1use anyhow::{Context as _, Result, anyhow};
2use collections::BTreeMap;
3use credentials_provider::CredentialsProvider;
4use futures::{FutureExt, Stream, StreamExt, future, future::BoxFuture};
5use google_ai::{
6 FunctionDeclaration, GenerateContentResponse, GoogleModelMode, Part, SystemInstruction,
7 ThinkingConfig, UsageMetadata,
8};
9use gpui::{AnyView, App, AsyncApp, Context, Entity, SharedString, Task, Window};
10use http_client::HttpClient;
11use language_model::{
12 AuthenticateError, ConfigurationViewTargetAgent, EnvVar, LanguageModelCompletionError,
13 LanguageModelCompletionEvent, LanguageModelToolChoice, LanguageModelToolSchemaFormat,
14 LanguageModelToolUse, LanguageModelToolUseId, MessageContent, StopReason,
15};
16use language_model::{
17 LanguageModel, LanguageModelId, LanguageModelName, LanguageModelProvider,
18 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
19 LanguageModelRequest, RateLimiter, Role,
20};
21use schemars::JsonSchema;
22use serde::{Deserialize, Serialize};
23pub use settings::GoogleAvailableModel as AvailableModel;
24use settings::{Settings, SettingsStore};
25use std::pin::Pin;
26use std::sync::{
27 Arc, LazyLock,
28 atomic::{self, AtomicU64},
29};
30use strum::IntoEnumIterator;
31use ui::{ButtonLink, ConfiguredApiCard, List, ListBulletItem, prelude::*};
32use ui_input::InputField;
33use util::ResultExt;
34
35use language_model::{ApiKey, ApiKeyState};
36
37const PROVIDER_ID: LanguageModelProviderId = language_model::GOOGLE_PROVIDER_ID;
38const PROVIDER_NAME: LanguageModelProviderName = language_model::GOOGLE_PROVIDER_NAME;
39
40#[derive(Default, Clone, Debug, PartialEq)]
41pub struct GoogleSettings {
42 pub api_url: String,
43 pub available_models: Vec<AvailableModel>,
44}
45
46#[derive(Clone, Copy, Debug, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
47#[serde(tag = "type", rename_all = "lowercase")]
48pub enum ModelMode {
49 #[default]
50 Default,
51 Thinking {
52 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
53 budget_tokens: Option<u32>,
54 },
55}
56
57pub struct GoogleLanguageModelProvider {
58 http_client: Arc<dyn HttpClient>,
59 state: Entity<State>,
60}
61
62pub struct State {
63 api_key_state: ApiKeyState,
64}
65
66const GEMINI_API_KEY_VAR_NAME: &str = "GEMINI_API_KEY";
67const GOOGLE_AI_API_KEY_VAR_NAME: &str = "GOOGLE_AI_API_KEY";
68
69static API_KEY_ENV_VAR: LazyLock<EnvVar> = LazyLock::new(|| {
70 // Try GEMINI_API_KEY first as primary, fallback to GOOGLE_AI_API_KEY
71 EnvVar::new(GEMINI_API_KEY_VAR_NAME.into()).or(EnvVar::new(GOOGLE_AI_API_KEY_VAR_NAME.into()))
72});
73
74impl State {
75 fn is_authenticated(&self) -> bool {
76 self.api_key_state.has_key()
77 }
78
79 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
80 let api_url = GoogleLanguageModelProvider::api_url(cx);
81 self.api_key_state
82 .store(api_url, api_key, |this| &mut this.api_key_state, cx)
83 }
84
85 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
86 let api_url = GoogleLanguageModelProvider::api_url(cx);
87 self.api_key_state
88 .load_if_needed(api_url, |this| &mut this.api_key_state, cx)
89 }
90}
91
92impl GoogleLanguageModelProvider {
93 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
94 let state = cx.new(|cx| {
95 cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
96 let api_url = Self::api_url(cx);
97 this.api_key_state
98 .handle_url_change(api_url, |this| &mut this.api_key_state, cx);
99 cx.notify();
100 })
101 .detach();
102 State {
103 api_key_state: ApiKeyState::new(Self::api_url(cx), (*API_KEY_ENV_VAR).clone()),
104 }
105 });
106
107 Self { http_client, state }
108 }
109
110 fn create_language_model(&self, model: google_ai::Model) -> Arc<dyn LanguageModel> {
111 Arc::new(GoogleLanguageModel {
112 id: LanguageModelId::from(model.id().to_string()),
113 model,
114 state: self.state.clone(),
115 http_client: self.http_client.clone(),
116 request_limiter: RateLimiter::new(4),
117 })
118 }
119
120 pub fn api_key_for_gemini_cli(cx: &mut App) -> Task<Result<String>> {
121 if let Some(key) = API_KEY_ENV_VAR.value.clone() {
122 return Task::ready(Ok(key));
123 }
124 let credentials_provider = <dyn CredentialsProvider>::global(cx);
125 let api_url = Self::api_url(cx).to_string();
126 cx.spawn(async move |cx| {
127 Ok(
128 ApiKey::load_from_system_keychain(&api_url, credentials_provider.as_ref(), cx)
129 .await?
130 .key()
131 .to_string(),
132 )
133 })
134 }
135
136 fn settings(cx: &App) -> &GoogleSettings {
137 &crate::AllLanguageModelSettings::get_global(cx).google
138 }
139
140 fn api_url(cx: &App) -> SharedString {
141 let api_url = &Self::settings(cx).api_url;
142 if api_url.is_empty() {
143 google_ai::API_URL.into()
144 } else {
145 SharedString::new(api_url.as_str())
146 }
147 }
148}
149
150impl LanguageModelProviderState for GoogleLanguageModelProvider {
151 type ObservableEntity = State;
152
153 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
154 Some(self.state.clone())
155 }
156}
157
158impl LanguageModelProvider for GoogleLanguageModelProvider {
159 fn id(&self) -> LanguageModelProviderId {
160 PROVIDER_ID
161 }
162
163 fn name(&self) -> LanguageModelProviderName {
164 PROVIDER_NAME
165 }
166
167 fn icon(&self) -> IconName {
168 IconName::AiGoogle
169 }
170
171 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
172 Some(self.create_language_model(google_ai::Model::default()))
173 }
174
175 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
176 Some(self.create_language_model(google_ai::Model::default_fast()))
177 }
178
179 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
180 let mut models = BTreeMap::default();
181
182 // Add base models from google_ai::Model::iter()
183 for model in google_ai::Model::iter() {
184 if !matches!(model, google_ai::Model::Custom { .. }) {
185 models.insert(model.id().to_string(), model);
186 }
187 }
188
189 // Override with available models from settings
190 for model in &GoogleLanguageModelProvider::settings(cx).available_models {
191 models.insert(
192 model.name.clone(),
193 google_ai::Model::Custom {
194 name: model.name.clone(),
195 display_name: model.display_name.clone(),
196 max_tokens: model.max_tokens,
197 mode: model.mode.unwrap_or_default(),
198 },
199 );
200 }
201
202 models
203 .into_values()
204 .map(|model| {
205 Arc::new(GoogleLanguageModel {
206 id: LanguageModelId::from(model.id().to_string()),
207 model,
208 state: self.state.clone(),
209 http_client: self.http_client.clone(),
210 request_limiter: RateLimiter::new(4),
211 }) as Arc<dyn LanguageModel>
212 })
213 .collect()
214 }
215
216 fn is_authenticated(&self, cx: &App) -> bool {
217 self.state.read(cx).is_authenticated()
218 }
219
220 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
221 self.state.update(cx, |state, cx| state.authenticate(cx))
222 }
223
224 fn configuration_view(
225 &self,
226 target_agent: language_model::ConfigurationViewTargetAgent,
227 window: &mut Window,
228 cx: &mut App,
229 ) -> AnyView {
230 cx.new(|cx| ConfigurationView::new(self.state.clone(), target_agent, window, cx))
231 .into()
232 }
233
234 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
235 self.state
236 .update(cx, |state, cx| state.set_api_key(None, cx))
237 }
238}
239
240pub struct GoogleLanguageModel {
241 id: LanguageModelId,
242 model: google_ai::Model,
243 state: Entity<State>,
244 http_client: Arc<dyn HttpClient>,
245 request_limiter: RateLimiter,
246}
247
248impl GoogleLanguageModel {
249 fn stream_completion(
250 &self,
251 request: google_ai::GenerateContentRequest,
252 cx: &AsyncApp,
253 ) -> BoxFuture<
254 'static,
255 Result<futures::stream::BoxStream<'static, Result<GenerateContentResponse>>>,
256 > {
257 let http_client = self.http_client.clone();
258
259 let Ok((api_key, api_url)) = self.state.read_with(cx, |state, cx| {
260 let api_url = GoogleLanguageModelProvider::api_url(cx);
261 (state.api_key_state.key(&api_url), api_url)
262 }) else {
263 return future::ready(Err(anyhow!("App state dropped"))).boxed();
264 };
265
266 async move {
267 let api_key = api_key.context("Missing Google API key")?;
268 let request = google_ai::stream_generate_content(
269 http_client.as_ref(),
270 &api_url,
271 &api_key,
272 request,
273 );
274 request.await.context("failed to stream completion")
275 }
276 .boxed()
277 }
278}
279
280impl LanguageModel for GoogleLanguageModel {
281 fn id(&self) -> LanguageModelId {
282 self.id.clone()
283 }
284
285 fn name(&self) -> LanguageModelName {
286 LanguageModelName::from(self.model.display_name().to_string())
287 }
288
289 fn provider_id(&self) -> LanguageModelProviderId {
290 PROVIDER_ID
291 }
292
293 fn provider_name(&self) -> LanguageModelProviderName {
294 PROVIDER_NAME
295 }
296
297 fn supports_tools(&self) -> bool {
298 self.model.supports_tools()
299 }
300
301 fn supports_images(&self) -> bool {
302 self.model.supports_images()
303 }
304
305 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
306 match choice {
307 LanguageModelToolChoice::Auto
308 | LanguageModelToolChoice::Any
309 | LanguageModelToolChoice::None => true,
310 }
311 }
312
313 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
314 LanguageModelToolSchemaFormat::JsonSchemaSubset
315 }
316
317 fn telemetry_id(&self) -> String {
318 format!("google/{}", self.model.request_id())
319 }
320
321 fn max_token_count(&self) -> u64 {
322 self.model.max_token_count()
323 }
324
325 fn max_output_tokens(&self) -> Option<u64> {
326 self.model.max_output_tokens()
327 }
328
329 fn count_tokens(
330 &self,
331 request: LanguageModelRequest,
332 cx: &App,
333 ) -> BoxFuture<'static, Result<u64>> {
334 let model_id = self.model.request_id().to_string();
335 let request = into_google(request, model_id, self.model.mode());
336 let http_client = self.http_client.clone();
337 let api_url = GoogleLanguageModelProvider::api_url(cx);
338 let api_key = self.state.read(cx).api_key_state.key(&api_url);
339
340 async move {
341 let Some(api_key) = api_key else {
342 return Err(LanguageModelCompletionError::NoApiKey {
343 provider: PROVIDER_NAME,
344 }
345 .into());
346 };
347 let response = google_ai::count_tokens(
348 http_client.as_ref(),
349 &api_url,
350 &api_key,
351 google_ai::CountTokensRequest {
352 generate_content_request: request,
353 },
354 )
355 .await?;
356 Ok(response.total_tokens)
357 }
358 .boxed()
359 }
360
361 fn stream_completion(
362 &self,
363 request: LanguageModelRequest,
364 cx: &AsyncApp,
365 ) -> BoxFuture<
366 'static,
367 Result<
368 futures::stream::BoxStream<
369 'static,
370 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
371 >,
372 LanguageModelCompletionError,
373 >,
374 > {
375 let request = into_google(
376 request,
377 self.model.request_id().to_string(),
378 self.model.mode(),
379 );
380 let request = self.stream_completion(request, cx);
381 let future = self.request_limiter.stream(async move {
382 let response = request.await.map_err(LanguageModelCompletionError::from)?;
383 Ok(GoogleEventMapper::new().map_stream(response))
384 });
385 async move { Ok(future.await?.boxed()) }.boxed()
386 }
387}
388
389pub fn into_google(
390 mut request: LanguageModelRequest,
391 model_id: String,
392 mode: GoogleModelMode,
393) -> google_ai::GenerateContentRequest {
394 fn map_content(content: Vec<MessageContent>) -> Vec<Part> {
395 content
396 .into_iter()
397 .flat_map(|content| match content {
398 language_model::MessageContent::Text(text) => {
399 if !text.is_empty() {
400 vec![Part::TextPart(google_ai::TextPart { text })]
401 } else {
402 vec![]
403 }
404 }
405 language_model::MessageContent::Thinking {
406 text: _,
407 signature: Some(signature),
408 } => {
409 if !signature.is_empty() {
410 vec![Part::ThoughtPart(google_ai::ThoughtPart {
411 thought: true,
412 thought_signature: signature,
413 })]
414 } else {
415 vec![]
416 }
417 }
418 language_model::MessageContent::Thinking { .. } => {
419 vec![]
420 }
421 language_model::MessageContent::RedactedThinking(_) => vec![],
422 language_model::MessageContent::Image(image) => {
423 vec![Part::InlineDataPart(google_ai::InlineDataPart {
424 inline_data: google_ai::GenerativeContentBlob {
425 mime_type: "image/png".to_string(),
426 data: image.source.to_string(),
427 },
428 })]
429 }
430 language_model::MessageContent::ToolUse(tool_use) => {
431 // Normalize empty string signatures to None
432 let thought_signature = tool_use.thought_signature.filter(|s| !s.is_empty());
433
434 vec![Part::FunctionCallPart(google_ai::FunctionCallPart {
435 function_call: google_ai::FunctionCall {
436 name: tool_use.name.to_string(),
437 args: tool_use.input,
438 },
439 thought_signature,
440 })]
441 }
442 language_model::MessageContent::ToolResult(tool_result) => {
443 match tool_result.content {
444 language_model::LanguageModelToolResultContent::Text(text) => {
445 vec![Part::FunctionResponsePart(
446 google_ai::FunctionResponsePart {
447 function_response: google_ai::FunctionResponse {
448 name: tool_result.tool_name.to_string(),
449 // The API expects a valid JSON object
450 response: serde_json::json!({
451 "output": text
452 }),
453 },
454 },
455 )]
456 }
457 language_model::LanguageModelToolResultContent::Image(image) => {
458 vec![
459 Part::FunctionResponsePart(google_ai::FunctionResponsePart {
460 function_response: google_ai::FunctionResponse {
461 name: tool_result.tool_name.to_string(),
462 // The API expects a valid JSON object
463 response: serde_json::json!({
464 "output": "Tool responded with an image"
465 }),
466 },
467 }),
468 Part::InlineDataPart(google_ai::InlineDataPart {
469 inline_data: google_ai::GenerativeContentBlob {
470 mime_type: "image/png".to_string(),
471 data: image.source.to_string(),
472 },
473 }),
474 ]
475 }
476 }
477 }
478 })
479 .collect()
480 }
481
482 let system_instructions = if request
483 .messages
484 .first()
485 .is_some_and(|msg| matches!(msg.role, Role::System))
486 {
487 let message = request.messages.remove(0);
488 Some(SystemInstruction {
489 parts: map_content(message.content),
490 })
491 } else {
492 None
493 };
494
495 google_ai::GenerateContentRequest {
496 model: google_ai::ModelName { model_id },
497 system_instruction: system_instructions,
498 contents: request
499 .messages
500 .into_iter()
501 .filter_map(|message| {
502 let parts = map_content(message.content);
503 if parts.is_empty() {
504 None
505 } else {
506 Some(google_ai::Content {
507 parts,
508 role: match message.role {
509 Role::User => google_ai::Role::User,
510 Role::Assistant => google_ai::Role::Model,
511 Role::System => google_ai::Role::User, // Google AI doesn't have a system role
512 },
513 })
514 }
515 })
516 .collect(),
517 generation_config: Some(google_ai::GenerationConfig {
518 candidate_count: Some(1),
519 stop_sequences: Some(request.stop),
520 max_output_tokens: None,
521 temperature: request.temperature.map(|t| t as f64).or(Some(1.0)),
522 thinking_config: match (request.thinking_allowed, mode) {
523 (true, GoogleModelMode::Thinking { budget_tokens }) => {
524 budget_tokens.map(|thinking_budget| ThinkingConfig { thinking_budget })
525 }
526 _ => None,
527 },
528 top_p: None,
529 top_k: None,
530 }),
531 safety_settings: None,
532 tools: (!request.tools.is_empty()).then(|| {
533 vec![google_ai::Tool {
534 function_declarations: request
535 .tools
536 .into_iter()
537 .map(|tool| FunctionDeclaration {
538 name: tool.name,
539 description: tool.description,
540 parameters: tool.input_schema,
541 })
542 .collect(),
543 }]
544 }),
545 tool_config: request.tool_choice.map(|choice| google_ai::ToolConfig {
546 function_calling_config: google_ai::FunctionCallingConfig {
547 mode: match choice {
548 LanguageModelToolChoice::Auto => google_ai::FunctionCallingMode::Auto,
549 LanguageModelToolChoice::Any => google_ai::FunctionCallingMode::Any,
550 LanguageModelToolChoice::None => google_ai::FunctionCallingMode::None,
551 },
552 allowed_function_names: None,
553 },
554 }),
555 }
556}
557
558pub struct GoogleEventMapper {
559 usage: UsageMetadata,
560 stop_reason: StopReason,
561}
562
563impl GoogleEventMapper {
564 pub fn new() -> Self {
565 Self {
566 usage: UsageMetadata::default(),
567 stop_reason: StopReason::EndTurn,
568 }
569 }
570
571 pub fn map_stream(
572 mut self,
573 events: Pin<Box<dyn Send + Stream<Item = Result<GenerateContentResponse>>>>,
574 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
575 {
576 events
577 .map(Some)
578 .chain(futures::stream::once(async { None }))
579 .flat_map(move |event| {
580 futures::stream::iter(match event {
581 Some(Ok(event)) => self.map_event(event),
582 Some(Err(error)) => {
583 vec![Err(LanguageModelCompletionError::from(error))]
584 }
585 None => vec![Ok(LanguageModelCompletionEvent::Stop(self.stop_reason))],
586 })
587 })
588 }
589
590 pub fn map_event(
591 &mut self,
592 event: GenerateContentResponse,
593 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
594 static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
595
596 let mut events: Vec<_> = Vec::new();
597 let mut wants_to_use_tool = false;
598 if let Some(usage_metadata) = event.usage_metadata {
599 update_usage(&mut self.usage, &usage_metadata);
600 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
601 convert_usage(&self.usage),
602 )))
603 }
604
605 if let Some(prompt_feedback) = event.prompt_feedback
606 && let Some(block_reason) = prompt_feedback.block_reason.as_deref()
607 {
608 self.stop_reason = match block_reason {
609 "SAFETY" | "OTHER" | "BLOCKLIST" | "PROHIBITED_CONTENT" | "IMAGE_SAFETY" => {
610 StopReason::Refusal
611 }
612 _ => {
613 log::error!("Unexpected Google block_reason: {block_reason}");
614 StopReason::Refusal
615 }
616 };
617 events.push(Ok(LanguageModelCompletionEvent::Stop(self.stop_reason)));
618
619 return events;
620 }
621
622 if let Some(candidates) = event.candidates {
623 for candidate in candidates {
624 if let Some(finish_reason) = candidate.finish_reason.as_deref() {
625 self.stop_reason = match finish_reason {
626 "STOP" => StopReason::EndTurn,
627 "MAX_TOKENS" => StopReason::MaxTokens,
628 _ => {
629 log::error!("Unexpected google finish_reason: {finish_reason}");
630 StopReason::EndTurn
631 }
632 };
633 }
634 candidate
635 .content
636 .parts
637 .into_iter()
638 .for_each(|part| match part {
639 Part::TextPart(text_part) => {
640 events.push(Ok(LanguageModelCompletionEvent::Text(text_part.text)))
641 }
642 Part::InlineDataPart(_) => {}
643 Part::FunctionCallPart(function_call_part) => {
644 wants_to_use_tool = true;
645 let name: Arc<str> = function_call_part.function_call.name.into();
646 let next_tool_id =
647 TOOL_CALL_COUNTER.fetch_add(1, atomic::Ordering::SeqCst);
648 let id: LanguageModelToolUseId =
649 format!("{}-{}", name, next_tool_id).into();
650
651 // Normalize empty string signatures to None
652 let thought_signature = function_call_part
653 .thought_signature
654 .filter(|s| !s.is_empty());
655
656 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
657 LanguageModelToolUse {
658 id,
659 name,
660 is_input_complete: true,
661 raw_input: function_call_part.function_call.args.to_string(),
662 input: function_call_part.function_call.args,
663 thought_signature,
664 },
665 )));
666 }
667 Part::FunctionResponsePart(_) => {}
668 Part::ThoughtPart(part) => {
669 events.push(Ok(LanguageModelCompletionEvent::Thinking {
670 text: "(Encrypted thought)".to_string(), // TODO: Can we populate this from thought summaries?
671 signature: Some(part.thought_signature),
672 }));
673 }
674 });
675 }
676 }
677
678 // Even when Gemini wants to use a Tool, the API
679 // responds with `finish_reason: STOP`
680 if wants_to_use_tool {
681 self.stop_reason = StopReason::ToolUse;
682 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
683 }
684 events
685 }
686}
687
688pub fn count_google_tokens(
689 request: LanguageModelRequest,
690 cx: &App,
691) -> BoxFuture<'static, Result<u64>> {
692 // We couldn't use the GoogleLanguageModelProvider to count tokens because the github copilot doesn't have the access to google_ai directly.
693 // So we have to use tokenizer from tiktoken_rs to count tokens.
694 cx.background_spawn(async move {
695 let messages = request
696 .messages
697 .into_iter()
698 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
699 role: match message.role {
700 Role::User => "user".into(),
701 Role::Assistant => "assistant".into(),
702 Role::System => "system".into(),
703 },
704 content: Some(message.string_contents()),
705 name: None,
706 function_call: None,
707 })
708 .collect::<Vec<_>>();
709
710 // Tiktoken doesn't yet support these models, so we manually use the
711 // same tokenizer as GPT-4.
712 tiktoken_rs::num_tokens_from_messages("gpt-4", &messages).map(|tokens| tokens as u64)
713 })
714 .boxed()
715}
716
717fn update_usage(usage: &mut UsageMetadata, new: &UsageMetadata) {
718 if let Some(prompt_token_count) = new.prompt_token_count {
719 usage.prompt_token_count = Some(prompt_token_count);
720 }
721 if let Some(cached_content_token_count) = new.cached_content_token_count {
722 usage.cached_content_token_count = Some(cached_content_token_count);
723 }
724 if let Some(candidates_token_count) = new.candidates_token_count {
725 usage.candidates_token_count = Some(candidates_token_count);
726 }
727 if let Some(tool_use_prompt_token_count) = new.tool_use_prompt_token_count {
728 usage.tool_use_prompt_token_count = Some(tool_use_prompt_token_count);
729 }
730 if let Some(thoughts_token_count) = new.thoughts_token_count {
731 usage.thoughts_token_count = Some(thoughts_token_count);
732 }
733 if let Some(total_token_count) = new.total_token_count {
734 usage.total_token_count = Some(total_token_count);
735 }
736}
737
738fn convert_usage(usage: &UsageMetadata) -> language_model::TokenUsage {
739 let prompt_tokens = usage.prompt_token_count.unwrap_or(0);
740 let cached_tokens = usage.cached_content_token_count.unwrap_or(0);
741 let input_tokens = prompt_tokens - cached_tokens;
742 let output_tokens = usage.candidates_token_count.unwrap_or(0);
743
744 language_model::TokenUsage {
745 input_tokens,
746 output_tokens,
747 cache_read_input_tokens: cached_tokens,
748 cache_creation_input_tokens: 0,
749 }
750}
751
752struct ConfigurationView {
753 api_key_editor: Entity<InputField>,
754 state: Entity<State>,
755 target_agent: language_model::ConfigurationViewTargetAgent,
756 load_credentials_task: Option<Task<()>>,
757}
758
759impl ConfigurationView {
760 fn new(
761 state: Entity<State>,
762 target_agent: language_model::ConfigurationViewTargetAgent,
763 window: &mut Window,
764 cx: &mut Context<Self>,
765 ) -> Self {
766 cx.observe(&state, |_, _, cx| {
767 cx.notify();
768 })
769 .detach();
770
771 let load_credentials_task = Some(cx.spawn_in(window, {
772 let state = state.clone();
773 async move |this, cx| {
774 if let Some(task) = state
775 .update(cx, |state, cx| state.authenticate(cx))
776 .log_err()
777 {
778 // We don't log an error, because "not signed in" is also an error.
779 let _ = task.await;
780 }
781 this.update(cx, |this, cx| {
782 this.load_credentials_task = None;
783 cx.notify();
784 })
785 .log_err();
786 }
787 }));
788
789 Self {
790 api_key_editor: cx.new(|cx| InputField::new(window, cx, "AIzaSy...")),
791 target_agent,
792 state,
793 load_credentials_task,
794 }
795 }
796
797 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
798 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
799 if api_key.is_empty() {
800 return;
801 }
802
803 // url changes can cause the editor to be displayed again
804 self.api_key_editor
805 .update(cx, |editor, cx| editor.set_text("", window, cx));
806
807 let state = self.state.clone();
808 cx.spawn_in(window, async move |_, cx| {
809 state
810 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))?
811 .await
812 })
813 .detach_and_log_err(cx);
814 }
815
816 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
817 self.api_key_editor
818 .update(cx, |editor, cx| editor.set_text("", window, cx));
819
820 let state = self.state.clone();
821 cx.spawn_in(window, async move |_, cx| {
822 state
823 .update(cx, |state, cx| state.set_api_key(None, cx))?
824 .await
825 })
826 .detach_and_log_err(cx);
827 }
828
829 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
830 !self.state.read(cx).is_authenticated()
831 }
832}
833
834impl Render for ConfigurationView {
835 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
836 let env_var_set = self.state.read(cx).api_key_state.is_from_env_var();
837 let configured_card_label = if env_var_set {
838 format!(
839 "API key set in {} environment variable",
840 API_KEY_ENV_VAR.name
841 )
842 } else {
843 let api_url = GoogleLanguageModelProvider::api_url(cx);
844 if api_url == google_ai::API_URL {
845 "API key configured".to_string()
846 } else {
847 format!("API key configured for {}", api_url)
848 }
849 };
850
851 if self.load_credentials_task.is_some() {
852 div()
853 .child(Label::new("Loading credentials..."))
854 .into_any_element()
855 } else if self.should_render_editor(cx) {
856 v_flex()
857 .size_full()
858 .on_action(cx.listener(Self::save_api_key))
859 .child(Label::new(format!("To use {}, you need to add an API key. Follow these steps:", match &self.target_agent {
860 ConfigurationViewTargetAgent::ZedAgent => "Zed's agent with Google AI".into(),
861 ConfigurationViewTargetAgent::Other(agent) => agent.clone(),
862 })))
863 .child(
864 List::new()
865 .child(
866 ListBulletItem::new("")
867 .child(Label::new("Create one by visiting"))
868 .child(ButtonLink::new("Google AI's console", "https://aistudio.google.com/app/apikey"))
869 )
870 .child(
871 ListBulletItem::new("Paste your API key below and hit enter to start using the agent")
872 )
873 )
874 .child(self.api_key_editor.clone())
875 .child(
876 Label::new(
877 format!("You can also assign the {GEMINI_API_KEY_VAR_NAME} environment variable and restart Zed."),
878 )
879 .size(LabelSize::Small).color(Color::Muted),
880 )
881 .into_any_element()
882 } else {
883 ConfiguredApiCard::new(configured_card_label)
884 .disabled(env_var_set)
885 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
886 .when(env_var_set, |this| {
887 this.tooltip_label(format!("To reset your API key, make sure {GEMINI_API_KEY_VAR_NAME} and {GOOGLE_AI_API_KEY_VAR_NAME} environment variables are unset."))
888 })
889 .into_any_element()
890 }
891 }
892}
893
894#[cfg(test)]
895mod tests {
896 use super::*;
897 use google_ai::{
898 Content, FunctionCall, FunctionCallPart, GenerateContentCandidate, GenerateContentResponse,
899 Part, Role as GoogleRole, TextPart,
900 };
901 use language_model::{LanguageModelToolUseId, MessageContent, Role};
902 use serde_json::json;
903
904 #[test]
905 fn test_function_call_with_signature_creates_tool_use_with_signature() {
906 let mut mapper = GoogleEventMapper::new();
907
908 let response = GenerateContentResponse {
909 candidates: Some(vec![GenerateContentCandidate {
910 index: Some(0),
911 content: Content {
912 parts: vec![Part::FunctionCallPart(FunctionCallPart {
913 function_call: FunctionCall {
914 name: "test_function".to_string(),
915 args: json!({"arg": "value"}),
916 },
917 thought_signature: Some("test_signature_123".to_string()),
918 })],
919 role: GoogleRole::Model,
920 },
921 finish_reason: None,
922 finish_message: None,
923 safety_ratings: None,
924 citation_metadata: None,
925 }]),
926 prompt_feedback: None,
927 usage_metadata: None,
928 };
929
930 let events = mapper.map_event(response);
931
932 assert_eq!(events.len(), 2); // ToolUse event + Stop event
933
934 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
935 assert_eq!(tool_use.name.as_ref(), "test_function");
936 assert_eq!(
937 tool_use.thought_signature.as_deref(),
938 Some("test_signature_123")
939 );
940 } else {
941 panic!("Expected ToolUse event");
942 }
943 }
944
945 #[test]
946 fn test_function_call_without_signature_has_none() {
947 let mut mapper = GoogleEventMapper::new();
948
949 let response = GenerateContentResponse {
950 candidates: Some(vec![GenerateContentCandidate {
951 index: Some(0),
952 content: Content {
953 parts: vec![Part::FunctionCallPart(FunctionCallPart {
954 function_call: FunctionCall {
955 name: "test_function".to_string(),
956 args: json!({"arg": "value"}),
957 },
958 thought_signature: None,
959 })],
960 role: GoogleRole::Model,
961 },
962 finish_reason: None,
963 finish_message: None,
964 safety_ratings: None,
965 citation_metadata: None,
966 }]),
967 prompt_feedback: None,
968 usage_metadata: None,
969 };
970
971 let events = mapper.map_event(response);
972
973 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
974 assert_eq!(tool_use.thought_signature, None);
975 } else {
976 panic!("Expected ToolUse event");
977 }
978 }
979
980 #[test]
981 fn test_empty_string_signature_normalized_to_none() {
982 let mut mapper = GoogleEventMapper::new();
983
984 let response = GenerateContentResponse {
985 candidates: Some(vec![GenerateContentCandidate {
986 index: Some(0),
987 content: Content {
988 parts: vec![Part::FunctionCallPart(FunctionCallPart {
989 function_call: FunctionCall {
990 name: "test_function".to_string(),
991 args: json!({"arg": "value"}),
992 },
993 thought_signature: Some("".to_string()),
994 })],
995 role: GoogleRole::Model,
996 },
997 finish_reason: None,
998 finish_message: None,
999 safety_ratings: None,
1000 citation_metadata: None,
1001 }]),
1002 prompt_feedback: None,
1003 usage_metadata: None,
1004 };
1005
1006 let events = mapper.map_event(response);
1007
1008 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1009 assert_eq!(tool_use.thought_signature, None);
1010 } else {
1011 panic!("Expected ToolUse event");
1012 }
1013 }
1014
1015 #[test]
1016 fn test_parallel_function_calls_preserve_signatures() {
1017 let mut mapper = GoogleEventMapper::new();
1018
1019 let response = GenerateContentResponse {
1020 candidates: Some(vec![GenerateContentCandidate {
1021 index: Some(0),
1022 content: Content {
1023 parts: vec![
1024 Part::FunctionCallPart(FunctionCallPart {
1025 function_call: FunctionCall {
1026 name: "function_1".to_string(),
1027 args: json!({"arg": "value1"}),
1028 },
1029 thought_signature: Some("signature_1".to_string()),
1030 }),
1031 Part::FunctionCallPart(FunctionCallPart {
1032 function_call: FunctionCall {
1033 name: "function_2".to_string(),
1034 args: json!({"arg": "value2"}),
1035 },
1036 thought_signature: None,
1037 }),
1038 ],
1039 role: GoogleRole::Model,
1040 },
1041 finish_reason: None,
1042 finish_message: None,
1043 safety_ratings: None,
1044 citation_metadata: None,
1045 }]),
1046 prompt_feedback: None,
1047 usage_metadata: None,
1048 };
1049
1050 let events = mapper.map_event(response);
1051
1052 assert_eq!(events.len(), 3); // 2 ToolUse events + Stop event
1053
1054 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1055 assert_eq!(tool_use.name.as_ref(), "function_1");
1056 assert_eq!(tool_use.thought_signature.as_deref(), Some("signature_1"));
1057 } else {
1058 panic!("Expected ToolUse event for function_1");
1059 }
1060
1061 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[1] {
1062 assert_eq!(tool_use.name.as_ref(), "function_2");
1063 assert_eq!(tool_use.thought_signature, None);
1064 } else {
1065 panic!("Expected ToolUse event for function_2");
1066 }
1067 }
1068
1069 #[test]
1070 fn test_tool_use_with_signature_converts_to_function_call_part() {
1071 let tool_use = language_model::LanguageModelToolUse {
1072 id: LanguageModelToolUseId::from("test_id"),
1073 name: "test_function".into(),
1074 raw_input: json!({"arg": "value"}).to_string(),
1075 input: json!({"arg": "value"}),
1076 is_input_complete: true,
1077 thought_signature: Some("test_signature_456".to_string()),
1078 };
1079
1080 let request = super::into_google(
1081 LanguageModelRequest {
1082 messages: vec![language_model::LanguageModelRequestMessage {
1083 role: Role::Assistant,
1084 content: vec![MessageContent::ToolUse(tool_use)],
1085 cache: false,
1086 reasoning_details: None,
1087 }],
1088 ..Default::default()
1089 },
1090 "gemini-2.5-flash".to_string(),
1091 GoogleModelMode::Default,
1092 );
1093
1094 assert_eq!(request.contents[0].parts.len(), 1);
1095 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1096 assert_eq!(fc_part.function_call.name, "test_function");
1097 assert_eq!(
1098 fc_part.thought_signature.as_deref(),
1099 Some("test_signature_456")
1100 );
1101 } else {
1102 panic!("Expected FunctionCallPart");
1103 }
1104 }
1105
1106 #[test]
1107 fn test_tool_use_without_signature_omits_field() {
1108 let tool_use = language_model::LanguageModelToolUse {
1109 id: LanguageModelToolUseId::from("test_id"),
1110 name: "test_function".into(),
1111 raw_input: json!({"arg": "value"}).to_string(),
1112 input: json!({"arg": "value"}),
1113 is_input_complete: true,
1114 thought_signature: None,
1115 };
1116
1117 let request = super::into_google(
1118 LanguageModelRequest {
1119 messages: vec![language_model::LanguageModelRequestMessage {
1120 role: Role::Assistant,
1121 content: vec![MessageContent::ToolUse(tool_use)],
1122 cache: false,
1123 reasoning_details: None,
1124 }],
1125 ..Default::default()
1126 },
1127 "gemini-2.5-flash".to_string(),
1128 GoogleModelMode::Default,
1129 );
1130
1131 assert_eq!(request.contents[0].parts.len(), 1);
1132 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1133 assert_eq!(fc_part.thought_signature, None);
1134 } else {
1135 panic!("Expected FunctionCallPart");
1136 }
1137 }
1138
1139 #[test]
1140 fn test_empty_signature_in_tool_use_normalized_to_none() {
1141 let tool_use = language_model::LanguageModelToolUse {
1142 id: LanguageModelToolUseId::from("test_id"),
1143 name: "test_function".into(),
1144 raw_input: json!({"arg": "value"}).to_string(),
1145 input: json!({"arg": "value"}),
1146 is_input_complete: true,
1147 thought_signature: Some("".to_string()),
1148 };
1149
1150 let request = super::into_google(
1151 LanguageModelRequest {
1152 messages: vec![language_model::LanguageModelRequestMessage {
1153 role: Role::Assistant,
1154 content: vec![MessageContent::ToolUse(tool_use)],
1155 cache: false,
1156 reasoning_details: None,
1157 }],
1158 ..Default::default()
1159 },
1160 "gemini-2.5-flash".to_string(),
1161 GoogleModelMode::Default,
1162 );
1163
1164 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1165 assert_eq!(fc_part.thought_signature, None);
1166 } else {
1167 panic!("Expected FunctionCallPart");
1168 }
1169 }
1170
1171 #[test]
1172 fn test_round_trip_preserves_signature() {
1173 let mut mapper = GoogleEventMapper::new();
1174
1175 // Simulate receiving a response from Google with a signature
1176 let response = GenerateContentResponse {
1177 candidates: Some(vec![GenerateContentCandidate {
1178 index: Some(0),
1179 content: Content {
1180 parts: vec![Part::FunctionCallPart(FunctionCallPart {
1181 function_call: FunctionCall {
1182 name: "test_function".to_string(),
1183 args: json!({"arg": "value"}),
1184 },
1185 thought_signature: Some("round_trip_sig".to_string()),
1186 })],
1187 role: GoogleRole::Model,
1188 },
1189 finish_reason: None,
1190 finish_message: None,
1191 safety_ratings: None,
1192 citation_metadata: None,
1193 }]),
1194 prompt_feedback: None,
1195 usage_metadata: None,
1196 };
1197
1198 let events = mapper.map_event(response);
1199
1200 let tool_use = if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1201 tool_use.clone()
1202 } else {
1203 panic!("Expected ToolUse event");
1204 };
1205
1206 // Convert back to Google format
1207 let request = super::into_google(
1208 LanguageModelRequest {
1209 messages: vec![language_model::LanguageModelRequestMessage {
1210 role: Role::Assistant,
1211 content: vec![MessageContent::ToolUse(tool_use)],
1212 cache: false,
1213 reasoning_details: None,
1214 }],
1215 ..Default::default()
1216 },
1217 "gemini-2.5-flash".to_string(),
1218 GoogleModelMode::Default,
1219 );
1220
1221 // Verify signature is preserved
1222 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1223 assert_eq!(fc_part.thought_signature.as_deref(), Some("round_trip_sig"));
1224 } else {
1225 panic!("Expected FunctionCallPart");
1226 }
1227 }
1228
1229 #[test]
1230 fn test_mixed_text_and_function_call_with_signature() {
1231 let mut mapper = GoogleEventMapper::new();
1232
1233 let response = GenerateContentResponse {
1234 candidates: Some(vec![GenerateContentCandidate {
1235 index: Some(0),
1236 content: Content {
1237 parts: vec![
1238 Part::TextPart(TextPart {
1239 text: "I'll help with that.".to_string(),
1240 }),
1241 Part::FunctionCallPart(FunctionCallPart {
1242 function_call: FunctionCall {
1243 name: "helper_function".to_string(),
1244 args: json!({"query": "help"}),
1245 },
1246 thought_signature: Some("mixed_sig".to_string()),
1247 }),
1248 ],
1249 role: GoogleRole::Model,
1250 },
1251 finish_reason: None,
1252 finish_message: None,
1253 safety_ratings: None,
1254 citation_metadata: None,
1255 }]),
1256 prompt_feedback: None,
1257 usage_metadata: None,
1258 };
1259
1260 let events = mapper.map_event(response);
1261
1262 assert_eq!(events.len(), 3); // Text event + ToolUse event + Stop event
1263
1264 if let Ok(LanguageModelCompletionEvent::Text(text)) = &events[0] {
1265 assert_eq!(text, "I'll help with that.");
1266 } else {
1267 panic!("Expected Text event");
1268 }
1269
1270 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[1] {
1271 assert_eq!(tool_use.name.as_ref(), "helper_function");
1272 assert_eq!(tool_use.thought_signature.as_deref(), Some("mixed_sig"));
1273 } else {
1274 panic!("Expected ToolUse event");
1275 }
1276 }
1277
1278 #[test]
1279 fn test_special_characters_in_signature_preserved() {
1280 let mut mapper = GoogleEventMapper::new();
1281
1282 let signature_with_special_chars = "sig<>\"'&%$#@!{}[]".to_string();
1283
1284 let response = GenerateContentResponse {
1285 candidates: Some(vec![GenerateContentCandidate {
1286 index: Some(0),
1287 content: Content {
1288 parts: vec![Part::FunctionCallPart(FunctionCallPart {
1289 function_call: FunctionCall {
1290 name: "test_function".to_string(),
1291 args: json!({"arg": "value"}),
1292 },
1293 thought_signature: Some(signature_with_special_chars.clone()),
1294 })],
1295 role: GoogleRole::Model,
1296 },
1297 finish_reason: None,
1298 finish_message: None,
1299 safety_ratings: None,
1300 citation_metadata: None,
1301 }]),
1302 prompt_feedback: None,
1303 usage_metadata: None,
1304 };
1305
1306 let events = mapper.map_event(response);
1307
1308 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1309 assert_eq!(
1310 tool_use.thought_signature.as_deref(),
1311 Some(signature_with_special_chars.as_str())
1312 );
1313 } else {
1314 panic!("Expected ToolUse event");
1315 }
1316 }
1317}