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::FivePointFour
314 | Model::FivePointFourPro
315 | Model::O1
316 | Model::O3 => 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_thinking(&self) -> bool {
334 self.model.reasoning_effort().is_some()
335 }
336
337 fn supports_split_token_display(&self) -> bool {
338 true
339 }
340
341 fn telemetry_id(&self) -> String {
342 format!("openai/{}", self.model.id())
343 }
344
345 fn max_token_count(&self) -> u64 {
346 self.model.max_token_count()
347 }
348
349 fn max_output_tokens(&self) -> Option<u64> {
350 self.model.max_output_tokens()
351 }
352
353 fn count_tokens(
354 &self,
355 request: LanguageModelRequest,
356 cx: &App,
357 ) -> BoxFuture<'static, Result<u64>> {
358 count_open_ai_tokens(request, self.model.clone(), cx)
359 }
360
361 fn stream_completion(
362 &self,
363 request: LanguageModelRequest,
364 cx: &AsyncApp,
365 ) -> BoxFuture<
366 'static,
367 Result<
368 futures::stream::BoxStream<
369 'static,
370 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
371 >,
372 LanguageModelCompletionError,
373 >,
374 > {
375 if self.model.supports_chat_completions() {
376 let request = into_open_ai(
377 request,
378 self.model.id(),
379 self.model.supports_parallel_tool_calls(),
380 self.model.supports_prompt_cache_key(),
381 self.max_output_tokens(),
382 self.model.reasoning_effort(),
383 );
384 let completions = self.stream_completion(request, cx);
385 async move {
386 let mapper = OpenAiEventMapper::new();
387 Ok(mapper.map_stream(completions.await?).boxed())
388 }
389 .boxed()
390 } else {
391 let request = into_open_ai_response(
392 request,
393 self.model.id(),
394 self.model.supports_parallel_tool_calls(),
395 self.model.supports_prompt_cache_key(),
396 self.max_output_tokens(),
397 self.model.reasoning_effort(),
398 );
399 let completions = self.stream_response(request, cx);
400 async move {
401 let mapper = OpenAiResponseEventMapper::new();
402 Ok(mapper.map_stream(completions.await?).boxed())
403 }
404 .boxed()
405 }
406 }
407}
408
409pub fn into_open_ai(
410 request: LanguageModelRequest,
411 model_id: &str,
412 supports_parallel_tool_calls: bool,
413 supports_prompt_cache_key: bool,
414 max_output_tokens: Option<u64>,
415 reasoning_effort: Option<ReasoningEffort>,
416) -> open_ai::Request {
417 let stream = !model_id.starts_with("o1-");
418
419 let mut messages = Vec::new();
420 for message in request.messages {
421 for content in message.content {
422 match content {
423 MessageContent::Text(text) | MessageContent::Thinking { text, .. } => {
424 let should_add = if message.role == Role::User {
425 // Including whitespace-only user messages can cause error with OpenAI compatible APIs
426 // See https://github.com/zed-industries/zed/issues/40097
427 !text.trim().is_empty()
428 } else {
429 !text.is_empty()
430 };
431 if should_add {
432 add_message_content_part(
433 open_ai::MessagePart::Text { text },
434 message.role,
435 &mut messages,
436 );
437 }
438 }
439 MessageContent::RedactedThinking(_) => {}
440 MessageContent::Image(image) => {
441 add_message_content_part(
442 open_ai::MessagePart::Image {
443 image_url: ImageUrl {
444 url: image.to_base64_url(),
445 detail: None,
446 },
447 },
448 message.role,
449 &mut messages,
450 );
451 }
452 MessageContent::ToolUse(tool_use) => {
453 let tool_call = open_ai::ToolCall {
454 id: tool_use.id.to_string(),
455 content: open_ai::ToolCallContent::Function {
456 function: open_ai::FunctionContent {
457 name: tool_use.name.to_string(),
458 arguments: serde_json::to_string(&tool_use.input)
459 .unwrap_or_default(),
460 },
461 },
462 };
463
464 if let Some(open_ai::RequestMessage::Assistant { tool_calls, .. }) =
465 messages.last_mut()
466 {
467 tool_calls.push(tool_call);
468 } else {
469 messages.push(open_ai::RequestMessage::Assistant {
470 content: None,
471 tool_calls: vec![tool_call],
472 });
473 }
474 }
475 MessageContent::ToolResult(tool_result) => {
476 let content = match &tool_result.content {
477 LanguageModelToolResultContent::Text(text) => {
478 vec![open_ai::MessagePart::Text {
479 text: text.to_string(),
480 }]
481 }
482 LanguageModelToolResultContent::Image(image) => {
483 vec![open_ai::MessagePart::Image {
484 image_url: ImageUrl {
485 url: image.to_base64_url(),
486 detail: None,
487 },
488 }]
489 }
490 };
491
492 messages.push(open_ai::RequestMessage::Tool {
493 content: content.into(),
494 tool_call_id: tool_result.tool_use_id.to_string(),
495 });
496 }
497 }
498 }
499 }
500
501 open_ai::Request {
502 model: model_id.into(),
503 messages,
504 stream,
505 stop: request.stop,
506 temperature: request.temperature.or(Some(1.0)),
507 max_completion_tokens: max_output_tokens,
508 parallel_tool_calls: if supports_parallel_tool_calls && !request.tools.is_empty() {
509 // Disable parallel tool calls, as the Agent currently expects a maximum of one per turn.
510 Some(false)
511 } else {
512 None
513 },
514 prompt_cache_key: if supports_prompt_cache_key {
515 request.thread_id
516 } else {
517 None
518 },
519 tools: request
520 .tools
521 .into_iter()
522 .map(|tool| open_ai::ToolDefinition::Function {
523 function: open_ai::FunctionDefinition {
524 name: tool.name,
525 description: Some(tool.description),
526 parameters: Some(tool.input_schema),
527 },
528 })
529 .collect(),
530 tool_choice: request.tool_choice.map(|choice| match choice {
531 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
532 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
533 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
534 }),
535 reasoning_effort,
536 }
537}
538
539pub fn into_open_ai_response(
540 request: LanguageModelRequest,
541 model_id: &str,
542 supports_parallel_tool_calls: bool,
543 supports_prompt_cache_key: bool,
544 max_output_tokens: Option<u64>,
545 reasoning_effort: Option<ReasoningEffort>,
546) -> ResponseRequest {
547 let stream = !model_id.starts_with("o1-");
548
549 let LanguageModelRequest {
550 thread_id,
551 prompt_id: _,
552 intent: _,
553 messages,
554 tools,
555 tool_choice,
556 stop: _,
557 temperature,
558 thinking_allowed: _,
559 thinking_effort: _,
560 } = request;
561
562 let mut input_items = Vec::new();
563 for (index, message) in messages.into_iter().enumerate() {
564 append_message_to_response_items(message, index, &mut input_items);
565 }
566
567 let tools: Vec<_> = tools
568 .into_iter()
569 .map(|tool| open_ai::responses::ToolDefinition::Function {
570 name: tool.name,
571 description: Some(tool.description),
572 parameters: Some(tool.input_schema),
573 strict: None,
574 })
575 .collect();
576
577 ResponseRequest {
578 model: model_id.into(),
579 input: input_items,
580 stream,
581 temperature,
582 top_p: None,
583 max_output_tokens,
584 parallel_tool_calls: if tools.is_empty() {
585 None
586 } else {
587 Some(supports_parallel_tool_calls)
588 },
589 tool_choice: tool_choice.map(|choice| match choice {
590 LanguageModelToolChoice::Auto => open_ai::ToolChoice::Auto,
591 LanguageModelToolChoice::Any => open_ai::ToolChoice::Required,
592 LanguageModelToolChoice::None => open_ai::ToolChoice::None,
593 }),
594 tools,
595 prompt_cache_key: if supports_prompt_cache_key {
596 thread_id
597 } else {
598 None
599 },
600 reasoning: reasoning_effort.map(|effort| open_ai::responses::ReasoningConfig { effort }),
601 }
602}
603
604fn append_message_to_response_items(
605 message: LanguageModelRequestMessage,
606 index: usize,
607 input_items: &mut Vec<ResponseInputItem>,
608) {
609 let mut content_parts: Vec<ResponseInputContent> = Vec::new();
610
611 for content in message.content {
612 match content {
613 MessageContent::Text(text) => {
614 push_response_text_part(&message.role, text, &mut content_parts);
615 }
616 MessageContent::Thinking { text, .. } => {
617 push_response_text_part(&message.role, text, &mut content_parts);
618 }
619 MessageContent::RedactedThinking(_) => {}
620 MessageContent::Image(image) => {
621 push_response_image_part(&message.role, image, &mut content_parts);
622 }
623 MessageContent::ToolUse(tool_use) => {
624 flush_response_parts(&message.role, index, &mut content_parts, input_items);
625 let call_id = tool_use.id.to_string();
626 input_items.push(ResponseInputItem::FunctionCall(ResponseFunctionCallItem {
627 call_id,
628 name: tool_use.name.to_string(),
629 arguments: tool_use.raw_input,
630 }));
631 }
632 MessageContent::ToolResult(tool_result) => {
633 flush_response_parts(&message.role, index, &mut content_parts, input_items);
634 input_items.push(ResponseInputItem::FunctionCallOutput(
635 ResponseFunctionCallOutputItem {
636 call_id: tool_result.tool_use_id.to_string(),
637 output: tool_result_output(&tool_result),
638 },
639 ));
640 }
641 }
642 }
643
644 flush_response_parts(&message.role, index, &mut content_parts, input_items);
645}
646
647fn push_response_text_part(
648 role: &Role,
649 text: impl Into<String>,
650 parts: &mut Vec<ResponseInputContent>,
651) {
652 let text = text.into();
653 if text.trim().is_empty() {
654 return;
655 }
656
657 match role {
658 Role::Assistant => parts.push(ResponseInputContent::OutputText {
659 text,
660 annotations: Vec::new(),
661 }),
662 _ => parts.push(ResponseInputContent::Text { text }),
663 }
664}
665
666fn push_response_image_part(
667 role: &Role,
668 image: LanguageModelImage,
669 parts: &mut Vec<ResponseInputContent>,
670) {
671 match role {
672 Role::Assistant => parts.push(ResponseInputContent::OutputText {
673 text: "[image omitted]".to_string(),
674 annotations: Vec::new(),
675 }),
676 _ => parts.push(ResponseInputContent::Image {
677 image_url: image.to_base64_url(),
678 }),
679 }
680}
681
682fn flush_response_parts(
683 role: &Role,
684 _index: usize,
685 parts: &mut Vec<ResponseInputContent>,
686 input_items: &mut Vec<ResponseInputItem>,
687) {
688 if parts.is_empty() {
689 return;
690 }
691
692 let item = ResponseInputItem::Message(ResponseMessageItem {
693 role: match role {
694 Role::User => open_ai::Role::User,
695 Role::Assistant => open_ai::Role::Assistant,
696 Role::System => open_ai::Role::System,
697 },
698 content: parts.clone(),
699 });
700
701 input_items.push(item);
702 parts.clear();
703}
704
705fn tool_result_output(result: &LanguageModelToolResult) -> String {
706 if let Some(output) = &result.output {
707 match output {
708 serde_json::Value::String(text) => text.clone(),
709 serde_json::Value::Null => String::new(),
710 _ => output.to_string(),
711 }
712 } else {
713 match &result.content {
714 LanguageModelToolResultContent::Text(text) => text.to_string(),
715 LanguageModelToolResultContent::Image(image) => image.to_base64_url(),
716 }
717 }
718}
719
720fn add_message_content_part(
721 new_part: open_ai::MessagePart,
722 role: Role,
723 messages: &mut Vec<open_ai::RequestMessage>,
724) {
725 match (role, messages.last_mut()) {
726 (Role::User, Some(open_ai::RequestMessage::User { content }))
727 | (
728 Role::Assistant,
729 Some(open_ai::RequestMessage::Assistant {
730 content: Some(content),
731 ..
732 }),
733 )
734 | (Role::System, Some(open_ai::RequestMessage::System { content, .. })) => {
735 content.push_part(new_part);
736 }
737 _ => {
738 messages.push(match role {
739 Role::User => open_ai::RequestMessage::User {
740 content: open_ai::MessageContent::from(vec![new_part]),
741 },
742 Role::Assistant => open_ai::RequestMessage::Assistant {
743 content: Some(open_ai::MessageContent::from(vec![new_part])),
744 tool_calls: Vec::new(),
745 },
746 Role::System => open_ai::RequestMessage::System {
747 content: open_ai::MessageContent::from(vec![new_part]),
748 },
749 });
750 }
751 }
752}
753
754pub struct OpenAiEventMapper {
755 tool_calls_by_index: HashMap<usize, RawToolCall>,
756}
757
758impl OpenAiEventMapper {
759 pub fn new() -> Self {
760 Self {
761 tool_calls_by_index: HashMap::default(),
762 }
763 }
764
765 pub fn map_stream(
766 mut self,
767 events: Pin<Box<dyn Send + Stream<Item = Result<ResponseStreamEvent>>>>,
768 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
769 {
770 events.flat_map(move |event| {
771 futures::stream::iter(match event {
772 Ok(event) => self.map_event(event),
773 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
774 })
775 })
776 }
777
778 pub fn map_event(
779 &mut self,
780 event: ResponseStreamEvent,
781 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
782 let mut events = Vec::new();
783 if let Some(usage) = event.usage {
784 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
785 input_tokens: usage.prompt_tokens,
786 output_tokens: usage.completion_tokens,
787 cache_creation_input_tokens: 0,
788 cache_read_input_tokens: 0,
789 })));
790 }
791
792 let Some(choice) = event.choices.first() else {
793 return events;
794 };
795
796 if let Some(delta) = choice.delta.as_ref() {
797 if let Some(reasoning_content) = delta.reasoning_content.clone() {
798 if !reasoning_content.is_empty() {
799 events.push(Ok(LanguageModelCompletionEvent::Thinking {
800 text: reasoning_content,
801 signature: None,
802 }));
803 }
804 }
805 if let Some(content) = delta.content.clone() {
806 if !content.is_empty() {
807 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
808 }
809 }
810
811 if let Some(tool_calls) = delta.tool_calls.as_ref() {
812 for tool_call in tool_calls {
813 let entry = self.tool_calls_by_index.entry(tool_call.index).or_default();
814
815 if let Some(tool_id) = tool_call.id.clone() {
816 entry.id = tool_id;
817 }
818
819 if let Some(function) = tool_call.function.as_ref() {
820 if let Some(name) = function.name.clone() {
821 entry.name = name;
822 }
823
824 if let Some(arguments) = function.arguments.clone() {
825 entry.arguments.push_str(&arguments);
826 }
827 }
828 }
829 }
830 }
831
832 match choice.finish_reason.as_deref() {
833 Some("stop") => {
834 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
835 }
836 Some("tool_calls") => {
837 events.extend(self.tool_calls_by_index.drain().map(|(_, tool_call)| {
838 match parse_tool_arguments(&tool_call.arguments) {
839 Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
840 LanguageModelToolUse {
841 id: tool_call.id.clone().into(),
842 name: tool_call.name.as_str().into(),
843 is_input_complete: true,
844 input,
845 raw_input: tool_call.arguments.clone(),
846 thought_signature: None,
847 },
848 )),
849 Err(error) => Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
850 id: tool_call.id.into(),
851 tool_name: tool_call.name.into(),
852 raw_input: tool_call.arguments.clone().into(),
853 json_parse_error: error.to_string(),
854 }),
855 }
856 }));
857
858 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
859 }
860 Some(stop_reason) => {
861 log::error!("Unexpected OpenAI stop_reason: {stop_reason:?}",);
862 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
863 }
864 None => {}
865 }
866
867 events
868 }
869}
870
871#[derive(Default)]
872struct RawToolCall {
873 id: String,
874 name: String,
875 arguments: String,
876}
877
878pub struct OpenAiResponseEventMapper {
879 function_calls_by_item: HashMap<String, PendingResponseFunctionCall>,
880 pending_stop_reason: Option<StopReason>,
881}
882
883#[derive(Default)]
884struct PendingResponseFunctionCall {
885 call_id: String,
886 name: Arc<str>,
887 arguments: String,
888}
889
890impl OpenAiResponseEventMapper {
891 pub fn new() -> Self {
892 Self {
893 function_calls_by_item: HashMap::default(),
894 pending_stop_reason: None,
895 }
896 }
897
898 pub fn map_stream(
899 mut self,
900 events: Pin<Box<dyn Send + Stream<Item = Result<ResponsesStreamEvent>>>>,
901 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
902 {
903 events.flat_map(move |event| {
904 futures::stream::iter(match event {
905 Ok(event) => self.map_event(event),
906 Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
907 })
908 })
909 }
910
911 pub fn map_event(
912 &mut self,
913 event: ResponsesStreamEvent,
914 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
915 match event {
916 ResponsesStreamEvent::OutputItemAdded { item, .. } => {
917 let mut events = Vec::new();
918
919 match &item {
920 ResponseOutputItem::Message(message) => {
921 if let Some(id) = &message.id {
922 events.push(Ok(LanguageModelCompletionEvent::StartMessage {
923 message_id: id.clone(),
924 }));
925 }
926 }
927 ResponseOutputItem::FunctionCall(function_call) => {
928 if let Some(item_id) = function_call.id.clone() {
929 let call_id = function_call
930 .call_id
931 .clone()
932 .or_else(|| function_call.id.clone())
933 .unwrap_or_else(|| item_id.clone());
934 let entry = PendingResponseFunctionCall {
935 call_id,
936 name: Arc::<str>::from(
937 function_call.name.clone().unwrap_or_default(),
938 ),
939 arguments: function_call.arguments.clone(),
940 };
941 self.function_calls_by_item.insert(item_id, entry);
942 }
943 }
944 ResponseOutputItem::Unknown => {}
945 }
946 events
947 }
948 ResponsesStreamEvent::OutputTextDelta { delta, .. } => {
949 if delta.is_empty() {
950 Vec::new()
951 } else {
952 vec![Ok(LanguageModelCompletionEvent::Text(delta))]
953 }
954 }
955 ResponsesStreamEvent::FunctionCallArgumentsDelta { item_id, delta, .. } => {
956 if let Some(entry) = self.function_calls_by_item.get_mut(&item_id) {
957 entry.arguments.push_str(&delta);
958 }
959 Vec::new()
960 }
961 ResponsesStreamEvent::FunctionCallArgumentsDone {
962 item_id, arguments, ..
963 } => {
964 if let Some(mut entry) = self.function_calls_by_item.remove(&item_id) {
965 if !arguments.is_empty() {
966 entry.arguments = arguments;
967 }
968 let raw_input = entry.arguments.clone();
969 self.pending_stop_reason = Some(StopReason::ToolUse);
970 match parse_tool_arguments(&entry.arguments) {
971 Ok(input) => {
972 vec![Ok(LanguageModelCompletionEvent::ToolUse(
973 LanguageModelToolUse {
974 id: LanguageModelToolUseId::from(entry.call_id.clone()),
975 name: entry.name.clone(),
976 is_input_complete: true,
977 input,
978 raw_input,
979 thought_signature: None,
980 },
981 ))]
982 }
983 Err(error) => {
984 vec![Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
985 id: LanguageModelToolUseId::from(entry.call_id.clone()),
986 tool_name: entry.name.clone(),
987 raw_input: Arc::<str>::from(raw_input),
988 json_parse_error: error.to_string(),
989 })]
990 }
991 }
992 } else {
993 Vec::new()
994 }
995 }
996 ResponsesStreamEvent::Completed { response } => {
997 self.handle_completion(response, StopReason::EndTurn)
998 }
999 ResponsesStreamEvent::Incomplete { response } => {
1000 let reason = response
1001 .status_details
1002 .as_ref()
1003 .and_then(|details| details.reason.as_deref());
1004 let stop_reason = match reason {
1005 Some("max_output_tokens") => StopReason::MaxTokens,
1006 Some("content_filter") => {
1007 self.pending_stop_reason = Some(StopReason::Refusal);
1008 StopReason::Refusal
1009 }
1010 _ => self
1011 .pending_stop_reason
1012 .take()
1013 .unwrap_or(StopReason::EndTurn),
1014 };
1015
1016 let mut events = Vec::new();
1017 if self.pending_stop_reason.is_none() {
1018 events.extend(self.emit_tool_calls_from_output(&response.output));
1019 }
1020 if let Some(usage) = response.usage.as_ref() {
1021 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1022 token_usage_from_response_usage(usage),
1023 )));
1024 }
1025 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1026 events
1027 }
1028 ResponsesStreamEvent::Failed { response } => {
1029 let message = response
1030 .status_details
1031 .and_then(|details| details.error)
1032 .map(|error| error.to_string())
1033 .unwrap_or_else(|| "response failed".to_string());
1034 vec![Err(LanguageModelCompletionError::Other(anyhow!(message)))]
1035 }
1036 ResponsesStreamEvent::Error { error }
1037 | ResponsesStreamEvent::GenericError { error } => {
1038 vec![Err(LanguageModelCompletionError::Other(anyhow!(format!(
1039 "{error:?}"
1040 ))))]
1041 }
1042 ResponsesStreamEvent::OutputTextDone { .. } => Vec::new(),
1043 ResponsesStreamEvent::OutputItemDone { .. }
1044 | ResponsesStreamEvent::ContentPartAdded { .. }
1045 | ResponsesStreamEvent::ContentPartDone { .. }
1046 | ResponsesStreamEvent::Created { .. }
1047 | ResponsesStreamEvent::InProgress { .. }
1048 | ResponsesStreamEvent::Unknown => Vec::new(),
1049 }
1050 }
1051
1052 fn handle_completion(
1053 &mut self,
1054 response: ResponsesSummary,
1055 default_reason: StopReason,
1056 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1057 let mut events = Vec::new();
1058
1059 if self.pending_stop_reason.is_none() {
1060 events.extend(self.emit_tool_calls_from_output(&response.output));
1061 }
1062
1063 if let Some(usage) = response.usage.as_ref() {
1064 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
1065 token_usage_from_response_usage(usage),
1066 )));
1067 }
1068
1069 let stop_reason = self.pending_stop_reason.take().unwrap_or(default_reason);
1070 events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
1071 events
1072 }
1073
1074 fn emit_tool_calls_from_output(
1075 &mut self,
1076 output: &[ResponseOutputItem],
1077 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
1078 let mut events = Vec::new();
1079 for item in output {
1080 if let ResponseOutputItem::FunctionCall(function_call) = item {
1081 let Some(call_id) = function_call
1082 .call_id
1083 .clone()
1084 .or_else(|| function_call.id.clone())
1085 else {
1086 log::error!(
1087 "Function call item missing both call_id and id: {:?}",
1088 function_call
1089 );
1090 continue;
1091 };
1092 let name: Arc<str> = Arc::from(function_call.name.clone().unwrap_or_default());
1093 let arguments = &function_call.arguments;
1094 self.pending_stop_reason = Some(StopReason::ToolUse);
1095 match parse_tool_arguments(arguments) {
1096 Ok(input) => {
1097 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
1098 LanguageModelToolUse {
1099 id: LanguageModelToolUseId::from(call_id.clone()),
1100 name: name.clone(),
1101 is_input_complete: true,
1102 input,
1103 raw_input: arguments.clone(),
1104 thought_signature: None,
1105 },
1106 )));
1107 }
1108 Err(error) => {
1109 events.push(Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
1110 id: LanguageModelToolUseId::from(call_id.clone()),
1111 tool_name: name.clone(),
1112 raw_input: Arc::<str>::from(arguments.clone()),
1113 json_parse_error: error.to_string(),
1114 }));
1115 }
1116 }
1117 }
1118 }
1119 events
1120 }
1121}
1122
1123fn token_usage_from_response_usage(usage: &ResponsesUsage) -> TokenUsage {
1124 TokenUsage {
1125 input_tokens: usage.input_tokens.unwrap_or_default(),
1126 output_tokens: usage.output_tokens.unwrap_or_default(),
1127 cache_creation_input_tokens: 0,
1128 cache_read_input_tokens: 0,
1129 }
1130}
1131
1132pub(crate) fn collect_tiktoken_messages(
1133 request: LanguageModelRequest,
1134) -> Vec<tiktoken_rs::ChatCompletionRequestMessage> {
1135 request
1136 .messages
1137 .into_iter()
1138 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
1139 role: match message.role {
1140 Role::User => "user".into(),
1141 Role::Assistant => "assistant".into(),
1142 Role::System => "system".into(),
1143 },
1144 content: Some(message.string_contents()),
1145 name: None,
1146 function_call: None,
1147 })
1148 .collect::<Vec<_>>()
1149}
1150
1151pub fn count_open_ai_tokens(
1152 request: LanguageModelRequest,
1153 model: Model,
1154 cx: &App,
1155) -> BoxFuture<'static, Result<u64>> {
1156 cx.background_spawn(async move {
1157 let messages = collect_tiktoken_messages(request);
1158 match model {
1159 Model::Custom { max_tokens, .. } => {
1160 let model = if max_tokens >= 100_000 {
1161 // If the max tokens is 100k or more, it likely uses the o200k_base tokenizer
1162 "gpt-4o"
1163 } else {
1164 // Otherwise fallback to gpt-4, since only cl100k_base and o200k_base are
1165 // supported with this tiktoken method
1166 "gpt-4"
1167 };
1168 tiktoken_rs::num_tokens_from_messages(model, &messages)
1169 }
1170 // Currently supported by tiktoken_rs
1171 // Sometimes tiktoken-rs is behind on model support. If that is the case, make a new branch
1172 // arm with an override. We enumerate all supported models here so that we can check if new
1173 // models are supported yet or not.
1174 Model::ThreePointFiveTurbo
1175 | Model::Four
1176 | Model::FourTurbo
1177 | Model::FourOmniMini
1178 | Model::FourPointOneNano
1179 | Model::O1
1180 | Model::O3
1181 | Model::O3Mini
1182 | Model::Five
1183 | Model::FiveCodex
1184 | Model::FiveMini
1185 | Model::FiveNano => tiktoken_rs::num_tokens_from_messages(model.id(), &messages),
1186 // GPT-5.1, 5.2, 5.2-codex, 5.3-codex, 5.4, and 5.4-pro don't have dedicated tiktoken support; use gpt-5 tokenizer
1187 Model::FivePointOne
1188 | Model::FivePointTwo
1189 | Model::FivePointTwoCodex
1190 | Model::FivePointThreeCodex
1191 | Model::FivePointFour
1192 | Model::FivePointFourPro => tiktoken_rs::num_tokens_from_messages("gpt-5", &messages),
1193 }
1194 .map(|tokens| tokens as u64)
1195 })
1196 .boxed()
1197}
1198
1199struct ConfigurationView {
1200 api_key_editor: Entity<InputField>,
1201 state: Entity<State>,
1202 load_credentials_task: Option<Task<()>>,
1203}
1204
1205impl ConfigurationView {
1206 fn new(state: Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
1207 let api_key_editor = cx.new(|cx| {
1208 InputField::new(
1209 window,
1210 cx,
1211 "sk-000000000000000000000000000000000000000000000000",
1212 )
1213 });
1214
1215 cx.observe(&state, |_, _, cx| {
1216 cx.notify();
1217 })
1218 .detach();
1219
1220 let load_credentials_task = Some(cx.spawn_in(window, {
1221 let state = state.clone();
1222 async move |this, cx| {
1223 if let Some(task) = Some(state.update(cx, |state, cx| state.authenticate(cx))) {
1224 // We don't log an error, because "not signed in" is also an error.
1225 let _ = task.await;
1226 }
1227 this.update(cx, |this, cx| {
1228 this.load_credentials_task = None;
1229 cx.notify();
1230 })
1231 .log_err();
1232 }
1233 }));
1234
1235 Self {
1236 api_key_editor,
1237 state,
1238 load_credentials_task,
1239 }
1240 }
1241
1242 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
1243 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
1244 if api_key.is_empty() {
1245 return;
1246 }
1247
1248 // url changes can cause the editor to be displayed again
1249 self.api_key_editor
1250 .update(cx, |editor, cx| editor.set_text("", window, cx));
1251
1252 let state = self.state.clone();
1253 cx.spawn_in(window, async move |_, cx| {
1254 state
1255 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
1256 .await
1257 })
1258 .detach_and_log_err(cx);
1259 }
1260
1261 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
1262 self.api_key_editor
1263 .update(cx, |input, cx| input.set_text("", window, cx));
1264
1265 let state = self.state.clone();
1266 cx.spawn_in(window, async move |_, cx| {
1267 state
1268 .update(cx, |state, cx| state.set_api_key(None, cx))
1269 .await
1270 })
1271 .detach_and_log_err(cx);
1272 }
1273
1274 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
1275 !self.state.read(cx).is_authenticated()
1276 }
1277}
1278
1279impl Render for ConfigurationView {
1280 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1281 let env_var_set = self.state.read(cx).api_key_state.is_from_env_var();
1282 let configured_card_label = if env_var_set {
1283 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable")
1284 } else {
1285 let api_url = OpenAiLanguageModelProvider::api_url(cx);
1286 if api_url == OPEN_AI_API_URL {
1287 "API key configured".to_string()
1288 } else {
1289 format!("API key configured for {}", api_url)
1290 }
1291 };
1292
1293 let api_key_section = if self.should_render_editor(cx) {
1294 v_flex()
1295 .on_action(cx.listener(Self::save_api_key))
1296 .child(Label::new("To use Zed's agent with OpenAI, you need to add an API key. Follow these steps:"))
1297 .child(
1298 List::new()
1299 .child(
1300 ListBulletItem::new("")
1301 .child(Label::new("Create one by visiting"))
1302 .child(ButtonLink::new("OpenAI's console", "https://platform.openai.com/api-keys"))
1303 )
1304 .child(
1305 ListBulletItem::new("Ensure your OpenAI account has credits")
1306 )
1307 .child(
1308 ListBulletItem::new("Paste your API key below and hit enter to start using the agent")
1309 ),
1310 )
1311 .child(self.api_key_editor.clone())
1312 .child(
1313 Label::new(format!(
1314 "You can also set the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed."
1315 ))
1316 .size(LabelSize::Small)
1317 .color(Color::Muted),
1318 )
1319 .child(
1320 Label::new(
1321 "Note that having a subscription for another service like GitHub Copilot won't work.",
1322 )
1323 .size(LabelSize::Small).color(Color::Muted),
1324 )
1325 .into_any_element()
1326 } else {
1327 ConfiguredApiCard::new(configured_card_label)
1328 .disabled(env_var_set)
1329 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
1330 .when(env_var_set, |this| {
1331 this.tooltip_label(format!("To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable."))
1332 })
1333 .into_any_element()
1334 };
1335
1336 let compatible_api_section = h_flex()
1337 .mt_1p5()
1338 .gap_0p5()
1339 .flex_wrap()
1340 .when(self.should_render_editor(cx), |this| {
1341 this.pt_1p5()
1342 .border_t_1()
1343 .border_color(cx.theme().colors().border_variant)
1344 })
1345 .child(
1346 h_flex()
1347 .gap_2()
1348 .child(
1349 Icon::new(IconName::Info)
1350 .size(IconSize::XSmall)
1351 .color(Color::Muted),
1352 )
1353 .child(Label::new("Zed also supports OpenAI-compatible models.")),
1354 )
1355 .child(
1356 Button::new("docs", "Learn More")
1357 .icon(IconName::ArrowUpRight)
1358 .icon_size(IconSize::Small)
1359 .icon_color(Color::Muted)
1360 .on_click(move |_, _window, cx| {
1361 cx.open_url("https://zed.dev/docs/ai/llm-providers#openai-api-compatible")
1362 }),
1363 );
1364
1365 if self.load_credentials_task.is_some() {
1366 div().child(Label::new("Loading credentials…")).into_any()
1367 } else {
1368 v_flex()
1369 .size_full()
1370 .child(api_key_section)
1371 .child(compatible_api_section)
1372 .into_any()
1373 }
1374 }
1375}
1376
1377#[cfg(test)]
1378mod tests {
1379 use futures::{StreamExt, executor::block_on};
1380 use gpui::TestAppContext;
1381 use language_model::{LanguageModelRequestMessage, LanguageModelRequestTool};
1382 use open_ai::responses::{
1383 ResponseFunctionToolCall, ResponseOutputItem, ResponseOutputMessage, ResponseStatusDetails,
1384 ResponseSummary, ResponseUsage, StreamEvent as ResponsesStreamEvent,
1385 };
1386 use pretty_assertions::assert_eq;
1387 use serde_json::json;
1388
1389 use super::*;
1390
1391 fn map_response_events(events: Vec<ResponsesStreamEvent>) -> Vec<LanguageModelCompletionEvent> {
1392 block_on(async {
1393 OpenAiResponseEventMapper::new()
1394 .map_stream(Box::pin(futures::stream::iter(events.into_iter().map(Ok))))
1395 .collect::<Vec<_>>()
1396 .await
1397 .into_iter()
1398 .map(Result::unwrap)
1399 .collect()
1400 })
1401 }
1402
1403 fn response_item_message(id: &str) -> ResponseOutputItem {
1404 ResponseOutputItem::Message(ResponseOutputMessage {
1405 id: Some(id.to_string()),
1406 role: Some("assistant".to_string()),
1407 status: Some("in_progress".to_string()),
1408 content: vec![],
1409 })
1410 }
1411
1412 fn response_item_function_call(id: &str, args: Option<&str>) -> ResponseOutputItem {
1413 ResponseOutputItem::FunctionCall(ResponseFunctionToolCall {
1414 id: Some(id.to_string()),
1415 status: Some("in_progress".to_string()),
1416 name: Some("get_weather".to_string()),
1417 call_id: Some("call_123".to_string()),
1418 arguments: args.map(|s| s.to_string()).unwrap_or_default(),
1419 })
1420 }
1421
1422 #[gpui::test]
1423 fn tiktoken_rs_support(cx: &TestAppContext) {
1424 let request = LanguageModelRequest {
1425 thread_id: None,
1426 prompt_id: None,
1427 intent: None,
1428 messages: vec![LanguageModelRequestMessage {
1429 role: Role::User,
1430 content: vec![MessageContent::Text("message".into())],
1431 cache: false,
1432 reasoning_details: None,
1433 }],
1434 tools: vec![],
1435 tool_choice: None,
1436 stop: vec![],
1437 temperature: None,
1438 thinking_allowed: true,
1439 thinking_effort: None,
1440 };
1441
1442 // Validate that all models are supported by tiktoken-rs
1443 for model in Model::iter() {
1444 let count = cx
1445 .foreground_executor()
1446 .block_on(count_open_ai_tokens(
1447 request.clone(),
1448 model,
1449 &cx.app.borrow(),
1450 ))
1451 .unwrap();
1452 assert!(count > 0);
1453 }
1454 }
1455
1456 #[test]
1457 fn responses_stream_maps_text_and_usage() {
1458 let events = vec![
1459 ResponsesStreamEvent::OutputItemAdded {
1460 output_index: 0,
1461 sequence_number: None,
1462 item: response_item_message("msg_123"),
1463 },
1464 ResponsesStreamEvent::OutputTextDelta {
1465 item_id: "msg_123".into(),
1466 output_index: 0,
1467 content_index: Some(0),
1468 delta: "Hello".into(),
1469 },
1470 ResponsesStreamEvent::Completed {
1471 response: ResponseSummary {
1472 usage: Some(ResponseUsage {
1473 input_tokens: Some(5),
1474 output_tokens: Some(3),
1475 total_tokens: Some(8),
1476 }),
1477 ..Default::default()
1478 },
1479 },
1480 ];
1481
1482 let mapped = map_response_events(events);
1483 assert!(matches!(
1484 mapped[0],
1485 LanguageModelCompletionEvent::StartMessage { ref message_id } if message_id == "msg_123"
1486 ));
1487 assert!(matches!(
1488 mapped[1],
1489 LanguageModelCompletionEvent::Text(ref text) if text == "Hello"
1490 ));
1491 assert!(matches!(
1492 mapped[2],
1493 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1494 input_tokens: 5,
1495 output_tokens: 3,
1496 ..
1497 })
1498 ));
1499 assert!(matches!(
1500 mapped[3],
1501 LanguageModelCompletionEvent::Stop(StopReason::EndTurn)
1502 ));
1503 }
1504
1505 #[test]
1506 fn into_open_ai_response_builds_complete_payload() {
1507 let tool_call_id = LanguageModelToolUseId::from("call-42");
1508 let tool_input = json!({ "city": "Boston" });
1509 let tool_arguments = serde_json::to_string(&tool_input).unwrap();
1510 let tool_use = LanguageModelToolUse {
1511 id: tool_call_id.clone(),
1512 name: Arc::from("get_weather"),
1513 raw_input: tool_arguments.clone(),
1514 input: tool_input,
1515 is_input_complete: true,
1516 thought_signature: None,
1517 };
1518 let tool_result = LanguageModelToolResult {
1519 tool_use_id: tool_call_id,
1520 tool_name: Arc::from("get_weather"),
1521 is_error: false,
1522 content: LanguageModelToolResultContent::Text(Arc::from("Sunny")),
1523 output: Some(json!({ "forecast": "Sunny" })),
1524 };
1525 let user_image = LanguageModelImage {
1526 source: SharedString::from("aGVsbG8="),
1527 size: None,
1528 };
1529 let expected_image_url = user_image.to_base64_url();
1530
1531 let request = LanguageModelRequest {
1532 thread_id: Some("thread-123".into()),
1533 prompt_id: None,
1534 intent: None,
1535 messages: vec![
1536 LanguageModelRequestMessage {
1537 role: Role::System,
1538 content: vec![MessageContent::Text("System context".into())],
1539 cache: false,
1540 reasoning_details: None,
1541 },
1542 LanguageModelRequestMessage {
1543 role: Role::User,
1544 content: vec![
1545 MessageContent::Text("Please check the weather.".into()),
1546 MessageContent::Image(user_image),
1547 ],
1548 cache: false,
1549 reasoning_details: None,
1550 },
1551 LanguageModelRequestMessage {
1552 role: Role::Assistant,
1553 content: vec![
1554 MessageContent::Text("Looking that up.".into()),
1555 MessageContent::ToolUse(tool_use),
1556 ],
1557 cache: false,
1558 reasoning_details: None,
1559 },
1560 LanguageModelRequestMessage {
1561 role: Role::Assistant,
1562 content: vec![MessageContent::ToolResult(tool_result)],
1563 cache: false,
1564 reasoning_details: None,
1565 },
1566 ],
1567 tools: vec![LanguageModelRequestTool {
1568 name: "get_weather".into(),
1569 description: "Fetches the weather".into(),
1570 input_schema: json!({ "type": "object" }),
1571 }],
1572 tool_choice: Some(LanguageModelToolChoice::Any),
1573 stop: vec!["<STOP>".into()],
1574 temperature: None,
1575 thinking_allowed: false,
1576 thinking_effort: None,
1577 };
1578
1579 let response = into_open_ai_response(
1580 request,
1581 "custom-model",
1582 true,
1583 true,
1584 Some(2048),
1585 Some(ReasoningEffort::Low),
1586 );
1587
1588 let serialized = serde_json::to_value(&response).unwrap();
1589 let expected = json!({
1590 "model": "custom-model",
1591 "input": [
1592 {
1593 "type": "message",
1594 "role": "system",
1595 "content": [
1596 { "type": "input_text", "text": "System context" }
1597 ]
1598 },
1599 {
1600 "type": "message",
1601 "role": "user",
1602 "content": [
1603 { "type": "input_text", "text": "Please check the weather." },
1604 { "type": "input_image", "image_url": expected_image_url }
1605 ]
1606 },
1607 {
1608 "type": "message",
1609 "role": "assistant",
1610 "content": [
1611 { "type": "output_text", "text": "Looking that up.", "annotations": [] }
1612 ]
1613 },
1614 {
1615 "type": "function_call",
1616 "call_id": "call-42",
1617 "name": "get_weather",
1618 "arguments": tool_arguments
1619 },
1620 {
1621 "type": "function_call_output",
1622 "call_id": "call-42",
1623 "output": "{\"forecast\":\"Sunny\"}"
1624 }
1625 ],
1626 "stream": true,
1627 "max_output_tokens": 2048,
1628 "parallel_tool_calls": true,
1629 "tool_choice": "required",
1630 "tools": [
1631 {
1632 "type": "function",
1633 "name": "get_weather",
1634 "description": "Fetches the weather",
1635 "parameters": { "type": "object" }
1636 }
1637 ],
1638 "prompt_cache_key": "thread-123",
1639 "reasoning": { "effort": "low" }
1640 });
1641
1642 assert_eq!(serialized, expected);
1643 }
1644
1645 #[test]
1646 fn responses_stream_maps_tool_calls() {
1647 let events = vec![
1648 ResponsesStreamEvent::OutputItemAdded {
1649 output_index: 0,
1650 sequence_number: None,
1651 item: response_item_function_call("item_fn", Some("{\"city\":\"Bos")),
1652 },
1653 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1654 item_id: "item_fn".into(),
1655 output_index: 0,
1656 delta: "ton\"}".into(),
1657 sequence_number: None,
1658 },
1659 ResponsesStreamEvent::FunctionCallArgumentsDone {
1660 item_id: "item_fn".into(),
1661 output_index: 0,
1662 arguments: "{\"city\":\"Boston\"}".into(),
1663 sequence_number: None,
1664 },
1665 ResponsesStreamEvent::Completed {
1666 response: ResponseSummary::default(),
1667 },
1668 ];
1669
1670 let mapped = map_response_events(events);
1671 assert!(matches!(
1672 mapped[0],
1673 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1674 ref id,
1675 ref name,
1676 ref raw_input,
1677 ..
1678 }) if id.to_string() == "call_123"
1679 && name.as_ref() == "get_weather"
1680 && raw_input == "{\"city\":\"Boston\"}"
1681 ));
1682 assert!(matches!(
1683 mapped[1],
1684 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1685 ));
1686 }
1687
1688 #[test]
1689 fn responses_stream_uses_max_tokens_stop_reason() {
1690 let events = vec![ResponsesStreamEvent::Incomplete {
1691 response: ResponseSummary {
1692 status_details: Some(ResponseStatusDetails {
1693 reason: Some("max_output_tokens".into()),
1694 r#type: Some("incomplete".into()),
1695 error: None,
1696 }),
1697 usage: Some(ResponseUsage {
1698 input_tokens: Some(10),
1699 output_tokens: Some(20),
1700 total_tokens: Some(30),
1701 }),
1702 ..Default::default()
1703 },
1704 }];
1705
1706 let mapped = map_response_events(events);
1707 assert!(matches!(
1708 mapped[0],
1709 LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1710 input_tokens: 10,
1711 output_tokens: 20,
1712 ..
1713 })
1714 ));
1715 assert!(matches!(
1716 mapped[1],
1717 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1718 ));
1719 }
1720
1721 #[test]
1722 fn responses_stream_handles_multiple_tool_calls() {
1723 let events = vec![
1724 ResponsesStreamEvent::OutputItemAdded {
1725 output_index: 0,
1726 sequence_number: None,
1727 item: response_item_function_call("item_fn1", Some("{\"city\":\"NYC\"}")),
1728 },
1729 ResponsesStreamEvent::FunctionCallArgumentsDone {
1730 item_id: "item_fn1".into(),
1731 output_index: 0,
1732 arguments: "{\"city\":\"NYC\"}".into(),
1733 sequence_number: None,
1734 },
1735 ResponsesStreamEvent::OutputItemAdded {
1736 output_index: 1,
1737 sequence_number: None,
1738 item: response_item_function_call("item_fn2", Some("{\"city\":\"LA\"}")),
1739 },
1740 ResponsesStreamEvent::FunctionCallArgumentsDone {
1741 item_id: "item_fn2".into(),
1742 output_index: 1,
1743 arguments: "{\"city\":\"LA\"}".into(),
1744 sequence_number: None,
1745 },
1746 ResponsesStreamEvent::Completed {
1747 response: ResponseSummary::default(),
1748 },
1749 ];
1750
1751 let mapped = map_response_events(events);
1752 assert_eq!(mapped.len(), 3);
1753 assert!(matches!(
1754 mapped[0],
1755 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1756 if raw_input == "{\"city\":\"NYC\"}"
1757 ));
1758 assert!(matches!(
1759 mapped[1],
1760 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1761 if raw_input == "{\"city\":\"LA\"}"
1762 ));
1763 assert!(matches!(
1764 mapped[2],
1765 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1766 ));
1767 }
1768
1769 #[test]
1770 fn responses_stream_handles_mixed_text_and_tool_calls() {
1771 let events = vec![
1772 ResponsesStreamEvent::OutputItemAdded {
1773 output_index: 0,
1774 sequence_number: None,
1775 item: response_item_message("msg_123"),
1776 },
1777 ResponsesStreamEvent::OutputTextDelta {
1778 item_id: "msg_123".into(),
1779 output_index: 0,
1780 content_index: Some(0),
1781 delta: "Let me check that".into(),
1782 },
1783 ResponsesStreamEvent::OutputItemAdded {
1784 output_index: 1,
1785 sequence_number: None,
1786 item: response_item_function_call("item_fn", Some("{\"query\":\"test\"}")),
1787 },
1788 ResponsesStreamEvent::FunctionCallArgumentsDone {
1789 item_id: "item_fn".into(),
1790 output_index: 1,
1791 arguments: "{\"query\":\"test\"}".into(),
1792 sequence_number: None,
1793 },
1794 ResponsesStreamEvent::Completed {
1795 response: ResponseSummary::default(),
1796 },
1797 ];
1798
1799 let mapped = map_response_events(events);
1800 assert!(matches!(
1801 mapped[0],
1802 LanguageModelCompletionEvent::StartMessage { .. }
1803 ));
1804 assert!(matches!(
1805 mapped[1],
1806 LanguageModelCompletionEvent::Text(ref text) if text == "Let me check that"
1807 ));
1808 assert!(matches!(
1809 mapped[2],
1810 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1811 if raw_input == "{\"query\":\"test\"}"
1812 ));
1813 assert!(matches!(
1814 mapped[3],
1815 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1816 ));
1817 }
1818
1819 #[test]
1820 fn responses_stream_handles_json_parse_error() {
1821 let events = vec![
1822 ResponsesStreamEvent::OutputItemAdded {
1823 output_index: 0,
1824 sequence_number: None,
1825 item: response_item_function_call("item_fn", Some("{invalid json")),
1826 },
1827 ResponsesStreamEvent::FunctionCallArgumentsDone {
1828 item_id: "item_fn".into(),
1829 output_index: 0,
1830 arguments: "{invalid json".into(),
1831 sequence_number: None,
1832 },
1833 ResponsesStreamEvent::Completed {
1834 response: ResponseSummary::default(),
1835 },
1836 ];
1837
1838 let mapped = map_response_events(events);
1839 assert!(matches!(
1840 mapped[0],
1841 LanguageModelCompletionEvent::ToolUseJsonParseError {
1842 ref raw_input,
1843 ..
1844 } if raw_input.as_ref() == "{invalid json"
1845 ));
1846 }
1847
1848 #[test]
1849 fn responses_stream_handles_incomplete_function_call() {
1850 let events = vec![
1851 ResponsesStreamEvent::OutputItemAdded {
1852 output_index: 0,
1853 sequence_number: None,
1854 item: response_item_function_call("item_fn", Some("{\"city\":")),
1855 },
1856 ResponsesStreamEvent::FunctionCallArgumentsDelta {
1857 item_id: "item_fn".into(),
1858 output_index: 0,
1859 delta: "\"Boston\"".into(),
1860 sequence_number: None,
1861 },
1862 ResponsesStreamEvent::Incomplete {
1863 response: ResponseSummary {
1864 status_details: Some(ResponseStatusDetails {
1865 reason: Some("max_output_tokens".into()),
1866 r#type: Some("incomplete".into()),
1867 error: None,
1868 }),
1869 output: vec![response_item_function_call(
1870 "item_fn",
1871 Some("{\"city\":\"Boston\"}"),
1872 )],
1873 ..Default::default()
1874 },
1875 },
1876 ];
1877
1878 let mapped = map_response_events(events);
1879 assert!(matches!(
1880 mapped[0],
1881 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1882 if raw_input == "{\"city\":\"Boston\"}"
1883 ));
1884 assert!(matches!(
1885 mapped[1],
1886 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1887 ));
1888 }
1889
1890 #[test]
1891 fn responses_stream_incomplete_does_not_duplicate_tool_calls() {
1892 let events = vec![
1893 ResponsesStreamEvent::OutputItemAdded {
1894 output_index: 0,
1895 sequence_number: None,
1896 item: response_item_function_call("item_fn", Some("{\"city\":\"Boston\"}")),
1897 },
1898 ResponsesStreamEvent::FunctionCallArgumentsDone {
1899 item_id: "item_fn".into(),
1900 output_index: 0,
1901 arguments: "{\"city\":\"Boston\"}".into(),
1902 sequence_number: None,
1903 },
1904 ResponsesStreamEvent::Incomplete {
1905 response: ResponseSummary {
1906 status_details: Some(ResponseStatusDetails {
1907 reason: Some("max_output_tokens".into()),
1908 r#type: Some("incomplete".into()),
1909 error: None,
1910 }),
1911 output: vec![response_item_function_call(
1912 "item_fn",
1913 Some("{\"city\":\"Boston\"}"),
1914 )],
1915 ..Default::default()
1916 },
1917 },
1918 ];
1919
1920 let mapped = map_response_events(events);
1921 assert_eq!(mapped.len(), 2);
1922 assert!(matches!(
1923 mapped[0],
1924 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1925 if raw_input == "{\"city\":\"Boston\"}"
1926 ));
1927 assert!(matches!(
1928 mapped[1],
1929 LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1930 ));
1931 }
1932
1933 #[test]
1934 fn responses_stream_handles_empty_tool_arguments() {
1935 // Test that tools with no arguments (empty string) are handled correctly
1936 let events = vec![
1937 ResponsesStreamEvent::OutputItemAdded {
1938 output_index: 0,
1939 sequence_number: None,
1940 item: response_item_function_call("item_fn", Some("")),
1941 },
1942 ResponsesStreamEvent::FunctionCallArgumentsDone {
1943 item_id: "item_fn".into(),
1944 output_index: 0,
1945 arguments: "".into(),
1946 sequence_number: None,
1947 },
1948 ResponsesStreamEvent::Completed {
1949 response: ResponseSummary::default(),
1950 },
1951 ];
1952
1953 let mapped = map_response_events(events);
1954 assert_eq!(mapped.len(), 2);
1955
1956 // Should produce a ToolUse event with an empty object
1957 assert!(matches!(
1958 &mapped[0],
1959 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1960 id,
1961 name,
1962 raw_input,
1963 input,
1964 ..
1965 }) if id.to_string() == "call_123"
1966 && name.as_ref() == "get_weather"
1967 && raw_input == ""
1968 && input.is_object()
1969 && input.as_object().unwrap().is_empty()
1970 ));
1971
1972 assert!(matches!(
1973 mapped[1],
1974 LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1975 ));
1976 }
1977}