pond-db 0.11.0

Lossless storage and hybrid search for sessions from any AI agent client
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
//! The HTTP+JSON and stdio-MCP transports: thin adapters over the shared wire
//! handlers. Both transports dispatch to the same handler functions - no
//! per-transport behavior divergence.
//!
//! HTTP exposes `POST /v1/search`, `POST /v1/get`, and `POST /v1/ingest`. MCP
//! exposes `pond_search` / `pond_get` (the kb-parity surface) plus
//! `pond_sql_query` (read-only SQL); ingest stays HTTP-only and CLI-only.

use std::sync::Arc;

use crate::{config::SearchConfig, embed::LazyEmbedder, sessions::Store};

/// Shared state handed to both transports. `embedder` holds a lazy handle:
/// the model isn't loaded until the first vector search asks for it, so
/// `pond mcp` idles at ~50 MB resident and only pays the ~600 MB load cost on
/// the first query that needs it (spec.md#search opt-in).
#[derive(Clone)]
pub struct AppState {
    pub store: Arc<Store>,
    pub embedder: Arc<LazyEmbedder>,
    pub search: SearchConfig,
}

pub mod http {
    //! axum HTTP+JSON server: `POST /v1/search`, `POST /v1/get`, and the `/mcp`
    //! route carrying rmcp's streamable-HTTP MCP transport.

    use std::net::{IpAddr, SocketAddr};

    use anyhow::Context;
    use axum::{
        Json, Router,
        extract::{DefaultBodyLimit, State},
        http::{HeaderValue, StatusCode},
        response::{IntoResponse, Response},
        routing::post,
    };
    use rmcp::transport::streamable_http_server::{
        StreamableHttpServerConfig, StreamableHttpService, session::local::LocalSessionManager,
    };
    use tokio::net::TcpListener;

    use super::AppState;
    use crate::{
        handlers::{pond_get, pond_ingest, pond_search},
        wire::{
            ErrorCode, GetEnvelope, GetRequest, IngestEnvelope, IngestRequest, SearchEnvelope,
            SearchRequest, default_namespace, new_request_id,
        },
    };

    /// HTTP body cap for `POST /v1/*` JSON handlers (spec.md#protocol): 8 MB.
    /// Replaces axum's 2 MB default - that default is more restrictive than the
    /// design's intent and would surface oversize ingests as a generic 413
    /// instead of pond's typed `validation_failed`.
    pub const HTTP_BODY_LIMIT_BYTES: usize = 8 * 1024 * 1024;

    /// Build the axum router: the `/v1/*` JSON handlers plus the nested `/mcp`
    /// streamable-HTTP MCP service. Public so the integration test can drive it
    /// without binding a socket.
    pub fn router(state: AppState) -> Router {
        let mcp_state = state.clone();
        let mcp = StreamableHttpService::new(
            move || Ok(super::mcp::PondMcp::new(mcp_state.clone())),
            LocalSessionManager::default().into(),
            StreamableHttpServerConfig::default(),
        );
        Router::new()
            .route("/v1/search", post(search))
            .route("/v1/get", post(get))
            .route("/v1/ingest", post(ingest))
            .layer(DefaultBodyLimit::max(HTTP_BODY_LIMIT_BYTES))
            .with_state(state)
            .nest_service("/mcp", mcp)
    }

    /// Bind and serve until ctrl-c. `--port 0` selects an OS-assigned free port;
    /// an unspecified host (`0.0.0.0` / `::`) logs a security notice because the
    /// personal pond is single-user and LAN exposure is opt-in (spec.md#scope).
    pub async fn serve(state: AppState, host: String, port: u16) -> anyhow::Result<()> {
        let ip: IpAddr = host
            .parse()
            .with_context(|| format!("invalid --host {host:?}"))?;
        if ip.is_unspecified() {
            tracing::warn!(
                %host,
                "binding to an unspecified address exposes pond on the LAN; \
                 the personal pond is single-user"
            );
        }
        let listener = TcpListener::bind(SocketAddr::new(ip, port))
            .await
            .with_context(|| format!("failed to bind {host}:{port}"))?;
        let local = listener
            .local_addr()
            .context("failed to read bound address")?;
        tracing::info!(%local, "pond serve listening (HTTP /v1/*, MCP /mcp)");
        axum::serve(listener, router(state))
            .with_graceful_shutdown(shutdown_signal())
            .await
            .context("axum server error")
    }

    async fn shutdown_signal() {
        let _ = tokio::signal::ctrl_c().await;
    }

    async fn search(
        State(state): State<AppState>,
        Json(mut request): Json<SearchRequest>,
    ) -> Response {
        request.namespace.get_or_insert_with(default_namespace);
        let envelope = pond_search(&state.store, &state.embedder, request, &state.search).await;
        let status = match &envelope {
            SearchEnvelope::Success(_) => StatusCode::OK,
            SearchEnvelope::Error(error) => status_for(&error.error.code),
        };
        with_request_id((status, Json(envelope)).into_response())
    }

    async fn get(State(state): State<AppState>, Json(mut request): Json<GetRequest>) -> Response {
        request.namespace.get_or_insert_with(default_namespace);
        let envelope = pond_get(&state.store, request).await;
        let status = match &envelope {
            GetEnvelope::Success(_) => StatusCode::OK,
            GetEnvelope::Error(error) => status_for(&error.error.code),
        };
        with_request_id((status, Json(envelope)).into_response())
    }

    async fn ingest(
        State(state): State<AppState>,
        Json(mut request): Json<IngestRequest>,
    ) -> Response {
        request.namespace.get_or_insert_with(default_namespace);
        let envelope = pond_ingest(&state.store, request).await;
        // Per-row errors in `results[]` are not request-level failures, so
        // the envelope success path always returns 200; only transport-level
        // failures (validation_failed, namespace_unknown, etc.) map to 4xx/5xx.
        let status = match &envelope {
            IngestEnvelope::Success(_) => StatusCode::OK,
            IngestEnvelope::Error(error) => status_for(&error.error.code),
        };
        with_request_id((status, Json(envelope)).into_response())
    }

    fn with_request_id(mut response: Response) -> Response {
        if let Ok(value) = HeaderValue::from_str(&new_request_id()) {
            response.headers_mut().insert("x-pond-request-id", value);
        }
        response
    }

    /// Map a wire error code to an HTTP status. The envelope body still carries
    /// the full typed error; the status is the coarse signal.
    fn status_for(code: &ErrorCode) -> StatusCode {
        match code {
            ErrorCode::ValidationFailed
            | ErrorCode::VersionUnsupported
            | ErrorCode::NamespaceUnknown => StatusCode::BAD_REQUEST,
            ErrorCode::NotFound => StatusCode::NOT_FOUND,
            ErrorCode::Conflict => StatusCode::CONFLICT,
            ErrorCode::StorageUnavailable => StatusCode::SERVICE_UNAVAILABLE,
            ErrorCode::Internal => StatusCode::INTERNAL_SERVER_ERROR,
        }
    }
}

pub mod mcp {
    //! The rmcp MCP layer: `pond_search` / `pond_get` / `pond_sql_query` tools
    //! and `schema://pond` / `schema://pond-sql` / `stats://pond` (plus
    //! `pond-sql-export://` export artifacts) resources, transport-agnostic.
    //! Mounted on stdio (via `pond mcp`) and on the `/mcp` HTTP route (via
    //! `pond serve`).

    use anyhow::Context;
    use base64::{Engine, engine::general_purpose::STANDARD};
    use rmcp::{
        ErrorData, RoleServer, ServerHandler, ServiceExt,
        handler::server::{router::tool::ToolRouter, wrapper::Parameters},
        model::{
            AnnotateAble, CallToolResult, Content, ErrorCode as JsonRpcErrorCode, Implementation,
            ListResourcesResult, ListToolsResult, Meta, PaginatedRequestParams, RawResource,
            ReadResourceRequestParams, ReadResourceResult, ResourceContents, ServerCapabilities,
            ServerInfo,
        },
        schemars,
        service::RequestContext,
        tool, tool_handler, tool_router,
        transport::stdio,
    };
    use serde::Deserialize;
    use uuid::Uuid;

    use super::AppState;
    use crate::{
        PROTOCOL_VERSION,
        handlers::pond_get as run_get,
        handlers::pond_search as run_search,
        sql,
        substrate::Table,
        wire::{
            ErrorCode as WireErrorCode, ErrorEnvelope, GetEnvelope, GetRequest, ProjectFilter,
            SearchEnvelope, SearchFilters, SearchModeWire, SearchRequest, SessionFrom, SortBy,
            default_namespace,
        },
    };

    /// Static documentation served as the `schema://pond` resource. Detail
    /// agents load on demand; the per-tool descriptions below stay tight.
    const SCHEMA_DOC: &str = "\
pond_search params: query (semantic - concepts, not project names), mode \
(vector default - matches on meaning | fts - matches exact whole words, BM25), \
sort_by (relevance default | recency), limit (returned sessions; default 10, \
max 200 - also the want-more knob, there is no pagination), project (path \
substring), session_id (exact session match - search within one session), \
from_date / to_date (YYYY-MM-DD). Subagents are excluded; reach them via \
pond_sql_query (parent_session_id).

pond_search response: a transcript. The first line states totals \
(`matched_total` is the message count before `limit` and byte-budget \
truncation), then results are grouped by session, best session first; within \
a session, matching messages are newest-first. Each hit is a `--- [n] score | \
role | time | message_id | project | agent | session ---` rule followed by its \
matched text (a ~600-char indexed window). `score` is in [0.0, 1.0] within one \
response (raw cosine for vector, normalized BM25 for fts). vector relevance \
ordering adds a gentle recency tiebreaker; sort_by=recency orders strictly \
newest-first and the response labels itself so. `has_more` warns the ranked \
set was cut by `limit` or the byte budget - raise `limit` to see the rest.

pond_search multilingual: pond's embedder (multilingual-e5-small) is trained \
for cross-lingual retrieval, so a vector query in language A can match indexed \
text in language B. fts is word-tokenized (the `simple` tokenizer with English \
stemming) and matches surface words only, so it stays within one language.

pond_get: message_id (the target message, marked `>`, plus message_context_before \
/ message_context_after conversational siblings each side, both default 3, like \
grep -B/-A) OR session_id (the whole session, conversational view). Output is a \
transcript - each message is a `--- [n] role | time | message_id ---` rule, \
then its text/content as real lines, then parts (`-> name [call_id]` tool \
call, `<- name [call_id] (ok|failed)` result). Session mode renders text plus \
one-line tool refs; fetch a message by message_id to expand full part bodies \
(reasoning included). session_limit defaults to 20; session_from=start|end picks \
which end the first page reads from (end = recent tail, e.g. post-compaction \
recovery). Bounded by a size budget: page with session_after_message_id (forward) \
or session_before_message_id (backward) using the id a page marker shows. A \
whole-session response also lists the session's subagents (each stored as its own \
session) in a footer; pass a listed id back as session_id to open it. Not for \
bulk export - use `pond copy --to <file>`.";

    /// Static documentation served as the `schema://pond-sql` resource: the
    /// table/column schema, dialect, function set, output modes, pagination
    /// pattern, drilling pattern, and worked examples for `pond_sql_query`.
    /// Loaded on demand so the tool description stays tight.
    ///
    /// TODO(#47): when the lance v8 FM-Index on parts.variant_data lands,
    /// tool-body substring search becomes `contains(variant_data, 'needle')`;
    /// update the routing guidance below (drop the "Never LIKE over parts ...
    /// no substring index (yet)" framing) and the timeout message in
    /// src/sql.rs.
    const SQL_SCHEMA_DOC: &str = "\
pond_sql_query runs ONE read-only SELECT (DataFusion SQL, PostgreSQL-compatible) \
over three registered tables. Read-only is hard-enforced: anything other than a \
single SELECT/WITH (or EXPLAIN of one) is rejected (no INSERT/UPDATE/DELETE/\
CREATE/DROP/COPY/SET).

Routing - pick the right surface before writing SQL:
- counts, group-by, time buckets, joins over metadata -> this tool, on \
messages/sessions.
- which tools ran / failed, tool params -> this tool, on parts (type = \
'tool_call' / 'tool_result'); worked example below.
- find text in conversations -> WHERE contains_tokens(search_text, '...') to \
filter, FROM fts('messages', ...) to rank (both below), or pond_search for \
meaning-based recall. Never LIKE over parts - tool bodies are JSON with no \
substring index (yet), and the conversational text is messages.search_text.
- read a transcript (a session, a message with context) -> pond_get, not SQL.

Tables and columns:
- messages(session_id text, message_id text, timestamp timestamp(us, UTC), role \
text {user|assistant|system|tool}, source_agent text, project text, content text \
NULL [system-role messages only], search_text text NULL [the conversational text - \
null for system/tool messages], embedding_model text NULL, options json). The \
embedding `vector` column exists but is never returned (omitted from results) and \
explicit projection of it is rejected; you may still filter on it in WHERE, e.g. \
`vector IS NOT NULL`. For semantic search, use pond_search.
- sessions(session_id text, parent_session_id text NULL, parent_message_id text \
NULL, source_agent text, created_at timestamp(us, UTC), project text, options json).
- parts(session_id text, message_id text, id text, ordinal int, type text \
{text|reasoning|file|tool_call|tool_result|tool_approval_request|\
tool_approval_response - exact strings, underscores not hyphens}, provenance \
text {conversational|injected}, variant_data json, options json). The verbatim \
part body lives in `variant_data`; its fields follow the part type, e.g. \
tool_call carries {call_id, name, params}, tool_result carries {call_id, name, \
is_failure, result}, text/reasoning carry {text}. FilePart binary payloads are \
not exposed in SQL.
Enum literals matter: a wrong value (e.g. 'tool-call') is valid SQL and silently \
returns zero rows. Discovery from SQL works too: SELECT table_name, column_name, \
data_type FROM information_schema.columns.

Join keys: messages.session_id = sessions.session_id; parts.session_id = \
messages.session_id AND parts.message_id = messages.message_id. Subagents are \
sessions whose source_agent matches '%/%' (e.g. 'claude-code/general-purpose').

Indexed (fast) filter columns: messages.session_id / source_agent; \
parts.session_id / message_id; sessions.session_id. \
Prefer equality/range predicates on these. Known limitation: prefix LIKE ('x%') and starts_with() FAIL \
on bitmap-indexed columns (messages.source_agent) with \"LIKE \
prefix queries are not supported for bitmap indexes\". Workarounds: equality, \
split_part(source_agent, '/', 1) = 'claude-code', or an infix pattern \
(LIKE '%/%' is fine - leading-wildcard patterns are not pushed to the index).

JSON columns (options, variant_data) are binary JSONB. Rules:
- NEVER CAST a JSON column (`variant_data::text` is rejected at plan time - the \
binary encoding can otherwise silently render as garbage). Stringify with \
json_extract(col, '$').
- A leading-wildcard LIKE over the whole document \
(`json_extract(variant_data, '$') LIKE '%...%'`) is rejected at plan time: it \
stringifies and scans every row and never finishes over parts. Match a single \
field (`json_extract(variant_data, '$.field') LIKE '...'`), scope to one session, \
or use contains_tokens for conversational text. (Substring search over tool \
bodies arrives with the FM-Index, #47.)
- json_extract(col, '$.a.b') takes a full JSONPath and returns JSON text of ANY \
value (objects/arrays serialize) - the right call for deeply nested or mixed-type \
fields, e.g. json_extract(variant_data, '$.params.command').
- json_get_string|json_get_int|json_get_float|json_get_bool(col, 'key', ...) walk \
a key path - json_get_string(options, 'anthropic', 'model') - array steps by \
numeric index. json_get_string serializes non-string values; the typed getters \
return NULL on a non-coercible value.
- json_get(col, 'key') returns JSONB for chaining: \
json_get_string(json_get(variant_data, 'params'), 'command').
- Also: json_array_contains(col, 'key', value), json_array_length(col, 'key').

Worked example - tool usage and failure rates over the last week:

  SELECT json_get_string(c.variant_data, 'name') AS tool,
         COUNT(*) AS calls,
         SUM(CASE WHEN json_get_bool(r.variant_data, 'is_failure') THEN 1 \
ELSE 0 END) AS failures
  FROM parts c
  JOIN messages m ON m.session_id = c.session_id AND m.message_id = c.message_id
  LEFT JOIN parts r ON r.session_id = c.session_id
   AND r.type = 'tool_result'
   AND json_get_string(r.variant_data, 'call_id') = \
json_get_string(c.variant_data, 'call_id')
  WHERE c.type = 'tool_call' AND m.timestamp >= now() - INTERVAL '7 days'
  GROUP BY tool ORDER BY calls DESC;

Full-text search in SQL is a pair - filter form and ranked form:
- Filtering (WHERE): contains_tokens(search_text, 'word1 word2') - true when the \
text contains ALL the words (split on punctuation/whitespace, case-sensitive \
tokens); accelerated by the FTS index. The right tool for exact strings, \
identifiers, and error messages - compose freely with other predicates: \
SELECT message_id FROM messages WHERE contains_tokens(search_text, 'OCC retry') \
AND project LIKE '%pond%'.
- Ranking (FROM): the fts() table function returns matches plus `_score` (BM25 \
relevance, a regular projectable column): SELECT message_id, _score, search_text \
FROM fts('messages', '{\"match\":{\"column\":\"search_text\",\"terms\":\"...\"}}') \
ORDER BY _score DESC - compose with WHERE/JOIN/GROUP BY around it. AND semantics: \
add \"operator\":\"And\" to the match; \"boolean\" queries (must/should/must_not \
over match clauses) also work. \"phrase\" queries are unavailable (index built \
without positions) - use contains_tokens or match + operator And, optionally with \
LIKE post-filters, for exact substrings.
fts() in WHERE is a plan-time error that points back here. Unlike pond_search, \
both forms cover subagent sessions (filter them out with WHERE NOT (source_agent \
LIKE '%/%') if unwanted). Vector/semantic search is NOT available in SQL; use \
pond_search for that.

Function quick-reference (exact DataFusion names so the model doesn't have to \
guess):
- aggregates: count, count(distinct ...), sum, avg, min, max, any_value, stddev, \
median, approx_distinct, approx_percentile_cont, array_agg, string_agg
- date/time: now(), date_trunc('day'|'hour'|'minute'|..., ts), date_part('year'|..., \
ts), date_bin(interval, ts, origin), to_char(ts, fmt), to_timestamp(text), \
extract(field FROM ts), age(t1, t2)
- intervals: `INTERVAL '7 days'`, `INTERVAL '1 hour'` (single-quoted, postgres-style)
- string: length, lower, upper, substr, position, split_part, regexp_like, \
regexp_match, regexp_replace, like, ilike, starts_with, ends_with, concat, \
concat_ws
- text search: contains_tokens(col, 'words') in WHERE; fts(table, query_json) in \
FROM (see above)
- numeric: round, floor, ceil, abs, sign, log, exp, power, sqrt
- conditional: CASE WHEN ... THEN ... ELSE ..., coalesce, nullif, greatest, least
- cast: CAST(x AS TYPE) or x::TYPE - but never on JSON columns (see the JSON \
rules above)
Quote identifiers with double quotes when they collide with keywords (e.g. \
\"timestamp\"); string literals use single quotes.

EXPLAIN is allowed: `EXPLAIN <query>` or `EXPLAIN ANALYZE <query>` returns the \
DataFusion plan (and per-operator timings for ANALYZE) so you can self-diagnose \
slow queries without leaving SQL.

Output modes (the `format` arg):
- text (default): a row-capped rendered ASCII table with a header showing \
`{total_rows} in {elapsed_ms} ms; showing {shown}` and, on truncation, a \
keyset-pagination hint.
- parquet | ndjson: write the FULL result set to a file and return a \
`pond-sql-export://<id>` resource link; read it via MCP resources/read. On a \
local/stdio install the response also names the on-disk path so you can open it \
directly with duckdb/polars.

Pagination - keyset (preferred):
Use ORDER BY on indexed columns plus a composite seek key for stable tie-breaking. \
The agent owns the cursor (the last sort value it saw); no server-side state.

  -- page 1: most recent 100 messages in pond
  SELECT message_id, timestamp, role, project
  FROM messages
  WHERE project LIKE '%pond%'
  ORDER BY timestamp DESC, message_id DESC
  LIMIT 100;

  -- page 2: pass back the last (timestamp, message_id) the agent saw
  SELECT message_id, timestamp, role, project
  FROM messages
  WHERE project LIKE '%pond%'
    AND (timestamp, message_id) < (TIMESTAMP '2026-06-05T08:14:22.123456Z', 'last-id')
  ORDER BY timestamp DESC, message_id DESC
  LIMIT 100;

Keyset stays stable across concurrent ingest (older rows don't shift) and uses \
the btree on `timestamp`/`message_id` directly. For known-bounded full results, skip \
pagination entirely: format=parquet writes everything in one call. OFFSET works \
but scans-and-discards prior rows and shifts pages under writes - prefer keyset.

Drilling from aggregates to content (instead of N round-trips of pond_get):
JOIN to messages/parts directly. Example - top 10 longest sessions with first \
user message:

  WITH top_sessions AS (
    SELECT session_id, COUNT(*) AS msgs
    FROM messages
    GROUP BY session_id
    ORDER BY msgs DESC
    LIMIT 10
  )
  SELECT ts.session_id, ts.msgs, s.project, s.source_agent,
         m.search_text AS first_user_msg
  FROM top_sessions ts
  JOIN sessions s ON s.session_id = ts.session_id
  LEFT JOIN messages m
    ON m.session_id = ts.session_id
   AND m.role = 'user'
   AND m.timestamp = (
     SELECT MIN(timestamp) FROM messages
     WHERE session_id = ts.session_id AND role = 'user'
   );

One call, agent picks exactly which columns to hydrate. When you want the \
pond_get-style rendered transcript (tool-call lines, subagent footer), call \
pond_get with the session_id - that's its job.

Examples (4 patterns the agent should recognize):

  -- 1. Activity by project this week
  SELECT project, COUNT(*) AS msgs, COUNT(DISTINCT session_id) AS sessions
  FROM messages
  WHERE timestamp >= now() - INTERVAL '7 days'
  GROUP BY project
  ORDER BY msgs DESC
  LIMIT 20;

  -- 2. Subagent breakdown
  SELECT source_agent, COUNT(*) AS n
  FROM sessions
  WHERE source_agent LIKE '%/%'
  GROUP BY source_agent
  ORDER BY n DESC;

  -- 3. Text filter in WHERE (all words must appear), composed with metadata
  SELECT message_id, timestamp, project, substr(search_text, 1, 120) AS preview
  FROM messages
  WHERE contains_tokens(search_text, 'race condition')
    AND timestamp >= now() - INTERVAL '30 days'
  ORDER BY timestamp DESC
  LIMIT 50;

  -- 4. BM25 search in FROM, joined with metadata, relevance-ranked
  SELECT m.session_id, m.timestamp, m.project, f._score, m.search_text
  FROM fts('messages', \
'{\"match\":{\"column\":\"search_text\",\"terms\":\"race condition\"}}') f
  JOIN messages m ON m.message_id = f.message_id
  WHERE m.project LIKE '%pond%'
  ORDER BY f._score DESC
  LIMIT 50;";

    /// `pond_search` MCP tool parameters.
    #[derive(Debug, Deserialize, schemars::JsonSchema)]
    struct McpSearchParams {
        /// What to search for: concepts and keywords. Keep it semantic - do
        /// not put project names in the query, use the `project` filter
        /// instead.
        query: String,
        /// Retrieval arm: "vector" (default - matches on meaning) or "fts"
        /// (matches exact whole words via BM25). Use vector for concepts/
        /// paraphrases, fts when you know the literal words. Falls back to fts
        /// when the store has no embeddings.
        #[serde(default)]
        mode: Option<String>,
        /// Result order: "relevance" (default - best match first) or "recency"
        /// (newest first; the response is labeled so you don't read rank-1 as
        /// the best match).
        #[serde(default)]
        sort_by: Option<String>,
        /// Max sessions to return. Default 10, server-capped at 200. This is
        /// also the "want more results" knob - raise it; there is no pagination.
        #[serde(default)]
        limit: Option<usize>,
        /// Filter to projects whose path contains this substring.
        #[serde(default)]
        project: Option<String>,
        /// Filter to one session (exact match) - search within a single,
        /// possibly long, session.
        #[serde(default)]
        session_id: Option<String>,
        /// Only messages on or after this date (YYYY-MM-DD).
        #[serde(default)]
        from_date: Option<String>,
        /// Only messages on or before this date (YYYY-MM-DD).
        #[serde(default)]
        to_date: Option<String>,
    }

    /// `pond_get` MCP tool parameters. Exactly one of `session_id` /
    /// `message_id` is required; the prefix on each param tells you which scope
    /// it applies to.
    #[derive(Debug, Deserialize, schemars::JsonSchema)]
    struct McpGetParams {
        /// Retrieve this whole session as a conversational transcript
        /// (user/assistant text + one-line tool/file refs). Use the `session_*`
        /// params with it.
        #[serde(default)]
        session_id: Option<String>,
        /// Max messages per session page. Default 20, max 1000.
        #[serde(default)]
        session_limit: Option<usize>,
        /// Which end to read the first session page from: "start" (oldest,
        /// default) or "end" (most recent, e.g. to recover context after
        /// compaction). Pages stay chronological.
        #[serde(default)]
        session_from: Option<String>,
        /// Page forward: a message id from a prior page's bottom marker;
        /// returns messages after it.
        #[serde(default)]
        session_after_message_id: Option<String>,
        /// Page backward: a message id from a prior page's top marker; returns
        /// messages before it.
        #[serde(default)]
        session_before_message_id: Option<String>,
        /// Retrieve this single message with its full parts (incl. tool_call /
        /// tool_result bodies). Use the `message_*` params with it. Mutually
        /// exclusive with session_id.
        #[serde(default)]
        message_id: Option<String>,
        /// Conversational sibling messages to include before the target
        /// (mirrors grep -B). Default 3.
        #[serde(default)]
        message_context_before: Option<usize>,
        /// Conversational sibling messages to include after the target
        /// (mirrors grep -A). Default 3.
        #[serde(default)]
        message_context_after: Option<usize>,
    }

    /// `pond_sql_query` MCP tool parameters.
    #[derive(Debug, Deserialize, schemars::JsonSchema)]
    struct McpSqlParams {
        /// One read-only SQL statement (DataFusion / PostgreSQL-compatible).
        /// SELECT/WITH only (or EXPLAIN of one); writes and side-effecting
        /// statements are rejected. Exact columns - messages(session_id,
        /// message_id, timestamp, role, source_agent, project, content
        /// [system-role only], search_text [the conversational text],
        /// embedding_model, options) | sessions(session_id,
        /// parent_session_id, parent_message_id, source_agent, created_at,
        /// project, options) | parts(session_id, message_id, id, ordinal,
        /// type, provenance, variant_data, options). parts.type enums use
        /// underscores: 'tool_call', 'tool_result', 'text', 'reasoning',
        /// 'file'. JSON columns (variant_data, options) are JSONB: read
        /// fields with json_extract(col, '$.a.b') or json_get_string(col,
        /// 'key', ...), never CAST them. Text search: WHERE
        /// contains_tokens(search_text, 'words') to filter, FROM
        /// fts('messages', '{...}') for BM25-ranked results. Control row count
        /// with SQL `LIMIT`; inline output is capped at 100 rows (use
        /// format=parquet|ndjson to get every row). See the `schema://pond-sql`
        /// resource for joins, JSON/FTS functions, pagination + drilling
        /// patterns, and worked examples.
        #[serde(alias = "sql")]
        query: String,
        /// Output format: "text" (default; rendered ASCII table with metrics
        /// footer, row-capped), "parquet", or "ndjson". For parquet/ndjson the
        /// full result set is written to a file and a `pond-sql-export://`
        /// resource link is returned (no truncation) - ndjson is the path for
        /// machine-readable JSON output.
        #[serde(default)]
        format: Option<String>,
    }

    fn parse_session_from(value: Option<String>) -> SessionFrom {
        match value.as_deref() {
            Some("end") => SessionFrom::End,
            _ => SessionFrom::Start,
        }
    }

    /// Parse the `mode` param; unknown / absent defaults to vector. Returns
    /// `None` for an explicit unknown value so the caller can reject it.
    fn parse_search_mode(value: Option<&str>) -> Option<SearchModeWire> {
        match value {
            None | Some("vector") => Some(SearchModeWire::Vector),
            Some("fts") => Some(SearchModeWire::Fts),
            Some(_) => None,
        }
    }

    fn parse_sort_by(value: Option<&str>) -> Option<SortBy> {
        match value {
            None | Some("relevance") => Some(SortBy::Relevance),
            Some("recency") => Some(SortBy::Recency),
            Some(_) => None,
        }
    }

    /// The pond MCP server: holds the shared state and the generated tool router.
    #[derive(Clone)]
    pub struct PondMcp {
        state: AppState,
        tool_router: ToolRouter<PondMcp>,
    }

    #[tool_router]
    impl PondMcp {
        pub fn new(state: AppState) -> Self {
            Self {
                state,
                tool_router: Self::tool_router(),
            }
        }

        #[tool(
            description = "Semantic search over stored conversation history. Pick the arm per \
                           query with `mode`: \"vector\" (default) matches on meaning - use it \
                           for concepts and paraphrases; \"fts\" matches exact whole words \
                           (BM25) - use it when you know the literal words. For symbols, \
                           substrings, identifiers, cross-session analytics, or subagent \
                           sessions, use pond_sql_query instead (this tool excludes subagents). \
                           Returns a readable transcript: a leading `key:` line explains the \
                           format and the first line states totals plus how many searchable \
                           messages the filters left in scope (the absence signal; searchable text \
                           is user/assistant conversational text by design - tool calls/results and \
                           reasoning are excluded as low-signal noise, so a gap there is expected, \
                           not a failure - reach tool output via pond_sql_query over \
                           parts.variant_data), then results are \
                           grouped by session, best session first; within a session, matching \
                           messages are newest-first. Each hit is a `--- [n] score | role | time \
                           | message_id | project | agent | session ---` rule followed by the \
                           matched text. Pass a returned `message_id` to `pond_get` for full \
                           text. Args: query (semantic - concepts, not project names), mode, \
                           sort_by (\"relevance\" default | \"recency\"), project / session_id / \
                           from_date / to_date to scope, limit to widen (no pagination - raise \
                           limit for more). Scores are relative within one response.",
            annotations(read_only_hint = true, idempotent_hint = true, open_world_hint = false)
        )]
        async fn pond_search(
            &self,
            Parameters(params): Parameters<McpSearchParams>,
        ) -> Result<CallToolResult, ErrorData> {
            let Some(mode) = parse_search_mode(params.mode.as_deref()) else {
                return Ok(CallToolResult::error(vec![Content::text(format!(
                    "unknown mode {:?}; use \"vector\" or \"fts\"",
                    params.mode.unwrap_or_default()
                ))]));
            };
            let Some(sort_by) = parse_sort_by(params.sort_by.as_deref()) else {
                return Ok(CallToolResult::error(vec![Content::text(format!(
                    "unknown sort_by {:?}; use \"relevance\" or \"recency\"",
                    params.sort_by.unwrap_or_default()
                ))]));
            };
            let request = SearchRequest {
                protocol_version: PROTOCOL_VERSION,
                namespace: Some(default_namespace()),
                query: params.query,
                mode,
                sort_by,
                filters: SearchFilters {
                    project: params.project.map(ProjectFilter::Contains),
                    session_id: params.session_id,
                    from_date: params.from_date,
                    to_date: params.to_date,
                    // min_score is intentionally not on the MCP surface; scores
                    // are response-relative, so a server-side threshold is a
                    // footgun for agent callers. CLI / HTTP still exposes it.
                    min_score: 0.0,
                },
                limit: params.limit.unwrap_or(10),
            };
            match run_search(
                &self.state.store,
                &self.state.embedder,
                request.clone(),
                &self.state.search,
            )
            .await
            {
                SearchEnvelope::Success(response) => Ok(tool_result(
                    crate::render::render_search_transcript(&response, &request),
                )),
                SearchEnvelope::Error(envelope) => Err(to_error_data(&envelope)),
            }
        }

        #[tool(
            description = "Retrieve stored conversation content as a readable transcript \
                           (a leading `key:` line explains the format). Pass exactly one of: \
                           session_id (the whole session as a conversational transcript - \
                           user/assistant text plus one-line tool/file refs; never inlines \
                           tool bodies) OR message_id (that one message with its full parts, \
                           incl. tool_call/tool_result bodies, marked `>`, plus its \
                           conversational neighbors). Params are prefixed by scope. session_*: \
                           session_limit (cap, default 20), session_from (\"start\"|\"end\"; \
                           \"end\" = most recent, e.g. to recover context after compaction), \
                           session_after_message_id / session_before_message_id (page down/up - \
                           pass the id a page marker shows). message_*: message_context_before / \
                           message_context_after (conversational neighbors each side, like \
                           grep -B/-A, default 3). A session_id response lists the session's \
                           subagents in a footer so you can open each. Tool/result lines render \
                           as `-> name [call_id]` / `<- name [call_id] (ok|failed)`. Not for \
                           bulk export - use `pond copy --to <file>`.",
            annotations(read_only_hint = true, idempotent_hint = true, open_world_hint = false)
        )]
        async fn pond_get(
            &self,
            Parameters(params): Parameters<McpGetParams>,
        ) -> Result<CallToolResult, ErrorData> {
            let request = GetRequest {
                protocol_version: PROTOCOL_VERSION,
                namespace: Some(default_namespace()),
                session_id: params.session_id,
                message_id: params.message_id,
                session_limit: params.session_limit.unwrap_or(20),
                session_from: parse_session_from(params.session_from),
                session_after_message_id: params.session_after_message_id,
                session_before_message_id: params.session_before_message_id,
                message_context_before: params.message_context_before.unwrap_or(3),
                message_context_after: params.message_context_after.unwrap_or(3),
            };
            match run_get(&self.state.store, request.clone()).await {
                GetEnvelope::Success(response) => {
                    let mut transcript = crate::render::render_get_transcript(&response, &request);
                    // Spawn-only subagents are stored as their own sessions
                    // (spec.md#datasets); surface them on the parent's first page
                    // so an agent can open each (otherwise they are undiscoverable
                    // from the MCP surface). Best-effort: a lookup failure just
                    // omits the footer rather than failing the get.
                    if request.message_id.is_none()
                        && request.session_after_message_id.is_none()
                        && request.session_before_message_id.is_none()
                        && let Ok(children) =
                            self.state.store.child_sessions(&response.session.id).await
                        && !children.is_empty()
                    {
                        transcript.push_str(&crate::render::render_subagents_footer(&children));
                    }
                    Ok(tool_result(transcript))
                }
                GetEnvelope::Error(envelope) => Err(to_error_data(&envelope)),
            }
        }

        #[tool(
            description = "Run ONE read-only SQL query (DataFusion / PostgreSQL-compatible) \
                           over the stored corpus as three tables: sessions, messages, parts. \
                           For filtering, joins, and aggregation (counts, group-by, time \
                           buckets) - the analytic complement to pond_search's semantic \
                           recall. SELECT/WITH only (or EXPLAIN of one); writes and side- \
                           effecting statements are rejected. The exact column lists are in \
                           the `query` parameter description - use those names, do not guess \
                           (column discovery also works: SELECT column_name FROM \
                           information_schema.columns WHERE table_name = 'messages'). \
                           Routing: metadata analytics -> SQL on messages/sessions; tool-call \
                           analytics -> parts WHERE type = 'tool_call' with \
                           json_get_string(variant_data, 'name'); text search -> WHERE \
                           contains_tokens(search_text, 'words') to filter or FROM \
                           fts('messages', '{...json...}') for BM25-ranked results, or \
                           pond_search for semantic recall; reading a transcript -> pond_get, \
                           not SQL. The embedding `vector` column is never returned (explicit \
                           projection is rejected; filtering in WHERE is fine). Control row \
                           count with SQL `LIMIT`; inline text output is capped at 100 rows. \
                           format defaults to text (a row-capped rendered table); set \
                           format=parquet|ndjson to write the full result to a file returned as \
                           a pond-sql-export:// resource (ndjson for machine-readable JSON). \
                           Read resource schema://pond-sql \
                           for joins, indexed columns, JSON access rules, the function \
                           quick-reference, pagination + drilling patterns, and worked \
                           examples.",
            annotations(read_only_hint = true, idempotent_hint = true, open_world_hint = false)
        )]
        async fn pond_sql_query(
            &self,
            Parameters(params): Parameters<McpSqlParams>,
        ) -> Result<CallToolResult, ErrorData> {
            let mode = match params.format.as_deref() {
                None | Some("text") => sql::Mode::Inline,
                Some("parquet") => sql::Mode::Export(sql::Format::Parquet),
                Some("ndjson") => sql::Mode::Export(sql::Format::Ndjson),
                Some(other) => {
                    return Ok(CallToolResult::error(vec![Content::text(format!(
                        "unknown format {other:?}; use \"text\", \"parquet\", or \"ndjson\""
                    ))]));
                }
            };
            let inline_rows = sql::DEFAULT_INLINE_ROWS;

            // Open only the tables the query names (spec.md#search): the slow
            // `parts.lance` open is pure waste for the common messages-only
            // query. The referenced tables are independent (per-table
            // caches/mutexes), so overlap their freshness/manifest fetches.
            let store = &self.state.store;
            let query = params.query.as_str();
            let tables = match tokio::try_join!(
                async {
                    anyhow::Ok(match sql::mentions_table(query, "sessions") {
                        true => Some(store.dataset(Table::Sessions).await?),
                        false => None,
                    })
                },
                async {
                    anyhow::Ok(match sql::mentions_table(query, "messages") {
                        true => Some(store.dataset(Table::Messages).await?),
                        false => None,
                    })
                },
                async {
                    anyhow::Ok(match sql::mentions_table(query, "parts") {
                        true => Some(store.dataset(Table::Parts).await?),
                        false => None,
                    })
                },
            ) {
                Ok((sessions, messages, parts)) => sql::Tables {
                    sessions,
                    messages,
                    parts,
                },
                Err(_) => {
                    return Err(ErrorData::internal_error(
                        "sql datasets unavailable".to_owned(),
                        None,
                    ));
                }
            };

            match sql::run(&tables, &params.query, mode, inline_rows).await {
                Ok(sql::Outcome::Inline(text)) => Ok(tool_result(text)),
                Ok(sql::Outcome::Export {
                    bytes,
                    format,
                    rows,
                    columns,
                }) => {
                    let name = format!("{}.{}", Uuid::now_v7(), format.ext());
                    match store.export_write(&name, &bytes).await {
                        Ok(_) => Ok(export_result(
                            store,
                            &name,
                            format,
                            rows,
                            &columns,
                            bytes.len(),
                        )),
                        Err(error) => Err(ErrorData::internal_error(
                            format!("export write failed: {error}"),
                            None,
                        )),
                    }
                }
                Err(sql::SqlError::Query(message)) => {
                    Ok(CallToolResult::error(vec![Content::text(message)]))
                }
                Err(sql::SqlError::Infra(error)) => Err(ErrorData::internal_error(
                    format!("sql execution failed: {error}"),
                    None,
                )),
            }
        }
    }

    // `router = self.tool_router` makes the generated `call_tool` / `list_tools`
    // read the cached router field; the bare-`#[tool_handler]` default rebuilds
    // the router via `Self::tool_router()` on every call instead.
    #[tool_handler(router = self.tool_router)]
    impl ServerHandler for PondMcp {
        fn get_info(&self) -> ServerInfo {
            ServerInfo::new(
                ServerCapabilities::builder()
                    .enable_tools()
                    .enable_resources()
                    .build(),
            )
            // rmcp's default `from_build_env` reports the rmcp crate (name +
            // version) - clients display the server's own identity, so set it.
            .with_server_info(Implementation::new("pond", env!("CARGO_PKG_VERSION")))
            .with_instructions(
                "pond recalls past agent sessions (Claude Code and others) - prior work, \
                 decisions, and context across sessions, not the live conversation. \
                 Workflow: pond_search to find relevant messages, then pond_get to read \
                 full text by message_id or a whole session by session_id; both return \
                 readable transcripts, not JSON. Scope with filters, not the query: project \
                 (path substring), session_id, source_agent, from_date / to_date - \
                 keep query semantic (concepts, not project names). Scores are relative \
                 within one response; there is no min_score. Subagents are stored as their \
                 own sessions (source_agent like \"claude-code/general-purpose\"); pond_get \
                 on a parent session lists them in a footer so you can open each. Recover \
                 context lost to compaction: find this session via pond_search (a distinctive \
                 recent topic + project + from_date=today), then pond_get(session_id, \
                 session_from=\"end\") for the recent pre-compaction turns. Deeper \
                 reference on demand: resource schema://pond (all filters + response format), \
                 stats://pond (corpus + embedding stats). For structured/analytic queries \
                 (filtering, joins, counts, group-by) use pond_sql_query: read-only SQL \
                 (SELECT only) over the sessions/messages/parts tables, with optional \
                 parquet/ndjson export; see resource schema://pond-sql. Search indexes only \
                 user/assistant conversational text by design (tool calls/results and \
                 reasoning are excluded as low-signal noise, not a bug), and a \
                 zero/weak result is not proof of absence - for exact strings, \
                 identifiers, or error messages run pond_sql_query with WHERE \
                 contains_tokens(search_text, 'words') (all words must match; \
                 index-accelerated), or FROM fts('messages', \
                 '{\"match\":{\"column\":\"search_text\",\"terms\":\"...\"}}') for \
                 BM25-ranked results; both cover subagent sessions too.",
            )
        }

        async fn list_resources(
            &self,
            _request: Option<PaginatedRequestParams>,
            _context: RequestContext<RoleServer>,
        ) -> Result<ListResourcesResult, ErrorData> {
            Ok(ListResourcesResult {
                resources: vec![
                    RawResource::new("schema://pond", "pond search schema").no_annotation(),
                    RawResource::new("schema://pond-sql", "pond SQL table schema").no_annotation(),
                    RawResource::new("stats://pond", "pond corpus stats").no_annotation(),
                ],
                next_cursor: None,
                meta: None,
            })
        }

        async fn read_resource(
            &self,
            request: ReadResourceRequestParams,
            _context: RequestContext<RoleServer>,
        ) -> Result<ReadResourceResult, ErrorData> {
            match request.uri.as_str() {
                "schema://pond" => Ok(ReadResourceResult::new(vec![ResourceContents::text(
                    SCHEMA_DOC,
                    request.uri,
                )])),
                "schema://pond-sql" => Ok(ReadResourceResult::new(vec![ResourceContents::text(
                    SQL_SCHEMA_DOC,
                    request.uri,
                )])),
                // `pond_sql_query` export artifacts: read the file pond wrote
                // (parquet -> base64 blob, ndjson -> text). The filename is
                // validated to a minted `<uuid>.<ext>` so the URI can't traverse.
                uri if uri.starts_with("pond-sql-export://") => {
                    let name = uri.trim_start_matches("pond-sql-export://").to_owned();
                    if !valid_export_name(&name) {
                        return Err(ErrorData::resource_not_found(
                            format!("invalid export id: {name}"),
                            None,
                        ));
                    }
                    let bytes = self.state.store.export_read(&name).await.map_err(|error| {
                        ErrorData::resource_not_found(format!("export not found: {error}"), None)
                    })?;
                    let contents = if name.ends_with(".ndjson") {
                        ResourceContents::text(
                            String::from_utf8_lossy(&bytes).into_owned(),
                            request.uri,
                        )
                        .with_mime_type("application/x-ndjson")
                    } else {
                        ResourceContents::blob(STANDARD.encode(&bytes), request.uri)
                            .with_mime_type("application/vnd.apache.parquet")
                    };
                    Ok(ReadResourceResult::new(vec![contents]))
                }
                "stats://pond" => {
                    let store = &self.state.store;
                    let map_err = |error: anyhow::Error| {
                        ErrorData::internal_error(format!("stats unavailable: {error}"), None)
                    };
                    let (sessions, messages, parts) = store.row_counts().await.map_err(&map_err)?;
                    let embedding = store.embedding_progress().await.map_err(&map_err)?;
                    let stale = store.stale_embedding_count().await.map_err(&map_err)?;
                    let indices = store.index_status().await.map_err(&map_err)?;

                    let embedded_percent = if embedding.total == 0 {
                        0.0
                    } else {
                        #[allow(clippy::cast_precision_loss)]
                        let pct = (embedding.embedded as f64 / embedding.total as f64) * 100.0;
                        (pct * 10.0).round() / 10.0
                    };
                    let index_rows = indices
                        .iter()
                        .map(|status| {
                            serde_json::json!({
                                "table": status.table.as_str(),
                                "intent": status.intent_name,
                                "exists": status.exists,
                                "fragments_covered": status.fragments_covered,
                                "unindexed_rows": status.unindexed_rows,
                            })
                        })
                        .collect::<Vec<_>>();

                    // spec.md#search: `search_text` is the conversational text
                    // (filtered of harness-injected parts at the adapter seam).
                    // `embedding.total` is the searchable population - that is
                    // the right denominator for "% embedded", not total messages.
                    let stats = serde_json::json!({
                        "corpus": {
                            "sessions": sessions,
                            "messages": messages,
                            "searchable_messages": embedding.total,
                            "parts": parts,
                        },
                        "embeddings": {
                            "model": embedding.model,
                            "embedded": embedding.embedded,
                            "searchable_total": embedding.total,
                            "embedded_percent": embedded_percent,
                            "stale_under_other_model": stale,
                        },
                        "indices": index_rows,
                    });
                    Ok(ReadResourceResult::new(vec![ResourceContents::text(
                        stats.to_string(),
                        request.uri,
                    )]))
                }
                other => Err(ErrorData::resource_not_found(
                    format!("unknown resource: {other}"),
                    None,
                )),
            }
        }

        async fn list_tools(
            &self,
            request: Option<PaginatedRequestParams>,
            context: RequestContext<RoleServer>,
        ) -> Result<ListToolsResult, ErrorData> {
            let _ = (request, context);
            let mut result = ListToolsResult {
                tools: self.tool_router.list_all(),
                next_cursor: None,
                meta: None,
            };
            annotate_tool_limits(&mut result);
            Ok(result)
        }
    }

    fn annotate_tool_limits(result: &mut ListToolsResult) {
        for tool in &mut result.tools {
            let chars = match tool.name.as_ref() {
                "pond_search" => 80_000,
                "pond_get" => 200_000,
                "pond_sql_query" => 80_000,
                _ => continue,
            };
            let mut meta = serde_json::Map::new();
            meta.insert(
                "anthropic/maxResultSizeChars".to_owned(),
                serde_json::json!(chars),
            );
            tool.meta = Some(Meta(meta));
        }
    }

    /// Run the stdio MCP server until the client disconnects. All diagnostics
    /// go to stderr (the shared `tracing` subscriber); stdout carries only
    /// JSON-RPC frames, written by rmcp's stdio transport (spec.md#scope).
    pub async fn serve_stdio(state: AppState) -> anyhow::Result<()> {
        let service = PondMcp::new(state)
            .serve(stdio())
            .await
            .context("failed to start stdio MCP server")?;
        service.waiting().await.context("stdio MCP server error")?;
        Ok(())
    }

    /// Build an MCP tool result from a rendered transcript. Deliberately text
    /// only: Claude Code surfaces `structuredContent` over the text block when
    /// both are present, which would shadow the transcript - the readable view
    /// is the whole point on the MCP surface. Programmatic clients that want the
    /// structured wire shape use the HTTP `/v1/*` JSON API instead.
    fn tool_result(transcript: String) -> CallToolResult {
        CallToolResult::success(vec![Content::text(transcript)])
    }

    /// Build the `pond_sql_query` export result: a text summary plus a
    /// `resource_link` to the artifact (the spec-canonical way to hand back a
    /// tool-produced file - the bytes ride `resources/read`, not the tool
    /// result, so they don't load into context unless the host fetches them).
    /// On a `file://` install the summary also names the on-disk path so a
    /// co-located agent can read it directly.
    fn export_result(
        store: &crate::sessions::Store,
        name: &str,
        format: sql::Format,
        rows: usize,
        columns: &[String],
        bytes: usize,
    ) -> CallToolResult {
        let uri = format!("pond-sql-export://{name}");
        let column_list = if columns.is_empty() {
            "(none)".to_owned()
        } else {
            columns.join(", ")
        };
        let mut summary = format!(
            "Exported {rows} row(s), {bytes} bytes ({}). Columns: {column_list}.\n\
             Fetch via MCP resources/read on {uri}.",
            format.ext()
        );
        if let Some(path) = store.export_local_path(name) {
            summary.push_str(&format!(
                "\nLocal file: {} - on this (stdio) install you can read it directly \
                 (e.g. duckdb, polars).",
                path.display()
            ));
        }
        let link = RawResource::new(uri, name.to_owned())
            .with_description(format!("pond SQL export ({}, {rows} rows)", format.ext()))
            .with_mime_type(format.mime().to_owned())
            .with_size(u32::try_from(bytes).unwrap_or(u32::MAX));
        CallToolResult::success(vec![Content::text(summary), Content::resource_link(link)])
    }

    /// Accept only the export filenames pond mints (`<uuid>.parquet|ndjson`),
    /// guarding the `pond-sql-export://` resource against path traversal.
    fn valid_export_name(name: &str) -> bool {
        let Some((stem, ext)) = name.rsplit_once('.') else {
            return false;
        };
        matches!(ext, "parquet" | "ndjson")
            && !stem.is_empty()
            && stem.bytes().all(|b| b.is_ascii_hexdigit() || b == b'-')
    }

    /// Map a wire error envelope to a JSON-RPC error. rmcp ships no app-level
    /// codes, so pond defines its own `-32000`-family set here. The `data`
    /// payload carries pond's canonical string code and a `retryable` flag
    /// (per spec.md#error-model) so MCP callers can branch on retry semantics
    /// without parsing message strings or knowing the JSON-RPC code mapping.
    fn to_error_data(envelope: &ErrorEnvelope) -> ErrorData {
        let (jsonrpc_code, pond_code, retryable) = match envelope.error.code {
            WireErrorCode::ValidationFailed => (-32010, "validation_failed", false),
            WireErrorCode::VersionUnsupported => (-32011, "version_unsupported", false),
            WireErrorCode::NotFound => (-32012, "not_found", false),
            WireErrorCode::NamespaceUnknown => (-32013, "namespace_unknown", false),
            WireErrorCode::StorageUnavailable => (-32014, "storage_unavailable", true),
            WireErrorCode::Conflict => (-32015, "conflict", true),
            WireErrorCode::Internal => (-32016, "internal", false),
        };
        let mut data = match &envelope.error.details {
            serde_json::Value::Object(map) => map.clone(),
            _ => serde_json::Map::new(),
        };
        data.insert("pond_code".to_owned(), serde_json::json!(pond_code));
        data.insert("retryable".to_owned(), serde_json::json!(retryable));
        ErrorData::new(
            JsonRpcErrorCode(jsonrpc_code),
            envelope.error.message.clone(),
            Some(serde_json::Value::Object(data)),
        )
    }

    #[cfg(test)]
    mod tests {
        #![allow(clippy::expect_used, clippy::unwrap_used)]

        use std::sync::Arc;

        use rmcp::model::{ErrorCode as JsonRpcErrorCode, Tool};

        use super::*;
        use crate::wire::{ErrorBody, ErrorCode};

        #[test]
        fn error_data_carries_code_and_retryability() {
            let cases = [
                (
                    ErrorCode::ValidationFailed,
                    -32010,
                    "validation_failed",
                    false,
                ),
                (
                    ErrorCode::VersionUnsupported,
                    -32011,
                    "version_unsupported",
                    false,
                ),
                (ErrorCode::NotFound, -32012, "not_found", false),
                (
                    ErrorCode::NamespaceUnknown,
                    -32013,
                    "namespace_unknown",
                    false,
                ),
                (
                    ErrorCode::StorageUnavailable,
                    -32014,
                    "storage_unavailable",
                    true,
                ),
                (ErrorCode::Conflict, -32015, "conflict", true),
                (ErrorCode::Internal, -32016, "internal", false),
            ];
            for (code, jsonrpc, pond_code, retryable) in cases {
                let error = to_error_data(&ErrorEnvelope {
                    error: ErrorBody {
                        code,
                        message: "boom".to_owned(),
                        details: serde_json::json!({"detail": 1}),
                    },
                });
                assert_eq!(error.code, JsonRpcErrorCode(jsonrpc));
                let data = error.data.unwrap();
                assert_eq!(data["detail"], serde_json::json!(1));
                assert_eq!(data["pond_code"], serde_json::json!(pond_code));
                assert_eq!(data["retryable"], serde_json::json!(retryable));
                assert!(
                    data.get("request_id").is_none(),
                    "MCP errors use JSON-RPC ids for correlation"
                );
            }
        }

        #[test]
        fn annotate_tool_limits_sets_anthropic_meta() {
            let schema = Arc::new(serde_json::Map::new());
            let mut result = ListToolsResult {
                tools: vec![
                    Tool::new("pond_search", "Search", Arc::clone(&schema)),
                    Tool::new("pond_get", "Get", Arc::clone(&schema)),
                ],
                next_cursor: None,
                meta: None,
            };
            annotate_tool_limits(&mut result);
            let value = |name: &str| {
                result
                    .tools
                    .iter()
                    .find(|tool| tool.name == name)
                    .and_then(|tool| tool.meta.as_ref())
                    .and_then(|meta| meta.0.get("anthropic/maxResultSizeChars"))
                    .and_then(serde_json::Value::as_i64)
            };
            assert_eq!(value("pond_search"), Some(80_000));
            assert_eq!(value("pond_get"), Some(200_000));
        }
    }
}