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::sync::{Arc, LazyLock};
32use strum::IntoEnumIterator;
33use ui::{ButtonLink, ConfiguredApiCard, List, ListBulletItem, prelude::*};
34use ui_input::InputField;
35use util::ResultExt;
36
37use crate::provider::util::parse_tool_arguments;
38
39const PROVIDER_ID: LanguageModelProviderId = language_model::OPEN_AI_PROVIDER_ID;
40const PROVIDER_NAME: LanguageModelProviderName = language_model::OPEN_AI_PROVIDER_NAME;
41
42const API_KEY_ENV_VAR_NAME: &str = "OPENAI_API_KEY";
43static API_KEY_ENV_VAR: LazyLock<EnvVar> = env_var!(API_KEY_ENV_VAR_NAME);
44
45#[derive(Default, Clone, Debug, PartialEq)]
46pub struct OpenAiSettings {
47 pub api_url: String,
48 pub available_models: Vec<AvailableModel>,
49}
50
51pub struct OpenAiLanguageModelProvider {
52 http_client: Arc<dyn HttpClient>,
53 state: Entity<State>,
54}
55
56pub struct State {
57 api_key_state: ApiKeyState,
58}
59
60impl State {
61 fn is_authenticated(&self) -> bool {
62 self.api_key_state.has_key()
63 }
64
65 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
66 let api_url = OpenAiLanguageModelProvider::api_url(cx);
67 self.api_key_state
68 .store(api_url, api_key, |this| &mut this.api_key_state, cx)
69 }
70
71 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
72 let api_url = OpenAiLanguageModelProvider::api_url(cx);
73 self.api_key_state
74 .load_if_needed(api_url, |this| &mut this.api_key_state, cx)
75 }
76}
77
78impl OpenAiLanguageModelProvider {
79 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
80 let state = cx.new(|cx| {
81 cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
82 let api_url = Self::api_url(cx);
83 this.api_key_state
84 .handle_url_change(api_url, |this| &mut this.api_key_state, cx);
85 cx.notify();
86 })
87 .detach();
88 State {
89 api_key_state: ApiKeyState::new(Self::api_url(cx), (*API_KEY_ENV_VAR).clone()),
90 }
91 });
92
93 Self { http_client, state }
94 }
95
96 fn create_language_model(&self, model: open_ai::Model) -> Arc<dyn LanguageModel> {
97 Arc::new(OpenAiLanguageModel {
98 id: LanguageModelId::from(model.id().to_string()),
99 model,
100 state: self.state.clone(),
101 http_client: self.http_client.clone(),
102 request_limiter: RateLimiter::new(4),
103 })
104 }
105
106 fn settings(cx: &App) -> &OpenAiSettings {
107 &crate::AllLanguageModelSettings::get_global(cx).openai
108 }
109
110 fn api_url(cx: &App) -> SharedString {
111 let api_url = &Self::settings(cx).api_url;
112 if api_url.is_empty() {
113 open_ai::OPEN_AI_API_URL.into()
114 } else {
115 SharedString::new(api_url.as_str())
116 }
117 }
118}
119
120impl LanguageModelProviderState for OpenAiLanguageModelProvider {
121 type ObservableEntity = State;
122
123 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
124 Some(self.state.clone())
125 }
126}
127
128impl LanguageModelProvider for OpenAiLanguageModelProvider {
129 fn id(&self) -> LanguageModelProviderId {
130 PROVIDER_ID
131 }
132
133 fn name(&self) -> LanguageModelProviderName {
134 PROVIDER_NAME
135 }
136
137 fn icon(&self) -> IconOrSvg {
138 IconOrSvg::Icon(IconName::AiOpenAi)
139 }
140
141 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
142 Some(self.create_language_model(open_ai::Model::default()))
143 }
144
145 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
146 Some(self.create_language_model(open_ai::Model::default_fast()))
147 }
148
149 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
150 let mut models = BTreeMap::default();
151
152 // Add base models from open_ai::Model::iter()
153 for model in open_ai::Model::iter() {
154 if !matches!(model, open_ai::Model::Custom { .. }) {
155 models.insert(model.id().to_string(), model);
156 }
157 }
158
159 // Override with available models from settings
160 for model in &OpenAiLanguageModelProvider::settings(cx).available_models {
161 models.insert(
162 model.name.clone(),
163 open_ai::Model::Custom {
164 name: model.name.clone(),
165 display_name: model.display_name.clone(),
166 max_tokens: model.max_tokens,
167 max_output_tokens: model.max_output_tokens,
168 max_completion_tokens: model.max_completion_tokens,
169 reasoning_effort: model.reasoning_effort.clone(),
170 supports_chat_completions: model.capabilities.chat_completions,
171 },
172 );
173 }
174
175 models
176 .into_values()
177 .map(|model| self.create_language_model(model))
178 .collect()
179 }
180
181 fn is_authenticated(&self, cx: &App) -> bool {
182 self.state.read(cx).is_authenticated()
183 }
184
185 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
186 self.state.update(cx, |state, cx| state.authenticate(cx))
187 }
188
189 fn configuration_view(
190 &self,
191 _target_agent: language_model::ConfigurationViewTargetAgent,
192 window: &mut Window,
193 cx: &mut App,
194 ) -> AnyView {
195 cx.new(|cx| ConfigurationView::new(self.state.clone(), window, cx))
196 .into()
197 }
198
199 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
200 self.state
201 .update(cx, |state, cx| state.set_api_key(None, cx))
202 }
203}
204
205pub struct OpenAiLanguageModel {
206 id: LanguageModelId,
207 model: open_ai::Model,
208 state: Entity<State>,
209 http_client: Arc<dyn HttpClient>,
210 request_limiter: RateLimiter,
211}
212
213impl OpenAiLanguageModel {
214 fn stream_completion(
215 &self,
216 request: open_ai::Request,
217 cx: &AsyncApp,
218 ) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>>>
219 {
220 let http_client = self.http_client.clone();
221
222 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
223 let api_url = OpenAiLanguageModelProvider::api_url(cx);
224 (state.api_key_state.key(&api_url), api_url)
225 });
226
227 let future = self.request_limiter.stream(async move {
228 let provider = PROVIDER_NAME;
229 let Some(api_key) = api_key else {
230 return Err(LanguageModelCompletionError::NoApiKey { provider });
231 };
232 let request = stream_completion(
233 http_client.as_ref(),
234 provider.0.as_str(),
235 &api_url,
236 &api_key,
237 request,
238 );
239 let response = request.await?;
240 Ok(response)
241 });
242
243 async move { Ok(future.await?.boxed()) }.boxed()
244 }
245
246 fn stream_response(
247 &self,
248 request: ResponseRequest,
249 cx: &AsyncApp,
250 ) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponsesStreamEvent>>>>
251 {
252 let http_client = self.http_client.clone();
253
254 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
255 let api_url = OpenAiLanguageModelProvider::api_url(cx);
256 (state.api_key_state.key(&api_url), api_url)
257 });
258
259 let provider = PROVIDER_NAME;
260 let future = self.request_limiter.stream(async move {
261 let Some(api_key) = api_key else {
262 return Err(LanguageModelCompletionError::NoApiKey { provider });
263 };
264 let request = stream_response(
265 http_client.as_ref(),
266 provider.0.as_str(),
267 &api_url,
268 &api_key,
269 request,
270 );
271 let response = request.await?;
272 Ok(response)
273 });
274
275 async move { Ok(future.await?.boxed()) }.boxed()
276 }
277}
278
279impl LanguageModel for OpenAiLanguageModel {
280 fn id(&self) -> LanguageModelId {
281 self.id.clone()
282 }
283
284 fn name(&self) -> LanguageModelName {
285 LanguageModelName::from(self.model.display_name().to_string())
286 }
287
288 fn provider_id(&self) -> LanguageModelProviderId {
289 PROVIDER_ID
290 }
291
292 fn provider_name(&self) -> LanguageModelProviderName {
293 PROVIDER_NAME
294 }
295
296 fn supports_tools(&self) -> bool {
297 true
298 }
299
300 fn supports_images(&self) -> bool {
301 use open_ai::Model;
302 match &self.model {
303 Model::FourOmniMini
304 | Model::FourPointOneNano
305 | Model::Five
306 | Model::FiveCodex
307 | Model::FiveMini
308 | Model::FiveNano
309 | Model::FivePointOne
310 | Model::FivePointTwo
311 | Model::FivePointTwoCodex
312 | Model::FivePointThreeCodex
313 | Model::O1
314 | Model::O3 => true,
315 Model::ThreePointFiveTurbo
316 | Model::Four
317 | Model::FourTurbo
318 | Model::O3Mini
319 | Model::Custom { .. } => false,
320 }
321 }
322
323 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
324 match choice {
325 LanguageModelToolChoice::Auto => true,
326 LanguageModelToolChoice::Any => true,
327 LanguageModelToolChoice::None => true,
328 }
329 }
330
331 fn supports_thinking(&self) -> bool {
332 self.model.reasoning_effort().is_some()
333 }
334
335 fn supports_split_token_display(&self) -> bool {
336 true
337 }
338
339 fn telemetry_id(&self) -> String {
340 format!("openai/{}", self.model.id())
341 }
342
343 fn max_token_count(&self) -> u64 {
344 self.model.max_token_count()
345 }
346
347 fn max_output_tokens(&self) -> Option<u64> {
348 self.model.max_output_tokens()
349 }
350
351 fn count_tokens(
352 &self,
353 request: LanguageModelRequest,
354 cx: &App,
355 ) -> BoxFuture<'static, Result<u64>> {
356 count_open_ai_tokens(request, self.model.clone(), cx)
357 }
358
359 fn stream_completion(
360 &self,
361 request: LanguageModelRequest,
362 cx: &AsyncApp,
363 ) -> BoxFuture<
364 'static,
365 Result<
366 futures::stream::BoxStream<
367 'static,
368 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
369 >,
370 LanguageModelCompletionError,
371 >,
372 > {
373 if self.model.supports_chat_completions() {
374 let request = into_open_ai(
375 request,
376 self.model.id(),
377 self.model.supports_parallel_tool_calls(),
378 self.model.supports_prompt_cache_key(),
379 self.max_output_tokens(),
380 self.model.reasoning_effort(),
381 );
382 let completions = self.stream_completion(request, cx);
383 async move {
384 let mapper = OpenAiEventMapper::new();
385 Ok(mapper.map_stream(completions.await?).boxed())
386 }
387 .boxed()
388 } else {
389 let request = into_open_ai_response(
390 request,
391 self.model.id(),
392 self.model.supports_parallel_tool_calls(),
393 self.model.supports_prompt_cache_key(),
394 self.max_output_tokens(),
395 self.model.reasoning_effort(),
396 );
397 let completions = self.stream_response(request, cx);
398 async move {
399 let mapper = OpenAiResponseEventMapper::new();
400 Ok(mapper.map_stream(completions.await?).boxed())
401 }
402 .boxed()
403 }
404 }
405}
406
407pub fn into_open_ai(
408 request: LanguageModelRequest,
409 model_id: &str,
410 supports_parallel_tool_calls: bool,
411 supports_prompt_cache_key: bool,
412 max_output_tokens: Option<u64>,
413 reasoning_effort: Option<ReasoningEffort>,
414) -> open_ai::Request {
415 let stream = !model_id.starts_with("o1-");
416
417 let mut messages = Vec::new();
418 for message in request.messages {
419 for content in message.content {
420 match content {
421 MessageContent::Text(text) | MessageContent::Thinking { text, .. } => {
422 let should_add = if message.role == Role::User {
423 // Including whitespace-only user messages can cause error with OpenAI compatible APIs
424 // See https://github.com/zed-industries/zed/issues/40097
425 !text.trim().is_empty()
426 } else {
427 !text.is_empty()
428 };
429 if should_add {
430 add_message_content_part(
431 open_ai::MessagePart::Text { text },
432 message.role,
433 &mut messages,
434 );
435 }
436 }
437 MessageContent::RedactedThinking(_) => {}
438 MessageContent::Image(image) => {
439 add_message_content_part(
440 open_ai::MessagePart::Image {
441 image_url: ImageUrl {
442 url: image.to_base64_url(),
443 detail: None,
444 },
445 },
446 message.role,
447 &mut messages,
448 );
449 }
450 MessageContent::ToolUse(tool_use) => {
451 let tool_call = open_ai::ToolCall {
452 id: tool_use.id.to_string(),
453 content: open_ai::ToolCallContent::Function {
454 function: open_ai::FunctionContent {
455 name: tool_use.name.to_string(),
456 arguments: serde_json::to_string(&tool_use.input)
457 .unwrap_or_default(),
458 },
459 },
460 };
461
462 if let Some(open_ai::RequestMessage::Assistant { tool_calls, .. }) =
463 messages.last_mut()
464 {
465 tool_calls.push(tool_call);
466 } else {
467 messages.push(open_ai::RequestMessage::Assistant {
468 content: None,
469 tool_calls: vec![tool_call],
470 });
471 }
472 }
473 MessageContent::ToolResult(tool_result) => {
474 let content = match &tool_result.content {
475 LanguageModelToolResultContent::Text(text) => {
476 vec![open_ai::MessagePart::Text {
477 text: text.to_string(),
478 }]
479 }
480 LanguageModelToolResultContent::Image(image) => {
481 vec![open_ai::MessagePart::Image {
482 image_url: ImageUrl {
483 url: image.to_base64_url(),
484 detail: None,
485 },
486 }]
487 }
488 };
489
490 messages.push(open_ai::RequestMessage::Tool {
491 content: content.into(),
492 tool_call_id: tool_result.tool_use_id.to_string(),
493 });
494 }
495 }
496 }
497 }
498
499 open_ai::Request {
500 model: model_id.into(),
501 messages,
502 stream,
503 stop: request.stop,
504 temperature: request.temperature.or(Some(1.0)),
505 max_completion_tokens: max_output_tokens,
506 parallel_tool_calls: if supports_parallel_tool_calls && !request.tools.is_empty() {
507 // Disable parallel tool calls, as the Agent currently expects a maximum of one per turn.
508 Some(false)
509 } else {
510 None
511 },
512 prompt_cache_key: if supports_prompt_cache_key {
513 request.thread_id
514 } else {
515 None
516 },
517 tools: request
518 .tools
519 .into_iter()
520 .map(|tool| open_ai::ToolDefinition::Function {
521 function: open_ai::FunctionDefinition {
522 name: tool.name,
523 description: Some(tool.description),
524 parameters: Some(tool.input_schema),
525 },
526 })
527 .collect(),
528 tool_choice: request.tool_choice.map(|choice| match choice {
529 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
530 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
531 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
532 }),
533 reasoning_effort,
534 }
535}
536
537pub fn into_open_ai_response(
538 request: LanguageModelRequest,
539 model_id: &str,
540 supports_parallel_tool_calls: bool,
541 supports_prompt_cache_key: bool,
542 max_output_tokens: Option<u64>,
543 reasoning_effort: Option<ReasoningEffort>,
544) -> ResponseRequest {
545 let stream = !model_id.starts_with("o1-");
546
547 let LanguageModelRequest {
548 thread_id,
549 prompt_id: _,
550 intent: _,
551 messages,
552 tools,
553 tool_choice,
554 stop: _,
555 temperature,
556 thinking_allowed: _,
557 thinking_effort: _,
558 } = request;
559
560 let mut input_items = Vec::new();
561 for (index, message) in messages.into_iter().enumerate() {
562 append_message_to_response_items(message, index, &mut input_items);
563 }
564
565 let tools: Vec<_> = tools
566 .into_iter()
567 .map(|tool| open_ai::responses::ToolDefinition::Function {
568 name: tool.name,
569 description: Some(tool.description),
570 parameters: Some(tool.input_schema),
571 strict: None,
572 })
573 .collect();
574
575 ResponseRequest {
576 model: model_id.into(),
577 input: input_items,
578 stream,
579 temperature,
580 top_p: None,
581 max_output_tokens,
582 parallel_tool_calls: if tools.is_empty() {
583 None
584 } else {
585 Some(supports_parallel_tool_calls)
586 },
587 tool_choice: tool_choice.map(|choice| match choice {
588 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
589 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
590 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
591 }),
592 tools,
593 prompt_cache_key: if supports_prompt_cache_key {
594 thread_id
595 } else {
596 None
597 },
598 reasoning: reasoning_effort.map(|effort| open_ai::responses::ReasoningConfig { effort }),
599 }
600}
601
602fn append_message_to_response_items(
603 message: LanguageModelRequestMessage,
604 index: usize,
605 input_items: &mut Vec<ResponseInputItem>,
606) {
607 let mut content_parts: Vec<ResponseInputContent> = Vec::new();
608
609 for content in message.content {
610 match content {
611 MessageContent::Text(text) => {
612 push_response_text_part(&message.role, text, &mut content_parts);
613 }
614 MessageContent::Thinking { text, .. } => {
615 push_response_text_part(&message.role, text, &mut content_parts);
616 }
617 MessageContent::RedactedThinking(_) => {}
618 MessageContent::Image(image) => {
619 push_response_image_part(&message.role, image, &mut content_parts);
620 }
621 MessageContent::ToolUse(tool_use) => {
622 flush_response_parts(&message.role, index, &mut content_parts, input_items);
623 let call_id = tool_use.id.to_string();
624 input_items.push(ResponseInputItem::FunctionCall(ResponseFunctionCallItem {
625 call_id,
626 name: tool_use.name.to_string(),
627 arguments: tool_use.raw_input,
628 }));
629 }
630 MessageContent::ToolResult(tool_result) => {
631 flush_response_parts(&message.role, index, &mut content_parts, input_items);
632 input_items.push(ResponseInputItem::FunctionCallOutput(
633 ResponseFunctionCallOutputItem {
634 call_id: tool_result.tool_use_id.to_string(),
635 output: tool_result_output(&tool_result),
636 },
637 ));
638 }
639 }
640 }
641
642 flush_response_parts(&message.role, index, &mut content_parts, input_items);
643}
644
645fn push_response_text_part(
646 role: &Role,
647 text: impl Into<String>,
648 parts: &mut Vec<ResponseInputContent>,
649) {
650 let text = text.into();
651 if text.trim().is_empty() {
652 return;
653 }
654
655 match role {
656 Role::Assistant => parts.push(ResponseInputContent::OutputText {
657 text,
658 annotations: Vec::new(),
659 }),
660 _ => parts.push(ResponseInputContent::Text { text }),
661 }
662}
663
664fn push_response_image_part(
665 role: &Role,
666 image: LanguageModelImage,
667 parts: &mut Vec<ResponseInputContent>,
668) {
669 match role {
670 Role::Assistant => parts.push(ResponseInputContent::OutputText {
671 text: "[image omitted]".to_string(),
672 annotations: Vec::new(),
673 }),
674 _ => parts.push(ResponseInputContent::Image {
675 image_url: image.to_base64_url(),
676 }),
677 }
678}
679
680fn flush_response_parts(
681 role: &Role,
682 _index: usize,
683 parts: &mut Vec<ResponseInputContent>,
684 input_items: &mut Vec<ResponseInputItem>,
685) {
686 if parts.is_empty() {
687 return;
688 }
689
690 let item = ResponseInputItem::Message(ResponseMessageItem {
691 role: match role {
692 Role::User => open_ai::Role::User,
693 Role::Assistant => open_ai::Role::Assistant,
694 Role::System => open_ai::Role::System,
695 },
696 content: parts.clone(),
697 });
698
699 input_items.push(item);
700 parts.clear();
701}
702
703fn tool_result_output(result: &LanguageModelToolResult) -> String {
704 if let Some(output) = &result.output {
705 match output {
706 serde_json::Value::String(text) => text.clone(),
707 serde_json::Value::Null => String::new(),
708 _ => output.to_string(),
709 }
710 } else {
711 match &result.content {
712 LanguageModelToolResultContent::Text(text) => text.to_string(),
713 LanguageModelToolResultContent::Image(image) => image.to_base64_url(),
714 }
715 }
716}
717
718fn add_message_content_part(
719 new_part: open_ai::MessagePart,
720 role: Role,
721 messages: &mut Vec<open_ai::RequestMessage>,
722) {
723 match (role, messages.last_mut()) {
724 (Role::User, Some(open_ai::RequestMessage::User { content }))
725 | (
726 Role::Assistant,
727 Some(open_ai::RequestMessage::Assistant {
728 content: Some(content),
729 ..
730 }),
731 )
732 | (Role::System, Some(open_ai::RequestMessage::System { content, .. })) => {
733 content.push_part(new_part);
734 }
735 _ => {
736 messages.push(match role {
737 Role::User => open_ai::RequestMessage::User {
738 content: open_ai::MessageContent::from(vec![new_part]),
739 },
740 Role::Assistant => open_ai::RequestMessage::Assistant {
741 content: Some(open_ai::MessageContent::from(vec![new_part])),
742 tool_calls: Vec::new(),
743 },
744 Role::System => open_ai::RequestMessage::System {
745 content: open_ai::MessageContent::from(vec![new_part]),
746 },
747 });
748 }
749 }
750}
751
752pub struct OpenAiEventMapper {
753 tool_calls_by_index: HashMap<usize, RawToolCall>,
754}
755
756impl OpenAiEventMapper {
757 pub fn new() -> Self {
758 Self {
759 tool_calls_by_index: HashMap::default(),
760 }
761 }
762
763 pub fn map_stream(
764 mut self,
765 events: Pin<Box<dyn Send + Stream<Item = Result<ResponseStreamEvent>>>>,
766 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
767 {
768 events.flat_map(move |event| {
769 futures::stream::iter(match event {
770 Ok(event) => self.map_event(event),
771 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
772 })
773 })
774 }
775
776 pub fn map_event(
777 &mut self,
778 event: ResponseStreamEvent,
779 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
780 let mut events = Vec::new();
781 if let Some(usage) = event.usage {
782 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
783 input_tokens: usage.prompt_tokens,
784 output_tokens: usage.completion_tokens,
785 cache_creation_input_tokens: 0,
786 cache_read_input_tokens: 0,
787 })));
788 }
789
790 let Some(choice) = event.choices.first() else {
791 return events;
792 };
793
794 if let Some(delta) = choice.delta.as_ref() {
795 if let Some(reasoning_content) = delta.reasoning_content.clone() {
796 if !reasoning_content.is_empty() {
797 events.push(Ok(LanguageModelCompletionEvent::Thinking {
798 text: reasoning_content,
799 signature: None,
800 }));
801 }
802 }
803 if let Some(content) = delta.content.clone() {
804 if !content.is_empty() {
805 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
806 }
807 }
808
809 if let Some(tool_calls) = delta.tool_calls.as_ref() {
810 for tool_call in tool_calls {
811 let entry = self.tool_calls_by_index.entry(tool_call.index).or_default();
812
813 if let Some(tool_id) = tool_call.id.clone() {
814 entry.id = tool_id;
815 }
816
817 if let Some(function) = tool_call.function.as_ref() {
818 if let Some(name) = function.name.clone() {
819 entry.name = name;
820 }
821
822 if let Some(arguments) = function.arguments.clone() {
823 entry.arguments.push_str(&arguments);
824 }
825 }
826 }
827 }
828 }
829
830 match choice.finish_reason.as_deref() {
831 Some("stop") => {
832 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
833 }
834 Some("tool_calls") => {
835 events.extend(self.tool_calls_by_index.drain().map(|(_, tool_call)| {
836 match parse_tool_arguments(&tool_call.arguments) {
837 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
838 LanguageModelToolUse {
839 id: tool_call.id.clone().into(),
840 name: tool_call.name.as_str().into(),
841 is_input_complete: true,
842 input,
843 raw_input: tool_call.arguments.clone(),
844 thought_signature: None,
845 },
846 )),
847 Err(error) => Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
848 id: tool_call.id.into(),
849 tool_name: tool_call.name.into(),
850 raw_input: tool_call.arguments.clone().into(),
851 json_parse_error: error.to_string(),
852 }),
853 }
854 }));
855
856 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
857 }
858 Some(stop_reason) => {
859 log::error!("Unexpected OpenAI stop_reason: {stop_reason:?}",);
860 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
861 }
862 None => {}
863 }
864
865 events
866 }
867}
868
869#[derive(Default)]
870struct RawToolCall {
871 id: String,
872 name: String,
873 arguments: String,
874}
875
876pub struct OpenAiResponseEventMapper {
877 function_calls_by_item: HashMap<String, PendingResponseFunctionCall>,
878 pending_stop_reason: Option<StopReason>,
879}
880
881#[derive(Default)]
882struct PendingResponseFunctionCall {
883 call_id: String,
884 name: Arc<str>,
885 arguments: String,
886}
887
888impl OpenAiResponseEventMapper {
889 pub fn new() -> Self {
890 Self {
891 function_calls_by_item: HashMap::default(),
892 pending_stop_reason: None,
893 }
894 }
895
896 pub fn map_stream(
897 mut self,
898 events: Pin<Box<dyn Send + Stream<Item = Result<ResponsesStreamEvent>>>>,
899 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
900 {
901 events.flat_map(move |event| {
902 futures::stream::iter(match event {
903 Ok(event) => self.map_event(event),
904 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
905 })
906 })
907 }
908
909 pub fn map_event(
910 &mut self,
911 event: ResponsesStreamEvent,
912 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
913 match event {
914 ResponsesStreamEvent::OutputItemAdded { item, .. } => {
915 let mut events = Vec::new();
916
917 match &item {
918 ResponseOutputItem::Message(message) => {
919 if let Some(id) = &message.id {
920 events.push(Ok(LanguageModelCompletionEvent::StartMessage {
921 message_id: id.clone(),
922 }));
923 }
924 }
925 ResponseOutputItem::FunctionCall(function_call) => {
926 if let Some(item_id) = function_call.id.clone() {
927 let call_id = function_call
928 .call_id
929 .clone()
930 .or_else(|| function_call.id.clone())
931 .unwrap_or_else(|| item_id.clone());
932 let entry = PendingResponseFunctionCall {
933 call_id,
934 name: Arc::<str>::from(
935 function_call.name.clone().unwrap_or_default(),
936 ),
937 arguments: function_call.arguments.clone(),
938 };
939 self.function_calls_by_item.insert(item_id, entry);
940 }
941 }
942 ResponseOutputItem::Unknown => {}
943 }
944 events
945 }
946 ResponsesStreamEvent::OutputTextDelta { delta, .. } => {
947 if delta.is_empty() {
948 Vec::new()
949 } else {
950 vec![Ok(LanguageModelCompletionEvent::Text(delta))]
951 }
952 }
953 ResponsesStreamEvent::FunctionCallArgumentsDelta { item_id, delta, .. } => {
954 if let Some(entry) = self.function_calls_by_item.get_mut(&item_id) {
955 entry.arguments.push_str(&delta);
956 }
957 Vec::new()
958 }
959 ResponsesStreamEvent::FunctionCallArgumentsDone {
960 item_id, arguments, ..
961 } => {
962 if let Some(mut entry) = self.function_calls_by_item.remove(&item_id) {
963 if !arguments.is_empty() {
964 entry.arguments = arguments;
965 }
966 let raw_input = entry.arguments.clone();
967 self.pending_stop_reason = Some(StopReason::ToolUse);
968 match parse_tool_arguments(&entry.arguments) {
969 Ok(input) => {
970 vec![Ok(LanguageModelCompletionEvent::ToolUse(
971 LanguageModelToolUse {
972 id: LanguageModelToolUseId::from(entry.call_id.clone()),
973 name: entry.name.clone(),
974 is_input_complete: true,
975 input,
976 raw_input,
977 thought_signature: None,
978 },
979 ))]
980 }
981 Err(error) => {
982 vec![Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
983 id: LanguageModelToolUseId::from(entry.call_id.clone()),
984 tool_name: entry.name.clone(),
985 raw_input: Arc::<str>::from(raw_input),
986 json_parse_error: error.to_string(),
987 })]
988 }
989 }
990 } else {
991 Vec::new()
992 }
993 }
994 ResponsesStreamEvent::Completed { response } => {
995 self.handle_completion(response, StopReason::EndTurn)
996 }
997 ResponsesStreamEvent::Incomplete { response } => {
998 let reason = response
999 .status_details
1000 .as_ref()
1001 .and_then(|details| details.reason.as_deref());
1002 let stop_reason = match reason {
1003 Some("max_output_tokens") => StopReason::MaxTokens,
1004 Some("content_filter") => {
1005 self.pending_stop_reason = Some(StopReason::Refusal);
1006 StopReason::Refusal
1007 }
1008 _ => self
1009 .pending_stop_reason
1010 .take()
1011 .unwrap_or(StopReason::EndTurn),
1012 };
1013
1014 let mut events = Vec::new();
1015 if self.pending_stop_reason.is_none() {
1016 events.extend(self.emit_tool_calls_from_output(&response.output));
1017 }
1018 if let Some(usage) = response.usage.as_ref() {
1019 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1020 token_usage_from_response_usage(usage),
1021 )));
1022 }
1023 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1024 events
1025 }
1026 ResponsesStreamEvent::Failed { response } => {
1027 let message = response
1028 .status_details
1029 .and_then(|details| details.error)
1030 .map(|error| error.to_string())
1031 .unwrap_or_else(|| "response failed".to_string());
1032 vec![Err(LanguageModelCompletionError::Other(anyhow!(message)))]
1033 }
1034 ResponsesStreamEvent::Error { error }
1035 | ResponsesStreamEvent::GenericError { error } => {
1036 vec![Err(LanguageModelCompletionError::Other(anyhow!(format!(
1037 "{error:?}"
1038 ))))]
1039 }
1040 ResponsesStreamEvent::OutputTextDone { .. } => Vec::new(),
1041 ResponsesStreamEvent::OutputItemDone { .. }
1042 | ResponsesStreamEvent::ContentPartAdded { .. }
1043 | ResponsesStreamEvent::ContentPartDone { .. }
1044 | ResponsesStreamEvent::Created { .. }
1045 | ResponsesStreamEvent::InProgress { .. }
1046 | ResponsesStreamEvent::Unknown => Vec::new(),
1047 }
1048 }
1049
1050 fn handle_completion(
1051 &mut self,
1052 response: ResponsesSummary,
1053 default_reason: StopReason,
1054 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1055 let mut events = Vec::new();
1056
1057 if self.pending_stop_reason.is_none() {
1058 events.extend(self.emit_tool_calls_from_output(&response.output));
1059 }
1060
1061 if let Some(usage) = response.usage.as_ref() {
1062 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1063 token_usage_from_response_usage(usage),
1064 )));
1065 }
1066
1067 let stop_reason = self.pending_stop_reason.take().unwrap_or(default_reason);
1068 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1069 events
1070 }
1071
1072 fn emit_tool_calls_from_output(
1073 &mut self,
1074 output: &[ResponseOutputItem],
1075 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1076 let mut events = Vec::new();
1077 for item in output {
1078 if let ResponseOutputItem::FunctionCall(function_call) = item {
1079 let Some(call_id) = function_call
1080 .call_id
1081 .clone()
1082 .or_else(|| function_call.id.clone())
1083 else {
1084 log::error!(
1085 "Function call item missing both call_id and id: {:?}",
1086 function_call
1087 );
1088 continue;
1089 };
1090 let name: Arc<str> = Arc::from(function_call.name.clone().unwrap_or_default());
1091 let arguments = &function_call.arguments;
1092 self.pending_stop_reason = Some(StopReason::ToolUse);
1093 match parse_tool_arguments(arguments) {
1094 Ok(input) => {
1095 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
1096 LanguageModelToolUse {
1097 id: LanguageModelToolUseId::from(call_id.clone()),
1098 name: name.clone(),
1099 is_input_complete: true,
1100 input,
1101 raw_input: arguments.clone(),
1102 thought_signature: None,
1103 },
1104 )));
1105 }
1106 Err(error) => {
1107 events.push(Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1108 id: LanguageModelToolUseId::from(call_id.clone()),
1109 tool_name: name.clone(),
1110 raw_input: Arc::<str>::from(arguments.clone()),
1111 json_parse_error: error.to_string(),
1112 }));
1113 }
1114 }
1115 }
1116 }
1117 events
1118 }
1119}
1120
1121fn token_usage_from_response_usage(usage: &ResponsesUsage) -> TokenUsage {
1122 TokenUsage {
1123 input_tokens: usage.input_tokens.unwrap_or_default(),
1124 output_tokens: usage.output_tokens.unwrap_or_default(),
1125 cache_creation_input_tokens: 0,
1126 cache_read_input_tokens: 0,
1127 }
1128}
1129
1130pub(crate) fn collect_tiktoken_messages(
1131 request: LanguageModelRequest,
1132) -> Vec<tiktoken_rs::ChatCompletionRequestMessage> {
1133 request
1134 .messages
1135 .into_iter()
1136 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
1137 role: match message.role {
1138 Role::User => "user".into(),
1139 Role::Assistant => "assistant".into(),
1140 Role::System => "system".into(),
1141 },
1142 content: Some(message.string_contents()),
1143 name: None,
1144 function_call: None,
1145 })
1146 .collect::<Vec<_>>()
1147}
1148
1149pub fn count_open_ai_tokens(
1150 request: LanguageModelRequest,
1151 model: Model,
1152 cx: &App,
1153) -> BoxFuture<'static, Result<u64>> {
1154 cx.background_spawn(async move {
1155 let messages = collect_tiktoken_messages(request);
1156 match model {
1157 Model::Custom { max_tokens, .. } => {
1158 let model = if max_tokens >= 100_000 {
1159 // If the max tokens is 100k or more, it likely uses the o200k_base tokenizer
1160 "gpt-4o"
1161 } else {
1162 // Otherwise fallback to gpt-4, since only cl100k_base and o200k_base are
1163 // supported with this tiktoken method
1164 "gpt-4"
1165 };
1166 tiktoken_rs::num_tokens_from_messages(model, &messages)
1167 }
1168 // Currently supported by tiktoken_rs
1169 // Sometimes tiktoken-rs is behind on model support. If that is the case, make a new branch
1170 // arm with an override. We enumerate all supported models here so that we can check if new
1171 // models are supported yet or not.
1172 Model::ThreePointFiveTurbo
1173 | Model::Four
1174 | Model::FourTurbo
1175 | Model::FourOmniMini
1176 | Model::FourPointOneNano
1177 | Model::O1
1178 | Model::O3
1179 | Model::O3Mini
1180 | Model::Five
1181 | Model::FiveCodex
1182 | Model::FiveMini
1183 | Model::FiveNano => tiktoken_rs::num_tokens_from_messages(model.id(), &messages),
1184 // GPT-5.1, 5.2, 5.2-codex, and 5.3-codex don't have dedicated tiktoken support; use gpt-5 tokenizer
1185 Model::FivePointOne
1186 | Model::FivePointTwo
1187 | Model::FivePointTwoCodex
1188 | Model::FivePointThreeCodex => {
1189 tiktoken_rs::num_tokens_from_messages("gpt-5", &messages)
1190 }
1191 }
1192 .map(|tokens| tokens as u64)
1193 })
1194 .boxed()
1195}
1196
1197struct ConfigurationView {
1198 api_key_editor: Entity<InputField>,
1199 state: Entity<State>,
1200 load_credentials_task: Option<Task<()>>,
1201}
1202
1203impl ConfigurationView {
1204 fn new(state: Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
1205 let api_key_editor = cx.new(|cx| {
1206 InputField::new(
1207 window,
1208 cx,
1209 "sk-000000000000000000000000000000000000000000000000",
1210 )
1211 });
1212
1213 cx.observe(&state, |_, _, cx| {
1214 cx.notify();
1215 })
1216 .detach();
1217
1218 let load_credentials_task = Some(cx.spawn_in(window, {
1219 let state = state.clone();
1220 async move |this, cx| {
1221 if let Some(task) = Some(state.update(cx, |state, cx| state.authenticate(cx))) {
1222 // We don't log an error, because "not signed in" is also an error.
1223 let _ = task.await;
1224 }
1225 this.update(cx, |this, cx| {
1226 this.load_credentials_task = None;
1227 cx.notify();
1228 })
1229 .log_err();
1230 }
1231 }));
1232
1233 Self {
1234 api_key_editor,
1235 state,
1236 load_credentials_task,
1237 }
1238 }
1239
1240 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
1241 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
1242 if api_key.is_empty() {
1243 return;
1244 }
1245
1246 // url changes can cause the editor to be displayed again
1247 self.api_key_editor
1248 .update(cx, |editor, cx| editor.set_text("", window, cx));
1249
1250 let state = self.state.clone();
1251 cx.spawn_in(window, async move |_, cx| {
1252 state
1253 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
1254 .await
1255 })
1256 .detach_and_log_err(cx);
1257 }
1258
1259 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
1260 self.api_key_editor
1261 .update(cx, |input, cx| input.set_text("", window, cx));
1262
1263 let state = self.state.clone();
1264 cx.spawn_in(window, async move |_, cx| {
1265 state
1266 .update(cx, |state, cx| state.set_api_key(None, cx))
1267 .await
1268 })
1269 .detach_and_log_err(cx);
1270 }
1271
1272 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
1273 !self.state.read(cx).is_authenticated()
1274 }
1275}
1276
1277impl Render for ConfigurationView {
1278 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1279 let env_var_set = self.state.read(cx).api_key_state.is_from_env_var();
1280 let configured_card_label = if env_var_set {
1281 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable")
1282 } else {
1283 let api_url = OpenAiLanguageModelProvider::api_url(cx);
1284 if api_url == OPEN_AI_API_URL {
1285 "API key configured".to_string()
1286 } else {
1287 format!("API key configured for {}", api_url)
1288 }
1289 };
1290
1291 let api_key_section = if self.should_render_editor(cx) {
1292 v_flex()
1293 .on_action(cx.listener(Self::save_api_key))
1294 .child(Label::new("To use Zed's agent with OpenAI, you need to add an API key. Follow these steps:"))
1295 .child(
1296 List::new()
1297 .child(
1298 ListBulletItem::new("")
1299 .child(Label::new("Create one by visiting"))
1300 .child(ButtonLink::new("OpenAI's console", "https://platform.openai.com/api-keys"))
1301 )
1302 .child(
1303 ListBulletItem::new("Ensure your OpenAI account has credits")
1304 )
1305 .child(
1306 ListBulletItem::new("Paste your API key below and hit enter to start using the agent")
1307 ),
1308 )
1309 .child(self.api_key_editor.clone())
1310 .child(
1311 Label::new(format!(
1312 "You can also set the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed."
1313 ))
1314 .size(LabelSize::Small)
1315 .color(Color::Muted),
1316 )
1317 .child(
1318 Label::new(
1319 "Note that having a subscription for another service like GitHub Copilot won't work.",
1320 )
1321 .size(LabelSize::Small).color(Color::Muted),
1322 )
1323 .into_any_element()
1324 } else {
1325 ConfiguredApiCard::new(configured_card_label)
1326 .disabled(env_var_set)
1327 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
1328 .when(env_var_set, |this| {
1329 this.tooltip_label(format!("To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable."))
1330 })
1331 .into_any_element()
1332 };
1333
1334 let compatible_api_section = h_flex()
1335 .mt_1p5()
1336 .gap_0p5()
1337 .flex_wrap()
1338 .when(self.should_render_editor(cx), |this| {
1339 this.pt_1p5()
1340 .border_t_1()
1341 .border_color(cx.theme().colors().border_variant)
1342 })
1343 .child(
1344 h_flex()
1345 .gap_2()
1346 .child(
1347 Icon::new(IconName::Info)
1348 .size(IconSize::XSmall)
1349 .color(Color::Muted),
1350 )
1351 .child(Label::new("Zed also supports OpenAI-compatible models.")),
1352 )
1353 .child(
1354 Button::new("docs", "Learn More")
1355 .icon(IconName::ArrowUpRight)
1356 .icon_size(IconSize::Small)
1357 .icon_color(Color::Muted)
1358 .on_click(move |_, _window, cx| {
1359 cx.open_url("https://zed.dev/docs/ai/llm-providers#openai-api-compatible")
1360 }),
1361 );
1362
1363 if self.load_credentials_task.is_some() {
1364 div().child(Label::new("Loading credentials…")).into_any()
1365 } else {
1366 v_flex()
1367 .size_full()
1368 .child(api_key_section)
1369 .child(compatible_api_section)
1370 .into_any()
1371 }
1372 }
1373}
1374
1375#[cfg(test)]
1376mod tests {
1377 use futures::{StreamExt, executor::block_on};
1378 use gpui::TestAppContext;
1379 use language_model::{LanguageModelRequestMessage, LanguageModelRequestTool};
1380 use open_ai::responses::{
1381 ResponseFunctionToolCall, ResponseOutputItem, ResponseOutputMessage, ResponseStatusDetails,
1382 ResponseSummary, ResponseUsage, StreamEvent as ResponsesStreamEvent,
1383 };
1384 use pretty_assertions::assert_eq;
1385 use serde_json::json;
1386
1387 use super::*;
1388
1389 fn map_response_events(events: Vec<ResponsesStreamEvent>) -> Vec<LanguageModelCompletionEvent> {
1390 block_on(async {
1391 OpenAiResponseEventMapper::new()
1392 .map_stream(Box::pin(futures::stream::iter(events.into_iter().map(Ok))))
1393 .collect::<Vec<_>>()
1394 .await
1395 .into_iter()
1396 .map(Result::unwrap)
1397 .collect()
1398 })
1399 }
1400
1401 fn response_item_message(id: &str) -> ResponseOutputItem {
1402 ResponseOutputItem::Message(ResponseOutputMessage {
1403 id: Some(id.to_string()),
1404 role: Some("assistant".to_string()),
1405 status: Some("in_progress".to_string()),
1406 content: vec![],
1407 })
1408 }
1409
1410 fn response_item_function_call(id: &str, args: Option<&str>) -> ResponseOutputItem {
1411 ResponseOutputItem::FunctionCall(ResponseFunctionToolCall {
1412 id: Some(id.to_string()),
1413 status: Some("in_progress".to_string()),
1414 name: Some("get_weather".to_string()),
1415 call_id: Some("call_123".to_string()),
1416 arguments: args.map(|s| s.to_string()).unwrap_or_default(),
1417 })
1418 }
1419
1420 #[gpui::test]
1421 fn tiktoken_rs_support(cx: &TestAppContext) {
1422 let request = LanguageModelRequest {
1423 thread_id: None,
1424 prompt_id: None,
1425 intent: None,
1426 messages: vec![LanguageModelRequestMessage {
1427 role: Role::User,
1428 content: vec![MessageContent::Text("message".into())],
1429 cache: false,
1430 reasoning_details: None,
1431 }],
1432 tools: vec![],
1433 tool_choice: None,
1434 stop: vec![],
1435 temperature: None,
1436 thinking_allowed: true,
1437 thinking_effort: None,
1438 };
1439
1440 // Validate that all models are supported by tiktoken-rs
1441 for model in Model::iter() {
1442 let count = cx
1443 .foreground_executor()
1444 .block_on(count_open_ai_tokens(
1445 request.clone(),
1446 model,
1447 &cx.app.borrow(),
1448 ))
1449 .unwrap();
1450 assert!(count > 0);
1451 }
1452 }
1453
1454 #[test]
1455 fn responses_stream_maps_text_and_usage() {
1456 let events = vec![
1457 ResponsesStreamEvent::OutputItemAdded {
1458 output_index: 0,
1459 sequence_number: None,
1460 item: response_item_message("msg_123"),
1461 },
1462 ResponsesStreamEvent::OutputTextDelta {
1463 item_id: "msg_123".into(),
1464 output_index: 0,
1465 content_index: Some(0),
1466 delta: "Hello".into(),
1467 },
1468 ResponsesStreamEvent::Completed {
1469 response: ResponseSummary {
1470 usage: Some(ResponseUsage {
1471 input_tokens: Some(5),
1472 output_tokens: Some(3),
1473 total_tokens: Some(8),
1474 }),
1475 ..Default::default()
1476 },
1477 },
1478 ];
1479
1480 let mapped = map_response_events(events);
1481 assert!(matches!(
1482 mapped[0],
1483 LanguageModelCompletionEvent::StartMessage { ref message_id } if message_id == "msg_123"
1484 ));
1485 assert!(matches!(
1486 mapped[1],
1487 LanguageModelCompletionEvent::Text(ref text) if text == "Hello"
1488 ));
1489 assert!(matches!(
1490 mapped[2],
1491 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1492 input_tokens: 5,
1493 output_tokens: 3,
1494 ..
1495 })
1496 ));
1497 assert!(matches!(
1498 mapped[3],
1499 LanguageModelCompletionEvent::Stop(StopReason::EndTurn)
1500 ));
1501 }
1502
1503 #[test]
1504 fn into_open_ai_response_builds_complete_payload() {
1505 let tool_call_id = LanguageModelToolUseId::from("call-42");
1506 let tool_input = json!({ "city": "Boston" });
1507 let tool_arguments = serde_json::to_string(&tool_input).unwrap();
1508 let tool_use = LanguageModelToolUse {
1509 id: tool_call_id.clone(),
1510 name: Arc::from("get_weather"),
1511 raw_input: tool_arguments.clone(),
1512 input: tool_input,
1513 is_input_complete: true,
1514 thought_signature: None,
1515 };
1516 let tool_result = LanguageModelToolResult {
1517 tool_use_id: tool_call_id,
1518 tool_name: Arc::from("get_weather"),
1519 is_error: false,
1520 content: LanguageModelToolResultContent::Text(Arc::from("Sunny")),
1521 output: Some(json!({ "forecast": "Sunny" })),
1522 };
1523 let user_image = LanguageModelImage {
1524 source: SharedString::from("aGVsbG8="),
1525 size: None,
1526 };
1527 let expected_image_url = user_image.to_base64_url();
1528
1529 let request = LanguageModelRequest {
1530 thread_id: Some("thread-123".into()),
1531 prompt_id: None,
1532 intent: None,
1533 messages: vec![
1534 LanguageModelRequestMessage {
1535 role: Role::System,
1536 content: vec![MessageContent::Text("System context".into())],
1537 cache: false,
1538 reasoning_details: None,
1539 },
1540 LanguageModelRequestMessage {
1541 role: Role::User,
1542 content: vec![
1543 MessageContent::Text("Please check the weather.".into()),
1544 MessageContent::Image(user_image),
1545 ],
1546 cache: false,
1547 reasoning_details: None,
1548 },
1549 LanguageModelRequestMessage {
1550 role: Role::Assistant,
1551 content: vec![
1552 MessageContent::Text("Looking that up.".into()),
1553 MessageContent::ToolUse(tool_use),
1554 ],
1555 cache: false,
1556 reasoning_details: None,
1557 },
1558 LanguageModelRequestMessage {
1559 role: Role::Assistant,
1560 content: vec![MessageContent::ToolResult(tool_result)],
1561 cache: false,
1562 reasoning_details: None,
1563 },
1564 ],
1565 tools: vec![LanguageModelRequestTool {
1566 name: "get_weather".into(),
1567 description: "Fetches the weather".into(),
1568 input_schema: json!({ "type": "object" }),
1569 use_input_streaming: false,
1570 }],
1571 tool_choice: Some(LanguageModelToolChoice::Any),
1572 stop: vec!["<STOP>".into()],
1573 temperature: None,
1574 thinking_allowed: false,
1575 thinking_effort: None,
1576 };
1577
1578 let response = into_open_ai_response(
1579 request,
1580 "custom-model",
1581 true,
1582 true,
1583 Some(2048),
1584 Some(ReasoningEffort::Low),
1585 );
1586
1587 let serialized = serde_json::to_value(&response).unwrap();
1588 let expected = json!({
1589 "model": "custom-model",
1590 "input": [
1591 {
1592 "type": "message",
1593 "role": "system",
1594 "content": [
1595 { "type": "input_text", "text": "System context" }
1596 ]
1597 },
1598 {
1599 "type": "message",
1600 "role": "user",
1601 "content": [
1602 { "type": "input_text", "text": "Please check the weather." },
1603 { "type": "input_image", "image_url": expected_image_url }
1604 ]
1605 },
1606 {
1607 "type": "message",
1608 "role": "assistant",
1609 "content": [
1610 { "type": "output_text", "text": "Looking that up.", "annotations": [] }
1611 ]
1612 },
1613 {
1614 "type": "function_call",
1615 "call_id": "call-42",
1616 "name": "get_weather",
1617 "arguments": tool_arguments
1618 },
1619 {
1620 "type": "function_call_output",
1621 "call_id": "call-42",
1622 "output": "{\"forecast\":\"Sunny\"}"
1623 }
1624 ],
1625 "stream": true,
1626 "max_output_tokens": 2048,
1627 "parallel_tool_calls": true,
1628 "tool_choice": "required",
1629 "tools": [
1630 {
1631 "type": "function",
1632 "name": "get_weather",
1633 "description": "Fetches the weather",
1634 "parameters": { "type": "object" }
1635 }
1636 ],
1637 "prompt_cache_key": "thread-123",
1638 "reasoning": { "effort": "low" }
1639 });
1640
1641 assert_eq!(serialized, expected);
1642 }
1643
1644 #[test]
1645 fn responses_stream_maps_tool_calls() {
1646 let events = vec![
1647 ResponsesStreamEvent::OutputItemAdded {
1648 output_index: 0,
1649 sequence_number: None,
1650 item: response_item_function_call("item_fn", Some("{\"city\":\"Bos")),
1651 },
1652 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1653 item_id: "item_fn".into(),
1654 output_index: 0,
1655 delta: "ton\"}".into(),
1656 sequence_number: None,
1657 },
1658 ResponsesStreamEvent::FunctionCallArgumentsDone {
1659 item_id: "item_fn".into(),
1660 output_index: 0,
1661 arguments: "{\"city\":\"Boston\"}".into(),
1662 sequence_number: None,
1663 },
1664 ResponsesStreamEvent::Completed {
1665 response: ResponseSummary::default(),
1666 },
1667 ];
1668
1669 let mapped = map_response_events(events);
1670 assert!(matches!(
1671 mapped[0],
1672 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1673 ref id,
1674 ref name,
1675 ref raw_input,
1676 ..
1677 }) if id.to_string() == "call_123"
1678 && name.as_ref() == "get_weather"
1679 && raw_input == "{\"city\":\"Boston\"}"
1680 ));
1681 assert!(matches!(
1682 mapped[1],
1683 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1684 ));
1685 }
1686
1687 #[test]
1688 fn responses_stream_uses_max_tokens_stop_reason() {
1689 let events = vec![ResponsesStreamEvent::Incomplete {
1690 response: ResponseSummary {
1691 status_details: Some(ResponseStatusDetails {
1692 reason: Some("max_output_tokens".into()),
1693 r#type: Some("incomplete".into()),
1694 error: None,
1695 }),
1696 usage: Some(ResponseUsage {
1697 input_tokens: Some(10),
1698 output_tokens: Some(20),
1699 total_tokens: Some(30),
1700 }),
1701 ..Default::default()
1702 },
1703 }];
1704
1705 let mapped = map_response_events(events);
1706 assert!(matches!(
1707 mapped[0],
1708 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1709 input_tokens: 10,
1710 output_tokens: 20,
1711 ..
1712 })
1713 ));
1714 assert!(matches!(
1715 mapped[1],
1716 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1717 ));
1718 }
1719
1720 #[test]
1721 fn responses_stream_handles_multiple_tool_calls() {
1722 let events = vec![
1723 ResponsesStreamEvent::OutputItemAdded {
1724 output_index: 0,
1725 sequence_number: None,
1726 item: response_item_function_call("item_fn1", Some("{\"city\":\"NYC\"}")),
1727 },
1728 ResponsesStreamEvent::FunctionCallArgumentsDone {
1729 item_id: "item_fn1".into(),
1730 output_index: 0,
1731 arguments: "{\"city\":\"NYC\"}".into(),
1732 sequence_number: None,
1733 },
1734 ResponsesStreamEvent::OutputItemAdded {
1735 output_index: 1,
1736 sequence_number: None,
1737 item: response_item_function_call("item_fn2", Some("{\"city\":\"LA\"}")),
1738 },
1739 ResponsesStreamEvent::FunctionCallArgumentsDone {
1740 item_id: "item_fn2".into(),
1741 output_index: 1,
1742 arguments: "{\"city\":\"LA\"}".into(),
1743 sequence_number: None,
1744 },
1745 ResponsesStreamEvent::Completed {
1746 response: ResponseSummary::default(),
1747 },
1748 ];
1749
1750 let mapped = map_response_events(events);
1751 assert_eq!(mapped.len(), 3);
1752 assert!(matches!(
1753 mapped[0],
1754 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1755 if raw_input == "{\"city\":\"NYC\"}"
1756 ));
1757 assert!(matches!(
1758 mapped[1],
1759 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1760 if raw_input == "{\"city\":\"LA\"}"
1761 ));
1762 assert!(matches!(
1763 mapped[2],
1764 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1765 ));
1766 }
1767
1768 #[test]
1769 fn responses_stream_handles_mixed_text_and_tool_calls() {
1770 let events = vec![
1771 ResponsesStreamEvent::OutputItemAdded {
1772 output_index: 0,
1773 sequence_number: None,
1774 item: response_item_message("msg_123"),
1775 },
1776 ResponsesStreamEvent::OutputTextDelta {
1777 item_id: "msg_123".into(),
1778 output_index: 0,
1779 content_index: Some(0),
1780 delta: "Let me check that".into(),
1781 },
1782 ResponsesStreamEvent::OutputItemAdded {
1783 output_index: 1,
1784 sequence_number: None,
1785 item: response_item_function_call("item_fn", Some("{\"query\":\"test\"}")),
1786 },
1787 ResponsesStreamEvent::FunctionCallArgumentsDone {
1788 item_id: "item_fn".into(),
1789 output_index: 1,
1790 arguments: "{\"query\":\"test\"}".into(),
1791 sequence_number: None,
1792 },
1793 ResponsesStreamEvent::Completed {
1794 response: ResponseSummary::default(),
1795 },
1796 ];
1797
1798 let mapped = map_response_events(events);
1799 assert!(matches!(
1800 mapped[0],
1801 LanguageModelCompletionEvent::StartMessage { .. }
1802 ));
1803 assert!(matches!(
1804 mapped[1],
1805 LanguageModelCompletionEvent::Text(ref text) if text == "Let me check that"
1806 ));
1807 assert!(matches!(
1808 mapped[2],
1809 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1810 if raw_input == "{\"query\":\"test\"}"
1811 ));
1812 assert!(matches!(
1813 mapped[3],
1814 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1815 ));
1816 }
1817
1818 #[test]
1819 fn responses_stream_handles_json_parse_error() {
1820 let events = vec![
1821 ResponsesStreamEvent::OutputItemAdded {
1822 output_index: 0,
1823 sequence_number: None,
1824 item: response_item_function_call("item_fn", Some("{invalid json")),
1825 },
1826 ResponsesStreamEvent::FunctionCallArgumentsDone {
1827 item_id: "item_fn".into(),
1828 output_index: 0,
1829 arguments: "{invalid json".into(),
1830 sequence_number: None,
1831 },
1832 ResponsesStreamEvent::Completed {
1833 response: ResponseSummary::default(),
1834 },
1835 ];
1836
1837 let mapped = map_response_events(events);
1838 assert!(matches!(
1839 mapped[0],
1840 LanguageModelCompletionEvent::ToolUseJsonParseError {
1841 ref raw_input,
1842 ..
1843 } if raw_input.as_ref() == "{invalid json"
1844 ));
1845 }
1846
1847 #[test]
1848 fn responses_stream_handles_incomplete_function_call() {
1849 let events = vec![
1850 ResponsesStreamEvent::OutputItemAdded {
1851 output_index: 0,
1852 sequence_number: None,
1853 item: response_item_function_call("item_fn", Some("{\"city\":")),
1854 },
1855 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1856 item_id: "item_fn".into(),
1857 output_index: 0,
1858 delta: "\"Boston\"".into(),
1859 sequence_number: None,
1860 },
1861 ResponsesStreamEvent::Incomplete {
1862 response: ResponseSummary {
1863 status_details: Some(ResponseStatusDetails {
1864 reason: Some("max_output_tokens".into()),
1865 r#type: Some("incomplete".into()),
1866 error: None,
1867 }),
1868 output: vec![response_item_function_call(
1869 "item_fn",
1870 Some("{\"city\":\"Boston\"}"),
1871 )],
1872 ..Default::default()
1873 },
1874 },
1875 ];
1876
1877 let mapped = map_response_events(events);
1878 assert!(matches!(
1879 mapped[0],
1880 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1881 if raw_input == "{\"city\":\"Boston\"}"
1882 ));
1883 assert!(matches!(
1884 mapped[1],
1885 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1886 ));
1887 }
1888
1889 #[test]
1890 fn responses_stream_incomplete_does_not_duplicate_tool_calls() {
1891 let events = vec![
1892 ResponsesStreamEvent::OutputItemAdded {
1893 output_index: 0,
1894 sequence_number: None,
1895 item: response_item_function_call("item_fn", Some("{\"city\":\"Boston\"}")),
1896 },
1897 ResponsesStreamEvent::FunctionCallArgumentsDone {
1898 item_id: "item_fn".into(),
1899 output_index: 0,
1900 arguments: "{\"city\":\"Boston\"}".into(),
1901 sequence_number: None,
1902 },
1903 ResponsesStreamEvent::Incomplete {
1904 response: ResponseSummary {
1905 status_details: Some(ResponseStatusDetails {
1906 reason: Some("max_output_tokens".into()),
1907 r#type: Some("incomplete".into()),
1908 error: None,
1909 }),
1910 output: vec![response_item_function_call(
1911 "item_fn",
1912 Some("{\"city\":\"Boston\"}"),
1913 )],
1914 ..Default::default()
1915 },
1916 },
1917 ];
1918
1919 let mapped = map_response_events(events);
1920 assert_eq!(mapped.len(), 2);
1921 assert!(matches!(
1922 mapped[0],
1923 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1924 if raw_input == "{\"city\":\"Boston\"}"
1925 ));
1926 assert!(matches!(
1927 mapped[1],
1928 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1929 ));
1930 }
1931
1932 #[test]
1933 fn responses_stream_handles_empty_tool_arguments() {
1934 // Test that tools with no arguments (empty string) are handled correctly
1935 let events = vec![
1936 ResponsesStreamEvent::OutputItemAdded {
1937 output_index: 0,
1938 sequence_number: None,
1939 item: response_item_function_call("item_fn", Some("")),
1940 },
1941 ResponsesStreamEvent::FunctionCallArgumentsDone {
1942 item_id: "item_fn".into(),
1943 output_index: 0,
1944 arguments: "".into(),
1945 sequence_number: None,
1946 },
1947 ResponsesStreamEvent::Completed {
1948 response: ResponseSummary::default(),
1949 },
1950 ];
1951
1952 let mapped = map_response_events(events);
1953 assert_eq!(mapped.len(), 2);
1954
1955 // Should produce a ToolUse event with an empty object
1956 assert!(matches!(
1957 &mapped[0],
1958 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1959 id,
1960 name,
1961 raw_input,
1962 input,
1963 ..
1964 }) if id.to_string() == "call_123"
1965 && name.as_ref() == "get_weather"
1966 && raw_input == ""
1967 && input.is_object()
1968 && input.as_object().unwrap().is_empty()
1969 ));
1970
1971 assert!(matches!(
1972 mapped[1],
1973 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1974 ));
1975 }
1976}