codewhale-tui 0.9.2

Terminal UI for open-source and open-weight coding models
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
//! Deferred tool catalog and built-in advanced tool helpers.
//!
//! The streaming turn loop owns when tools are offered or executed. This module
//! owns the catalog-level policy around deferred loading, tool search, missing
//! tool suggestions, and the small set of built-in advanced tools that are not
//! registered by the normal runtime tool registry.

use std::collections::HashSet;
use std::path::Path;
use std::time::Duration;

use serde_json::{Value, json};

use crate::mcp::McpPool;
use crate::model_profile::ToolSurfaceBudget;
use crate::models::Tool;
use crate::tools::spec::{ToolError, ToolResult, optional_str, optional_u64, required_str};
use crate::tui::app::AppMode;

use crate::dependencies::ExternalTool;
use crate::regex_cache::compile_user_regex;

pub(super) const MULTI_TOOL_PARALLEL_NAME: &str = "multi_tool_use.parallel";
pub(super) const REQUEST_USER_INPUT_NAME: &str = "request_user_input";
pub(super) const CODE_EXECUTION_TOOL_NAME: &str = "code_execution";
const CODE_EXECUTION_TOOL_TYPE: &str = "code_execution_20250825";
pub(super) use crate::tools::js_execution::JS_EXECUTION_TOOL_NAME;
pub(super) const TOOL_SEARCH_NAME: &str = "tool_search";
const TOOL_SEARCH_TYPE: &str = "tool_search_20251119";
const LEGACY_TOOL_SEARCH_REGEX_NAME: &str = "tool_search_tool_regex";
const LEGACY_TOOL_SEARCH_BM25_NAME: &str = "tool_search_tool_bm25";
const TOOL_SEARCH_DEFAULT_MAX_RESULTS: usize = 20;
const TOOL_SEARCH_MAX_RESULTS_LIMIT: usize = 100;

pub(super) fn is_tool_search_tool(name: &str) -> bool {
    matches!(
        name,
        TOOL_SEARCH_NAME | LEGACY_TOOL_SEARCH_REGEX_NAME | LEGACY_TOOL_SEARCH_BM25_NAME
    )
}

pub(super) const DEFAULT_ACTIVE_NATIVE_TOOLS: &[&str] = &[
    // #4625: the model-facing shell tool is `Bash`; legacy `exec_shell*`
    // names are hidden compat aliases and must not be default-active.
    "Bash",
    "File",
    "Git",
    "Run",
    "agent",
    "remember",
    // Piagent phase B: the model-facing durable-task tool is `tasks`; the
    // legacy `task_create`/`task_list`/`task_read` names it replaces are
    // hidden compat aliases and must not be default-active.
    "tasks",
    "work_update",
];

const CORE_ACTION_TOOL_FALLBACKS: &[CoreActionToolFallback] = &[
    CoreActionToolFallback {
        name: "Bash",
        description: "Run shell commands in the workspace.",
        unavailable_reason: "Not present in the current model-visible catalog. Interactive Agent sessions expose shell by default unless allow_shell = false; noninteractive and durable profiles require allow_shell = true. Plan mode hides shell, and command tool allow/deny gates can also block it.",
    },
    CoreActionToolFallback {
        name: "File",
        description: "Read, search, and modify workspace files.",
        unavailable_reason: "Not present in the current model-visible catalog. File reads are available in Plan and Agent modes; write and edit actions require an executable mode, while patch also requires the apply_patch feature.",
    },
];

#[derive(Debug, Clone, Copy)]
struct CoreActionToolFallback {
    name: &'static str,
    description: &'static str,
    unavailable_reason: &'static str,
}

/// Pre-computed lowercased haystack + name for each fallback; built once.
struct CachedFallback {
    fallback: CoreActionToolFallback,
    haystack: String,
    name_lower: String,
}

static CACHED_FALLBACKS: std::sync::OnceLock<Vec<CachedFallback>> = std::sync::OnceLock::new();

fn cached_fallbacks() -> &'static [CachedFallback] {
    CACHED_FALLBACKS.get_or_init(|| {
        CORE_ACTION_TOOL_FALLBACKS
            .iter()
            .map(|f| CachedFallback {
                fallback: *f,
                haystack: format!(
                    "{}\n{}\n{}",
                    f.name.to_lowercase(),
                    f.description.to_lowercase(),
                    f.unavailable_reason.to_lowercase(),
                ),
                name_lower: f.name.to_lowercase(),
            })
            .collect()
    })
}

/// Membership index over [`DEFAULT_ACTIVE_NATIVE_TOOLS`], built once for the
/// process lifetime. The array stays the source of truth for *ordered*
/// iteration (see [`tool_catalog_consistency_issues`] and
/// `engine::default_active_native_tool_names`); this set only accelerates the
/// hot membership check in [`should_default_defer_tool`], which runs once per
/// catalog tool on every catalog rebuild (i.e. per turn) — an O(n·m) linear
/// scan over the array collapses to O(1) hashed lookups.
static DEFAULT_ACTIVE_NATIVE_TOOLS_SET: std::sync::OnceLock<HashSet<&'static str>> =
    std::sync::OnceLock::new();

fn default_active_native_tools_set() -> &'static HashSet<&'static str> {
    DEFAULT_ACTIVE_NATIVE_TOOLS_SET
        .get_or_init(|| DEFAULT_ACTIVE_NATIVE_TOOLS.iter().copied().collect())
}

pub(super) fn should_default_defer_tool(name: &str, always_load: &HashSet<String>) -> bool {
    if always_load.contains(name) {
        return false;
    }

    if is_tool_search_tool(name) {
        return false;
    }

    // Membership-only test (no ordering dependency): the side set built from
    // DEFAULT_ACTIVE_NATIVE_TOOLS returns identical hit/miss results as the
    // former `.iter().any(...)` linear scan.
    !default_active_native_tools_set().contains(name)
}

pub(super) fn apply_native_tool_deferral(catalog: &mut [Tool], always_load: &HashSet<String>) {
    for tool in catalog {
        tool.defer_loading = Some(should_default_defer_tool(&tool.name, always_load));
    }
}

fn should_keep_mcp_tool_loaded(name: &str) -> bool {
    matches!(
        name,
        "list_mcp_resources"
            | "list_mcp_resource_templates"
            | "mcp_read_resource"
            | "read_mcp_resource"
            | "mcp_get_prompt"
    )
}

pub(super) fn apply_mcp_tool_deferral(
    catalog: &mut [Tool],
    mode: AppMode,
    always_load: &HashSet<String>,
) {
    for tool in catalog {
        if always_load.contains(&tool.name) {
            tool.defer_loading = Some(false);
            continue;
        }
        tool.defer_loading =
            Some(mode != AppMode::Yolo && !should_keep_mcp_tool_loaded(&tool.name));
    }
}

/// Build the model tool catalog from native and MCP tool lists.
///
/// **Catalog-head stability invariant.** The head of the catalog (all
/// non-deferred tools) must remain byte-identical across mode toggles
/// (Plan ↔ Agent ↔ YOLO) for tools that are common to both modes.
/// Deferred tool activations append to the tail and never reorder the
/// head. This invariant is critical for DeepSeek's KV prefix cache:
/// the tools array is part of the immutable prefix, and any byte-level
/// change in the head forces a full re-prefill on the next turn.
#[cfg(test)]
pub(super) fn build_model_tool_catalog(
    native_tools: Vec<Tool>,
    mcp_tools: Vec<Tool>,
    mode: AppMode,
    always_load: &HashSet<String>,
) -> Vec<Tool> {
    build_model_tool_catalog_with_surface(
        native_tools,
        mcp_tools,
        mode,
        always_load,
        ToolSurfaceBudget::Standard,
    )
}

pub(super) fn build_model_tool_catalog_with_surface(
    mut native_tools: Vec<Tool>,
    mut mcp_tools: Vec<Tool>,
    mode: AppMode,
    always_load: &HashSet<String>,
    surface_budget: ToolSurfaceBudget,
) -> Vec<Tool> {
    apply_native_tool_deferral(&mut native_tools, always_load);
    apply_mcp_tool_deferral(&mut mcp_tools, mode, always_load);
    apply_tool_surface_budget(&mut native_tools, surface_budget, always_load);
    apply_tool_surface_budget(&mut mcp_tools, surface_budget, always_load);
    // Sort each partition by name for prefix-cache stability (#263). The
    // upstream `to_api_tools()` already sorts the registry's HashMap output;
    // this catalog is built from caller-supplied Vecs which the test harness
    // and (future) caller refactors may not pre-sort. Built-ins stay as a
    // contiguous prefix ahead of MCP tools so adding/removing an MCP tool
    // never shifts a built-in's position.
    native_tools.sort_by(|a, b| a.name.cmp(&b.name));
    mcp_tools.sort_by(|a, b| a.name.cmp(&b.name));
    native_tools.extend(mcp_tools);
    native_tools
}

fn apply_tool_surface_budget(
    catalog: &mut [Tool],
    surface_budget: ToolSurfaceBudget,
    always_load: &HashSet<String>,
) {
    if !matches!(surface_budget, ToolSurfaceBudget::Compact) {
        return;
    }
    for tool in catalog {
        if always_load.contains(&tool.name) {
            continue;
        }
        if matches!(tool.name.as_str(), "agent" | "Run" | "tasks" | "Web") {
            tool.defer_loading = Some(true);
        }
    }
}

/// Whether two tool-surface budgets currently produce the same catalog.
///
/// Runs [`apply_tool_surface_budget`] over `catalog` under both budgets and
/// compares the results. `/preview-request` publishes the Standard-vs-Full
/// answer as a derived field so the truthful "these are currently collapsed"
/// disclosure cannot drift from the code: the day the shaper narrows Standard
/// differently from Full, this starts returning `false` on its own.
pub(super) fn surface_budgets_produce_same_catalog(
    catalog: &[Tool],
    always_load: &HashSet<String>,
    left_budget: ToolSurfaceBudget,
    right_budget: ToolSurfaceBudget,
) -> bool {
    let mut left = catalog.to_vec();
    let mut right = catalog.to_vec();
    apply_tool_surface_budget(&mut left, left_budget, always_load);
    apply_tool_surface_budget(&mut right, right_budget, always_load);
    serde_json::to_string(&left).ok() == serde_json::to_string(&right).ok()
}

pub(super) fn ensure_advanced_tooling(
    catalog: &mut Vec<Tool>,
    mode: AppMode,
    always_load: &HashSet<String>,
) {
    // code_execution depends on a locally-installed Python interpreter
    // (python3 / python / py -3). Before v0.8.31, the tool was always
    // advertised and would fail at execution time on Windows where
    // `python3` isn't on PATH — the model treated the tool as reliable
    // once it appeared in the catalog. We now probe at catalog-build
    // time and only advertise when an interpreter resolves. See
    // `crate::dependencies::resolve_python_interpreter` for the probe.
    if mode != AppMode::Plan
        && !catalog.iter().any(|t| t.name == CODE_EXECUTION_TOOL_NAME)
        && crate::dependencies::resolve_python_interpreter().is_some()
    {
        catalog.push(Tool {
            tool_type: Some(CODE_EXECUTION_TOOL_TYPE.to_string()),
            name: CODE_EXECUTION_TOOL_NAME.to_string(),
            description: "Execute Python code in a local sandboxed runtime and return stdout/stderr/return_code as JSON.".to_string(),
            input_schema: json!({
                "type": "object",
                "properties": {
                    "code": { "type": "string", "description": "Python source code to execute." }
                },
                "required": ["code"]
            }),
            allowed_callers: Some(vec!["direct".to_string()]),
            defer_loading: Some(should_default_defer_tool(
                CODE_EXECUTION_TOOL_NAME,
                always_load,
            )),
            input_examples: None,
            strict: None,
            cache_control: None,
        });
    }

    // js_execution mirrors code_execution: gate on Node.js being
    // present locally so the model never sees a runtime it can't
    // actually use. Plan mode hides shell/exec surfaces (including
    // both interpreter tools) by construction; Agent / YOLO advertise
    // the tool only when `resolve_node()` succeeds.
    if mode != AppMode::Plan
        && !catalog.iter().any(|t| t.name == JS_EXECUTION_TOOL_NAME)
        && crate::dependencies::resolve_node().is_some()
    {
        let mut tool = crate::tools::js_execution::js_execution_tool_definition();
        tool.defer_loading = Some(should_default_defer_tool(&tool.name, always_load));
        catalog.push(tool);
    }

    if !catalog.iter().any(|t| t.name == TOOL_SEARCH_NAME) {
        catalog.push(Tool {
            tool_type: Some(TOOL_SEARCH_TYPE.to_string()),
            name: TOOL_SEARCH_NAME.to_string(),
            description: "Search deferred tool definitions and return matching tool references.".to_string(),
            input_schema: json!({
                "type": "object",
                "properties": {
                    "query": { "type": "string", "description": "Search query for tool discovery." },
                    "match": {
                        "type": "string",
                        "enum": ["bm25", "regex"],
                        "default": "bm25",
                        "description": "Matching algorithm: bm25 for natural-language matching, regex for a regular expression over tool names/descriptions/schema."
                    },
                    "max_results": {
                        "type": "integer",
                        "minimum": 1,
                        "maximum": TOOL_SEARCH_MAX_RESULTS_LIMIT,
                        "default": TOOL_SEARCH_DEFAULT_MAX_RESULTS,
                        "description": "Maximum number of matching tool references to return."
                    }
                },
                "required": ["query"]
            }),
            allowed_callers: Some(vec!["direct".to_string()]),
            defer_loading: Some(false),
            input_examples: None,
            strict: None,
            cache_control: None,
        });
    }
}

pub(super) fn initial_active_tools(catalog: &[Tool]) -> HashSet<String> {
    let mut active = HashSet::new();
    for tool in catalog {
        if !tool.defer_loading.unwrap_or(false) || is_tool_search_tool(&tool.name) {
            active.insert(tool.name.clone());
        }
    }
    if active.is_empty()
        && !catalog.is_empty()
        && let Some(first) = catalog.first()
    {
        active.insert(first.name.clone());
    }
    active
}

fn active_tool_list_from_catalog(catalog: &[Tool], active: &HashSet<String>) -> Vec<Tool> {
    // Two-pass for prefix-cache stability (#263). Always-loaded tools come
    // first in their stable catalog order; tools that started life deferred
    // and were activated mid-conversation by ToolSearch get appended at the
    // tail. Otherwise activating a deferred tool shifts every later tool's
    // byte offset and busts the cached prefix from that point onwards.
    let catalog_len = catalog.len();
    let mut head: Vec<Tool> = Vec::with_capacity(catalog_len);
    let mut tail: Vec<Tool> = Vec::with_capacity(catalog_len);
    for tool in catalog {
        if !active.contains(&tool.name) {
            continue;
        }
        if tool.defer_loading.unwrap_or(false) {
            tail.push(tool.clone());
        } else {
            head.push(tool.clone());
        }
    }
    head.extend(tail);
    head
}

pub(super) fn active_tools_for_step(catalog: &[Tool], active: &HashSet<String>) -> Vec<Tool> {
    active_tool_list_from_catalog(catalog, active)
}

/// The exact tool state the next model request would carry.
///
/// This is the single answer to "what tools would the next turn send?".
/// [`super::turn_loop`] seeds its mutable per-step state from it, and
/// `/preview-request` reports [`Self::active`] verbatim. Nothing else may
/// re-derive tool selection — in particular, the session's *last* catalog is
/// both stale and pre-activation, so it is never a substitute for this.
#[derive(Debug, Clone)]
pub(super) struct TurnToolPlan {
    /// Full catalog after mode/always-load repair, including deferred tools.
    pub(super) catalog: Vec<Tool>,
    /// Names active at the start of the turn.
    pub(super) active_names: HashSet<String>,
    /// The catalog subset that would actually be serialized into the request.
    /// `None` when the turn would send no `tools` field at all.
    pub(super) active: Option<Vec<Tool>>,
}

/// The `tools` field of one outbound request, from a catalog and the set of
/// currently-active tool names.
///
/// Shared by [`plan_turn_tools`] (turn seed and `/preview-request`) and by the
/// per-step rebuild inside the turn loop, so activating a deferred tool
/// mid-turn goes through exactly one code path.
pub(super) fn active_tools_for_request(
    catalog: &[Tool],
    active: &HashSet<String>,
    strict_tool_mode: bool,
) -> Option<Vec<Tool>> {
    if catalog.is_empty() {
        return None;
    }
    let mut tools = active_tools_for_step(catalog, active);
    if strict_tool_mode {
        crate::tools::schema_sanitize::prepare_tools_for_strict_mode(&mut tools);
    }
    Some(tools)
}

/// Compute [`TurnToolPlan`] from a freshly built catalog.
///
/// `tools` is the catalog produced by `build_model_tool_catalog_with_surface`
/// plus the gate and permission-posture filters — i.e. exactly the value the
/// engine hands to `handle_deepseek_turn`.
pub(super) fn plan_turn_tools(
    tools: Option<Vec<Tool>>,
    mode: AppMode,
    always_load: &HashSet<String>,
    dynamic_active_tools: &[&'static str],
    strict_tool_mode: bool,
) -> TurnToolPlan {
    let mut catalog = tools.unwrap_or_default();
    if !catalog.is_empty() {
        ensure_advanced_tooling(&mut catalog, mode, always_load);
    }
    let mut active_names = initial_active_tools(&catalog);
    active_names.extend(dynamic_active_tools.iter().map(|name| (*name).to_string()));
    let active = active_tools_for_request(&catalog, &active_names, strict_tool_mode);
    TurnToolPlan {
        catalog,
        active_names,
        active,
    }
}

fn tool_search_haystack(tool: &Tool) -> String {
    format!(
        "{}\n{}\n{}",
        tool.name.to_lowercase(),
        tool.description.to_lowercase(),
        tool.input_schema.to_string().to_lowercase()
    )
}

fn catalog_contains_tool(catalog: &[Tool], name: &str) -> bool {
    catalog.iter().any(|tool| tool.name == name)
}

fn unavailable_core_action_tools_with_regex(
    catalog: &[Tool],
    query: &str,
    max_results: usize,
) -> Result<Vec<CoreActionToolFallback>, ToolError> {
    if max_results == 0 {
        return Ok(Vec::new());
    }
    let regex = compile_user_regex(query)
        .map_err(|err| ToolError::invalid_input(format!("Invalid regex query: {err}")))?;
    Ok(cached_fallbacks()
        .iter()
        .filter(|cf| !catalog_contains_tool(catalog, cf.fallback.name))
        .filter(|cf| regex.is_match(&cf.haystack))
        .take(max_results)
        .map(|cf| cf.fallback)
        .collect())
}

fn unavailable_core_action_tools_with_bm25_like(
    catalog: &[Tool],
    query: &str,
    max_results: usize,
) -> Vec<CoreActionToolFallback> {
    if max_results == 0 {
        return Vec::new();
    }
    let terms: Vec<String> = query
        .split_whitespace()
        .map(|term| term.trim().to_lowercase())
        .filter(|term| !term.is_empty())
        .collect();
    if terms.is_empty() {
        return Vec::new();
    }

    let mut scored: Vec<(i64, CoreActionToolFallback)> = Vec::new();
    for cf in cached_fallbacks() {
        if catalog_contains_tool(catalog, cf.fallback.name) {
            continue;
        }
        let hay = &cf.haystack;
        let name = &cf.name_lower;
        let mut score = 0i64;
        for term in &terms {
            if hay.contains(term) {
                score += 1;
            }
            if name.contains(term) {
                score += 2;
            }
        }
        if score > 0 {
            scored.push((score, cf.fallback));
        }
    }
    scored.sort_by(|a, b| b.0.cmp(&a.0).then_with(|| a.1.name.cmp(b.1.name)));
    scored
        .into_iter()
        .take(max_results)
        .map(|(_, fallback)| fallback)
        .collect()
}

fn discover_tools_with_regex(
    catalog: &[Tool],
    query: &str,
    max_results: usize,
) -> Result<Vec<String>, ToolError> {
    let regex = compile_user_regex(query)
        .map_err(|err| ToolError::invalid_input(format!("Invalid regex query: {err}")))?;

    let mut matches = Vec::new();
    for tool in catalog {
        if is_tool_search_tool(&tool.name) {
            continue;
        }
        let hay = tool_search_haystack(tool);
        if regex.is_match(&hay) {
            matches.push(tool.name.clone());
        }
        if matches.len() >= max_results {
            break;
        }
    }
    Ok(matches)
}

fn discover_tools_with_bm25_like(catalog: &[Tool], query: &str, max_results: usize) -> Vec<String> {
    let terms: Vec<String> = query
        .split_whitespace()
        .map(|term| term.trim().to_lowercase())
        .filter(|term| !term.is_empty())
        .collect();
    if terms.is_empty() {
        return Vec::new();
    }

    let mut scored: Vec<(i64, String)> = Vec::new();
    for tool in catalog {
        if is_tool_search_tool(&tool.name) {
            continue;
        }
        let hay = tool_search_haystack(tool);
        let mut score = 0i64;
        for term in &terms {
            if hay.contains(term) {
                score += 1;
            }
            if tool.name.to_lowercase().contains(term) {
                score += 2;
            }
        }
        if score > 0 {
            scored.push((score, tool.name.clone()));
        }
    }
    scored.sort_by(|a, b| b.0.cmp(&a.0).then_with(|| a.1.cmp(&b.1)));
    scored
        .into_iter()
        .take(max_results)
        .map(|(_, name)| name)
        .collect()
}

fn edit_distance(a: &str, b: &str) -> usize {
    if a == b {
        return 0;
    }
    if a.is_empty() {
        return b.chars().count();
    }
    if b.is_empty() {
        return a.chars().count();
    }

    let b_chars: Vec<char> = b.chars().collect();
    let mut prev: Vec<usize> = (0..=b_chars.len()).collect();
    let mut curr = vec![0usize; b_chars.len() + 1];

    for (i, a_ch) in a.chars().enumerate() {
        curr[0] = i + 1;
        for (j, b_ch) in b_chars.iter().enumerate() {
            let cost = if a_ch == *b_ch { 0 } else { 1 };
            let delete = prev[j + 1] + 1;
            let insert = curr[j] + 1;
            let substitute = prev[j] + cost;
            curr[j + 1] = delete.min(insert).min(substitute);
        }
        std::mem::swap(&mut prev, &mut curr);
    }

    prev[b_chars.len()]
}

fn suggest_tool_names(catalog: &[Tool], requested: &str, limit: usize) -> Vec<String> {
    let requested = requested.trim().to_ascii_lowercase();
    if requested.is_empty() || limit == 0 {
        return Vec::new();
    }

    let mut candidates: Vec<(u8, usize, String)> = Vec::new();
    for tool in catalog {
        let candidate = tool.name.to_ascii_lowercase();
        let prefix_match = candidate.starts_with(&requested) || requested.starts_with(&candidate);
        let contains_match = candidate.contains(&requested) || requested.contains(&candidate);
        let distance = edit_distance(&candidate, &requested);
        let close_typo = distance <= 3;

        if !(prefix_match || contains_match || close_typo) {
            continue;
        }

        let rank = if prefix_match {
            0
        } else if contains_match {
            1
        } else {
            2
        };
        candidates.push((rank, distance, tool.name.clone()));
    }

    candidates.sort_by(|a, b| {
        a.0.cmp(&b.0)
            .then_with(|| a.1.cmp(&b.1))
            .then_with(|| a.2.cmp(&b.2))
    });
    candidates.dedup_by(|a, b| a.2 == b.2);
    candidates
        .into_iter()
        .take(limit)
        .map(|(_, _, name)| name)
        .collect()
}

/// Catalog tools the engine injects itself rather than registering, plus the
/// legacy tool-search spellings. Exposed so the read-only request projection
/// can label their provenance as `synthetic` from the same source of truth as
/// [`is_synthetic_catalog_tool`] instead of guessing.
///
/// MCP-contributed names are deliberately *not* here: those resolve through the
/// real pool, and stay unknown when the pool did not resolve them.
/// [`MULTI_TOOL_PARALLEL_NAME`] is not here either — it is a call name the model
/// may emit, never a catalog entry, so it can never appear in a transmitted
/// tool array and has no catalog provenance to report.
pub(super) fn default_synthetic_catalog_tool_names() -> Vec<String> {
    let mut names: Vec<String> = vec![
        TOOL_SEARCH_NAME.to_string(),
        LEGACY_TOOL_SEARCH_REGEX_NAME.to_string(),
        LEGACY_TOOL_SEARCH_BM25_NAME.to_string(),
        CODE_EXECUTION_TOOL_NAME.to_string(),
        JS_EXECUTION_TOOL_NAME.to_string(),
    ];
    names.sort();
    names.dedup();
    names
}

fn is_synthetic_catalog_tool(name: &str) -> bool {
    is_tool_search_tool(name)
        || matches!(name, CODE_EXECUTION_TOOL_NAME | JS_EXECUTION_TOOL_NAME)
        || McpPool::is_mcp_tool(name)
}

pub(super) fn tool_catalog_consistency_issues(
    catalog: &[Tool],
    registry: &crate::tools::ToolRegistry,
) -> Vec<String> {
    let catalog_names = catalog
        .iter()
        .map(|tool| tool.name.as_str())
        .collect::<HashSet<_>>();
    let registry_api_tools = registry.to_api_tools();
    let registry_model_visible_names = registry_api_tools
        .iter()
        .map(|tool| tool.name.as_str())
        .collect::<HashSet<_>>();
    let mut issues = Vec::new();

    for tool in catalog {
        if is_synthetic_catalog_tool(&tool.name) {
            continue;
        }
        if !registry.contains(&tool.name) {
            issues.push(format!(
                "catalog advertises '{}' but no registered handler exists",
                tool.name
            ));
        }
    }

    for name in DEFAULT_ACTIVE_NATIVE_TOOLS {
        if registry_model_visible_names.contains(name) && !catalog_names.contains(name) {
            issues.push(format!(
                "registered core tool '{name}' is missing from the model/search catalog"
            ));
        }
    }

    issues.sort();
    issues
}

pub(super) fn missing_tool_error_message(tool_name: &str, catalog: &[Tool]) -> String {
    // Dogfood A5 (#4092): models mid-checklist sometimes emit each list entry
    // as its own tool call named `item`/`todo`/... . Fuzzy suggestions are
    // actively misleading there ("Did you mean: note, tts?"); name the actual
    // fix instead.
    if matches!(
        tool_name,
        "item" | "items" | "todo" | "todos" | "checklist" | "checklist_item" | "plan_item"
    ) {
        return format!(
            "Tool '{tool_name}' is not available in the current tool catalog. \
             Checklist entries are not separate tool calls — write the whole list \
             in one `work_update` call with a `todos` array of \
             {{content, status}} objects."
        );
    }
    let suggestions = suggest_tool_names(catalog, tool_name, 3);
    let shell_hint = if is_shell_tool_name(tool_name) {
        Some(shell_tool_allow_shell_hint())
    } else {
        None
    };
    if suggestions.is_empty() {
        if let Some(shell_hint) = shell_hint {
            return format!(
                "Tool '{tool_name}' is not available in the current tool catalog. \
                 {shell_hint}, or use {TOOL_SEARCH_NAME} with a short query."
            );
        }
        return format!(
            "Tool '{tool_name}' is not available in the current tool catalog. \
             Verify mode/feature flags, or use {TOOL_SEARCH_NAME} with a short query."
        );
    }

    let suggestion_text = format!("Did you mean: {}?", suggestions.join(", "));
    if let Some(shell_hint) = shell_hint {
        return format!(
            "Tool '{tool_name}' is not available in the current tool catalog. \
             {suggestion_text} {shell_hint}. \
             You can also use {TOOL_SEARCH_NAME} to discover tools."
        );
    }

    format!(
        "Tool '{tool_name}' is not available in the current tool catalog. \
         {suggestion_text} You can also use {TOOL_SEARCH_NAME} to discover tools."
    )
}

fn shell_tool_allow_shell_hint() -> &'static str {
    "Shell tools are absent because this session or profile disabled shell access, \
     commonly via top-level `allow_shell = false` or Plan mode. \
     Interactive Act mode exposes shell by default with approval gating unless disabled. \
     Run `/config allow_shell true` for this session or add `--save` for future sessions; \
     the next turn will expose shell again"
}

fn is_shell_tool_name(tool_name: &str) -> bool {
    matches!(
        tool_name,
        "exec_shell"
            | "exec_shell_wait"
            | "exec_shell_interact"
            | "task_shell_start"
            | "task_shell_wait"
    )
}

#[cfg(test)]
pub(super) fn maybe_activate_requested_deferred_tool(
    tool_name: &str,
    catalog: &[Tool],
    active_tools: &mut HashSet<String>,
) -> bool {
    let Some(def) = catalog.iter().find(|def| def.name == tool_name) else {
        return false;
    };

    if !def.defer_loading.unwrap_or(false) || active_tools.contains(tool_name) {
        return false;
    }

    active_tools.insert(tool_name.to_string())
}

pub(super) fn maybe_hydrate_requested_deferred_tool(
    tool_name: &str,
    tool_input: &Value,
    catalog: &[Tool],
    active_tools_at_batch_start: &HashSet<String>,
    hydrated_tools_this_batch: &mut HashSet<String>,
) -> Option<ToolResult> {
    let def = catalog.iter().find(|def| def.name == tool_name)?;

    if !def.defer_loading.unwrap_or(false) || active_tools_at_batch_start.contains(tool_name) {
        return None;
    }

    hydrated_tools_this_batch.insert(tool_name.to_string());
    Some(deferred_tool_schema_hydration_result(def, tool_input))
}

#[cfg(test)]
pub(super) fn preflight_requested_deferred_tool(
    tool_name: &str,
    tool_input: &Value,
    catalog: &[Tool],
    active_tools: &mut HashSet<String>,
) -> Option<ToolResult> {
    let active_tools_at_batch_start = active_tools.clone();
    let mut hydrated_tools_this_batch = HashSet::new();
    let result = maybe_hydrate_requested_deferred_tool(
        tool_name,
        tool_input,
        catalog,
        &active_tools_at_batch_start,
        &mut hydrated_tools_this_batch,
    );
    active_tools.extend(hydrated_tools_this_batch);
    result
}

fn deferred_tool_schema_hydration_result(tool: &Tool, tool_input: &Value) -> ToolResult {
    let expected = schema_fields(&tool.input_schema);
    let required = schema_required_fields(&tool.input_schema);
    let received = received_field_names(tool_input);
    let missing = required
        .iter()
        .filter(|field| !received.contains(field))
        .cloned()
        .collect::<Vec<_>>();
    let unexpected = received
        .iter()
        .filter(|field| !expected.iter().any(|expected| &expected.name == *field))
        .cloned()
        .collect::<Vec<_>>();
    let corrections = likely_field_corrections(&received, &expected, &tool.name);

    let mut lines = vec![
        format!("Tool `{}` was deferred and has now been loaded.", tool.name),
        String::new(),
        "The tool was not executed. Retry with the loaded schema.".to_string(),
        String::new(),
        "Expected fields:".to_string(),
    ];
    if expected.is_empty() {
        lines.push("  (none)".to_string());
    } else {
        for field in &expected {
            let required_marker = if required.contains(&field.name) {
                " required"
            } else {
                ""
            };
            lines.push(format!(
                "  {}: {}{}",
                field.name, field.kind, required_marker
            ));
        }
    }
    lines.push(String::new());
    lines.push("Received fields:".to_string());
    if received.is_empty() {
        lines.push("  (none)".to_string());
    } else {
        lines.push(format!("  {}", received.join(", ")));
    }
    if !missing.is_empty() {
        lines.push(String::new());
        lines.push("Missing required fields:".to_string());
        lines.push(format!("  {}", missing.join(", ")));
    }
    if !unexpected.is_empty() {
        lines.push(String::new());
        lines.push("Unexpected fields:".to_string());
        lines.push(format!("  {}", unexpected.join(", ")));
    }
    if !corrections.is_empty() {
        lines.push(String::new());
        lines.push("Likely corrections:".to_string());
        for correction in &corrections {
            lines.push(format!("  {correction}"));
        }
    }

    ToolResult::success(lines.join("\n")).with_metadata(json!({
        "event": "tool.schema_hydrated",
        "tool": tool.name,
        "executed": false,
        "retry_required": true,
        "reason": "deferred_tool_first_use",
        "deferred_tool_loaded": true,
        "tool_name": tool.name,
        "expected_fields": expected.iter().map(|field| field.name.clone()).collect::<Vec<_>>(),
        "received_fields": received,
        "missing_required_fields": missing,
        "unexpected_fields": unexpected,
        "likely_corrections": corrections,
    }))
}

#[derive(Debug, Clone)]
struct SchemaField {
    name: String,
    kind: String,
}

fn schema_fields(schema: &Value) -> Vec<SchemaField> {
    let Some(properties) = schema.get("properties").and_then(Value::as_object) else {
        return Vec::new();
    };
    let mut fields = properties
        .iter()
        .map(|(name, spec)| SchemaField {
            name: name.clone(),
            kind: schema_type_label(spec),
        })
        .collect::<Vec<_>>();
    fields.sort_by(|a, b| a.name.cmp(&b.name));
    fields
}

fn schema_required_fields(schema: &Value) -> Vec<String> {
    let mut required = schema
        .get("required")
        .and_then(Value::as_array)
        .into_iter()
        .flatten()
        .filter_map(|value| value.as_str().map(str::to_string))
        .collect::<Vec<_>>();
    required.sort();
    required
}

fn schema_type_label(spec: &Value) -> String {
    let Some(kind) = spec.get("type").and_then(Value::as_str) else {
        return "value".to_string();
    };
    if let Some(values) = spec.get("enum").and_then(Value::as_array) {
        let labels = values.iter().filter_map(Value::as_str).collect::<Vec<_>>();
        if !labels.is_empty() {
            return format!("{kind} ({})", labels.join(" | "));
        }
    }
    kind.to_string()
}

fn received_field_names(input: &Value) -> Vec<String> {
    let mut fields = input
        .as_object()
        .map(|object| object.keys().cloned().collect::<Vec<_>>())
        .unwrap_or_default();
    fields.sort();
    fields
}

fn likely_field_corrections(
    received: &[String],
    expected: &[SchemaField],
    tool_name: &str,
) -> Vec<String> {
    let has_expected = |name: &str| expected.iter().any(|field| field.name == name);
    let has_received = |name: &str| received.iter().any(|field| field == name);
    let mut corrections = Vec::new();

    if has_received("old_string") && has_expected("search") {
        corrections.push("old_string -> search".to_string());
    } else if has_received("old_str") && has_expected("search") {
        corrections.push("old_str -> search".to_string());
    }
    if has_received("new_string") && has_expected("replace") {
        corrections.push("new_string -> replace".to_string());
    } else if has_received("new_str") && has_expected("replace") {
        corrections.push("new_str -> replace".to_string());
    } else if has_received("replacement") && has_expected("replace") {
        corrections.push("replacement -> replace".to_string());
    }
    if matches!(tool_name, "checklist_update" | "todo_update") && has_received("todos") {
        corrections.push(
            "Use work_update to replace the full list, or retry checklist_update/todo_update with id and status."
                .to_string(),
        );
    }
    // RLM source fields are easy to misname (#2659). rlm_open takes exactly one
    // of file_path / content / url / session_object; nudge common wrong names
    // toward those. The unified `rlm` tool carries the same fields for
    // action=open, so it gets the same correction.
    if matches!(tool_name, "rlm_open" | "rlm") {
        for wrong in [
            "prompt",
            "resident_file",
            "text",
            "body",
            "path",
            "file",
            "source",
        ] {
            if has_received(wrong)
                && !has_received("file_path")
                && !has_received("content")
                && !has_received("url")
                && !has_received("session_object")
            {
                corrections.push(format!("{wrong} -> file_path (local file), content (inline text), url, or session_object"));
            }
        }
    }
    corrections
}

pub(super) fn execute_tool_search(
    tool_name: &str,
    input: &serde_json::Value,
    catalog: &[Tool],
    active_tools: &mut HashSet<String>,
) -> Result<ToolResult, ToolError> {
    let query = required_str(input, "query")?;
    let match_kind = match tool_name {
        LEGACY_TOOL_SEARCH_REGEX_NAME => "regex",
        LEGACY_TOOL_SEARCH_BM25_NAME => "bm25",
        _ => optional_str(input, "match").unwrap_or("bm25"),
    };
    if !matches!(match_kind, "bm25" | "regex") {
        return Err(ToolError::invalid_input(format!(
            "Unsupported match algorithm '{match_kind}'. Expected one of: bm25, regex"
        )));
    }
    let max_results = usize::try_from(optional_u64(
        input,
        "max_results",
        TOOL_SEARCH_DEFAULT_MAX_RESULTS as u64,
    ))
    .unwrap_or(TOOL_SEARCH_DEFAULT_MAX_RESULTS)
    .clamp(1, TOOL_SEARCH_MAX_RESULTS_LIMIT);
    let discovered = if match_kind == "regex" {
        discover_tools_with_regex(catalog, query, max_results)?
    } else {
        discover_tools_with_bm25_like(catalog, query, max_results)
    };
    let remaining_results = max_results.saturating_sub(discovered.len());
    let unavailable = if match_kind == "regex" {
        unavailable_core_action_tools_with_regex(catalog, query, remaining_results)?
    } else {
        unavailable_core_action_tools_with_bm25_like(catalog, query, remaining_results)
    };

    for name in &discovered {
        active_tools.insert(name.clone());
    }

    let references = discovered
        .iter()
        .map(|name| json!({"type": "tool_reference", "tool_name": name}))
        .collect::<Vec<_>>();
    let unavailable_references = unavailable
        .iter()
        .map(|fallback| {
            json!({
                "type": "unavailable_tool_reference",
                "tool_name": fallback.name,
                "reason": fallback.unavailable_reason,
            })
        })
        .collect::<Vec<_>>();

    let payload = json!({
        "type": "tool_search_tool_search_result",
        "tool_references": references,
        "unavailable_tool_references": unavailable_references.clone(),
    });

    Ok(ToolResult {
        content: serde_json::to_string(&payload).unwrap_or_else(|_| payload.to_string()),
        success: true,
        metadata: Some(json!({
            "tool_references": discovered,
            "unavailable_tool_references": unavailable_references,
        })),
    })
}

pub(super) async fn execute_code_execution_tool(
    input: &serde_json::Value,
    workspace: &Path,
) -> Result<ToolResult, ToolError> {
    let code = required_str(input, "code")?;

    // Resolve the locally-installed Python interpreter we cached at
    // catalog-build time. If it's absent now (somehow registered but
    // disappeared between startup and this call — concurrent uninstall,
    // PATH change, etc.) the ExternalTool::tokio_command() will return
    // None and we fail fast with a clear message.
    //
    // Write the code to a temp file and execute it as a script rather
    // than passing it via `-c "<code>"`. Reasons:
    //   * `-c` has length limits (argv) on Windows.
    //   * Multiline code with quote nesting is brittle through `-c`.
    //   * Tracebacks reference a real filename instead of `<string>`,
    //     so the model can interpret line numbers correctly.
    // Tempfile lives only for the duration of this execution; Drop
    // removes it. We use `.py` so any shebang / encoding-sniffer
    // logic in the interpreter behaves normally.
    let temp_dir = tempfile::tempdir()
        .map_err(|e| ToolError::execution_failed(format!("tempdir failed: {e}")))?;
    let script_path = temp_dir.path().join("code_execution.py");
    tokio::fs::write(&script_path, code)
        .await
        .map_err(|e| ToolError::execution_failed(format!("tempfile write failed: {e}")))?;

    let mut cmd = crate::dependencies::Python::tokio_command().ok_or_else(|| {
        ToolError::execution_failed(
            "code_execution: Python interpreter became unavailable".to_string(),
        )
    })?;
    cmd.arg(&script_path).current_dir(workspace);

    let output = tokio::time::timeout(Duration::from_secs(120), cmd.output())
        .await
        .map_err(|_| ToolError::Timeout { seconds: 120 })
        .and_then(|res| res.map_err(|e| ToolError::execution_failed(e.to_string())))?;

    let stdout = String::from_utf8_lossy(&output.stdout).to_string();
    let stderr = String::from_utf8_lossy(&output.stderr).to_string();
    let return_code = output.status.code().unwrap_or(-1);
    let success = output.status.success();
    let payload = json!({
        "type": "code_execution_result",
        "stdout": stdout,
        "stderr": stderr,
        "return_code": return_code,
        "content": [],
    });

    Ok(ToolResult {
        content: serde_json::to_string(&payload).unwrap_or_else(|_| payload.to_string()),
        success,
        metadata: Some(payload),
    })
}

#[cfg(test)]
mod synthetic_name_tests {
    use super::{default_synthetic_catalog_tool_names, is_synthetic_catalog_tool};

    /// The published synthetic-name list and the predicate that classifies a
    /// catalog entry as synthetic must agree. A name that appears in the list
    /// but is not classified synthetic would let the request projection report
    /// a provenance the engine itself disputes.
    #[test]
    fn published_synthetic_names_agree_with_the_synthetic_predicate() {
        let names = default_synthetic_catalog_tool_names();
        assert!(!names.is_empty());
        for name in &names {
            assert!(
                is_synthetic_catalog_tool(name),
                "'{name}' is published as synthetic but the predicate disagrees"
            );
        }
        let mut sorted = names.clone();
        sorted.sort();
        sorted.dedup();
        assert_eq!(names, sorted, "the list must be sorted and deduplicated");

        // MCP names resolve through the real pool, so they are deliberately
        // absent here even though the predicate accepts them.
        assert!(!names.iter().any(|name| name.starts_with("mcp_")));

        // `multi_tool_use.parallel` is a call name, never a catalog entry, so
        // it has no catalog provenance and must not be published as synthetic.
        assert!(
            !names
                .iter()
                .any(|name| name == super::MULTI_TOOL_PARALLEL_NAME)
        );
    }
}