File size: 32,810 Bytes
e48a2d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
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
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
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
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
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
582
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
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
package server

import (
	"context"
	"encoding/json"
	"io"
	"net/http"
	"net/http/httptest"
	"strings"
	"testing"

	"github.com/misakano7545/workbuddy2api-panel/internal/auth"
	"github.com/misakano7545/workbuddy2api-panel/internal/upstream"
)

// fakeContinueUpstream 按调用次数返回脚本化 SSE 并记录每次实际发出的请求体。
// 注意计数口径:单测里「段1」是手工喂给 reader 的(不经 fake),所以 fake 的
// call=1 就是第一次续写请求——script 按「第几次续写」编写。
func fakeContinueUpstream(t *testing.T, script func(call int) (int, string)) (*upstream.Client, *[]string) {
	t.Helper()
	reqs := &[]string{}
	call := 0
	up := &upstream.Client{
		HTTP: &http.Client{Transport: roundTripFunc(func(r *http.Request) (*http.Response, error) {
			call++
			raw, _ := io.ReadAll(r.Body)
			*reqs = append(*reqs, string(raw))
			status, body := script(call)
			return &http.Response{
				StatusCode: status,
				Header:     http.Header{"Content-Type": []string{"text/event-stream"}},
				Body:       io.NopCloser(strings.NewReader(body)),
			}, nil
		})},
		ChatBaseCN:    "https://fake.example",
		BillingBaseCN: "https://fake.example",
	}
	return up, reqs
}

// 段1:正文截断(finish_reason=length),带 usage。
const testSeg1Text = `data: {"id":"cmpl-1","choices":[{"index":0,"delta":{"role":"assistant","content":"你好"}}]}` + "\n\n" +
	`data: {"id":"cmpl-1","choices":[{"index":0,"delta":{"content":"世界"}}]}` + "\n\n" +
	`data: {"id":"cmpl-1","choices":[{"index":0,"delta":{},"finish_reason":"length"}],"usage":{"prompt_tokens":10,"completion_tokens":5,"total_tokens":15}}` + "\n\n" +
	"data: [DONE]\n\n"

// 续写段:自然收尾(finish_reason=stop),带 usage。
const testSeg2Text = `data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"content":"继续"}}]}` + "\n\n" +
	`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{},"finish_reason":"stop"}],"usage":{"prompt_tokens":20,"completion_tokens":7,"total_tokens":27}}` + "\n\n" +
	"data: [DONE]\n\n"

func newTestContinueReader(t *testing.T, up *upstream.Client, body string, rc io.ReadCloser) *continueReader {
	t.Helper()
	acct := &auth.Auth{UID: "u1", AccessToken: "at-1", ExpiresAt: 9999999999}
	return newContinueReader(context.Background(), up, acct, []byte(body), "1.2.3.4", upstream.ChatMeta{}, rc)
}

// TestContinueReaderTextCutContinues 正文截断 → 同模型续写;客户端可见流是一条
// 连续输出:两段正文都在、终态只留 stop、length 不外泄、usage 为跨段合计。
func TestContinueReaderTextCutContinues(t *testing.T) {
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r := newTestContinueReader(t, up, `{"model":"glm-5.2","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg1Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if len(*reqs) != 1 {
		t.Fatalf("续写请求数=%d want 1", len(*reqs))
	}
	for _, want := range []string{"你好", "世界", "继续", `"finish_reason":"stop"`} {
		if !strings.Contains(s, want) {
			t.Fatalf("输出缺少 %q:\n%s", want, s)
		}
	}
	if strings.Contains(s, `"finish_reason":"length"`) {
		t.Fatalf("length 终态不应外泄:\n%s", s)
	}
	// usage:恰好一条,prompt 取末段(20),completion 跨段累计(5+7=12),
	// total 按 prompt+completion 重算(32)——跨段 prompt 求和会把续写段
	// 携带的整个上下文重复计入,客户端按 usage 判上下文压力会提前压缩。
	var usageFrames int
	var usage map[string]any
	for _, f := range strings.Split(s, "\n\n") {
		line, ok := strings.CutPrefix(f, "data: ")
		if !ok || line == "" || line == "[DONE]" {
			continue
		}
		var obj map[string]any
		if json.Unmarshal([]byte(line), &obj) != nil {
			continue
		}
		if u, ok := obj["usage"].(map[string]any); ok {
			usageFrames++
			usage = u
		}
	}
	if usageFrames != 1 {
		t.Fatalf("usage 帧数=%d want 1:\n%s", usageFrames, s)
	}
	if usage["prompt_tokens"] != float64(20) || usage["completion_tokens"] != float64(12) || usage["total_tokens"] != float64(32) {
		t.Fatalf("usage 合并=%v want prompt=20 completion=12 total=32", usage)
	}
	// 续写请求:同模型 + assistant 携带已输出内容 + 续写指令 + 原始消息仍在。
	req := (*reqs)[0]
	for _, want := range []string{`"model":"glm-5.2"`, "你好世界", "从中断处继续输出剩余内容", `"content":"hi"`} {
		if !strings.Contains(req, want) {
			t.Fatalf("续写请求缺少 %q:\n%s", want, req)
		}
	}
	if !strings.Contains(req, `"role":"assistant"`) {
		t.Fatalf("续写请求缺 assistant 消息:\n%s", req)
	}
}

// TestContinueReaderToolCutNoContinue 工具调用分片在场 → 不续写(零续写请求),
// finish=length 原样外泄(交回既有截断语义)。
func TestContinueReaderToolCutNoContinue(t *testing.T) {
	seg := `data: {"id":"c","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"call_1","function":{"name":"Bash","arguments":"{\"cmd\":"}}]}}]}` + "\n\n" +
		`data: {"id":"c","choices":[{"index":0,"delta":{},"finish_reason":"length"}]}` + "\n\n" +
		"data: [DONE]\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(seg)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if len(*reqs) != 0 {
		t.Fatalf("续写请求数=%d want 0(工具截断不续)", len(*reqs))
	}
	if !strings.Contains(string(out), `"finish_reason":"length"`) {
		t.Fatalf("length 终态应原样回放:\n%s", out)
	}
}

// TestContinueReaderMidToolCallCutNoRetry 断流落在工具调用参数中间 → 不盲重试。
// 与 finish=length 那条同一条理由(本文件开头写明):工具参数已被切,重试体只带
// 正文/推理当 assistant 前缀(continuationBody 不含残缺 tool_calls),模型会重发一遍
// 调用,客户端在同一块里拼接两次参数 → 半截 JSON 拼成垃圾,比"不续"更糟。
// 形态刻意带上正文(模型先说话再调工具是常态),否则零输出守卫会掩盖这条路径。
func TestContinueReaderMidToolCallCutNoRetry(t *testing.T) {
	cut := `data: {"id":"c","choices":[{"index":0,"delta":{"content":"好的,我来执行"}}]}` + "\n\n" +
		`data: {"id":"c","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"call_1","function":{"name":"Bash","arguments":"{\"cmd\":"}}]}}]}` + "\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(cut)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if len(*reqs) != 0 {
		t.Fatalf("工具调用中断流不得盲重试,实际发了 %d 次续写请求", len(*reqs))
	}
	// 客户端仍要知道这一回合是残缺的:补截断终态(恰好一个)。
	if n := strings.Count(string(out), `"finish_reason":"length"`); n != 1 {
		t.Fatalf("截断终态帧应恰好一个,got=%d:\n%s", n, out)
	}
	if !strings.Contains(string(out), `"arguments":"{\"cmd\":"`) {
		t.Fatalf("已发出的工具分片应原样保留:\n%s", out)
	}
}

// TestContinueReaderExplicitLimitNoContinue 客户端显式输出限额 → 截断是它的
// 预期语义,不续写。
func TestContinueReaderExplicitLimitNoContinue(t *testing.T) {
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r := newTestContinueReader(t, up, `{"model":"m","max_tokens":1024,"messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg1Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if len(*reqs) != 0 {
		t.Fatalf("续写请求数=%d want 0(显式限额不续)", len(*reqs))
	}
	if !strings.Contains(string(out), `"finish_reason":"length"`) {
		t.Fatalf("length 终态应原样回放:\n%s", out)
	}
}

// TestContinueReaderMainSegCutBlindRetry 主段断流(有正文、无 finish、无 [DONE])也做
// 一次盲重试:断流不是「请求本身失败」——它已经产出过内容,重试请求体把已收内容作为
// assistant 前缀(幂等不重复),客户端因此看到的是接续而不是半截回合。线上实测这是
// Codex「莫名中断、要手动继续」的主因。
func TestContinueReaderMainSegCutBlindRetry(t *testing.T) {
	cut := `data: {"id":"c1","model":"m","choices":[{"index":0,"delta":{"role":"assistant"}}]}` + "\n\n" +
		`data: {"id":"c1","model":"m","choices":[{"index":0,"delta":{"content":"半截正文"}}]}` + "\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(cut)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if len(*reqs) != 1 {
		t.Fatalf("主段断流应发一次盲重试,got=%d", len(*reqs))
	}
	if !strings.Contains((*reqs)[0], "半截正文") {
		t.Fatalf("重试请求应携带已收内容(幂等不重复):\n%s", (*reqs)[0])
	}
	for _, want := range []string{"半截正文", "继续", `"finish_reason":"stop"`} {
		if !strings.Contains(s, want) {
			t.Fatalf("客户端可见流缺 %q:\n%s", want, s)
		}
	}
	if strings.Contains(s, `"finish_reason":"length"`) {
		t.Fatalf("续上之后 length 终态不应外泄:\n%s", s)
	}
}

// TestContinueReaderMainSegCutRetryFailsEmitsTerminal 盲重试拿不到 200 → 降级补截断
// 终态(恰好一个),客户端仍能判断这个回合是残缺的。
func TestContinueReaderMainSegCutRetryFailsEmitsTerminal(t *testing.T) {
	cut := `data: {"id":"c1","model":"m","choices":[{"index":0,"delta":{"content":"半截正文"}}]}` + "\n\n"
	up, _ := fakeContinueUpstream(t, func(call int) (int, string) { return 500, `{"code":11102,"msg":"boom"}` })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(cut)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if n := strings.Count(s, `"finish_reason":"length"`); n != 1 {
		t.Fatalf("截断终态帧应恰好一个,got=%d:\n%s", n, s)
	}
	if idx := strings.Index(s, `"finish_reason":"length"`); idx < strings.Index(s, "半截正文") {
		t.Fatalf("终态帧必须在正文之后:\n%s", s)
	}
}

// TestContinueReaderCutNoContentNoTerminal 断流但零模型输出(只有 role 帧)→ 不补
// 终态:补了会让 StreamHint 的 validEvents 从 0 变 1,把空流本该收敛的 502 观测
// 伪装成「有内容的 200」。空流判据必须留在上游侧。
func TestContinueReaderCutNoContentNoTerminal(t *testing.T) {
	onlyRole := `data: {"id":"c1","model":"m","choices":[{"index":0,"delta":{"role":"assistant"}}]}` + "\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(onlyRole)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if strings.Contains(string(out), "finish_reason") {
		t.Fatalf("零输出断流不得补终态(会让空流 502 观测失真):\n%s", out)
	}
	// 也没有可续的东西:盲重试的资格要求「已经产出过内容」。
	if len(*reqs) != 0 {
		t.Fatalf("零输出断流不该发盲重试,got=%d", len(*reqs))
	}
}

// TestContinueReaderNormalEndNoDuplicateTerminal 正常收尾(finish=stop + [DONE]):
// 不得多补一条终态(客户端会看到两个 finish_reason)。
func TestContinueReaderNormalEndNoDuplicateTerminal(t *testing.T) {
	up, _ := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r := newTestContinueReader(t, up, `{"model":"m","max_tokens":9,"messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg2Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if n := strings.Count(string(out), "finish_reason"); n != 1 {
		t.Fatalf("终态帧应恰好一个,got=%d:\n%s", n, out)
	}
}

// TestStreamCutTruncatedAcrossAllProtocols 同一份断流,三条适配器都必须如实标截断:
// 补在 continue 这一处接缝(唯一共用层),三条路径各自把 length 映射到自己的词表。
// 这是用户报的 bug 的端到端回归:此前 chat 裸 [DONE]、responses 标 completed、
// messages 标 end_turn——agent 客户端(Codex/CC)把半截回合当正常结束。
func TestStreamCutTruncatedAcrossAllProtocols(t *testing.T) {
	cut := `data: {"id":"c1","object":"chat.completion.chunk","model":"glm-5.2","choices":[{"index":0,"delta":{"role":"assistant"}}]}` + "\n\n" +
		`data: {"id":"c1","object":"chat.completion.chunk","model":"glm-5.2","choices":[{"index":0,"delta":{"content":"半截"}}]}` + "\n\n"
	cases := []struct {
		name, path, body string
		want             []string
		reject           []string
	}{
		{
			"chat", "/v1/chat/completions",
			`{"model":"glm-5.2","stream":true,"messages":[{"role":"user","content":"hi"}]}`,
			[]string{`"finish_reason":"length"`, "data: [DONE]"}, nil,
		},
		{
			"responses", "/v1/responses",
			`{"model":"glm-5.2","stream":true,"input":"hi"}`,
			[]string{`"type":"response.incomplete"`, `"reason":"max_output_tokens"`},
			[]string{`"type":"response.completed"`},
		},
		{
			"messages", "/v1/messages",
			`{"model":"glm-5.2","stream":true,"max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`,
			[]string{`"stop_reason":"max_tokens"`},
			[]string{`"stop_reason":"end_turn"`},
		},
	}
	for _, tc := range cases {
		up := newFakeUpstream(t, func(string) (int, string, bool) { return 200, cut, true })
		p := testPoolWith(&auth.Auth{UID: "u1", AccessToken: "at1", ExpiresAt: 9999999999})
		h := NewHandler(Config{Pool: p, Upstream: up})
		rec := httptest.NewRecorder()
		h.ServeHTTP(rec, httptest.NewRequest("POST", tc.path, strings.NewReader(tc.body)))
		body := rec.Body.String()
		for _, w := range tc.want {
			if !strings.Contains(body, w) {
				t.Errorf("%s: 期望收尾 %s,实际:\n%s", tc.name, w, body)
			}
		}
		for _, r := range tc.reject {
			if strings.Contains(body, r) {
				t.Errorf("%s: 断流不得伪装成正常完成(%s):\n%s", tc.name, r, body)
			}
		}
	}
}

// TestStreamCutMidToolCallMarksTruncated 断流落在工具调用参数中间(agent 客户端最
// 常见的形态):必须仍然补截断终态(guard 里 toolSeen 就是为此),且不与既有工具帧
// 重复/错序——客户端据此能判断这个调用是残缺的,而不是拿半截 JSON 去执行。
func TestStreamCutMidToolCallMarksTruncated(t *testing.T) {
	cut := `data: {"id":"c1","object":"chat.completion.chunk","model":"glm-5.2","choices":[{"index":0,"delta":{"role":"assistant"}}]}` + "\n\n" +
		`data: {"id":"c1","object":"chat.completion.chunk","model":"glm-5.2","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"call_1","type":"function","function":{"name":"Bash","arguments":"{\"cmd\":"}}]}}]}` + "\n\n"
	up := newFakeUpstream(t, func(string) (int, string, bool) { return 200, cut, true })
	p := testPoolWith(&auth.Auth{UID: "u1", AccessToken: "at1", ExpiresAt: 9999999999})
	h := NewHandler(Config{Pool: p, Upstream: up})
	rec := httptest.NewRecorder()
	h.ServeHTTP(rec, httptest.NewRequest("POST", "/v1/chat/completions",
		strings.NewReader(`{"model":"glm-5.2","stream":true,"messages":[{"role":"user","content":"hi"}]}`)))
	body := rec.Body.String()
	if n := strings.Count(body, `"finish_reason":"length"`); n != 1 {
		t.Fatalf("工具调用中断流应恰好一个截断终态,got=%d:\n%s", n, body)
	}
	if !strings.Contains(body, `"tool_calls"`) {
		t.Fatalf("工具调用帧应保留:\n%s", body)
	}
	if strings.Index(body, `"tool_calls"`) > strings.Index(body, `"finish_reason":"length"`) {
		t.Fatalf("终态必须落在工具调用帧之后:\n%s", body)
	}
}

// TestResponsesStreamCutAutoContinued 端到端(用户报的那个症状):上游流半路断掉
// (有正文、无 finish_reason、无 [DONE])→ 网关自动带已收内容续写一次 → 客户端看到的
// 是一条连续输出、以 response.completed 正常收尾,而不是「半截 + 要手动继续」。
func TestResponsesStreamCutAutoContinued(t *testing.T) {
	cut := `data: {"id":"c1","model":"m","choices":[{"index":0,"delta":{"content":"半截"}}]}` + "\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) {
		if call == 1 {
			return 200, cut
		}
		return 200, testSeg2Text
	})
	p := testPoolWith(&auth.Auth{UID: "u1", AccessToken: "at1", ExpiresAt: 9999999999})
	h := NewHandler(Config{Pool: p, Upstream: up})
	rec := httptest.NewRecorder()
	h.ServeHTTP(rec, httptest.NewRequest("POST", "/v1/responses",
		strings.NewReader(`{"model":"m","stream":true,"input":"hi"}`)))
	body := rec.Body.String()
	if len(*reqs) != 2 {
		t.Fatalf("应恰好 1 次自动续写(共 2 次上游调用),got=%d", len(*reqs))
	}
	if !strings.Contains((*reqs)[1], "半截") {
		t.Fatalf("续写请求应带已收内容(幂等不重复):\n%s", (*reqs)[1])
	}
	for _, want := range []string{"半截", "继续", "response.completed"} {
		if !strings.Contains(body, want) {
			t.Fatalf("客户端可见流缺 %q:\n%s", want, body)
		}
	}
	for _, bad := range []string{"response.incomplete", "max_output_tokens"} {
		if strings.Contains(body, bad) {
			t.Fatalf("续上之后不得报截断(%s):\n%s", bad, body)
		}
	}
}

// TestNonStreamCutTruncatedAcrossAllProtocols 同一根因在非流式一侧:客户端不带
// stream 时,网关向**自己**聚合上游流(上游禁非流式),Aggregate 的 finish_reason
// 就是客户端看到的那个。responses / messages 的非流式按同一字段推导 status /
// stop_reason,所以 Aggregate 一处修正三边同时生效——这条测试盯住那条传导链。
func TestNonStreamCutTruncatedAcrossAllProtocols(t *testing.T) {
	cut := `data: {"id":"c1","object":"chat.completion.chunk","model":"glm-5.2","choices":[{"index":0,"delta":{"role":"assistant"}}]}` + "\n\n" +
		`data: {"id":"c1","object":"chat.completion.chunk","model":"glm-5.2","choices":[{"index":0,"delta":{"content":"半截"}}]}` + "\n\n"
	cases := []struct {
		name, path, body string
		want             []string
		reject           []string
	}{
		{
			"chat", "/v1/chat/completions",
			`{"model":"glm-5.2","stream":false,"messages":[{"role":"user","content":"hi"}]}`,
			[]string{`"finish_reason":"length"`},
			[]string{`"finish_reason":"stop"`},
		},
		{
			"responses", "/v1/responses",
			`{"model":"glm-5.2","stream":false,"input":"hi"}`,
			[]string{`"status":"incomplete"`, `"reason":"max_output_tokens"`},
			[]string{`"status":"completed"`},
		},
		{
			"messages", "/v1/messages",
			`{"model":"glm-5.2","stream":false,"max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`,
			[]string{`"stop_reason":"max_tokens"`},
			[]string{`"stop_reason":"end_turn"`},
		},
	}
	for _, tc := range cases {
		up := newFakeUpstream(t, func(string) (int, string, bool) { return 200, cut, true })
		p := testPoolWith(&auth.Auth{UID: "u1", AccessToken: "at1", ExpiresAt: 9999999999})
		h := NewHandler(Config{Pool: p, Upstream: up})
		rec := httptest.NewRecorder()
		h.ServeHTTP(rec, httptest.NewRequest("POST", tc.path, strings.NewReader(tc.body)))
		body := rec.Body.String()
		for _, w := range tc.want {
			if !strings.Contains(body, w) {
				t.Errorf("%s: 期望 %s,实际:\n%s", tc.name, w, body)
			}
		}
		for _, r := range tc.reject {
			if strings.Contains(body, r) {
				t.Errorf("%s: 断流不得伪装成正常完成(%s):\n%s", tc.name, r, body)
			}
		}
	}
}

// TestContinueReaderSeg2FailureDegrades 续写请求失败 → 降级为截断终态:
// tail 回放(length 在场)、不挂死、账目仍记(段1 usage 合成帧在场)。
func TestContinueReaderSeg2FailureDegrades(t *testing.T) {
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) {
		return 500, `{"code":11102,"msg":"boom"}`
	})
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg1Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if len(*reqs) != 1 {
		t.Fatalf("续写请求数=%d want 1", len(*reqs))
	}
	if !strings.Contains(s, `"finish_reason":"length"`) {
		t.Fatalf("降级后 length 终态应在场:\n%s", s)
	}
	if !strings.Contains(s, `"completion_tokens":5`) {
		t.Fatalf("段1 usage 应合成在场:\n%s", s)
	}
}

// TestContinueReaderSeamReasoningDropped 正文已开始后的续写段 reasoning 不外泄
// (仅累计,供下一段请求带回):不能让 responses 侧在 message item 还开着时开
// 第二个 reasoning item(否则 codex 报 "OutputTextDelta without active item")。
// 正文未开始时相反——照常外泄,见 TestContinueReaderSeg2ReasoningForwarded。
func TestContinueReaderSeamReasoningDropped(t *testing.T) {
	seg1 := `data: {"id":"cmpl-1","choices":[{"index":0,"delta":{"reasoning_content":"公开思考"}}]}` + "\n\n" + testSeg1Text
	seg2 := `data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"reasoning_content":"内部思考"}}]}` + "\n\n" +
		`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"content":"继续"}}]}` + "\n\n" +
		`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}` + "\n\n" +
		"data: [DONE]\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, seg2 })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(seg1)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if !strings.Contains(s, "公开思考") {
		t.Fatalf("段1 思考应外泄:\n%s", s)
	}
	if strings.Contains(s, "内部思考") {
		t.Fatalf("续写段思考不应外泄:\n%s", s)
	}
	if !strings.Contains(s, "继续") {
		t.Fatalf("续写段正文应外泄:\n%s", s)
	}
	// 续写请求应携带第一段累计的思考(上下文一致性;段2 的思考若产生段3 请求才携带)。
	if !strings.Contains((*reqs)[0], "公开思考") {
		t.Fatalf("续写请求应携带累计思考:\n%s", (*reqs)[0])
	}
}

// TestContinueReaderCapAtMaxSegments 连续截断封顶:每段都以 length 结束时,
// 最多 1 次原生 + (maxContinueSegments-1) 次续写,不无限续。
func TestContinueReaderCapAtMaxSegments(t *testing.T) {
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg1Text })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg1Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if len(*reqs) != maxContinueSegments-1 {
		t.Fatalf("续写请求数=%d want %d(封顶)", len(*reqs), maxContinueSegments-1)
	}
	if !strings.Contains(string(out), `"finish_reason":"length"`) {
		t.Fatalf("封顶后 length 终态应在场:\n%s", out)
	}
}

// TestContinueReaderSeg2ReasoningForwarded 正文尚未开始(text=0B)时的续写段
// 思考照常外泄:此时的截断全是「纯思考烧满」形态,放行思考既续着同一个
// reasoning item(零协议风险),又消掉客户端侧静默(codex stream_idle_timeout
// 会在长静默时判 responseStreamDisconnected 掐线,实测静默约 122s 断开)。
func TestContinueReaderSeg2ReasoningForwarded(t *testing.T) {
	// 段1:纯思考截断(正文 0B —— 线上观测到的全部续写场景)。
	seg1 := `data: {"id":"cmpl-1","choices":[{"index":0,"delta":{"reasoning_content":"公开思考"}}]}` + "\n\n" +
		`data: {"id":"cmpl-1","choices":[{"index":0,"delta":{},"finish_reason":"length"}]}` + "\n\n" +
		"data: [DONE]\n\n"
	seg2 := `data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"reasoning_content":"内部思考"}}]}` + "\n\n" +
		`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"content":"继续"}}]}` + "\n\n" +
		`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}` + "\n\n" +
		"data: [DONE]\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, seg2 })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(seg1)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if len(*reqs) != 1 {
		t.Fatalf("续写请求数=%d want 1", len(*reqs))
	}
	if !strings.Contains(s, "内部思考") {
		t.Fatalf("正文未开始时续写段思考应放行(消静默):\n%s", s)
	}
	if strings.Index(s, "内部思考") > strings.Index(s, "继续") {
		t.Fatalf("思考应先于正文出现:\n%s", s)
	}
	// 放行不等于放弃累计:首次续写请求应携带段1 累计思考(上下文一致性;
	// 段2 的思考若再触发续写请求才会携带)。
	req := (*reqs)[0]
	if !strings.Contains(req, "公开思考") {
		t.Fatalf("续写请求应携带累计思考:\n%s", req)
	}
}

// TestContinueReaderKeepaliveOnEmptyFrames 残余静默源(续写段的全空 no-op 帧)
// 按 keepaliveEvery 补 SSE 注释帧——chat 路径直达客户端喂活连接。注释帧是
// SSE 规范 no-op(": keepalive"),不得影响正文与终态。
func TestContinueReaderKeepaliveOnEmptyFrames(t *testing.T) {
	old := keepaliveEvery
	keepaliveEvery = 0 // 确定性:每条被吞帧都补,不依赖墙钟
	t.Cleanup(func() { keepaliveEvery = old })

	// 段1:带正文截断(text>0,续写段思考会走剥离分支)。
	// 段2:全空 no-op 帧(残余静默源)→ 触发 keepalive。
	empty := `data: {"id":"cmpl-2","choices":[{"index":0,"delta":{}}]}` + "\n\n"
	seg2 := empty +
		`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"reasoning_content":"内部思考"}}]}` + "\n\n" +
		`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"content":"继续"}}]}` + "\n\n" +
		`data: {"id":"cmpl-2","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}` + "\n\n" +
		"data: [DONE]\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) { return 200, seg2 })
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg1Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if len(*reqs) != 1 {
		t.Fatalf("续写请求数=%d want 1", len(*reqs))
	}
	if n := strings.Count(s, ": keepalive"); n < 1 {
		t.Fatalf("keepalive 注释帧=%d want ≥1:\n%s", n, s)
	}
	if strings.Contains(s, "内部思考") {
		t.Fatalf("正文已开始后续写段思考不应外泄:\n%s", s)
	}
	// 注释帧不能破坏正常收尾:正文与 stop 终态照常在。
	for _, want := range []string{"你好", "世界", "继续", `"finish_reason":"stop"`} {
		if !strings.Contains(s, want) {
			t.Fatalf("输出缺少 %q:\n%s", want, s)
		}
	}
}

// TestContinueInStreamErrorFrameNoRetry 上游把失败塞在流里(HTTP 200 内嵌
// {code,msg,displayMsg}、无 choices 无 error)时:按显式失败终止,不续写、不盲重试
// ——撞的是已拦下这个账号的策略,重试只是白跑一次;此前这种帧不置 errSeen,日志会把它
// 记成「段无 finish 结束(疑似上游断流)」并让续写段白跑一次盲重试。
func TestContinueInStreamErrorFrameNoRetry(t *testing.T) {
	blocked := `data: {"code":11140,"msg":"request illegal","displayMsg":"请求被安全策略拦截"}` + "\n\n"

	// ① 续写段(seg=1)收到拦截帧:不盲重试(首次请求由测试直接喂入,故只应有 1 次
	// 取流请求 = 那次续写;未修时这里会变成 2 —— 多一次白撞同一账号的盲重试)。
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) {
		if call == 1 {
			return 200, blocked
		}
		return 200, testSeg2Text
	})
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg1Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if len(*reqs) != 1 {
		t.Fatalf("拦截帧不得触发盲重试:取流请求数=%d want 1(未修时为 2)", len(*reqs))
	}
	if s := string(out); !strings.Contains(s, "11140") || !strings.Contains(s, "request illegal") {
		t.Fatalf("拦截帧应如实透传给客户端:\n%s", s)
	}

	// ② 主段(seg=0)收到拦截帧:同样按显式失败终止,不续写。
	// 主段由初始流喂入(截到 length 终态之前 + 拦截帧 + EOF),故取流请求数应为 0。
	mainSeg := testSeg1Text[:strings.Index(testSeg1Text, `data: {"id":"cmpl-1","choices":[{"index":0,"delta":{},"finish_reason":"length"}]`)] + blocked
	up2, reqs2 := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	r2 := newTestContinueReader(t, up2, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(mainSeg)))
	out2, err := io.ReadAll(r2)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	if len(*reqs2) != 0 {
		t.Fatalf("主段拦截不得发起续写:取流请求数=%d want 0(未修时为 1)", len(*reqs2))
	}
	if s := string(out2); !strings.Contains(s, "11140") || !strings.Contains(s, "你好") {
		t.Fatalf("主段拦截:正文照常 + 拦截帧如实透传,实际:\n%s", s)
	}
}

// TestContinueReaderSeg2StreamCutBlindRetry 续写段被上游断流(无 finish_reason
// 直接 EOF)时盲重试一次:重试请求体与断流段逐字节相同(已收内容都进 assistant
// 前缀,幂等不重复),客户端最终拿到完整内容。线上 17:50 的「32k 处截断」正是
// 这条路径(15 次续写里 1 次上游断流)。
func TestContinueReaderSeg2StreamCutBlindRetry(t *testing.T) {
	// 断流段:有正文、无 finish_reason、无 [DONE],直接 EOF。
	segCut := `data: {"id":"cmpl-2","choices":[{"index":0,"delta":{"content":"部分"}}]}` + "\n\n"
	up, reqs := fakeContinueUpstream(t, func(call int) (int, string) {
		if call == 1 {
			return 200, segCut
		}
		return 200, testSeg2Text
	})
	r := newTestContinueReader(t, up, `{"model":"m","messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(testSeg1Text)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if len(*reqs) != 2 {
		t.Fatalf("盲重试后应有 2 次续写请求,实际 %d", len(*reqs))
	}
	if !strings.Contains((*reqs)[1], "部分") {
		t.Fatalf("重试请求应携带断流段已收内容(幂等不重复):\n%s", (*reqs)[1])
	}
	for _, want := range []string{"你好", "世界", "部分", "继续", `"finish_reason":"stop"`} {
		if !strings.Contains(s, want) {
			t.Fatalf("客户端可见流缺 %q:\n%s", want, s)
		}
	}
	if strings.Contains(s, `"finish_reason":"length"`) {
		t.Fatalf("续写成功后 length 终态不应外泄:\n%s", s)
	}
}

func TestBareModelOfStripsRealmPrefix(t *testing.T) {
	cases := []struct{ body, want string }{
		{`{"model":"cn:glm-5.2"}`, "glm-5.2"},
		{`{"model":"global:gpt-x"}`, "gpt-x"},
		{`{"model":"glm-5.2"}`, "glm-5.2"},
		{`{"model":"weird:model"}`, "weird:model"},
		{`not json`, "?"},
	}
	for _, c := range cases {
		if got := bareModelOf([]byte(c.body)); got != c.want {
			t.Errorf("bareModelOf(%s) = %q, want %q", c.body, got, c.want)
		}
	}
}

// 请求体带路由前缀时,断流补的截断终态帧仍须是裸模型名。
func TestContinueFrameUsesBareModel(t *testing.T) {
	up, _ := fakeContinueUpstream(t, func(call int) (int, string) { return 200, testSeg2Text })
	// 显式 max_tokens → limitSet,断流后不续写,直接补截断终态帧。
	cut := `data: {"id":"c1","choices":[{"index":0,"delta":{"content":"半截"}}]}` + "\n\n"
	r := newTestContinueReader(t, up, `{"model":"cn:glm-5.2","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`,
		io.NopCloser(strings.NewReader(cut)))
	out, err := io.ReadAll(r)
	if err != nil {
		t.Fatalf("read: %v", err)
	}
	s := string(out)
	if !strings.Contains(s, `"model":"glm-5.2"`) {
		t.Fatalf("截断终态帧缺少裸模型名:\n%s", s)
	}
	if strings.Contains(s, `"model":"cn:glm-5.2"`) {
		t.Fatalf("截断终态帧泄露了路由前缀:\n%s", s)
	}
	if !strings.Contains(s, `"finish_reason":"length"`) {
		t.Fatalf("截断终态帧缺少 length:\n%s", s)
	}
}