, cx)
471 }
472
473 fn stream_completion(
474 &self,
475 request: LanguageModelRequest,
476 cx: &AsyncApp,
477 ) -> BoxFuture<
478 'static,
479 Result<
480 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
481 LanguageModelCompletionError,
482 >,
483 > {
484 let request = into_anthropic(
485 request,
486 self.model.request_id().into(),
487 self.model.default_temperature(),
488 self.model.max_output_tokens(),
489 self.model.mode(),
490 );
491 let request = self.stream_completion(request, cx);
492 let future = self.request_limiter.stream(async move {
493 let response = request.await?;
494 Ok(AnthropicEventMapper::new().map_stream(response))
495 });
496 async move { Ok(future.await?.boxed()) }.boxed()
497 }
498
499 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
500 self.model
501 .cache_configuration()
502 .map(|config| LanguageModelCacheConfiguration {
503 max_cache_anchors: config.max_cache_anchors,
504 should_speculate: config.should_speculate,
505 min_total_token: config.min_total_token,
506 })
507 }
508}
509
510pub fn into_anthropic(
511 request: LanguageModelRequest,
512 model: String,
513 default_temperature: f32,
514 max_output_tokens: u64,
515 mode: AnthropicModelMode,
516) -> anthropic::Request {
517 let mut new_messages: Vec<anthropic::Message> = Vec::new();
518 let mut system_message = String::new();
519
520 for message in request.messages {
521 if message.contents_empty() {
522 continue;
523 }
524
525 match message.role {
526 Role::User | Role::Assistant => {
527 let mut anthropic_message_content: Vec<anthropic::RequestContent> = message
528 .content
529 .into_iter()
530 .filter_map(|content| match content {
531 MessageContent::Text(text) => {
532 if !text.is_empty() {
533 Some(anthropic::RequestContent::Text {
534 text,
535 cache_control: None,
536 })
537 } else {
538 None
539 }
540 }
541 MessageContent::Thinking {
542 text: thinking,
543 signature,
544 } => {
545 if !thinking.is_empty() {
546 Some(anthropic::RequestContent::Thinking {
547 thinking,
548 signature: signature.unwrap_or_default(),
549 cache_control: None,
550 })
551 } else {
552 None
553 }
554 }
555 MessageContent::RedactedThinking(data) => {
556 if !data.is_empty() {
557 Some(anthropic::RequestContent::RedactedThinking {
558 data: String::from_utf8(data).ok()?,
559 })
560 } else {
561 None
562 }
563 }
564 MessageContent::Image(image) => Some(anthropic::RequestContent::Image {
565 source: anthropic::ImageSource {
566 source_type: "base64".to_string(),
567 media_type: "image/png".to_string(),
568 data: image.source.to_string(),
569 },
570 cache_control: None,
571 }),
572 MessageContent::ToolUse(tool_use) => {
573 Some(anthropic::RequestContent::ToolUse {
574 id: tool_use.id.to_string(),
575 name: tool_use.name.to_string(),
576 input: tool_use.input,
577 cache_control: None,
578 })
579 }
580 MessageContent::ToolResult(tool_result) => {
581 Some(anthropic::RequestContent::ToolResult {
582 tool_use_id: tool_result.tool_use_id.to_string(),
583 is_error: tool_result.is_error,
584 content: match tool_result.content {
585 LanguageModelToolResultContent::Text(text) => {
586 ToolResultContent::Plain(text.to_string())
587 }
588 LanguageModelToolResultContent::Image(image) => {
589 ToolResultContent::Multipart(vec![ToolResultPart::Image {
590 source: anthropic::ImageSource {
591 source_type: "base64".to_string(),
592 media_type: "image/png".to_string(),
593 data: image.source.to_string(),
594 },
595 }])
596 }
597 },
598 cache_control: None,
599 })
600 }
601 })
602 .collect();
603 let anthropic_role = match message.role {
604 Role::User => anthropic::Role::User,
605 Role::Assistant => anthropic::Role::Assistant,
606 Role::System => unreachable!("System role should never occur here"),
607 };
608 if let Some(last_message) = new_messages.last_mut() {
609 if last_message.role == anthropic_role {
610 last_message.content.extend(anthropic_message_content);
611 continue;
612 }
613 }
614
615 // Mark the last segment of the message as cached
616 if message.cache {
617 let cache_control_value = Some(anthropic::CacheControl {
618 cache_type: anthropic::CacheControlType::Ephemeral,
619 });
620 for message_content in anthropic_message_content.iter_mut().rev() {
621 match message_content {
622 anthropic::RequestContent::RedactedThinking { .. } => {
623 // Caching is not possible, fallback to next message
624 }
625 anthropic::RequestContent::Text { cache_control, .. }
626 | anthropic::RequestContent::Thinking { cache_control, .. }
627 | anthropic::RequestContent::Image { cache_control, .. }
628 | anthropic::RequestContent::ToolUse { cache_control, .. }
629 | anthropic::RequestContent::ToolResult { cache_control, .. } => {
630 *cache_control = cache_control_value;
631 break;
632 }
633 }
634 }
635 }
636
637 new_messages.push(anthropic::Message {
638 role: anthropic_role,
639 content: anthropic_message_content,
640 });
641 }
642 Role::System => {
643 if !system_message.is_empty() {
644 system_message.push_str("\n\n");
645 }
646 system_message.push_str(&message.string_contents());
647 }
648 }
649 }
650
651 anthropic::Request {
652 model,
653 messages: new_messages,
654 max_tokens: max_output_tokens,
655 system: if system_message.is_empty() {
656 None
657 } else {
658 Some(anthropic::StringOrContents::String(system_message))
659 },
660 thinking: if let AnthropicModelMode::Thinking { budget_tokens } = mode {
661 Some(anthropic::Thinking::Enabled { budget_tokens })
662 } else {
663 None
664 },
665 tools: request
666 .tools
667 .into_iter()
668 .map(|tool| anthropic::Tool {
669 name: tool.name,
670 description: tool.description,
671 input_schema: tool.input_schema,
672 })
673 .collect(),
674 tool_choice: request.tool_choice.map(|choice| match choice {
675 LanguageModelToolChoice::Auto => anthropic::ToolChoice::Auto,
676 LanguageModelToolChoice::Any => anthropic::ToolChoice::Any,
677 LanguageModelToolChoice::None => anthropic::ToolChoice::None,
678 }),
679 metadata: None,
680 stop_sequences: Vec::new(),
681 temperature: request.temperature.or(Some(default_temperature)),
682 top_k: None,
683 top_p: None,
684 }
685}
686
687pub struct AnthropicEventMapper {
688 tool_uses_by_index: HashMap<usize, RawToolUse>,
689 usage: Usage,
690 stop_reason: StopReason,
691}
692
693impl AnthropicEventMapper {
694 pub fn new() -> Self {
695 Self {
696 tool_uses_by_index: HashMap::default(),
697 usage: Usage::default(),
698 stop_reason: StopReason::EndTurn,
699 }
700 }
701
702 pub fn map_stream(
703 mut self,
704 events: Pin<Box<dyn Send + Stream<Item = Result<Event, AnthropicError>>>>,
705 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
706 {
707 events.flat_map(move |event| {
708 futures::stream::iter(match event {
709 Ok(event) => self.map_event(event),
710 Err(error) => vec![Err(error.into())],
711 })
712 })
713 }
714
715 pub fn map_event(
716 &mut self,
717 event: Event,
718 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
719 match event {
720 Event::ContentBlockStart {
721 index,
722 content_block,
723 } => match content_block {
724 ResponseContent::Text { text } => {
725 vec![Ok(LanguageModelCompletionEvent::Text(text))]
726 }
727 ResponseContent::Thinking { thinking } => {
728 vec![Ok(LanguageModelCompletionEvent::Thinking {
729 text: thinking,
730 signature: None,
731 })]
732 }
733 ResponseContent::RedactedThinking { .. } => {
734 // Redacted thinking is encrypted and not accessible to the user, see:
735 // https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#suggestions-for-handling-redacted-thinking-in-production
736 Vec::new()
737 }
738 ResponseContent::ToolUse { id, name, .. } => {
739 self.tool_uses_by_index.insert(
740 index,
741 RawToolUse {
742 id,
743 name,
744 input_json: String::new(),
745 },
746 );
747 Vec::new()
748 }
749 },
750 Event::ContentBlockDelta { index, delta } => match delta {
751 ContentDelta::TextDelta { text } => {
752 vec![Ok(LanguageModelCompletionEvent::Text(text))]
753 }
754 ContentDelta::ThinkingDelta { thinking } => {
755 vec![Ok(LanguageModelCompletionEvent::Thinking {
756 text: thinking,
757 signature: None,
758 })]
759 }
760 ContentDelta::SignatureDelta { signature } => {
761 vec![Ok(LanguageModelCompletionEvent::Thinking {
762 text: "".to_string(),
763 signature: Some(signature),
764 })]
765 }
766 ContentDelta::InputJsonDelta { partial_json } => {
767 if let Some(tool_use) = self.tool_uses_by_index.get_mut(&index) {
768 tool_use.input_json.push_str(&partial_json);
769
770 // Try to convert invalid (incomplete) JSON into
771 // valid JSON that serde can accept, e.g. by closing
772 // unclosed delimiters. This way, we can update the
773 // UI with whatever has been streamed back so far.
774 if let Ok(input) = serde_json::Value::from_str(
775 &partial_json_fixer::fix_json(&tool_use.input_json),
776 ) {
777 return vec![Ok(LanguageModelCompletionEvent::ToolUse(
778 LanguageModelToolUse {
779 id: tool_use.id.clone().into(),
780 name: tool_use.name.clone().into(),
781 is_input_complete: false,
782 raw_input: tool_use.input_json.clone(),
783 input,
784 },
785 ))];
786 }
787 }
788 return vec![];
789 }
790 },
791 Event::ContentBlockStop { index } => {
792 if let Some(tool_use) = self.tool_uses_by_index.remove(&index) {
793 let input_json = tool_use.input_json.trim();
794 let input_value = if input_json.is_empty() {
795 Ok(serde_json::Value::Object(serde_json::Map::default()))
796 } else {
797 serde_json::Value::from_str(input_json)
798 };
799 let event_result = match input_value {
800 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
801 LanguageModelToolUse {
802 id: tool_use.id.into(),
803 name: tool_use.name.into(),
804 is_input_complete: true,
805 input,
806 raw_input: tool_use.input_json.clone(),
807 },
808 )),
809 Err(json_parse_err) => Err(LanguageModelCompletionError::BadInputJson {
810 id: tool_use.id.into(),
811 tool_name: tool_use.name.into(),
812 raw_input: input_json.into(),
813 json_parse_error: json_parse_err.to_string(),
814 }),
815 };
816
817 vec![event_result]
818 } else {
819 Vec::new()
820 }
821 }
822 Event::MessageStart { message } => {
823 update_usage(&mut self.usage, &message.usage);
824 vec![
825 Ok(LanguageModelCompletionEvent::UsageUpdate(convert_usage(
826 &self.usage,
827 ))),
828 Ok(LanguageModelCompletionEvent::StartMessage {
829 message_id: message.id,
830 }),
831 ]
832 }
833 Event::MessageDelta { delta, usage } => {
834 update_usage(&mut self.usage, &usage);
835 if let Some(stop_reason) = delta.stop_reason.as_deref() {
836 self.stop_reason = match stop_reason {
837 "end_turn" => StopReason::EndTurn,
838 "max_tokens" => StopReason::MaxTokens,
839 "tool_use" => StopReason::ToolUse,
840 "refusal" => StopReason::Refusal,
841 _ => {
842 log::error!("Unexpected anthropic stop_reason: {stop_reason}");
843 StopReason::EndTurn
844 }
845 };
846 }
847 vec![Ok(LanguageModelCompletionEvent::UsageUpdate(
848 convert_usage(&self.usage),
849 ))]
850 }
851 Event::MessageStop => {
852 vec![Ok(LanguageModelCompletionEvent::Stop(self.stop_reason))]
853 }
854 Event::Error { error } => {
855 vec![Err(error.into())]
856 }
857 _ => Vec::new(),
858 }
859 }
860}
861
862struct RawToolUse {
863 id: String,
864 name: String,
865 input_json: String,
866}
867
868/// Updates usage data by preferring counts from `new`.
869fn update_usage(usage: &mut Usage, new: &Usage) {
870 if let Some(input_tokens) = new.input_tokens {
871 usage.input_tokens = Some(input_tokens);
872 }
873 if let Some(output_tokens) = new.output_tokens {
874 usage.output_tokens = Some(output_tokens);
875 }
876 if let Some(cache_creation_input_tokens) = new.cache_creation_input_tokens {
877 usage.cache_creation_input_tokens = Some(cache_creation_input_tokens);
878 }
879 if let Some(cache_read_input_tokens) = new.cache_read_input_tokens {
880 usage.cache_read_input_tokens = Some(cache_read_input_tokens);
881 }
882}
883
884fn convert_usage(usage: &Usage) -> language_model::TokenUsage {
885 language_model::TokenUsage {
886 input_tokens: usage.input_tokens.unwrap_or(0),
887 output_tokens: usage.output_tokens.unwrap_or(0),
888 cache_creation_input_tokens: usage.cache_creation_input_tokens.unwrap_or(0),
889 cache_read_input_tokens: usage.cache_read_input_tokens.unwrap_or(0),
890 }
891}
892
893struct ConfigurationView {
894 api_key_editor: Entity<Editor>,
895 state: gpui::Entity<State>,
896 load_credentials_task: Option<Task<()>>,
897}
898
899impl ConfigurationView {
900 const PLACEHOLDER_TEXT: &'static str = "sk-ant-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
901
902 fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
903 cx.observe(&state, |_, _, cx| {
904 cx.notify();
905 })
906 .detach();
907
908 let load_credentials_task = Some(cx.spawn({
909 let state = state.clone();
910 async move |this, cx| {
911 if let Some(task) = state
912 .update(cx, |state, cx| state.authenticate(cx))
913 .log_err()
914 {
915 // We don't log an error, because "not signed in" is also an error.
916 let _ = task.await;
917 }
918 this.update(cx, |this, cx| {
919 this.load_credentials_task = None;
920 cx.notify();
921 })
922 .log_err();
923 }
924 }));
925
926 Self {
927 api_key_editor: cx.new(|cx| {
928 let mut editor = Editor::single_line(window, cx);
929 editor.set_placeholder_text(Self::PLACEHOLDER_TEXT, cx);
930 editor
931 }),
932 state,
933 load_credentials_task,
934 }
935 }
936
937 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
938 let api_key = self.api_key_editor.read(cx).text(cx);
939 if api_key.is_empty() {
940 return;
941 }
942
943 let state = self.state.clone();
944 cx.spawn_in(window, async move |_, cx| {
945 state
946 .update(cx, |state, cx| state.set_api_key(api_key, cx))?
947 .await
948 })
949 .detach_and_log_err(cx);
950
951 cx.notify();
952 }
953
954 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
955 self.api_key_editor
956 .update(cx, |editor, cx| editor.set_text("", window, cx));
957
958 let state = self.state.clone();
959 cx.spawn_in(window, async move |_, cx| {
960 state.update(cx, |state, cx| state.reset_api_key(cx))?.await
961 })
962 .detach_and_log_err(cx);
963
964 cx.notify();
965 }
966
967 fn render_api_key_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
968 let settings = ThemeSettings::get_global(cx);
969 let text_style = TextStyle {
970 color: cx.theme().colors().text,
971 font_family: settings.ui_font.family.clone(),
972 font_features: settings.ui_font.features.clone(),
973 font_fallbacks: settings.ui_font.fallbacks.clone(),
974 font_size: rems(0.875).into(),
975 font_weight: settings.ui_font.weight,
976 font_style: FontStyle::Normal,
977 line_height: relative(1.3),
978 white_space: WhiteSpace::Normal,
979 ..Default::default()
980 };
981 EditorElement::new(
982 &self.api_key_editor,
983 EditorStyle {
984 background: cx.theme().colors().editor_background,
985 local_player: cx.theme().players().local(),
986 text: text_style,
987 ..Default::default()
988 },
989 )
990 }
991
992 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
993 !self.state.read(cx).is_authenticated()
994 }
995}
996
997impl Render for ConfigurationView {
998 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
999 let env_var_set = self.state.read(cx).api_key_from_env;
1000
1001 if self.load_credentials_task.is_some() {
1002 div().child(Label::new("Loading credentials...")).into_any()
1003 } else if self.should_render_editor(cx) {
1004 v_flex()
1005 .size_full()
1006 .on_action(cx.listener(Self::save_api_key))
1007 .child(Label::new("To use Zed's assistant with Anthropic, you need to add an API key. Follow these steps:"))
1008 .child(
1009 List::new()
1010 .child(
1011 InstructionListItem::new(
1012 "Create one by visiting",
1013 Some("Anthropic's settings"),
1014 Some("https://console.anthropic.com/settings/keys")
1015 )
1016 )
1017 .child(
1018 InstructionListItem::text_only("Paste your API key below and hit enter to start using the assistant")
1019 )
1020 )
1021 .child(
1022 h_flex()
1023 .w_full()
1024 .my_2()
1025 .px_2()
1026 .py_1()
1027 .bg(cx.theme().colors().editor_background)
1028 .border_1()
1029 .border_color(cx.theme().colors().border)
1030 .rounded_sm()
1031 .child(self.render_api_key_editor(cx)),
1032 )
1033 .child(
1034 Label::new(
1035 format!("You can also assign the {ANTHROPIC_API_KEY_VAR} environment variable and restart Zed."),
1036 )
1037 .size(LabelSize::Small)
1038 .color(Color::Muted),
1039 )
1040 .into_any()
1041 } else {
1042 h_flex()
1043 .mt_1()
1044 .p_1()
1045 .justify_between()
1046 .rounded_md()
1047 .border_1()
1048 .border_color(cx.theme().colors().border)
1049 .bg(cx.theme().colors().background)
1050 .child(
1051 h_flex()
1052 .gap_1()
1053 .child(Icon::new(IconName::Check).color(Color::Success))
1054 .child(Label::new(if env_var_set {
1055 format!("API key set in {ANTHROPIC_API_KEY_VAR} environment variable.")
1056 } else {
1057 "API key configured.".to_string()
1058 })),
1059 )
1060 .child(
1061 Button::new("reset-key", "Reset Key")
1062 .label_size(LabelSize::Small)
1063 .icon(Some(IconName::Trash))
1064 .icon_size(IconSize::Small)
1065 .icon_position(IconPosition::Start)
1066 .disabled(env_var_set)
1067 .when(env_var_set, |this| {
1068 this.tooltip(Tooltip::text(format!("To reset your API key, unset the {ANTHROPIC_API_KEY_VAR} environment variable.")))
1069 })
1070 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx))),
1071 )
1072 .into_any()
1073 }
1074 }
1075}
1076
1077#[cfg(test)]
1078mod tests {
1079 use super::*;
1080 use anthropic::AnthropicModelMode;
1081 use language_model::{LanguageModelRequestMessage, MessageContent};
1082
1083 #[test]
1084 fn test_cache_control_only_on_last_segment() {
1085 let request = LanguageModelRequest {
1086 messages: vec![LanguageModelRequestMessage {
1087 role: Role::User,
1088 content: vec![
1089 MessageContent::Text("Some prompt".to_string()),
1090 MessageContent::Image(language_model::LanguageModelImage::empty()),
1091 MessageContent::Image(language_model::LanguageModelImage::empty()),
1092 MessageContent::Image(language_model::LanguageModelImage::empty()),
1093 MessageContent::Image(language_model::LanguageModelImage::empty()),
1094 ],
1095 cache: true,
1096 }],
1097 thread_id: None,
1098 prompt_id: None,
1099 intent: None,
1100 mode: None,
1101 stop: vec![],
1102 temperature: None,
1103 tools: vec![],
1104 tool_choice: None,
1105 };
1106
1107 let anthropic_request = into_anthropic(
1108 request,
1109 "claude-3-5-sonnet".to_string(),
1110 0.7,
1111 4096,
1112 AnthropicModelMode::Default,
1113 );
1114
1115 assert_eq!(anthropic_request.messages.len(), 1);
1116
1117 let message = &anthropic_request.messages[0];
1118 assert_eq!(message.content.len(), 5);
1119
1120 assert!(matches!(
1121 message.content[0],
1122 anthropic::RequestContent::Text {
1123 cache_control: None,
1124 ..
1125 }
1126 ));
1127 for i in 1..3 {
1128 assert!(matches!(
1129 message.content[i],
1130 anthropic::RequestContent::Image {
1131 cache_control: None,
1132 ..
1133 }
1134 ));
1135 }
1136
1137 assert!(matches!(
1138 message.content[4],
1139 anthropic::RequestContent::Image {
1140 cache_control: Some(anthropic::CacheControl {
1141 cache_type: anthropic::CacheControlType::Ephemeral,
1142 }),
1143 ..
1144 }
1145 ));
1146 }
1147}