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