1use anthropic::{
2 ANTHROPIC_API_URL, AnthropicError, AnthropicModelMode, ContentDelta, CountTokensRequest, Event,
3 ResponseContent, ToolResultContent, ToolResultPart, Usage,
4};
5use anyhow::Result;
6use collections::{BTreeMap, HashMap};
7use futures::{FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream};
8use gpui::{AnyView, App, AsyncApp, Context, Entity, Task};
9use http_client::HttpClient;
10use language_model::{
11 ApiKeyState, AuthenticateError, ConfigurationViewTargetAgent, EnvVar, IconOrSvg, LanguageModel,
12 LanguageModelCacheConfiguration, LanguageModelCompletionError, LanguageModelCompletionEvent,
13 LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
14 LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,
15 LanguageModelToolChoice, LanguageModelToolResultContent, LanguageModelToolUse, MessageContent,
16 RateLimiter, Role, StopReason, env_var,
17};
18use settings::{Settings, SettingsStore};
19use std::pin::Pin;
20use std::str::FromStr;
21use std::sync::{Arc, LazyLock};
22use strum::IntoEnumIterator;
23use ui::{ButtonLink, ConfiguredApiCard, List, ListBulletItem, prelude::*};
24use ui_input::InputField;
25use util::ResultExt;
26
27use crate::provider::util::parse_tool_arguments;
28
29pub use settings::AnthropicAvailableModel as AvailableModel;
30
31const PROVIDER_ID: LanguageModelProviderId = language_model::ANTHROPIC_PROVIDER_ID;
32const PROVIDER_NAME: LanguageModelProviderName = language_model::ANTHROPIC_PROVIDER_NAME;
33
34#[derive(Default, Clone, Debug, PartialEq)]
35pub struct AnthropicSettings {
36 pub api_url: String,
37 /// Extend Zed's list of Anthropic models.
38 pub available_models: Vec<AvailableModel>,
39}
40
41pub struct AnthropicLanguageModelProvider {
42 http_client: Arc<dyn HttpClient>,
43 state: Entity<State>,
44}
45
46const API_KEY_ENV_VAR_NAME: &str = "ANTHROPIC_API_KEY";
47static API_KEY_ENV_VAR: LazyLock<EnvVar> = env_var!(API_KEY_ENV_VAR_NAME);
48
49pub struct State {
50 api_key_state: ApiKeyState,
51}
52
53impl State {
54 fn is_authenticated(&self) -> bool {
55 self.api_key_state.has_key()
56 }
57
58 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
59 let api_url = AnthropicLanguageModelProvider::api_url(cx);
60 self.api_key_state
61 .store(api_url, api_key, |this| &mut this.api_key_state, cx)
62 }
63
64 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
65 let api_url = AnthropicLanguageModelProvider::api_url(cx);
66 self.api_key_state
67 .load_if_needed(api_url, |this| &mut this.api_key_state, cx)
68 }
69}
70
71impl AnthropicLanguageModelProvider {
72 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
73 let state = cx.new(|cx| {
74 cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
75 let api_url = Self::api_url(cx);
76 this.api_key_state
77 .handle_url_change(api_url, |this| &mut this.api_key_state, cx);
78 cx.notify();
79 })
80 .detach();
81 State {
82 api_key_state: ApiKeyState::new(Self::api_url(cx), (*API_KEY_ENV_VAR).clone()),
83 }
84 });
85
86 Self { http_client, state }
87 }
88
89 fn create_language_model(&self, model: anthropic::Model) -> Arc<dyn LanguageModel> {
90 Arc::new(AnthropicModel {
91 id: LanguageModelId::from(model.id().to_string()),
92 model,
93 state: self.state.clone(),
94 http_client: self.http_client.clone(),
95 request_limiter: RateLimiter::new(4),
96 })
97 }
98
99 fn settings(cx: &App) -> &AnthropicSettings {
100 &crate::AllLanguageModelSettings::get_global(cx).anthropic
101 }
102
103 fn api_url(cx: &App) -> SharedString {
104 let api_url = &Self::settings(cx).api_url;
105 if api_url.is_empty() {
106 ANTHROPIC_API_URL.into()
107 } else {
108 SharedString::new(api_url.as_str())
109 }
110 }
111}
112
113impl LanguageModelProviderState for AnthropicLanguageModelProvider {
114 type ObservableEntity = State;
115
116 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
117 Some(self.state.clone())
118 }
119}
120
121impl LanguageModelProvider for AnthropicLanguageModelProvider {
122 fn id(&self) -> LanguageModelProviderId {
123 PROVIDER_ID
124 }
125
126 fn name(&self) -> LanguageModelProviderName {
127 PROVIDER_NAME
128 }
129
130 fn icon(&self) -> IconOrSvg {
131 IconOrSvg::Icon(IconName::AiAnthropic)
132 }
133
134 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
135 Some(self.create_language_model(anthropic::Model::default()))
136 }
137
138 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
139 Some(self.create_language_model(anthropic::Model::default_fast()))
140 }
141
142 fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {
143 [
144 anthropic::Model::ClaudeSonnet4_5,
145 anthropic::Model::ClaudeSonnet4_5Thinking,
146 ]
147 .into_iter()
148 .map(|model| self.create_language_model(model))
149 .collect()
150 }
151
152 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
153 let mut models = BTreeMap::default();
154
155 // Add base models from anthropic::Model::iter()
156 for model in anthropic::Model::iter() {
157 if !matches!(model, anthropic::Model::Custom { .. }) {
158 models.insert(model.id().to_string(), model);
159 }
160 }
161
162 // Override with available models from settings
163 for model in &AnthropicLanguageModelProvider::settings(cx).available_models {
164 models.insert(
165 model.name.clone(),
166 anthropic::Model::Custom {
167 name: model.name.clone(),
168 display_name: model.display_name.clone(),
169 max_tokens: model.max_tokens,
170 tool_override: model.tool_override.clone(),
171 cache_configuration: model.cache_configuration.as_ref().map(|config| {
172 anthropic::AnthropicModelCacheConfiguration {
173 max_cache_anchors: config.max_cache_anchors,
174 should_speculate: config.should_speculate,
175 min_total_token: config.min_total_token,
176 }
177 }),
178 max_output_tokens: model.max_output_tokens,
179 default_temperature: model.default_temperature,
180 extra_beta_headers: model.extra_beta_headers.clone(),
181 mode: model.mode.unwrap_or_default().into(),
182 },
183 );
184 }
185
186 models
187 .into_values()
188 .map(|model| self.create_language_model(model))
189 .collect()
190 }
191
192 fn is_authenticated(&self, cx: &App) -> bool {
193 self.state.read(cx).is_authenticated()
194 }
195
196 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
197 self.state.update(cx, |state, cx| state.authenticate(cx))
198 }
199
200 fn configuration_view(
201 &self,
202 target_agent: ConfigurationViewTargetAgent,
203 window: &mut Window,
204 cx: &mut App,
205 ) -> AnyView {
206 cx.new(|cx| ConfigurationView::new(self.state.clone(), target_agent, window, cx))
207 .into()
208 }
209
210 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
211 self.state
212 .update(cx, |state, cx| state.set_api_key(None, cx))
213 }
214}
215
216pub struct AnthropicModel {
217 id: LanguageModelId,
218 model: anthropic::Model,
219 state: Entity<State>,
220 http_client: Arc<dyn HttpClient>,
221 request_limiter: RateLimiter,
222}
223
224fn to_anthropic_content(content: MessageContent) -> Option<anthropic::RequestContent> {
225 match content {
226 MessageContent::Text(text) => {
227 let text = if text.chars().last().is_some_and(|c| c.is_whitespace()) {
228 text.trim_end().to_string()
229 } else {
230 text
231 };
232 if !text.is_empty() {
233 Some(anthropic::RequestContent::Text {
234 text,
235 cache_control: None,
236 })
237 } else {
238 None
239 }
240 }
241 MessageContent::Thinking {
242 text: thinking,
243 signature,
244 } => {
245 if let Some(signature) = signature
246 && !thinking.is_empty()
247 {
248 Some(anthropic::RequestContent::Thinking {
249 thinking,
250 signature,
251 cache_control: None,
252 })
253 } else {
254 None
255 }
256 }
257 MessageContent::RedactedThinking(data) => {
258 if !data.is_empty() {
259 Some(anthropic::RequestContent::RedactedThinking { data })
260 } else {
261 None
262 }
263 }
264 MessageContent::Image(image) => Some(anthropic::RequestContent::Image {
265 source: anthropic::ImageSource {
266 source_type: "base64".to_string(),
267 media_type: "image/png".to_string(),
268 data: image.source.to_string(),
269 },
270 cache_control: None,
271 }),
272 MessageContent::ToolUse(tool_use) => Some(anthropic::RequestContent::ToolUse {
273 id: tool_use.id.to_string(),
274 name: tool_use.name.to_string(),
275 input: tool_use.input,
276 cache_control: None,
277 }),
278 MessageContent::ToolResult(tool_result) => Some(anthropic::RequestContent::ToolResult {
279 tool_use_id: tool_result.tool_use_id.to_string(),
280 is_error: tool_result.is_error,
281 content: match tool_result.content {
282 LanguageModelToolResultContent::Text(text) => {
283 ToolResultContent::Plain(text.to_string())
284 }
285 LanguageModelToolResultContent::Image(image) => {
286 ToolResultContent::Multipart(vec![ToolResultPart::Image {
287 source: anthropic::ImageSource {
288 source_type: "base64".to_string(),
289 media_type: "image/png".to_string(),
290 data: image.source.to_string(),
291 },
292 }])
293 }
294 },
295 cache_control: None,
296 }),
297 }
298}
299
300/// Convert a LanguageModelRequest to an Anthropic CountTokensRequest.
301pub fn into_anthropic_count_tokens_request(
302 request: LanguageModelRequest,
303 model: String,
304 mode: AnthropicModelMode,
305) -> CountTokensRequest {
306 let mut new_messages: Vec<anthropic::Message> = Vec::new();
307 let mut system_message = String::new();
308
309 for message in request.messages {
310 if message.contents_empty() {
311 continue;
312 }
313
314 match message.role {
315 Role::User | Role::Assistant => {
316 let anthropic_message_content: Vec<anthropic::RequestContent> = message
317 .content
318 .into_iter()
319 .filter_map(to_anthropic_content)
320 .collect();
321 let anthropic_role = match message.role {
322 Role::User => anthropic::Role::User,
323 Role::Assistant => anthropic::Role::Assistant,
324 Role::System => unreachable!("System role should never occur here"),
325 };
326 if anthropic_message_content.is_empty() {
327 continue;
328 }
329
330 if let Some(last_message) = new_messages.last_mut()
331 && last_message.role == anthropic_role
332 {
333 last_message.content.extend(anthropic_message_content);
334 continue;
335 }
336
337 new_messages.push(anthropic::Message {
338 role: anthropic_role,
339 content: anthropic_message_content,
340 });
341 }
342 Role::System => {
343 if !system_message.is_empty() {
344 system_message.push_str("\n\n");
345 }
346 system_message.push_str(&message.string_contents());
347 }
348 }
349 }
350
351 CountTokensRequest {
352 model,
353 messages: new_messages,
354 system: if system_message.is_empty() {
355 None
356 } else {
357 Some(anthropic::StringOrContents::String(system_message))
358 },
359 thinking: if request.thinking_allowed
360 && let AnthropicModelMode::Thinking { budget_tokens } = mode
361 {
362 Some(anthropic::Thinking::Enabled { budget_tokens })
363 } else {
364 None
365 },
366 tools: request
367 .tools
368 .into_iter()
369 .map(|tool| anthropic::Tool {
370 name: tool.name,
371 description: tool.description,
372 input_schema: tool.input_schema,
373 })
374 .collect(),
375 tool_choice: request.tool_choice.map(|choice| match choice {
376 LanguageModelToolChoice::Auto => anthropic::ToolChoice::Auto,
377 LanguageModelToolChoice::Any => anthropic::ToolChoice::Any,
378 LanguageModelToolChoice::None => anthropic::ToolChoice::None,
379 }),
380 }
381}
382
383/// Estimate tokens using tiktoken. Used as a fallback when the API is unavailable,
384/// or by providers (like Zed Cloud) that don't have direct Anthropic API access.
385pub fn count_anthropic_tokens_with_tiktoken(request: LanguageModelRequest) -> Result<u64> {
386 let messages = request.messages;
387 let mut tokens_from_images = 0;
388 let mut string_messages = Vec::with_capacity(messages.len());
389
390 for message in messages {
391 let mut string_contents = String::new();
392
393 for content in message.content {
394 match content {
395 MessageContent::Text(text) => {
396 string_contents.push_str(&text);
397 }
398 MessageContent::Thinking { .. } => {
399 // Thinking blocks are not included in the input token count.
400 }
401 MessageContent::RedactedThinking(_) => {
402 // Thinking blocks are not included in the input token count.
403 }
404 MessageContent::Image(image) => {
405 tokens_from_images += image.estimate_tokens();
406 }
407 MessageContent::ToolUse(_tool_use) => {
408 // TODO: Estimate token usage from tool uses.
409 }
410 MessageContent::ToolResult(tool_result) => match &tool_result.content {
411 LanguageModelToolResultContent::Text(text) => {
412 string_contents.push_str(text);
413 }
414 LanguageModelToolResultContent::Image(image) => {
415 tokens_from_images += image.estimate_tokens();
416 }
417 },
418 }
419 }
420
421 if !string_contents.is_empty() {
422 string_messages.push(tiktoken_rs::ChatCompletionRequestMessage {
423 role: match message.role {
424 Role::User => "user".into(),
425 Role::Assistant => "assistant".into(),
426 Role::System => "system".into(),
427 },
428 content: Some(string_contents),
429 name: None,
430 function_call: None,
431 });
432 }
433 }
434
435 // Tiktoken doesn't yet support these models, so we manually use the
436 // same tokenizer as GPT-4.
437 tiktoken_rs::num_tokens_from_messages("gpt-4", &string_messages)
438 .map(|tokens| (tokens + tokens_from_images) as u64)
439}
440
441impl AnthropicModel {
442 fn stream_completion(
443 &self,
444 request: anthropic::Request,
445 cx: &AsyncApp,
446 ) -> BoxFuture<
447 'static,
448 Result<
449 BoxStream<'static, Result<anthropic::Event, AnthropicError>>,
450 LanguageModelCompletionError,
451 >,
452 > {
453 let http_client = self.http_client.clone();
454
455 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
456 let api_url = AnthropicLanguageModelProvider::api_url(cx);
457 (state.api_key_state.key(&api_url), api_url)
458 });
459
460 let beta_headers = self.model.beta_headers();
461
462 async move {
463 let Some(api_key) = api_key else {
464 return Err(LanguageModelCompletionError::NoApiKey {
465 provider: PROVIDER_NAME,
466 });
467 };
468 let request = anthropic::stream_completion(
469 http_client.as_ref(),
470 &api_url,
471 &api_key,
472 request,
473 beta_headers,
474 );
475 request.await.map_err(Into::into)
476 }
477 .boxed()
478 }
479}
480
481impl LanguageModel for AnthropicModel {
482 fn id(&self) -> LanguageModelId {
483 self.id.clone()
484 }
485
486 fn name(&self) -> LanguageModelName {
487 LanguageModelName::from(self.model.display_name().to_string())
488 }
489
490 fn provider_id(&self) -> LanguageModelProviderId {
491 PROVIDER_ID
492 }
493
494 fn provider_name(&self) -> LanguageModelProviderName {
495 PROVIDER_NAME
496 }
497
498 fn supports_tools(&self) -> bool {
499 true
500 }
501
502 fn supports_images(&self) -> bool {
503 true
504 }
505
506 fn supports_streaming_tools(&self) -> bool {
507 true
508 }
509
510 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
511 match choice {
512 LanguageModelToolChoice::Auto
513 | LanguageModelToolChoice::Any
514 | LanguageModelToolChoice::None => true,
515 }
516 }
517
518 fn telemetry_id(&self) -> String {
519 format!("anthropic/{}", self.model.id())
520 }
521
522 fn api_key(&self, cx: &App) -> Option<String> {
523 self.state.read_with(cx, |state, cx| {
524 let api_url = AnthropicLanguageModelProvider::api_url(cx);
525 state.api_key_state.key(&api_url).map(|key| key.to_string())
526 })
527 }
528
529 fn max_token_count(&self) -> u64 {
530 self.model.max_token_count()
531 }
532
533 fn max_output_tokens(&self) -> Option<u64> {
534 Some(self.model.max_output_tokens())
535 }
536
537 fn count_tokens(
538 &self,
539 request: LanguageModelRequest,
540 cx: &App,
541 ) -> BoxFuture<'static, Result<u64>> {
542 let http_client = self.http_client.clone();
543 let model_id = self.model.request_id().to_string();
544 let mode = self.model.mode();
545
546 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
547 let api_url = AnthropicLanguageModelProvider::api_url(cx);
548 (
549 state.api_key_state.key(&api_url).map(|k| k.to_string()),
550 api_url.to_string(),
551 )
552 });
553
554 async move {
555 // If no API key, fall back to tiktoken estimation
556 let Some(api_key) = api_key else {
557 return count_anthropic_tokens_with_tiktoken(request);
558 };
559
560 let count_request =
561 into_anthropic_count_tokens_request(request.clone(), model_id, mode);
562
563 match anthropic::count_tokens(http_client.as_ref(), &api_url, &api_key, count_request)
564 .await
565 {
566 Ok(response) => Ok(response.input_tokens),
567 Err(err) => {
568 log::error!(
569 "Anthropic count_tokens API failed, falling back to tiktoken: {err:?}"
570 );
571 count_anthropic_tokens_with_tiktoken(request)
572 }
573 }
574 }
575 .boxed()
576 }
577
578 fn stream_completion(
579 &self,
580 request: LanguageModelRequest,
581 cx: &AsyncApp,
582 ) -> BoxFuture<
583 'static,
584 Result<
585 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
586 LanguageModelCompletionError,
587 >,
588 > {
589 let request = into_anthropic(
590 request,
591 self.model.request_id().into(),
592 self.model.default_temperature(),
593 self.model.max_output_tokens(),
594 self.model.mode(),
595 );
596 let request = self.stream_completion(request, cx);
597 let future = self.request_limiter.stream(async move {
598 let response = request.await?;
599 Ok(AnthropicEventMapper::new().map_stream(response))
600 });
601 async move { Ok(future.await?.boxed()) }.boxed()
602 }
603
604 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
605 self.model
606 .cache_configuration()
607 .map(|config| LanguageModelCacheConfiguration {
608 max_cache_anchors: config.max_cache_anchors,
609 should_speculate: config.should_speculate,
610 min_total_token: config.min_total_token,
611 })
612 }
613}
614
615pub fn into_anthropic(
616 request: LanguageModelRequest,
617 model: String,
618 default_temperature: f32,
619 max_output_tokens: u64,
620 mode: AnthropicModelMode,
621) -> anthropic::Request {
622 let mut new_messages: Vec<anthropic::Message> = Vec::new();
623 let mut system_message = String::new();
624
625 for message in request.messages {
626 if message.contents_empty() {
627 continue;
628 }
629
630 match message.role {
631 Role::User | Role::Assistant => {
632 let mut anthropic_message_content: Vec<anthropic::RequestContent> = message
633 .content
634 .into_iter()
635 .filter_map(to_anthropic_content)
636 .collect();
637 let anthropic_role = match message.role {
638 Role::User => anthropic::Role::User,
639 Role::Assistant => anthropic::Role::Assistant,
640 Role::System => unreachable!("System role should never occur here"),
641 };
642 if anthropic_message_content.is_empty() {
643 continue;
644 }
645
646 if let Some(last_message) = new_messages.last_mut()
647 && last_message.role == anthropic_role
648 {
649 last_message.content.extend(anthropic_message_content);
650 continue;
651 }
652
653 // Mark the last segment of the message as cached
654 if message.cache {
655 let cache_control_value = Some(anthropic::CacheControl {
656 cache_type: anthropic::CacheControlType::Ephemeral,
657 });
658 for message_content in anthropic_message_content.iter_mut().rev() {
659 match message_content {
660 anthropic::RequestContent::RedactedThinking { .. } => {
661 // Caching is not possible, fallback to next message
662 }
663 anthropic::RequestContent::Text { cache_control, .. }
664 | anthropic::RequestContent::Thinking { cache_control, .. }
665 | anthropic::RequestContent::Image { cache_control, .. }
666 | anthropic::RequestContent::ToolUse { cache_control, .. }
667 | anthropic::RequestContent::ToolResult { cache_control, .. } => {
668 *cache_control = cache_control_value;
669 break;
670 }
671 }
672 }
673 }
674
675 new_messages.push(anthropic::Message {
676 role: anthropic_role,
677 content: anthropic_message_content,
678 });
679 }
680 Role::System => {
681 if !system_message.is_empty() {
682 system_message.push_str("\n\n");
683 }
684 system_message.push_str(&message.string_contents());
685 }
686 }
687 }
688
689 anthropic::Request {
690 model,
691 messages: new_messages,
692 max_tokens: max_output_tokens,
693 system: if system_message.is_empty() {
694 None
695 } else {
696 Some(anthropic::StringOrContents::String(system_message))
697 },
698 thinking: if request.thinking_allowed
699 && let AnthropicModelMode::Thinking { budget_tokens } = mode
700 {
701 Some(anthropic::Thinking::Enabled { budget_tokens })
702 } else {
703 None
704 },
705 tools: request
706 .tools
707 .into_iter()
708 .map(|tool| anthropic::Tool {
709 name: tool.name,
710 description: tool.description,
711 input_schema: tool.input_schema,
712 })
713 .collect(),
714 tool_choice: request.tool_choice.map(|choice| match choice {
715 LanguageModelToolChoice::Auto => anthropic::ToolChoice::Auto,
716 LanguageModelToolChoice::Any => anthropic::ToolChoice::Any,
717 LanguageModelToolChoice::None => anthropic::ToolChoice::None,
718 }),
719 metadata: None,
720 output_config: None,
721 stop_sequences: Vec::new(),
722 temperature: request.temperature.or(Some(default_temperature)),
723 top_k: None,
724 top_p: None,
725 }
726}
727
728pub struct AnthropicEventMapper {
729 tool_uses_by_index: HashMap<usize, RawToolUse>,
730 usage: Usage,
731 stop_reason: StopReason,
732}
733
734impl AnthropicEventMapper {
735 pub fn new() -> Self {
736 Self {
737 tool_uses_by_index: HashMap::default(),
738 usage: Usage::default(),
739 stop_reason: StopReason::EndTurn,
740 }
741 }
742
743 pub fn map_stream(
744 mut self,
745 events: Pin<Box<dyn Send + Stream<Item = Result<Event, AnthropicError>>>>,
746 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
747 {
748 events.flat_map(move |event| {
749 futures::stream::iter(match event {
750 Ok(event) => self.map_event(event),
751 Err(error) => vec![Err(error.into())],
752 })
753 })
754 }
755
756 pub fn map_event(
757 &mut self,
758 event: Event,
759 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
760 match event {
761 Event::ContentBlockStart {
762 index,
763 content_block,
764 } => match content_block {
765 ResponseContent::Text { text } => {
766 vec![Ok(LanguageModelCompletionEvent::Text(text))]
767 }
768 ResponseContent::Thinking { thinking } => {
769 vec![Ok(LanguageModelCompletionEvent::Thinking {
770 text: thinking,
771 signature: None,
772 })]
773 }
774 ResponseContent::RedactedThinking { data } => {
775 vec![Ok(LanguageModelCompletionEvent::RedactedThinking { data })]
776 }
777 ResponseContent::ToolUse { id, name, .. } => {
778 self.tool_uses_by_index.insert(
779 index,
780 RawToolUse {
781 id,
782 name,
783 input_json: String::new(),
784 },
785 );
786 Vec::new()
787 }
788 },
789 Event::ContentBlockDelta { index, delta } => match delta {
790 ContentDelta::TextDelta { text } => {
791 vec![Ok(LanguageModelCompletionEvent::Text(text))]
792 }
793 ContentDelta::ThinkingDelta { thinking } => {
794 vec![Ok(LanguageModelCompletionEvent::Thinking {
795 text: thinking,
796 signature: None,
797 })]
798 }
799 ContentDelta::SignatureDelta { signature } => {
800 vec![Ok(LanguageModelCompletionEvent::Thinking {
801 text: "".to_string(),
802 signature: Some(signature),
803 })]
804 }
805 ContentDelta::InputJsonDelta { partial_json } => {
806 if let Some(tool_use) = self.tool_uses_by_index.get_mut(&index) {
807 tool_use.input_json.push_str(&partial_json);
808
809 // Try to convert invalid (incomplete) JSON into
810 // valid JSON that serde can accept, e.g. by closing
811 // unclosed delimiters. This way, we can update the
812 // UI with whatever has been streamed back so far.
813 if let Ok(input) = serde_json::Value::from_str(
814 &partial_json_fixer::fix_json(&tool_use.input_json),
815 ) {
816 return vec![Ok(LanguageModelCompletionEvent::ToolUse(
817 LanguageModelToolUse {
818 id: tool_use.id.clone().into(),
819 name: tool_use.name.clone().into(),
820 is_input_complete: false,
821 raw_input: tool_use.input_json.clone(),
822 input,
823 thought_signature: None,
824 },
825 ))];
826 }
827 }
828 vec![]
829 }
830 },
831 Event::ContentBlockStop { index } => {
832 if let Some(tool_use) = self.tool_uses_by_index.remove(&index) {
833 let input_json = tool_use.input_json.trim();
834 let event_result = match parse_tool_arguments(input_json) {
835 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
836 LanguageModelToolUse {
837 id: tool_use.id.into(),
838 name: tool_use.name.into(),
839 is_input_complete: true,
840 input,
841 raw_input: tool_use.input_json.clone(),
842 thought_signature: None,
843 },
844 )),
845 Err(json_parse_err) => {
846 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
847 id: tool_use.id.into(),
848 tool_name: tool_use.name.into(),
849 raw_input: input_json.into(),
850 json_parse_error: json_parse_err.to_string(),
851 })
852 }
853 };
854
855 vec![event_result]
856 } else {
857 Vec::new()
858 }
859 }
860 Event::MessageStart { message } => {
861 update_usage(&mut self.usage, &message.usage);
862 vec![
863 Ok(LanguageModelCompletionEvent::UsageUpdate(convert_usage(
864 &self.usage,
865 ))),
866 Ok(LanguageModelCompletionEvent::StartMessage {
867 message_id: message.id,
868 }),
869 ]
870 }
871 Event::MessageDelta { delta, usage } => {
872 update_usage(&mut self.usage, &usage);
873 if let Some(stop_reason) = delta.stop_reason.as_deref() {
874 self.stop_reason = match stop_reason {
875 "end_turn" => StopReason::EndTurn,
876 "max_tokens" => StopReason::MaxTokens,
877 "tool_use" => StopReason::ToolUse,
878 "refusal" => StopReason::Refusal,
879 _ => {
880 log::error!("Unexpected anthropic stop_reason: {stop_reason}");
881 StopReason::EndTurn
882 }
883 };
884 }
885 vec![Ok(LanguageModelCompletionEvent::UsageUpdate(
886 convert_usage(&self.usage),
887 ))]
888 }
889 Event::MessageStop => {
890 vec![Ok(LanguageModelCompletionEvent::Stop(self.stop_reason))]
891 }
892 Event::Error { error } => {
893 vec![Err(error.into())]
894 }
895 _ => Vec::new(),
896 }
897 }
898}
899
900struct RawToolUse {
901 id: String,
902 name: String,
903 input_json: String,
904}
905
906/// Updates usage data by preferring counts from `new`.
907fn update_usage(usage: &mut Usage, new: &Usage) {
908 if let Some(input_tokens) = new.input_tokens {
909 usage.input_tokens = Some(input_tokens);
910 }
911 if let Some(output_tokens) = new.output_tokens {
912 usage.output_tokens = Some(output_tokens);
913 }
914 if let Some(cache_creation_input_tokens) = new.cache_creation_input_tokens {
915 usage.cache_creation_input_tokens = Some(cache_creation_input_tokens);
916 }
917 if let Some(cache_read_input_tokens) = new.cache_read_input_tokens {
918 usage.cache_read_input_tokens = Some(cache_read_input_tokens);
919 }
920}
921
922fn convert_usage(usage: &Usage) -> language_model::TokenUsage {
923 language_model::TokenUsage {
924 input_tokens: usage.input_tokens.unwrap_or(0),
925 output_tokens: usage.output_tokens.unwrap_or(0),
926 cache_creation_input_tokens: usage.cache_creation_input_tokens.unwrap_or(0),
927 cache_read_input_tokens: usage.cache_read_input_tokens.unwrap_or(0),
928 }
929}
930
931struct ConfigurationView {
932 api_key_editor: Entity<InputField>,
933 state: Entity<State>,
934 load_credentials_task: Option<Task<()>>,
935 target_agent: ConfigurationViewTargetAgent,
936}
937
938impl ConfigurationView {
939 const PLACEHOLDER_TEXT: &'static str = "sk-ant-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
940
941 fn new(
942 state: Entity<State>,
943 target_agent: ConfigurationViewTargetAgent,
944 window: &mut Window,
945 cx: &mut Context<Self>,
946 ) -> Self {
947 cx.observe(&state, |_, _, cx| {
948 cx.notify();
949 })
950 .detach();
951
952 let load_credentials_task = Some(cx.spawn({
953 let state = state.clone();
954 async move |this, cx| {
955 let task = state.update(cx, |state, cx| state.authenticate(cx));
956 // We don't log an error, because "not signed in" is also an error.
957 let _ = task.await;
958 this.update(cx, |this, cx| {
959 this.load_credentials_task = None;
960 cx.notify();
961 })
962 .log_err();
963 }
964 }));
965
966 Self {
967 api_key_editor: cx.new(|cx| InputField::new(window, cx, Self::PLACEHOLDER_TEXT)),
968 state,
969 load_credentials_task,
970 target_agent,
971 }
972 }
973
974 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
975 let api_key = self.api_key_editor.read(cx).text(cx);
976 if api_key.is_empty() {
977 return;
978 }
979
980 // url changes can cause the editor to be displayed again
981 self.api_key_editor
982 .update(cx, |editor, cx| editor.set_text("", window, cx));
983
984 let state = self.state.clone();
985 cx.spawn_in(window, async move |_, cx| {
986 state
987 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
988 .await
989 })
990 .detach_and_log_err(cx);
991 }
992
993 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
994 self.api_key_editor
995 .update(cx, |editor, cx| editor.set_text("", window, cx));
996
997 let state = self.state.clone();
998 cx.spawn_in(window, async move |_, cx| {
999 state
1000 .update(cx, |state, cx| state.set_api_key(None, cx))
1001 .await
1002 })
1003 .detach_and_log_err(cx);
1004 }
1005
1006 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
1007 !self.state.read(cx).is_authenticated()
1008 }
1009}
1010
1011impl Render for ConfigurationView {
1012 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1013 let env_var_set = self.state.read(cx).api_key_state.is_from_env_var();
1014 let configured_card_label = if env_var_set {
1015 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable")
1016 } else {
1017 let api_url = AnthropicLanguageModelProvider::api_url(cx);
1018 if api_url == ANTHROPIC_API_URL {
1019 "API key configured".to_string()
1020 } else {
1021 format!("API key configured for {}", api_url)
1022 }
1023 };
1024
1025 if self.load_credentials_task.is_some() {
1026 div()
1027 .child(Label::new("Loading credentials..."))
1028 .into_any_element()
1029 } else if self.should_render_editor(cx) {
1030 v_flex()
1031 .size_full()
1032 .on_action(cx.listener(Self::save_api_key))
1033 .child(Label::new(format!("To use {}, you need to add an API key. Follow these steps:", match &self.target_agent {
1034 ConfigurationViewTargetAgent::ZedAgent => "Zed's agent with Anthropic".into(),
1035 ConfigurationViewTargetAgent::Other(agent) => agent.clone(),
1036 })))
1037 .child(
1038 List::new()
1039 .child(
1040 ListBulletItem::new("")
1041 .child(Label::new("Create one by visiting"))
1042 .child(ButtonLink::new("Anthropic's settings", "https://console.anthropic.com/settings/keys"))
1043 )
1044 .child(
1045 ListBulletItem::new("Paste your API key below and hit enter to start using the agent")
1046 )
1047 )
1048 .child(self.api_key_editor.clone())
1049 .child(
1050 Label::new(
1051 format!("You can also set the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed."),
1052 )
1053 .size(LabelSize::Small)
1054 .color(Color::Muted)
1055 .mt_0p5(),
1056 )
1057 .into_any_element()
1058 } else {
1059 ConfiguredApiCard::new(configured_card_label)
1060 .disabled(env_var_set)
1061 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
1062 .when(env_var_set, |this| {
1063 this.tooltip_label(format!(
1064 "To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable."
1065 ))
1066 })
1067 .into_any_element()
1068 }
1069 }
1070}
1071
1072#[cfg(test)]
1073mod tests {
1074 use super::*;
1075 use anthropic::AnthropicModelMode;
1076 use language_model::{LanguageModelRequestMessage, MessageContent};
1077
1078 #[test]
1079 fn test_cache_control_only_on_last_segment() {
1080 let request = LanguageModelRequest {
1081 messages: vec![LanguageModelRequestMessage {
1082 role: Role::User,
1083 content: vec![
1084 MessageContent::Text("Some prompt".to_string()),
1085 MessageContent::Image(language_model::LanguageModelImage::empty()),
1086 MessageContent::Image(language_model::LanguageModelImage::empty()),
1087 MessageContent::Image(language_model::LanguageModelImage::empty()),
1088 MessageContent::Image(language_model::LanguageModelImage::empty()),
1089 ],
1090 cache: true,
1091 reasoning_details: None,
1092 }],
1093 thread_id: None,
1094 prompt_id: None,
1095 intent: None,
1096 stop: vec![],
1097 temperature: None,
1098 tools: vec![],
1099 tool_choice: None,
1100 thinking_allowed: true,
1101 thinking_effort: None,
1102 };
1103
1104 let anthropic_request = into_anthropic(
1105 request,
1106 "claude-3-5-sonnet".to_string(),
1107 0.7,
1108 4096,
1109 AnthropicModelMode::Default,
1110 );
1111
1112 assert_eq!(anthropic_request.messages.len(), 1);
1113
1114 let message = &anthropic_request.messages[0];
1115 assert_eq!(message.content.len(), 5);
1116
1117 assert!(matches!(
1118 message.content[0],
1119 anthropic::RequestContent::Text {
1120 cache_control: None,
1121 ..
1122 }
1123 ));
1124 for i in 1..3 {
1125 assert!(matches!(
1126 message.content[i],
1127 anthropic::RequestContent::Image {
1128 cache_control: None,
1129 ..
1130 }
1131 ));
1132 }
1133
1134 assert!(matches!(
1135 message.content[4],
1136 anthropic::RequestContent::Image {
1137 cache_control: Some(anthropic::CacheControl {
1138 cache_type: anthropic::CacheControlType::Ephemeral,
1139 }),
1140 ..
1141 }
1142 ));
1143 }
1144
1145 fn request_with_assistant_content(
1146 assistant_content: Vec<MessageContent>,
1147 ) -> anthropic::Request {
1148 let mut request = LanguageModelRequest {
1149 messages: vec![LanguageModelRequestMessage {
1150 role: Role::User,
1151 content: vec![MessageContent::Text("Hello".to_string())],
1152 cache: false,
1153 reasoning_details: None,
1154 }],
1155 thinking_effort: None,
1156 thread_id: None,
1157 prompt_id: None,
1158 intent: None,
1159 stop: vec![],
1160 temperature: None,
1161 tools: vec![],
1162 tool_choice: None,
1163 thinking_allowed: true,
1164 };
1165 request.messages.push(LanguageModelRequestMessage {
1166 role: Role::Assistant,
1167 content: assistant_content,
1168 cache: false,
1169 reasoning_details: None,
1170 });
1171 into_anthropic(
1172 request,
1173 "claude-sonnet-4-5".to_string(),
1174 1.0,
1175 16000,
1176 AnthropicModelMode::Thinking {
1177 budget_tokens: Some(10000),
1178 },
1179 )
1180 }
1181
1182 #[test]
1183 fn test_unsigned_thinking_blocks_stripped() {
1184 let result = request_with_assistant_content(vec![
1185 MessageContent::Thinking {
1186 text: "Cancelled mid-think, no signature".to_string(),
1187 signature: None,
1188 },
1189 MessageContent::Text("Some response text".to_string()),
1190 ]);
1191
1192 let assistant_message = result
1193 .messages
1194 .iter()
1195 .find(|m| m.role == anthropic::Role::Assistant)
1196 .expect("assistant message should still exist");
1197
1198 assert_eq!(
1199 assistant_message.content.len(),
1200 1,
1201 "Only the text content should remain; unsigned thinking block should be stripped"
1202 );
1203 assert!(matches!(
1204 &assistant_message.content[0],
1205 anthropic::RequestContent::Text { text, .. } if text == "Some response text"
1206 ));
1207 }
1208
1209 #[test]
1210 fn test_signed_thinking_blocks_preserved() {
1211 let result = request_with_assistant_content(vec![
1212 MessageContent::Thinking {
1213 text: "Completed thinking".to_string(),
1214 signature: Some("valid-signature".to_string()),
1215 },
1216 MessageContent::Text("Response".to_string()),
1217 ]);
1218
1219 let assistant_message = result
1220 .messages
1221 .iter()
1222 .find(|m| m.role == anthropic::Role::Assistant)
1223 .expect("assistant message should exist");
1224
1225 assert_eq!(
1226 assistant_message.content.len(),
1227 2,
1228 "Both the signed thinking block and text should be preserved"
1229 );
1230 assert!(matches!(
1231 &assistant_message.content[0],
1232 anthropic::RequestContent::Thinking { thinking, signature, .. }
1233 if thinking == "Completed thinking" && signature == "valid-signature"
1234 ));
1235 }
1236
1237 #[test]
1238 fn test_only_unsigned_thinking_block_omits_entire_message() {
1239 let result = request_with_assistant_content(vec![MessageContent::Thinking {
1240 text: "Cancelled before any text or signature".to_string(),
1241 signature: None,
1242 }]);
1243
1244 let assistant_messages: Vec<_> = result
1245 .messages
1246 .iter()
1247 .filter(|m| m.role == anthropic::Role::Assistant)
1248 .collect();
1249
1250 assert_eq!(
1251 assistant_messages.len(),
1252 0,
1253 "An assistant message whose only content was an unsigned thinking block \
1254 should be omitted entirely"
1255 );
1256 }
1257}