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::FivePointTwoCodex
314 | Model::O1
315 | Model::O3
316 | Model::O4Mini => true,
317 Model::ThreePointFiveTurbo
318 | Model::Four
319 | Model::FourTurbo
320 | Model::O3Mini
321 | Model::Custom { .. } => false,
322 }
323 }
324
325 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
326 match choice {
327 LanguageModelToolChoice::Auto => true,
328 LanguageModelToolChoice::Any => true,
329 LanguageModelToolChoice::None => true,
330 }
331 }
332
333 fn supports_split_token_display(&self) -> bool {
334 true
335 }
336
337 fn telemetry_id(&self) -> String {
338 format!("openai/{}", self.model.id())
339 }
340
341 fn max_token_count(&self) -> u64 {
342 self.model.max_token_count()
343 }
344
345 fn max_output_tokens(&self) -> Option<u64> {
346 self.model.max_output_tokens()
347 }
348
349 fn count_tokens(
350 &self,
351 request: LanguageModelRequest,
352 cx: &App,
353 ) -> BoxFuture<'static, Result<u64>> {
354 count_open_ai_tokens(request, self.model.clone(), cx)
355 }
356
357 fn stream_completion(
358 &self,
359 request: LanguageModelRequest,
360 cx: &AsyncApp,
361 ) -> BoxFuture<
362 'static,
363 Result<
364 futures::stream::BoxStream<
365 'static,
366 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
367 >,
368 LanguageModelCompletionError,
369 >,
370 > {
371 if self.model.supports_chat_completions() {
372 let request = into_open_ai(
373 request,
374 self.model.id(),
375 self.model.supports_parallel_tool_calls(),
376 self.model.supports_prompt_cache_key(),
377 self.max_output_tokens(),
378 self.model.reasoning_effort(),
379 );
380 let completions = self.stream_completion(request, cx);
381 async move {
382 let mapper = OpenAiEventMapper::new();
383 Ok(mapper.map_stream(completions.await?).boxed())
384 }
385 .boxed()
386 } else {
387 let request = into_open_ai_response(
388 request,
389 self.model.id(),
390 self.model.supports_parallel_tool_calls(),
391 self.model.supports_prompt_cache_key(),
392 self.max_output_tokens(),
393 self.model.reasoning_effort(),
394 );
395 let completions = self.stream_response(request, cx);
396 async move {
397 let mapper = OpenAiResponseEventMapper::new();
398 Ok(mapper.map_stream(completions.await?).boxed())
399 }
400 .boxed()
401 }
402 }
403}
404
405pub fn into_open_ai(
406 request: LanguageModelRequest,
407 model_id: &str,
408 supports_parallel_tool_calls: bool,
409 supports_prompt_cache_key: bool,
410 max_output_tokens: Option<u64>,
411 reasoning_effort: Option<ReasoningEffort>,
412) -> open_ai::Request {
413 let stream = !model_id.starts_with("o1-");
414
415 let mut messages = Vec::new();
416 for message in request.messages {
417 for content in message.content {
418 match content {
419 MessageContent::Text(text) | MessageContent::Thinking { text, .. } => {
420 let should_add = if message.role == Role::User {
421 // Including whitespace-only user messages can cause error with OpenAI compatible APIs
422 // See https://github.com/zed-industries/zed/issues/40097
423 !text.trim().is_empty()
424 } else {
425 !text.is_empty()
426 };
427 if should_add {
428 add_message_content_part(
429 open_ai::MessagePart::Text { text },
430 message.role,
431 &mut messages,
432 );
433 }
434 }
435 MessageContent::RedactedThinking(_) => {}
436 MessageContent::Image(image) => {
437 add_message_content_part(
438 open_ai::MessagePart::Image {
439 image_url: ImageUrl {
440 url: image.to_base64_url(),
441 detail: None,
442 },
443 },
444 message.role,
445 &mut messages,
446 );
447 }
448 MessageContent::ToolUse(tool_use) => {
449 let tool_call = open_ai::ToolCall {
450 id: tool_use.id.to_string(),
451 content: open_ai::ToolCallContent::Function {
452 function: open_ai::FunctionContent {
453 name: tool_use.name.to_string(),
454 arguments: serde_json::to_string(&tool_use.input)
455 .unwrap_or_default(),
456 },
457 },
458 };
459
460 if let Some(open_ai::RequestMessage::Assistant { tool_calls, .. }) =
461 messages.last_mut()
462 {
463 tool_calls.push(tool_call);
464 } else {
465 messages.push(open_ai::RequestMessage::Assistant {
466 content: None,
467 tool_calls: vec![tool_call],
468 });
469 }
470 }
471 MessageContent::ToolResult(tool_result) => {
472 let content = match &tool_result.content {
473 LanguageModelToolResultContent::Text(text) => {
474 vec![open_ai::MessagePart::Text {
475 text: text.to_string(),
476 }]
477 }
478 LanguageModelToolResultContent::Image(image) => {
479 vec![open_ai::MessagePart::Image {
480 image_url: ImageUrl {
481 url: image.to_base64_url(),
482 detail: None,
483 },
484 }]
485 }
486 };
487
488 messages.push(open_ai::RequestMessage::Tool {
489 content: content.into(),
490 tool_call_id: tool_result.tool_use_id.to_string(),
491 });
492 }
493 }
494 }
495 }
496
497 open_ai::Request {
498 model: model_id.into(),
499 messages,
500 stream,
501 stop: request.stop,
502 temperature: request.temperature.or(Some(1.0)),
503 max_completion_tokens: max_output_tokens,
504 parallel_tool_calls: if supports_parallel_tool_calls && !request.tools.is_empty() {
505 // Disable parallel tool calls, as the Agent currently expects a maximum of one per turn.
506 Some(false)
507 } else {
508 None
509 },
510 prompt_cache_key: if supports_prompt_cache_key {
511 request.thread_id
512 } else {
513 None
514 },
515 tools: request
516 .tools
517 .into_iter()
518 .map(|tool| open_ai::ToolDefinition::Function {
519 function: open_ai::FunctionDefinition {
520 name: tool.name,
521 description: Some(tool.description),
522 parameters: Some(tool.input_schema),
523 },
524 })
525 .collect(),
526 tool_choice: request.tool_choice.map(|choice| match choice {
527 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
528 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
529 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
530 }),
531 reasoning_effort,
532 }
533}
534
535pub fn into_open_ai_response(
536 request: LanguageModelRequest,
537 model_id: &str,
538 supports_parallel_tool_calls: bool,
539 supports_prompt_cache_key: bool,
540 max_output_tokens: Option<u64>,
541 reasoning_effort: Option<ReasoningEffort>,
542) -> ResponseRequest {
543 let stream = !model_id.starts_with("o1-");
544
545 let LanguageModelRequest {
546 thread_id,
547 prompt_id: _,
548 intent: _,
549 messages,
550 tools,
551 tool_choice,
552 stop: _,
553 temperature,
554 thinking_allowed: _,
555 } = request;
556
557 let mut input_items = Vec::new();
558 for (index, message) in messages.into_iter().enumerate() {
559 append_message_to_response_items(message, index, &mut input_items);
560 }
561
562 let tools: Vec<_> = tools
563 .into_iter()
564 .map(|tool| open_ai::responses::ToolDefinition::Function {
565 name: tool.name,
566 description: Some(tool.description),
567 parameters: Some(tool.input_schema),
568 strict: None,
569 })
570 .collect();
571
572 ResponseRequest {
573 model: model_id.into(),
574 input: input_items,
575 stream,
576 temperature,
577 top_p: None,
578 max_output_tokens,
579 parallel_tool_calls: if tools.is_empty() {
580 None
581 } else {
582 Some(supports_parallel_tool_calls)
583 },
584 tool_choice: tool_choice.map(|choice| match choice {
585 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
586 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
587 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
588 }),
589 tools,
590 prompt_cache_key: if supports_prompt_cache_key {
591 thread_id
592 } else {
593 None
594 },
595 reasoning: reasoning_effort.map(|effort| open_ai::responses::ReasoningConfig { effort }),
596 }
597}
598
599fn append_message_to_response_items(
600 message: LanguageModelRequestMessage,
601 index: usize,
602 input_items: &mut Vec<ResponseInputItem>,
603) {
604 let mut content_parts: Vec<ResponseInputContent> = Vec::new();
605
606 for content in message.content {
607 match content {
608 MessageContent::Text(text) => {
609 push_response_text_part(&message.role, text, &mut content_parts);
610 }
611 MessageContent::Thinking { text, .. } => {
612 push_response_text_part(&message.role, text, &mut content_parts);
613 }
614 MessageContent::RedactedThinking(_) => {}
615 MessageContent::Image(image) => {
616 push_response_image_part(&message.role, image, &mut content_parts);
617 }
618 MessageContent::ToolUse(tool_use) => {
619 flush_response_parts(&message.role, index, &mut content_parts, input_items);
620 let call_id = tool_use.id.to_string();
621 input_items.push(ResponseInputItem::FunctionCall(ResponseFunctionCallItem {
622 call_id,
623 name: tool_use.name.to_string(),
624 arguments: tool_use.raw_input,
625 }));
626 }
627 MessageContent::ToolResult(tool_result) => {
628 flush_response_parts(&message.role, index, &mut content_parts, input_items);
629 input_items.push(ResponseInputItem::FunctionCallOutput(
630 ResponseFunctionCallOutputItem {
631 call_id: tool_result.tool_use_id.to_string(),
632 output: tool_result_output(&tool_result),
633 },
634 ));
635 }
636 }
637 }
638
639 flush_response_parts(&message.role, index, &mut content_parts, input_items);
640}
641
642fn push_response_text_part(
643 role: &Role,
644 text: impl Into<String>,
645 parts: &mut Vec<ResponseInputContent>,
646) {
647 let text = text.into();
648 if text.trim().is_empty() {
649 return;
650 }
651
652 match role {
653 Role::Assistant => parts.push(ResponseInputContent::OutputText {
654 text,
655 annotations: Vec::new(),
656 }),
657 _ => parts.push(ResponseInputContent::Text { text }),
658 }
659}
660
661fn push_response_image_part(
662 role: &Role,
663 image: LanguageModelImage,
664 parts: &mut Vec<ResponseInputContent>,
665) {
666 match role {
667 Role::Assistant => parts.push(ResponseInputContent::OutputText {
668 text: "[image omitted]".to_string(),
669 annotations: Vec::new(),
670 }),
671 _ => parts.push(ResponseInputContent::Image {
672 image_url: image.to_base64_url(),
673 }),
674 }
675}
676
677fn flush_response_parts(
678 role: &Role,
679 _index: usize,
680 parts: &mut Vec<ResponseInputContent>,
681 input_items: &mut Vec<ResponseInputItem>,
682) {
683 if parts.is_empty() {
684 return;
685 }
686
687 let item = ResponseInputItem::Message(ResponseMessageItem {
688 role: match role {
689 Role::User => open_ai::Role::User,
690 Role::Assistant => open_ai::Role::Assistant,
691 Role::System => open_ai::Role::System,
692 },
693 content: parts.clone(),
694 });
695
696 input_items.push(item);
697 parts.clear();
698}
699
700fn tool_result_output(result: &LanguageModelToolResult) -> String {
701 if let Some(output) = &result.output {
702 match output {
703 serde_json::Value::String(text) => text.clone(),
704 serde_json::Value::Null => String::new(),
705 _ => output.to_string(),
706 }
707 } else {
708 match &result.content {
709 LanguageModelToolResultContent::Text(text) => text.to_string(),
710 LanguageModelToolResultContent::Image(image) => image.to_base64_url(),
711 }
712 }
713}
714
715fn add_message_content_part(
716 new_part: open_ai::MessagePart,
717 role: Role,
718 messages: &mut Vec<open_ai::RequestMessage>,
719) {
720 match (role, messages.last_mut()) {
721 (Role::User, Some(open_ai::RequestMessage::User { content }))
722 | (
723 Role::Assistant,
724 Some(open_ai::RequestMessage::Assistant {
725 content: Some(content),
726 ..
727 }),
728 )
729 | (Role::System, Some(open_ai::RequestMessage::System { content, .. })) => {
730 content.push_part(new_part);
731 }
732 _ => {
733 messages.push(match role {
734 Role::User => open_ai::RequestMessage::User {
735 content: open_ai::MessageContent::from(vec![new_part]),
736 },
737 Role::Assistant => open_ai::RequestMessage::Assistant {
738 content: Some(open_ai::MessageContent::from(vec![new_part])),
739 tool_calls: Vec::new(),
740 },
741 Role::System => open_ai::RequestMessage::System {
742 content: open_ai::MessageContent::from(vec![new_part]),
743 },
744 });
745 }
746 }
747}
748
749pub struct OpenAiEventMapper {
750 tool_calls_by_index: HashMap<usize, RawToolCall>,
751}
752
753impl OpenAiEventMapper {
754 pub fn new() -> Self {
755 Self {
756 tool_calls_by_index: HashMap::default(),
757 }
758 }
759
760 pub fn map_stream(
761 mut self,
762 events: Pin<Box<dyn Send + Stream<Item = Result<ResponseStreamEvent>>>>,
763 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
764 {
765 events.flat_map(move |event| {
766 futures::stream::iter(match event {
767 Ok(event) => self.map_event(event),
768 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
769 })
770 })
771 }
772
773 pub fn map_event(
774 &mut self,
775 event: ResponseStreamEvent,
776 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
777 let mut events = Vec::new();
778 if let Some(usage) = event.usage {
779 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
780 input_tokens: usage.prompt_tokens,
781 output_tokens: usage.completion_tokens,
782 cache_creation_input_tokens: 0,
783 cache_read_input_tokens: 0,
784 })));
785 }
786
787 let Some(choice) = event.choices.first() else {
788 return events;
789 };
790
791 if let Some(delta) = choice.delta.as_ref() {
792 if let Some(reasoning_content) = delta.reasoning_content.clone() {
793 if !reasoning_content.is_empty() {
794 events.push(Ok(LanguageModelCompletionEvent::Thinking {
795 text: reasoning_content,
796 signature: None,
797 }));
798 }
799 }
800 if let Some(content) = delta.content.clone() {
801 if !content.is_empty() {
802 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
803 }
804 }
805
806 if let Some(tool_calls) = delta.tool_calls.as_ref() {
807 for tool_call in tool_calls {
808 let entry = self.tool_calls_by_index.entry(tool_call.index).or_default();
809
810 if let Some(tool_id) = tool_call.id.clone() {
811 entry.id = tool_id;
812 }
813
814 if let Some(function) = tool_call.function.as_ref() {
815 if let Some(name) = function.name.clone() {
816 entry.name = name;
817 }
818
819 if let Some(arguments) = function.arguments.clone() {
820 entry.arguments.push_str(&arguments);
821 }
822 }
823 }
824 }
825 }
826
827 match choice.finish_reason.as_deref() {
828 Some("stop") => {
829 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
830 }
831 Some("tool_calls") => {
832 events.extend(self.tool_calls_by_index.drain().map(|(_, tool_call)| {
833 match serde_json::Value::from_str(&tool_call.arguments) {
834 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
835 LanguageModelToolUse {
836 id: tool_call.id.clone().into(),
837 name: tool_call.name.as_str().into(),
838 is_input_complete: true,
839 input,
840 raw_input: tool_call.arguments.clone(),
841 thought_signature: None,
842 },
843 )),
844 Err(error) => Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
845 id: tool_call.id.into(),
846 tool_name: tool_call.name.into(),
847 raw_input: tool_call.arguments.clone().into(),
848 json_parse_error: error.to_string(),
849 }),
850 }
851 }));
852
853 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
854 }
855 Some(stop_reason) => {
856 log::error!("Unexpected OpenAI stop_reason: {stop_reason:?}",);
857 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
858 }
859 None => {}
860 }
861
862 events
863 }
864}
865
866#[derive(Default)]
867struct RawToolCall {
868 id: String,
869 name: String,
870 arguments: String,
871}
872
873pub struct OpenAiResponseEventMapper {
874 function_calls_by_item: HashMap<String, PendingResponseFunctionCall>,
875 pending_stop_reason: Option<StopReason>,
876}
877
878#[derive(Default)]
879struct PendingResponseFunctionCall {
880 call_id: String,
881 name: Arc<str>,
882 arguments: String,
883}
884
885impl OpenAiResponseEventMapper {
886 pub fn new() -> Self {
887 Self {
888 function_calls_by_item: HashMap::default(),
889 pending_stop_reason: None,
890 }
891 }
892
893 pub fn map_stream(
894 mut self,
895 events: Pin<Box<dyn Send + Stream<Item = Result<ResponsesStreamEvent>>>>,
896 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
897 {
898 events.flat_map(move |event| {
899 futures::stream::iter(match event {
900 Ok(event) => self.map_event(event),
901 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
902 })
903 })
904 }
905
906 pub fn map_event(
907 &mut self,
908 event: ResponsesStreamEvent,
909 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
910 match event {
911 ResponsesStreamEvent::OutputItemAdded { item, .. } => {
912 let mut events = Vec::new();
913
914 match &item {
915 ResponseOutputItem::Message(message) => {
916 if let Some(id) = &message.id {
917 events.push(Ok(LanguageModelCompletionEvent::StartMessage {
918 message_id: id.clone(),
919 }));
920 }
921 }
922 ResponseOutputItem::FunctionCall(function_call) => {
923 if let Some(item_id) = function_call.id.clone() {
924 let call_id = function_call
925 .call_id
926 .clone()
927 .or_else(|| function_call.id.clone())
928 .unwrap_or_else(|| item_id.clone());
929 let entry = PendingResponseFunctionCall {
930 call_id,
931 name: Arc::<str>::from(
932 function_call.name.clone().unwrap_or_default(),
933 ),
934 arguments: function_call.arguments.clone(),
935 };
936 self.function_calls_by_item.insert(item_id, entry);
937 }
938 }
939 ResponseOutputItem::Unknown => {}
940 }
941 events
942 }
943 ResponsesStreamEvent::OutputTextDelta { delta, .. } => {
944 if delta.is_empty() {
945 Vec::new()
946 } else {
947 vec![Ok(LanguageModelCompletionEvent::Text(delta))]
948 }
949 }
950 ResponsesStreamEvent::FunctionCallArgumentsDelta { item_id, delta, .. } => {
951 if let Some(entry) = self.function_calls_by_item.get_mut(&item_id) {
952 entry.arguments.push_str(&delta);
953 }
954 Vec::new()
955 }
956 ResponsesStreamEvent::FunctionCallArgumentsDone {
957 item_id, arguments, ..
958 } => {
959 if let Some(mut entry) = self.function_calls_by_item.remove(&item_id) {
960 if !arguments.is_empty() {
961 entry.arguments = arguments;
962 }
963 let raw_input = entry.arguments.clone();
964 self.pending_stop_reason = Some(StopReason::ToolUse);
965 match serde_json::from_str::<serde_json::Value>(&entry.arguments) {
966 Ok(input) => {
967 vec![Ok(LanguageModelCompletionEvent::ToolUse(
968 LanguageModelToolUse {
969 id: LanguageModelToolUseId::from(entry.call_id.clone()),
970 name: entry.name.clone(),
971 is_input_complete: true,
972 input,
973 raw_input,
974 thought_signature: None,
975 },
976 ))]
977 }
978 Err(error) => {
979 vec![Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
980 id: LanguageModelToolUseId::from(entry.call_id.clone()),
981 tool_name: entry.name.clone(),
982 raw_input: Arc::<str>::from(raw_input),
983 json_parse_error: error.to_string(),
984 })]
985 }
986 }
987 } else {
988 Vec::new()
989 }
990 }
991 ResponsesStreamEvent::Completed { response } => {
992 self.handle_completion(response, StopReason::EndTurn)
993 }
994 ResponsesStreamEvent::Incomplete { response } => {
995 let reason = response
996 .status_details
997 .as_ref()
998 .and_then(|details| details.reason.as_deref());
999 let stop_reason = match reason {
1000 Some("max_output_tokens") => StopReason::MaxTokens,
1001 Some("content_filter") => {
1002 self.pending_stop_reason = Some(StopReason::Refusal);
1003 StopReason::Refusal
1004 }
1005 _ => self
1006 .pending_stop_reason
1007 .take()
1008 .unwrap_or(StopReason::EndTurn),
1009 };
1010
1011 let mut events = Vec::new();
1012 if self.pending_stop_reason.is_none() {
1013 events.extend(self.emit_tool_calls_from_output(&response.output));
1014 }
1015 if let Some(usage) = response.usage.as_ref() {
1016 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1017 token_usage_from_response_usage(usage),
1018 )));
1019 }
1020 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1021 events
1022 }
1023 ResponsesStreamEvent::Failed { response } => {
1024 let message = response
1025 .status_details
1026 .and_then(|details| details.error)
1027 .map(|error| error.to_string())
1028 .unwrap_or_else(|| "response failed".to_string());
1029 vec![Err(LanguageModelCompletionError::Other(anyhow!(message)))]
1030 }
1031 ResponsesStreamEvent::Error { error }
1032 | ResponsesStreamEvent::GenericError { error } => {
1033 vec![Err(LanguageModelCompletionError::Other(anyhow!(format!(
1034 "{error:?}"
1035 ))))]
1036 }
1037 ResponsesStreamEvent::OutputTextDone { .. } => Vec::new(),
1038 ResponsesStreamEvent::OutputItemDone { .. }
1039 | ResponsesStreamEvent::ContentPartAdded { .. }
1040 | ResponsesStreamEvent::ContentPartDone { .. }
1041 | ResponsesStreamEvent::Created { .. }
1042 | ResponsesStreamEvent::InProgress { .. }
1043 | ResponsesStreamEvent::Unknown => Vec::new(),
1044 }
1045 }
1046
1047 fn handle_completion(
1048 &mut self,
1049 response: ResponsesSummary,
1050 default_reason: StopReason,
1051 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1052 let mut events = Vec::new();
1053
1054 if self.pending_stop_reason.is_none() {
1055 events.extend(self.emit_tool_calls_from_output(&response.output));
1056 }
1057
1058 if let Some(usage) = response.usage.as_ref() {
1059 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1060 token_usage_from_response_usage(usage),
1061 )));
1062 }
1063
1064 let stop_reason = self.pending_stop_reason.take().unwrap_or(default_reason);
1065 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1066 events
1067 }
1068
1069 fn emit_tool_calls_from_output(
1070 &mut self,
1071 output: &[ResponseOutputItem],
1072 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1073 let mut events = Vec::new();
1074 for item in output {
1075 if let ResponseOutputItem::FunctionCall(function_call) = item {
1076 let Some(call_id) = function_call
1077 .call_id
1078 .clone()
1079 .or_else(|| function_call.id.clone())
1080 else {
1081 log::error!(
1082 "Function call item missing both call_id and id: {:?}",
1083 function_call
1084 );
1085 continue;
1086 };
1087 let name: Arc<str> = Arc::from(function_call.name.clone().unwrap_or_default());
1088 let arguments = &function_call.arguments;
1089 if !arguments.is_empty() {
1090 self.pending_stop_reason = Some(StopReason::ToolUse);
1091 match serde_json::from_str::<serde_json::Value>(arguments) {
1092 Ok(input) => {
1093 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
1094 LanguageModelToolUse {
1095 id: LanguageModelToolUseId::from(call_id.clone()),
1096 name: name.clone(),
1097 is_input_complete: true,
1098 input,
1099 raw_input: arguments.clone(),
1100 thought_signature: None,
1101 },
1102 )));
1103 }
1104 Err(error) => {
1105 events.push(Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1106 id: LanguageModelToolUseId::from(call_id.clone()),
1107 tool_name: name.clone(),
1108 raw_input: Arc::<str>::from(arguments.clone()),
1109 json_parse_error: error.to_string(),
1110 }));
1111 }
1112 }
1113 }
1114 }
1115 }
1116 events
1117 }
1118}
1119
1120fn token_usage_from_response_usage(usage: &ResponsesUsage) -> TokenUsage {
1121 TokenUsage {
1122 input_tokens: usage.input_tokens.unwrap_or_default(),
1123 output_tokens: usage.output_tokens.unwrap_or_default(),
1124 cache_creation_input_tokens: 0,
1125 cache_read_input_tokens: 0,
1126 }
1127}
1128
1129pub(crate) fn collect_tiktoken_messages(
1130 request: LanguageModelRequest,
1131) -> Vec<tiktoken_rs::ChatCompletionRequestMessage> {
1132 request
1133 .messages
1134 .into_iter()
1135 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
1136 role: match message.role {
1137 Role::User => "user".into(),
1138 Role::Assistant => "assistant".into(),
1139 Role::System => "system".into(),
1140 },
1141 content: Some(message.string_contents()),
1142 name: None,
1143 function_call: None,
1144 })
1145 .collect::<Vec<_>>()
1146}
1147
1148pub fn count_open_ai_tokens(
1149 request: LanguageModelRequest,
1150 model: Model,
1151 cx: &App,
1152) -> BoxFuture<'static, Result<u64>> {
1153 cx.background_spawn(async move {
1154 let messages = collect_tiktoken_messages(request);
1155 match model {
1156 Model::Custom { max_tokens, .. } => {
1157 let model = if max_tokens >= 100_000 {
1158 // If the max tokens is 100k or more, it is likely the o200k_base tokenizer from gpt4o
1159 "gpt-4o"
1160 } else {
1161 // Otherwise fallback to gpt-4, since only cl100k_base and o200k_base are
1162 // supported with this tiktoken method
1163 "gpt-4"
1164 };
1165 tiktoken_rs::num_tokens_from_messages(model, &messages)
1166 }
1167 // Currently supported by tiktoken_rs
1168 // Sometimes tiktoken-rs is behind on model support. If that is the case, make a new branch
1169 // arm with an override. We enumerate all supported models here so that we can check if new
1170 // models are supported yet or not.
1171 Model::ThreePointFiveTurbo
1172 | Model::Four
1173 | Model::FourTurbo
1174 | Model::FourOmni
1175 | Model::FourOmniMini
1176 | Model::FourPointOne
1177 | Model::FourPointOneMini
1178 | Model::FourPointOneNano
1179 | Model::O1
1180 | Model::O3
1181 | Model::O3Mini
1182 | Model::O4Mini
1183 | Model::Five
1184 | Model::FiveCodex
1185 | Model::FiveMini
1186 | Model::FiveNano => tiktoken_rs::num_tokens_from_messages(model.id(), &messages),
1187 // GPT-5.1, 5.2, and 5.2-codex don't have dedicated tiktoken support; use gpt-5 tokenizer
1188 Model::FivePointOne | Model::FivePointTwo | Model::FivePointTwoCodex => {
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 };
1438
1439 // Validate that all models are supported by tiktoken-rs
1440 for model in Model::iter() {
1441 let count = cx
1442 .foreground_executor()
1443 .block_on(count_open_ai_tokens(
1444 request.clone(),
1445 model,
1446 &cx.app.borrow(),
1447 ))
1448 .unwrap();
1449 assert!(count > 0);
1450 }
1451 }
1452
1453 #[test]
1454 fn responses_stream_maps_text_and_usage() {
1455 let events = vec![
1456 ResponsesStreamEvent::OutputItemAdded {
1457 output_index: 0,
1458 sequence_number: None,
1459 item: response_item_message("msg_123"),
1460 },
1461 ResponsesStreamEvent::OutputTextDelta {
1462 item_id: "msg_123".into(),
1463 output_index: 0,
1464 content_index: Some(0),
1465 delta: "Hello".into(),
1466 },
1467 ResponsesStreamEvent::Completed {
1468 response: ResponseSummary {
1469 usage: Some(ResponseUsage {
1470 input_tokens: Some(5),
1471 output_tokens: Some(3),
1472 total_tokens: Some(8),
1473 }),
1474 ..Default::default()
1475 },
1476 },
1477 ];
1478
1479 let mapped = map_response_events(events);
1480 assert!(matches!(
1481 mapped[0],
1482 LanguageModelCompletionEvent::StartMessage { ref message_id } if message_id == "msg_123"
1483 ));
1484 assert!(matches!(
1485 mapped[1],
1486 LanguageModelCompletionEvent::Text(ref text) if text == "Hello"
1487 ));
1488 assert!(matches!(
1489 mapped[2],
1490 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1491 input_tokens: 5,
1492 output_tokens: 3,
1493 ..
1494 })
1495 ));
1496 assert!(matches!(
1497 mapped[3],
1498 LanguageModelCompletionEvent::Stop(StopReason::EndTurn)
1499 ));
1500 }
1501
1502 #[test]
1503 fn into_open_ai_response_builds_complete_payload() {
1504 let tool_call_id = LanguageModelToolUseId::from("call-42");
1505 let tool_input = json!({ "city": "Boston" });
1506 let tool_arguments = serde_json::to_string(&tool_input).unwrap();
1507 let tool_use = LanguageModelToolUse {
1508 id: tool_call_id.clone(),
1509 name: Arc::from("get_weather"),
1510 raw_input: tool_arguments.clone(),
1511 input: tool_input,
1512 is_input_complete: true,
1513 thought_signature: None,
1514 };
1515 let tool_result = LanguageModelToolResult {
1516 tool_use_id: tool_call_id,
1517 tool_name: Arc::from("get_weather"),
1518 is_error: false,
1519 content: LanguageModelToolResultContent::Text(Arc::from("Sunny")),
1520 output: Some(json!({ "forecast": "Sunny" })),
1521 };
1522 let user_image = LanguageModelImage {
1523 source: SharedString::from("aGVsbG8="),
1524 size: None,
1525 };
1526 let expected_image_url = user_image.to_base64_url();
1527
1528 let request = LanguageModelRequest {
1529 thread_id: Some("thread-123".into()),
1530 prompt_id: None,
1531 intent: None,
1532 messages: vec![
1533 LanguageModelRequestMessage {
1534 role: Role::System,
1535 content: vec![MessageContent::Text("System context".into())],
1536 cache: false,
1537 reasoning_details: None,
1538 },
1539 LanguageModelRequestMessage {
1540 role: Role::User,
1541 content: vec![
1542 MessageContent::Text("Please check the weather.".into()),
1543 MessageContent::Image(user_image),
1544 ],
1545 cache: false,
1546 reasoning_details: None,
1547 },
1548 LanguageModelRequestMessage {
1549 role: Role::Assistant,
1550 content: vec![
1551 MessageContent::Text("Looking that up.".into()),
1552 MessageContent::ToolUse(tool_use),
1553 ],
1554 cache: false,
1555 reasoning_details: None,
1556 },
1557 LanguageModelRequestMessage {
1558 role: Role::Assistant,
1559 content: vec![MessageContent::ToolResult(tool_result)],
1560 cache: false,
1561 reasoning_details: None,
1562 },
1563 ],
1564 tools: vec![LanguageModelRequestTool {
1565 name: "get_weather".into(),
1566 description: "Fetches the weather".into(),
1567 input_schema: json!({ "type": "object" }),
1568 }],
1569 tool_choice: Some(LanguageModelToolChoice::Any),
1570 stop: vec!["<STOP>".into()],
1571 temperature: None,
1572 thinking_allowed: false,
1573 };
1574
1575 let response = into_open_ai_response(
1576 request,
1577 "custom-model",
1578 true,
1579 true,
1580 Some(2048),
1581 Some(ReasoningEffort::Low),
1582 );
1583
1584 let serialized = serde_json::to_value(&response).unwrap();
1585 let expected = json!({
1586 "model": "custom-model",
1587 "input": [
1588 {
1589 "type": "message",
1590 "role": "system",
1591 "content": [
1592 { "type": "input_text", "text": "System context" }
1593 ]
1594 },
1595 {
1596 "type": "message",
1597 "role": "user",
1598 "content": [
1599 { "type": "input_text", "text": "Please check the weather." },
1600 { "type": "input_image", "image_url": expected_image_url }
1601 ]
1602 },
1603 {
1604 "type": "message",
1605 "role": "assistant",
1606 "content": [
1607 { "type": "output_text", "text": "Looking that up.", "annotations": [] }
1608 ]
1609 },
1610 {
1611 "type": "function_call",
1612 "call_id": "call-42",
1613 "name": "get_weather",
1614 "arguments": tool_arguments
1615 },
1616 {
1617 "type": "function_call_output",
1618 "call_id": "call-42",
1619 "output": "{\"forecast\":\"Sunny\"}"
1620 }
1621 ],
1622 "stream": true,
1623 "max_output_tokens": 2048,
1624 "parallel_tool_calls": true,
1625 "tool_choice": "required",
1626 "tools": [
1627 {
1628 "type": "function",
1629 "name": "get_weather",
1630 "description": "Fetches the weather",
1631 "parameters": { "type": "object" }
1632 }
1633 ],
1634 "prompt_cache_key": "thread-123",
1635 "reasoning": { "effort": "low" }
1636 });
1637
1638 assert_eq!(serialized, expected);
1639 }
1640
1641 #[test]
1642 fn responses_stream_maps_tool_calls() {
1643 let events = vec![
1644 ResponsesStreamEvent::OutputItemAdded {
1645 output_index: 0,
1646 sequence_number: None,
1647 item: response_item_function_call("item_fn", Some("{\"city\":\"Bos")),
1648 },
1649 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1650 item_id: "item_fn".into(),
1651 output_index: 0,
1652 delta: "ton\"}".into(),
1653 sequence_number: None,
1654 },
1655 ResponsesStreamEvent::FunctionCallArgumentsDone {
1656 item_id: "item_fn".into(),
1657 output_index: 0,
1658 arguments: "{\"city\":\"Boston\"}".into(),
1659 sequence_number: None,
1660 },
1661 ResponsesStreamEvent::Completed {
1662 response: ResponseSummary::default(),
1663 },
1664 ];
1665
1666 let mapped = map_response_events(events);
1667 assert!(matches!(
1668 mapped[0],
1669 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1670 ref id,
1671 ref name,
1672 ref raw_input,
1673 ..
1674 }) if id.to_string() == "call_123"
1675 && name.as_ref() == "get_weather"
1676 && raw_input == "{\"city\":\"Boston\"}"
1677 ));
1678 assert!(matches!(
1679 mapped[1],
1680 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1681 ));
1682 }
1683
1684 #[test]
1685 fn responses_stream_uses_max_tokens_stop_reason() {
1686 let events = vec![ResponsesStreamEvent::Incomplete {
1687 response: ResponseSummary {
1688 status_details: Some(ResponseStatusDetails {
1689 reason: Some("max_output_tokens".into()),
1690 r#type: Some("incomplete".into()),
1691 error: None,
1692 }),
1693 usage: Some(ResponseUsage {
1694 input_tokens: Some(10),
1695 output_tokens: Some(20),
1696 total_tokens: Some(30),
1697 }),
1698 ..Default::default()
1699 },
1700 }];
1701
1702 let mapped = map_response_events(events);
1703 assert!(matches!(
1704 mapped[0],
1705 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1706 input_tokens: 10,
1707 output_tokens: 20,
1708 ..
1709 })
1710 ));
1711 assert!(matches!(
1712 mapped[1],
1713 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1714 ));
1715 }
1716
1717 #[test]
1718 fn responses_stream_handles_multiple_tool_calls() {
1719 let events = vec![
1720 ResponsesStreamEvent::OutputItemAdded {
1721 output_index: 0,
1722 sequence_number: None,
1723 item: response_item_function_call("item_fn1", Some("{\"city\":\"NYC\"}")),
1724 },
1725 ResponsesStreamEvent::FunctionCallArgumentsDone {
1726 item_id: "item_fn1".into(),
1727 output_index: 0,
1728 arguments: "{\"city\":\"NYC\"}".into(),
1729 sequence_number: None,
1730 },
1731 ResponsesStreamEvent::OutputItemAdded {
1732 output_index: 1,
1733 sequence_number: None,
1734 item: response_item_function_call("item_fn2", Some("{\"city\":\"LA\"}")),
1735 },
1736 ResponsesStreamEvent::FunctionCallArgumentsDone {
1737 item_id: "item_fn2".into(),
1738 output_index: 1,
1739 arguments: "{\"city\":\"LA\"}".into(),
1740 sequence_number: None,
1741 },
1742 ResponsesStreamEvent::Completed {
1743 response: ResponseSummary::default(),
1744 },
1745 ];
1746
1747 let mapped = map_response_events(events);
1748 assert_eq!(mapped.len(), 3);
1749 assert!(matches!(
1750 mapped[0],
1751 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1752 if raw_input == "{\"city\":\"NYC\"}"
1753 ));
1754 assert!(matches!(
1755 mapped[1],
1756 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1757 if raw_input == "{\"city\":\"LA\"}"
1758 ));
1759 assert!(matches!(
1760 mapped[2],
1761 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1762 ));
1763 }
1764
1765 #[test]
1766 fn responses_stream_handles_mixed_text_and_tool_calls() {
1767 let events = vec![
1768 ResponsesStreamEvent::OutputItemAdded {
1769 output_index: 0,
1770 sequence_number: None,
1771 item: response_item_message("msg_123"),
1772 },
1773 ResponsesStreamEvent::OutputTextDelta {
1774 item_id: "msg_123".into(),
1775 output_index: 0,
1776 content_index: Some(0),
1777 delta: "Let me check that".into(),
1778 },
1779 ResponsesStreamEvent::OutputItemAdded {
1780 output_index: 1,
1781 sequence_number: None,
1782 item: response_item_function_call("item_fn", Some("{\"query\":\"test\"}")),
1783 },
1784 ResponsesStreamEvent::FunctionCallArgumentsDone {
1785 item_id: "item_fn".into(),
1786 output_index: 1,
1787 arguments: "{\"query\":\"test\"}".into(),
1788 sequence_number: None,
1789 },
1790 ResponsesStreamEvent::Completed {
1791 response: ResponseSummary::default(),
1792 },
1793 ];
1794
1795 let mapped = map_response_events(events);
1796 assert!(matches!(
1797 mapped[0],
1798 LanguageModelCompletionEvent::StartMessage { .. }
1799 ));
1800 assert!(matches!(
1801 mapped[1],
1802 LanguageModelCompletionEvent::Text(ref text) if text == "Let me check that"
1803 ));
1804 assert!(matches!(
1805 mapped[2],
1806 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1807 if raw_input == "{\"query\":\"test\"}"
1808 ));
1809 assert!(matches!(
1810 mapped[3],
1811 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1812 ));
1813 }
1814
1815 #[test]
1816 fn responses_stream_handles_json_parse_error() {
1817 let events = vec![
1818 ResponsesStreamEvent::OutputItemAdded {
1819 output_index: 0,
1820 sequence_number: None,
1821 item: response_item_function_call("item_fn", Some("{invalid json")),
1822 },
1823 ResponsesStreamEvent::FunctionCallArgumentsDone {
1824 item_id: "item_fn".into(),
1825 output_index: 0,
1826 arguments: "{invalid json".into(),
1827 sequence_number: None,
1828 },
1829 ResponsesStreamEvent::Completed {
1830 response: ResponseSummary::default(),
1831 },
1832 ];
1833
1834 let mapped = map_response_events(events);
1835 assert!(matches!(
1836 mapped[0],
1837 LanguageModelCompletionEvent::ToolUseJsonParseError {
1838 ref raw_input,
1839 ..
1840 } if raw_input.as_ref() == "{invalid json"
1841 ));
1842 }
1843
1844 #[test]
1845 fn responses_stream_handles_incomplete_function_call() {
1846 let events = vec![
1847 ResponsesStreamEvent::OutputItemAdded {
1848 output_index: 0,
1849 sequence_number: None,
1850 item: response_item_function_call("item_fn", Some("{\"city\":")),
1851 },
1852 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1853 item_id: "item_fn".into(),
1854 output_index: 0,
1855 delta: "\"Boston\"".into(),
1856 sequence_number: None,
1857 },
1858 ResponsesStreamEvent::Incomplete {
1859 response: ResponseSummary {
1860 status_details: Some(ResponseStatusDetails {
1861 reason: Some("max_output_tokens".into()),
1862 r#type: Some("incomplete".into()),
1863 error: None,
1864 }),
1865 output: vec![response_item_function_call(
1866 "item_fn",
1867 Some("{\"city\":\"Boston\"}"),
1868 )],
1869 ..Default::default()
1870 },
1871 },
1872 ];
1873
1874 let mapped = map_response_events(events);
1875 assert!(matches!(
1876 mapped[0],
1877 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1878 if raw_input == "{\"city\":\"Boston\"}"
1879 ));
1880 assert!(matches!(
1881 mapped[1],
1882 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1883 ));
1884 }
1885
1886 #[test]
1887 fn responses_stream_incomplete_does_not_duplicate_tool_calls() {
1888 let events = vec![
1889 ResponsesStreamEvent::OutputItemAdded {
1890 output_index: 0,
1891 sequence_number: None,
1892 item: response_item_function_call("item_fn", Some("{\"city\":\"Boston\"}")),
1893 },
1894 ResponsesStreamEvent::FunctionCallArgumentsDone {
1895 item_id: "item_fn".into(),
1896 output_index: 0,
1897 arguments: "{\"city\":\"Boston\"}".into(),
1898 sequence_number: None,
1899 },
1900 ResponsesStreamEvent::Incomplete {
1901 response: ResponseSummary {
1902 status_details: Some(ResponseStatusDetails {
1903 reason: Some("max_output_tokens".into()),
1904 r#type: Some("incomplete".into()),
1905 error: None,
1906 }),
1907 output: vec![response_item_function_call(
1908 "item_fn",
1909 Some("{\"city\":\"Boston\"}"),
1910 )],
1911 ..Default::default()
1912 },
1913 },
1914 ];
1915
1916 let mapped = map_response_events(events);
1917 assert_eq!(mapped.len(), 2);
1918 assert!(matches!(
1919 mapped[0],
1920 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1921 if raw_input == "{\"city\":\"Boston\"}"
1922 ));
1923 assert!(matches!(
1924 mapped[1],
1925 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1926 ));
1927 }
1928}