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 mode: _,
542 messages,
543 tools,
544 tool_choice,
545 stop: _,
546 temperature,
547 thinking_allowed: _,
548 } = request;
549
550 let mut input_items = Vec::new();
551 for (index, message) in messages.into_iter().enumerate() {
552 append_message_to_response_items(message, index, &mut input_items);
553 }
554
555 let tools: Vec<_> = tools
556 .into_iter()
557 .map(|tool| open_ai::responses::ToolDefinition::Function {
558 name: tool.name,
559 description: Some(tool.description),
560 parameters: Some(tool.input_schema),
561 strict: None,
562 })
563 .collect();
564
565 ResponseRequest {
566 model: model_id.into(),
567 input: input_items,
568 stream,
569 temperature,
570 top_p: None,
571 max_output_tokens,
572 parallel_tool_calls: if tools.is_empty() {
573 None
574 } else {
575 Some(supports_parallel_tool_calls)
576 },
577 tool_choice: tool_choice.map(|choice| match choice {
578 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
579 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
580 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
581 }),
582 tools,
583 prompt_cache_key: if supports_prompt_cache_key {
584 thread_id
585 } else {
586 None
587 },
588 reasoning: reasoning_effort.map(|effort| open_ai::responses::ReasoningConfig { effort }),
589 }
590}
591
592fn append_message_to_response_items(
593 message: LanguageModelRequestMessage,
594 index: usize,
595 input_items: &mut Vec<ResponseInputItem>,
596) {
597 let mut content_parts: Vec<ResponseInputContent> = Vec::new();
598
599 for content in message.content {
600 match content {
601 MessageContent::Text(text) => {
602 push_response_text_part(&message.role, text, &mut content_parts);
603 }
604 MessageContent::Thinking { text, .. } => {
605 push_response_text_part(&message.role, text, &mut content_parts);
606 }
607 MessageContent::RedactedThinking(_) => {}
608 MessageContent::Image(image) => {
609 push_response_image_part(&message.role, image, &mut content_parts);
610 }
611 MessageContent::ToolUse(tool_use) => {
612 flush_response_parts(&message.role, index, &mut content_parts, input_items);
613 let call_id = tool_use.id.to_string();
614 input_items.push(ResponseInputItem::FunctionCall(ResponseFunctionCallItem {
615 call_id,
616 name: tool_use.name.to_string(),
617 arguments: tool_use.raw_input,
618 }));
619 }
620 MessageContent::ToolResult(tool_result) => {
621 flush_response_parts(&message.role, index, &mut content_parts, input_items);
622 input_items.push(ResponseInputItem::FunctionCallOutput(
623 ResponseFunctionCallOutputItem {
624 call_id: tool_result.tool_use_id.to_string(),
625 output: tool_result_output(&tool_result),
626 },
627 ));
628 }
629 }
630 }
631
632 flush_response_parts(&message.role, index, &mut content_parts, input_items);
633}
634
635fn push_response_text_part(
636 role: &Role,
637 text: impl Into<String>,
638 parts: &mut Vec<ResponseInputContent>,
639) {
640 let text = text.into();
641 if text.trim().is_empty() {
642 return;
643 }
644
645 match role {
646 Role::Assistant => parts.push(ResponseInputContent::OutputText {
647 text,
648 annotations: Vec::new(),
649 }),
650 _ => parts.push(ResponseInputContent::Text { text }),
651 }
652}
653
654fn push_response_image_part(
655 role: &Role,
656 image: LanguageModelImage,
657 parts: &mut Vec<ResponseInputContent>,
658) {
659 match role {
660 Role::Assistant => parts.push(ResponseInputContent::OutputText {
661 text: "[image omitted]".to_string(),
662 annotations: Vec::new(),
663 }),
664 _ => parts.push(ResponseInputContent::Image {
665 image_url: image.to_base64_url(),
666 }),
667 }
668}
669
670fn flush_response_parts(
671 role: &Role,
672 _index: usize,
673 parts: &mut Vec<ResponseInputContent>,
674 input_items: &mut Vec<ResponseInputItem>,
675) {
676 if parts.is_empty() {
677 return;
678 }
679
680 let item = ResponseInputItem::Message(ResponseMessageItem {
681 role: match role {
682 Role::User => open_ai::Role::User,
683 Role::Assistant => open_ai::Role::Assistant,
684 Role::System => open_ai::Role::System,
685 },
686 content: parts.clone(),
687 });
688
689 input_items.push(item);
690 parts.clear();
691}
692
693fn tool_result_output(result: &LanguageModelToolResult) -> String {
694 if let Some(output) = &result.output {
695 match output {
696 serde_json::Value::String(text) => text.clone(),
697 serde_json::Value::Null => String::new(),
698 _ => output.to_string(),
699 }
700 } else {
701 match &result.content {
702 LanguageModelToolResultContent::Text(text) => text.to_string(),
703 LanguageModelToolResultContent::Image(image) => image.to_base64_url(),
704 }
705 }
706}
707
708fn add_message_content_part(
709 new_part: open_ai::MessagePart,
710 role: Role,
711 messages: &mut Vec<open_ai::RequestMessage>,
712) {
713 match (role, messages.last_mut()) {
714 (Role::User, Some(open_ai::RequestMessage::User { content }))
715 | (
716 Role::Assistant,
717 Some(open_ai::RequestMessage::Assistant {
718 content: Some(content),
719 ..
720 }),
721 )
722 | (Role::System, Some(open_ai::RequestMessage::System { content, .. })) => {
723 content.push_part(new_part);
724 }
725 _ => {
726 messages.push(match role {
727 Role::User => open_ai::RequestMessage::User {
728 content: open_ai::MessageContent::from(vec![new_part]),
729 },
730 Role::Assistant => open_ai::RequestMessage::Assistant {
731 content: Some(open_ai::MessageContent::from(vec![new_part])),
732 tool_calls: Vec::new(),
733 },
734 Role::System => open_ai::RequestMessage::System {
735 content: open_ai::MessageContent::from(vec![new_part]),
736 },
737 });
738 }
739 }
740}
741
742pub struct OpenAiEventMapper {
743 tool_calls_by_index: HashMap<usize, RawToolCall>,
744}
745
746impl OpenAiEventMapper {
747 pub fn new() -> Self {
748 Self {
749 tool_calls_by_index: HashMap::default(),
750 }
751 }
752
753 pub fn map_stream(
754 mut self,
755 events: Pin<Box<dyn Send + Stream<Item = Result<ResponseStreamEvent>>>>,
756 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
757 {
758 events.flat_map(move |event| {
759 futures::stream::iter(match event {
760 Ok(event) => self.map_event(event),
761 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
762 })
763 })
764 }
765
766 pub fn map_event(
767 &mut self,
768 event: ResponseStreamEvent,
769 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
770 let mut events = Vec::new();
771 if let Some(usage) = event.usage {
772 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
773 input_tokens: usage.prompt_tokens,
774 output_tokens: usage.completion_tokens,
775 cache_creation_input_tokens: 0,
776 cache_read_input_tokens: 0,
777 })));
778 }
779
780 let Some(choice) = event.choices.first() else {
781 return events;
782 };
783
784 if let Some(delta) = choice.delta.as_ref() {
785 if let Some(content) = delta.content.clone() {
786 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
787 }
788
789 if let Some(tool_calls) = delta.tool_calls.as_ref() {
790 for tool_call in tool_calls {
791 let entry = self.tool_calls_by_index.entry(tool_call.index).or_default();
792
793 if let Some(tool_id) = tool_call.id.clone() {
794 entry.id = tool_id;
795 }
796
797 if let Some(function) = tool_call.function.as_ref() {
798 if let Some(name) = function.name.clone() {
799 entry.name = name;
800 }
801
802 if let Some(arguments) = function.arguments.clone() {
803 entry.arguments.push_str(&arguments);
804 }
805 }
806 }
807 }
808 }
809
810 match choice.finish_reason.as_deref() {
811 Some("stop") => {
812 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
813 }
814 Some("tool_calls") => {
815 events.extend(self.tool_calls_by_index.drain().map(|(_, tool_call)| {
816 match serde_json::Value::from_str(&tool_call.arguments) {
817 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
818 LanguageModelToolUse {
819 id: tool_call.id.clone().into(),
820 name: tool_call.name.as_str().into(),
821 is_input_complete: true,
822 input,
823 raw_input: tool_call.arguments.clone(),
824 thought_signature: None,
825 },
826 )),
827 Err(error) => Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
828 id: tool_call.id.into(),
829 tool_name: tool_call.name.into(),
830 raw_input: tool_call.arguments.clone().into(),
831 json_parse_error: error.to_string(),
832 }),
833 }
834 }));
835
836 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
837 }
838 Some(stop_reason) => {
839 log::error!("Unexpected OpenAI stop_reason: {stop_reason:?}",);
840 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
841 }
842 None => {}
843 }
844
845 events
846 }
847}
848
849#[derive(Default)]
850struct RawToolCall {
851 id: String,
852 name: String,
853 arguments: String,
854}
855
856pub struct OpenAiResponseEventMapper {
857 function_calls_by_item: HashMap<String, PendingResponseFunctionCall>,
858 pending_stop_reason: Option<StopReason>,
859}
860
861#[derive(Default)]
862struct PendingResponseFunctionCall {
863 call_id: String,
864 name: Arc<str>,
865 arguments: String,
866}
867
868impl OpenAiResponseEventMapper {
869 pub fn new() -> Self {
870 Self {
871 function_calls_by_item: HashMap::default(),
872 pending_stop_reason: None,
873 }
874 }
875
876 pub fn map_stream(
877 mut self,
878 events: Pin<Box<dyn Send + Stream<Item = Result<ResponsesStreamEvent>>>>,
879 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
880 {
881 events.flat_map(move |event| {
882 futures::stream::iter(match event {
883 Ok(event) => self.map_event(event),
884 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
885 })
886 })
887 }
888
889 pub fn map_event(
890 &mut self,
891 event: ResponsesStreamEvent,
892 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
893 match event {
894 ResponsesStreamEvent::OutputItemAdded { item, .. } => {
895 let mut events = Vec::new();
896
897 match &item {
898 ResponseOutputItem::Message(message) => {
899 if let Some(id) = &message.id {
900 events.push(Ok(LanguageModelCompletionEvent::StartMessage {
901 message_id: id.clone(),
902 }));
903 }
904 }
905 ResponseOutputItem::FunctionCall(function_call) => {
906 if let Some(item_id) = function_call.id.clone() {
907 let call_id = function_call
908 .call_id
909 .clone()
910 .or_else(|| function_call.id.clone())
911 .unwrap_or_else(|| item_id.clone());
912 let entry = PendingResponseFunctionCall {
913 call_id,
914 name: Arc::<str>::from(
915 function_call.name.clone().unwrap_or_default(),
916 ),
917 arguments: function_call.arguments.clone(),
918 };
919 self.function_calls_by_item.insert(item_id, entry);
920 }
921 }
922 ResponseOutputItem::Unknown => {}
923 }
924 events
925 }
926 ResponsesStreamEvent::OutputTextDelta { delta, .. } => {
927 if delta.is_empty() {
928 Vec::new()
929 } else {
930 vec![Ok(LanguageModelCompletionEvent::Text(delta))]
931 }
932 }
933 ResponsesStreamEvent::FunctionCallArgumentsDelta { item_id, delta, .. } => {
934 if let Some(entry) = self.function_calls_by_item.get_mut(&item_id) {
935 entry.arguments.push_str(&delta);
936 }
937 Vec::new()
938 }
939 ResponsesStreamEvent::FunctionCallArgumentsDone {
940 item_id, arguments, ..
941 } => {
942 if let Some(mut entry) = self.function_calls_by_item.remove(&item_id) {
943 if !arguments.is_empty() {
944 entry.arguments = arguments;
945 }
946 let raw_input = entry.arguments.clone();
947 self.pending_stop_reason = Some(StopReason::ToolUse);
948 match serde_json::from_str::<serde_json::Value>(&entry.arguments) {
949 Ok(input) => {
950 vec![Ok(LanguageModelCompletionEvent::ToolUse(
951 LanguageModelToolUse {
952 id: LanguageModelToolUseId::from(entry.call_id.clone()),
953 name: entry.name.clone(),
954 is_input_complete: true,
955 input,
956 raw_input,
957 thought_signature: None,
958 },
959 ))]
960 }
961 Err(error) => {
962 vec![Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
963 id: LanguageModelToolUseId::from(entry.call_id.clone()),
964 tool_name: entry.name.clone(),
965 raw_input: Arc::<str>::from(raw_input),
966 json_parse_error: error.to_string(),
967 })]
968 }
969 }
970 } else {
971 Vec::new()
972 }
973 }
974 ResponsesStreamEvent::Completed { response } => {
975 self.handle_completion(response, StopReason::EndTurn)
976 }
977 ResponsesStreamEvent::Incomplete { response } => {
978 let reason = response
979 .status_details
980 .as_ref()
981 .and_then(|details| details.reason.as_deref());
982 let stop_reason = match reason {
983 Some("max_output_tokens") => StopReason::MaxTokens,
984 Some("content_filter") => {
985 self.pending_stop_reason = Some(StopReason::Refusal);
986 StopReason::Refusal
987 }
988 _ => self
989 .pending_stop_reason
990 .take()
991 .unwrap_or(StopReason::EndTurn),
992 };
993
994 let mut events = Vec::new();
995 if self.pending_stop_reason.is_none() {
996 events.extend(self.emit_tool_calls_from_output(&response.output));
997 }
998 if let Some(usage) = response.usage.as_ref() {
999 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1000 token_usage_from_response_usage(usage),
1001 )));
1002 }
1003 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1004 events
1005 }
1006 ResponsesStreamEvent::Failed { response } => {
1007 let message = response
1008 .status_details
1009 .and_then(|details| details.error)
1010 .map(|error| error.to_string())
1011 .unwrap_or_else(|| "response failed".to_string());
1012 vec![Err(LanguageModelCompletionError::Other(anyhow!(message)))]
1013 }
1014 ResponsesStreamEvent::Error { error }
1015 | ResponsesStreamEvent::GenericError { error } => {
1016 vec![Err(LanguageModelCompletionError::Other(anyhow!(format!(
1017 "{error:?}"
1018 ))))]
1019 }
1020 ResponsesStreamEvent::OutputTextDone { .. } => Vec::new(),
1021 ResponsesStreamEvent::OutputItemDone { .. }
1022 | ResponsesStreamEvent::ContentPartAdded { .. }
1023 | ResponsesStreamEvent::ContentPartDone { .. }
1024 | ResponsesStreamEvent::Created { .. }
1025 | ResponsesStreamEvent::InProgress { .. }
1026 | ResponsesStreamEvent::Unknown => Vec::new(),
1027 }
1028 }
1029
1030 fn handle_completion(
1031 &mut self,
1032 response: ResponsesSummary,
1033 default_reason: StopReason,
1034 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1035 let mut events = Vec::new();
1036
1037 if self.pending_stop_reason.is_none() {
1038 events.extend(self.emit_tool_calls_from_output(&response.output));
1039 }
1040
1041 if let Some(usage) = response.usage.as_ref() {
1042 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1043 token_usage_from_response_usage(usage),
1044 )));
1045 }
1046
1047 let stop_reason = self.pending_stop_reason.take().unwrap_or(default_reason);
1048 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1049 events
1050 }
1051
1052 fn emit_tool_calls_from_output(
1053 &mut self,
1054 output: &[ResponseOutputItem],
1055 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1056 let mut events = Vec::new();
1057 for item in output {
1058 if let ResponseOutputItem::FunctionCall(function_call) = item {
1059 let Some(call_id) = function_call
1060 .call_id
1061 .clone()
1062 .or_else(|| function_call.id.clone())
1063 else {
1064 log::error!(
1065 "Function call item missing both call_id and id: {:?}",
1066 function_call
1067 );
1068 continue;
1069 };
1070 let name: Arc<str> = Arc::from(function_call.name.clone().unwrap_or_default());
1071 let arguments = &function_call.arguments;
1072 if !arguments.is_empty() {
1073 self.pending_stop_reason = Some(StopReason::ToolUse);
1074 match serde_json::from_str::<serde_json::Value>(arguments) {
1075 Ok(input) => {
1076 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
1077 LanguageModelToolUse {
1078 id: LanguageModelToolUseId::from(call_id.clone()),
1079 name: name.clone(),
1080 is_input_complete: true,
1081 input,
1082 raw_input: arguments.clone(),
1083 thought_signature: None,
1084 },
1085 )));
1086 }
1087 Err(error) => {
1088 events.push(Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1089 id: LanguageModelToolUseId::from(call_id.clone()),
1090 tool_name: name.clone(),
1091 raw_input: Arc::<str>::from(arguments.clone()),
1092 json_parse_error: error.to_string(),
1093 }));
1094 }
1095 }
1096 }
1097 }
1098 }
1099 events
1100 }
1101}
1102
1103fn token_usage_from_response_usage(usage: &ResponsesUsage) -> TokenUsage {
1104 TokenUsage {
1105 input_tokens: usage.input_tokens.unwrap_or_default(),
1106 output_tokens: usage.output_tokens.unwrap_or_default(),
1107 cache_creation_input_tokens: 0,
1108 cache_read_input_tokens: 0,
1109 }
1110}
1111
1112pub(crate) fn collect_tiktoken_messages(
1113 request: LanguageModelRequest,
1114) -> Vec<tiktoken_rs::ChatCompletionRequestMessage> {
1115 request
1116 .messages
1117 .into_iter()
1118 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
1119 role: match message.role {
1120 Role::User => "user".into(),
1121 Role::Assistant => "assistant".into(),
1122 Role::System => "system".into(),
1123 },
1124 content: Some(message.string_contents()),
1125 name: None,
1126 function_call: None,
1127 })
1128 .collect::<Vec<_>>()
1129}
1130
1131pub fn count_open_ai_tokens(
1132 request: LanguageModelRequest,
1133 model: Model,
1134 cx: &App,
1135) -> BoxFuture<'static, Result<u64>> {
1136 cx.background_spawn(async move {
1137 let messages = collect_tiktoken_messages(request);
1138 match model {
1139 Model::Custom { max_tokens, .. } => {
1140 let model = if max_tokens >= 100_000 {
1141 // If the max tokens is 100k or more, it is likely the o200k_base tokenizer from gpt4o
1142 "gpt-4o"
1143 } else {
1144 // Otherwise fallback to gpt-4, since only cl100k_base and o200k_base are
1145 // supported with this tiktoken method
1146 "gpt-4"
1147 };
1148 tiktoken_rs::num_tokens_from_messages(model, &messages)
1149 }
1150 // Currently supported by tiktoken_rs
1151 // Sometimes tiktoken-rs is behind on model support. If that is the case, make a new branch
1152 // arm with an override. We enumerate all supported models here so that we can check if new
1153 // models are supported yet or not.
1154 Model::ThreePointFiveTurbo
1155 | Model::Four
1156 | Model::FourTurbo
1157 | Model::FourOmni
1158 | Model::FourOmniMini
1159 | Model::FourPointOne
1160 | Model::FourPointOneMini
1161 | Model::FourPointOneNano
1162 | Model::O1
1163 | Model::O3
1164 | Model::O3Mini
1165 | Model::O4Mini
1166 | Model::Five
1167 | Model::FiveCodex
1168 | Model::FiveMini
1169 | Model::FiveNano => tiktoken_rs::num_tokens_from_messages(model.id(), &messages),
1170 // GPT-5.1 and 5.2 don't have dedicated tiktoken support; use gpt-5 tokenizer
1171 Model::FivePointOne | Model::FivePointTwo => {
1172 tiktoken_rs::num_tokens_from_messages("gpt-5", &messages)
1173 }
1174 }
1175 .map(|tokens| tokens as u64)
1176 })
1177 .boxed()
1178}
1179
1180struct ConfigurationView {
1181 api_key_editor: Entity<InputField>,
1182 state: Entity<State>,
1183 load_credentials_task: Option<Task<()>>,
1184}
1185
1186impl ConfigurationView {
1187 fn new(state: Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
1188 let api_key_editor = cx.new(|cx| {
1189 InputField::new(
1190 window,
1191 cx,
1192 "sk-000000000000000000000000000000000000000000000000",
1193 )
1194 });
1195
1196 cx.observe(&state, |_, _, cx| {
1197 cx.notify();
1198 })
1199 .detach();
1200
1201 let load_credentials_task = Some(cx.spawn_in(window, {
1202 let state = state.clone();
1203 async move |this, cx| {
1204 if let Some(task) = Some(state.update(cx, |state, cx| state.authenticate(cx))) {
1205 // We don't log an error, because "not signed in" is also an error.
1206 let _ = task.await;
1207 }
1208 this.update(cx, |this, cx| {
1209 this.load_credentials_task = None;
1210 cx.notify();
1211 })
1212 .log_err();
1213 }
1214 }));
1215
1216 Self {
1217 api_key_editor,
1218 state,
1219 load_credentials_task,
1220 }
1221 }
1222
1223 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
1224 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
1225 if api_key.is_empty() {
1226 return;
1227 }
1228
1229 // url changes can cause the editor to be displayed again
1230 self.api_key_editor
1231 .update(cx, |editor, cx| editor.set_text("", window, cx));
1232
1233 let state = self.state.clone();
1234 cx.spawn_in(window, async move |_, cx| {
1235 state
1236 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
1237 .await
1238 })
1239 .detach_and_log_err(cx);
1240 }
1241
1242 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
1243 self.api_key_editor
1244 .update(cx, |input, cx| input.set_text("", window, cx));
1245
1246 let state = self.state.clone();
1247 cx.spawn_in(window, async move |_, cx| {
1248 state
1249 .update(cx, |state, cx| state.set_api_key(None, cx))
1250 .await
1251 })
1252 .detach_and_log_err(cx);
1253 }
1254
1255 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
1256 !self.state.read(cx).is_authenticated()
1257 }
1258}
1259
1260impl Render for ConfigurationView {
1261 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1262 let env_var_set = self.state.read(cx).api_key_state.is_from_env_var();
1263 let configured_card_label = if env_var_set {
1264 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable")
1265 } else {
1266 let api_url = OpenAiLanguageModelProvider::api_url(cx);
1267 if api_url == OPEN_AI_API_URL {
1268 "API key configured".to_string()
1269 } else {
1270 format!("API key configured for {}", api_url)
1271 }
1272 };
1273
1274 let api_key_section = if self.should_render_editor(cx) {
1275 v_flex()
1276 .on_action(cx.listener(Self::save_api_key))
1277 .child(Label::new("To use Zed's agent with OpenAI, you need to add an API key. Follow these steps:"))
1278 .child(
1279 List::new()
1280 .child(
1281 ListBulletItem::new("")
1282 .child(Label::new("Create one by visiting"))
1283 .child(ButtonLink::new("OpenAI's console", "https://platform.openai.com/api-keys"))
1284 )
1285 .child(
1286 ListBulletItem::new("Ensure your OpenAI account has credits")
1287 )
1288 .child(
1289 ListBulletItem::new("Paste your API key below and hit enter to start using the agent")
1290 ),
1291 )
1292 .child(self.api_key_editor.clone())
1293 .child(
1294 Label::new(format!(
1295 "You can also set the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed."
1296 ))
1297 .size(LabelSize::Small)
1298 .color(Color::Muted),
1299 )
1300 .child(
1301 Label::new(
1302 "Note that having a subscription for another service like GitHub Copilot won't work.",
1303 )
1304 .size(LabelSize::Small).color(Color::Muted),
1305 )
1306 .into_any_element()
1307 } else {
1308 ConfiguredApiCard::new(configured_card_label)
1309 .disabled(env_var_set)
1310 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
1311 .when(env_var_set, |this| {
1312 this.tooltip_label(format!("To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable."))
1313 })
1314 .into_any_element()
1315 };
1316
1317 let compatible_api_section = h_flex()
1318 .mt_1p5()
1319 .gap_0p5()
1320 .flex_wrap()
1321 .when(self.should_render_editor(cx), |this| {
1322 this.pt_1p5()
1323 .border_t_1()
1324 .border_color(cx.theme().colors().border_variant)
1325 })
1326 .child(
1327 h_flex()
1328 .gap_2()
1329 .child(
1330 Icon::new(IconName::Info)
1331 .size(IconSize::XSmall)
1332 .color(Color::Muted),
1333 )
1334 .child(Label::new("Zed also supports OpenAI-compatible models.")),
1335 )
1336 .child(
1337 Button::new("docs", "Learn More")
1338 .icon(IconName::ArrowUpRight)
1339 .icon_size(IconSize::Small)
1340 .icon_color(Color::Muted)
1341 .on_click(move |_, _window, cx| {
1342 cx.open_url("https://zed.dev/docs/ai/llm-providers#openai-api-compatible")
1343 }),
1344 );
1345
1346 if self.load_credentials_task.is_some() {
1347 div().child(Label::new("Loading credentials…")).into_any()
1348 } else {
1349 v_flex()
1350 .size_full()
1351 .child(api_key_section)
1352 .child(compatible_api_section)
1353 .into_any()
1354 }
1355 }
1356}
1357
1358#[cfg(test)]
1359mod tests {
1360 use futures::{StreamExt, executor::block_on};
1361 use gpui::TestAppContext;
1362 use language_model::{LanguageModelRequestMessage, LanguageModelRequestTool};
1363 use open_ai::responses::{
1364 ResponseFunctionToolCall, ResponseOutputItem, ResponseOutputMessage, ResponseStatusDetails,
1365 ResponseSummary, ResponseUsage, StreamEvent as ResponsesStreamEvent,
1366 };
1367 use pretty_assertions::assert_eq;
1368 use serde_json::json;
1369
1370 use super::*;
1371
1372 fn map_response_events(events: Vec<ResponsesStreamEvent>) -> Vec<LanguageModelCompletionEvent> {
1373 block_on(async {
1374 OpenAiResponseEventMapper::new()
1375 .map_stream(Box::pin(futures::stream::iter(events.into_iter().map(Ok))))
1376 .collect::<Vec<_>>()
1377 .await
1378 .into_iter()
1379 .map(Result::unwrap)
1380 .collect()
1381 })
1382 }
1383
1384 fn response_item_message(id: &str) -> ResponseOutputItem {
1385 ResponseOutputItem::Message(ResponseOutputMessage {
1386 id: Some(id.to_string()),
1387 role: Some("assistant".to_string()),
1388 status: Some("in_progress".to_string()),
1389 content: vec![],
1390 })
1391 }
1392
1393 fn response_item_function_call(id: &str, args: Option<&str>) -> ResponseOutputItem {
1394 ResponseOutputItem::FunctionCall(ResponseFunctionToolCall {
1395 id: Some(id.to_string()),
1396 status: Some("in_progress".to_string()),
1397 name: Some("get_weather".to_string()),
1398 call_id: Some("call_123".to_string()),
1399 arguments: args.map(|s| s.to_string()).unwrap_or_default(),
1400 })
1401 }
1402
1403 #[gpui::test]
1404 fn tiktoken_rs_support(cx: &TestAppContext) {
1405 let request = LanguageModelRequest {
1406 thread_id: None,
1407 prompt_id: None,
1408 intent: None,
1409 mode: None,
1410 messages: vec![LanguageModelRequestMessage {
1411 role: Role::User,
1412 content: vec![MessageContent::Text("message".into())],
1413 cache: false,
1414 reasoning_details: None,
1415 }],
1416 tools: vec![],
1417 tool_choice: None,
1418 stop: vec![],
1419 temperature: None,
1420 thinking_allowed: true,
1421 };
1422
1423 // Validate that all models are supported by tiktoken-rs
1424 for model in Model::iter() {
1425 let count = cx
1426 .foreground_executor()
1427 .block_on(count_open_ai_tokens(
1428 request.clone(),
1429 model,
1430 &cx.app.borrow(),
1431 ))
1432 .unwrap();
1433 assert!(count > 0);
1434 }
1435 }
1436
1437 #[test]
1438 fn responses_stream_maps_text_and_usage() {
1439 let events = vec![
1440 ResponsesStreamEvent::OutputItemAdded {
1441 output_index: 0,
1442 sequence_number: None,
1443 item: response_item_message("msg_123"),
1444 },
1445 ResponsesStreamEvent::OutputTextDelta {
1446 item_id: "msg_123".into(),
1447 output_index: 0,
1448 content_index: Some(0),
1449 delta: "Hello".into(),
1450 },
1451 ResponsesStreamEvent::Completed {
1452 response: ResponseSummary {
1453 usage: Some(ResponseUsage {
1454 input_tokens: Some(5),
1455 output_tokens: Some(3),
1456 total_tokens: Some(8),
1457 }),
1458 ..Default::default()
1459 },
1460 },
1461 ];
1462
1463 let mapped = map_response_events(events);
1464 assert!(matches!(
1465 mapped[0],
1466 LanguageModelCompletionEvent::StartMessage { ref message_id } if message_id == "msg_123"
1467 ));
1468 assert!(matches!(
1469 mapped[1],
1470 LanguageModelCompletionEvent::Text(ref text) if text == "Hello"
1471 ));
1472 assert!(matches!(
1473 mapped[2],
1474 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1475 input_tokens: 5,
1476 output_tokens: 3,
1477 ..
1478 })
1479 ));
1480 assert!(matches!(
1481 mapped[3],
1482 LanguageModelCompletionEvent::Stop(StopReason::EndTurn)
1483 ));
1484 }
1485
1486 #[test]
1487 fn into_open_ai_response_builds_complete_payload() {
1488 let tool_call_id = LanguageModelToolUseId::from("call-42");
1489 let tool_input = json!({ "city": "Boston" });
1490 let tool_arguments = serde_json::to_string(&tool_input).unwrap();
1491 let tool_use = LanguageModelToolUse {
1492 id: tool_call_id.clone(),
1493 name: Arc::from("get_weather"),
1494 raw_input: tool_arguments.clone(),
1495 input: tool_input,
1496 is_input_complete: true,
1497 thought_signature: None,
1498 };
1499 let tool_result = LanguageModelToolResult {
1500 tool_use_id: tool_call_id,
1501 tool_name: Arc::from("get_weather"),
1502 is_error: false,
1503 content: LanguageModelToolResultContent::Text(Arc::from("Sunny")),
1504 output: Some(json!({ "forecast": "Sunny" })),
1505 };
1506 let user_image = LanguageModelImage {
1507 source: SharedString::from("aGVsbG8="),
1508 size: None,
1509 };
1510 let expected_image_url = user_image.to_base64_url();
1511
1512 let request = LanguageModelRequest {
1513 thread_id: Some("thread-123".into()),
1514 prompt_id: None,
1515 intent: None,
1516 mode: None,
1517 messages: vec![
1518 LanguageModelRequestMessage {
1519 role: Role::System,
1520 content: vec![MessageContent::Text("System context".into())],
1521 cache: false,
1522 reasoning_details: None,
1523 },
1524 LanguageModelRequestMessage {
1525 role: Role::User,
1526 content: vec![
1527 MessageContent::Text("Please check the weather.".into()),
1528 MessageContent::Image(user_image),
1529 ],
1530 cache: false,
1531 reasoning_details: None,
1532 },
1533 LanguageModelRequestMessage {
1534 role: Role::Assistant,
1535 content: vec![
1536 MessageContent::Text("Looking that up.".into()),
1537 MessageContent::ToolUse(tool_use),
1538 ],
1539 cache: false,
1540 reasoning_details: None,
1541 },
1542 LanguageModelRequestMessage {
1543 role: Role::Assistant,
1544 content: vec![MessageContent::ToolResult(tool_result)],
1545 cache: false,
1546 reasoning_details: None,
1547 },
1548 ],
1549 tools: vec![LanguageModelRequestTool {
1550 name: "get_weather".into(),
1551 description: "Fetches the weather".into(),
1552 input_schema: json!({ "type": "object" }),
1553 }],
1554 tool_choice: Some(LanguageModelToolChoice::Any),
1555 stop: vec!["<STOP>".into()],
1556 temperature: None,
1557 thinking_allowed: false,
1558 };
1559
1560 let response = into_open_ai_response(
1561 request,
1562 "custom-model",
1563 true,
1564 true,
1565 Some(2048),
1566 Some(ReasoningEffort::Low),
1567 );
1568
1569 let serialized = serde_json::to_value(&response).unwrap();
1570 let expected = json!({
1571 "model": "custom-model",
1572 "input": [
1573 {
1574 "type": "message",
1575 "role": "system",
1576 "content": [
1577 { "type": "input_text", "text": "System context" }
1578 ]
1579 },
1580 {
1581 "type": "message",
1582 "role": "user",
1583 "content": [
1584 { "type": "input_text", "text": "Please check the weather." },
1585 { "type": "input_image", "image_url": expected_image_url }
1586 ]
1587 },
1588 {
1589 "type": "message",
1590 "role": "assistant",
1591 "content": [
1592 { "type": "output_text", "text": "Looking that up.", "annotations": [] }
1593 ]
1594 },
1595 {
1596 "type": "function_call",
1597 "call_id": "call-42",
1598 "name": "get_weather",
1599 "arguments": tool_arguments
1600 },
1601 {
1602 "type": "function_call_output",
1603 "call_id": "call-42",
1604 "output": "{\"forecast\":\"Sunny\"}"
1605 }
1606 ],
1607 "stream": true,
1608 "max_output_tokens": 2048,
1609 "parallel_tool_calls": true,
1610 "tool_choice": "required",
1611 "tools": [
1612 {
1613 "type": "function",
1614 "name": "get_weather",
1615 "description": "Fetches the weather",
1616 "parameters": { "type": "object" }
1617 }
1618 ],
1619 "prompt_cache_key": "thread-123",
1620 "reasoning": { "effort": "low" }
1621 });
1622
1623 assert_eq!(serialized, expected);
1624 }
1625
1626 #[test]
1627 fn responses_stream_maps_tool_calls() {
1628 let events = vec![
1629 ResponsesStreamEvent::OutputItemAdded {
1630 output_index: 0,
1631 sequence_number: None,
1632 item: response_item_function_call("item_fn", Some("{\"city\":\"Bos")),
1633 },
1634 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1635 item_id: "item_fn".into(),
1636 output_index: 0,
1637 delta: "ton\"}".into(),
1638 sequence_number: None,
1639 },
1640 ResponsesStreamEvent::FunctionCallArgumentsDone {
1641 item_id: "item_fn".into(),
1642 output_index: 0,
1643 arguments: "{\"city\":\"Boston\"}".into(),
1644 sequence_number: None,
1645 },
1646 ResponsesStreamEvent::Completed {
1647 response: ResponseSummary::default(),
1648 },
1649 ];
1650
1651 let mapped = map_response_events(events);
1652 assert!(matches!(
1653 mapped[0],
1654 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1655 ref id,
1656 ref name,
1657 ref raw_input,
1658 ..
1659 }) if id.to_string() == "call_123"
1660 && name.as_ref() == "get_weather"
1661 && raw_input == "{\"city\":\"Boston\"}"
1662 ));
1663 assert!(matches!(
1664 mapped[1],
1665 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1666 ));
1667 }
1668
1669 #[test]
1670 fn responses_stream_uses_max_tokens_stop_reason() {
1671 let events = vec![ResponsesStreamEvent::Incomplete {
1672 response: ResponseSummary {
1673 status_details: Some(ResponseStatusDetails {
1674 reason: Some("max_output_tokens".into()),
1675 r#type: Some("incomplete".into()),
1676 error: None,
1677 }),
1678 usage: Some(ResponseUsage {
1679 input_tokens: Some(10),
1680 output_tokens: Some(20),
1681 total_tokens: Some(30),
1682 }),
1683 ..Default::default()
1684 },
1685 }];
1686
1687 let mapped = map_response_events(events);
1688 assert!(matches!(
1689 mapped[0],
1690 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1691 input_tokens: 10,
1692 output_tokens: 20,
1693 ..
1694 })
1695 ));
1696 assert!(matches!(
1697 mapped[1],
1698 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1699 ));
1700 }
1701
1702 #[test]
1703 fn responses_stream_handles_multiple_tool_calls() {
1704 let events = vec![
1705 ResponsesStreamEvent::OutputItemAdded {
1706 output_index: 0,
1707 sequence_number: None,
1708 item: response_item_function_call("item_fn1", Some("{\"city\":\"NYC\"}")),
1709 },
1710 ResponsesStreamEvent::FunctionCallArgumentsDone {
1711 item_id: "item_fn1".into(),
1712 output_index: 0,
1713 arguments: "{\"city\":\"NYC\"}".into(),
1714 sequence_number: None,
1715 },
1716 ResponsesStreamEvent::OutputItemAdded {
1717 output_index: 1,
1718 sequence_number: None,
1719 item: response_item_function_call("item_fn2", Some("{\"city\":\"LA\"}")),
1720 },
1721 ResponsesStreamEvent::FunctionCallArgumentsDone {
1722 item_id: "item_fn2".into(),
1723 output_index: 1,
1724 arguments: "{\"city\":\"LA\"}".into(),
1725 sequence_number: None,
1726 },
1727 ResponsesStreamEvent::Completed {
1728 response: ResponseSummary::default(),
1729 },
1730 ];
1731
1732 let mapped = map_response_events(events);
1733 assert_eq!(mapped.len(), 3);
1734 assert!(matches!(
1735 mapped[0],
1736 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1737 if raw_input == "{\"city\":\"NYC\"}"
1738 ));
1739 assert!(matches!(
1740 mapped[1],
1741 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1742 if raw_input == "{\"city\":\"LA\"}"
1743 ));
1744 assert!(matches!(
1745 mapped[2],
1746 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1747 ));
1748 }
1749
1750 #[test]
1751 fn responses_stream_handles_mixed_text_and_tool_calls() {
1752 let events = vec![
1753 ResponsesStreamEvent::OutputItemAdded {
1754 output_index: 0,
1755 sequence_number: None,
1756 item: response_item_message("msg_123"),
1757 },
1758 ResponsesStreamEvent::OutputTextDelta {
1759 item_id: "msg_123".into(),
1760 output_index: 0,
1761 content_index: Some(0),
1762 delta: "Let me check that".into(),
1763 },
1764 ResponsesStreamEvent::OutputItemAdded {
1765 output_index: 1,
1766 sequence_number: None,
1767 item: response_item_function_call("item_fn", Some("{\"query\":\"test\"}")),
1768 },
1769 ResponsesStreamEvent::FunctionCallArgumentsDone {
1770 item_id: "item_fn".into(),
1771 output_index: 1,
1772 arguments: "{\"query\":\"test\"}".into(),
1773 sequence_number: None,
1774 },
1775 ResponsesStreamEvent::Completed {
1776 response: ResponseSummary::default(),
1777 },
1778 ];
1779
1780 let mapped = map_response_events(events);
1781 assert!(matches!(
1782 mapped[0],
1783 LanguageModelCompletionEvent::StartMessage { .. }
1784 ));
1785 assert!(matches!(
1786 mapped[1],
1787 LanguageModelCompletionEvent::Text(ref text) if text == "Let me check that"
1788 ));
1789 assert!(matches!(
1790 mapped[2],
1791 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1792 if raw_input == "{\"query\":\"test\"}"
1793 ));
1794 assert!(matches!(
1795 mapped[3],
1796 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1797 ));
1798 }
1799
1800 #[test]
1801 fn responses_stream_handles_json_parse_error() {
1802 let events = vec![
1803 ResponsesStreamEvent::OutputItemAdded {
1804 output_index: 0,
1805 sequence_number: None,
1806 item: response_item_function_call("item_fn", Some("{invalid json")),
1807 },
1808 ResponsesStreamEvent::FunctionCallArgumentsDone {
1809 item_id: "item_fn".into(),
1810 output_index: 0,
1811 arguments: "{invalid json".into(),
1812 sequence_number: None,
1813 },
1814 ResponsesStreamEvent::Completed {
1815 response: ResponseSummary::default(),
1816 },
1817 ];
1818
1819 let mapped = map_response_events(events);
1820 assert!(matches!(
1821 mapped[0],
1822 LanguageModelCompletionEvent::ToolUseJsonParseError {
1823 ref raw_input,
1824 ..
1825 } if raw_input.as_ref() == "{invalid json"
1826 ));
1827 }
1828
1829 #[test]
1830 fn responses_stream_handles_incomplete_function_call() {
1831 let events = vec![
1832 ResponsesStreamEvent::OutputItemAdded {
1833 output_index: 0,
1834 sequence_number: None,
1835 item: response_item_function_call("item_fn", Some("{\"city\":")),
1836 },
1837 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1838 item_id: "item_fn".into(),
1839 output_index: 0,
1840 delta: "\"Boston\"".into(),
1841 sequence_number: None,
1842 },
1843 ResponsesStreamEvent::Incomplete {
1844 response: ResponseSummary {
1845 status_details: Some(ResponseStatusDetails {
1846 reason: Some("max_output_tokens".into()),
1847 r#type: Some("incomplete".into()),
1848 error: None,
1849 }),
1850 output: vec![response_item_function_call(
1851 "item_fn",
1852 Some("{\"city\":\"Boston\"}"),
1853 )],
1854 ..Default::default()
1855 },
1856 },
1857 ];
1858
1859 let mapped = map_response_events(events);
1860 assert!(matches!(
1861 mapped[0],
1862 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1863 if raw_input == "{\"city\":\"Boston\"}"
1864 ));
1865 assert!(matches!(
1866 mapped[1],
1867 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1868 ));
1869 }
1870
1871 #[test]
1872 fn responses_stream_incomplete_does_not_duplicate_tool_calls() {
1873 let events = vec![
1874 ResponsesStreamEvent::OutputItemAdded {
1875 output_index: 0,
1876 sequence_number: None,
1877 item: response_item_function_call("item_fn", Some("{\"city\":\"Boston\"}")),
1878 },
1879 ResponsesStreamEvent::FunctionCallArgumentsDone {
1880 item_id: "item_fn".into(),
1881 output_index: 0,
1882 arguments: "{\"city\":\"Boston\"}".into(),
1883 sequence_number: None,
1884 },
1885 ResponsesStreamEvent::Incomplete {
1886 response: ResponseSummary {
1887 status_details: Some(ResponseStatusDetails {
1888 reason: Some("max_output_tokens".into()),
1889 r#type: Some("incomplete".into()),
1890 error: None,
1891 }),
1892 output: vec![response_item_function_call(
1893 "item_fn",
1894 Some("{\"city\":\"Boston\"}"),
1895 )],
1896 ..Default::default()
1897 },
1898 },
1899 ];
1900
1901 let mapped = map_response_events(events);
1902 assert_eq!(mapped.len(), 2);
1903 assert!(matches!(
1904 mapped[0],
1905 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1906 if raw_input == "{\"city\":\"Boston\"}"
1907 ));
1908 assert!(matches!(
1909 mapped[1],
1910 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1911 ));
1912 }
1913}