vtcode-core 0.142.1

Core library for VT Code - a Rust-based terminal coding agent
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
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
use anyhow::{Context, Result};
use chrono::{DateTime, Utc};
use serde_json::{Value, json};
use std::fmt::Write;
use vtcode_commons::llm::FinishReason;
use vtcode_config::constants::context::DEFAULT_COMPACTION_TRIGGER_RATIO;

use crate::config::types::{ReasoningEffortLevel, VerbosityLevel};
use crate::exec::events::CompactionMode;
use crate::llm::provider::{LLMProvider, LLMRequest, Message, MessageContent, MessageRole, ResponsesCompactionOptions};
use crate::llm::utils::truncate_to_token_limit;

pub mod auto;
pub mod memory_envelope;
pub mod prefire;
pub mod summarizer;
pub mod two_pass;

pub use crate::compaction::prefire::{AsyncCompactionCache, PrefireState};

pub const SUPPRESS_NONE: u8 = 0;
pub const SUPPRESS_TURN: u8 = 1;
pub const SUPPRESS_STICKY: u8 = 2;
pub const SUPPRESS_UNTIL_SUCCESS: u8 = 3;

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum SuppressReason {
    CreditBlock,
    Size,
    Auth,
    Schema,
    Other,
}

impl SuppressReason {
    fn suppress_state(self) -> u8 {
        match self {
            SuppressReason::Size | SuppressReason::Schema => SUPPRESS_STICKY,
            SuppressReason::CreditBlock | SuppressReason::Auth => SUPPRESS_UNTIL_SUCCESS,
            SuppressReason::Other => SUPPRESS_TURN,
        }
    }
}

/// Classify a deterministic compaction failure's error text into a fixed
/// [`SuppressReason`] (drives telemetry + sticky-vs-per-turn scope).
pub(crate) fn classify_suppress_reason(error_msg: &str) -> SuppressReason {
    let m = error_msg.to_ascii_lowercase();
    if m.contains("spending-limit")
        || m.contains("spending limit")
        || m.contains("out of credits")
        || m.contains("usage balance exhausted")
        || m.contains("usage limit reached")
    {
        SuppressReason::CreditBlock
    } else if m.contains("context length") || m.contains("too many tokens") {
        SuppressReason::Size
    } else if m.contains("status 401") || m.contains("unauthorized") {
        SuppressReason::Auth
    } else if m.contains("invalid_request_error") {
        SuppressReason::Schema
    } else {
        SuppressReason::Other
    }
}

const DEFAULT_COMPACTION_TARGET_THRESHOLD: f64 = 0.50;
const DEFAULT_COMPACTION_KEEP_LAST_MESSAGES: usize = 10;
const DEFAULT_RETAINED_USER_MESSAGE_TOKENS: usize = 20_000;
const DEFAULT_RETAINED_USER_MESSAGES: usize = 6;
/// Internal continuity budget. This is deliberately not configuration: changing
/// it changes the shape of every compacted request and therefore provider cache
/// behavior.
const CONTINUITY_TAIL_TARGET_TOKENS: usize = 20_000;
/// Keep history below the model window after reserving space for the system
/// prompt, memory envelope, summary framing, and the next response.
const COMPACTION_CONTEXT_OVERHEAD_FRACTION_DENOMINATOR: usize = 8;
const COMPACTION_CONTEXT_FIXED_OVERHEAD_TOKENS: usize = 512;
const SUMMARY_PREFIX: &str = "Previous conversation summary:\n";
const ABSTRACT_PREFIX: &str = "Earlier context (abstract):\n";
const DETAIL_PREFIX: &str = "Recent context (summary):\n";

/// Default summarization prompt. Structures the summary for continuity: after
/// reading it, the next context must feel like a seamless continuation, not a
/// fresh start. Kept as a `const` so `CompactionConfig::default` does not
/// re-allocate a ~1KB literal on every construction.
const DEFAULT_SUMMARY_PROMPT: &str = "Summarize the conversation so far using this exact structure. The goal is continuity: after reading this summary, the next context must feel like a seamless continuation of the same task, not a fresh start.\n\n## Goal\n[What the user is trying to accomplish]\n\n## Constraints & Preferences\n- [Requirements, preferences, or constraints from the user]\n\n## What I Was Just Doing\n[The single most recent action in progress: what step the agent was executing, which tool or edit was underway, and where it left off. This is the continuity anchor.]\n\n## Last Action & Result\n[The last completed action and its outcome (success, error, or partial). Include the exact error or status if relevant.]\n\n## Progress\n### Done\n- [Completed work]\n\n### In Progress\n- [Current work]\n\n### Blocked\n- [Blocking issues, if any]\n\n## Key Decisions\n- **[Decision]**: [Reason]\n\n## Next Steps\n1. [Most important next step]\n\n## Critical Context\n- [Facts needed to continue]\n\nKeep it concise and actionable. Always preserve the current task objective and acceptance criteria, file paths that were read or modified, test results and error messages, and decisions with their reasoning.";

/// Compaction configuration for context window management.
#[derive(Debug, Clone)]
pub struct CompactionConfig {
    /// Threshold (0.0-1.0) at which to trigger compaction.
    pub trigger_threshold: f64,
    /// Target usage ratio (0.0-1.0) after compaction.
    pub target_threshold: f64,
    /// Prompt for summarization.
    pub summary_prompt: String,
    /// Legacy short-circuit used to skip local compaction for tiny histories.
    pub keep_last_messages: usize,
    /// Total token budget reserved for retaining real user messages verbatim.
    pub retained_user_message_tokens: usize,
    /// Maximum number of recent user messages to retain verbatim.
    pub retained_user_messages: usize,
    /// Force local summarization even for short histories and providers with native compaction.
    pub always_summarize: bool,
    /// Enable hierarchical summarization (multi-level pyramid).
    ///
    /// When `true`, compaction produces three tiers instead of a flat summary:
    /// - **Abstract**: oldest turns compressed into 1-2 sentences
    /// - **Detail**: middle turns summarized into a paragraph
    /// - **Verbatim**: most recent turns kept as-is
    ///
    /// When `false` (default), all old turns become a single flat summary.
    pub hierarchical: bool,
    /// Auto-compaction suppression state. `SUPPRESS_NONE` (0) means compaction
    /// is allowed; any other value gates automatic compaction until the
    /// appropriate clearing event (success, model switch, etc.).
    pub auto_compact_suppressed: u8,
}

impl Default for CompactionConfig {
    fn default() -> Self {
        Self {
            trigger_threshold: DEFAULT_COMPACTION_TRIGGER_RATIO,
            target_threshold: DEFAULT_COMPACTION_TARGET_THRESHOLD,
            summary_prompt: DEFAULT_SUMMARY_PROMPT.to_string(),
            keep_last_messages: DEFAULT_COMPACTION_KEEP_LAST_MESSAGES,
            retained_user_message_tokens: DEFAULT_RETAINED_USER_MESSAGE_TOKENS,
            retained_user_messages: DEFAULT_RETAINED_USER_MESSAGES,
            always_summarize: false,
            hierarchical: false,
            auto_compact_suppressed: SUPPRESS_NONE,
        }
    }
}

/// Compact conversation history using the configured summarizer.
#[cfg_attr(feature = "profiling", hotpath::measure)]
pub async fn compact_history(
    provider: &dyn LLMProvider,
    model: &str,
    history: &[Message],
    config: &CompactionConfig,
) -> Result<Vec<Message>> {
    if history.is_empty() {
        return Ok(Vec::new());
    }

    if !config.always_summarize && history.len() <= config.keep_last_messages {
        return Ok(history.to_vec());
    }

    if !config.always_summarize && provider.supports_responses_compaction(model) {
        let compacted = provider
            .compact_history(model, history)
            .await
            .context("Failed to compact history via Responses compact endpoint")?;
        return Ok(bound_compacted_history_to_context(
            normalize_provider_compacted_history(compacted, history),
            provider,
            model,
        ));
    }

    let effective_config = context_bounded_compaction_config(provider, model, history, config);
    let (summary_history, _) = split_continuity_history(history);
    let summary_prompt = build_summary_prompt(summary_history, &effective_config.summary_prompt);
    let request = LLMRequest {
        messages: std::sync::Arc::new(vec![Message::user(summary_prompt)]),
        model: model.to_string(),
        ..Default::default()
    };

    let response = provider
        .generate(request)
        .await
        .context("Failed to generate compaction summary")?;

    let summary = response.content.unwrap_or_default().trim().to_string();
    Ok(bound_compacted_history_to_context(
        build_local_compacted_history(
            history,
            &summary,
            effective_config.retained_user_message_tokens,
            effective_config.retained_user_messages,
            // Keep the same protocol-safe continuity tail as the live/manual
            // paths. Forked histories must not lose the newest working turn.
            true,
        ),
        provider,
        model,
    ))
}

/// How the manual `/compact` command compacts for a given provider/model.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum CompactionStrategy {
    /// Provider exposes a standalone on-demand compaction endpoint
    /// (OpenAI `/responses/compact`). Delegates to `LLMProvider::compact_history_with_options`.
    NativeStandalone,
    /// Provider compacts inline via request fields, threshold-triggered
    /// (Anthropic `compact_20260112`). Invoked through `LLMProvider::generate`
    /// with `context_management` set and `pause_after_compaction`.
    NativeInline,
    /// Universal fallback: summarize history via `LLMProvider::generate` and
    /// rebuild as a summary message plus retained recent user messages.
    /// Works for every provider.
    Local,
}

/// Select the manual-compaction strategy for a provider/model.
///
/// `NativeStandalone` when the provider opts in via `supports_manual_openai_compaction`
/// (e.g. OpenAI `/responses/compact`), `NativeInline` when the provider reports
/// inline compaction support via `supports_native_inline_compaction` (e.g.
/// Anthropic `compact_20260112`), otherwise `Local`.
///
/// Note: `supports_responses_compaction` is intentionally *not* the discriminator
/// for `NativeInline`. It is overloaded — true for both OpenAI-compatible
/// standalone compaction and Anthropic inline compaction — so OpenAI-compatible
/// custom endpoints (which report it but cannot serve an Anthropic
/// `compact_20260112` edit) would otherwise be misrouted to `NativeInline` and
/// waste a rejected `generate` call before falling back to `Local`.
pub fn manual_compaction_strategy(provider: &dyn LLMProvider, model: &str) -> CompactionStrategy {
    if provider.supports_manual_openai_compaction(model) {
        CompactionStrategy::NativeStandalone
    } else if provider.supports_native_inline_compaction(model) {
        CompactionStrategy::NativeInline
    } else {
        CompactionStrategy::Local
    }
}

/// Universally meaningful manual-compaction options.
///
/// Provider-specific extras (OpenAI `service_tier` / `prompt_cache_key` / `store` /
/// `include`) are intentionally absent: the manual `/compact` command exposes only
/// the options that apply across every provider.
#[derive(Debug, Clone, Default, PartialEq, Eq)]
pub struct ManualCompactionOptions {
    /// Overrides the default summary/compaction prompt when set.
    pub instructions: Option<String>,
    /// Caps the summary/compaction output length on every provider.
    pub max_output_tokens: Option<u32>,
    /// Optional reasoning effort override for the compaction pass.
    pub reasoning_effort: Option<ReasoningEffortLevel>,
    /// Optional verbosity override for the compaction output.
    pub verbosity: Option<VerbosityLevel>,
}

impl From<ManualCompactionOptions> for ResponsesCompactionOptions {
    fn from(options: ManualCompactionOptions) -> Self {
        Self {
            instructions: options.instructions,
            max_output_tokens: options.max_output_tokens,
            reasoning_effort: options.reasoning_effort,
            verbosity: options.verbosity,
            responses_include: None,
            response_store: None,
            service_tier: None,
            prompt_cache_key: None,
        }
    }
}

impl CompactionConfig {
    /// Return a config with the manual options' instructions applied as the
    /// summary prompt override. The remaining option fields
    /// (`max_output_tokens`, `reasoning_effort`, `verbosity`) are applied to the
    /// summary `LLMRequest` directly by `summarize_locally`, not stored here.
    fn with_manual_overrides(self, options: &ManualCompactionOptions) -> Self {
        let summary_prompt = options
            .instructions
            .clone()
            .map(|instructions| instructions.trim().to_string())
            .filter(|instructions| !instructions.is_empty())
            .unwrap_or(self.summary_prompt);
        Self { summary_prompt, ..self }
    }
}

/// Compact history for the manual `/compact` command using provider-native
/// compaction when available, falling back to local summarization otherwise.
///
/// Returns the compacted messages and the `CompactionMode` that produced them
/// (`Provider` for native compaction, `Local` for client-side summarization).
#[cfg_attr(feature = "profiling", hotpath::measure)]
pub async fn compact_history_manual(
    provider: &dyn LLMProvider,
    model: &str,
    history: &[Message],
    config: &CompactionConfig,
    options: &ManualCompactionOptions,
) -> Result<(Vec<Message>, CompactionMode)> {
    if history.is_empty() {
        return Ok((Vec::new(), CompactionMode::Local));
    }
    match manual_compaction_strategy(provider, model) {
        CompactionStrategy::NativeStandalone => {
            let responses_options: ResponsesCompactionOptions = options.clone().into();
            let compacted = provider
                .compact_history_with_options(model, history, &responses_options)
                .await
                .context("Failed to compact history via provider-native compaction")?;
            Ok((
                bound_compacted_history_to_context(
                    normalize_provider_compacted_history(compacted, history),
                    provider,
                    model,
                ),
                CompactionMode::Provider,
            ))
        }
        CompactionStrategy::NativeInline => {
            compact_history_native_inline(provider, model, history, config, options).await
        }
        CompactionStrategy::Local => {
            let compacted = summarize_locally(provider, model, history, config, options).await?;
            Ok((compacted, CompactionMode::Local))
        }
    }
}

/// Native inline compaction (Anthropic `compact_20260112`).
///
/// Forces a compaction pass by setting the minimum trigger threshold with
/// `pause_after_compaction: true`, so the response contains only the compaction
/// block. If compaction does not fire (history below the provider's minimum
/// trigger, currently 50k tokens for Anthropic), transparently falls back to
/// local summarization so the manual command always succeeds.
async fn compact_history_native_inline(
    provider: &dyn LLMProvider,
    model: &str,
    history: &[Message],
    config: &CompactionConfig,
    options: &ManualCompactionOptions,
) -> Result<(Vec<Message>, CompactionMode)> {
    const ANTHROPIC_COMPACT_TRIGGER_FLOOR: u64 = 50_000;

    let mut compact_edit = serde_json::Map::new();
    compact_edit.insert("type".to_string(), json!("compact_20260112"));
    compact_edit
        .insert("trigger".to_string(), json!({ "type": "input_tokens", "value": ANTHROPIC_COMPACT_TRIGGER_FLOOR }));
    compact_edit.insert("pause_after_compaction".to_string(), json!(true));
    if let Some(instructions) = options
        .instructions
        .as_ref()
        .map(|instructions| instructions.trim())
        .filter(|instructions| !instructions.is_empty())
    {
        compact_edit.insert("instructions".to_string(), json!(instructions));
    }

    let request = LLMRequest {
        messages: std::sync::Arc::new(history.to_vec()),
        model: model.to_string(),
        context_management: Some(json!({ "edits": [Value::Object(compact_edit)] })),
        max_tokens: options.max_output_tokens,
        reasoning_effort: options.reasoning_effort,
        verbosity: options.verbosity,
        ..Default::default()
    };

    // The inline compaction request is Anthropic-specific (`compact_20260112`).
    // If the provider does not actually support inline compaction (e.g. it was
    // selected because it exposes a different standalone Responses compact
    // endpoint) the request may be rejected. Per the manual `/compact` contract
    // ("always succeeds"), swallow the inline error and fall back to local
    // summarization rather than aborting the whole command.
    let response = match provider.generate(request).await {
        Ok(response) => response,
        Err(error) => {
            tracing::warn!(
                error = %error,
                "provider-native inline compaction request failed; \
                 falling back to local summarization"
            );
            let compacted = summarize_locally(provider, model, history, config, options).await?;
            return Ok((compacted, CompactionMode::Local));
        }
    };

    if response.finish_reason == FinishReason::Pause
        && let Some(summary) = response
            .compaction
            .as_ref()
            .map(|summary| summary.trim())
            .filter(|summary| !summary.is_empty())
    {
        let effective_config = context_bounded_compaction_config(provider, model, history, config);
        let compacted = build_summary_compacted_history(history, summary, &effective_config, true);
        return Ok((bound_compacted_history_to_context(compacted, provider, model), CompactionMode::Provider));
    }

    // Compaction did not fire (e.g. history below the minimum trigger threshold);
    // fall back to local summarization so the manual command always succeeds.
    let compacted = summarize_locally(provider, model, history, config, options).await?;
    Ok((compacted, CompactionMode::Local))
}

/// Local (provider-agnostic) summarization compaction.
///
/// Builds a summary prompt from the history, asks the provider to summarize via
/// `generate`, and rebuilds the history as a summary system message plus the
/// retained recent user messages. Applies the manual options to the summary
/// request.
///
/// When `config.hierarchical` is `true`, delegates to
/// [`summarize_locally_hierarchical`] which produces a multi-tier pyramid
/// (abstract + detail + verbatim) instead of a flat summary.
async fn summarize_locally(
    provider: &dyn LLMProvider,
    model: &str,
    history: &[Message],
    config: &CompactionConfig,
    options: &ManualCompactionOptions,
) -> Result<Vec<Message>> {
    if config.hierarchical {
        return summarize_locally_hierarchical(provider, model, history, config, options).await;
    }

    let effective_config =
        context_bounded_compaction_config(provider, model, history, &config.clone().with_manual_overrides(options));
    let (summary_history, _) = split_continuity_history(history);
    let summary_prompt = build_summary_prompt(summary_history, &effective_config.summary_prompt);
    let request = LLMRequest {
        messages: std::sync::Arc::new(vec![Message::user(summary_prompt)]),
        model: model.to_string(),
        max_tokens: options.max_output_tokens,
        reasoning_effort: options.reasoning_effort,
        verbosity: options.verbosity,
        ..Default::default()
    };

    let response = provider
        .generate(request)
        .await
        .context("Failed to generate compaction summary")?;

    let summary = response.content.unwrap_or_default().trim().to_string();
    Ok(bound_compacted_history_to_context(
        build_summary_compacted_history(history, summary, &effective_config, true),
        provider,
        model,
    ))
}

/// Hierarchical local summarization: abstract + detail + verbatim pyramid.
///
/// Splits the history into three bands and summarizes each with a different
/// compression target:
/// - **Abstract** (oldest third): 1-2 sentence overview
/// - **Detail** (middle third): paragraph-level summary
/// - **Verbatim** (newest third): kept as-is plus importance-weighted retained messages
///
/// This follows the hierarchical summarization strategy from the context window
/// management literature: recent turns verbatim, older turns as paragraph
/// summaries, oldest turns as a single abstract.
async fn summarize_locally_hierarchical(
    provider: &dyn LLMProvider,
    model: &str,
    history: &[Message],
    config: &CompactionConfig,
    options: &ManualCompactionOptions,
) -> Result<Vec<Message>> {
    let effective_config =
        context_bounded_compaction_config(provider, model, history, &config.clone().with_manual_overrides(options));
    let (summary_history, _) = split_continuity_history(history);

    // Split history into three bands at roughly equal thirds.
    let total = summary_history.len();
    let band_size = total / 3;
    let abstract_end = band_size;
    let detail_end = band_size * 2;

    // Band 1 (oldest): compress into 1-2 sentence abstract.
    let abstract_band = &summary_history[..abstract_end];
    let abstract_prompt = format!(
        "In 1-2 sentences, what was the overall goal and major progress in this \
         portion of the conversation?\n\n{}",
        build_summary_prompt(abstract_band, ""),
    );
    let abstract_request = LLMRequest {
        messages: std::sync::Arc::new(vec![Message::user(abstract_prompt)]),
        model: model.to_string(),
        max_tokens: Some(150),
        reasoning_effort: options.reasoning_effort,
        verbosity: options.verbosity,
        ..Default::default()
    };
    let abstract_response = provider
        .generate(abstract_request)
        .await
        .context("Failed to generate abstract summary")?;
    let abstract_summary = abstract_response.content.unwrap_or_default().trim().to_string();

    // Band 2 (middle): paragraph-level summary using the full summary prompt.
    let detail_band = &summary_history[abstract_end..detail_end];
    let detail_prompt = build_summary_prompt(detail_band, &effective_config.summary_prompt);
    let detail_request = LLMRequest {
        messages: std::sync::Arc::new(vec![Message::user(detail_prompt)]),
        model: model.to_string(),
        max_tokens: options.max_output_tokens,
        reasoning_effort: options.reasoning_effort,
        verbosity: options.verbosity,
        ..Default::default()
    };
    let detail_response = provider
        .generate(detail_request)
        .await
        .context("Failed to generate detail summary")?;
    let detail_summary = detail_response.content.unwrap_or_default().trim().to_string();

    // Band 3 (newest): retain verbatim via the bounded protocol tail.
    let recent_band = &summary_history[detail_end..];
    let retained = collect_retained_user_messages(
        recent_band,
        effective_config.retained_user_message_tokens,
        effective_config.retained_user_messages,
    );

    // Assemble: [abstract, detail, ...retained_recent, ...continuity_tail]
    let mut new_history = Vec::with_capacity(2 + retained.len());
    new_history.push(Message::system(format!("{ABSTRACT_PREFIX}{abstract_summary}")));
    new_history.push(Message::system(format!("{DETAIL_PREFIX}{detail_summary}")));
    new_history.extend(retained);
    // Live compaction: retain the most recent turn verbatim for continuity.
    for message in continuity_tail(history) {
        new_history.push(message.clone());
    }
    Ok(bound_compacted_history_to_context(new_history, provider, model))
}

pub(crate) fn build_summary_prompt(history: &[Message], instructions: &str) -> String {
    // Pre-size for the header plus every (non-empty) message body, avoiding
    // repeated reallocations while the summary prompt is assembled.
    let estimated_len =
        instructions.len() + history.iter().map(|m| m.content.as_text().len()).sum::<usize>() + history.len() * 16;
    let mut formatted = String::with_capacity(estimated_len);
    let now: DateTime<Utc> = Utc::now();
    let _ = writeln!(&mut formatted, "Summary requested at {}.\n{}", now.to_rfc3339(), instructions);

    for message in history {
        let role = match message.role {
            MessageRole::System => "system",
            MessageRole::User => "user",
            MessageRole::Assistant => "assistant",
            MessageRole::Tool => "tool",
        };
        let content = message.content.as_text();
        if content.trim().is_empty() {
            continue;
        }
        let _ = writeln!(&mut formatted, "\n[{}]\n{}", role, content.trim());
    }

    formatted
}

fn compaction_history_budget(provider: &dyn LLMProvider, model: &str) -> Option<usize> {
    let context_size = provider.effective_context_size(model);
    (context_size > 0).then(|| {
        context_size
            .saturating_sub(context_size / COMPACTION_CONTEXT_OVERHEAD_FRACTION_DENOMINATOR)
            .saturating_sub(COMPACTION_CONTEXT_FIXED_OVERHEAD_TOKENS)
    })
}

fn context_bounded_compaction_config(
    provider: &dyn LLMProvider,
    model: &str,
    history: &[Message],
    config: &CompactionConfig,
) -> CompactionConfig {
    let mut bounded = config.clone();
    if let Some(history_budget) = compaction_history_budget(provider, model) {
        let continuity_tokens = continuity_tail(history).iter().map(Message::estimate_tokens).sum::<usize>();
        bounded.retained_user_message_tokens = bounded
            .retained_user_message_tokens
            .min(history_budget.saturating_sub(continuity_tokens));
    }
    bounded
}

/// Apply the model-window budget to provider-native results as well as local
/// results. Native providers can return a large prefix or metadata-heavy tool
/// calls, so retaining the tail alone is not sufficient to guarantee that the
/// next request fits.
fn bound_compacted_history_to_context(
    compacted: Vec<Message>,
    provider: &dyn LLMProvider,
    model: &str,
) -> Vec<Message> {
    let Some(history_budget) = compaction_history_budget(provider, model) else {
        return compacted;
    };
    if compacted.iter().map(Message::estimate_tokens).sum::<usize>() <= history_budget {
        return compacted;
    }

    let Some((tail_start, tail_end, _)) = continuity_tail_selection(&compacted) else {
        return compacted
            .into_iter()
            .scan(history_budget, |remaining, message| {
                if *remaining < 4 {
                    return None;
                }
                let bounded = bounded_message_preview(&message, *remaining);
                let used = bounded.estimate_tokens();
                if used > *remaining {
                    return None;
                }
                *remaining -= used;
                Some(bounded)
            })
            .collect();
    };

    let raw_tail = &compacted[tail_start..tail_end];
    let raw_tail_tokens = raw_tail.iter().map(Message::estimate_tokens).sum::<usize>();
    let tail = if raw_tail_tokens > history_budget {
        bounded_protocol_group(raw_tail, history_budget.max(4))
    } else {
        raw_tail.to_vec()
    };
    let tail_tokens = tail.iter().map(Message::estimate_tokens).sum::<usize>();
    let mut remaining = history_budget.saturating_sub(tail_tokens);
    let mut bounded_prefix = Vec::new();

    for (index, message) in compacted[..tail_start].iter().enumerate() {
        if remaining < 4 {
            break;
        }
        // Keep the leading summary/envelope message readable, but cap it so
        // the newest complete protocol groups retain priority.
        let message_budget = if index == 0 { remaining.min(4_096) } else { remaining };
        let bounded = bounded_message_preview(message, message_budget);
        let used = bounded.estimate_tokens();
        if used > remaining {
            if index == 0 {
                let fallback = Message::system(truncate_to_token_limit(
                    message.content.as_text().as_ref(),
                    remaining.saturating_sub(4),
                ));
                let fallback_tokens = fallback.estimate_tokens();
                if fallback_tokens <= remaining {
                    bounded_prefix.push(fallback);
                }
            }
            break;
        }
        remaining -= used;
        bounded_prefix.push(bounded);
    }

    bounded_prefix.extend(tail);
    bounded_prefix
}

pub(crate) fn build_local_compacted_history(
    history: &[Message],
    summary: &str,
    retained_user_message_tokens: usize,
    retained_user_messages: usize,
    include_continuity_tail: bool,
) -> Vec<Message> {
    let (retention_history, continuity) = if include_continuity_tail {
        split_continuity_history(history)
    } else {
        (history, Vec::new())
    };
    let retained_users =
        collect_retained_user_messages(retention_history, retained_user_message_tokens, retained_user_messages);
    let mut new_history = Vec::with_capacity(retained_users.len().saturating_add(1));
    new_history.push(Message::system(format!("{SUMMARY_PREFIX}{}", summary.trim())));
    new_history.extend(retained_users);

    // Continuity anchor: retain the newest complete protocol groups verbatim
    // within the fixed tail budget. Duplicate message text is still valid
    // across turns, so preserve the sequence by index rather than deduplicating
    // on role/content.
    if include_continuity_tail {
        for message in continuity {
            new_history.push(message);
        }
    }
    new_history
}

/// Return the newest complete user-anchored protocol groups that fit the fixed
/// continuity budget. The returned messages are owned because an oversized
/// individual group may need a bounded preview.
fn continuity_tail(history: &[Message]) -> Vec<Message> {
    let Some((start, end, oversized)) = continuity_tail_selection(history) else {
        return Vec::new();
    };
    if oversized {
        bounded_protocol_group(&history[start..end], CONTINUITY_TAIL_TARGET_TOKENS)
    } else {
        history[start..end].to_vec()
    }
}

/// Find one contiguous suffix of complete protocol groups. An incomplete
/// trailing assistant tool-call group is truncated at the assistant message,
/// preserving its user anchor while excluding the invalid protocol suffix.
fn continuity_tail_selection(history: &[Message]) -> Option<(usize, usize, bool)> {
    if history.is_empty() {
        return None;
    }
    let group_starts: Vec<usize> = history
        .iter()
        .enumerate()
        .filter_map(|(index, message)| (message.role == MessageRole::User).then_some(index))
        .collect();
    let last_group_index = group_starts.len().saturating_sub(1);
    let last_group_start = *group_starts.last()?;
    let last_group_end = history.len();
    let last_group = &history[last_group_start..last_group_end];
    let last_group_prefix_end = complete_protocol_group_prefix(last_group);
    let tail_end = last_group_start.saturating_add(last_group_prefix_end);

    let mut selected_start = None;
    let mut estimated_tokens = 0usize;
    for group_index in (0..group_starts.len()).rev() {
        let start = group_starts[group_index];
        let natural_end = group_starts.get(group_index + 1).copied().unwrap_or(history.len());
        let end = if group_index == last_group_index {
            tail_end
        } else {
            natural_end
        };
        if start == end {
            continue;
        }
        let group = &history[start..end];
        if group_index != last_group_index && !protocol_group_is_complete(group) {
            break;
        }
        let group_tokens = group.iter().map(Message::estimate_tokens).sum::<usize>();
        if selected_start.is_none() && group_tokens > CONTINUITY_TAIL_TARGET_TOKENS {
            return Some((start, end, true));
        }
        if estimated_tokens.saturating_add(group_tokens) > CONTINUITY_TAIL_TARGET_TOKENS {
            break;
        }
        selected_start = Some(start);
        estimated_tokens += group_tokens;
    }

    selected_start.map(|start| (start, tail_end, false))
}

fn protocol_group_is_complete(group: &[Message]) -> bool {
    complete_protocol_group_prefix(group) == group.len()
}

/// Return the length of the valid protocol prefix in a user-anchored group.
/// When a tool call is still pending, the prefix ends before the assistant
/// message that introduced it. This handles parallel tool calls and prevents
/// a partially answered group from entering the continuity tail.
fn complete_protocol_group_prefix(group: &[Message]) -> usize {
    if group.first().is_none_or(|message| message.role != MessageRole::User) {
        return 0;
    }

    let mut pending_tool_call_ids: Vec<&str> = Vec::new();
    let mut pending_origin = None;

    for (index, message) in group.iter().enumerate() {
        if !pending_tool_call_ids.is_empty() {
            if message.role != MessageRole::Tool {
                return pending_origin.unwrap_or(index);
            }
            let Some(tool_call_id) = message.tool_call_id.as_deref() else {
                return pending_origin.unwrap_or(index);
            };
            let Some(pending_index) = pending_tool_call_ids.iter().position(|pending_id| *pending_id == tool_call_id)
            else {
                return pending_origin.unwrap_or(index);
            };
            pending_tool_call_ids.swap_remove(pending_index);
            if pending_tool_call_ids.is_empty() {
                pending_origin = None;
            }
            continue;
        }

        // Older persisted histories may contain a tool result without the
        // assistant call metadata that introduced it. Preserve that legacy
        // pair rather than discarding an otherwise complete protocol group.
        if message.role == MessageRole::Tool {
            continue;
        }

        if message.role == MessageRole::Assistant
            && let Some(tool_calls) = message.tool_calls.as_ref()
            && !tool_calls.is_empty()
        {
            for (call_index, call) in tool_calls.iter().enumerate() {
                if call.id.is_empty() || tool_calls[..call_index].iter().any(|prior| prior.id == call.id) {
                    return index;
                }
            }
            pending_origin = Some(index);
            pending_tool_call_ids.extend(tool_calls.iter().map(|call| call.id.as_str()));
        }
    }

    pending_origin.unwrap_or(group.len())
}

fn split_continuity_history(history: &[Message]) -> (&[Message], Vec<Message>) {
    let Some((tail_start, _, _)) = continuity_tail_selection(history) else {
        return (history, Vec::new());
    };
    (&history[..tail_start], continuity_tail(history))
}

fn build_summary_compacted_history(
    history: &[Message],
    summary: impl AsRef<str>,
    config: &CompactionConfig,
    include_continuity_tail: bool,
) -> Vec<Message> {
    let (retention_history, continuity) = split_continuity_history(history);
    let retained_users = collect_retained_user_messages(
        retention_history,
        config.retained_user_message_tokens,
        config.retained_user_messages,
    );
    let mut compacted = Vec::with_capacity(retained_users.len().saturating_add(1));
    compacted.push(Message::system(format!("{SUMMARY_PREFIX}{}", summary.as_ref().trim())));
    compacted.extend(retained_users);
    if include_continuity_tail {
        for message in continuity {
            compacted.push(message);
        }
    }
    compacted
}

fn bounded_protocol_group(group: &[Message], token_budget: usize) -> Vec<Message> {
    let per_message_budget = (token_budget / group.len().max(1)).max(4);
    group
        .iter()
        .map(|message| bounded_message_preview(message, per_message_budget))
        .collect()
}

fn bounded_message_preview(message: &Message, token_budget: usize) -> Message {
    if message.estimate_tokens() <= token_budget {
        return message.clone();
    }
    let mut preview = message.clone();
    let available = token_budget.saturating_sub(4);
    let text = truncate_to_token_limit(message.content.as_text().as_ref(), available);
    preview.content = MessageContent::Text(text);

    // Tool-call arguments and provider metadata are part of the message's
    // token footprint even when `content` is empty. Preserve call IDs and
    // function names so the protocol group remains correlatable, but replace
    // oversized argument payloads with valid minimal JSON and drop optional
    // reasoning/signature metadata.
    preview.reasoning = preview
        .reasoning
        .map(|reasoning| truncate_to_token_limit(&reasoning, available.min(1_024)));
    preview.reasoning_details = None;
    preview.metadata = None;
    preview.origin_tool = None;
    if let Some(tool_calls) = preview.tool_calls.as_mut() {
        for call in tool_calls {
            if let Some(function) = call.function.as_mut()
                && function.arguments.len() > available.saturating_mul(4)
            {
                function.arguments = "{}".to_string();
            }
            if let Some(text) = call.text.as_mut() {
                *text = truncate_to_token_limit(text, available.min(1_024));
            }
            call.thought_signature = None;
        }
    }

    if preview.estimate_tokens() > token_budget {
        preview.content = MessageContent::Text(String::new());
        preview.reasoning = None;
        preview.reasoning_details = None;
        if let Some(tool_calls) = preview.tool_calls.as_mut() {
            for call in tool_calls {
                if let Some(function) = call.function.as_mut() {
                    function.arguments = "{}".to_string();
                }
                call.text = None;
                call.thought_signature = None;
            }
        }
    }
    preview
}

/// Keep the provider's compacted prefix, but apply the same bounded protocol
/// tail rules used by local compaction. A provider may return a summary only;
/// that remains valid, while malformed trailing tool calls are discarded.
fn normalize_provider_compacted_history(compacted: Vec<Message>, fallback_history: &[Message]) -> Vec<Message> {
    let mut compacted = compacted;
    while compacted.last().is_some_and(|message| {
        message.role == MessageRole::Assistant && message.tool_calls.as_ref().is_some_and(|calls| !calls.is_empty())
    }) {
        compacted.pop();
    }
    let selection = continuity_tail_selection(&compacted);
    let tail = selection.map(|(start, end, oversized)| {
        if oversized {
            bounded_protocol_group(&compacted[start..end], CONTINUITY_TAIL_TARGET_TOKENS)
        } else {
            compacted[start..end].to_vec()
        }
    });
    let mut normalized = if let Some((tail_start, _, _)) = selection {
        // `tail_end` may be before the physical end when a provider returned
        // only part of a tool-result group. Discard that invalid suffix by
        // using the selection's actual start, rather than subtracting the
        // selected tail length from the physical vector length.
        compacted[..tail_start].to_vec()
    } else {
        compacted
    };
    if tail.as_ref().is_none_or(Vec::is_empty) {
        // Some native endpoints return only their summary. Keep the newest
        // source protocol groups in that case rather than silently losing the
        // continuity anchor. Preserve sequence identity: two steering intents
        // may intentionally have identical text but distinct metadata IDs.
        normalized.extend(continuity_tail(fallback_history));
    } else if let Some(tail) = tail {
        normalized.extend(tail);
    }
    normalized
}

fn collect_retained_user_messages(history: &[Message], token_budget: usize, max_messages: usize) -> Vec<Message> {
    if token_budget == 0 || max_messages == 0 {
        return Vec::new();
    }

    // Phase 1: select up to `max_messages` user messages, scored by importance.
    let total = history.len();
    let mut user_scored: Vec<(usize, f64, &Message)> = history
        .iter()
        .enumerate()
        .filter(|(_, m)| m.role == MessageRole::User && !m.content.trim().is_empty())
        .map(|(i, m)| {
            let score = score_message(m, i, total);
            (i, score, m)
        })
        .collect();
    user_scored.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));

    let mut selected: Vec<(usize, Message)> = Vec::with_capacity(max_messages.min(history.len()));
    let mut remaining = token_budget;

    for (original_idx, _score, message) in &user_scored {
        if selected.len() >= max_messages {
            break;
        }
        let estimated = message.estimate_tokens();
        if estimated <= remaining {
            selected.push((*original_idx, (*message).clone()));
            remaining = remaining.saturating_sub(estimated);
            continue;
        }
        if let Some(truncated) = truncate_user_message(message, remaining) {
            selected.push((*original_idx, truncated));
        }
        break;
    }

    // Phase 2: if budget remains, add high-value non-user messages (tool
    // results, assistant tool calls) that fit within the remaining capacity.
    if selected.len() < max_messages && remaining > 0 {
        let mut non_user_scored: Vec<(usize, f64, &Message)> = history
            .iter()
            .enumerate()
            .filter(|(_, m)| m.role != MessageRole::User && m.role != MessageRole::System && is_retainable_message(m))
            .map(|(i, m)| {
                let score = score_message(m, i, total);
                (i, score, m)
            })
            .collect();
        non_user_scored.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));

        for (original_idx, _score, message) in &non_user_scored {
            if selected.len() >= max_messages {
                break;
            }
            let estimated = message.estimate_tokens();
            if estimated <= remaining {
                selected.push((*original_idx, (*message).clone()));
                remaining = remaining.saturating_sub(estimated);
            }
        }
    }

    // Re-sort by original conversation order, then enforce tool-call/turn
    // coherence so the compacted history is valid to send back to a provider.
    selected.sort_by_key(|(idx, _)| *idx);
    coherence_tool_call_pairs(history, &selected)
        .into_iter()
        .map(|(_, msg)| msg)
        .collect()
}

/// Keep retained tool-call turns internally consistent.
///
/// A `Tool` message references a tool call the model must have seen, and an
/// `Assistant` message that still carries `tool_calls` must be followed by the
/// results those calls produced. Sending either without its counterpart is
/// invalid: providers reject unmatched tool calls, and orphaned tool results
/// reference a call the model never observed. This pass:
///
/// - **Force-keeps** the `Tool` messages immediately following any retained
///   `Assistant` that carries `tool_calls`, so the model observes each call's
///   return value. A complete turn ends with its results, which survive even
///   if they push past the soft `max_messages` cap.
/// - **Drops** a retained `Tool` message whose calling `Assistant` (the message
///   directly before it in `history`) was *not* retained — an orphaned result
///   the model cannot reconcile.
///
/// Tool results that follow a plain `Assistant` (no `tool_calls`) are ordinary
/// turn output and are kept exactly as selected.
fn coherence_tool_call_pairs(history: &[Message], selected: &[(usize, Message)]) -> Vec<(usize, Message)> {
    let mut keep: std::collections::HashSet<usize> = selected.iter().map(|(i, _)| *i).collect();

    for (idx, msg) in selected {
        if msg.role == MessageRole::Assistant && msg.tool_calls.as_ref().is_some_and(|calls| !calls.is_empty()) {
            let mut j = *idx + 1;
            while let Some(next) = history.get(j) {
                if next.role == MessageRole::Tool {
                    keep.insert(j);
                    j += 1;
                } else {
                    break;
                }
            }
        }
    }

    // Emit every index in `keep` in original history order, dropping orphaned
    // Tool results whose calling Assistant turn was not retained.
    let mut indices: Vec<usize> = keep.iter().copied().collect();
    indices.sort_unstable();

    indices
        .into_iter()
        .filter(|idx| {
            let msg = &history[*idx];
            if msg.role != MessageRole::Tool {
                return true;
            }
            // Walk backward through this contiguous result run to decide
            // coherence against the calling assistant turn.
            let mut cursor = *idx;
            loop {
                match history.get(cursor) {
                    Some(m) if m.role == MessageRole::Tool => {
                        cursor = match cursor.checked_sub(1) {
                            Some(c) => c,
                            None => break,
                        };
                    }
                    Some(m)
                        if m.role == MessageRole::Assistant && m.tool_calls.as_ref().is_some_and(|c| !c.is_empty()) =>
                    {
                        // Reached the calling assistant: coherent only if it
                        // was retained.
                        return keep.contains(&cursor);
                    }
                    _ => {
                        // Plain assistant or boundary: ordinary output.
                        return true;
                    }
                }
            }
            true
        })
        .map(|idx| (idx, history[idx].clone()))
        .collect()
}

/// Score a message for importance-weighted retention during compaction.
///
/// Uses a weighted combination of content importance and recency:
/// - Messages containing errors, corrections, or tool results score higher
/// - Recent messages get a recency bonus
/// - Assistant messages with tool calls are moderately important
fn score_message(message: &Message, index: usize, total: usize) -> f64 {
    let content = message.content.as_text();
    let content_lower = content.to_lowercase();

    // Importance weight based on content signals.
    let importance = match message.role {
        MessageRole::User => {
            if contains_error_signal(&content_lower) {
                3.0
            } else if contains_correction_signal(&content_lower) {
                2.5
            } else {
                1.0
            }
        }
        MessageRole::Tool => {
            // Tool results contain factual data the model may need.
            2.0
        }
        MessageRole::Assistant => {
            if message.tool_calls.is_some() {
                // Assistant messages with tool calls show action taken.
                0.5
            } else {
                0.1
            }
        }
        MessageRole::System => 0.0,
    };

    // Recency bonus: linear from 0.0 (oldest) to 1.0 (newest).
    let recency = if total > 0 { index as f64 / total as f64 } else { 0.0 };

    importance + recency
}

/// Check if content contains error or failure signals.
fn contains_error_signal(content: &str) -> bool {
    content.contains("error")
        || content.contains("failed")
        || content.contains("failure")
        || content.contains("panic")
        || content.contains("bug")
        || content.contains("broken")
        || content.contains("regression")
}

/// Check if content contains user correction signals.
fn contains_correction_signal(content: &str) -> bool {
    content.contains("no,")
        || content.contains("wrong")
        || content.contains("actually")
        || content.contains("fix")
        || content.contains("instead")
        || content.contains("should be")
        || content.contains("don't")
}

/// Whether a message is worth retaining during compaction.
fn is_retainable_message(message: &Message) -> bool {
    match message.role {
        MessageRole::User => !message.content.trim().is_empty(),
        MessageRole::Tool => !message.content.trim().is_empty(),
        MessageRole::Assistant => {
            // Retain assistant messages that contain tool calls (action history).
            message.tool_calls.is_some()
        }
        MessageRole::System => false,
    }
}

fn truncate_user_message(message: &Message, token_budget: usize) -> Option<Message> {
    if token_budget <= 4 {
        return None;
    }

    let available_content_tokens = token_budget.saturating_sub(4);
    let truncated = truncate_to_token_limit(message.content.as_text().as_ref(), available_content_tokens);
    let trimmed = truncated.trim();
    if trimmed.is_empty() {
        return None;
    }

    Some(Message::user(trimmed.to_string()))
}

#[cfg(test)]
mod tests {
    use super::{
        CompactionConfig, ManualCompactionOptions, compact_history, compact_history_manual, continuity_tail,
        manual_compaction_strategy,
    };
    use crate::config::types::{ReasoningEffortLevel, VerbosityLevel};
    use crate::exec::events::CompactionMode;
    use crate::llm::provider::{
        LLMError, LLMProvider, LLMRequest, LLMResponse, Message, MessageRole, ResponsesCompactionOptions,
    };
    use async_trait::async_trait;
    use std::sync::Mutex;
    use vtcode_commons::llm::{FinishReason, ToolCall};

    struct StubProvider;

    struct NativeCompactionProvider;

    /// Provider that opts into the standalone manual-compaction path
    /// (`supports_manual_openai_compaction -> true`), e.g. OpenAI `/responses/compact`.
    struct ManualStandaloneProvider {
        last_options: Mutex<Option<ResponsesCompactionOptions>>,
    }

    /// Inline-compaction-capable provider (`supports_responses_compaction -> true`,
    /// `supports_manual_openai_compaction -> false`), e.g. Anthropic `compact_20260112`.
    /// Returns a `Pause` finish with a compaction block so the inline path succeeds.
    struct InlinePauseProvider {
        last_request: Mutex<Option<LLMRequest>>,
    }

    /// Capturing provider with no native support; used to assert the Local summary
    /// request carries the manual options.
    struct CapturingProvider {
        last_request: Mutex<Option<LLMRequest>>,
    }

    /// Inline-dispatched provider whose inline `generate` rejects the Anthropic
    /// `compact_20260112` edit. Models providers that report
    /// `supports_responses_compaction` but are not Anthropic-style inline
    /// compactors; the dispatch must fall back to Local rather than aborting.
    struct InlineRejectingProvider;

    /// Models an OpenAI-compatible custom endpoint (non-`api.openai.com` host or
    /// `provider_key_override`): it exposes the Responses API
    /// (`supports_responses_compaction == true`) but neither the standalone
    /// `/responses/compact` endpoint nor Anthropic inline compaction. The dispatch
    /// must pick `Local` rather than misrouting it to `NativeInline` (which would
    /// send an Anthropic `compact_20260112` edit only to be rejected).
    struct CompatibleEndpointProvider;

    #[async_trait]
    impl LLMProvider for StubProvider {
        fn name(&self) -> &str {
            "stub"
        }

        async fn generate(&self, _request: LLMRequest) -> Result<LLMResponse, LLMError> {
            Ok(LLMResponse::new("stub-model", "summary"))
        }

        fn supported_models(&self) -> Vec<String> {
            vec!["stub-model".to_string()]
        }

        fn validate_request(&self, _request: &LLMRequest) -> Result<(), LLMError> {
            Ok(())
        }

        fn effective_context_size(&self, _model: &str) -> usize {
            32_768
        }
    }

    #[async_trait]
    impl LLMProvider for NativeCompactionProvider {
        fn name(&self) -> &str {
            "native"
        }

        async fn generate(&self, _request: LLMRequest) -> Result<LLMResponse, LLMError> {
            Ok(LLMResponse::new("stub-model", "summary"))
        }

        fn supported_models(&self) -> Vec<String> {
            vec!["stub-model".to_string()]
        }

        fn validate_request(&self, _request: &LLMRequest) -> Result<(), LLMError> {
            Ok(())
        }

        fn supports_responses_compaction(&self, _model: &str) -> bool {
            true
        }

        async fn compact_history(&self, _model: &str, _history: &[Message]) -> Result<Vec<Message>, LLMError> {
            Ok(vec![Message::system("provider compacted".to_string())])
        }
    }

    #[async_trait]
    impl LLMProvider for ManualStandaloneProvider {
        fn name(&self) -> &str {
            "manual-standalone"
        }

        async fn generate(&self, _request: LLMRequest) -> Result<LLMResponse, LLMError> {
            Ok(LLMResponse::new("stub-model", "summary"))
        }

        fn supported_models(&self) -> Vec<String> {
            vec!["stub-model".to_string()]
        }

        fn validate_request(&self, _request: &LLMRequest) -> Result<(), LLMError> {
            Ok(())
        }

        fn supports_manual_openai_compaction(&self, _model: &str) -> bool {
            true
        }

        async fn compact_history_with_options(
            &self,
            _model: &str,
            _history: &[Message],
            options: &ResponsesCompactionOptions,
        ) -> Result<Vec<Message>, LLMError> {
            *self.last_options.lock().unwrap_or_else(std::sync::PoisonError::into_inner) = Some(options.clone());
            Ok(vec![Message::system("provider standalone compacted".to_string())])
        }
    }

    #[async_trait]
    impl LLMProvider for InlinePauseProvider {
        fn name(&self) -> &str {
            "inline-pause"
        }

        async fn generate(&self, request: LLMRequest) -> Result<LLMResponse, LLMError> {
            *self.last_request.lock().unwrap_or_else(std::sync::PoisonError::into_inner) = Some(request);
            let mut response = LLMResponse::new("stub-model", "compacted by provider");
            response.finish_reason = FinishReason::Pause;
            response.compaction = Some("provider compaction summary".to_string());
            Ok(response)
        }

        fn supported_models(&self) -> Vec<String> {
            vec!["stub-model".to_string()]
        }

        fn validate_request(&self, _request: &LLMRequest) -> Result<(), LLMError> {
            Ok(())
        }

        fn supports_responses_compaction(&self, _model: &str) -> bool {
            true
        }

        fn supports_native_inline_compaction(&self, _model: &str) -> bool {
            true
        }
    }

    #[async_trait]
    impl LLMProvider for CapturingProvider {
        fn name(&self) -> &str {
            "capturing"
        }

        async fn generate(&self, request: LLMRequest) -> Result<LLMResponse, LLMError> {
            *self.last_request.lock().unwrap_or_else(std::sync::PoisonError::into_inner) = Some(request);
            Ok(LLMResponse::new("stub-model", "summary"))
        }

        fn supported_models(&self) -> Vec<String> {
            vec!["stub-model".to_string()]
        }

        fn validate_request(&self, _request: &LLMRequest) -> Result<(), LLMError> {
            Ok(())
        }
    }

    #[async_trait]
    impl LLMProvider for InlineRejectingProvider {
        fn name(&self) -> &str {
            "inline-rejecting"
        }

        async fn generate(&self, request: LLMRequest) -> Result<LLMResponse, LLMError> {
            // Reject only the inline compaction request (carries the Anthropic
            // `compact_20260112` edit); the Local summary request must succeed.
            if request.context_management.is_some() {
                return Err(LLMError::Provider {
                    message: "provider rejected inline compact edit".to_string(),
                    metadata: None,
                });
            }
            Ok(LLMResponse::new("stub-model", "summary"))
        }

        fn supported_models(&self) -> Vec<String> {
            vec!["stub-model".to_string()]
        }

        fn validate_request(&self, _request: &LLMRequest) -> Result<(), LLMError> {
            Ok(())
        }

        fn supports_responses_compaction(&self, _model: &str) -> bool {
            true
        }

        fn supports_native_inline_compaction(&self, _model: &str) -> bool {
            true
        }
    }

    #[async_trait]
    impl LLMProvider for CompatibleEndpointProvider {
        fn name(&self) -> &str {
            "compatible-endpoint"
        }

        async fn generate(&self, _request: LLMRequest) -> Result<LLMResponse, LLMError> {
            Ok(LLMResponse::new("stub-model", "summary"))
        }

        fn supported_models(&self) -> Vec<String> {
            vec!["stub-model".to_string()]
        }

        fn validate_request(&self, _request: &LLMRequest) -> Result<(), LLMError> {
            Ok(())
        }

        // Reports Responses API support but neither standalone nor inline
        // compaction (defaults: supports_manual_openai_compaction and
        // supports_native_inline_compaction are both false).
        fn supports_responses_compaction(&self, _model: &str) -> bool {
            true
        }
    }

    fn sample_history() -> Vec<Message> {
        vec![
            Message::assistant("setup".to_string()),
            Message::user("first request".to_string()),
            Message::assistant("working".to_string()),
            Message::user("second request".to_string()),
        ]
    }

    /// Build an assistant message that carries a (single) pending tool call.
    fn assistant_with_calls(content: &str, call_id: &str) -> Message {
        let mut message = Message::assistant(content.to_string());
        message.tool_calls = Some(vec![ToolCall {
            id: call_id.to_string(),
            call_type: "function".to_string(),
            function: None,
            text: None,
            thought_signature: None,
        }]);
        message
    }

    #[test]
    fn collect_retained_keeps_tool_result_with_its_assistant() {
        // When the assistant tool-call turn is retained, its tool result must
        // survive so the turn stays coherent (the model sees each call's return).
        let history = vec![
            Message::user("u1".to_string()),
            assistant_with_calls("calling tool", "c1"),
            Message::tool_response("c1".to_string(), "r1".to_string()),
            Message::user("u2".to_string()),
        ];
        let retained = super::collect_retained_user_messages(&history, 20_000, 4);
        assert!(retained.iter().any(|m| m.content.as_text().contains("u1")));
        assert!(retained.iter().any(|m| m.content.as_text().contains("u2")));
        assert!(
            retained.iter().any(|m| m.content.as_text().contains("r1")),
            "tool result paired with its retained assistant must survive"
        );
    }

    #[test]
    fn collect_retained_drops_orphaned_tool_result() {
        // If the assistant tool-call turn is dropped (over the retention cap),
        // its tool result is orphaned — the model never saw the call — and must
        // not survive compaction, because an orphaned result is invalid to send
        // to a provider.
        let history = vec![
            Message::user("u1".to_string()),
            assistant_with_calls("calling tool", "c1"),
            Message::tool_response("c1".to_string(), "r1".to_string()),
            Message::user("u2".to_string()),
        ];
        let retained = super::collect_retained_user_messages(&history, 20_000, 3);
        assert!(retained.iter().any(|m| m.content.as_text().contains("u1")));
        assert!(retained.iter().any(|m| m.content.as_text().contains("u2")));
        assert!(!retained.iter().any(|m| m.content.as_text().contains("r1")), "orphaned tool result must be dropped");
    }

    #[tokio::test]
    async fn manual_compaction_strategy_picks_local_for_plain_provider() {
        assert_eq!(manual_compaction_strategy(&StubProvider, "stub-model"), super::CompactionStrategy::Local);
    }

    #[tokio::test]
    async fn manual_compaction_strategy_picks_native_standalone_for_manual_provider() {
        let provider = ManualStandaloneProvider { last_options: Mutex::new(None) };
        assert_eq!(manual_compaction_strategy(&provider, "stub-model"), super::CompactionStrategy::NativeStandalone);
    }

    #[tokio::test]
    async fn manual_compaction_strategy_picks_native_inline_for_responses_capable_provider() {
        let provider = InlinePauseProvider { last_request: Mutex::new(None) };
        assert_eq!(manual_compaction_strategy(&provider, "stub-model"), super::CompactionStrategy::NativeInline);
    }

    #[tokio::test]
    async fn manual_compaction_strategy_picks_local_for_compatible_endpoint() {
        // OpenAI-compatible custom endpoints report `supports_responses_compaction`
        // but cannot serve standalone `/responses/compact` or Anthropic inline
        // compaction; they must route to Local, not NativeInline.
        assert_eq!(
            manual_compaction_strategy(&CompatibleEndpointProvider, "stub-model"),
            super::CompactionStrategy::Local
        );
    }

    #[tokio::test]
    async fn compact_history_manual_uses_local_summary_for_plain_provider() {
        let history = sample_history();
        let config = CompactionConfig {
            always_summarize: true,
            ..CompactionConfig::default()
        };

        let (compacted, mode) =
            compact_history_manual(&StubProvider, "stub-model", &history, &config, &ManualCompactionOptions::default())
                .await
                .expect("manual compaction");

        assert_eq!(mode, CompactionMode::Local);
        assert_eq!(compacted.len(), 4);
        assert_eq!(compacted[0].content.as_text(), "Previous conversation summary:\nsummary");
        assert_eq!(compacted[1].content.as_text(), "first request");
        assert_eq!(compacted[2].content.as_text(), "working");
        assert_eq!(compacted[3].content.as_text(), "second request");
    }

    #[tokio::test]
    async fn compact_history_manual_uses_native_standalone_for_manual_provider() {
        let history = sample_history();
        let config = CompactionConfig::default();
        let provider = ManualStandaloneProvider { last_options: Mutex::new(None) };

        let (compacted, mode) =
            compact_history_manual(&provider, "stub-model", &history, &config, &ManualCompactionOptions::default())
                .await
                .expect("manual compaction");

        assert_eq!(mode, CompactionMode::Provider);
        assert_eq!(compacted.len(), 4);
        assert_eq!(compacted[0].content.as_text(), "provider standalone compacted");
        assert_eq!(compacted[1].content.as_text(), "first request");
        assert_eq!(compacted[2].content.as_text(), "working");
        assert_eq!(compacted[3].content.as_text(), "second request");
    }

    #[tokio::test]
    async fn compact_history_manual_passes_options_to_native_standalone() {
        let history = sample_history();
        let config = CompactionConfig::default();
        let provider = ManualStandaloneProvider { last_options: Mutex::new(None) };
        let options = ManualCompactionOptions {
            instructions: Some("keep only decisions".to_string()),
            max_output_tokens: Some(256),
            reasoning_effort: Some(ReasoningEffortLevel::Minimal),
            verbosity: Some(VerbosityLevel::High),
        };

        let (_compacted, mode) = compact_history_manual(&provider, "stub-model", &history, &config, &options)
            .await
            .expect("manual compaction");

        assert_eq!(mode, CompactionMode::Provider);
        let captured = provider.last_options.lock().unwrap().clone().expect("captured options");
        assert_eq!(captured.instructions.as_deref(), Some("keep only decisions"));
        assert_eq!(captured.max_output_tokens, Some(256));
        assert_eq!(captured.reasoning_effort, Some(ReasoningEffortLevel::Minimal));
        assert_eq!(captured.verbosity, Some(VerbosityLevel::High));
    }

    #[tokio::test]
    async fn compact_history_manual_uses_native_inline_when_pause_and_compaction_present() {
        let history = sample_history();
        let config = CompactionConfig::default();
        let provider = InlinePauseProvider { last_request: Mutex::new(None) };

        let (compacted, mode) =
            compact_history_manual(&provider, "stub-model", &history, &config, &ManualCompactionOptions::default())
                .await
                .expect("manual compaction");

        assert_eq!(mode, CompactionMode::Provider);
        assert_eq!(compacted.len(), 4);
        assert_eq!(compacted[0].content.as_text(), "Previous conversation summary:\nprovider compaction summary");
        assert_eq!(compacted[1].content.as_text(), "first request");
        assert_eq!(compacted[2].content.as_text(), "working");
        assert_eq!(compacted[3].content.as_text(), "second request");

        // The inline request must carry the `compact_20260112` edit with a forced
        // pause so the provider actually performs compaction on demand.
        let captured = provider.last_request.lock().unwrap().clone().expect("captured inline request");
        let context_management = captured
            .context_management
            .as_ref()
            .expect("context_management set on inline compaction request");
        let edit = &context_management["edits"][0];
        assert_eq!(edit["type"].as_str(), Some("compact_20260112"));
        assert_eq!(edit["pause_after_compaction"].as_bool(), Some(true));
        assert_eq!(edit["trigger"]["value"].as_u64(), Some(50_000));
    }

    #[tokio::test]
    async fn compact_history_manual_inline_request_carries_instructions_when_provided() {
        let history = sample_history();
        let config = CompactionConfig::default();
        let provider = InlinePauseProvider { last_request: Mutex::new(None) };
        let options = ManualCompactionOptions {
            instructions: Some("  keep only decisions  ".to_string()),
            ..ManualCompactionOptions::default()
        };

        let (_compacted, _mode) = compact_history_manual(&provider, "stub-model", &history, &config, &options)
            .await
            .expect("manual compaction");

        let captured = provider.last_request.lock().unwrap().clone().expect("captured inline request");
        let edit = &captured.context_management.as_ref().expect("context_management")["edits"][0];
        assert_eq!(edit["instructions"].as_str(), Some("keep only decisions"));
    }

    #[tokio::test]
    async fn compact_history_manual_falls_back_to_local_when_inline_compaction_not_fired() {
        let history = sample_history();
        let config = CompactionConfig::default();

        // NativeCompactionProvider is inline-capable but its `generate` returns a
        // normal `Stop` with no compaction block, so the inline attempt cannot
        // fire and the dispatch must transparently fall back to Local.
        let (compacted, mode) = compact_history_manual(
            &NativeCompactionProvider,
            "stub-model",
            &history,
            &config,
            &ManualCompactionOptions::default(),
        )
        .await
        .expect("manual compaction");

        assert_eq!(mode, CompactionMode::Local);
        assert_eq!(compacted.len(), 4);
        assert_eq!(compacted[0].content.as_text(), "Previous conversation summary:\nsummary");
    }

    #[tokio::test]
    async fn compact_history_manual_falls_back_to_local_when_inline_request_errors() {
        let history = sample_history();
        let config = CompactionConfig::default();

        // A provider dispatched to NativeInline that rejects the Anthropic
        // `compact_20260112` edit must not abort the whole command; the dispatch
        // falls back to Local summarization (the manual `/compact` contract:
        // always succeeds).
        let (compacted, mode) = compact_history_manual(
            &InlineRejectingProvider,
            "stub-model",
            &history,
            &config,
            &ManualCompactionOptions::default(),
        )
        .await
        .expect("manual compaction should fall back to local");

        assert_eq!(mode, CompactionMode::Local);
        assert_eq!(compacted.len(), 4);
        assert_eq!(compacted[0].content.as_text(), "Previous conversation summary:\nsummary");
    }

    #[tokio::test]
    async fn compact_history_manual_applies_options_to_local_summary_request() {
        let history = sample_history();
        let config = CompactionConfig {
            always_summarize: true,
            ..CompactionConfig::default()
        };
        let provider = CapturingProvider { last_request: Mutex::new(None) };
        let options = ManualCompactionOptions {
            instructions: Some("KEEP DECISIONS ONLY".to_string()),
            max_output_tokens: Some(128),
            reasoning_effort: Some(ReasoningEffortLevel::Minimal),
            verbosity: Some(VerbosityLevel::High),
        };

        let (compacted, mode) = compact_history_manual(&provider, "stub-model", &history, &config, &options)
            .await
            .expect("manual compaction");

        assert_eq!(mode, CompactionMode::Local);
        let captured = provider.last_request.lock().unwrap().clone().expect("captured summary request");
        assert_eq!(captured.max_tokens, Some(128));
        assert_eq!(captured.reasoning_effort, Some(ReasoningEffortLevel::Minimal));
        assert_eq!(captured.verbosity, Some(VerbosityLevel::High));
        // The custom instructions override the default summary prompt.
        let prompt = captured.messages[0].content.as_text();
        assert!(prompt.contains("KEEP DECISIONS ONLY"));
        assert!(!prompt.contains("acceptance criteria"));
        assert_eq!(compacted[0].content.as_text(), "Previous conversation summary:\nsummary");
    }

    #[tokio::test]
    async fn compact_history_manual_returns_empty_for_empty_history() {
        let (compacted, mode) = compact_history_manual(
            &StubProvider,
            "stub-model",
            &[],
            &CompactionConfig::default(),
            &ManualCompactionOptions::default(),
        )
        .await
        .expect("manual compaction");

        assert!(compacted.is_empty());
        assert_eq!(mode, CompactionMode::Local);
    }

    #[tokio::test]
    async fn compact_history_rebuilds_history_around_summary_and_important_messages() {
        let history = vec![
            Message::assistant("setup".to_string()),
            Message::user("first request".to_string()),
            Message::assistant("working".to_string()),
            Message::tool_response("call-1".to_string(), "done".to_string()),
            Message::user("second request".to_string()),
            Message::assistant("final reply".to_string()),
        ];
        let config = CompactionConfig {
            always_summarize: true,
            ..CompactionConfig::default()
        };

        let compacted = compact_history(&StubProvider, "stub-model", &history, &config)
            .await
            .expect("compacted history");

        // Summary plus the complete newest protocol groups. The assistant/tool
        // messages remain paired with their user anchors in the continuity
        // tail.
        assert_eq!(compacted.len(), 6);
        assert_eq!(compacted[0].content.as_text(), "Previous conversation summary:\nsummary");
        assert_eq!(compacted[1].content.as_text(), "first request");
        assert_eq!(compacted[2].content.as_text(), "working");
        assert_eq!(compacted[3].content.as_text(), "done");
        assert_eq!(compacted[4].content.as_text(), "second request");
        assert_eq!(compacted[5].content.as_text(), "final reply");
    }

    #[tokio::test]
    async fn compact_history_preserves_continuity_tail_over_retention_budget() {
        let history = vec![
            Message::user("alpha beta gamma delta epsilon zeta".to_string()),
            Message::assistant("ack".to_string()),
            Message::user("newest request".to_string()),
        ];
        let config = CompactionConfig {
            always_summarize: true,
            retained_user_message_tokens: 8,
            ..CompactionConfig::default()
        };

        let compacted = compact_history(&StubProvider, "stub-model", &history, &config)
            .await
            .expect("compacted history");

        assert_eq!(compacted.len(), 4);
        assert_eq!(compacted[1].content.as_text(), "alpha beta gamma delta epsilon zeta");
        assert_eq!(compacted[2].content.as_text(), "ack");
        assert_eq!(compacted[3].content.as_text(), "newest request");
    }

    #[tokio::test]
    async fn compacted_history_respects_model_context_budget() {
        let mut history = Vec::new();
        for index in 0..24 {
            history.push(Message::user(format!("request-{index} {}", "context ".repeat(1_200))));
            history.push(Message::assistant(format!("completed request {index}")));
        }
        let config = CompactionConfig {
            always_summarize: true,
            ..CompactionConfig::default()
        };

        let compacted = compact_history(&StubProvider, "stub-model", &history, &config)
            .await
            .expect("compacted history");
        let estimated_tokens = compacted.iter().map(Message::estimate_tokens).sum::<usize>();

        assert!(estimated_tokens <= 32_768 - 512);
    }

    #[tokio::test]
    async fn compact_history_caps_retained_user_message_count() {
        let history = vec![
            Message::user("first request".to_string()),
            Message::assistant("ack".to_string()),
            Message::user("second request".to_string()),
            Message::assistant("ack".to_string()),
            Message::user("third request".to_string()),
            Message::assistant("ack".to_string()),
            Message::user("fourth request".to_string()),
            Message::assistant("ack".to_string()),
            Message::user("fifth request".to_string()),
        ];
        let config = CompactionConfig {
            always_summarize: true,
            retained_user_messages: 4,
            ..CompactionConfig::default()
        };

        let compacted = compact_history(&StubProvider, "stub-model", &history, &config)
            .await
            .expect("compacted history");

        let continuity_tail = compacted
            .iter()
            .skip(1)
            .map(|message| message.content.as_text().to_string())
            .collect::<Vec<_>>();
        assert_eq!(continuity_tail.len(), history.len());
        assert_eq!(continuity_tail[0], "first request");
        assert_eq!(continuity_tail[8], "fifth request");
    }

    #[tokio::test]
    async fn compact_history_forces_local_summary_when_always_summarize_is_enabled() {
        let history = vec![
            Message::user("first request".to_string()),
            Message::assistant("working".to_string()),
            Message::user("second request".to_string()),
        ];
        let config = CompactionConfig {
            always_summarize: true,
            ..CompactionConfig::default()
        };

        let compacted = compact_history(&NativeCompactionProvider, "stub-model", &history, &config)
            .await
            .expect("compacted history");

        assert_eq!(compacted.len(), 4);
        assert_eq!(compacted[0].content.as_text(), "Previous conversation summary:\nsummary");
        assert_eq!(compacted[1].content.as_text(), "first request");
        assert_eq!(compacted[2].content.as_text(), "working");
        assert_eq!(compacted[3].content.as_text(), "second request");
    }

    #[test]
    fn default_summary_prompt_preserves_required_compaction_context() {
        let prompt = CompactionConfig::default().summary_prompt;

        assert!(prompt.contains("acceptance criteria"));
        assert!(prompt.contains("file paths that were read or modified"));
        assert!(prompt.contains("test results and error messages"));
        assert!(prompt.contains("decisions with their reasoning"));
    }

    #[test]
    fn continuity_tail_keeps_complete_turn_but_drops_unmatched_tool_call() {
        // A completed turn: user -> assistant(tool call) -> tool result. The
        // tail must keep the whole turn intact (the tool result makes the
        // trailing assistant tool call valid to send).
        let complete = vec![
            Message::user("do the thing".into()),
            {
                let mut m = Message::assistant("calling".into());
                m.tool_calls = Some(vec![ToolCall::function("c1".into(), "run".into(), "{}".into())]);
                m
            },
            Message::tool_response("c1".into(), "ran".into()),
        ];
        assert_eq!(continuity_tail(&complete).len(), 3);

        // An interrupted turn: user -> assistant(tool call) with no tool
        // result. Sending the trailing assistant message to a provider is
        // invalid, so the tail must drop it and keep only the user message.
        let interrupted = vec![Message::user("do the thing".into()), {
            let mut m = Message::assistant("calling".into());
            m.tool_calls = Some(vec![ToolCall::function("c1".into(), "run".into(), "{}".into())]);
            m
        }];
        let tail = continuity_tail(&interrupted);
        assert_eq!(tail.len(), 1);
        assert_eq!(tail[0].role, MessageRole::User);

        let parallel_complete = vec![
            Message::user("run both".into()),
            Message::assistant_with_tools(
                "calling".into(),
                vec![
                    ToolCall::function("c1".into(), "run".into(), "{}".into()),
                    ToolCall::function("c2".into(), "run".into(), "{}".into()),
                ],
            ),
            Message::tool_response("c1".into(), "first result".into()),
            Message::tool_response("c2".into(), "second result".into()),
        ];
        assert_eq!(continuity_tail(&parallel_complete).len(), 4);

        let parallel_interrupted = parallel_complete[..3].to_vec();
        let tail = continuity_tail(&parallel_interrupted);
        assert_eq!(tail.len(), 1, "a missing parallel result drops the whole assistant call");
        assert_eq!(tail[0].content.as_text(), "run both");
    }

    #[test]
    fn continuity_tail_keeps_newest_complete_groups_within_budget() {
        let mut history = Vec::new();
        for group_index in 0..12 {
            history.push(Message::user(format!("group-{group_index} {}", "alpha beta gamma delta ".repeat(2_000))));
            history.push(Message::assistant(format!("completed group {group_index}")));
        }

        let tail = continuity_tail(&history);
        let estimated_tokens = tail.iter().map(Message::estimate_tokens).sum::<usize>();

        assert!(estimated_tokens <= super::CONTINUITY_TAIL_TARGET_TOKENS);
        assert_eq!(tail.first().map(|message| message.role), Some(MessageRole::User));
        assert!(tail.iter().any(|message| message.content.as_text().contains("group-11")));
        assert!(!tail.iter().any(|message| message.content.as_text().contains("group-0 ")));
        assert_eq!(tail.len() % 2, 0, "protocol groups must remain atomic");
    }

    #[test]
    fn continuity_tail_bounds_an_oversized_newest_group() {
        let history = vec![
            Message::user("u".repeat(100_000)),
            Message::assistant("a".repeat(100_000)),
        ];

        let tail = continuity_tail(&history);

        assert_eq!(tail.len(), 2);
        assert!(tail.iter().map(Message::estimate_tokens).sum::<usize>() <= super::CONTINUITY_TAIL_TARGET_TOKENS);
    }

    #[test]
    fn continuity_tail_bounds_tool_call_metadata_without_breaking_correlation() {
        let mut assistant = Message::assistant(String::new());
        assistant.tool_calls = Some(vec![ToolCall::function(
            "call-large".into(),
            "run_command".into(),
            format!("{{\"command\":\"{}\"}}", "x".repeat(300_000)),
        )]);
        let history = vec![
            Message::user("run the command".into()),
            assistant,
            Message::tool_response("call-large".into(), "completed".into()),
        ];

        let tail = continuity_tail(&history);
        assert!(tail.iter().map(Message::estimate_tokens).sum::<usize>() <= super::CONTINUITY_TAIL_TARGET_TOKENS);
        let call = tail[1].tool_calls.as_ref().expect("tool call should remain").first().unwrap();
        assert_eq!(call.id, "call-large");
        assert_eq!(call.function.as_ref().unwrap().arguments, "{}");
        assert_eq!(tail[2].tool_call_id.as_deref(), Some("call-large"));
    }

    #[test]
    fn provider_normalization_falls_back_to_source_tail_and_drops_pending_call() {
        let source = vec![
            Message::user("latest request".into()),
            Message::assistant("finished".into()),
        ];
        let provider_output = vec![Message::system("provider summary".into())];

        let normalized = super::normalize_provider_compacted_history(provider_output, &source);

        assert_eq!(normalized.len(), 3);
        assert_eq!(normalized[0].content.as_text(), "provider summary");
        assert_eq!(normalized[1].content.as_text(), "latest request");
        assert_eq!(normalized[2].content.as_text(), "finished");

        let malformed = vec![Message::system("summary".into()), {
            let mut message = Message::assistant("pending".into());
            message.tool_calls = Some(vec![ToolCall::function("c1".into(), "run".into(), "{}".into())]);
            message
        }];
        let normalized = super::normalize_provider_compacted_history(malformed, &source);
        assert_eq!(normalized.len(), 3);
        assert_eq!(normalized[0].content.as_text(), "summary");
        assert_eq!(normalized[1].content.as_text(), "latest request");

        let partial = vec![
            Message::system("provider summary".into()),
            Message::user("older request".into()),
            Message::assistant("older result".into()),
            Message::user("latest request".into()),
            Message::assistant_with_tools(
                "calling".into(),
                vec![
                    ToolCall::function("c1".into(), "run".into(), "{}".into()),
                    ToolCall::function("c2".into(), "run".into(), "{}".into()),
                ],
            ),
            Message::tool_response("c1".into(), "partial result".into()),
        ];
        let normalized = super::normalize_provider_compacted_history(partial, &source);
        assert_eq!(normalized[0].content.as_text(), "provider summary");
        assert_eq!(normalized.last().unwrap().content.as_text(), "latest request");
        assert!(!normalized.iter().any(|message| message.tool_call_id.as_deref() == Some("c1")));
        assert!(!normalized.iter().any(|message| message.tool_calls.is_some()));
    }

    #[test]
    fn provider_normalization_preserves_duplicate_tagged_intents() {
        use vtcode_commons::message_metadata::MessageMetadata;

        let mut first = Message::user("same follow-up".into());
        first.metadata = Some(MessageMetadata::user_input(1, first.estimate_tokens()).with_intent_id("intent-1"));
        let mut second = Message::user("same follow-up".into());
        second.metadata = Some(MessageMetadata::user_input(2, second.estimate_tokens()).with_intent_id("intent-2"));
        let source = vec![first, second];

        let normalized = super::normalize_provider_compacted_history(vec![Message::system("summary".into())], &source);
        let intent_ids = normalized
            .iter()
            .filter_map(|message| message.metadata.as_ref().and_then(|metadata| metadata.intent_id()))
            .collect::<Vec<_>>();

        assert_eq!(intent_ids, vec!["intent-1", "intent-2"]);
    }

    /// History-growth verify-item: the local summarization prompt must exclude
    /// `Message.reasoning` / `reasoning_details` and serialize only the visible
    /// text content (`content.as_text()`). Reasoning traces are large and
    /// ephemeral; including them in every compaction summary would bloat the
    /// post-compaction context and re-inject stale chain-of-thought. The
    /// continuity tail preserves provider-protocol reasoning (Anthropic
    /// thinking signatures, OpenAI reasoning items) separately and is NOT
    /// summarized, so stripping reasoning here is safe and correct. This test
    /// pins that invariant so a change to `build_summary_prompt` cannot
    /// accidentally start including reasoning.
    #[test]
    fn build_summary_prompt_excludes_reasoning_traces() {
        use super::build_summary_prompt;

        let instructions = "Summarize the conversation.";
        // Assistant message carrying a reasoning trace alongside its visible
        // content. If `build_summary_prompt` ever reads `reasoning`, the
        // summary would contain "SECRET_REASONING" and "raw-only reasoning".
        let history = vec![
            Message::user("What is 2+2?".to_string()),
            Message::assistant("The answer is 4.".to_string())
                .with_reasoning(Some("SECRET_REASONING: I computed 2+2=4.".to_string())),
        ];

        let prompt = build_summary_prompt(&history, instructions);

        assert!(prompt.contains("The answer is 4."), "visible assistant text must appear in the summary prompt");
        assert!(
            !prompt.contains("SECRET_REASONING"),
            "Message.reasoning must NOT be included in the summary prompt -- \
             it would bloat every compaction pass with ephemeral chain-of-thought"
        );
        assert!(prompt.contains("Summarize the conversation."), "instructions must appear in the summary prompt");
    }

    #[test]
    fn coherence_force_keeps_tool_results_not_in_selection() {
        // Regression: `coherence_tool_call_pairs` force-keeps the Tool results
        // that follow a retained Assistant-with-tool_calls, but the previous
        // implementation derived its output from `selected` only, so those
        // force-kept messages silently vanished. A compacted history can then
        // carry an Assistant tool-call with no result, which providers reject.
        use super::coherence_tool_call_pairs;

        let history = vec![
            Message::user("check the code".to_string()),
            Message::assistant("Looking...".to_string()).with_tool_calls(vec![ToolCall {
                id: "call_1".to_string(),
                call_type: "function".to_string(),
                function: Some(crate::llm::provider::FunctionCall {
                    namespace: None,
                    name: "read_file".to_string(),
                    arguments: "{}".to_string(),
                }),
                text: None,
                thought_signature: None,
            }]),
            Message::tool_response("call_1".to_string(), "tool result that must survive".to_string()),
            Message::assistant("Done.".to_string()),
        ];

        // Selection keeps only the assistant-with-tool-calls turn, dropping the
        // following Tool result from the budget-driven selection.
        let selected = vec![(1, history[1].clone())];
        let result = coherence_tool_call_pairs(&history, &selected);

        let tool_results = result
            .iter()
            .filter(|(_, m)| m.role == MessageRole::Tool)
            .map(|(idx, m)| (*idx, m.content.as_text().to_string()))
            .collect::<Vec<_>>();

        assert_eq!(
            tool_results,
            vec![(2usize, "tool result that must survive".to_string())],
            "force-kept tool result must survive compaction even when not selected"
        );
        // History order must be preserved with the assistant preceding its result.
        let indices = result.iter().map(|(idx, _)| *idx).collect::<Vec<_>>();
        assert_eq!(indices, vec![1, 2]);
    }

    #[test]
    fn coherence_drops_orphaned_tool_results_of_unretained_assistant() {
        use super::coherence_tool_call_pairs;

        let history = vec![
            Message::user("check".to_string()),
            Message::assistant("Looking...".to_string()).with_tool_calls(vec![ToolCall {
                id: "call_1".to_string(),
                call_type: "function".to_string(),
                function: Some(crate::llm::provider::FunctionCall {
                    namespace: None,
                    name: "read_file".to_string(),
                    arguments: "{}".to_string(),
                }),
                text: None,
                thought_signature: None,
            }]),
            Message::tool_response("call_1".to_string(), "orphan result".to_string()),
        ];

        // The calling assistant was NOT retained; its orphaned Tool result must
        // be dropped so we never emit a result the model never saw a call for.
        let selected = vec![(0, history[0].clone())];
        let result = coherence_tool_call_pairs(&history, &selected);

        assert_eq!(result.len(), 1, "orphaned tool result must be dropped");
        assert_eq!(result[0].0, 0);
    }
}