1use anyhow::{Result, anyhow};
2use collections::{BTreeMap, HashMap};
3use futures::Stream;
4use futures::{FutureExt, StreamExt, future::BoxFuture};
5use gpui::{AnyView, App, AsyncApp, Context, Entity, SharedString, Task, Window};
6use http_client::HttpClient;
7use language_model::{
8 ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel, LanguageModelCompletionError,
9 LanguageModelCompletionEvent, LanguageModelId, LanguageModelImage, LanguageModelName,
10 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
11 LanguageModelProviderState, LanguageModelRequest, LanguageModelRequestMessage,
12 LanguageModelToolChoice, LanguageModelToolResult, LanguageModelToolResultContent,
13 LanguageModelToolUse, LanguageModelToolUseId, MessageContent, RateLimiter, Role, StopReason,
14 TokenUsage, env_var,
15};
16use menu;
17use open_ai::responses::{
18 ResponseFunctionCallItem, ResponseFunctionCallOutputItem, ResponseInputContent,
19 ResponseInputItem, ResponseMessageItem,
20};
21use open_ai::{
22 ImageUrl, Model, OPEN_AI_API_URL, ReasoningEffort, ResponseStreamEvent,
23 responses::{
24 Request as ResponseRequest, ResponseOutputItem, ResponseSummary as ResponsesSummary,
25 ResponseUsage as ResponsesUsage, StreamEvent as ResponsesStreamEvent, stream_response,
26 },
27 stream_completion,
28};
29use settings::{OpenAiAvailableModel as AvailableModel, Settings, SettingsStore};
30use std::pin::Pin;
31use std::str::FromStr as _;
32use std::sync::{Arc, LazyLock};
33use strum::IntoEnumIterator;
34use ui::{ButtonLink, ConfiguredApiCard, List, ListBulletItem, prelude::*};
35use ui_input::InputField;
36use util::ResultExt;
37
38const PROVIDER_ID: LanguageModelProviderId = language_model::OPEN_AI_PROVIDER_ID;
39const PROVIDER_NAME: LanguageModelProviderName = language_model::OPEN_AI_PROVIDER_NAME;
40
41const API_KEY_ENV_VAR_NAME: &str = "OPENAI_API_KEY";
42static API_KEY_ENV_VAR: LazyLock<EnvVar> = env_var!(API_KEY_ENV_VAR_NAME);
43
44#[derive(Default, Clone, Debug, PartialEq)]
45pub struct OpenAiSettings {
46 pub api_url: String,
47 pub available_models: Vec<AvailableModel>,
48}
49
50pub struct OpenAiLanguageModelProvider {
51 http_client: Arc<dyn HttpClient>,
52 state: Entity<State>,
53}
54
55pub struct State {
56 api_key_state: ApiKeyState,
57}
58
59impl State {
60 fn is_authenticated(&self) -> bool {
61 self.api_key_state.has_key()
62 }
63
64 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
65 let api_url = OpenAiLanguageModelProvider::api_url(cx);
66 self.api_key_state
67 .store(api_url, api_key, |this| &mut this.api_key_state, cx)
68 }
69
70 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
71 let api_url = OpenAiLanguageModelProvider::api_url(cx);
72 self.api_key_state
73 .load_if_needed(api_url, |this| &mut this.api_key_state, cx)
74 }
75}
76
77impl OpenAiLanguageModelProvider {
78 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
79 let state = cx.new(|cx| {
80 cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
81 let api_url = Self::api_url(cx);
82 this.api_key_state
83 .handle_url_change(api_url, |this| &mut this.api_key_state, cx);
84 cx.notify();
85 })
86 .detach();
87 State {
88 api_key_state: ApiKeyState::new(Self::api_url(cx), (*API_KEY_ENV_VAR).clone()),
89 }
90 });
91
92 Self { http_client, state }
93 }
94
95 fn create_language_model(&self, model: open_ai::Model) -> Arc<dyn LanguageModel> {
96 Arc::new(OpenAiLanguageModel {
97 id: LanguageModelId::from(model.id().to_string()),
98 model,
99 state: self.state.clone(),
100 http_client: self.http_client.clone(),
101 request_limiter: RateLimiter::new(4),
102 })
103 }
104
105 fn settings(cx: &App) -> &OpenAiSettings {
106 &crate::AllLanguageModelSettings::get_global(cx).openai
107 }
108
109 fn api_url(cx: &App) -> SharedString {
110 let api_url = &Self::settings(cx).api_url;
111 if api_url.is_empty() {
112 open_ai::OPEN_AI_API_URL.into()
113 } else {
114 SharedString::new(api_url.as_str())
115 }
116 }
117}
118
119impl LanguageModelProviderState for OpenAiLanguageModelProvider {
120 type ObservableEntity = State;
121
122 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
123 Some(self.state.clone())
124 }
125}
126
127impl LanguageModelProvider for OpenAiLanguageModelProvider {
128 fn id(&self) -> LanguageModelProviderId {
129 PROVIDER_ID
130 }
131
132 fn name(&self) -> LanguageModelProviderName {
133 PROVIDER_NAME
134 }
135
136 fn icon(&self) -> IconOrSvg {
137 IconOrSvg::Icon(IconName::AiOpenAi)
138 }
139
140 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
141 Some(self.create_language_model(open_ai::Model::default()))
142 }
143
144 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
145 Some(self.create_language_model(open_ai::Model::default_fast()))
146 }
147
148 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
149 let mut models = BTreeMap::default();
150
151 // Add base models from open_ai::Model::iter()
152 for model in open_ai::Model::iter() {
153 if !matches!(model, open_ai::Model::Custom { .. }) {
154 models.insert(model.id().to_string(), model);
155 }
156 }
157
158 // Override with available models from settings
159 for model in &OpenAiLanguageModelProvider::settings(cx).available_models {
160 models.insert(
161 model.name.clone(),
162 open_ai::Model::Custom {
163 name: model.name.clone(),
164 display_name: model.display_name.clone(),
165 max_tokens: model.max_tokens,
166 max_output_tokens: model.max_output_tokens,
167 max_completion_tokens: model.max_completion_tokens,
168 reasoning_effort: model.reasoning_effort.clone(),
169 supports_chat_completions: model.capabilities.chat_completions,
170 },
171 );
172 }
173
174 models
175 .into_values()
176 .map(|model| self.create_language_model(model))
177 .collect()
178 }
179
180 fn is_authenticated(&self, cx: &App) -> bool {
181 self.state.read(cx).is_authenticated()
182 }
183
184 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
185 self.state.update(cx, |state, cx| state.authenticate(cx))
186 }
187
188 fn configuration_view(
189 &self,
190 _target_agent: language_model::ConfigurationViewTargetAgent,
191 window: &mut Window,
192 cx: &mut App,
193 ) -> AnyView {
194 cx.new(|cx| ConfigurationView::new(self.state.clone(), window, cx))
195 .into()
196 }
197
198 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
199 self.state
200 .update(cx, |state, cx| state.set_api_key(None, cx))
201 }
202}
203
204pub struct OpenAiLanguageModel {
205 id: LanguageModelId,
206 model: open_ai::Model,
207 state: Entity<State>,
208 http_client: Arc<dyn HttpClient>,
209 request_limiter: RateLimiter,
210}
211
212impl OpenAiLanguageModel {
213 fn stream_completion(
214 &self,
215 request: open_ai::Request,
216 cx: &AsyncApp,
217 ) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>>>
218 {
219 let http_client = self.http_client.clone();
220
221 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
222 let api_url = OpenAiLanguageModelProvider::api_url(cx);
223 (state.api_key_state.key(&api_url), api_url)
224 });
225
226 let future = self.request_limiter.stream(async move {
227 let provider = PROVIDER_NAME;
228 let Some(api_key) = api_key else {
229 return Err(LanguageModelCompletionError::NoApiKey { provider });
230 };
231 let request = stream_completion(
232 http_client.as_ref(),
233 provider.0.as_str(),
234 &api_url,
235 &api_key,
236 request,
237 );
238 let response = request.await?;
239 Ok(response)
240 });
241
242 async move { Ok(future.await?.boxed()) }.boxed()
243 }
244
245 fn stream_response(
246 &self,
247 request: ResponseRequest,
248 cx: &AsyncApp,
249 ) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponsesStreamEvent>>>>
250 {
251 let http_client = self.http_client.clone();
252
253 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
254 let api_url = OpenAiLanguageModelProvider::api_url(cx);
255 (state.api_key_state.key(&api_url), api_url)
256 });
257
258 let provider = PROVIDER_NAME;
259 let future = self.request_limiter.stream(async move {
260 let Some(api_key) = api_key else {
261 return Err(LanguageModelCompletionError::NoApiKey { provider });
262 };
263 let request = stream_response(
264 http_client.as_ref(),
265 provider.0.as_str(),
266 &api_url,
267 &api_key,
268 request,
269 );
270 let response = request.await?;
271 Ok(response)
272 });
273
274 async move { Ok(future.await?.boxed()) }.boxed()
275 }
276}
277
278impl LanguageModel for OpenAiLanguageModel {
279 fn id(&self) -> LanguageModelId {
280 self.id.clone()
281 }
282
283 fn name(&self) -> LanguageModelName {
284 LanguageModelName::from(self.model.display_name().to_string())
285 }
286
287 fn provider_id(&self) -> LanguageModelProviderId {
288 PROVIDER_ID
289 }
290
291 fn provider_name(&self) -> LanguageModelProviderName {
292 PROVIDER_NAME
293 }
294
295 fn supports_tools(&self) -> bool {
296 true
297 }
298
299 fn supports_images(&self) -> bool {
300 use open_ai::Model;
301 match &self.model {
302 Model::FourOmni
303 | Model::FourOmniMini
304 | Model::FourPointOne
305 | Model::FourPointOneMini
306 | Model::FourPointOneNano
307 | Model::Five
308 | Model::FiveCodex
309 | Model::FiveMini
310 | Model::FiveNano
311 | Model::FivePointOne
312 | Model::FivePointTwo
313 | Model::O1
314 | Model::O3
315 | Model::O4Mini => true,
316 Model::ThreePointFiveTurbo
317 | Model::Four
318 | Model::FourTurbo
319 | Model::O3Mini
320 | Model::Custom { .. } => false,
321 }
322 }
323
324 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
325 match choice {
326 LanguageModelToolChoice::Auto => true,
327 LanguageModelToolChoice::Any => true,
328 LanguageModelToolChoice::None => true,
329 }
330 }
331
332 fn supports_split_token_display(&self) -> bool {
333 true
334 }
335
336 fn telemetry_id(&self) -> String {
337 format!("openai/{}", self.model.id())
338 }
339
340 fn max_token_count(&self) -> u64 {
341 self.model.max_token_count()
342 }
343
344 fn max_output_tokens(&self) -> Option<u64> {
345 self.model.max_output_tokens()
346 }
347
348 fn count_tokens(
349 &self,
350 request: LanguageModelRequest,
351 cx: &App,
352 ) -> BoxFuture<'static, Result<u64>> {
353 count_open_ai_tokens(request, self.model.clone(), cx)
354 }
355
356 fn stream_completion(
357 &self,
358 request: LanguageModelRequest,
359 cx: &AsyncApp,
360 ) -> BoxFuture<
361 'static,
362 Result<
363 futures::stream::BoxStream<
364 'static,
365 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
366 >,
367 LanguageModelCompletionError,
368 >,
369 > {
370 if self.model.supports_chat_completions() {
371 let request = into_open_ai(
372 request,
373 self.model.id(),
374 self.model.supports_parallel_tool_calls(),
375 self.model.supports_prompt_cache_key(),
376 self.max_output_tokens(),
377 self.model.reasoning_effort(),
378 );
379 let completions = self.stream_completion(request, cx);
380 async move {
381 let mapper = OpenAiEventMapper::new();
382 Ok(mapper.map_stream(completions.await?).boxed())
383 }
384 .boxed()
385 } else {
386 let request = into_open_ai_response(
387 request,
388 self.model.id(),
389 self.model.supports_parallel_tool_calls(),
390 self.model.supports_prompt_cache_key(),
391 self.max_output_tokens(),
392 self.model.reasoning_effort(),
393 );
394 let completions = self.stream_response(request, cx);
395 async move {
396 let mapper = OpenAiResponseEventMapper::new();
397 Ok(mapper.map_stream(completions.await?).boxed())
398 }
399 .boxed()
400 }
401 }
402}
403
404pub fn into_open_ai(
405 request: LanguageModelRequest,
406 model_id: &str,
407 supports_parallel_tool_calls: bool,
408 supports_prompt_cache_key: bool,
409 max_output_tokens: Option<u64>,
410 reasoning_effort: Option<ReasoningEffort>,
411) -> open_ai::Request {
412 let stream = !model_id.starts_with("o1-");
413
414 let mut messages = Vec::new();
415 for message in request.messages {
416 for content in message.content {
417 match content {
418 MessageContent::Text(text) | MessageContent::Thinking { text, .. } => {
419 if !text.trim().is_empty() {
420 add_message_content_part(
421 open_ai::MessagePart::Text { text },
422 message.role,
423 &mut messages,
424 );
425 }
426 }
427 MessageContent::RedactedThinking(_) => {}
428 MessageContent::Image(image) => {
429 add_message_content_part(
430 open_ai::MessagePart::Image {
431 image_url: ImageUrl {
432 url: image.to_base64_url(),
433 detail: None,
434 },
435 },
436 message.role,
437 &mut messages,
438 );
439 }
440 MessageContent::ToolUse(tool_use) => {
441 let tool_call = open_ai::ToolCall {
442 id: tool_use.id.to_string(),
443 content: open_ai::ToolCallContent::Function {
444 function: open_ai::FunctionContent {
445 name: tool_use.name.to_string(),
446 arguments: serde_json::to_string(&tool_use.input)
447 .unwrap_or_default(),
448 },
449 },
450 };
451
452 if let Some(open_ai::RequestMessage::Assistant { tool_calls, .. }) =
453 messages.last_mut()
454 {
455 tool_calls.push(tool_call);
456 } else {
457 messages.push(open_ai::RequestMessage::Assistant {
458 content: None,
459 tool_calls: vec![tool_call],
460 });
461 }
462 }
463 MessageContent::ToolResult(tool_result) => {
464 let content = match &tool_result.content {
465 LanguageModelToolResultContent::Text(text) => {
466 vec![open_ai::MessagePart::Text {
467 text: text.to_string(),
468 }]
469 }
470 LanguageModelToolResultContent::Image(image) => {
471 vec![open_ai::MessagePart::Image {
472 image_url: ImageUrl {
473 url: image.to_base64_url(),
474 detail: None,
475 },
476 }]
477 }
478 };
479
480 messages.push(open_ai::RequestMessage::Tool {
481 content: content.into(),
482 tool_call_id: tool_result.tool_use_id.to_string(),
483 });
484 }
485 }
486 }
487 }
488
489 open_ai::Request {
490 model: model_id.into(),
491 messages,
492 stream,
493 stop: request.stop,
494 temperature: request.temperature.or(Some(1.0)),
495 max_completion_tokens: max_output_tokens,
496 parallel_tool_calls: if supports_parallel_tool_calls && !request.tools.is_empty() {
497 // Disable parallel tool calls, as the Agent currently expects a maximum of one per turn.
498 Some(false)
499 } else {
500 None
501 },
502 prompt_cache_key: if supports_prompt_cache_key {
503 request.thread_id
504 } else {
505 None
506 },
507 tools: request
508 .tools
509 .into_iter()
510 .map(|tool| open_ai::ToolDefinition::Function {
511 function: open_ai::FunctionDefinition {
512 name: tool.name,
513 description: Some(tool.description),
514 parameters: Some(tool.input_schema),
515 },
516 })
517 .collect(),
518 tool_choice: request.tool_choice.map(|choice| match choice {
519 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
520 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
521 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
522 }),
523 reasoning_effort,
524 }
525}
526
527pub fn into_open_ai_response(
528 request: LanguageModelRequest,
529 model_id: &str,
530 supports_parallel_tool_calls: bool,
531 supports_prompt_cache_key: bool,
532 max_output_tokens: Option<u64>,
533 reasoning_effort: Option<ReasoningEffort>,
534) -> ResponseRequest {
535 let stream = !model_id.starts_with("o1-");
536
537 let LanguageModelRequest {
538 thread_id,
539 prompt_id: _,
540 intent: _,
541 messages,
542 tools,
543 tool_choice,
544 stop: _,
545 temperature,
546 thinking_allowed: _,
547 } = request;
548
549 let mut input_items = Vec::new();
550 for (index, message) in messages.into_iter().enumerate() {
551 append_message_to_response_items(message, index, &mut input_items);
552 }
553
554 let tools: Vec<_> = tools
555 .into_iter()
556 .map(|tool| open_ai::responses::ToolDefinition::Function {
557 name: tool.name,
558 description: Some(tool.description),
559 parameters: Some(tool.input_schema),
560 strict: None,
561 })
562 .collect();
563
564 ResponseRequest {
565 model: model_id.into(),
566 input: input_items,
567 stream,
568 temperature,
569 top_p: None,
570 max_output_tokens,
571 parallel_tool_calls: if tools.is_empty() {
572 None
573 } else {
574 Some(supports_parallel_tool_calls)
575 },
576 tool_choice: tool_choice.map(|choice| match choice {
577 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
578 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
579 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
580 }),
581 tools,
582 prompt_cache_key: if supports_prompt_cache_key {
583 thread_id
584 } else {
585 None
586 },
587 reasoning: reasoning_effort.map(|effort| open_ai::responses::ReasoningConfig { effort }),
588 }
589}
590
591fn append_message_to_response_items(
592 message: LanguageModelRequestMessage,
593 index: usize,
594 input_items: &mut Vec<ResponseInputItem>,
595) {
596 let mut content_parts: Vec<ResponseInputContent> = Vec::new();
597
598 for content in message.content {
599 match content {
600 MessageContent::Text(text) => {
601 push_response_text_part(&message.role, text, &mut content_parts);
602 }
603 MessageContent::Thinking { text, .. } => {
604 push_response_text_part(&message.role, text, &mut content_parts);
605 }
606 MessageContent::RedactedThinking(_) => {}
607 MessageContent::Image(image) => {
608 push_response_image_part(&message.role, image, &mut content_parts);
609 }
610 MessageContent::ToolUse(tool_use) => {
611 flush_response_parts(&message.role, index, &mut content_parts, input_items);
612 let call_id = tool_use.id.to_string();
613 input_items.push(ResponseInputItem::FunctionCall(ResponseFunctionCallItem {
614 call_id,
615 name: tool_use.name.to_string(),
616 arguments: tool_use.raw_input,
617 }));
618 }
619 MessageContent::ToolResult(tool_result) => {
620 flush_response_parts(&message.role, index, &mut content_parts, input_items);
621 input_items.push(ResponseInputItem::FunctionCallOutput(
622 ResponseFunctionCallOutputItem {
623 call_id: tool_result.tool_use_id.to_string(),
624 output: tool_result_output(&tool_result),
625 },
626 ));
627 }
628 }
629 }
630
631 flush_response_parts(&message.role, index, &mut content_parts, input_items);
632}
633
634fn push_response_text_part(
635 role: &Role,
636 text: impl Into<String>,
637 parts: &mut Vec<ResponseInputContent>,
638) {
639 let text = text.into();
640 if text.trim().is_empty() {
641 return;
642 }
643
644 match role {
645 Role::Assistant => parts.push(ResponseInputContent::OutputText {
646 text,
647 annotations: Vec::new(),
648 }),
649 _ => parts.push(ResponseInputContent::Text { text }),
650 }
651}
652
653fn push_response_image_part(
654 role: &Role,
655 image: LanguageModelImage,
656 parts: &mut Vec<ResponseInputContent>,
657) {
658 match role {
659 Role::Assistant => parts.push(ResponseInputContent::OutputText {
660 text: "[image omitted]".to_string(),
661 annotations: Vec::new(),
662 }),
663 _ => parts.push(ResponseInputContent::Image {
664 image_url: image.to_base64_url(),
665 }),
666 }
667}
668
669fn flush_response_parts(
670 role: &Role,
671 _index: usize,
672 parts: &mut Vec<ResponseInputContent>,
673 input_items: &mut Vec<ResponseInputItem>,
674) {
675 if parts.is_empty() {
676 return;
677 }
678
679 let item = ResponseInputItem::Message(ResponseMessageItem {
680 role: match role {
681 Role::User => open_ai::Role::User,
682 Role::Assistant => open_ai::Role::Assistant,
683 Role::System => open_ai::Role::System,
684 },
685 content: parts.clone(),
686 });
687
688 input_items.push(item);
689 parts.clear();
690}
691
692fn tool_result_output(result: &LanguageModelToolResult) -> String {
693 if let Some(output) = &result.output {
694 match output {
695 serde_json::Value::String(text) => text.clone(),
696 serde_json::Value::Null => String::new(),
697 _ => output.to_string(),
698 }
699 } else {
700 match &result.content {
701 LanguageModelToolResultContent::Text(text) => text.to_string(),
702 LanguageModelToolResultContent::Image(image) => image.to_base64_url(),
703 }
704 }
705}
706
707fn add_message_content_part(
708 new_part: open_ai::MessagePart,
709 role: Role,
710 messages: &mut Vec<open_ai::RequestMessage>,
711) {
712 match (role, messages.last_mut()) {
713 (Role::User, Some(open_ai::RequestMessage::User { content }))
714 | (
715 Role::Assistant,
716 Some(open_ai::RequestMessage::Assistant {
717 content: Some(content),
718 ..
719 }),
720 )
721 | (Role::System, Some(open_ai::RequestMessage::System { content, .. })) => {
722 content.push_part(new_part);
723 }
724 _ => {
725 messages.push(match role {
726 Role::User => open_ai::RequestMessage::User {
727 content: open_ai::MessageContent::from(vec![new_part]),
728 },
729 Role::Assistant => open_ai::RequestMessage::Assistant {
730 content: Some(open_ai::MessageContent::from(vec![new_part])),
731 tool_calls: Vec::new(),
732 },
733 Role::System => open_ai::RequestMessage::System {
734 content: open_ai::MessageContent::from(vec![new_part]),
735 },
736 });
737 }
738 }
739}
740
741pub struct OpenAiEventMapper {
742 tool_calls_by_index: HashMap<usize, RawToolCall>,
743}
744
745impl OpenAiEventMapper {
746 pub fn new() -> Self {
747 Self {
748 tool_calls_by_index: HashMap::default(),
749 }
750 }
751
752 pub fn map_stream(
753 mut self,
754 events: Pin<Box<dyn Send + Stream<Item = Result<ResponseStreamEvent>>>>,
755 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
756 {
757 events.flat_map(move |event| {
758 futures::stream::iter(match event {
759 Ok(event) => self.map_event(event),
760 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
761 })
762 })
763 }
764
765 pub fn map_event(
766 &mut self,
767 event: ResponseStreamEvent,
768 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
769 let mut events = Vec::new();
770 if let Some(usage) = event.usage {
771 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
772 input_tokens: usage.prompt_tokens,
773 output_tokens: usage.completion_tokens,
774 cache_creation_input_tokens: 0,
775 cache_read_input_tokens: 0,
776 })));
777 }
778
779 let Some(choice) = event.choices.first() else {
780 return events;
781 };
782
783 if let Some(delta) = choice.delta.as_ref() {
784 if let Some(content) = delta.content.clone() {
785 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
786 }
787
788 if let Some(tool_calls) = delta.tool_calls.as_ref() {
789 for tool_call in tool_calls {
790 let entry = self.tool_calls_by_index.entry(tool_call.index).or_default();
791
792 if let Some(tool_id) = tool_call.id.clone() {
793 entry.id = tool_id;
794 }
795
796 if let Some(function) = tool_call.function.as_ref() {
797 if let Some(name) = function.name.clone() {
798 entry.name = name;
799 }
800
801 if let Some(arguments) = function.arguments.clone() {
802 entry.arguments.push_str(&arguments);
803 }
804 }
805 }
806 }
807 }
808
809 match choice.finish_reason.as_deref() {
810 Some("stop") => {
811 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
812 }
813 Some("tool_calls") => {
814 events.extend(self.tool_calls_by_index.drain().map(|(_, tool_call)| {
815 match serde_json::Value::from_str(&tool_call.arguments) {
816 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
817 LanguageModelToolUse {
818 id: tool_call.id.clone().into(),
819 name: tool_call.name.as_str().into(),
820 is_input_complete: true,
821 input,
822 raw_input: tool_call.arguments.clone(),
823 thought_signature: None,
824 },
825 )),
826 Err(error) => Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
827 id: tool_call.id.into(),
828 tool_name: tool_call.name.into(),
829 raw_input: tool_call.arguments.clone().into(),
830 json_parse_error: error.to_string(),
831 }),
832 }
833 }));
834
835 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
836 }
837 Some(stop_reason) => {
838 log::error!("Unexpected OpenAI stop_reason: {stop_reason:?}",);
839 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
840 }
841 None => {}
842 }
843
844 events
845 }
846}
847
848#[derive(Default)]
849struct RawToolCall {
850 id: String,
851 name: String,
852 arguments: String,
853}
854
855pub struct OpenAiResponseEventMapper {
856 function_calls_by_item: HashMap<String, PendingResponseFunctionCall>,
857 pending_stop_reason: Option<StopReason>,
858}
859
860#[derive(Default)]
861struct PendingResponseFunctionCall {
862 call_id: String,
863 name: Arc<str>,
864 arguments: String,
865}
866
867impl OpenAiResponseEventMapper {
868 pub fn new() -> Self {
869 Self {
870 function_calls_by_item: HashMap::default(),
871 pending_stop_reason: None,
872 }
873 }
874
875 pub fn map_stream(
876 mut self,
877 events: Pin<Box<dyn Send + Stream<Item = Result<ResponsesStreamEvent>>>>,
878 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
879 {
880 events.flat_map(move |event| {
881 futures::stream::iter(match event {
882 Ok(event) => self.map_event(event),
883 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
884 })
885 })
886 }
887
888 pub fn map_event(
889 &mut self,
890 event: ResponsesStreamEvent,
891 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
892 match event {
893 ResponsesStreamEvent::OutputItemAdded { item, .. } => {
894 let mut events = Vec::new();
895
896 match &item {
897 ResponseOutputItem::Message(message) => {
898 if let Some(id) = &message.id {
899 events.push(Ok(LanguageModelCompletionEvent::StartMessage {
900 message_id: id.clone(),
901 }));
902 }
903 }
904 ResponseOutputItem::FunctionCall(function_call) => {
905 if let Some(item_id) = function_call.id.clone() {
906 let call_id = function_call
907 .call_id
908 .clone()
909 .or_else(|| function_call.id.clone())
910 .unwrap_or_else(|| item_id.clone());
911 let entry = PendingResponseFunctionCall {
912 call_id,
913 name: Arc::<str>::from(
914 function_call.name.clone().unwrap_or_default(),
915 ),
916 arguments: function_call.arguments.clone(),
917 };
918 self.function_calls_by_item.insert(item_id, entry);
919 }
920 }
921 ResponseOutputItem::Unknown => {}
922 }
923 events
924 }
925 ResponsesStreamEvent::OutputTextDelta { delta, .. } => {
926 if delta.is_empty() {
927 Vec::new()
928 } else {
929 vec![Ok(LanguageModelCompletionEvent::Text(delta))]
930 }
931 }
932 ResponsesStreamEvent::FunctionCallArgumentsDelta { item_id, delta, .. } => {
933 if let Some(entry) = self.function_calls_by_item.get_mut(&item_id) {
934 entry.arguments.push_str(&delta);
935 }
936 Vec::new()
937 }
938 ResponsesStreamEvent::FunctionCallArgumentsDone {
939 item_id, arguments, ..
940 } => {
941 if let Some(mut entry) = self.function_calls_by_item.remove(&item_id) {
942 if !arguments.is_empty() {
943 entry.arguments = arguments;
944 }
945 let raw_input = entry.arguments.clone();
946 self.pending_stop_reason = Some(StopReason::ToolUse);
947 match serde_json::from_str::<serde_json::Value>(&entry.arguments) {
948 Ok(input) => {
949 vec![Ok(LanguageModelCompletionEvent::ToolUse(
950 LanguageModelToolUse {
951 id: LanguageModelToolUseId::from(entry.call_id.clone()),
952 name: entry.name.clone(),
953 is_input_complete: true,
954 input,
955 raw_input,
956 thought_signature: None,
957 },
958 ))]
959 }
960 Err(error) => {
961 vec![Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
962 id: LanguageModelToolUseId::from(entry.call_id.clone()),
963 tool_name: entry.name.clone(),
964 raw_input: Arc::<str>::from(raw_input),
965 json_parse_error: error.to_string(),
966 })]
967 }
968 }
969 } else {
970 Vec::new()
971 }
972 }
973 ResponsesStreamEvent::Completed { response } => {
974 self.handle_completion(response, StopReason::EndTurn)
975 }
976 ResponsesStreamEvent::Incomplete { response } => {
977 let reason = response
978 .status_details
979 .as_ref()
980 .and_then(|details| details.reason.as_deref());
981 let stop_reason = match reason {
982 Some("max_output_tokens") => StopReason::MaxTokens,
983 Some("content_filter") => {
984 self.pending_stop_reason = Some(StopReason::Refusal);
985 StopReason::Refusal
986 }
987 _ => self
988 .pending_stop_reason
989 .take()
990 .unwrap_or(StopReason::EndTurn),
991 };
992
993 let mut events = Vec::new();
994 if self.pending_stop_reason.is_none() {
995 events.extend(self.emit_tool_calls_from_output(&response.output));
996 }
997 if let Some(usage) = response.usage.as_ref() {
998 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
999 token_usage_from_response_usage(usage),
1000 )));
1001 }
1002 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1003 events
1004 }
1005 ResponsesStreamEvent::Failed { response } => {
1006 let message = response
1007 .status_details
1008 .and_then(|details| details.error)
1009 .map(|error| error.to_string())
1010 .unwrap_or_else(|| "response failed".to_string());
1011 vec![Err(LanguageModelCompletionError::Other(anyhow!(message)))]
1012 }
1013 ResponsesStreamEvent::Error { error }
1014 | ResponsesStreamEvent::GenericError { error } => {
1015 vec![Err(LanguageModelCompletionError::Other(anyhow!(format!(
1016 "{error:?}"
1017 ))))]
1018 }
1019 ResponsesStreamEvent::OutputTextDone { .. } => Vec::new(),
1020 ResponsesStreamEvent::OutputItemDone { .. }
1021 | ResponsesStreamEvent::ContentPartAdded { .. }
1022 | ResponsesStreamEvent::ContentPartDone { .. }
1023 | ResponsesStreamEvent::Created { .. }
1024 | ResponsesStreamEvent::InProgress { .. }
1025 | ResponsesStreamEvent::Unknown => Vec::new(),
1026 }
1027 }
1028
1029 fn handle_completion(
1030 &mut self,
1031 response: ResponsesSummary,
1032 default_reason: StopReason,
1033 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1034 let mut events = Vec::new();
1035
1036 if self.pending_stop_reason.is_none() {
1037 events.extend(self.emit_tool_calls_from_output(&response.output));
1038 }
1039
1040 if let Some(usage) = response.usage.as_ref() {
1041 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1042 token_usage_from_response_usage(usage),
1043 )));
1044 }
1045
1046 let stop_reason = self.pending_stop_reason.take().unwrap_or(default_reason);
1047 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1048 events
1049 }
1050
1051 fn emit_tool_calls_from_output(
1052 &mut self,
1053 output: &[ResponseOutputItem],
1054 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1055 let mut events = Vec::new();
1056 for item in output {
1057 if let ResponseOutputItem::FunctionCall(function_call) = item {
1058 let Some(call_id) = function_call
1059 .call_id
1060 .clone()
1061 .or_else(|| function_call.id.clone())
1062 else {
1063 log::error!(
1064 "Function call item missing both call_id and id: {:?}",
1065 function_call
1066 );
1067 continue;
1068 };
1069 let name: Arc<str> = Arc::from(function_call.name.clone().unwrap_or_default());
1070 let arguments = &function_call.arguments;
1071 if !arguments.is_empty() {
1072 self.pending_stop_reason = Some(StopReason::ToolUse);
1073 match serde_json::from_str::<serde_json::Value>(arguments) {
1074 Ok(input) => {
1075 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
1076 LanguageModelToolUse {
1077 id: LanguageModelToolUseId::from(call_id.clone()),
1078 name: name.clone(),
1079 is_input_complete: true,
1080 input,
1081 raw_input: arguments.clone(),
1082 thought_signature: None,
1083 },
1084 )));
1085 }
1086 Err(error) => {
1087 events.push(Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1088 id: LanguageModelToolUseId::from(call_id.clone()),
1089 tool_name: name.clone(),
1090 raw_input: Arc::<str>::from(arguments.clone()),
1091 json_parse_error: error.to_string(),
1092 }));
1093 }
1094 }
1095 }
1096 }
1097 }
1098 events
1099 }
1100}
1101
1102fn token_usage_from_response_usage(usage: &ResponsesUsage) -> TokenUsage {
1103 TokenUsage {
1104 input_tokens: usage.input_tokens.unwrap_or_default(),
1105 output_tokens: usage.output_tokens.unwrap_or_default(),
1106 cache_creation_input_tokens: 0,
1107 cache_read_input_tokens: 0,
1108 }
1109}
1110
1111pub(crate) fn collect_tiktoken_messages(
1112 request: LanguageModelRequest,
1113) -> Vec<tiktoken_rs::ChatCompletionRequestMessage> {
1114 request
1115 .messages
1116 .into_iter()
1117 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
1118 role: match message.role {
1119 Role::User => "user".into(),
1120 Role::Assistant => "assistant".into(),
1121 Role::System => "system".into(),
1122 },
1123 content: Some(message.string_contents()),
1124 name: None,
1125 function_call: None,
1126 })
1127 .collect::<Vec<_>>()
1128}
1129
1130pub fn count_open_ai_tokens(
1131 request: LanguageModelRequest,
1132 model: Model,
1133 cx: &App,
1134) -> BoxFuture<'static, Result<u64>> {
1135 cx.background_spawn(async move {
1136 let messages = collect_tiktoken_messages(request);
1137 match model {
1138 Model::Custom { max_tokens, .. } => {
1139 let model = if max_tokens >= 100_000 {
1140 // If the max tokens is 100k or more, it is likely the o200k_base tokenizer from gpt4o
1141 "gpt-4o"
1142 } else {
1143 // Otherwise fallback to gpt-4, since only cl100k_base and o200k_base are
1144 // supported with this tiktoken method
1145 "gpt-4"
1146 };
1147 tiktoken_rs::num_tokens_from_messages(model, &messages)
1148 }
1149 // Currently supported by tiktoken_rs
1150 // Sometimes tiktoken-rs is behind on model support. If that is the case, make a new branch
1151 // arm with an override. We enumerate all supported models here so that we can check if new
1152 // models are supported yet or not.
1153 Model::ThreePointFiveTurbo
1154 | Model::Four
1155 | Model::FourTurbo
1156 | Model::FourOmni
1157 | Model::FourOmniMini
1158 | Model::FourPointOne
1159 | Model::FourPointOneMini
1160 | Model::FourPointOneNano
1161 | Model::O1
1162 | Model::O3
1163 | Model::O3Mini
1164 | Model::O4Mini
1165 | Model::Five
1166 | Model::FiveCodex
1167 | Model::FiveMini
1168 | Model::FiveNano => tiktoken_rs::num_tokens_from_messages(model.id(), &messages),
1169 // GPT-5.1 and 5.2 don't have dedicated tiktoken support; use gpt-5 tokenizer
1170 Model::FivePointOne | Model::FivePointTwo => {
1171 tiktoken_rs::num_tokens_from_messages("gpt-5", &messages)
1172 }
1173 }
1174 .map(|tokens| tokens as u64)
1175 })
1176 .boxed()
1177}
1178
1179struct ConfigurationView {
1180 api_key_editor: Entity<InputField>,
1181 state: Entity<State>,
1182 load_credentials_task: Option<Task<()>>,
1183}
1184
1185impl ConfigurationView {
1186 fn new(state: Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
1187 let api_key_editor = cx.new(|cx| {
1188 InputField::new(
1189 window,
1190 cx,
1191 "sk-000000000000000000000000000000000000000000000000",
1192 )
1193 });
1194
1195 cx.observe(&state, |_, _, cx| {
1196 cx.notify();
1197 })
1198 .detach();
1199
1200 let load_credentials_task = Some(cx.spawn_in(window, {
1201 let state = state.clone();
1202 async move |this, cx| {
1203 if let Some(task) = Some(state.update(cx, |state, cx| state.authenticate(cx))) {
1204 // We don't log an error, because "not signed in" is also an error.
1205 let _ = task.await;
1206 }
1207 this.update(cx, |this, cx| {
1208 this.load_credentials_task = None;
1209 cx.notify();
1210 })
1211 .log_err();
1212 }
1213 }));
1214
1215 Self {
1216 api_key_editor,
1217 state,
1218 load_credentials_task,
1219 }
1220 }
1221
1222 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
1223 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
1224 if api_key.is_empty() {
1225 return;
1226 }
1227
1228 // url changes can cause the editor to be displayed again
1229 self.api_key_editor
1230 .update(cx, |editor, cx| editor.set_text("", window, cx));
1231
1232 let state = self.state.clone();
1233 cx.spawn_in(window, async move |_, cx| {
1234 state
1235 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
1236 .await
1237 })
1238 .detach_and_log_err(cx);
1239 }
1240
1241 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
1242 self.api_key_editor
1243 .update(cx, |input, cx| input.set_text("", window, cx));
1244
1245 let state = self.state.clone();
1246 cx.spawn_in(window, async move |_, cx| {
1247 state
1248 .update(cx, |state, cx| state.set_api_key(None, cx))
1249 .await
1250 })
1251 .detach_and_log_err(cx);
1252 }
1253
1254 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
1255 !self.state.read(cx).is_authenticated()
1256 }
1257}
1258
1259impl Render for ConfigurationView {
1260 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1261 let env_var_set = self.state.read(cx).api_key_state.is_from_env_var();
1262 let configured_card_label = if env_var_set {
1263 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable")
1264 } else {
1265 let api_url = OpenAiLanguageModelProvider::api_url(cx);
1266 if api_url == OPEN_AI_API_URL {
1267 "API key configured".to_string()
1268 } else {
1269 format!("API key configured for {}", api_url)
1270 }
1271 };
1272
1273 let api_key_section = if self.should_render_editor(cx) {
1274 v_flex()
1275 .on_action(cx.listener(Self::save_api_key))
1276 .child(Label::new("To use Zed's agent with OpenAI, you need to add an API key. Follow these steps:"))
1277 .child(
1278 List::new()
1279 .child(
1280 ListBulletItem::new("")
1281 .child(Label::new("Create one by visiting"))
1282 .child(ButtonLink::new("OpenAI's console", "https://platform.openai.com/api-keys"))
1283 )
1284 .child(
1285 ListBulletItem::new("Ensure your OpenAI account has credits")
1286 )
1287 .child(
1288 ListBulletItem::new("Paste your API key below and hit enter to start using the agent")
1289 ),
1290 )
1291 .child(self.api_key_editor.clone())
1292 .child(
1293 Label::new(format!(
1294 "You can also set the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed."
1295 ))
1296 .size(LabelSize::Small)
1297 .color(Color::Muted),
1298 )
1299 .child(
1300 Label::new(
1301 "Note that having a subscription for another service like GitHub Copilot won't work.",
1302 )
1303 .size(LabelSize::Small).color(Color::Muted),
1304 )
1305 .into_any_element()
1306 } else {
1307 ConfiguredApiCard::new(configured_card_label)
1308 .disabled(env_var_set)
1309 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
1310 .when(env_var_set, |this| {
1311 this.tooltip_label(format!("To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable."))
1312 })
1313 .into_any_element()
1314 };
1315
1316 let compatible_api_section = h_flex()
1317 .mt_1p5()
1318 .gap_0p5()
1319 .flex_wrap()
1320 .when(self.should_render_editor(cx), |this| {
1321 this.pt_1p5()
1322 .border_t_1()
1323 .border_color(cx.theme().colors().border_variant)
1324 })
1325 .child(
1326 h_flex()
1327 .gap_2()
1328 .child(
1329 Icon::new(IconName::Info)
1330 .size(IconSize::XSmall)
1331 .color(Color::Muted),
1332 )
1333 .child(Label::new("Zed also supports OpenAI-compatible models.")),
1334 )
1335 .child(
1336 Button::new("docs", "Learn More")
1337 .icon(IconName::ArrowUpRight)
1338 .icon_size(IconSize::Small)
1339 .icon_color(Color::Muted)
1340 .on_click(move |_, _window, cx| {
1341 cx.open_url("https://zed.dev/docs/ai/llm-providers#openai-api-compatible")
1342 }),
1343 );
1344
1345 if self.load_credentials_task.is_some() {
1346 div().child(Label::new("Loading credentials…")).into_any()
1347 } else {
1348 v_flex()
1349 .size_full()
1350 .child(api_key_section)
1351 .child(compatible_api_section)
1352 .into_any()
1353 }
1354 }
1355}
1356
1357#[cfg(test)]
1358mod tests {
1359 use futures::{StreamExt, executor::block_on};
1360 use gpui::TestAppContext;
1361 use language_model::{LanguageModelRequestMessage, LanguageModelRequestTool};
1362 use open_ai::responses::{
1363 ResponseFunctionToolCall, ResponseOutputItem, ResponseOutputMessage, ResponseStatusDetails,
1364 ResponseSummary, ResponseUsage, StreamEvent as ResponsesStreamEvent,
1365 };
1366 use pretty_assertions::assert_eq;
1367 use serde_json::json;
1368
1369 use super::*;
1370
1371 fn map_response_events(events: Vec<ResponsesStreamEvent>) -> Vec<LanguageModelCompletionEvent> {
1372 block_on(async {
1373 OpenAiResponseEventMapper::new()
1374 .map_stream(Box::pin(futures::stream::iter(events.into_iter().map(Ok))))
1375 .collect::<Vec<_>>()
1376 .await
1377 .into_iter()
1378 .map(Result::unwrap)
1379 .collect()
1380 })
1381 }
1382
1383 fn response_item_message(id: &str) -> ResponseOutputItem {
1384 ResponseOutputItem::Message(ResponseOutputMessage {
1385 id: Some(id.to_string()),
1386 role: Some("assistant".to_string()),
1387 status: Some("in_progress".to_string()),
1388 content: vec![],
1389 })
1390 }
1391
1392 fn response_item_function_call(id: &str, args: Option<&str>) -> ResponseOutputItem {
1393 ResponseOutputItem::FunctionCall(ResponseFunctionToolCall {
1394 id: Some(id.to_string()),
1395 status: Some("in_progress".to_string()),
1396 name: Some("get_weather".to_string()),
1397 call_id: Some("call_123".to_string()),
1398 arguments: args.map(|s| s.to_string()).unwrap_or_default(),
1399 })
1400 }
1401
1402 #[gpui::test]
1403 fn tiktoken_rs_support(cx: &TestAppContext) {
1404 let request = LanguageModelRequest {
1405 thread_id: None,
1406 prompt_id: None,
1407 intent: None,
1408 messages: vec![LanguageModelRequestMessage {
1409 role: Role::User,
1410 content: vec![MessageContent::Text("message".into())],
1411 cache: false,
1412 reasoning_details: None,
1413 }],
1414 tools: vec![],
1415 tool_choice: None,
1416 stop: vec![],
1417 temperature: None,
1418 thinking_allowed: true,
1419 };
1420
1421 // Validate that all models are supported by tiktoken-rs
1422 for model in Model::iter() {
1423 let count = cx
1424 .foreground_executor()
1425 .block_on(count_open_ai_tokens(
1426 request.clone(),
1427 model,
1428 &cx.app.borrow(),
1429 ))
1430 .unwrap();
1431 assert!(count > 0);
1432 }
1433 }
1434
1435 #[test]
1436 fn responses_stream_maps_text_and_usage() {
1437 let events = vec![
1438 ResponsesStreamEvent::OutputItemAdded {
1439 output_index: 0,
1440 sequence_number: None,
1441 item: response_item_message("msg_123"),
1442 },
1443 ResponsesStreamEvent::OutputTextDelta {
1444 item_id: "msg_123".into(),
1445 output_index: 0,
1446 content_index: Some(0),
1447 delta: "Hello".into(),
1448 },
1449 ResponsesStreamEvent::Completed {
1450 response: ResponseSummary {
1451 usage: Some(ResponseUsage {
1452 input_tokens: Some(5),
1453 output_tokens: Some(3),
1454 total_tokens: Some(8),
1455 }),
1456 ..Default::default()
1457 },
1458 },
1459 ];
1460
1461 let mapped = map_response_events(events);
1462 assert!(matches!(
1463 mapped[0],
1464 LanguageModelCompletionEvent::StartMessage { ref message_id } if message_id == "msg_123"
1465 ));
1466 assert!(matches!(
1467 mapped[1],
1468 LanguageModelCompletionEvent::Text(ref text) if text == "Hello"
1469 ));
1470 assert!(matches!(
1471 mapped[2],
1472 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1473 input_tokens: 5,
1474 output_tokens: 3,
1475 ..
1476 })
1477 ));
1478 assert!(matches!(
1479 mapped[3],
1480 LanguageModelCompletionEvent::Stop(StopReason::EndTurn)
1481 ));
1482 }
1483
1484 #[test]
1485 fn into_open_ai_response_builds_complete_payload() {
1486 let tool_call_id = LanguageModelToolUseId::from("call-42");
1487 let tool_input = json!({ "city": "Boston" });
1488 let tool_arguments = serde_json::to_string(&tool_input).unwrap();
1489 let tool_use = LanguageModelToolUse {
1490 id: tool_call_id.clone(),
1491 name: Arc::from("get_weather"),
1492 raw_input: tool_arguments.clone(),
1493 input: tool_input,
1494 is_input_complete: true,
1495 thought_signature: None,
1496 };
1497 let tool_result = LanguageModelToolResult {
1498 tool_use_id: tool_call_id,
1499 tool_name: Arc::from("get_weather"),
1500 is_error: false,
1501 content: LanguageModelToolResultContent::Text(Arc::from("Sunny")),
1502 output: Some(json!({ "forecast": "Sunny" })),
1503 };
1504 let user_image = LanguageModelImage {
1505 source: SharedString::from("aGVsbG8="),
1506 size: None,
1507 };
1508 let expected_image_url = user_image.to_base64_url();
1509
1510 let request = LanguageModelRequest {
1511 thread_id: Some("thread-123".into()),
1512 prompt_id: None,
1513 intent: None,
1514 messages: vec![
1515 LanguageModelRequestMessage {
1516 role: Role::System,
1517 content: vec![MessageContent::Text("System context".into())],
1518 cache: false,
1519 reasoning_details: None,
1520 },
1521 LanguageModelRequestMessage {
1522 role: Role::User,
1523 content: vec![
1524 MessageContent::Text("Please check the weather.".into()),
1525 MessageContent::Image(user_image),
1526 ],
1527 cache: false,
1528 reasoning_details: None,
1529 },
1530 LanguageModelRequestMessage {
1531 role: Role::Assistant,
1532 content: vec![
1533 MessageContent::Text("Looking that up.".into()),
1534 MessageContent::ToolUse(tool_use),
1535 ],
1536 cache: false,
1537 reasoning_details: None,
1538 },
1539 LanguageModelRequestMessage {
1540 role: Role::Assistant,
1541 content: vec![MessageContent::ToolResult(tool_result)],
1542 cache: false,
1543 reasoning_details: None,
1544 },
1545 ],
1546 tools: vec![LanguageModelRequestTool {
1547 name: "get_weather".into(),
1548 description: "Fetches the weather".into(),
1549 input_schema: json!({ "type": "object" }),
1550 }],
1551 tool_choice: Some(LanguageModelToolChoice::Any),
1552 stop: vec!["<STOP>".into()],
1553 temperature: None,
1554 thinking_allowed: false,
1555 };
1556
1557 let response = into_open_ai_response(
1558 request,
1559 "custom-model",
1560 true,
1561 true,
1562 Some(2048),
1563 Some(ReasoningEffort::Low),
1564 );
1565
1566 let serialized = serde_json::to_value(&response).unwrap();
1567 let expected = json!({
1568 "model": "custom-model",
1569 "input": [
1570 {
1571 "type": "message",
1572 "role": "system",
1573 "content": [
1574 { "type": "input_text", "text": "System context" }
1575 ]
1576 },
1577 {
1578 "type": "message",
1579 "role": "user",
1580 "content": [
1581 { "type": "input_text", "text": "Please check the weather." },
1582 { "type": "input_image", "image_url": expected_image_url }
1583 ]
1584 },
1585 {
1586 "type": "message",
1587 "role": "assistant",
1588 "content": [
1589 { "type": "output_text", "text": "Looking that up.", "annotations": [] }
1590 ]
1591 },
1592 {
1593 "type": "function_call",
1594 "call_id": "call-42",
1595 "name": "get_weather",
1596 "arguments": tool_arguments
1597 },
1598 {
1599 "type": "function_call_output",
1600 "call_id": "call-42",
1601 "output": "{\"forecast\":\"Sunny\"}"
1602 }
1603 ],
1604 "stream": true,
1605 "max_output_tokens": 2048,
1606 "parallel_tool_calls": true,
1607 "tool_choice": "required",
1608 "tools": [
1609 {
1610 "type": "function",
1611 "name": "get_weather",
1612 "description": "Fetches the weather",
1613 "parameters": { "type": "object" }
1614 }
1615 ],
1616 "prompt_cache_key": "thread-123",
1617 "reasoning": { "effort": "low" }
1618 });
1619
1620 assert_eq!(serialized, expected);
1621 }
1622
1623 #[test]
1624 fn responses_stream_maps_tool_calls() {
1625 let events = vec![
1626 ResponsesStreamEvent::OutputItemAdded {
1627 output_index: 0,
1628 sequence_number: None,
1629 item: response_item_function_call("item_fn", Some("{\"city\":\"Bos")),
1630 },
1631 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1632 item_id: "item_fn".into(),
1633 output_index: 0,
1634 delta: "ton\"}".into(),
1635 sequence_number: None,
1636 },
1637 ResponsesStreamEvent::FunctionCallArgumentsDone {
1638 item_id: "item_fn".into(),
1639 output_index: 0,
1640 arguments: "{\"city\":\"Boston\"}".into(),
1641 sequence_number: None,
1642 },
1643 ResponsesStreamEvent::Completed {
1644 response: ResponseSummary::default(),
1645 },
1646 ];
1647
1648 let mapped = map_response_events(events);
1649 assert!(matches!(
1650 mapped[0],
1651 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1652 ref id,
1653 ref name,
1654 ref raw_input,
1655 ..
1656 }) if id.to_string() == "call_123"
1657 && name.as_ref() == "get_weather"
1658 && raw_input == "{\"city\":\"Boston\"}"
1659 ));
1660 assert!(matches!(
1661 mapped[1],
1662 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1663 ));
1664 }
1665
1666 #[test]
1667 fn responses_stream_uses_max_tokens_stop_reason() {
1668 let events = vec![ResponsesStreamEvent::Incomplete {
1669 response: ResponseSummary {
1670 status_details: Some(ResponseStatusDetails {
1671 reason: Some("max_output_tokens".into()),
1672 r#type: Some("incomplete".into()),
1673 error: None,
1674 }),
1675 usage: Some(ResponseUsage {
1676 input_tokens: Some(10),
1677 output_tokens: Some(20),
1678 total_tokens: Some(30),
1679 }),
1680 ..Default::default()
1681 },
1682 }];
1683
1684 let mapped = map_response_events(events);
1685 assert!(matches!(
1686 mapped[0],
1687 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1688 input_tokens: 10,
1689 output_tokens: 20,
1690 ..
1691 })
1692 ));
1693 assert!(matches!(
1694 mapped[1],
1695 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1696 ));
1697 }
1698
1699 #[test]
1700 fn responses_stream_handles_multiple_tool_calls() {
1701 let events = vec![
1702 ResponsesStreamEvent::OutputItemAdded {
1703 output_index: 0,
1704 sequence_number: None,
1705 item: response_item_function_call("item_fn1", Some("{\"city\":\"NYC\"}")),
1706 },
1707 ResponsesStreamEvent::FunctionCallArgumentsDone {
1708 item_id: "item_fn1".into(),
1709 output_index: 0,
1710 arguments: "{\"city\":\"NYC\"}".into(),
1711 sequence_number: None,
1712 },
1713 ResponsesStreamEvent::OutputItemAdded {
1714 output_index: 1,
1715 sequence_number: None,
1716 item: response_item_function_call("item_fn2", Some("{\"city\":\"LA\"}")),
1717 },
1718 ResponsesStreamEvent::FunctionCallArgumentsDone {
1719 item_id: "item_fn2".into(),
1720 output_index: 1,
1721 arguments: "{\"city\":\"LA\"}".into(),
1722 sequence_number: None,
1723 },
1724 ResponsesStreamEvent::Completed {
1725 response: ResponseSummary::default(),
1726 },
1727 ];
1728
1729 let mapped = map_response_events(events);
1730 assert_eq!(mapped.len(), 3);
1731 assert!(matches!(
1732 mapped[0],
1733 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1734 if raw_input == "{\"city\":\"NYC\"}"
1735 ));
1736 assert!(matches!(
1737 mapped[1],
1738 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1739 if raw_input == "{\"city\":\"LA\"}"
1740 ));
1741 assert!(matches!(
1742 mapped[2],
1743 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1744 ));
1745 }
1746
1747 #[test]
1748 fn responses_stream_handles_mixed_text_and_tool_calls() {
1749 let events = vec![
1750 ResponsesStreamEvent::OutputItemAdded {
1751 output_index: 0,
1752 sequence_number: None,
1753 item: response_item_message("msg_123"),
1754 },
1755 ResponsesStreamEvent::OutputTextDelta {
1756 item_id: "msg_123".into(),
1757 output_index: 0,
1758 content_index: Some(0),
1759 delta: "Let me check that".into(),
1760 },
1761 ResponsesStreamEvent::OutputItemAdded {
1762 output_index: 1,
1763 sequence_number: None,
1764 item: response_item_function_call("item_fn", Some("{\"query\":\"test\"}")),
1765 },
1766 ResponsesStreamEvent::FunctionCallArgumentsDone {
1767 item_id: "item_fn".into(),
1768 output_index: 1,
1769 arguments: "{\"query\":\"test\"}".into(),
1770 sequence_number: None,
1771 },
1772 ResponsesStreamEvent::Completed {
1773 response: ResponseSummary::default(),
1774 },
1775 ];
1776
1777 let mapped = map_response_events(events);
1778 assert!(matches!(
1779 mapped[0],
1780 LanguageModelCompletionEvent::StartMessage { .. }
1781 ));
1782 assert!(matches!(
1783 mapped[1],
1784 LanguageModelCompletionEvent::Text(ref text) if text == "Let me check that"
1785 ));
1786 assert!(matches!(
1787 mapped[2],
1788 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1789 if raw_input == "{\"query\":\"test\"}"
1790 ));
1791 assert!(matches!(
1792 mapped[3],
1793 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1794 ));
1795 }
1796
1797 #[test]
1798 fn responses_stream_handles_json_parse_error() {
1799 let events = vec![
1800 ResponsesStreamEvent::OutputItemAdded {
1801 output_index: 0,
1802 sequence_number: None,
1803 item: response_item_function_call("item_fn", Some("{invalid json")),
1804 },
1805 ResponsesStreamEvent::FunctionCallArgumentsDone {
1806 item_id: "item_fn".into(),
1807 output_index: 0,
1808 arguments: "{invalid json".into(),
1809 sequence_number: None,
1810 },
1811 ResponsesStreamEvent::Completed {
1812 response: ResponseSummary::default(),
1813 },
1814 ];
1815
1816 let mapped = map_response_events(events);
1817 assert!(matches!(
1818 mapped[0],
1819 LanguageModelCompletionEvent::ToolUseJsonParseError {
1820 ref raw_input,
1821 ..
1822 } if raw_input.as_ref() == "{invalid json"
1823 ));
1824 }
1825
1826 #[test]
1827 fn responses_stream_handles_incomplete_function_call() {
1828 let events = vec![
1829 ResponsesStreamEvent::OutputItemAdded {
1830 output_index: 0,
1831 sequence_number: None,
1832 item: response_item_function_call("item_fn", Some("{\"city\":")),
1833 },
1834 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1835 item_id: "item_fn".into(),
1836 output_index: 0,
1837 delta: "\"Boston\"".into(),
1838 sequence_number: None,
1839 },
1840 ResponsesStreamEvent::Incomplete {
1841 response: ResponseSummary {
1842 status_details: Some(ResponseStatusDetails {
1843 reason: Some("max_output_tokens".into()),
1844 r#type: Some("incomplete".into()),
1845 error: None,
1846 }),
1847 output: vec![response_item_function_call(
1848 "item_fn",
1849 Some("{\"city\":\"Boston\"}"),
1850 )],
1851 ..Default::default()
1852 },
1853 },
1854 ];
1855
1856 let mapped = map_response_events(events);
1857 assert!(matches!(
1858 mapped[0],
1859 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1860 if raw_input == "{\"city\":\"Boston\"}"
1861 ));
1862 assert!(matches!(
1863 mapped[1],
1864 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1865 ));
1866 }
1867
1868 #[test]
1869 fn responses_stream_incomplete_does_not_duplicate_tool_calls() {
1870 let events = vec![
1871 ResponsesStreamEvent::OutputItemAdded {
1872 output_index: 0,
1873 sequence_number: None,
1874 item: response_item_function_call("item_fn", Some("{\"city\":\"Boston\"}")),
1875 },
1876 ResponsesStreamEvent::FunctionCallArgumentsDone {
1877 item_id: "item_fn".into(),
1878 output_index: 0,
1879 arguments: "{\"city\":\"Boston\"}".into(),
1880 sequence_number: None,
1881 },
1882 ResponsesStreamEvent::Incomplete {
1883 response: ResponseSummary {
1884 status_details: Some(ResponseStatusDetails {
1885 reason: Some("max_output_tokens".into()),
1886 r#type: Some("incomplete".into()),
1887 error: None,
1888 }),
1889 output: vec![response_item_function_call(
1890 "item_fn",
1891 Some("{\"city\":\"Boston\"}"),
1892 )],
1893 ..Default::default()
1894 },
1895 },
1896 ];
1897
1898 let mapped = map_response_events(events);
1899 assert_eq!(mapped.len(), 2);
1900 assert!(matches!(
1901 mapped[0],
1902 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1903 if raw_input == "{\"city\":\"Boston\"}"
1904 ));
1905 assert!(matches!(
1906 mapped[1],
1907 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1908 ));
1909 }
1910}