cpa-exec 0.4.4

Part of cliproxy: upstream provider executors
Documentation
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
704
705
706
707
708
709
710
711
712
713
714
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
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
//! Differential tests against `tests/fixtures/openai_compat_go.json`, produced by the real
//! Go executor (tests/reference/openai_compat). Each scenario replays the same scripted
//! upstream answer to the Rust executor and compares the raw upstream request and the
//! executor's output. Expected values come only from Go.

use std::sync::{Arc, Mutex};

use cpa_core::exec::Caller;
use serde_json::Value;
use tokio::io::{AsyncBufReadExt, AsyncReadExt, AsyncWriteExt, BufReader};

use super::*;
use crate::openai_compat_usage::Logged;

const FIXTURE: &str = include_str!("../tests/fixtures/openai_compat_go.json");

/// Shared helpers whose real port has not landed: scenarios that need them are skipped
/// and listed, so integration can remove an entry and see the scenario run.
const PENDING: &[&str] = &[];

/// One-shot raw HTTP/1.1 capture server answering with the scripted response.
struct Mock {
    addr: String,
    raw: Arc<Mutex<Option<Vec<u8>>>>,
}

impl Mock {
    async fn start(up: &Value) -> Self {
        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let addr = listener.local_addr().unwrap().to_string();
        let raw: Arc<Mutex<Option<Vec<u8>>>> = Arc::default();
        let sink = raw.clone();
        let status = up["status"].as_u64().unwrap() as u16;
        let headers: Vec<(String, String)> = up["headers"]
            .as_array()
            .map(|hs| {
                hs.iter()
                    .map(|h| (h[0].as_str().unwrap().to_owned(), h[1].as_str().unwrap().to_owned()))
                    .collect()
            })
            .unwrap_or_default();
        let mut body = up["body"].as_str().unwrap().as_bytes().to_vec();
        if up["gzip"].as_bool() == Some(true) {
            use async_compression::tokio::bufread::GzipEncoder;
            let mut encoder = GzipEncoder::new(std::io::Cursor::new(body));
            let mut out = Vec::new();
            encoder.read_to_end(&mut out).await.unwrap();
            body = out;
        }
        let no_length = up["no_length"].as_bool() == Some(true);
        tokio::spawn(async move {
            let Ok((socket, _)) = listener.accept().await else {
                return;
            };
            let (read, mut write) = socket.into_split();
            let mut reader = BufReader::new(read);
            let mut captured = Vec::new();
            let mut length = 0;
            loop {
                let mut line = Vec::new();
                if reader.read_until(b'\n', &mut line).await.unwrap_or(0) == 0 {
                    break;
                }
                captured.extend_from_slice(&line);
                let text = String::from_utf8_lossy(&line).to_ascii_lowercase();
                if let Some(v) = text.strip_prefix("content-length:") {
                    length = v.trim().parse().unwrap_or(0);
                }
                if line == b"\r\n" {
                    break;
                }
            }
            let mut payload = vec![0; length];
            reader.read_exact(&mut payload).await.unwrap();
            captured.extend_from_slice(&payload);
            *sink.lock().unwrap() = Some(captured);
            let reason = http::StatusCode::from_u16(status)
                .ok()
                .and_then(|s| s.canonical_reason())
                .unwrap_or("");
            let mut out = format!("HTTP/1.1 {status} {reason}\r\n").into_bytes();
            for (n, v) in &headers {
                out.extend_from_slice(format!("{n}: {v}\r\n").as_bytes());
            }
            if !no_length {
                out.extend_from_slice(format!("Content-Length: {}\r\n", body.len()).as_bytes());
            }
            out.extend_from_slice(b"Connection: close\r\n\r\n");
            out.extend_from_slice(&body);
            let _ = write.write_all(&out).await;
            let _ = write.shutdown().await;
        });
        Self { addr, raw }
    }

    fn request(&self) -> Option<Vec<u8>> {
        let raw = self.raw.lock().unwrap().clone()?;
        let mut out = replace(&raw, self.addr.as_bytes(), b"UPSTREAM");
        // Go's multipart writer picks a random 60-hex-digit boundary; so does Rust.
        let marker = b"boundary=";
        if let Some(i) = out.windows(marker.len()).position(|w| w == marker)
            && let Some(boundary) = out.get(i + marker.len()..i + marker.len() + 60)
            && boundary.iter().all(u8::is_ascii_hexdigit)
        {
            let boundary = boundary.to_vec();
            out = replace(&out, &boundary, b"BOUNDARY");
        }
        Some(out)
    }
}

fn replace(haystack: &[u8], from: &[u8], to: &[u8]) -> Vec<u8> {
    let mut out = Vec::with_capacity(haystack.len());
    let mut i = 0;
    while i < haystack.len() {
        if haystack[i..].starts_with(from) {
            out.extend_from_slice(to);
            i += from.len();
        } else {
            out.push(haystack[i]);
            i += 1;
        }
    }
    out
}

fn credential(s: &Value, cfg: &Config, addr: &str) -> Credential {
    let index = s["config_auth"].as_i64().unwrap();
    if index >= 0 {
        let compat: Vec<Credential> = cpa_core::config::credentials::from_config(cfg)
            .into_iter()
            .filter(|c| {
                c.attributes.get("compat_name").is_some_and(|n| !n.is_empty()) || c.provider == "openai-compatibility"
            })
            .collect();
        return compat[index as usize].clone();
    }
    let metadata = s["metadata"].as_object().cloned().unwrap_or_default();
    let mut c = Credential::from_file(
        std::path::Path::new("/fixture"),
        std::path::Path::new("/fixture/compat.json"),
        {
            let mut m = metadata;
            m.insert("type".into(), s["provider"].clone());
            m
        },
    )
    .unwrap();
    for (k, v) in s["attributes"].as_object().into_iter().flatten() {
        c.attributes
            .insert(k.clone(), v.as_str().unwrap().replace("UPSTREAM", addr));
    }
    c
}

fn response_format(s: &Value) -> Format {
    let source = Format::parse(s["source"].as_str().unwrap()).unwrap();
    s["response"].as_str().and_then(Format::parse).unwrap_or(source)
}

fn request(s: &Value, usage: cpa_core::exec::UsageSink) -> (ExecRequest, String) {
    let op = s["op"].as_str().unwrap();
    let source = Format::parse(s["source"].as_str().unwrap()).unwrap();
    let response = s["response"].as_str().and_then(Format::parse).unwrap_or(source);
    let model = s["model"].as_str().unwrap().to_owned();
    let payload = match s["payload_b64"].as_str() {
        Some(b64) => {
            Bytes::from(base64::engine::Engine::decode(&base64::engine::general_purpose::STANDARD, b64).unwrap())
        }
        None => Bytes::from(s["payload"].as_str().unwrap().to_owned()),
    };
    // Go sets opts.OriginalRequest only when the scenario has one.
    let original = s["original"]
        .as_str()
        .map(|o| Bytes::from(o.to_owned()))
        .unwrap_or_default();
    let mut headers = http::HeaderMap::new();
    for (k, v) in s["headers"].as_object().into_iter().flatten() {
        headers.insert(
            http::HeaderName::from_bytes(k.as_bytes()).unwrap(),
            v.as_str().unwrap().parse().unwrap(),
        );
    }
    if let Some(ct) = s["content_type"].as_str() {
        headers.insert(http::header::CONTENT_TYPE, ct.parse().unwrap());
    }
    // Go's executor-side fallback (EnsureSessionContext -> CanonicalSessionID) for a
    // request without conductor metadata, which is how the fixtures were generated.
    let session_body: &[u8] = if original.is_empty() { &payload } else { &original };
    let meta = cpa_common::session::Meta {
        execution_session: s["execution_session"].as_str(),
        derived: s["derived_session"].as_str(),
    };
    let session = Some(cpa_common::session::extract_session_id(&headers, session_body, &meta))
        .filter(|id| !id.is_empty())
        .map(|id| cpa_common::session::bound_session_identity(&id));
    let req = ExecRequest {
        operation: if op == "count" {
            Operation::CountTokens
        } else {
            Operation::Generate
        },
        source_format: source,
        response_format: response,
        stream_payloads: false,
        requested_model: s["requested_model"].as_str().unwrap_or(&model).to_owned(),
        model,
        original_body: original,
        body: payload,
        stream: op.ends_with("stream") || s["stream"].as_bool() == Some(true),
        alt: s["alt"].as_str().map(str::to_owned),
        session,
        execution_session: s["execution_session"].as_str().map(str::to_owned),
        derived_session: s["derived_session"].as_str().map(str::to_owned),
        resolved_model: bound_model(&s["resolved_model"]),
        usage,
        request_path: String::new(),
        headers,
        caller: Caller {
            principal: "fake-client-key".into(),
            source: "authorization",
            provider: String::new(),
        },
    };
    (req, op.to_owned())
}

/// The dispatch loop's binding (cpa-server capabilities::resolve), as Go's conductor
/// bound it in the generator.
pub(crate) fn bound_model(v: &Value) -> Option<cpa_core::exec::ResolvedModel> {
    let raw = v.as_object()?.clone();
    Some(cpa_core::exec::ResolvedModel {
        info: cpa_core::registry::ModelInfo::from_raw(raw).expect("resolved model info"),
        source: cpa_core::exec::ResolvedSource::ApiKey,
    })
}

fn error_json(e: &ExecError) -> Value {
    serde_json::json!({
        "status": e.status,
        "message": String::from_utf8_lossy(&e.body),
        "retry_after_ms": e.retry_after.map_or(-1, |d| d.as_millis() as i64),
        // The server's reading of IsCredentialScoped (classify::credential_scoped).
        "credential_scoped": e.scope == FailureScope::Credential && e.status == 429,
        // A plain Go error (no status) is a transport-scoped 500 here.
        "plain": e.scope == FailureScope::Transport,
    })
}

fn want_bytes(s: &Value) -> Option<Vec<u8>> {
    match s["request_b64"].as_str() {
        Some(b64) => Some(base64::engine::Engine::decode(&base64::engine::general_purpose::STANDARD, b64).unwrap()),
        None => s["request"].as_str().map(|r| r.as_bytes().to_vec()),
    }
}

#[tokio::test]
async fn go_reference_scenarios() {
    // The `responses_eof_*` scenarios (0ea4e1dc) and `stream_multiline_with_done`
    // (f8e4a840: the served model comes from accepted frames) are recorded at v8.0.35.
    let fixture = crate::payload_barrier::with_v8_0_35(
        FIXTURE,
        include_str!("../tests/fixtures/openai_compat_go_v8.0.35.json"),
        "scenarios",
    );
    let scenarios = fixture["scenarios"].as_array().unwrap();
    assert!(scenarios.len() >= 60, "fixture lost scenarios");
    let executor = OpenAICompatExecutor::default();
    let mut skipped = Vec::new();
    for s in scenarios {
        let name = s["name"].as_str().unwrap();
        let needs: Vec<&str> = s["needs"]
            .as_array()
            .map(|n| n.iter().filter_map(Value::as_str).collect())
            .unwrap_or_default();
        if needs.iter().any(|n| PENDING.contains(n)) {
            skipped.push(name);
            continue;
        }
        let mock = match s.get("upstream").filter(|u| !u.is_null()) {
            Some(up) => Some(Mock::start(up).await),
            None => None,
        };
        let addr = mock.as_ref().map_or("127.0.0.1:9".to_owned(), |m| m.addr.clone());
        let cfg = Config::parse(&s["config"].as_str().unwrap_or_default().replace("UPSTREAM", &addr)).unwrap();
        let cred = credential(s, &cfg, &addr);
        let (recorder, capture, sink) = crate::openai_compat_usage::sinks();
        let (req, op) = request(s, sink);
        let result = if op.starts_with("images") {
            let path = s["request_path"].as_str().unwrap_or_default().to_owned();
            executor.images(&cred, req, &path, &cfg).await
        } else {
            executor.execute(&cred, req, &cfg).await
        };
        let mut output = None;
        let mut chunks = Vec::new();
        let mut error = None;
        match result {
            Err(e) => error = Some(error_json(&e)),
            Ok(response) => match response.body {
                ResponseBody::Buffered(bytes) => output = Some(String::from_utf8_lossy(&bytes).into_owned()),
                ResponseBody::Stream(mut stream) => {
                    let mut joined = Vec::new();
                    while let Some(item) = stream.next().await {
                        // An error is terminal: nothing may follow it.
                        assert!(error.is_none(), "{name}: stream item after an error");
                        match item {
                            Ok(bytes) => {
                                joined.extend_from_slice(&bytes);
                                chunks.push(String::from_utf8_lossy(&bytes).into_owned());
                            }
                            Err(e) => error = Some(error_json(&e)),
                        }
                    }
                    if op == "images_stream" {
                        output = Some(String::from_utf8_lossy(&joined).into_owned());
                        chunks.clear();
                    }
                }
            },
        }
        if let Some(mock) = &mock {
            let want = match s["request_b64"].as_str() {
                Some(b64) => {
                    Some(base64::engine::Engine::decode(&base64::engine::general_purpose::STANDARD, b64).unwrap())
                }
                None => s["request"].as_str().map(|r| r.as_bytes().to_vec()),
            };
            assert_eq!(
                mock.request().map(|r| String::from_utf8_lossy(&r).into_owned()),
                want.map(|r| String::from_utf8_lossy(&r).into_owned()),
                "{name}: upstream request"
            );
            assert_eq!(mock.request(), want_bytes(s), "{name}: upstream request bytes");
            let raw = mock.raw.lock().unwrap().clone().unwrap();
            let target = String::from_utf8_lossy(&raw)
                .split(' ')
                .nth(1)
                .unwrap_or_default()
                .to_owned();
            let up = &s["upstream"];
            let status = up["status"].as_u64().unwrap() as u16;
            let ok = (200..300).contains(&status);
            let attempt = crate::openai_compat_usage::HttpAttempt {
                raw: &raw,
                url: &format!("http://{}{target}", mock.addr),
                provider: s["provider"].as_str().unwrap(),
                credential: &cred,
                logged_body: None,
                status,
                // Go's transport drops Content-Encoding from a body it decompressed.
                headers: up["headers"]
                    .as_array()
                    .into_iter()
                    .flatten()
                    .map(|h| (h[0].as_str().unwrap().to_owned(), h[1].as_str().unwrap().to_owned()))
                    .filter(|(n, _)| !(up["gzip"] == true && n.eq_ignore_ascii_case("content-encoding")))
                    .collect(),
                body: up["body"].as_str().unwrap(),
                logged: match op.as_str() {
                    "stream" if ok => Logged::Lines,
                    "images_stream" if ok => Logged::Reads,
                    _ => Logged::Whole,
                },
            };
            let events = capture.take();
            let problems = crate::openai_compat_usage::http_capture_problems(&events, &attempt);
            assert!(problems.is_empty(), "{name}: {problems:#?}");
            if op == "stream"
                && ok
                && let Some(e) = &error
            {
                stream_failure_logged(name, e, &events, &recorder.0.lock().unwrap());
            }
        } else {
            assert_eq!(
                capture.take(),
                Vec::new(),
                "{name}: capture without an upstream attempt"
            );
        }
        assert_eq!(output.as_deref(), s["output"].as_str(), "{name}: output");
        // Go's chunks are what the client's route frames; the Rust translator contract
        // emits the framed events, one per item. Responses routes join chunks into frames
        // (responsesSSEFramer), so those compare frame by frame. The route's frame repairs
        // (repairErrorPayload) are server logic and not modelled here.
        let go_chunks: Vec<&str> = s["chunks"]
            .as_array()
            .into_iter()
            .flatten()
            .filter_map(Value::as_str)
            .collect();
        if response_format(s) == Format::OpenAIResponse {
            let mut framer = cpa_translate::stream::ResponsesFramer::default();
            let mut want: Vec<bytes::Bytes> = Vec::new();
            for chunk in &go_chunks {
                want.extend(framer.write(chunk.as_bytes()));
            }
            want.extend(framer.flush());
            let want: Vec<String> = want.iter().map(|f| String::from_utf8_lossy(f).into_owned()).collect();
            assert_eq!(chunks, want, "{name}: stream frames");
        } else {
            let want_chunks: Vec<String> = go_chunks
                .iter()
                .filter_map(|c| cpa_translate::stream::frame(response_format(s), c.as_bytes()))
                .map(|f| String::from_utf8_lossy(&f).into_owned())
                .collect();
            assert_eq!(chunks, want_chunks, "{name}: stream chunks");
        }
        // Go's published usage record, derived from what the executor reported.
        if let Some(want) = s.get("usage").filter(|u| !u.is_null()) {
            let got = crate::openai_compat_usage::derived(&recorder.0.lock().unwrap(), want["failed"] == true);
            assert_eq!(&got, want, "{name}: usage");
        }
        let mut want_error = s["error"].clone();
        if !want_error.is_null() {
            let plain = want_error["status"] == 0;
            if plain {
                // A plain Go error: the handler answers 500.
                want_error["status"] = 500.into();
            }
            want_error["plain"] = plain.into();
        }
        assert_eq!(error.unwrap_or(Value::Null), want_error, "{name}: error");
    }
    assert_eq!(skipped, Vec::<&str>::new());
}

/// Go's images paths read the whole response before checking the status
/// (openai_compat_executor.go executeImages and executeImagesStream): a non-2xx body that
/// ends early is logged (`RecordAPIResponseError`) and its read error returned, where the
/// chat paths ignore it.
#[tokio::test]
async fn images_error_body_read_failure_is_logged_and_returned() {
    use crate::openai_compat_usage::Captured;
    let fixture: Value = serde_json::from_str(FIXTURE).unwrap();
    let executor = OpenAICompatExecutor::default();
    for name in ["images_error_status", "images_stream_error_status"] {
        let s = fixture["scenarios"]
            .as_array()
            .unwrap()
            .iter()
            .find(|s| s["name"] == name)
            .unwrap();
        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let addr = listener.local_addr().unwrap().to_string();
        tokio::spawn(async move {
            let (socket, _) = listener.accept().await.unwrap();
            let (read, mut write) = socket.into_split();
            let mut reader = BufReader::new(read);
            let mut length = 0;
            loop {
                let mut line = String::new();
                if reader.read_line(&mut line).await.unwrap_or(0) == 0 || line == "\r\n" {
                    break;
                }
                if let Some(v) = line.to_ascii_lowercase().strip_prefix("content-length:") {
                    length = v.trim().parse().unwrap_or(0);
                }
            }
            let mut body = vec![0; length];
            let _ = reader.read_exact(&mut body).await;
            // 100 bytes declared, 8 sent, then the connection ends.
            let _ = write
                .write_all(b"HTTP/1.1 429 Too Many Requests\r\nContent-Length: 100\r\n\r\n{\"error\"")
                .await;
            let _ = write.shutdown().await;
        });
        let cfg = Config::parse(&s["config"].as_str().unwrap().replace("UPSTREAM", &addr)).unwrap();
        let cred = credential(s, &cfg, &addr);
        let (_recorder, capture, sink) = crate::openai_compat_usage::sinks();
        let (req, _) = request(s, sink);
        let path = s["request_path"].as_str().unwrap().to_owned();
        let error = executor.images(&cred, req, &path, &cfg).await.err().expect(name);
        assert_eq!(
            error.scope,
            FailureScope::Transport,
            "{name}: the read error, not the 429"
        );
        let events = capture.take();
        assert!(
            matches!(
                events.as_slice(),
                [Captured::Request { .. }, Captured::Metadata(429, _), Captured::Error(_)]
            ),
            "{name}: {events:?}"
        );
    }
}

/// Scenarios whose stream fails on an upstream error payload: Go logs and publishes a
/// fixed text instead (`publishStreamError(err, true)`).
const PAYLOAD_ERRORS: &[&str] = &[
    "stream_named_error_event",
    "stream_data_error_with_status",
    "stream_data_type_failed",
    "stream_data_top_level_code_message",
    "stream_plain_json_after_blank_lines",
    "stream_data_error_fractional_status_string",
    "stream_data_error_overflowing_status_string",
];

/// Scenarios whose apply_patch input fails only at EOF: `EndApplyPatchStream` publishes
/// the failure without logging it.
const EOF_TOOL_FAILURES: &[&str] = &["apply_patch_stream_truncated"];

/// Go's stream failure: `RecordAPIResponseError` and `PublishFailure` with the logged
/// error, before the client gets the real one.
fn stream_failure_logged(
    name: &str,
    error: &Value,
    events: &[crate::openai_compat_usage::Captured],
    reports: &[crate::openai_compat_usage::Report],
) {
    use crate::openai_compat_usage::{Captured, Report};
    let message = error["message"].as_str().unwrap();
    let logged = if PAYLOAD_ERRORS.contains(&name) {
        "upstream stream returned an error payload"
    } else {
        message
    };
    let tail = if EOF_TOOL_FAILURES.contains(&name) {
        None
    } else {
        Some(Captured::Error(logged.to_owned()))
    };
    assert_eq!(
        events.last().filter(|e| matches!(e, Captured::Error(_))),
        tail.as_ref(),
        "{name}: logged"
    );
    let published = reports.iter().find_map(|r| match r {
        Report::PublishFailure(status, body) | Report::PublishFailureWithUsage(status, body) => {
            Some((*status, body.as_str()))
        }
        _ => None,
    });
    let status = error["status"].as_u64().unwrap() as u16;
    assert_eq!(published, Some((status, logged)), "{name}: published failure");
}

#[test]
fn provider_keys() {
    assert!(handles("openai-compatibility"));
    assert!(handles("openai-compatible-acme"));
    assert!(!handles("openai"));
    assert!(!handles("codex"));
}

fn b64(s: &Value) -> Vec<u8> {
    base64::engine::Engine::decode(&base64::engine::general_purpose::STANDARD, s.as_str().unwrap()).unwrap()
}

/// Go standard-library answers recorded by the generator (`vectors`).
#[test]
fn go_primitive_vectors() {
    use crate::openai_compat_go as go;
    use crate::openai_compat_multipart::{MediaType, parse_media_type};
    let fixture: Value = serde_json::from_str(FIXTURE).unwrap();
    let v = &fixture["vectors"];
    for case in v["http_time"].as_array().unwrap() {
        let raw = case[0].as_str().unwrap();
        let got = go::parse_http_time(raw).map(|t| match t.duration_since(std::time::UNIX_EPOCH) {
            Ok(d) => d.as_nanos() as i64,
            Err(e) => -(e.duration().as_nanos() as i64),
        });
        assert_eq!(got, case[1].as_i64(), "http.ParseTime({raw:?})");
    }
    for case in v["json_valid"].as_array().unwrap() {
        let input = b64(&case[0]);
        assert_eq!(
            go::json_valid(&input),
            case[1].as_bool().unwrap(),
            "json.Valid({:?})",
            String::from_utf8_lossy(&input[..input.len().min(40)])
        );
    }
    for case in v["gjson_int"].as_array().unwrap() {
        let raw = case[0].as_str().unwrap();
        assert_eq!(
            Some(gj::parse(raw.as_bytes()).int()),
            case[1].as_i64(),
            "gjson Int({raw})"
        );
    }
    for case in v["trim_space"].as_array().unwrap() {
        assert_eq!(go::trim_space(&b64(&case[0])), b64(&case[1]).as_slice());
    }
    for case in v["media_type"].as_array().unwrap() {
        let input = case["input"].as_str().unwrap();
        let (media, error) = (case["media"].as_str().unwrap(), case["error"].as_str().unwrap());
        let want = if error.is_empty() {
            let params = case["params"]
                .as_object()
                .unwrap()
                .iter()
                .map(|(k, v)| (k.clone(), v.as_str().unwrap().to_owned()))
                .collect();
            MediaType::Ok(media.to_owned(), params)
        } else if media.is_empty() {
            match parse_media_type(input) {
                MediaType::Invalid(reason) => {
                    assert_eq!(reason, error, "mime.ParseMediaType({input:?})");
                    MediaType::Invalid(reason)
                }
                _ => MediaType::Invalid("<expected invalid>"),
            }
        } else {
            MediaType::BadParams(media.to_owned())
        };
        assert_eq!(parse_media_type(input), want, "mime.ParseMediaType({input:?})");
    }
    for case in v["token_counts"].as_array().unwrap() {
        let (model, payload) = (case[0].as_str().unwrap(), case[1].as_str().unwrap());
        assert_eq!(
            crate::openai_compat_payload::count_chat_tokens(model, payload.as_bytes()),
            Ok(case[2].as_i64().unwrap()),
            "CountOpenAIChatTokens({model:?}, {payload})"
        );
    }
}

/// A translator holding one unfinished Responses frame.
struct Pending;

impl StreamTranslator for Pending {
    fn event(&mut self, _: &[u8]) -> Result<Vec<Bytes>, cpa_translate::Error> {
        Ok(vec![])
    }

    fn finish(&mut self) -> Result<Vec<Bytes>, cpa_translate::Error> {
        Ok(vec![])
    }

    fn flush_frames(&mut self) -> Vec<Bytes> {
        vec![Bytes::from_static(b"event: pending\n\n")]
    }
}

#[tokio::test]
async fn terminal_errors_flush_pending_frames_first() {
    // Go's responsesSSEFramer flushes before the handler writes a terminal error.
    for (lines, status) in [
        (
            vec![
                Ok(Bytes::from_static(b"data: {\"error\":{\"status\":429}}")),
                Ok(Bytes::new()),
            ],
            429,
        ),
        (vec![Err(ExecError::local(502, FailureScope::Transport, "cut"))], 502),
    ] {
        let lines: ExecStream = futures_util::stream::iter(lines).boxed();
        let out: Vec<_> = frames(
            lines,
            Box::new(Pending),
            ClaudeInputTokens::new(Format::OpenAI, Format::OpenAI, Format::OpenAI, Bytes::new()),
            Format::OpenAIResponse,
            Default::default(),
            Default::default(),
        )
        .collect()
        .await;
        assert_eq!(out.len(), 2);
        assert_eq!(out[0].as_ref().unwrap(), &Bytes::from_static(b"event: pending\n\n"));
        assert_eq!(out[1].as_ref().unwrap_err().status, status);
    }
}

/// Go 868e137 gives CountTokens the payload barrier: the rules run after thinking,
/// before the local tokenizer, so a configured message changes the count.
#[test]
fn count_tokens_sees_the_payload_rules() {
    let s = serde_json::json!({
        "op": "count", "source": "openai", "model": "gpt-4o", "provider": "openai-compatibility", "config_auth": -1,
        "payload": r#"{"model":"gpt-4o","messages":[{"role":"user","content":"hi"}]}"#,
    });
    let count = |config: &str| {
        let cfg = Config::parse(config).unwrap();
        let (req, _) = request(&s, Default::default());
        let out = super::count_tokens(&credential(&s, &cfg, "127.0.0.1:9"), &req, &cfg).unwrap();
        let ResponseBody::Buffered(bytes) = out.body else {
            panic!("count answers are buffered")
        };
        gj::get(&bytes, "usage.prompt_tokens").int()
    };
    let plain = count("");
    let configured = count(
        "payload:\n  override:\n    - models: [{name: gpt-4o}]\n      params: {messages.0.content: one two three four five six seven eight nine ten}\n",
    );
    assert!(plain > 0);
    assert!(configured > plain + 5, "{plain} -> {configured}");
}

/// Go 868e137: the chat rules run after the built-in stream options, so a filter on
/// `stream_options` removes the `include_usage` the executor sets.
#[tokio::test]
async fn chat_payload_rules_run_after_the_stream_options() {
    let fixture: Value = serde_json::from_str(FIXTURE).unwrap();
    let mut s = fixture["scenarios"]
        .as_array()
        .unwrap()
        .iter()
        .find(|s| s["name"] == "stream_basic")
        .unwrap()
        .clone();
    let rule = "\npayload:\n  filter:\n    - models: [{name: acme-chat}]\n      params: [stream_options]\n";
    s["config"] = Value::String(format!("{}{rule}", s["config"].as_str().unwrap()));
    let mock = Mock::start(&s["upstream"]).await;
    let cfg = Config::parse(&s["config"].as_str().unwrap().replace("UPSTREAM", &mock.addr)).unwrap();
    let cred = credential(&s, &cfg, &mock.addr);
    let (req, _) = request(&s, Default::default());
    let ResponseBody::Stream(mut stream) = OpenAICompatExecutor::default()
        .execute(&cred, req, &cfg)
        .await
        .unwrap()
        .body
    else {
        panic!("a stream")
    };
    while stream.next().await.is_some() {}
    let wire = String::from_utf8(mock.request().unwrap()).unwrap();
    assert!(wire.contains("\r\n\r\n{"), "{wire}");
    assert!(!wire.contains("stream_options"), "{wire}");
}

/// Go 868e137 (`ApplyMediaPayloadConfig`): the rules see a multipart edit as JSON and
/// the form is rebuilt from their result, files kept.
#[tokio::test]
async fn images_multipart_payload_rules_run_last() {
    let fixture: Value = serde_json::from_str(FIXTURE).unwrap();
    let mut s = fixture["scenarios"]
        .as_array()
        .unwrap()
        .iter()
        .find(|s| s["name"] == "images_edits_multipart_rewritten")
        .unwrap()
        .clone();
    let rule = "\npayload:\n  override:\n    - models: [{name: \"*\", protocol: openai}]\n      params: {size: auto}\n";
    s["config"] = Value::String(format!("{}{rule}", s["config"].as_str().unwrap()));
    let mock = Mock::start(&s["upstream"]).await;
    let cfg = Config::parse(&s["config"].as_str().unwrap().replace("UPSTREAM", &mock.addr)).unwrap();
    let cred = credential(&s, &cfg, &mock.addr);
    let (req, _) = request(&s, Default::default());
    let path = s["request_path"].as_str().unwrap().to_owned();
    OpenAICompatExecutor::default()
        .images(&cred, req, &path, &cfg)
        .await
        .unwrap();
    let wire = mock.request().unwrap();
    let at = wire.windows(4).position(|w| w == b"\r\n\r\n").unwrap();
    let head = String::from_utf8_lossy(&wire[..at]).into_owned();
    let content_type = head.lines().find_map(|l| l.strip_prefix("Content-Type: ")).unwrap();
    let boundary = content_type.strip_prefix("multipart/form-data; boundary=").unwrap();
    let parts = crate::openai_compat_multipart::parts(&wire[at + 4..], boundary).unwrap();
    let size = parts.iter().find(|p| p.name == "size").expect("size part");
    assert_eq!(size.body, b"auto");
    assert!(parts.iter().any(|p| !p.filename.is_empty()), "{head}");
}

/// Go ignores a refusal only the original (baseline) body hits
/// (TranslateRequestPairWithAPIKeyModelCompatibilityAndUpdateIntent): the working body is
/// the one sent, and only its own refusal refuses the request.
#[tokio::test]
async fn a_refused_baseline_alone_does_not_refuse_the_request() {
    let fixture: Value = serde_json::from_str(FIXTURE).unwrap();
    let mut s = fixture["scenarios"]
        .as_array()
        .unwrap()
        .iter()
        .find(|s| s["name"] == "claude_not_compat_drops_thinking")
        .unwrap()
        .clone();
    s["original"] = Value::String(
        r#"{"model":"chat","max_tokens":10,"messages":[{"role":"user","content":[{"type":"container_upload","source":{"type":"file","file_id":"f1"}}]}]}"#.into(),
    );
    let run = |s: Value| async move {
        let mock = Mock::start(&s["upstream"]).await;
        let cfg = Config::parse(&s["config"].as_str().unwrap().replace("UPSTREAM", &mock.addr)).unwrap();
        let cred = credential(&s, &cfg, &mock.addr);
        let (req, _) = request(&s, Default::default());
        let ok = OpenAICompatExecutor::default().execute(&cred, req, &cfg).await.is_ok();
        (ok, mock.request().map(|w| String::from_utf8(w).unwrap()))
    };
    let (ok, wire) = run(s.clone()).await;
    assert!(ok);
    assert!(wire.unwrap().contains(r#""content":"go""#));
    s["payload"] = s["original"].clone();
    assert_eq!(run(s).await, (false, None));
}

#[test]
fn compaction_summary_budget_is_configurable() {
    let cfg = Config::parse("compaction:\n  summary-max-tokens: 16384\n  summary-reasoning-effort: high\n").unwrap();
    let (max, effort) = summary_budget(&cfg);
    assert_eq!(max, 16384);
    assert_eq!(effort.as_deref(), Some("high"));

    let mut body = br#"{"model":"m","max_output_tokens":16,"reasoning":{"effort":"xhigh"}}"#.to_vec();
    apply_summary_budget(&mut body, max, effort.as_deref());
    let v: Value = serde_json::from_slice(&body).unwrap();
    assert_eq!(v["max_output_tokens"], 16384);
    assert_eq!(v["reasoning"]["effort"], "high");
}

#[test]
fn compaction_summary_budget_defaults_and_never_lowers() {
    let cfg = Config::parse("config-version: 8\n").unwrap();
    let (max, effort) = summary_budget(&cfg);
    assert_eq!(max, DEFAULT_SUMMARY_MAX_TOKENS);
    assert_eq!(effort, None);

    let mut body = br#"{"max_output_tokens":32768,"reasoning":{"effort":"xhigh"}}"#.to_vec();
    apply_summary_budget(&mut body, max, None);
    let v: Value = serde_json::from_slice(&body).unwrap();
    assert_eq!(v["max_output_tokens"], 32768);
    assert_eq!(v["reasoning"]["effort"], "low");

    let mut body = br#"{"max_output_tokens":16}"#.to_vec();
    apply_summary_budget(&mut body, max, None);
    let v: Value = serde_json::from_slice(&body).unwrap();
    assert_eq!(v["max_output_tokens"], 8192);
    assert!(v.get("reasoning").is_none());
}

#[test]
fn compaction_summary_budget_keep_and_zero() {
    let cfg = Config::parse("compaction: {summary-max-tokens: 0, summary-reasoning-effort: keep}\n").unwrap();
    let (max, effort) = summary_budget(&cfg);
    assert_eq!(max, 0);
    assert_eq!(effort.as_deref(), Some("keep"));
    let mut body = br#"{"max_output_tokens":16,"reasoning":{"effort":"xhigh"}}"#.to_vec();
    apply_summary_budget(&mut body, max, effort.as_deref());
    let v: Value = serde_json::from_slice(&body).unwrap();
    assert_eq!(v["max_output_tokens"], 16);
    assert_eq!(v["reasoning"]["effort"], "xhigh");
}

const SUMMARY_ANSWER: &str = r#"{"id":"mock","object":"chat.completion","created":0,"model":"deepseek-flash","choices":[{"index":0,"message":{"role":"assistant","content":"mock summary"},"finish_reason":"stop"}],"usage":{"prompt_tokens":19,"completion_tokens":10,"total_tokens":29}}"#;
const TRIGGER_PAYLOAD: &str = r#"{"model":"gpt-6.1-sol","input":[{"type":"message","role":"user","content":[{"type":"input_text","text":"hello"}]},{"type":"compaction_trigger"}],"reasoning":{"effort":"xhigh"},"max_output_tokens":16}"#;

/// Runs `payload` against one OpenAI-compatible key whose upstream answers with
/// `answer`. `{base}` in `config` becomes `base` with `ADDR` replaced by the mock's
/// address, `http://ADDR` by default.
async fn compaction_run(
    config: &str,
    base: Option<&str>,
    payload: &str,
    stream: bool,
    alt: Option<&str>,
    answer: &str,
) -> (Result<ExecResponse, ExecError>, Mock) {
    let mock = Mock::start(&serde_json::json!({"status": 200, "body": answer})).await;
    let base = base.map_or_else(|| format!("http://{}", mock.addr), |b| b.replace("ADDR", &mock.addr));
    let cfg = Config::parse(&config.replace("{base}", &base)).unwrap();
    let scenario = serde_json::json!({
        "config_auth": 0,
        "op": "generate",
        "source": "openai-response",
        "model": "deepseek-flash",
        "requested_model": "gpt-6.1-sol",
        "payload": payload,
        "stream": stream,
        "alt": alt,
    });
    let cred = credential(&scenario, &cfg, &mock.addr);
    let (req, _) = request(&scenario, Default::default());
    (OpenAICompatExecutor::default().execute(&cred, req, &cfg).await, mock)
}

const COMPACTION_CONFIG: &str = "config-version: 8\n\
     compaction:\n  summary-max-tokens: 16384\n  summary-reasoning-effort: high\n\
     api-keys:\n  openai-compatibility:\n    - name: mock\n      base-url: {base}\n      \
     keys: [{api-key: k}]\n      models:\n        - name: deepseek-flash\n          alias: gpt-6.1-sol\n";

fn wire_body(mock: &Mock) -> Value {
    let raw = String::from_utf8(mock.request().unwrap()).unwrap();
    serde_json::from_str(raw.rsplit_once("\r\n\r\n").unwrap().1).unwrap()
}

/// The compaction trigger is answered with exactly one sealed capsule, and the
/// internal summary request uses the configurable budget instead of the caller's
/// tiny `max_output_tokens`. The answer reports the summary's Chat usage, input and
/// output apart.
#[tokio::test]
async fn compaction_trigger_returns_one_capsule_and_uses_summary_budget() {
    let (resp, mock) = compaction_run(COMPACTION_CONFIG, None, TRIGGER_PAYLOAD, false, None, SUMMARY_ANSWER).await;
    let body = match resp.unwrap().body {
        ResponseBody::Buffered(body) => body,
        ResponseBody::Stream(_) => panic!("compaction answer must be buffered"),
    };
    let answer: Value = serde_json::from_slice(&body).unwrap();
    let output = answer["output"].as_array().unwrap();
    assert_eq!(output.len(), 1);
    assert_eq!(output[0]["type"], "compaction");
    assert!(
        output[0]["encrypted_content"]
            .as_str()
            .unwrap()
            .starts_with("cpa-ag-compact-v1:")
    );
    assert_eq!(
        (
            &answer["usage"]["input_tokens"],
            &answer["usage"]["output_tokens"],
            &answer["usage"]["total_tokens"]
        ),
        (&Value::from(19), &Value::from(10), &Value::from(29))
    );
    let wire = wire_body(&mock);
    assert_eq!(wire["max_tokens"], 16384);
    assert_eq!(wire["reasoning_effort"], "high");
}

/// A streamed trigger gets the same capsule as events, and `response.completed`
/// carries the summary's Chat usage, input and output apart.
#[tokio::test]
async fn compaction_trigger_stream_reports_chat_usage() {
    let (resp, _mock) = compaction_run(COMPACTION_CONFIG, None, TRIGGER_PAYLOAD, true, None, SUMMARY_ANSWER).await;
    let ResponseBody::Stream(stream) = resp.unwrap().body else {
        panic!("a streamed trigger is answered with events");
    };
    let frames: Vec<Bytes> = stream.map(Result::unwrap).collect().await;
    let text = String::from_utf8(frames.concat()).unwrap();
    let completed = text
        .split("\n\n")
        .find_map(|frame| frame.strip_prefix("event: response.completed\ndata: "))
        .unwrap();
    let usage = &serde_json::from_str::<Value>(completed).unwrap()["response"]["usage"];
    assert_eq!(
        (&usage["input_tokens"], &usage["output_tokens"], &usage["total_tokens"]),
        (&Value::from(19), &Value::from(10), &Value::from(29))
    );
}

/// The summary request is prepared like any chat request: a model that takes
/// `max_completion_tokens` gets the floor there, the payload rules run last, and a
/// configured effort OpenAI does not define (`max`) is not sent.
#[tokio::test]
async fn compaction_summary_uses_normal_request_preparation() {
    let config = COMPACTION_CONFIG
        .replace("summary-reasoning-effort: high", "summary-reasoning-effort: max")
        .replace(
            "alias: gpt-6.1-sol\n",
            "alias: gpt-6.1-sol\n          use-max-completion-tokens: true\n",
        )
        + "payload:\n  override:\n    - models: [{name: deepseek-flash}]\n      params: {temperature: 0.25}\n";
    let (resp, mock) = compaction_run(&config, None, TRIGGER_PAYLOAD, false, None, SUMMARY_ANSWER).await;
    resp.unwrap();
    let wire = wire_body(&mock);
    assert_eq!(wire["max_completion_tokens"], 16384, "{wire}");
    assert!(wire.get("max_tokens").is_none(), "{wire}");
    assert_eq!(wire["temperature"], 0.25, "{wire}");
    assert_eq!(wire["reasoning_effort"], "low", "{wire}");
}

/// A base URL Go parses without a host (`http:127.0.0.1:port`) is refused for the
/// summary request as for any chat request, before the key is sent anywhere.
#[tokio::test]
async fn compaction_summary_refuses_a_url_without_a_host() {
    let (resp, mock) = compaction_run(
        COMPACTION_CONFIG,
        Some("http:ADDR"),
        TRIGGER_PAYLOAD,
        false,
        None,
        SUMMARY_ANSWER,
    )
    .await;
    let Err(e) = resp else {
        panic!("a URL without a host must be refused");
    };
    let message = String::from_utf8_lossy(&e.body).into_owned();
    assert!(message.ends_with("http: no Host in request URL"), "{message}");
    assert!(mock.request().is_none(), "the summary request reached the upstream");
}

/// `/responses/compact` keeps the upstream's own compaction items and expands only
/// CPA capsules.
#[tokio::test]
async fn compact_alt_keeps_native_compaction_items() {
    let capsule = crate::antigravity_compaction::seal("earlier work", "deepseek-flash").unwrap();
    let native = r#"{"type":"compaction","encrypted_content":"gAAAAnative-opaque"}"#;
    let payload = format!(
        r#"{{"model":"gpt-6.1-sol","input":[{{"type":"message","role":"user","content":"hi"}},{native},{{"type":"compaction","encrypted_content":"{capsule}"}}]}}"#
    );
    let answer = r#"{"id":"resp_1","object":"response.compaction","usage":{"input_tokens":1,"output_tokens":2,"total_tokens":3}}"#;
    let (resp, mock) = compaction_run(COMPACTION_CONFIG, None, &payload, false, Some(COMPACT_ALT), answer).await;
    resp.unwrap();
    let raw = String::from_utf8(mock.request().unwrap()).unwrap();
    assert!(raw.starts_with("POST /responses/compact "), "{raw}");
    let input = wire_body(&mock)["input"].clone();
    assert_eq!(input.as_array().unwrap().len(), 3, "{input}");
    assert_eq!(input[1], serde_json::from_str::<Value>(native).unwrap());
    assert_eq!(input[2]["role"], "developer");
    assert_eq!(
        input[2]["content"][0]["text"],
        "Context summary from previous turns:\nearlier work"
    );
}

/// One request from `source` through an `acme` key holding `acme-compat` (alias `cc`)
/// and `acme-chat` (alias `chat`); returns the upstream body. `model` is the execution
/// model and may carry a thinking suffix. The binding is what the server resolves for a
/// config model with `thinking: {levels: [none, low, medium, high]}` (never
/// user-defined), plus `is_compat` when `compat`.
async fn compat_reasoning_run(model: &str, compat: bool, source: &str, payload: &str) -> Value {
    let answer = r#"{"id":"x","object":"chat.completion","created":0,"model":"m","choices":[{"index":0,"message":{"role":"assistant","content":"ok"},"finish_reason":"stop"}]}"#;
    let mock = Mock::start(&serde_json::json!({"status": 200, "body": answer})).await;
    let config = format!(
        "config-version: 8\napi-keys:\n  openai-compatibility:\n    - name: acme\n      base-url: http://{addr}/v1\n      \
         keys: [{{api-key: k}}]\n      models:\n        - name: acme-compat\n          alias: cc\n          is-compat: true\n        \
         - name: acme-chat\n          alias: chat\n",
        addr = mock.addr
    );
    let cfg = Config::parse(&config).unwrap();
    let base = parse_suffix(model).model_name;
    let mut resolved = serde_json::json!({"created": 0, "id": base, "object": "", "owned_by": "", "type": "openai-compatibility", "thinking": {"zero_allowed": true, "levels": ["none", "low", "medium", "high"]}});
    if compat {
        resolved["is_compat"] = Value::Bool(true);
    }
    let scenario = serde_json::json!({
        "config_auth": 0,
        "op": "execute",
        "source": source,
        "model": model,
        "requested_model": if compat { "cc" } else { "chat" },
        "payload": payload,
        "resolved_model": resolved,
    });
    let cred = credential(&scenario, &cfg, &mock.addr);
    let (req, _) = request(&scenario, Default::default());
    OpenAICompatExecutor::default().execute(&cred, req, &cfg).await.unwrap();
    let raw = String::from_utf8(mock.request().unwrap()).unwrap();
    serde_json::from_str(raw.rsplit_once("\r\n\r\n").unwrap().1).unwrap()
}

/// Go 0e0cab0d (v8.0.64): for an `is-compat` model, an assistant tool-call turn without
/// thinking text carries `reasoning_content: " "` exactly when the resolved effort is
/// on, and one with thinking keeps its text. The thinking rules have the last word: a
/// `(none)` suffix on an enabled body removes the translator's placeholder, and a
/// `(high)` suffix on a disabled body adds one. Other models and non-Claude clients are
/// left alone.
#[tokio::test]
async fn compat_tool_call_turns_get_placeholder_reasoning() {
    let claude = |thinking: &str| {
        format!(
            r#"{{"model":"cc","max_tokens":10,"thinking":{thinking},"messages":[{{"role":"user","content":"read"}},{{"role":"assistant","content":[{{"type":"tool_use","id":"call_1","name":"Read","input":{{}}}}]}},{{"role":"user","content":[{{"type":"tool_result","tool_use_id":"call_1","content":"a"}}]}},{{"role":"assistant","content":[{{"type":"thinking","thinking":"plan it","signature":"sig"}},{{"type":"tool_use","id":"call_2","name":"Read","input":{{}}}}]}},{{"role":"user","content":[{{"type":"tool_result","tool_use_id":"call_2","content":"b"}}]}}]}}"#
        )
    };
    let enabled = claude(r#"{"type":"enabled","budget_tokens":1024}"#);
    let disabled = claude(r#"{"type":"disabled"}"#);
    let turn = |wire: &Value, call: &str| -> Value {
        let turn = wire["messages"]
            .as_array()
            .unwrap()
            .iter()
            .find(|m| m["tool_calls"][0]["id"] == call);
        turn.unwrap_or_else(|| panic!("no turn calling {call}: {wire}")).clone()
    };
    // (effort sent upstream, reasoning_content of the turn without thinking text)
    let seen = |wire: &Value| {
        (
            wire["reasoning_effort"].clone(),
            turn(wire, "call_1")["reasoning_content"].clone(),
        )
    };

    let wire = compat_reasoning_run("acme-compat", true, "claude", &enabled).await;
    assert_eq!(seen(&wire), (Value::from("low"), Value::from(" ")), "{wire}");
    assert_eq!(turn(&wire, "call_2")["reasoning_content"], "plan it", "{wire}");

    let wire = compat_reasoning_run("acme-compat", true, "claude", &disabled).await;
    assert_eq!(seen(&wire), (Value::from("none"), Value::Null), "{wire}");

    let wire = compat_reasoning_run("acme-compat(none)", true, "claude", &enabled).await;
    assert_eq!(seen(&wire), (Value::from("none"), Value::Null), "{wire}");
    assert_eq!(turn(&wire, "call_2")["reasoning_content"], "plan it", "{wire}");

    let wire = compat_reasoning_run("acme-compat(high)", true, "claude", &disabled).await;
    assert_eq!(seen(&wire), (Value::from("high"), Value::from(" ")), "{wire}");

    let wire = compat_reasoning_run("acme-chat", false, "claude", &enabled).await;
    assert_eq!(turn(&wire, "call_1").get("reasoning_content"), None, "{wire}");

    // A native OpenAI client's blank reasoning_content is its own and is sent as is.
    let openai = r#"{"model":"cc","reasoning_effort":"high","messages":[{"role":"user","content":"read"},{"role":"assistant","content":"","reasoning_content":"","tool_calls":[{"id":"call_1","type":"function","function":{"name":"Read","arguments":"{}"}}]},{"role":"tool","tool_call_id":"call_1","content":"a"}]}"#;
    let wire = compat_reasoning_run("acme-compat", true, "openai", openai).await;
    assert_eq!(seen(&wire), (Value::from("high"), Value::from("")), "{wire}");
}