streaming.rs 24.7 KB
Newer Older
1
2
3
4
//! Streaming infrastructure for /v1/responses endpoint

use std::collections::HashMap;

5
use axum::{body::Body, http::StatusCode, response::Response};
6
7
8
use bytes::Bytes;
use serde_json::json;
use tokio::sync::mpsc;
9
use tokio_stream::wrappers::UnboundedReceiverStream;
10
11
use uuid::Uuid;

12
13
14
15
use crate::{
    mcp,
    protocols::{chat::ChatCompletionStreamResponse, responses::ResponsesRequest},
};
16

17
pub enum OutputItemType {
18
19
20
    Message,
    McpListTools,
    McpCall,
21
    FunctionCall,
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
    Reasoning,
}

/// Status of an output item
#[derive(Debug, Clone, PartialEq)]
enum ItemStatus {
    InProgress,
    Completed,
}

/// State tracking for a single output item
#[derive(Debug, Clone)]
struct OutputItemState {
    output_index: usize,
    status: ItemStatus,
37
    item_data: Option<serde_json::Value>,
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
}

/// OpenAI-compatible event emitter for /v1/responses streaming
///
/// Manages state and sequence numbers to emit proper event types:
/// - response.created
/// - response.in_progress
/// - response.output_item.added
/// - response.content_part.added
/// - response.output_text.delta (multiple)
/// - response.output_text.done
/// - response.content_part.done
/// - response.output_item.done
/// - response.completed
/// - response.mcp_list_tools.in_progress
/// - response.mcp_list_tools.completed
/// - response.mcp_call.in_progress
/// - response.mcp_call_arguments.delta
/// - response.mcp_call_arguments.done
/// - response.mcp_call.completed
/// - response.mcp_call.failed
59
pub struct ResponseStreamEventEmitter {
60
    sequence_number: u64,
61
    pub response_id: String,
62
63
64
65
66
67
68
69
70
71
    model: String,
    created_at: u64,
    message_id: String,
    accumulated_text: String,
    has_emitted_created: bool,
    has_emitted_in_progress: bool,
    has_emitted_output_item_added: bool,
    has_emitted_content_part_added: bool,
    // MCP call tracking
    mcp_call_accumulated_args: HashMap<String, String>,
72
    // Output item tracking
73
74
75
76
    output_items: Vec<OutputItemState>,
    next_output_index: usize,
    current_message_output_index: Option<usize>, // Tracks output_index of current message
    current_item_id: Option<String>,             // Tracks item_id of current item
77
    original_request: Option<ResponsesRequest>,
78
79
80
}

impl ResponseStreamEventEmitter {
81
    pub fn new(response_id: String, model: String, created_at: u64) -> Self {
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
        let message_id = format!("msg_{}", Uuid::new_v4());

        Self {
            sequence_number: 0,
            response_id,
            model,
            created_at,
            message_id,
            accumulated_text: String::new(),
            has_emitted_created: false,
            has_emitted_in_progress: false,
            has_emitted_output_item_added: false,
            has_emitted_content_part_added: false,
            mcp_call_accumulated_args: HashMap::new(),
            output_items: Vec::new(),
            next_output_index: 0,
            current_message_output_index: None,
            current_item_id: None,
100
            original_request: None,
101
102
103
        }
    }

104
105
106
107
108
    /// Set the original request for including all fields in response.completed
    pub fn set_original_request(&mut self, request: ResponsesRequest) {
        self.original_request = Some(request);
    }

109
110
111
112
113
114
    fn next_sequence(&mut self) -> u64 {
        let seq = self.sequence_number;
        self.sequence_number += 1;
        seq
    }

115
    pub fn emit_created(&mut self) -> serde_json::Value {
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
        self.has_emitted_created = true;
        json!({
            "type": "response.created",
            "sequence_number": self.next_sequence(),
            "response": {
                "id": self.response_id,
                "object": "response",
                "created_at": self.created_at,
                "status": "in_progress",
                "model": self.model,
                "output": []
            }
        })
    }

131
    pub fn emit_in_progress(&mut self) -> serde_json::Value {
132
133
134
135
136
137
138
139
140
141
142
143
        self.has_emitted_in_progress = true;
        json!({
            "type": "response.in_progress",
            "sequence_number": self.next_sequence(),
            "response": {
                "id": self.response_id,
                "object": "response",
                "status": "in_progress"
            }
        })
    }

144
    pub fn emit_content_part_added(
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
        &mut self,
        output_index: usize,
        item_id: &str,
        content_index: usize,
    ) -> serde_json::Value {
        self.has_emitted_content_part_added = true;
        json!({
            "type": "response.content_part.added",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "content_index": content_index,
            "part": {
                "type": "text",
                "text": ""
            }
        })
    }

164
    pub fn emit_text_delta(
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
        &mut self,
        delta: &str,
        output_index: usize,
        item_id: &str,
        content_index: usize,
    ) -> serde_json::Value {
        self.accumulated_text.push_str(delta);
        json!({
            "type": "response.output_text.delta",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "content_index": content_index,
            "delta": delta
        })
    }

182
    pub fn emit_text_done(
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
        &mut self,
        output_index: usize,
        item_id: &str,
        content_index: usize,
    ) -> serde_json::Value {
        json!({
            "type": "response.output_text.done",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "content_index": content_index,
            "text": self.accumulated_text.clone()
        })
    }

198
    pub fn emit_content_part_done(
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
        &mut self,
        output_index: usize,
        item_id: &str,
        content_index: usize,
    ) -> serde_json::Value {
        json!({
            "type": "response.content_part.done",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "content_index": content_index,
            "part": {
                "type": "text",
                "text": self.accumulated_text.clone()
            }
        })
    }

217
    pub fn emit_completed(&mut self, usage: Option<&serde_json::Value>) -> serde_json::Value {
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
        // Build output array from tracked items
        let output: Vec<serde_json::Value> = self
            .output_items
            .iter()
            .filter_map(|item| {
                if item.status == ItemStatus::Completed {
                    item.item_data.clone()
                } else {
                    None
                }
            })
            .collect();

        // If no items were tracked (legacy path), fall back to generic message
        let output = if output.is_empty() {
            vec![json!({
                "id": self.message_id.clone(),
                "type": "message",
                "role": "assistant",
                "content": [{
                    "type": "text",
                    "text": self.accumulated_text.clone()
240
                }]
241
242
243
244
245
246
247
248
249
250
251
252
253
            })]
        } else {
            output
        };

        // Build base response object
        let mut response_obj = json!({
            "id": self.response_id,
            "object": "response",
            "created_at": self.created_at,
            "status": "completed",
            "model": self.model,
            "output": output
254
255
        });

256
        // Add usage if provided
257
        if let Some(usage_val) = usage {
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
            response_obj["usage"] = usage_val.clone();
        }

        // Add all original request fields if available
        if let Some(ref req) = self.original_request {
            Self::add_optional_field(&mut response_obj, "instructions", &req.instructions);
            Self::add_optional_field(
                &mut response_obj,
                "max_output_tokens",
                &req.max_output_tokens,
            );
            Self::add_optional_field(&mut response_obj, "max_tool_calls", &req.max_tool_calls);
            Self::add_optional_field(
                &mut response_obj,
                "previous_response_id",
                &req.previous_response_id,
            );
            Self::add_optional_field(&mut response_obj, "reasoning", &req.reasoning);
            Self::add_optional_field(&mut response_obj, "temperature", &req.temperature);
            Self::add_optional_field(&mut response_obj, "top_p", &req.top_p);
            Self::add_optional_field(&mut response_obj, "truncation", &req.truncation);
            Self::add_optional_field(&mut response_obj, "user", &req.user);

            response_obj["parallel_tool_calls"] = json!(req.parallel_tool_calls.unwrap_or(true));
            response_obj["store"] = json!(req.store.unwrap_or(true));
            response_obj["tools"] = json!(req.tools.as_ref().unwrap_or(&vec![]));
            response_obj["metadata"] = json!(req.metadata.as_ref().unwrap_or(&Default::default()));

            // tool_choice: serialize if present, otherwise use "auto"
            if let Some(ref tc) = req.tool_choice {
                response_obj["tool_choice"] = json!(tc);
            } else {
                response_obj["tool_choice"] = json!("auto");
            }
292
293
        }

294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
        json!({
            "type": "response.completed",
            "sequence_number": self.next_sequence(),
            "response": response_obj
        })
    }

    /// Helper to add optional fields to JSON object
    fn add_optional_field<T: serde::Serialize>(
        obj: &mut serde_json::Value,
        key: &str,
        value: &Option<T>,
    ) {
        if let Some(val) = value {
            obj[key] = json!(val);
        }
310
311
312
313
314
315
    }

    // ========================================================================
    // MCP Event Emission Methods
    // ========================================================================

316
    pub fn emit_mcp_list_tools_in_progress(&mut self, output_index: usize) -> serde_json::Value {
317
318
319
320
321
322
323
        json!({
            "type": "response.mcp_list_tools.in_progress",
            "sequence_number": self.next_sequence(),
            "output_index": output_index
        })
    }

324
    pub fn emit_mcp_list_tools_completed(
325
326
        &mut self,
        output_index: usize,
327
        tools: &[mcp::Tool],
328
329
330
331
332
    ) -> serde_json::Value {
        let tool_items: Vec<_> = tools
            .iter()
            .map(|t| {
                json!({
333
334
335
                    "name": &t.name,
                    "description": &t.description,
                    "input_schema": t.input_schema.clone()
336
337
338
339
340
341
342
343
344
345
346
347
                })
            })
            .collect();

        json!({
            "type": "response.mcp_list_tools.completed",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "tools": tool_items
        })
    }

348
    pub fn emit_mcp_call_in_progress(
349
350
351
352
353
354
355
356
357
358
359
360
        &mut self,
        output_index: usize,
        item_id: &str,
    ) -> serde_json::Value {
        json!({
            "type": "response.mcp_call.in_progress",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id
        })
    }

361
    pub fn emit_mcp_call_arguments_delta(
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
        &mut self,
        output_index: usize,
        item_id: &str,
        delta: &str,
    ) -> serde_json::Value {
        // Accumulate arguments for this call
        self.mcp_call_accumulated_args
            .entry(item_id.to_string())
            .or_default()
            .push_str(delta);

        json!({
            "type": "response.mcp_call_arguments.delta",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "delta": delta
        })
    }

382
    pub fn emit_mcp_call_arguments_done(
383
384
385
386
387
388
389
390
391
392
393
394
395
396
        &mut self,
        output_index: usize,
        item_id: &str,
        arguments: &str,
    ) -> serde_json::Value {
        json!({
            "type": "response.mcp_call_arguments.done",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "arguments": arguments
        })
    }

397
    pub fn emit_mcp_call_completed(
398
399
400
401
402
403
404
405
406
407
408
409
        &mut self,
        output_index: usize,
        item_id: &str,
    ) -> serde_json::Value {
        json!({
            "type": "response.mcp_call.completed",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id
        })
    }

410
    pub fn emit_mcp_call_failed(
411
412
413
414
415
416
417
418
419
420
421
422
423
424
        &mut self,
        output_index: usize,
        item_id: &str,
        error: &str,
    ) -> serde_json::Value {
        json!({
            "type": "response.mcp_call.failed",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "error": error
        })
    }

425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
    // ========================================================================
    // Function Call Event Emission Methods
    // ========================================================================

    pub fn emit_function_call_arguments_delta(
        &mut self,
        output_index: usize,
        item_id: &str,
        delta: &str,
    ) -> serde_json::Value {
        json!({
            "type": "response.function_call_arguments.delta",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "delta": delta
        })
    }

    pub fn emit_function_call_arguments_done(
        &mut self,
        output_index: usize,
        item_id: &str,
        arguments: &str,
    ) -> serde_json::Value {
        json!({
            "type": "response.function_call_arguments.done",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item_id": item_id,
            "arguments": arguments
        })
    }

459
460
461
462
463
    // ========================================================================
    // Output Item Wrapper Events
    // ========================================================================

    /// Emit response.output_item.added event
464
    pub fn emit_output_item_added(
465
466
467
468
469
470
471
472
473
474
475
476
477
        &mut self,
        output_index: usize,
        item: &serde_json::Value,
    ) -> serde_json::Value {
        json!({
            "type": "response.output_item.added",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item": item
        })
    }

    /// Emit response.output_item.done event
478
    pub fn emit_output_item_done(
479
480
481
482
        &mut self,
        output_index: usize,
        item: &serde_json::Value,
    ) -> serde_json::Value {
483
484
485
        // Store the item data for later use in emit_completed
        self.store_output_item_data(output_index, item.clone());

486
487
488
489
490
491
492
493
494
495
496
497
498
499
        json!({
            "type": "response.output_item.done",
            "sequence_number": self.next_sequence(),
            "output_index": output_index,
            "item": item
        })
    }

    /// Generate unique ID for item type
    fn generate_item_id(prefix: &str) -> String {
        format!("{}_{}", prefix, Uuid::new_v4().to_string().replace("-", ""))
    }

    /// Allocate next output index and track item
500
    pub fn allocate_output_index(&mut self, item_type: OutputItemType) -> (usize, String) {
501
502
503
504
505
506
        let index = self.next_output_index;
        self.next_output_index += 1;

        let id_prefix = match &item_type {
            OutputItemType::McpListTools => "mcpl",
            OutputItemType::McpCall => "mcp",
507
            OutputItemType::FunctionCall => "fc",
508
509
510
511
512
513
514
515
516
            OutputItemType::Message => "msg",
            OutputItemType::Reasoning => "rs",
        };

        let id = Self::generate_item_id(id_prefix);

        self.output_items.push(OutputItemState {
            output_index: index,
            status: ItemStatus::InProgress,
517
            item_data: None,
518
519
520
521
522
        });

        (index, id)
    }

523
    /// Mark output item as completed and store its data
524
    pub fn complete_output_item(&mut self, output_index: usize) {
525
526
527
528
529
530
531
532
533
        if let Some(item) = self
            .output_items
            .iter_mut()
            .find(|i| i.output_index == output_index)
        {
            item.status = ItemStatus::Completed;
        }
    }

534
535
536
537
538
539
540
541
542
543
544
    /// Store output item data when emitting output_item.done
    pub fn store_output_item_data(&mut self, output_index: usize, item_data: serde_json::Value) {
        if let Some(item) = self
            .output_items
            .iter_mut()
            .find(|i| i.output_index == output_index)
        {
            item.item_data = Some(item_data);
        }
    }

545
546
547
548
    /// Emit reasoning item wrapper events (added + done)
    ///
    /// Reasoning items in OpenAI format are simple placeholders emitted between tool iterations.
    /// They don't have streaming content - just wrapper events with empty/null content.
549
    pub fn emit_reasoning_item(
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
        &mut self,
        tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
        reasoning_content: Option<String>,
    ) -> Result<(), String> {
        // Allocate output index and generate ID
        let (output_index, item_id) = self.allocate_output_index(OutputItemType::Reasoning);

        // Build reasoning item structure
        let item = json!({
            "id": item_id,
            "type": "reasoning",
            "summary": [],
            "content": reasoning_content,
            "encrypted_content": null,
            "status": null
        });

        // Emit output_item.added
        let added_event = self.emit_output_item_added(output_index, &item);
        self.send_event(&added_event, tx)?;

        // Immediately emit output_item.done (no streaming for reasoning)
        let done_event = self.emit_output_item_done(output_index, &item);
        self.send_event(&done_event, tx)?;

        // Mark as completed
        self.complete_output_item(output_index);

        Ok(())
    }

    /// Process a chunk and emit appropriate events
582
    pub fn process_chunk(
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
        &mut self,
        chunk: &ChatCompletionStreamResponse,
        tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
    ) -> Result<(), String> {
        // Process content if present
        if let Some(choice) = chunk.choices.first() {
            if let Some(content) = &choice.delta.content {
                if !content.is_empty() {
                    // Allocate output_index and item_id for this message item (once per message)
                    if self.current_item_id.is_none() {
                        let (output_index, item_id) =
                            self.allocate_output_index(OutputItemType::Message);

                        // Build message item structure
                        let item = json!({
                            "id": item_id,
                            "type": "message",
                            "role": "assistant",
                            "content": []
                        });

                        // Emit output_item.added
                        let event = self.emit_output_item_added(output_index, &item);
                        self.send_event(&event, tx)?;
                        self.has_emitted_output_item_added = true;

                        // Store for subsequent events
                        self.current_item_id = Some(item_id);
                        self.current_message_output_index = Some(output_index);
                    }

                    let output_index = self.current_message_output_index.unwrap();
                    let item_id = self.current_item_id.clone().unwrap(); // Clone to avoid borrow checker issues
                    let content_index = 0; // Single content part for now

                    // Emit content_part.added before first delta
                    if !self.has_emitted_content_part_added {
                        let event =
                            self.emit_content_part_added(output_index, &item_id, content_index);
                        self.send_event(&event, tx)?;
                        self.has_emitted_content_part_added = true;
                    }

                    // Emit text delta
                    let event =
                        self.emit_text_delta(content, output_index, &item_id, content_index);
                    self.send_event(&event, tx)?;
                }
            }

            // Check for finish_reason to emit completion events
            if let Some(reason) = &choice.finish_reason {
                if reason == "stop" || reason == "length" {
                    let output_index = self.current_message_output_index.unwrap();
                    let item_id = self.current_item_id.clone().unwrap(); // Clone to avoid borrow checker issues
                    let content_index = 0;

                    // Emit closing events
                    if self.has_emitted_content_part_added {
                        let event = self.emit_text_done(output_index, &item_id, content_index);
                        self.send_event(&event, tx)?;
                        let event =
                            self.emit_content_part_done(output_index, &item_id, content_index);
                        self.send_event(&event, tx)?;
                    }

                    if self.has_emitted_output_item_added {
                        // Build complete message item for output_item.done
                        let item = json!({
                            "id": item_id,
                            "type": "message",
                            "role": "assistant",
                            "content": [{
                                "type": "text",
                                "text": self.accumulated_text.clone()
                            }]
                        });
                        let event = self.emit_output_item_done(output_index, &item);
                        self.send_event(&event, tx)?;
                    }

                    // Mark item as completed
                    self.complete_output_item(output_index);
                }
            }
        }

        Ok(())
    }

673
    pub fn send_event(
674
675
676
677
678
679
680
        &self,
        event: &serde_json::Value,
        tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
    ) -> Result<(), String> {
        let event_json = serde_json::to_string(event)
            .map_err(|e| format!("Failed to serialize event: {}", e))?;

681
682
683
684
685
686
687
688
689
690
        // Extract event type from the JSON for SSE event field
        let event_type = event
            .get("type")
            .and_then(|v| v.as_str())
            .unwrap_or("message");

        // Format as SSE with event: field
        let sse_message = format!("event: {}\ndata: {}\n\n", event_type, event_json);

        if tx.send(Ok(Bytes::from(sse_message))).is_err() {
691
692
693
694
695
            return Err("Client disconnected".to_string());
        }

        Ok(())
    }
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714

    /// Send event and log any errors (typically client disconnect)
    ///
    /// This is a convenience method for streaming scenarios where client
    /// disconnection is expected and should be logged but not fail the operation.
    /// Returns true if sent successfully, false if client disconnected.
    pub fn send_event_best_effort(
        &self,
        event: &serde_json::Value,
        tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
    ) -> bool {
        match self.send_event(event, tx) {
            Ok(()) => true,
            Err(e) => {
                tracing::debug!("Failed to send event (likely client disconnect): {}", e);
                false
            }
        }
    }
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750

    /// Emit an error event
    ///
    /// Creates and sends an error event with the given error message.
    /// Uses OpenAI's error event format.
    /// Use this for terminal errors that should abort the streaming response.
    pub fn emit_error(
        &mut self,
        error_msg: &str,
        error_code: Option<&str>,
        tx: &mpsc::UnboundedSender<Result<Bytes, std::io::Error>>,
    ) {
        let event = json!({
            "type": "error",
            "code": error_code.unwrap_or("internal_error"),
            "message": error_msg,
            "param": null,
            "sequence_number": self.next_sequence()
        });
        let sse_data = format!("data: {}\n\n", serde_json::to_string(&event).unwrap());
        let _ = tx.send(Ok(Bytes::from(sse_data)));
    }
}

/// Build a Server-Sent Events (SSE) response
///
/// Creates a Response with proper SSE headers and streaming body.
pub fn build_sse_response(rx: mpsc::UnboundedReceiver<Result<Bytes, std::io::Error>>) -> Response {
    let stream = UnboundedReceiverStream::new(rx);
    Response::builder()
        .status(StatusCode::OK)
        .header("Content-Type", "text/event-stream")
        .header("Cache-Control", "no-cache")
        .header("Connection", "keep-alive")
        .body(Body::from_stream(stream))
        .unwrap()
751
}