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//! LLM-based session compaction.
//!
//! When a session's message count exceeds a threshold, the compactor
//! uses an LLM to summarize older messages into a concise summary,
//! keeping only the most recent messages intact. This prevents context
//! windows from growing unboundedly while preserving key information.
//!
//! Supports three summarization stages:
//! 1. Full single-pass summarization (fastest, best quality)
//! 2. Adaptive chunked summarization with merge (handles large histories)
//! 3. Minimal fallback without LLM (when summarization is unavailable)
use crate::llm_driver::{CompletionRequest, LlmDriver};
use crate::str_utils::safe_truncate_str;
use openfang_memory::session::Session;
use openfang_types::message::{ContentBlock, Message, MessageContent, Role};
use openfang_types::tool::ToolDefinition;
use serde::Serialize;
use std::sync::Arc;
use tracing::{info, warn};
/// Configuration for session compaction.
#[derive(Debug, Clone)]
pub struct CompactionConfig {
/// Compact when session message count exceeds this.
pub threshold: usize,
/// Number of recent messages to keep verbatim (not summarized).
pub keep_recent: usize,
/// Maximum tokens for the summary generation.
pub max_summary_tokens: u32,
/// Base ratio of messages to process per chunk (0.0-1.0).
pub base_chunk_ratio: f64,
/// Minimum chunk ratio (floor for adaptive computation).
pub min_chunk_ratio: f64,
/// Safety margin multiplier for token estimation inaccuracy.
pub safety_margin: f64,
/// Overhead tokens reserved for summarization prompt itself.
pub summarization_overhead_tokens: u32,
/// Maximum input chars per summarization chunk.
pub max_chunk_chars: usize,
/// Maximum retry attempts for summarization.
pub max_retries: u32,
/// Trigger compaction when estimated tokens exceed this fraction of context_window_tokens.
pub token_threshold_ratio: f64,
/// Model context window size in tokens.
pub context_window_tokens: usize,
}
impl Default for CompactionConfig {
fn default() -> Self {
Self {
threshold: 30,
keep_recent: 10,
max_summary_tokens: 1024,
base_chunk_ratio: 0.4,
min_chunk_ratio: 0.15,
safety_margin: 1.2,
summarization_overhead_tokens: 4096,
max_chunk_chars: 80_000,
max_retries: 3,
token_threshold_ratio: 0.7,
context_window_tokens: 200_000,
}
}
}
/// Result of a compaction operation.
#[derive(Debug)]
pub struct CompactionResult {
/// LLM-generated summary of the compacted messages.
pub summary: String,
/// Messages to keep (the most recent ones).
pub kept_messages: Vec<Message>,
/// Number of messages that were compacted (summarized).
pub compacted_count: usize,
/// Number of chunks used (1 = single-pass, >1 = chunked).
pub chunks_used: u32,
/// Whether fallback was used (LLM unavailable).
pub used_fallback: bool,
}
/// Check whether a session needs compaction (message-count trigger).
pub fn needs_compaction(session: &Session, config: &CompactionConfig) -> bool {
session.messages.len() > config.threshold
}
/// Estimate token count for a set of messages, optional system prompt, and tool definitions.
///
/// Uses the chars/4 heuristic — not exact, but good enough for budget gating.
pub fn estimate_token_count(
messages: &[Message],
system_prompt: Option<&str>,
tools: Option<&[openfang_types::tool::ToolDefinition]>,
) -> usize {
let mut chars: usize = 0;
// System prompt
if let Some(sp) = system_prompt {
chars += sp.len();
}
// Messages
for msg in messages {
chars += msg.content.text_length();
// Per-message overhead (role label, framing tokens)
chars += 16;
}
// Tool definitions (JSON schema is the biggest contributor)
if let Some(tool_defs) = tools {
for tool in tool_defs {
chars += tool.name.len() + tool.description.len();
if let Ok(schema_str) = serde_json::to_string(&tool.input_schema) {
chars += schema_str.len();
}
}
}
// chars / 4 heuristic
chars / 4
}
/// Check whether estimated tokens exceed the compaction threshold.
///
/// Returns true if `estimated_tokens > context_window * token_threshold_ratio`.
pub fn needs_compaction_by_tokens(estimated_tokens: usize, config: &CompactionConfig) -> bool {
let threshold = (config.context_window_tokens as f64 * config.token_threshold_ratio) as usize;
estimated_tokens > threshold
}
// ---------------------------------------------------------------------------
// Context Report
// ---------------------------------------------------------------------------
/// Context window pressure level.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)]
#[serde(rename_all = "lowercase")]
pub enum ContextPressure {
/// < 50% usage
Low,
/// 50–70% usage
Medium,
/// 70–85% usage
High,
/// > 85% usage
Critical,
}
impl ContextPressure {
fn from_percent(pct: f64) -> Self {
if pct > 85.0 {
Self::Critical
} else if pct > 70.0 {
Self::High
} else if pct > 50.0 {
Self::Medium
} else {
Self::Low
}
}
/// CSS-friendly color name.
pub fn color(&self) -> &'static str {
match self {
Self::Low => "green",
Self::Medium => "yellow",
Self::High => "orange",
Self::Critical => "red",
}
}
}
/// Token breakdown by source.
#[derive(Debug, Clone, Serialize)]
pub struct ContextBreakdown {
pub system_prompt_tokens: usize,
pub message_tokens: usize,
pub tool_definition_tokens: usize,
}
/// Context window usage report.
#[derive(Debug, Clone, Serialize)]
pub struct ContextReport {
pub estimated_tokens: usize,
pub context_window: usize,
pub usage_percent: f64,
pub pressure: ContextPressure,
pub message_count: usize,
pub breakdown: ContextBreakdown,
pub recommendation: String,
}
/// Generate a context window usage report.
pub fn generate_context_report(
messages: &[Message],
system_prompt: Option<&str>,
tools: Option<&[ToolDefinition]>,
context_window: usize,
) -> ContextReport {
// Break down token estimates by source
let sp_tokens = system_prompt.map_or(0, |s| s.len() / 4);
let msg_tokens = {
let mut chars: usize = 0;
for msg in messages {
chars += msg.content.text_length() + 16;
}
chars / 4
};
let tool_tokens = tools.map_or(0, |defs| {
let mut chars: usize = 0;
for t in defs {
chars += t.name.len() + t.description.len();
if let Ok(s) = serde_json::to_string(&t.input_schema) {
chars += s.len();
}
}
chars / 4
});
let total = sp_tokens + msg_tokens + tool_tokens;
let cw = context_window.max(1);
let pct = (total as f64 / cw as f64 * 100.0).min(100.0);
let pressure = ContextPressure::from_percent(pct);
let recommendation = match pressure {
ContextPressure::Low => "Context usage is healthy.".to_string(),
ContextPressure::Medium => {
"Consider using /compact if the conversation grows longer.".to_string()
}
ContextPressure::High => {
"Context is getting full. Use /compact to summarize older messages.".to_string()
}
ContextPressure::Critical => {
"Context is nearly full! Use /compact or /new immediately.".to_string()
}
};
ContextReport {
estimated_tokens: total,
context_window: cw,
usage_percent: (pct * 10.0).round() / 10.0, // 1 decimal
pressure,
message_count: messages.len(),
breakdown: ContextBreakdown {
system_prompt_tokens: sp_tokens,
message_tokens: msg_tokens,
tool_definition_tokens: tool_tokens,
},
recommendation,
}
}
/// Format a context report as human-readable text with ASCII progress bar.
pub fn format_context_report(report: &ContextReport) -> String {
let bar_len: usize = 20;
let filled = ((report.usage_percent / 100.0) * bar_len as f64).round() as usize;
let empty = bar_len.saturating_sub(filled);
let bar: String = std::iter::repeat_n('█', filled)
.chain(std::iter::repeat_n('░', empty))
.collect();
format!(
"**Context Usage:** {bar} {:.1}% ({} / {} tokens)\n\n\
**Breakdown:**\n\
- System prompt: ~{} tokens\n\
- Messages ({}): ~{} tokens\n\
- Tool definitions: ~{} tokens\n\n\
**Pressure:** {:?}\n\
**Recommendation:** {}",
report.usage_percent,
report.estimated_tokens,
report.context_window,
report.breakdown.system_prompt_tokens,
report.message_count,
report.breakdown.message_tokens,
report.breakdown.tool_definition_tokens,
report.pressure,
report.recommendation,
)
}
// ---------------------------------------------------------------------------
// Adaptive Chunking
// ---------------------------------------------------------------------------
/// Compute adaptive chunk ratio based on average message size.
///
/// Shorter messages get larger chunks (more context per summary).
/// Longer messages get smaller chunks (each message has more info to summarize).
fn compute_adaptive_chunk_ratio(messages: &[Message], config: &CompactionConfig) -> f64 {
if messages.is_empty() {
return config.base_chunk_ratio;
}
let avg_len = messages
.iter()
.map(|m| m.content.text_length())
.sum::<usize>() as f64
/ messages.len() as f64;
// Heuristic: longer messages → smaller ratio (fewer per chunk)
let ratio = if avg_len > 1000.0 {
config.min_chunk_ratio
} else if avg_len > 500.0 {
(config.base_chunk_ratio + config.min_chunk_ratio) / 2.0
} else {
config.base_chunk_ratio
};
ratio.clamp(config.min_chunk_ratio, config.base_chunk_ratio)
}
/// Check if a single message is oversized (> 50% of max_chunk_chars).
///
/// Oversized messages should be summarized individually rather than in chunks
/// to avoid exceeding context window limits.
fn is_oversized(message: &Message, config: &CompactionConfig) -> bool {
message.content.text_length() > config.max_chunk_chars / 2
}
/// Build conversation text from a slice of messages (block-aware).
///
/// Handles all content block types: text, tool use, tool result, image, unknown.
/// Oversized messages are truncated inline with a marker.
fn build_conversation_text(messages: &[Message], config: &CompactionConfig) -> String {
let mut conversation_text = String::new();
for msg in messages {
let role_label = match msg.role {
Role::User => "User",
Role::Assistant => "Assistant",
Role::System => "System",
};
// If a single message is oversized, truncate its contribution
let oversized = is_oversized(msg, config);
match &msg.content {
MessageContent::Text(s) => {
if !s.is_empty() {
if oversized {
let limit = config.max_chunk_chars / 4;
let truncated = if s.len() > limit {
format!("{}...[truncated from {} chars]", safe_truncate_str(s, limit), s.len())
} else {
s.clone()
};
conversation_text.push_str(&format!("{role_label}: {truncated}\n\n"));
} else {
conversation_text.push_str(&format!("{role_label}: {s}\n\n"));
}
}
}
MessageContent::Blocks(blocks) => {
for block in blocks {
match block {
ContentBlock::Text { text } => {
if !text.is_empty() {
if oversized && text.len() > config.max_chunk_chars / 4 {
let limit = config.max_chunk_chars / 4;
conversation_text.push_str(&format!(
"{role_label}: {}...[truncated from {} chars]\n\n",
safe_truncate_str(text, limit),
text.len()
));
} else {
conversation_text
.push_str(&format!("{role_label}: {text}\n\n"));
}
}
}
ContentBlock::ToolUse { name, input, .. } => {
let input_str = serde_json::to_string(input).unwrap_or_default();
let input_preview = if input_str.len() > 200 {
format!("{}...", safe_truncate_str(&input_str, 200))
} else {
input_str
};
conversation_text.push_str(&format!(
"[Used tool '{name}' with params: {input_preview}]\n\n"
));
}
ContentBlock::ToolResult {
content, is_error, ..
} => {
let status = if *is_error { "ERROR" } else { "OK" };
// Strip base64 blobs and injection markers before compaction
let cleaned = crate::session_repair::strip_tool_result_details(content);
let preview = if cleaned.len() > 2000 {
format!("{}...", safe_truncate_str(&cleaned, 2000))
} else {
cleaned
};
conversation_text
.push_str(&format!("[Tool result ({status}): {preview}]\n\n"));
}
ContentBlock::Image { media_type, .. } => {
conversation_text.push_str(&format!("[Image: {media_type}]\n\n"));
}
ContentBlock::Thinking { .. } => {}
ContentBlock::Unknown => {}
}
}
}
}
}
conversation_text
}
/// Summarize a slice of messages using the LLM.
///
/// Builds the conversation text, applies chunking limits, and calls the LLM
/// with a summarization prompt. Retries on transient failures.
async fn summarize_messages(
driver: Arc<dyn LlmDriver>,
model: &str,
messages: &[Message],
config: &CompactionConfig,
) -> Result<String, String> {
let mut conversation_text = build_conversation_text(messages, config);
// Truncate if exceeding max_chunk_chars (with safety margin)
let effective_max = (config.max_chunk_chars as f64 / config.safety_margin) as usize;
if conversation_text.len() > effective_max {
// Keep the tail (most recent) which is usually more important
conversation_text =
conversation_text[conversation_text.len() - effective_max..].to_string();
}
let summarize_prompt = format!(
"Summarize the following conversation preserving key facts, decisions, user preferences, \
and important context. Be concise but thorough. Output only the summary, no preamble.\n\n\
---\n{conversation_text}---"
);
let request = CompletionRequest {
model: model.to_string(),
messages: vec![Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::Text {
text: summarize_prompt,
}]),
}],
tools: vec![],
max_tokens: config.max_summary_tokens,
temperature: 0.3,
system: Some(
"You are a conversation summarizer. Produce a concise summary that captures \
all key facts, decisions, and context from the conversation."
.to_string(),
),
thinking: None,
};
// Retry logic for transient failures
let mut last_error = String::new();
for attempt in 0..config.max_retries {
match driver.complete(request.clone()).await {
Ok(response) => {
let summary = response.text();
if summary.is_empty() {
last_error = "LLM returned empty summary".to_string();
warn!(attempt, "Empty summary from LLM, retrying");
continue;
}
return Ok(summary);
}
Err(e) => {
last_error = format!("LLM summarization failed: {e}");
if attempt + 1 < config.max_retries {
warn!(attempt, error = %e, "Summarization attempt failed, retrying");
}
}
}
}
Err(last_error)
}
/// Summarize messages in adaptive chunks, then merge the per-chunk summaries.
///
/// Splits messages into chunks based on adaptive ratio (accounting for message size),
/// summarizes each chunk independently, then merges all chunk summaries with a final
/// LLM call into one cohesive summary.
async fn summarize_in_chunks(
driver: Arc<dyn LlmDriver>,
model: &str,
messages: &[Message],
config: &CompactionConfig,
) -> Result<String, String> {
let chunk_ratio = compute_adaptive_chunk_ratio(messages, config);
let chunk_size = (messages.len() as f64 * chunk_ratio).ceil() as usize;
let chunk_size = chunk_size.max(5); // minimum 5 messages per chunk
info!(
total = messages.len(),
chunk_size, chunk_ratio, "Starting chunked summarization"
);
let mut summaries = Vec::new();
let mut success_count = 0usize;
let mut last_chunk_error = String::new();
for (i, chunk) in messages.chunks(chunk_size).enumerate() {
match summarize_messages(driver.clone(), model, chunk, config).await {
Ok(summary) => {
info!(chunk = i, summary_len = summary.len(), "Chunk summarized");
summaries.push(summary);
success_count += 1;
}
Err(e) => {
// If a single chunk fails, note it and continue with remaining chunks.
// A partial summary is better than none.
warn!(chunk = i, error = %e, "Chunk summarization failed, skipping");
last_chunk_error = e;
summaries.push(format!(
"[Chunk {}: {} messages, summarization unavailable]",
i + 1,
chunk.len()
));
}
}
}
// If ALL chunks failed, propagate the error to trigger fallback
if success_count == 0 {
return Err(format!(
"All {} chunks failed to summarize: {last_chunk_error}",
summaries.len()
));
}
if summaries.is_empty() {
return Err("No chunks were summarized".to_string());
}
if summaries.len() == 1 {
return Ok(summaries.into_iter().next().unwrap());
}
// Merge summaries with another LLM call
let merge_prompt = format!(
"Merge these {} conversation summaries into one concise, coherent summary. \
Preserve all key facts, decisions, and context. Output only the merged summary.\n\n{}",
summaries.len(),
summaries
.iter()
.enumerate()
.map(|(i, s)| format!("--- Part {} ---\n{}", i + 1, s))
.collect::<Vec<_>>()
.join("\n\n")
);
let merge_request = CompletionRequest {
model: model.to_string(),
messages: vec![Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::Text { text: merge_prompt }]),
}],
tools: vec![],
max_tokens: config.max_summary_tokens,
temperature: 0.3,
system: Some(
"You are a conversation summarizer. Merge the provided partial summaries \
into a single cohesive summary."
.to_string(),
),
thinking: None,
};
match driver.complete(merge_request).await {
Ok(response) => {
let merged = response.text();
if merged.is_empty() {
// Fall back to concatenating the per-chunk summaries
Ok(summaries.join("\n\n"))
} else {
Ok(merged)
}
}
Err(e) => {
warn!(error = %e, "Merge summarization failed, concatenating chunks");
// Fallback: just concatenate the chunk summaries
Ok(summaries.join("\n\n"))
}
}
}
/// Compact a session by summarizing older messages with an LLM.
///
/// Takes all messages except the most recent `keep_recent` and uses a
/// multi-stage approach to produce a concise summary:
///
/// 1. **Full summarization**: tries to summarize all older messages in one pass
/// 2. **Chunked summarization**: splits into adaptive chunks, summarizes each,
/// then merges the chunk summaries
/// 3. **Minimal fallback**: if LLM is unavailable, produces a placeholder note
///
/// Returns the summary, the kept messages, and metadata about the operation.
pub async fn compact_session(
driver: Arc<dyn LlmDriver>,
model: &str,
session: &Session,
config: &CompactionConfig,
) -> Result<CompactionResult, String> {
let msg_count = session.messages.len();
if msg_count <= config.keep_recent {
return Ok(CompactionResult {
summary: String::new(),
kept_messages: session.messages.clone(),
compacted_count: 0,
chunks_used: 0,
used_fallback: false,
});
}
let split_at = msg_count.saturating_sub(config.keep_recent);
let to_compact = &session.messages[..split_at];
let kept = &session.messages[split_at..];
info!(
total = msg_count,
compacting = to_compact.len(),
keeping = kept.len(),
"Compacting session messages"
);
let kept_messages = kept.to_vec();
let compacted_count = to_compact.len();
// Stage 1: Try full single-pass summarization
match summarize_messages(driver.clone(), model, to_compact, config).await {
Ok(summary) => {
info!(
summary_len = summary.len(),
compacted = compacted_count,
"Session compaction complete (single-pass)"
);
return Ok(CompactionResult {
summary,
kept_messages,
compacted_count,
chunks_used: 1,
used_fallback: false,
});
}
Err(e) => {
warn!(error = %e, "Full summarization failed, trying chunked approach");
}
}
// Stage 2: Chunked summarization with adaptive ratio
match summarize_in_chunks(driver.clone(), model, to_compact, config).await {
Ok(summary) => {
let chunk_ratio = compute_adaptive_chunk_ratio(to_compact, config);
let chunk_size = (to_compact.len() as f64 * chunk_ratio).ceil() as usize;
let chunk_size = chunk_size.max(5);
let num_chunks = (to_compact.len() as f64 / chunk_size as f64).ceil() as u32;
info!(
summary_len = summary.len(),
compacted = compacted_count,
chunks = num_chunks,
"Session compaction complete (chunked)"
);
return Ok(CompactionResult {
summary,
kept_messages,
compacted_count,
chunks_used: num_chunks.max(1),
used_fallback: false,
});
}
Err(e) => {
warn!(error = %e, "Chunked summarization failed, using minimal fallback");
}
}
// Stage 3: Minimal fallback -- note what was compacted without LLM
let minimal = format!(
"[Session compacted: {} messages removed. Recent {} messages preserved. \
Summarization was unavailable.]",
to_compact.len(),
kept_messages.len()
);
warn!(
compacted = compacted_count,
"Using fallback compaction (no LLM summary)"
);
Ok(CompactionResult {
summary: minimal,
kept_messages,
compacted_count,
chunks_used: 0,
used_fallback: true,
})
}
#[cfg(test)]
mod tests {
use super::*;
use openfang_types::message::TokenUsage;
#[test]
fn test_needs_compaction_below_threshold() {
let session = Session {
id: openfang_types::agent::SessionId::new(),
agent_id: openfang_types::agent::AgentId::new(),
messages: vec![Message::user("hello")],
context_window_tokens: 0,
label: None,
};
let config = CompactionConfig::default();
assert!(!needs_compaction(&session, &config));
}
#[test]
fn test_needs_compaction_above_threshold() {
let messages: Vec<Message> = (0..100)
.map(|i| Message::user(format!("msg {i}")))
.collect();
let session = Session {
id: openfang_types::agent::SessionId::new(),
agent_id: openfang_types::agent::AgentId::new(),
messages,
context_window_tokens: 0,
label: None,
};
let config = CompactionConfig::default();
assert!(needs_compaction(&session, &config));
}
#[test]
fn test_compaction_config_defaults() {
let config = CompactionConfig::default();
assert_eq!(config.threshold, 30);
assert_eq!(config.keep_recent, 10);
assert_eq!(config.max_summary_tokens, 1024);
assert!((config.token_threshold_ratio - 0.7).abs() < f64::EPSILON);
assert_eq!(config.context_window_tokens, 200_000);
}
#[tokio::test]
async fn test_compact_session_few_messages() {
use crate::llm_driver::{CompletionResponse, LlmError};
use async_trait::async_trait;
struct FakeDriver;
#[async_trait]
impl LlmDriver for FakeDriver {
async fn complete(
&self,
_req: CompletionRequest,
) -> Result<CompletionResponse, LlmError> {
Ok(CompletionResponse {
content: vec![ContentBlock::Text {
text: "Summary of conversation".to_string(),
}],
stop_reason: openfang_types::message::StopReason::EndTurn,
tool_calls: vec![],
usage: TokenUsage {
input_tokens: 100,
output_tokens: 50,
},
})
}
}
let session = Session {
id: openfang_types::agent::SessionId::new(),
agent_id: openfang_types::agent::AgentId::new(),
messages: vec![Message::user("hello"), Message::assistant("hi")],
context_window_tokens: 0,
label: None,
};
let config = CompactionConfig {
threshold: 30,
keep_recent: 10,
max_summary_tokens: 1024,
..CompactionConfig::default()
};
// With only 2 messages and keep_recent=10, nothing should be compacted
let result = compact_session(Arc::new(FakeDriver), "test-model", &session, &config)
.await
.unwrap();
assert_eq!(result.compacted_count, 0);
assert_eq!(result.kept_messages.len(), 2);
assert_eq!(result.chunks_used, 0);
assert!(!result.used_fallback);
}
#[tokio::test]
async fn test_compact_includes_tool_calls() {
use crate::llm_driver::{CompletionResponse, LlmError};
use async_trait::async_trait;
struct FakeDriver;
#[async_trait]
impl LlmDriver for FakeDriver {
async fn complete(
&self,
req: CompletionRequest,
) -> Result<CompletionResponse, LlmError> {
// Verify the input includes tool call information
let input_text = req.messages[0].content.text_content();
assert!(
input_text.contains("web_search"),
"Should include tool name"
);
assert!(
input_text.contains("Tool result"),
"Should include tool result"
);
Ok(CompletionResponse {
content: vec![ContentBlock::Text {
text: "Summary with tools".to_string(),
}],
stop_reason: openfang_types::message::StopReason::EndTurn,
tool_calls: vec![],
usage: TokenUsage {
input_tokens: 100,
output_tokens: 50,
},
})
}
}
let mut messages: Vec<Message> = Vec::new();
// Add enough messages to trigger compaction (keep_recent = 5 for this test)
for _ in 0..8 {
messages.push(Message::user("Query"));
}
// Insert a tool use + result pair early in the history
messages[1] = Message {
role: Role::Assistant,
content: MessageContent::Blocks(vec![ContentBlock::ToolUse {
id: "tu-1".to_string(),
name: "web_search".to_string(),
input: serde_json::json!({"query": "test"}),
}]),
};
messages[2] = Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::ToolResult {
tool_use_id: "tu-1".to_string(),
content: "Search results here".to_string(),
is_error: false,
}]),
};
let session = Session {
id: openfang_types::agent::SessionId::new(),
agent_id: openfang_types::agent::AgentId::new(),
messages,
context_window_tokens: 0,
label: None,
};
let config = CompactionConfig {
threshold: 5,
keep_recent: 3,
max_summary_tokens: 512,
..CompactionConfig::default()
};
let result = compact_session(Arc::new(FakeDriver), "test-model", &session, &config)
.await
.unwrap();
assert!(result.compacted_count > 0);
assert!(result.summary.contains("tools"));
assert_eq!(result.chunks_used, 1);
assert!(!result.used_fallback);
}
#[test]
fn test_compact_truncates_large_tool_input() {
// Verify that the block-aware builder truncates large tool inputs
let large_input = serde_json::json!({"data": "x".repeat(500)});
let input_str = serde_json::to_string(&large_input).unwrap();
// The builder truncates to 200 chars
assert!(input_str.len() > 200);
// Just verify the truncation logic works correctly
let preview = if input_str.len() > 200 {
format!("{}...", safe_truncate_str(&input_str, 200))
} else {
input_str.clone()
};
assert!(preview.len() < input_str.len());
assert!(preview.ends_with("..."));
}
#[tokio::test]
async fn test_compact_session_many_messages() {
use crate::llm_driver::{CompletionResponse, LlmError};
use async_trait::async_trait;
struct FakeDriver;
#[async_trait]
impl LlmDriver for FakeDriver {
async fn complete(
&self,
_req: CompletionRequest,
) -> Result<CompletionResponse, LlmError> {
Ok(CompletionResponse {
content: vec![ContentBlock::Text {
text: "Summary: discussed topics 0 through 79".to_string(),
}],
stop_reason: openfang_types::message::StopReason::EndTurn,
tool_calls: vec![],
usage: TokenUsage {
input_tokens: 500,
output_tokens: 100,
},
})
}
}
let messages: Vec<Message> = (0..100)
.map(|i| Message::user(format!("Message about topic {i}")))
.collect();
let session = Session {
id: openfang_types::agent::SessionId::new(),
agent_id: openfang_types::agent::AgentId::new(),
messages,
context_window_tokens: 0,
label: None,
};
let config = CompactionConfig {
threshold: 30,
keep_recent: 10,
max_summary_tokens: 1024,
..CompactionConfig::default()
};
let result = compact_session(Arc::new(FakeDriver), "test-model", &session, &config)
.await
.unwrap();
assert_eq!(result.compacted_count, 90);
assert_eq!(result.kept_messages.len(), 10);
assert!(result.summary.contains("Summary"));
assert_eq!(result.chunks_used, 1);
assert!(!result.used_fallback);
}
// --- New tests ---
#[test]
fn test_adaptive_chunk_ratio_short_messages() {
let config = CompactionConfig::default();
let messages: Vec<Message> = (0..50).map(|i| Message::user(format!("msg {i}"))).collect();
let ratio = compute_adaptive_chunk_ratio(&messages, &config);
// Short messages (~6 chars) → should get the base (largest) ratio
assert!(
(ratio - config.base_chunk_ratio).abs() < f64::EPSILON,
"Short messages should use base ratio, got {ratio}"
);
}
#[test]
fn test_adaptive_chunk_ratio_long_messages() {
let config = CompactionConfig::default();
let messages: Vec<Message> = (0..20).map(|_| Message::user("x".repeat(1500))).collect();
let ratio = compute_adaptive_chunk_ratio(&messages, &config);
// Long messages (1500 chars) → should use min ratio
assert!(
(ratio - config.min_chunk_ratio).abs() < f64::EPSILON,
"Long messages should use min ratio, got {ratio}"
);
}
#[test]
fn test_adaptive_chunk_ratio_medium_messages() {
let config = CompactionConfig::default();
let messages: Vec<Message> = (0..20).map(|_| Message::user("y".repeat(700))).collect();
let ratio = compute_adaptive_chunk_ratio(&messages, &config);
let expected = (config.base_chunk_ratio + config.min_chunk_ratio) / 2.0;
assert!(
(ratio - expected).abs() < f64::EPSILON,
"Medium messages should use middle ratio, got {ratio}"
);
}
#[test]
fn test_adaptive_chunk_ratio_empty() {
let config = CompactionConfig::default();
let messages: Vec<Message> = vec![];
let ratio = compute_adaptive_chunk_ratio(&messages, &config);
assert!(
(ratio - config.base_chunk_ratio).abs() < f64::EPSILON,
"Empty messages should default to base ratio"
);
}
#[test]
fn test_oversized_message_detection() {
let config = CompactionConfig::default();
// max_chunk_chars default is 80_000, so threshold is 40_000
let small_msg = Message::user("short");
assert!(!is_oversized(&small_msg, &config));
let large_msg = Message::user("x".repeat(50_000));
assert!(is_oversized(&large_msg, &config));
// Boundary: exactly at threshold
let boundary_msg = Message::user("x".repeat(40_000));
assert!(!is_oversized(&boundary_msg, &config));
let just_over = Message::user("x".repeat(40_001));
assert!(is_oversized(&just_over, &config));
}
#[test]
fn test_compaction_config_new_defaults() {
let config = CompactionConfig::default();
assert_eq!(config.threshold, 30);
assert_eq!(config.keep_recent, 10);
assert_eq!(config.max_summary_tokens, 1024);
assert!((config.base_chunk_ratio - 0.4).abs() < f64::EPSILON);
assert!((config.min_chunk_ratio - 0.15).abs() < f64::EPSILON);
assert!((config.safety_margin - 1.2).abs() < f64::EPSILON);
assert_eq!(config.summarization_overhead_tokens, 4096);
assert_eq!(config.max_chunk_chars, 80_000);
assert_eq!(config.max_retries, 3);
assert!((config.token_threshold_ratio - 0.7).abs() < f64::EPSILON);
assert_eq!(config.context_window_tokens, 200_000);
}
#[tokio::test]
async fn test_fallback_on_llm_failure() {
use crate::llm_driver::{CompletionResponse, LlmError};
use async_trait::async_trait;
struct FailingDriver;
#[async_trait]
impl LlmDriver for FailingDriver {
async fn complete(
&self,
_req: CompletionRequest,
) -> Result<CompletionResponse, LlmError> {
Err(LlmError::Http("connection refused".to_string()))
}
}
let messages: Vec<Message> = (0..30)
.map(|i| Message::user(format!("Message {i}")))
.collect();
let session = Session {
id: openfang_types::agent::SessionId::new(),
agent_id: openfang_types::agent::AgentId::new(),
messages,
context_window_tokens: 0,
label: None,
};
let config = CompactionConfig {
threshold: 10,
keep_recent: 5,
max_summary_tokens: 512,
max_retries: 1, // fast failure
..CompactionConfig::default()
};
let result = compact_session(Arc::new(FailingDriver), "test-model", &session, &config)
.await
.unwrap();
assert!(result.used_fallback, "Should have used fallback");
assert_eq!(result.chunks_used, 0, "Fallback uses 0 chunks");
assert!(
result.summary.contains("Summarization was unavailable"),
"Fallback summary should indicate unavailability"
);
assert!(
result.summary.contains("25 messages removed"),
"Should state how many messages removed, got: {}",
result.summary
);
assert_eq!(result.compacted_count, 25);
assert_eq!(result.kept_messages.len(), 5);
}
#[tokio::test]
async fn test_chunked_summarization_splits_correctly() {
use crate::llm_driver::{CompletionResponse, LlmError};
use async_trait::async_trait;
use std::sync::atomic::{AtomicU32, Ordering};
static CALL_COUNT: AtomicU32 = AtomicU32::new(0);
struct CountingDriver;
#[async_trait]
impl LlmDriver for CountingDriver {
async fn complete(
&self,
_req: CompletionRequest,
) -> Result<CompletionResponse, LlmError> {
let n = CALL_COUNT.fetch_add(1, Ordering::SeqCst);
Ok(CompletionResponse {
content: vec![ContentBlock::Text {
text: format!("Chunk summary {n}"),
}],
stop_reason: openfang_types::message::StopReason::EndTurn,
tool_calls: vec![],
usage: TokenUsage {
input_tokens: 50,
output_tokens: 20,
},
})
}
}
// Reset counter
CALL_COUNT.store(0, Ordering::SeqCst);
let messages: Vec<Message> = (0..20)
.map(|i| Message::user(format!("Message {i}")))
.collect();
let config = CompactionConfig::default();
let result =
summarize_in_chunks(Arc::new(CountingDriver), "test-model", &messages, &config)
.await
.unwrap();
let calls = CALL_COUNT.load(Ordering::SeqCst);
// With base_chunk_ratio=0.4, chunk_size = ceil(20*0.4) = 8, so 3 chunks + 1 merge = 4 calls
assert!(
calls >= 2,
"Should have made multiple LLM calls for chunked summary, got {calls}"
);
assert!(!result.is_empty(), "Should produce a summary");
}
#[test]
fn test_compaction_result_new_fields() {
let result = CompactionResult {
summary: "test".to_string(),
kept_messages: vec![],
compacted_count: 10,
chunks_used: 3,
used_fallback: false,
};
assert_eq!(result.chunks_used, 3);
assert!(!result.used_fallback);
let fallback_result = CompactionResult {
summary: "fallback".to_string(),
kept_messages: vec![],
compacted_count: 5,
chunks_used: 0,
used_fallback: true,
};
assert_eq!(fallback_result.chunks_used, 0);
assert!(fallback_result.used_fallback);
}
#[test]
fn test_build_conversation_text_handles_all_blocks() {
let config = CompactionConfig::default();
let messages = vec![
Message::user("Hello"),
Message {
role: Role::Assistant,
content: MessageContent::Blocks(vec![
ContentBlock::Text {
text: "Let me search".to_string(),
},
ContentBlock::ToolUse {
id: "tu-1".to_string(),
name: "web_search".to_string(),
input: serde_json::json!({"query": "rust"}),
},
]),
},
Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::ToolResult {
tool_use_id: "tu-1".to_string(),
content: "Results found".to_string(),
is_error: false,
}]),
},
Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::Image {
media_type: "image/png".to_string(),
data: "base64data".to_string(),
}]),
},
];
let text = build_conversation_text(&messages, &config);
assert!(text.contains("User: Hello"));
assert!(text.contains("Assistant: Let me search"));
assert!(text.contains("web_search"));
assert!(text.contains("Tool result (OK)"));
assert!(text.contains("[Image: image/png]"));
}
#[test]
fn test_build_conversation_text_truncates_oversized() {
let config = CompactionConfig {
max_chunk_chars: 1000, // small limit for testing
..CompactionConfig::default()
};
let large_msg = Message::user("x".repeat(2000));
let messages = vec![large_msg];
let text = build_conversation_text(&messages, &config);
// Should be truncated since 2000 > 1000/2 = 500 (oversized threshold)
assert!(
text.contains("truncated from"),
"Oversized message should be truncated, got: {}",
&text[..text.len().min(200)]
);
}
#[test]
fn test_estimate_token_count_basic() {
let messages = vec![
Message::user("Hello world"), // 11 chars + 16 overhead = 27
Message::assistant("Hi there"), // 8 chars + 16 overhead = 24
];
let tokens = estimate_token_count(&messages, None, None);
// (11 + 16 + 8 + 16) / 4 = 12 (approx)
assert!(tokens > 0);
assert!(tokens < 100);
}
#[test]
fn test_estimate_token_count_with_system_prompt() {
let messages = vec![Message::user("hi")];
let system = "You are a helpful assistant. ".repeat(100); // ~2800 chars
let tokens_without = estimate_token_count(&messages, None, None);
let tokens_with = estimate_token_count(&messages, Some(&system), None);
assert!(tokens_with > tokens_without);
}
#[test]
fn test_estimate_token_count_with_tools() {
use openfang_types::tool::ToolDefinition;
let messages = vec![Message::user("hi")];
let tools = vec![ToolDefinition {
name: "web_search".into(),
description: "Search the web for information".into(),
input_schema: serde_json::json!({"type": "object", "properties": {"query": {"type": "string"}}}),
}];
let tokens_without = estimate_token_count(&messages, None, None);
let tokens_with = estimate_token_count(&messages, None, Some(&tools));
assert!(tokens_with > tokens_without);
}
#[test]
fn test_needs_compaction_by_tokens_below() {
let config = CompactionConfig::default();
// 70% of 200_000 = 140_000
assert!(!needs_compaction_by_tokens(100_000, &config));
}
#[test]
fn test_needs_compaction_by_tokens_above() {
let config = CompactionConfig::default();
// 70% of 200_000 = 140_000
assert!(needs_compaction_by_tokens(150_000, &config));
}
#[test]
fn test_context_pressure_from_percent() {
assert_eq!(ContextPressure::from_percent(30.0), ContextPressure::Low);
assert_eq!(ContextPressure::from_percent(55.0), ContextPressure::Medium);
assert_eq!(ContextPressure::from_percent(75.0), ContextPressure::High);
assert_eq!(
ContextPressure::from_percent(90.0),
ContextPressure::Critical
);
}
#[test]
fn test_generate_context_report_basic() {
let messages = vec![Message::user("Hello world"), Message::assistant("Hi there")];
let report = generate_context_report(&messages, Some("You are helpful."), None, 200_000);
assert!(report.estimated_tokens > 0);
assert!(report.usage_percent < 1.0); // tiny messages
assert_eq!(report.pressure, ContextPressure::Low);
assert_eq!(report.message_count, 2);
assert!(report.breakdown.system_prompt_tokens > 0);
assert!(report.breakdown.message_tokens > 0);
}
#[test]
fn test_generate_context_report_critical() {
// Create enough messages to push past 85%
let big_msg = "x".repeat(800_000); // 200K tokens at chars/4
let messages = vec![Message::user(big_msg)];
let report = generate_context_report(&messages, None, None, 200_000);
assert_eq!(report.pressure, ContextPressure::Critical);
assert!(report.usage_percent > 85.0);
}
#[test]
fn test_format_context_report() {
let messages = vec![Message::user("hi")];
let report = generate_context_report(&messages, Some("system"), None, 200_000);
let formatted = format_context_report(&report);
assert!(formatted.contains("Context Usage"));
assert!(formatted.contains("Breakdown"));
assert!(formatted.contains("Pressure"));
}
#[test]
fn test_compaction_strips_base64_blobs() {
let config = CompactionConfig::default();
let blob = "A".repeat(2000);
let tool_content = format!("result: {blob}");
let messages = vec![Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::ToolResult {
tool_use_id: "t1".to_string(),
content: tool_content,
is_error: false,
}]),
}];
let text = build_conversation_text(&messages, &config);
// The base64 blob should be stripped/replaced by session_repair
assert!(text.contains("[base64 blob"));
assert!(!text.contains(&"A".repeat(2000)));
}
#[test]
fn test_compaction_applies_2k_cap() {
let config = CompactionConfig::default();
// Create a tool result larger than 2K but without base64 blobs
let large_result = "word ".repeat(500); // ~2500 chars of non-base64 text
let messages = vec![Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::ToolResult {
tool_use_id: "t2".to_string(),
content: large_result,
is_error: false,
}]),
}];
let text = build_conversation_text(&messages, &config);
// Should be capped at ~2000 chars (plus the "..." suffix)
let result_part = text.split("[Tool result (OK): ").nth(1).unwrap_or("");
// The result_part includes trailing "]\n\n", so just check it's under 2100
assert!(
result_part.len() < 2100,
"result_part len = {}",
result_part.len()
);
}
#[test]
fn test_compaction_short_results_unchanged() {
let config = CompactionConfig::default();
let short_result = "Success: 42 records processed";
let messages = vec![Message {
role: Role::User,
content: MessageContent::Blocks(vec![ContentBlock::ToolResult {
tool_use_id: "t3".to_string(),
content: short_result.to_string(),
is_error: false,
}]),
}];
let text = build_conversation_text(&messages, &config);
assert!(text.contains(short_result));
}
}