completion.rs

   1use anyhow::{Result, anyhow};
   2use collections::HashMap;
   3use futures::{Stream, StreamExt};
   4use language_model_core::{
   5    LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelImage,
   6    LanguageModelRequest, LanguageModelRequestMessage, LanguageModelToolChoice,
   7    LanguageModelToolResultContent, LanguageModelToolUse, LanguageModelToolUseId, MessageContent,
   8    Role, StopReason, TokenUsage,
   9    util::{fix_streamed_json, parse_tool_arguments},
  10};
  11use std::pin::Pin;
  12use std::sync::Arc;
  13
  14use crate::responses::{
  15    Request as ResponseRequest, ResponseFunctionCallItem, ResponseFunctionCallOutputContent,
  16    ResponseFunctionCallOutputItem, ResponseInputContent, ResponseInputItem, ResponseMessageItem,
  17    ResponseOutputItem, ResponseSummary as ResponsesSummary, ResponseUsage as ResponsesUsage,
  18    StreamEvent as ResponsesStreamEvent,
  19};
  20use crate::{
  21    FunctionContent, FunctionDefinition, ImageUrl, MessagePart, ReasoningEffort,
  22    ResponseStreamEvent, ToolCall, ToolCallContent,
  23};
  24
  25pub fn into_open_ai(
  26    request: LanguageModelRequest,
  27    model_id: &str,
  28    supports_parallel_tool_calls: bool,
  29    supports_prompt_cache_key: bool,
  30    max_output_tokens: Option<u64>,
  31    reasoning_effort: Option<ReasoningEffort>,
  32    interleaved_reasoning: bool,
  33) -> crate::Request {
  34    let stream = !model_id.starts_with("o1-");
  35
  36    let mut messages = Vec::new();
  37    let mut current_reasoning: Option<String> = None;
  38    for message in request.messages {
  39        for content in message.content {
  40            match content {
  41                MessageContent::Thinking { text, .. } if interleaved_reasoning => {
  42                    current_reasoning.get_or_insert_default().push_str(&text);
  43                }
  44                MessageContent::Text(text) | MessageContent::Thinking { text, .. } => {
  45                    let should_add = if message.role == Role::User {
  46                        // Including whitespace-only user messages can cause error with OpenAI compatible APIs
  47                        // See https://github.com/zed-industries/zed/issues/40097
  48                        !text.trim().is_empty()
  49                    } else {
  50                        !text.is_empty()
  51                    };
  52                    if should_add {
  53                        add_message_content_part(
  54                            MessagePart::Text { text },
  55                            message.role,
  56                            &mut messages,
  57                        );
  58                        if let Some(reasoning) = current_reasoning.take() {
  59                            if let Some(crate::RequestMessage::Assistant {
  60                                reasoning_content,
  61                                ..
  62                            }) = messages.last_mut()
  63                            {
  64                                *reasoning_content = Some(reasoning);
  65                            }
  66                        }
  67                    }
  68                }
  69                MessageContent::RedactedThinking(_) => {}
  70                MessageContent::Image(image) => {
  71                    add_message_content_part(
  72                        MessagePart::Image {
  73                            image_url: ImageUrl {
  74                                url: image.to_base64_url(),
  75                                detail: None,
  76                            },
  77                        },
  78                        message.role,
  79                        &mut messages,
  80                    );
  81                }
  82                MessageContent::ToolUse(tool_use) => {
  83                    let tool_call = ToolCall {
  84                        id: tool_use.id.to_string(),
  85                        content: ToolCallContent::Function {
  86                            function: FunctionContent {
  87                                name: tool_use.name.to_string(),
  88                                arguments: serde_json::to_string(&tool_use.input)
  89                                    .unwrap_or_default(),
  90                            },
  91                        },
  92                    };
  93
  94                    if let Some(crate::RequestMessage::Assistant { tool_calls, .. }) =
  95                        messages.last_mut()
  96                    {
  97                        tool_calls.push(tool_call);
  98                    } else {
  99                        messages.push(crate::RequestMessage::Assistant {
 100                            content: None,
 101                            tool_calls: vec![tool_call],
 102                            reasoning_content: current_reasoning.take(),
 103                        });
 104                    }
 105                }
 106                MessageContent::ToolResult(tool_result) => {
 107                    let content: Vec<MessagePart> = tool_result
 108                        .content
 109                        .iter()
 110                        .map(|part| match part {
 111                            LanguageModelToolResultContent::Text(text) => MessagePart::Text {
 112                                text: text.to_string(),
 113                            },
 114                            LanguageModelToolResultContent::Image(image) => MessagePart::Image {
 115                                image_url: ImageUrl {
 116                                    url: image.to_base64_url(),
 117                                    detail: None,
 118                                },
 119                            },
 120                        })
 121                        .collect();
 122
 123                    messages.push(crate::RequestMessage::Tool {
 124                        content: content.into(),
 125                        tool_call_id: tool_result.tool_use_id.to_string(),
 126                    });
 127                }
 128            }
 129        }
 130    }
 131
 132    crate::Request {
 133        model: model_id.into(),
 134        messages,
 135        stream,
 136        stream_options: if stream {
 137            Some(crate::StreamOptions::default())
 138        } else {
 139            None
 140        },
 141        stop: request.stop,
 142        temperature: request.temperature.or(Some(1.0)),
 143        max_completion_tokens: max_output_tokens,
 144        parallel_tool_calls: if supports_parallel_tool_calls && !request.tools.is_empty() {
 145            Some(supports_parallel_tool_calls)
 146        } else {
 147            None
 148        },
 149        prompt_cache_key: if supports_prompt_cache_key {
 150            request.thread_id
 151        } else {
 152            None
 153        },
 154        tools: request
 155            .tools
 156            .into_iter()
 157            .map(|tool| crate::ToolDefinition::Function {
 158                function: FunctionDefinition {
 159                    name: tool.name,
 160                    description: Some(tool.description),
 161                    parameters: Some(tool.input_schema),
 162                },
 163            })
 164            .collect(),
 165        tool_choice: request.tool_choice.map(|choice| match choice {
 166            LanguageModelToolChoice::Auto => crate::ToolChoice::Auto,
 167            LanguageModelToolChoice::Any => crate::ToolChoice::Required,
 168            LanguageModelToolChoice::None => crate::ToolChoice::None,
 169        }),
 170        reasoning_effort,
 171    }
 172}
 173
 174pub fn into_open_ai_response(
 175    request: LanguageModelRequest,
 176    model_id: &str,
 177    supports_parallel_tool_calls: bool,
 178    supports_prompt_cache_key: bool,
 179    max_output_tokens: Option<u64>,
 180    reasoning_effort: Option<ReasoningEffort>,
 181) -> ResponseRequest {
 182    let stream = !model_id.starts_with("o1-");
 183
 184    let LanguageModelRequest {
 185        thread_id,
 186        prompt_id: _,
 187        intent: _,
 188        messages,
 189        tools,
 190        tool_choice,
 191        stop: _,
 192        temperature,
 193        thinking_allowed: _,
 194        thinking_effort: _,
 195        speed: _,
 196    } = request;
 197
 198    let mut input_items = Vec::new();
 199    for (index, message) in messages.into_iter().enumerate() {
 200        append_message_to_response_items(message, index, &mut input_items);
 201    }
 202
 203    let tools: Vec<_> = tools
 204        .into_iter()
 205        .map(|tool| crate::responses::ToolDefinition::Function {
 206            name: tool.name,
 207            description: Some(tool.description),
 208            parameters: Some(tool.input_schema),
 209            strict: None,
 210        })
 211        .collect();
 212
 213    ResponseRequest {
 214        model: model_id.into(),
 215        input: input_items,
 216        stream,
 217        temperature,
 218        top_p: None,
 219        max_output_tokens,
 220        parallel_tool_calls: if tools.is_empty() {
 221            None
 222        } else {
 223            Some(supports_parallel_tool_calls)
 224        },
 225        tool_choice: tool_choice.map(|choice| match choice {
 226            LanguageModelToolChoice::Auto => crate::ToolChoice::Auto,
 227            LanguageModelToolChoice::Any => crate::ToolChoice::Required,
 228            LanguageModelToolChoice::None => crate::ToolChoice::None,
 229        }),
 230        tools,
 231        prompt_cache_key: if supports_prompt_cache_key {
 232            thread_id
 233        } else {
 234            None
 235        },
 236        reasoning: reasoning_effort.map(|effort| crate::responses::ReasoningConfig {
 237            effort,
 238            summary: Some(crate::responses::ReasoningSummaryMode::Auto),
 239        }),
 240    }
 241}
 242
 243fn append_message_to_response_items(
 244    message: LanguageModelRequestMessage,
 245    index: usize,
 246    input_items: &mut Vec<ResponseInputItem>,
 247) {
 248    let mut content_parts: Vec<ResponseInputContent> = Vec::new();
 249
 250    for content in message.content {
 251        match content {
 252            MessageContent::Text(text) => {
 253                push_response_text_part(&message.role, text, &mut content_parts);
 254            }
 255            MessageContent::Thinking { text, .. } => {
 256                push_response_text_part(&message.role, text, &mut content_parts);
 257            }
 258            MessageContent::RedactedThinking(_) => {}
 259            MessageContent::Image(image) => {
 260                push_response_image_part(&message.role, image, &mut content_parts);
 261            }
 262            MessageContent::ToolUse(tool_use) => {
 263                flush_response_parts(&message.role, index, &mut content_parts, input_items);
 264                let call_id = tool_use.id.to_string();
 265                input_items.push(ResponseInputItem::FunctionCall(ResponseFunctionCallItem {
 266                    call_id,
 267                    name: tool_use.name.to_string(),
 268                    arguments: tool_use.raw_input,
 269                }));
 270            }
 271            MessageContent::ToolResult(tool_result) => {
 272                flush_response_parts(&message.role, index, &mut content_parts, input_items);
 273                let output = match tool_result.content.as_slice() {
 274                    [LanguageModelToolResultContent::Text(text)] => {
 275                        ResponseFunctionCallOutputContent::Text(text.to_string())
 276                    }
 277                    _ => {
 278                        let parts = tool_result
 279                            .content
 280                            .into_iter()
 281                            .map(|part| match part {
 282                                LanguageModelToolResultContent::Text(text) => {
 283                                    ResponseInputContent::Text {
 284                                        text: text.to_string(),
 285                                    }
 286                                }
 287                                LanguageModelToolResultContent::Image(image) => {
 288                                    ResponseInputContent::Image {
 289                                        image_url: image.to_base64_url(),
 290                                    }
 291                                }
 292                            })
 293                            .collect();
 294                        ResponseFunctionCallOutputContent::List(parts)
 295                    }
 296                };
 297                input_items.push(ResponseInputItem::FunctionCallOutput(
 298                    ResponseFunctionCallOutputItem {
 299                        call_id: tool_result.tool_use_id.to_string(),
 300                        output,
 301                    },
 302                ));
 303            }
 304        }
 305    }
 306
 307    flush_response_parts(&message.role, index, &mut content_parts, input_items);
 308}
 309
 310fn push_response_text_part(
 311    role: &Role,
 312    text: impl Into<String>,
 313    parts: &mut Vec<ResponseInputContent>,
 314) {
 315    let text = text.into();
 316    if text.trim().is_empty() {
 317        return;
 318    }
 319
 320    match role {
 321        Role::Assistant => parts.push(ResponseInputContent::OutputText {
 322            text,
 323            annotations: Vec::new(),
 324        }),
 325        _ => parts.push(ResponseInputContent::Text { text }),
 326    }
 327}
 328
 329fn push_response_image_part(
 330    role: &Role,
 331    image: LanguageModelImage,
 332    parts: &mut Vec<ResponseInputContent>,
 333) {
 334    match role {
 335        Role::Assistant => parts.push(ResponseInputContent::OutputText {
 336            text: "[image omitted]".to_string(),
 337            annotations: Vec::new(),
 338        }),
 339        _ => parts.push(ResponseInputContent::Image {
 340            image_url: image.to_base64_url(),
 341        }),
 342    }
 343}
 344
 345fn flush_response_parts(
 346    role: &Role,
 347    _index: usize,
 348    parts: &mut Vec<ResponseInputContent>,
 349    input_items: &mut Vec<ResponseInputItem>,
 350) {
 351    if parts.is_empty() {
 352        return;
 353    }
 354
 355    let item = ResponseInputItem::Message(ResponseMessageItem {
 356        role: match role {
 357            Role::User => crate::Role::User,
 358            Role::Assistant => crate::Role::Assistant,
 359            Role::System => crate::Role::System,
 360        },
 361        content: parts.clone(),
 362    });
 363
 364    input_items.push(item);
 365    parts.clear();
 366}
 367
 368fn add_message_content_part(
 369    new_part: MessagePart,
 370    role: Role,
 371    messages: &mut Vec<crate::RequestMessage>,
 372) {
 373    match (role, messages.last_mut()) {
 374        (Role::User, Some(crate::RequestMessage::User { content }))
 375        | (
 376            Role::Assistant,
 377            Some(crate::RequestMessage::Assistant {
 378                content: Some(content),
 379                ..
 380            }),
 381        )
 382        | (Role::System, Some(crate::RequestMessage::System { content, .. })) => {
 383            content.push_part(new_part);
 384        }
 385        _ => {
 386            messages.push(match role {
 387                Role::User => crate::RequestMessage::User {
 388                    content: crate::MessageContent::from(vec![new_part]),
 389                },
 390                Role::Assistant => crate::RequestMessage::Assistant {
 391                    content: Some(crate::MessageContent::from(vec![new_part])),
 392                    tool_calls: Vec::new(),
 393                    reasoning_content: None,
 394                },
 395                Role::System => crate::RequestMessage::System {
 396                    content: crate::MessageContent::from(vec![new_part]),
 397                },
 398            });
 399        }
 400    }
 401}
 402
 403pub struct OpenAiEventMapper {
 404    tool_calls_by_index: HashMap<usize, RawToolCall>,
 405}
 406
 407impl OpenAiEventMapper {
 408    pub fn new() -> Self {
 409        Self {
 410            tool_calls_by_index: HashMap::default(),
 411        }
 412    }
 413
 414    pub fn map_stream(
 415        mut self,
 416        events: Pin<Box<dyn Send + Stream<Item = Result<ResponseStreamEvent>>>>,
 417    ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
 418    {
 419        events.flat_map(move |event| {
 420            futures::stream::iter(match event {
 421                Ok(event) => self.map_event(event),
 422                Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
 423            })
 424        })
 425    }
 426
 427    pub fn map_event(
 428        &mut self,
 429        event: ResponseStreamEvent,
 430    ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
 431        let mut events = Vec::new();
 432        if let Some(usage) = event.usage {
 433            events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
 434                input_tokens: usage.prompt_tokens,
 435                output_tokens: usage.completion_tokens,
 436                cache_creation_input_tokens: 0,
 437                cache_read_input_tokens: 0,
 438            })));
 439        }
 440
 441        let Some(choice) = event.choices.first() else {
 442            return events;
 443        };
 444
 445        if let Some(delta) = choice.delta.as_ref() {
 446            if let Some(reasoning_content) = delta.reasoning_content.clone() {
 447                if !reasoning_content.is_empty() {
 448                    events.push(Ok(LanguageModelCompletionEvent::Thinking {
 449                        text: reasoning_content,
 450                        signature: None,
 451                    }));
 452                }
 453            }
 454            if let Some(content) = delta.content.clone() {
 455                if !content.is_empty() {
 456                    events.push(Ok(LanguageModelCompletionEvent::Text(content)));
 457                }
 458            }
 459
 460            if let Some(tool_calls) = delta.tool_calls.as_ref() {
 461                for tool_call in tool_calls {
 462                    let entry = self.tool_calls_by_index.entry(tool_call.index).or_default();
 463
 464                    if let Some(tool_id) = tool_call.id.clone() {
 465                        entry.id = tool_id;
 466                    }
 467
 468                    if let Some(function) = tool_call.function.as_ref() {
 469                        if let Some(name) = function.name.clone() {
 470                            entry.name = name;
 471                        }
 472
 473                        if let Some(arguments) = function.arguments.clone() {
 474                            entry.arguments.push_str(&arguments);
 475                        }
 476                    }
 477
 478                    if !entry.id.is_empty() && !entry.name.is_empty() {
 479                        if let Ok(input) = serde_json::from_str::<serde_json::Value>(
 480                            &fix_streamed_json(&entry.arguments),
 481                        ) {
 482                            events.push(Ok(LanguageModelCompletionEvent::ToolUse(
 483                                LanguageModelToolUse {
 484                                    id: entry.id.clone().into(),
 485                                    name: entry.name.as_str().into(),
 486                                    is_input_complete: false,
 487                                    input,
 488                                    raw_input: entry.arguments.clone(),
 489                                    thought_signature: None,
 490                                },
 491                            )));
 492                        }
 493                    }
 494                }
 495            }
 496        }
 497
 498        match choice.finish_reason.as_deref() {
 499            Some("stop") => {
 500                events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
 501            }
 502            Some("tool_calls") => {
 503                events.extend(self.tool_calls_by_index.drain().map(|(_, tool_call)| {
 504                    match parse_tool_arguments(&tool_call.arguments) {
 505                        Ok(input) => Ok(LanguageModelCompletionEvent::ToolUse(
 506                            LanguageModelToolUse {
 507                                id: tool_call.id.clone().into(),
 508                                name: tool_call.name.as_str().into(),
 509                                is_input_complete: true,
 510                                input,
 511                                raw_input: tool_call.arguments.clone(),
 512                                thought_signature: None,
 513                            },
 514                        )),
 515                        Err(error) => Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
 516                            id: tool_call.id.into(),
 517                            tool_name: tool_call.name.into(),
 518                            raw_input: tool_call.arguments.clone().into(),
 519                            json_parse_error: error.to_string(),
 520                        }),
 521                    }
 522                }));
 523
 524                events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
 525            }
 526            Some(stop_reason) => {
 527                log::error!("Unexpected OpenAI stop_reason: {stop_reason:?}",);
 528                events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
 529            }
 530            None => {}
 531        }
 532
 533        events
 534    }
 535}
 536
 537#[derive(Default)]
 538struct RawToolCall {
 539    id: String,
 540    name: String,
 541    arguments: String,
 542}
 543
 544pub struct OpenAiResponseEventMapper {
 545    function_calls_by_item: HashMap<String, PendingResponseFunctionCall>,
 546    pending_stop_reason: Option<StopReason>,
 547}
 548
 549#[derive(Default)]
 550struct PendingResponseFunctionCall {
 551    call_id: String,
 552    name: Arc<str>,
 553    arguments: String,
 554}
 555
 556impl OpenAiResponseEventMapper {
 557    pub fn new() -> Self {
 558        Self {
 559            function_calls_by_item: HashMap::default(),
 560            pending_stop_reason: None,
 561        }
 562    }
 563
 564    pub fn map_stream(
 565        mut self,
 566        events: Pin<Box<dyn Send + Stream<Item = Result<ResponsesStreamEvent>>>>,
 567    ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
 568    {
 569        events.flat_map(move |event| {
 570            futures::stream::iter(match event {
 571                Ok(event) => self.map_event(event),
 572                Err(error) => vec![Err(LanguageModelCompletionError::from(anyhow!(error)))],
 573            })
 574        })
 575    }
 576
 577    pub fn map_event(
 578        &mut self,
 579        event: ResponsesStreamEvent,
 580    ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
 581        match event {
 582            ResponsesStreamEvent::OutputItemAdded { item, .. } => {
 583                let mut events = Vec::new();
 584
 585                match &item {
 586                    ResponseOutputItem::Message(message) => {
 587                        if let Some(id) = &message.id {
 588                            events.push(Ok(LanguageModelCompletionEvent::StartMessage {
 589                                message_id: id.clone(),
 590                            }));
 591                        }
 592                    }
 593                    ResponseOutputItem::FunctionCall(function_call) => {
 594                        if let Some(item_id) = function_call.id.clone() {
 595                            let call_id = function_call
 596                                .call_id
 597                                .clone()
 598                                .or_else(|| function_call.id.clone())
 599                                .unwrap_or_else(|| item_id.clone());
 600                            let entry = PendingResponseFunctionCall {
 601                                call_id,
 602                                name: Arc::<str>::from(
 603                                    function_call.name.clone().unwrap_or_default(),
 604                                ),
 605                                arguments: function_call.arguments.clone(),
 606                            };
 607                            self.function_calls_by_item.insert(item_id, entry);
 608                        }
 609                    }
 610                    ResponseOutputItem::Reasoning(_) | ResponseOutputItem::Unknown => {}
 611                }
 612                events
 613            }
 614            ResponsesStreamEvent::ReasoningSummaryTextDelta { delta, .. } => {
 615                if delta.is_empty() {
 616                    Vec::new()
 617                } else {
 618                    vec![Ok(LanguageModelCompletionEvent::Thinking {
 619                        text: delta,
 620                        signature: None,
 621                    })]
 622                }
 623            }
 624            ResponsesStreamEvent::OutputTextDelta { delta, .. } => {
 625                if delta.is_empty() {
 626                    Vec::new()
 627                } else {
 628                    vec![Ok(LanguageModelCompletionEvent::Text(delta))]
 629                }
 630            }
 631            ResponsesStreamEvent::FunctionCallArgumentsDelta { item_id, delta, .. } => {
 632                if let Some(entry) = self.function_calls_by_item.get_mut(&item_id) {
 633                    entry.arguments.push_str(&delta);
 634                    if let Ok(input) = serde_json::from_str::<serde_json::Value>(
 635                        &fix_streamed_json(&entry.arguments),
 636                    ) {
 637                        return vec![Ok(LanguageModelCompletionEvent::ToolUse(
 638                            LanguageModelToolUse {
 639                                id: LanguageModelToolUseId::from(entry.call_id.clone()),
 640                                name: entry.name.clone(),
 641                                is_input_complete: false,
 642                                input,
 643                                raw_input: entry.arguments.clone(),
 644                                thought_signature: None,
 645                            },
 646                        ))];
 647                    }
 648                }
 649                Vec::new()
 650            }
 651            ResponsesStreamEvent::FunctionCallArgumentsDone {
 652                item_id, arguments, ..
 653            } => {
 654                if let Some(mut entry) = self.function_calls_by_item.remove(&item_id) {
 655                    if !arguments.is_empty() {
 656                        entry.arguments = arguments;
 657                    }
 658                    let raw_input = entry.arguments.clone();
 659                    self.pending_stop_reason = Some(StopReason::ToolUse);
 660                    match parse_tool_arguments(&entry.arguments) {
 661                        Ok(input) => {
 662                            vec![Ok(LanguageModelCompletionEvent::ToolUse(
 663                                LanguageModelToolUse {
 664                                    id: LanguageModelToolUseId::from(entry.call_id.clone()),
 665                                    name: entry.name.clone(),
 666                                    is_input_complete: true,
 667                                    input,
 668                                    raw_input,
 669                                    thought_signature: None,
 670                                },
 671                            ))]
 672                        }
 673                        Err(error) => {
 674                            vec![Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
 675                                id: LanguageModelToolUseId::from(entry.call_id.clone()),
 676                                tool_name: entry.name.clone(),
 677                                raw_input: Arc::<str>::from(raw_input),
 678                                json_parse_error: error.to_string(),
 679                            })]
 680                        }
 681                    }
 682                } else {
 683                    Vec::new()
 684                }
 685            }
 686            ResponsesStreamEvent::Completed { response } => {
 687                self.handle_completion(response, StopReason::EndTurn)
 688            }
 689            ResponsesStreamEvent::Incomplete { response } => {
 690                let reason = response
 691                    .status_details
 692                    .as_ref()
 693                    .and_then(|details| details.reason.as_deref());
 694                let stop_reason = match reason {
 695                    Some("max_output_tokens") => StopReason::MaxTokens,
 696                    Some("content_filter") => {
 697                        self.pending_stop_reason = Some(StopReason::Refusal);
 698                        StopReason::Refusal
 699                    }
 700                    _ => self
 701                        .pending_stop_reason
 702                        .take()
 703                        .unwrap_or(StopReason::EndTurn),
 704                };
 705
 706                let mut events = Vec::new();
 707                if self.pending_stop_reason.is_none() {
 708                    events.extend(self.emit_tool_calls_from_output(&response.output));
 709                }
 710                if let Some(usage) = response.usage.as_ref() {
 711                    events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
 712                        token_usage_from_response_usage(usage),
 713                    )));
 714                }
 715                events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
 716                events
 717            }
 718            ResponsesStreamEvent::Failed { response } => {
 719                let message = response
 720                    .status_details
 721                    .and_then(|details| details.error)
 722                    .map(|error| error.to_string())
 723                    .unwrap_or_else(|| "response failed".to_string());
 724                vec![Err(LanguageModelCompletionError::Other(anyhow!(message)))]
 725            }
 726            ResponsesStreamEvent::Error { error }
 727            | ResponsesStreamEvent::GenericError { error } => {
 728                vec![Err(LanguageModelCompletionError::Other(anyhow!(
 729                    error.message
 730                )))]
 731            }
 732            ResponsesStreamEvent::ReasoningSummaryPartAdded { summary_index, .. } => {
 733                if summary_index > 0 {
 734                    vec![Ok(LanguageModelCompletionEvent::Thinking {
 735                        text: "\n\n".to_string(),
 736                        signature: None,
 737                    })]
 738                } else {
 739                    Vec::new()
 740                }
 741            }
 742            ResponsesStreamEvent::OutputTextDone { .. }
 743            | ResponsesStreamEvent::OutputItemDone { .. }
 744            | ResponsesStreamEvent::ContentPartAdded { .. }
 745            | ResponsesStreamEvent::ContentPartDone { .. }
 746            | ResponsesStreamEvent::ReasoningSummaryTextDone { .. }
 747            | ResponsesStreamEvent::ReasoningSummaryPartDone { .. }
 748            | ResponsesStreamEvent::Created { .. }
 749            | ResponsesStreamEvent::InProgress { .. }
 750            | ResponsesStreamEvent::Unknown => Vec::new(),
 751        }
 752    }
 753
 754    fn handle_completion(
 755        &mut self,
 756        response: ResponsesSummary,
 757        default_reason: StopReason,
 758    ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
 759        let mut events = Vec::new();
 760
 761        if self.pending_stop_reason.is_none() {
 762            events.extend(self.emit_tool_calls_from_output(&response.output));
 763        }
 764
 765        if let Some(usage) = response.usage.as_ref() {
 766            events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
 767                token_usage_from_response_usage(usage),
 768            )));
 769        }
 770
 771        let stop_reason = self.pending_stop_reason.take().unwrap_or(default_reason);
 772        events.push(Ok(LanguageModelCompletionEvent::Stop(stop_reason)));
 773        events
 774    }
 775
 776    fn emit_tool_calls_from_output(
 777        &mut self,
 778        output: &[ResponseOutputItem],
 779    ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
 780        let mut events = Vec::new();
 781        for item in output {
 782            if let ResponseOutputItem::FunctionCall(function_call) = item {
 783                let Some(call_id) = function_call
 784                    .call_id
 785                    .clone()
 786                    .or_else(|| function_call.id.clone())
 787                else {
 788                    log::error!(
 789                        "Function call item missing both call_id and id: {:?}",
 790                        function_call
 791                    );
 792                    continue;
 793                };
 794                let name: Arc<str> = Arc::from(function_call.name.clone().unwrap_or_default());
 795                let arguments = &function_call.arguments;
 796                self.pending_stop_reason = Some(StopReason::ToolUse);
 797                match parse_tool_arguments(arguments) {
 798                    Ok(input) => {
 799                        events.push(Ok(LanguageModelCompletionEvent::ToolUse(
 800                            LanguageModelToolUse {
 801                                id: LanguageModelToolUseId::from(call_id.clone()),
 802                                name: name.clone(),
 803                                is_input_complete: true,
 804                                input,
 805                                raw_input: arguments.clone(),
 806                                thought_signature: None,
 807                            },
 808                        )));
 809                    }
 810                    Err(error) => {
 811                        events.push(Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
 812                            id: LanguageModelToolUseId::from(call_id.clone()),
 813                            tool_name: name.clone(),
 814                            raw_input: Arc::<str>::from(arguments.clone()),
 815                            json_parse_error: error.to_string(),
 816                        }));
 817                    }
 818                }
 819            }
 820        }
 821        events
 822    }
 823}
 824
 825fn token_usage_from_response_usage(usage: &ResponsesUsage) -> TokenUsage {
 826    TokenUsage {
 827        input_tokens: usage.input_tokens.unwrap_or_default(),
 828        output_tokens: usage.output_tokens.unwrap_or_default(),
 829        cache_creation_input_tokens: 0,
 830        cache_read_input_tokens: 0,
 831    }
 832}
 833
 834#[cfg(test)]
 835mod tests {
 836    use crate::responses::{
 837        ReasoningSummaryPart, ResponseFunctionToolCall, ResponseOutputItem, ResponseOutputMessage,
 838        ResponseReasoningItem, ResponseStatusDetails, ResponseSummary, ResponseUsage,
 839        StreamEvent as ResponsesStreamEvent,
 840    };
 841    use futures::{StreamExt, executor::block_on};
 842    use language_model_core::{
 843        LanguageModelImage, LanguageModelRequestMessage, LanguageModelRequestTool,
 844        LanguageModelToolResult, LanguageModelToolResultContent, LanguageModelToolUse,
 845        LanguageModelToolUseId, SharedString,
 846    };
 847    use pretty_assertions::assert_eq;
 848    use serde_json::json;
 849
 850    use super::*;
 851
 852    fn map_response_events(events: Vec<ResponsesStreamEvent>) -> Vec<LanguageModelCompletionEvent> {
 853        block_on(async {
 854            OpenAiResponseEventMapper::new()
 855                .map_stream(Box::pin(futures::stream::iter(events.into_iter().map(Ok))))
 856                .collect::<Vec<_>>()
 857                .await
 858                .into_iter()
 859                .map(Result::unwrap)
 860                .collect()
 861        })
 862    }
 863
 864    fn response_item_message(id: &str) -> ResponseOutputItem {
 865        ResponseOutputItem::Message(ResponseOutputMessage {
 866            id: Some(id.to_string()),
 867            role: Some("assistant".to_string()),
 868            status: Some("in_progress".to_string()),
 869            content: vec![],
 870        })
 871    }
 872
 873    fn response_item_function_call(id: &str, args: Option<&str>) -> ResponseOutputItem {
 874        ResponseOutputItem::FunctionCall(ResponseFunctionToolCall {
 875            id: Some(id.to_string()),
 876            status: Some("in_progress".to_string()),
 877            name: Some("get_weather".to_string()),
 878            call_id: Some("call_123".to_string()),
 879            arguments: args.map(|s| s.to_string()).unwrap_or_default(),
 880        })
 881    }
 882
 883    #[test]
 884    fn responses_stream_maps_text_and_usage() {
 885        let events = vec![
 886            ResponsesStreamEvent::OutputItemAdded {
 887                output_index: 0,
 888                sequence_number: None,
 889                item: response_item_message("msg_123"),
 890            },
 891            ResponsesStreamEvent::OutputTextDelta {
 892                item_id: "msg_123".into(),
 893                output_index: 0,
 894                content_index: Some(0),
 895                delta: "Hello".into(),
 896            },
 897            ResponsesStreamEvent::Completed {
 898                response: ResponseSummary {
 899                    usage: Some(ResponseUsage {
 900                        input_tokens: Some(5),
 901                        output_tokens: Some(3),
 902                        total_tokens: Some(8),
 903                    }),
 904                    ..Default::default()
 905                },
 906            },
 907        ];
 908
 909        let mapped = map_response_events(events);
 910        assert!(matches!(
 911            mapped[0],
 912            LanguageModelCompletionEvent::StartMessage { ref message_id } if message_id == "msg_123"
 913        ));
 914        assert!(matches!(
 915            mapped[1],
 916            LanguageModelCompletionEvent::Text(ref text) if text == "Hello"
 917        ));
 918        assert!(matches!(
 919            mapped[2],
 920            LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
 921                input_tokens: 5,
 922                output_tokens: 3,
 923                ..
 924            })
 925        ));
 926        assert!(matches!(
 927            mapped[3],
 928            LanguageModelCompletionEvent::Stop(StopReason::EndTurn)
 929        ));
 930    }
 931
 932    #[test]
 933    fn into_open_ai_response_builds_complete_payload() {
 934        let tool_call_id = LanguageModelToolUseId::from("call-42");
 935        let tool_input = json!({ "city": "Boston" });
 936        let tool_arguments = serde_json::to_string(&tool_input).unwrap();
 937        let tool_use = LanguageModelToolUse {
 938            id: tool_call_id.clone(),
 939            name: Arc::from("get_weather"),
 940            raw_input: tool_arguments.clone(),
 941            input: tool_input,
 942            is_input_complete: true,
 943            thought_signature: None,
 944        };
 945        let tool_result = LanguageModelToolResult {
 946            tool_use_id: tool_call_id,
 947            tool_name: Arc::from("get_weather"),
 948            is_error: false,
 949            content: vec![LanguageModelToolResultContent::Text(Arc::from("Sunny"))],
 950            output: Some(json!({ "forecast": "Sunny" })),
 951        };
 952        let user_image = LanguageModelImage {
 953            source: SharedString::from("aGVsbG8="),
 954            size: None,
 955        };
 956        let expected_image_url = user_image.to_base64_url();
 957
 958        let request = LanguageModelRequest {
 959            thread_id: Some("thread-123".into()),
 960            prompt_id: None,
 961            intent: None,
 962            messages: vec![
 963                LanguageModelRequestMessage {
 964                    role: Role::System,
 965                    content: vec![MessageContent::Text("System context".into())],
 966                    cache: false,
 967                    reasoning_details: None,
 968                },
 969                LanguageModelRequestMessage {
 970                    role: Role::User,
 971                    content: vec![
 972                        MessageContent::Text("Please check the weather.".into()),
 973                        MessageContent::Image(user_image),
 974                    ],
 975                    cache: false,
 976                    reasoning_details: None,
 977                },
 978                LanguageModelRequestMessage {
 979                    role: Role::Assistant,
 980                    content: vec![
 981                        MessageContent::Text("Looking that up.".into()),
 982                        MessageContent::ToolUse(tool_use),
 983                    ],
 984                    cache: false,
 985                    reasoning_details: None,
 986                },
 987                LanguageModelRequestMessage {
 988                    role: Role::Assistant,
 989                    content: vec![MessageContent::ToolResult(tool_result)],
 990                    cache: false,
 991                    reasoning_details: None,
 992                },
 993            ],
 994            tools: vec![LanguageModelRequestTool {
 995                name: "get_weather".into(),
 996                description: "Fetches the weather".into(),
 997                input_schema: json!({ "type": "object" }),
 998                use_input_streaming: false,
 999            }],
1000            tool_choice: Some(LanguageModelToolChoice::Any),
1001            stop: vec!["<STOP>".into()],
1002            temperature: None,
1003            thinking_allowed: false,
1004            thinking_effort: None,
1005            speed: None,
1006        };
1007
1008        let response = into_open_ai_response(
1009            request,
1010            "custom-model",
1011            true,
1012            true,
1013            Some(2048),
1014            Some(ReasoningEffort::Low),
1015        );
1016
1017        let serialized = serde_json::to_value(&response).unwrap();
1018        let expected = json!({
1019            "model": "custom-model",
1020            "input": [
1021                {
1022                    "type": "message",
1023                    "role": "system",
1024                    "content": [
1025                        { "type": "input_text", "text": "System context" }
1026                    ]
1027                },
1028                {
1029                    "type": "message",
1030                    "role": "user",
1031                    "content": [
1032                        { "type": "input_text", "text": "Please check the weather." },
1033                        { "type": "input_image", "image_url": expected_image_url }
1034                    ]
1035                },
1036                {
1037                    "type": "message",
1038                    "role": "assistant",
1039                    "content": [
1040                        { "type": "output_text", "text": "Looking that up.", "annotations": [] }
1041                    ]
1042                },
1043                {
1044                    "type": "function_call",
1045                    "call_id": "call-42",
1046                    "name": "get_weather",
1047                    "arguments": tool_arguments
1048                },
1049                {
1050                    "type": "function_call_output",
1051                    "call_id": "call-42",
1052                    "output": "Sunny"
1053                }
1054            ],
1055            "stream": true,
1056            "max_output_tokens": 2048,
1057            "parallel_tool_calls": true,
1058            "tool_choice": "required",
1059            "tools": [
1060                {
1061                    "type": "function",
1062                    "name": "get_weather",
1063                    "description": "Fetches the weather",
1064                    "parameters": { "type": "object" }
1065                }
1066            ],
1067            "prompt_cache_key": "thread-123",
1068            "reasoning": { "effort": "low", "summary": "auto" }
1069        });
1070
1071        assert_eq!(serialized, expected);
1072    }
1073
1074    #[test]
1075    fn responses_stream_maps_tool_calls() {
1076        let events = vec![
1077            ResponsesStreamEvent::OutputItemAdded {
1078                output_index: 0,
1079                sequence_number: None,
1080                item: response_item_function_call("item_fn", Some("{\"city\":\"Bos")),
1081            },
1082            ResponsesStreamEvent::FunctionCallArgumentsDelta {
1083                item_id: "item_fn".into(),
1084                output_index: 0,
1085                delta: "ton\"}".into(),
1086                sequence_number: None,
1087            },
1088            ResponsesStreamEvent::FunctionCallArgumentsDone {
1089                item_id: "item_fn".into(),
1090                output_index: 0,
1091                arguments: "{\"city\":\"Boston\"}".into(),
1092                sequence_number: None,
1093            },
1094            ResponsesStreamEvent::Completed {
1095                response: ResponseSummary::default(),
1096            },
1097        ];
1098
1099        let mapped = map_response_events(events);
1100        assert_eq!(mapped.len(), 3);
1101        assert!(matches!(
1102            mapped[0],
1103            LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1104                is_input_complete: false,
1105                ..
1106            })
1107        ));
1108        assert!(matches!(
1109            mapped[1],
1110            LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1111                ref id,
1112                ref name,
1113                ref raw_input,
1114                is_input_complete: true,
1115                ..
1116            }) if id.to_string() == "call_123"
1117                && name.as_ref() == "get_weather"
1118                && raw_input == "{\"city\":\"Boston\"}"
1119        ));
1120        assert!(matches!(
1121            mapped[2],
1122            LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1123        ));
1124    }
1125
1126    #[test]
1127    fn responses_stream_uses_max_tokens_stop_reason() {
1128        let events = vec![ResponsesStreamEvent::Incomplete {
1129            response: ResponseSummary {
1130                status_details: Some(ResponseStatusDetails {
1131                    reason: Some("max_output_tokens".into()),
1132                    r#type: Some("incomplete".into()),
1133                    error: None,
1134                }),
1135                usage: Some(ResponseUsage {
1136                    input_tokens: Some(10),
1137                    output_tokens: Some(20),
1138                    total_tokens: Some(30),
1139                }),
1140                ..Default::default()
1141            },
1142        }];
1143
1144        let mapped = map_response_events(events);
1145        assert!(matches!(
1146            mapped[0],
1147            LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
1148                input_tokens: 10,
1149                output_tokens: 20,
1150                ..
1151            })
1152        ));
1153        assert!(matches!(
1154            mapped[1],
1155            LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1156        ));
1157    }
1158
1159    #[test]
1160    fn responses_stream_handles_multiple_tool_calls() {
1161        let events = vec![
1162            ResponsesStreamEvent::OutputItemAdded {
1163                output_index: 0,
1164                sequence_number: None,
1165                item: response_item_function_call("item_fn1", Some("{\"city\":\"NYC\"}")),
1166            },
1167            ResponsesStreamEvent::FunctionCallArgumentsDone {
1168                item_id: "item_fn1".into(),
1169                output_index: 0,
1170                arguments: "{\"city\":\"NYC\"}".into(),
1171                sequence_number: None,
1172            },
1173            ResponsesStreamEvent::OutputItemAdded {
1174                output_index: 1,
1175                sequence_number: None,
1176                item: response_item_function_call("item_fn2", Some("{\"city\":\"LA\"}")),
1177            },
1178            ResponsesStreamEvent::FunctionCallArgumentsDone {
1179                item_id: "item_fn2".into(),
1180                output_index: 1,
1181                arguments: "{\"city\":\"LA\"}".into(),
1182                sequence_number: None,
1183            },
1184            ResponsesStreamEvent::Completed {
1185                response: ResponseSummary::default(),
1186            },
1187        ];
1188
1189        let mapped = map_response_events(events);
1190        assert_eq!(mapped.len(), 3);
1191        assert!(matches!(
1192            mapped[0],
1193            LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1194            if raw_input == "{\"city\":\"NYC\"}"
1195        ));
1196        assert!(matches!(
1197            mapped[1],
1198            LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. })
1199            if raw_input == "{\"city\":\"LA\"}"
1200        ));
1201        assert!(matches!(
1202            mapped[2],
1203            LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1204        ));
1205    }
1206
1207    #[test]
1208    fn responses_stream_handles_mixed_text_and_tool_calls() {
1209        let events = vec![
1210            ResponsesStreamEvent::OutputItemAdded {
1211                output_index: 0,
1212                sequence_number: None,
1213                item: response_item_message("msg_123"),
1214            },
1215            ResponsesStreamEvent::OutputTextDelta {
1216                item_id: "msg_123".into(),
1217                output_index: 0,
1218                content_index: Some(0),
1219                delta: "Let me check that".into(),
1220            },
1221            ResponsesStreamEvent::OutputItemAdded {
1222                output_index: 1,
1223                sequence_number: None,
1224                item: response_item_function_call("item_fn", Some("{\"query\":\"test\"}")),
1225            },
1226            ResponsesStreamEvent::FunctionCallArgumentsDone {
1227                item_id: "item_fn".into(),
1228                output_index: 1,
1229                arguments: "{\"query\":\"test\"}".into(),
1230                sequence_number: None,
1231            },
1232            ResponsesStreamEvent::Completed {
1233                response: ResponseSummary::default(),
1234            },
1235        ];
1236
1237        let mapped = map_response_events(events);
1238        assert!(matches!(
1239            mapped[0],
1240            LanguageModelCompletionEvent::StartMessage { .. }
1241        ));
1242        assert!(
1243            matches!(mapped[1], LanguageModelCompletionEvent::Text(ref text) if text == "Let me check that")
1244        );
1245        assert!(
1246            matches!(mapped[2], LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. }) if raw_input == "{\"query\":\"test\"}")
1247        );
1248        assert!(matches!(
1249            mapped[3],
1250            LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1251        ));
1252    }
1253
1254    #[test]
1255    fn responses_stream_handles_json_parse_error() {
1256        let events = vec![
1257            ResponsesStreamEvent::OutputItemAdded {
1258                output_index: 0,
1259                sequence_number: None,
1260                item: response_item_function_call("item_fn", Some("{invalid json")),
1261            },
1262            ResponsesStreamEvent::FunctionCallArgumentsDone {
1263                item_id: "item_fn".into(),
1264                output_index: 0,
1265                arguments: "{invalid json".into(),
1266                sequence_number: None,
1267            },
1268            ResponsesStreamEvent::Completed {
1269                response: ResponseSummary::default(),
1270            },
1271        ];
1272
1273        let mapped = map_response_events(events);
1274        assert!(matches!(
1275            mapped[0],
1276            LanguageModelCompletionEvent::ToolUseJsonParseError { ref raw_input, .. }
1277            if raw_input.as_ref() == "{invalid json"
1278        ));
1279    }
1280
1281    #[test]
1282    fn responses_stream_handles_incomplete_function_call() {
1283        let events = vec![
1284            ResponsesStreamEvent::OutputItemAdded {
1285                output_index: 0,
1286                sequence_number: None,
1287                item: response_item_function_call("item_fn", Some("{\"city\":")),
1288            },
1289            ResponsesStreamEvent::FunctionCallArgumentsDelta {
1290                item_id: "item_fn".into(),
1291                output_index: 0,
1292                delta: "\"Boston\"".into(),
1293                sequence_number: None,
1294            },
1295            ResponsesStreamEvent::Incomplete {
1296                response: ResponseSummary {
1297                    status_details: Some(ResponseStatusDetails {
1298                        reason: Some("max_output_tokens".into()),
1299                        r#type: Some("incomplete".into()),
1300                        error: None,
1301                    }),
1302                    output: vec![response_item_function_call(
1303                        "item_fn",
1304                        Some("{\"city\":\"Boston\"}"),
1305                    )],
1306                    ..Default::default()
1307                },
1308            },
1309        ];
1310
1311        let mapped = map_response_events(events);
1312        assert_eq!(mapped.len(), 3);
1313        assert!(matches!(
1314            mapped[0],
1315            LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1316                is_input_complete: false,
1317                ..
1318            })
1319        ));
1320        assert!(
1321            matches!(mapped[1], LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, is_input_complete: true, .. }) if raw_input == "{\"city\":\"Boston\"}")
1322        );
1323        assert!(matches!(
1324            mapped[2],
1325            LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1326        ));
1327    }
1328
1329    #[test]
1330    fn responses_stream_incomplete_does_not_duplicate_tool_calls() {
1331        let events = vec![
1332            ResponsesStreamEvent::OutputItemAdded {
1333                output_index: 0,
1334                sequence_number: None,
1335                item: response_item_function_call("item_fn", Some("{\"city\":\"Boston\"}")),
1336            },
1337            ResponsesStreamEvent::FunctionCallArgumentsDone {
1338                item_id: "item_fn".into(),
1339                output_index: 0,
1340                arguments: "{\"city\":\"Boston\"}".into(),
1341                sequence_number: None,
1342            },
1343            ResponsesStreamEvent::Incomplete {
1344                response: ResponseSummary {
1345                    status_details: Some(ResponseStatusDetails {
1346                        reason: Some("max_output_tokens".into()),
1347                        r#type: Some("incomplete".into()),
1348                        error: None,
1349                    }),
1350                    output: vec![response_item_function_call(
1351                        "item_fn",
1352                        Some("{\"city\":\"Boston\"}"),
1353                    )],
1354                    ..Default::default()
1355                },
1356            },
1357        ];
1358
1359        let mapped = map_response_events(events);
1360        assert_eq!(mapped.len(), 2);
1361        assert!(
1362            matches!(mapped[0], LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse { ref raw_input, .. }) if raw_input == "{\"city\":\"Boston\"}")
1363        );
1364        assert!(matches!(
1365            mapped[1],
1366            LanguageModelCompletionEvent::Stop(StopReason::MaxTokens)
1367        ));
1368    }
1369
1370    #[test]
1371    fn responses_stream_handles_empty_tool_arguments() {
1372        let events = vec![
1373            ResponsesStreamEvent::OutputItemAdded {
1374                output_index: 0,
1375                sequence_number: None,
1376                item: response_item_function_call("item_fn", Some("")),
1377            },
1378            ResponsesStreamEvent::FunctionCallArgumentsDone {
1379                item_id: "item_fn".into(),
1380                output_index: 0,
1381                arguments: "".into(),
1382                sequence_number: None,
1383            },
1384            ResponsesStreamEvent::Completed {
1385                response: ResponseSummary::default(),
1386            },
1387        ];
1388
1389        let mapped = map_response_events(events);
1390        assert_eq!(mapped.len(), 2);
1391        assert!(matches!(
1392            &mapped[0],
1393            LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1394                id, name, raw_input, input, ..
1395            }) if id.to_string() == "call_123"
1396                && name.as_ref() == "get_weather"
1397                && raw_input == ""
1398                && input.is_object()
1399                && input.as_object().unwrap().is_empty()
1400        ));
1401        assert!(matches!(
1402            mapped[1],
1403            LanguageModelCompletionEvent::Stop(StopReason::ToolUse)
1404        ));
1405    }
1406
1407    #[test]
1408    fn responses_stream_emits_partial_tool_use_events() {
1409        let events = vec![
1410            ResponsesStreamEvent::OutputItemAdded {
1411                output_index: 0,
1412                sequence_number: None,
1413                item: ResponseOutputItem::FunctionCall(
1414                    crate::responses::ResponseFunctionToolCall {
1415                        id: Some("item_fn".to_string()),
1416                        status: Some("in_progress".to_string()),
1417                        name: Some("get_weather".to_string()),
1418                        call_id: Some("call_abc".to_string()),
1419                        arguments: String::new(),
1420                    },
1421                ),
1422            },
1423            ResponsesStreamEvent::FunctionCallArgumentsDelta {
1424                item_id: "item_fn".into(),
1425                output_index: 0,
1426                delta: "{\"city\":\"Bos".into(),
1427                sequence_number: None,
1428            },
1429            ResponsesStreamEvent::FunctionCallArgumentsDelta {
1430                item_id: "item_fn".into(),
1431                output_index: 0,
1432                delta: "ton\"}".into(),
1433                sequence_number: None,
1434            },
1435            ResponsesStreamEvent::FunctionCallArgumentsDone {
1436                item_id: "item_fn".into(),
1437                output_index: 0,
1438                arguments: "{\"city\":\"Boston\"}".into(),
1439                sequence_number: None,
1440            },
1441            ResponsesStreamEvent::Completed {
1442                response: ResponseSummary::default(),
1443            },
1444        ];
1445
1446        let mapped = map_response_events(events);
1447        assert!(mapped.len() >= 3);
1448
1449        let complete_tool_use = mapped.iter().find(|e| {
1450            matches!(
1451                e,
1452                LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1453                    is_input_complete: true,
1454                    ..
1455                })
1456            )
1457        });
1458        assert!(
1459            complete_tool_use.is_some(),
1460            "should have a complete tool use event"
1461        );
1462
1463        let tool_uses: Vec<_> = mapped
1464            .iter()
1465            .filter(|e| matches!(e, LanguageModelCompletionEvent::ToolUse(_)))
1466            .collect();
1467        assert!(
1468            tool_uses.len() >= 2,
1469            "should have at least one partial and one complete event"
1470        );
1471        assert!(matches!(
1472            tool_uses.last().unwrap(),
1473            LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
1474                is_input_complete: true,
1475                ..
1476            })
1477        ));
1478    }
1479
1480    #[test]
1481    fn responses_stream_maps_reasoning_summary_deltas() {
1482        let events = vec![
1483            ResponsesStreamEvent::OutputItemAdded {
1484                output_index: 0,
1485                sequence_number: None,
1486                item: ResponseOutputItem::Reasoning(ResponseReasoningItem {
1487                    id: Some("rs_123".into()),
1488                    summary: vec![],
1489                }),
1490            },
1491            ResponsesStreamEvent::ReasoningSummaryPartAdded {
1492                item_id: "rs_123".into(),
1493                output_index: 0,
1494                summary_index: 0,
1495            },
1496            ResponsesStreamEvent::ReasoningSummaryTextDelta {
1497                item_id: "rs_123".into(),
1498                output_index: 0,
1499                delta: "Thinking about".into(),
1500            },
1501            ResponsesStreamEvent::ReasoningSummaryTextDelta {
1502                item_id: "rs_123".into(),
1503                output_index: 0,
1504                delta: " the answer".into(),
1505            },
1506            ResponsesStreamEvent::ReasoningSummaryTextDone {
1507                item_id: "rs_123".into(),
1508                output_index: 0,
1509                text: "Thinking about the answer".into(),
1510            },
1511            ResponsesStreamEvent::ReasoningSummaryPartDone {
1512                item_id: "rs_123".into(),
1513                output_index: 0,
1514                summary_index: 0,
1515            },
1516            ResponsesStreamEvent::ReasoningSummaryPartAdded {
1517                item_id: "rs_123".into(),
1518                output_index: 0,
1519                summary_index: 1,
1520            },
1521            ResponsesStreamEvent::ReasoningSummaryTextDelta {
1522                item_id: "rs_123".into(),
1523                output_index: 0,
1524                delta: "Second part".into(),
1525            },
1526            ResponsesStreamEvent::ReasoningSummaryTextDone {
1527                item_id: "rs_123".into(),
1528                output_index: 0,
1529                text: "Second part".into(),
1530            },
1531            ResponsesStreamEvent::ReasoningSummaryPartDone {
1532                item_id: "rs_123".into(),
1533                output_index: 0,
1534                summary_index: 1,
1535            },
1536            ResponsesStreamEvent::OutputItemDone {
1537                output_index: 0,
1538                sequence_number: None,
1539                item: ResponseOutputItem::Reasoning(ResponseReasoningItem {
1540                    id: Some("rs_123".into()),
1541                    summary: vec![
1542                        ReasoningSummaryPart::SummaryText {
1543                            text: "Thinking about the answer".into(),
1544                        },
1545                        ReasoningSummaryPart::SummaryText {
1546                            text: "Second part".into(),
1547                        },
1548                    ],
1549                }),
1550            },
1551            ResponsesStreamEvent::OutputItemAdded {
1552                output_index: 1,
1553                sequence_number: None,
1554                item: response_item_message("msg_456"),
1555            },
1556            ResponsesStreamEvent::OutputTextDelta {
1557                item_id: "msg_456".into(),
1558                output_index: 1,
1559                content_index: Some(0),
1560                delta: "The answer is 42".into(),
1561            },
1562            ResponsesStreamEvent::Completed {
1563                response: ResponseSummary::default(),
1564            },
1565        ];
1566
1567        let mapped = map_response_events(events);
1568
1569        let thinking_events: Vec<_> = mapped
1570            .iter()
1571            .filter(|e| matches!(e, LanguageModelCompletionEvent::Thinking { .. }))
1572            .collect();
1573        assert_eq!(
1574            thinking_events.len(),
1575            4,
1576            "expected 4 thinking events, got {:?}",
1577            thinking_events
1578        );
1579        assert!(
1580            matches!(&thinking_events[0], LanguageModelCompletionEvent::Thinking { text, .. } if text == "Thinking about")
1581        );
1582        assert!(
1583            matches!(&thinking_events[1], LanguageModelCompletionEvent::Thinking { text, .. } if text == " the answer")
1584        );
1585        assert!(
1586            matches!(&thinking_events[2], LanguageModelCompletionEvent::Thinking { text, .. } if text == "\n\n"),
1587            "expected separator between summary parts"
1588        );
1589        assert!(
1590            matches!(&thinking_events[3], LanguageModelCompletionEvent::Thinking { text, .. } if text == "Second part")
1591        );
1592
1593        assert!(mapped.iter().any(
1594            |e| matches!(e, LanguageModelCompletionEvent::Text(t) if t == "The answer is 42")
1595        ));
1596    }
1597
1598    #[test]
1599    fn responses_stream_maps_reasoning_from_done_only() {
1600        let events = vec![
1601            ResponsesStreamEvent::OutputItemAdded {
1602                output_index: 0,
1603                sequence_number: None,
1604                item: ResponseOutputItem::Reasoning(ResponseReasoningItem {
1605                    id: Some("rs_789".into()),
1606                    summary: vec![],
1607                }),
1608            },
1609            ResponsesStreamEvent::OutputItemDone {
1610                output_index: 0,
1611                sequence_number: None,
1612                item: ResponseOutputItem::Reasoning(ResponseReasoningItem {
1613                    id: Some("rs_789".into()),
1614                    summary: vec![ReasoningSummaryPart::SummaryText {
1615                        text: "Summary without deltas".into(),
1616                    }],
1617                }),
1618            },
1619            ResponsesStreamEvent::Completed {
1620                response: ResponseSummary::default(),
1621            },
1622        ];
1623
1624        let mapped = map_response_events(events);
1625        assert!(
1626            !mapped
1627                .iter()
1628                .any(|e| matches!(e, LanguageModelCompletionEvent::Thinking { .. })),
1629            "OutputItemDone reasoning should not produce Thinking events"
1630        );
1631    }
1632
1633    #[test]
1634    fn into_open_ai_interleaved_reasoning() {
1635        let tool_use_id = LanguageModelToolUseId::from("call-1");
1636        let tool_input = json!({"query": "foo"});
1637        let tool_arguments = serde_json::to_string(&tool_input).unwrap();
1638        let tool_use = LanguageModelToolUse {
1639            id: tool_use_id.clone(),
1640            name: Arc::from("search"),
1641            raw_input: tool_arguments.clone(),
1642            input: tool_input,
1643            is_input_complete: true,
1644            thought_signature: None,
1645        };
1646        let tool_result = LanguageModelToolResult {
1647            tool_use_id: tool_use_id,
1648            tool_name: Arc::from("search"),
1649            is_error: false,
1650            content: vec![LanguageModelToolResultContent::Text(Arc::from("result"))],
1651            output: None,
1652        };
1653        let request = LanguageModelRequest {
1654            thread_id: None,
1655            prompt_id: None,
1656            intent: None,
1657            messages: vec![
1658                LanguageModelRequestMessage {
1659                    role: Role::User,
1660                    content: vec![MessageContent::Text("search for something".into())],
1661                    cache: false,
1662                    reasoning_details: None,
1663                },
1664                LanguageModelRequestMessage {
1665                    role: Role::Assistant,
1666                    content: vec![
1667                        MessageContent::Thinking {
1668                            text: "I should search".into(),
1669                            signature: None,
1670                        },
1671                        MessageContent::Text("Searching now.".into()),
1672                        MessageContent::ToolUse(tool_use),
1673                    ],
1674                    cache: false,
1675                    reasoning_details: None,
1676                },
1677                LanguageModelRequestMessage {
1678                    role: Role::Assistant,
1679                    content: vec![MessageContent::ToolResult(tool_result)],
1680                    cache: false,
1681                    reasoning_details: None,
1682                },
1683            ],
1684            tools: vec![],
1685            tool_choice: None,
1686            stop: vec![],
1687            temperature: None,
1688            thinking_allowed: true,
1689            thinking_effort: None,
1690            speed: None,
1691        };
1692
1693        let result = into_open_ai(request.clone(), "model", false, false, None, None, true);
1694        assert_eq!(
1695            serde_json::to_value(&result).unwrap()["messages"],
1696            json!([
1697                {"role": "user", "content": "search for something"},
1698                {
1699                    "role": "assistant",
1700                    "content": "Searching now.",
1701                    "tool_calls": [{"id": "call-1", "type": "function", "function": {"name": "search", "arguments": tool_arguments}}],
1702                    "reasoning_content": "I should search"
1703                },
1704                {"role": "tool", "content": "result", "tool_call_id": "call-1"}
1705            ])
1706        );
1707
1708        let result = into_open_ai(request, "model", false, false, None, None, false);
1709        assert_eq!(
1710            serde_json::to_value(&result).unwrap()["messages"],
1711            json!([
1712                {"role": "user", "content": "search for something"},
1713                {
1714                    "role": "assistant",
1715                    "content": [
1716                        {"type": "text", "text": "I should search"},
1717                        {"type": "text", "text": "Searching now."}
1718                    ],
1719                    "tool_calls": [{"id": "call-1", "type": "function", "function": {"name": "search", "arguments": tool_arguments}}]
1720                },
1721                {"role": "tool", "content": "result", "tool_call_id": "call-1"}
1722            ])
1723        );
1724    }
1725}