meerkat-core 0.6.0

Core agent logic for Meerkat (no I/O deps)
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
//! Agent runner interface.

use crate::budget::Budget;
use crate::error::{AgentError, ToolError};
use crate::event::AgentEvent;
use crate::hooks::{HookInvocation, HookPoint};
use crate::ops::{ToolDispatchOutcome, ToolDispatchTimeoutPolicy};
use crate::retry::RetryPolicy;
use crate::service::TurnToolOverlay;
use crate::session::{PendingSystemContextAppend, Session};
use crate::state::LoopState;
#[cfg(target_arch = "wasm32")]
use crate::tokio;
use crate::tool_scope::{
    EXTERNAL_TOOL_FILTER_METADATA_KEY, ExternalToolSurfaceBaseState,
    ExternalToolSurfaceDeltaOperation, ExternalToolSurfaceDeltaPhase,
    ExternalToolSurfaceEntrySnapshot, ExternalToolSurfaceSnapshot, ToolFilter, ToolScopeRevision,
    ToolScopeStageError,
};
use crate::turn_execution_authority::{
    TurnPrimitiveKind, TurnTerminalCauseKind, TurnTerminalOutcome,
};
use crate::types::{ContentInput, Message, RunResult, ToolCallView, ToolNameSet};
use async_trait::async_trait;
use serde_json::value::to_raw_value;
use std::collections::HashSet;
use std::sync::Arc;
use tokio::sync::mpsc;

use super::{Agent, AgentBuilder, AgentLlmClient, AgentSessionStore, AgentToolDispatcher};

fn dispatcher_knows_tool<T>(dispatcher: &T, name: &str) -> bool
where
    T: AgentToolDispatcher + ?Sized,
{
    if dispatcher.tool_catalog_capabilities().exact_catalog {
        dispatcher
            .tool_catalog()
            .iter()
            .any(|entry| entry.tool.name == name)
    } else {
        dispatcher.tools().iter().any(|tool| tool.name == name)
    }
}

fn precheck_visible_tool_call<T>(
    dispatcher: &T,
    visible_names: &ToolNameSet,
    name: &str,
) -> Result<(), ToolError>
where
    T: AgentToolDispatcher + ?Sized,
{
    if visible_names.contains(name) {
        return Ok(());
    }
    if dispatcher_knows_tool(dispatcher, name) {
        return Err(ToolError::access_denied(name));
    }
    Err(ToolError::not_found(name))
}

fn runtime_execution_snapshot(
    handle: &dyn crate::TurnStateHandle,
    applied_cursor: crate::completion_feed::CompletionSeq,
) -> Option<crate::AgentExecutionSnapshot> {
    let snapshot = handle.snapshot();
    let turn_phase = snapshot.turn_phase;
    // Typed handle contract: primitive_kind / terminal_outcome are
    // `Option<TurnPrimitiveKind>` / `Option<TurnTerminalOutcome>`. `None`
    // on the handle means "no primitive / no terminal outcome recorded
    // yet"; collapse to the typed `None` variant for downstream
    // consumers.
    let primitive_kind = snapshot.primitive_kind.unwrap_or(TurnPrimitiveKind::None);
    let terminal_outcome = snapshot
        .terminal_outcome
        .unwrap_or(TurnTerminalOutcome::None);
    let pending_operation_ids = if snapshot.pending_op_refs.is_empty() {
        None
    } else {
        Some(
            snapshot
                .pending_op_refs
                .iter()
                .map(|op_ref| op_ref.operation_id.clone())
                .collect(),
        )
    };
    let barrier_operation_ids = snapshot.barrier_operation_ids.iter().cloned().collect();

    Some(crate::AgentExecutionSnapshot {
        loop_state: snapshot.loop_state,
        turn_phase,
        active_run_id: snapshot.active_run_id,
        primitive_kind,
        admitted_content_shape: snapshot.admitted_content_shape,
        vision_enabled: snapshot.vision_enabled,
        image_tool_results_enabled: snapshot.image_tool_results_enabled,
        tool_calls_pending: u32::try_from(snapshot.tool_calls_pending).ok()?,
        pending_operation_ids,
        barrier_operation_ids,
        has_barrier_ops: snapshot.has_barrier_ops,
        barrier_satisfied: snapshot.barrier_satisfied,
        boundary_count: u32::try_from(snapshot.boundary_count).ok()?,
        cancel_after_boundary: snapshot.cancel_after_boundary,
        terminal_outcome,
        terminal_cause_kind: snapshot.terminal_cause_kind,
        extraction_attempts: u32::try_from(snapshot.extraction_attempts).ok()?,
        max_extraction_retries: u32::try_from(snapshot.max_extraction_retries).ok()?,
        applied_cursor,
    })
}

fn runtime_external_tool_surface_snapshot(
    handle: &dyn crate::ExternalToolSurfaceHandle,
) -> Option<ExternalToolSurfaceSnapshot> {
    let snapshot = handle.diagnostic_snapshot();
    let phase = snapshot.surface_phase;
    let visible_surfaces = snapshot.visible_surfaces;
    let snapshot_epoch = snapshot.snapshot_epoch;
    let snapshot_aligned_epoch = snapshot.snapshot_aligned_epoch;
    let mut entries = Vec::with_capacity(snapshot.entries.len());
    for entry in snapshot.entries {
        entries.push(ExternalToolSurfaceEntrySnapshot {
            visible: visible_surfaces.contains(&entry.surface_id),
            surface_id: entry.surface_id,
            // Typed handle contract: DSL projects a typed enum. `None`
            // means the DSL never recorded a value for this surface, so
            // the projection defaults to `Absent` / `None` per the
            // contract invariants (no state is equivalent to the zero
            // variant).
            base_state: entry
                .base_state
                .unwrap_or(ExternalToolSurfaceBaseState::Absent),
            has_removal_timing: entry.removal_draining_since_ms.is_some()
                || entry.removal_timeout_at_ms.is_some()
                || entry.removal_applied_at_turn.is_some(),
            pending_op: entry.pending_op,
            staged_op: entry.staged_op,
            staged_intent_sequence: entry.staged_intent_sequence.unwrap_or(0),
            pending_task_sequence: entry.pending_task_sequence.unwrap_or(0),
            pending_lineage_sequence: entry.pending_lineage_sequence.unwrap_or(0),
            inflight_call_count: entry.inflight_calls,
            last_delta_operation: entry
                .last_delta_operation
                .unwrap_or(ExternalToolSurfaceDeltaOperation::None),
            last_delta_phase: entry
                .last_delta_phase
                .unwrap_or(ExternalToolSurfaceDeltaPhase::None),
        });
    }
    Some(ExternalToolSurfaceSnapshot {
        phase,
        snapshot_epoch,
        snapshot_aligned_epoch,
        entries,
    })
}

/// Minimal runner interface for an Agent.
#[cfg_attr(target_arch = "wasm32", async_trait(?Send))]
#[cfg_attr(not(target_arch = "wasm32"), async_trait)]
pub trait AgentRunner: Send {
    async fn run(&mut self, prompt: ContentInput) -> Result<RunResult, AgentError>;

    async fn run_with_events(
        &mut self,
        prompt: ContentInput,
        tx: mpsc::Sender<AgentEvent>,
    ) -> Result<RunResult, AgentError>;
}

impl<C, T, S> Agent<C, T, S>
where
    C: AgentLlmClient + ?Sized,
    T: AgentToolDispatcher + ?Sized + 'static,
    S: AgentSessionStore + ?Sized,
{
    /// Stage an external tool visibility filter update for subsequent turns.
    pub fn stage_external_tool_filter(
        &mut self,
        filter: ToolFilter,
    ) -> Result<ToolScopeRevision, ToolScopeStageError> {
        // Durable visibility intent is machine-owned (or owned by the local
        // fallback owner for standalone builds). ToolScope only routes the
        // staging request and rebuilds the visible-tool projection.
        let handle = self.tool_scope.handle();
        let revision = handle.stage_external_filter(filter)?;
        let _ = handle.staged_revision();
        if let Ok(visibility_state) = self.tool_scope.visibility_state() {
            if let Err(err) = self.session.set_tool_visibility_state(visibility_state) {
                tracing::warn!(
                    error = %err,
                    "failed to persist staged canonical tool visibility state"
                );
            }
            self.session
                .remove_metadata(EXTERNAL_TOOL_FILTER_METADATA_KEY);
        }
        Ok(revision)
    }

    /// Set or clear a per-turn flow tool overlay.
    pub fn set_flow_tool_overlay(
        &mut self,
        overlay: Option<TurnToolOverlay>,
    ) -> Result<(), ToolScopeStageError> {
        let handle = self.tool_scope.handle();
        if let Some(overlay) = overlay {
            let allow = overlay
                .allowed_tools
                .map(|tools| tools.into_iter().collect::<HashSet<_>>());
            let deny = overlay
                .blocked_tools
                .unwrap_or_default()
                .into_iter()
                .collect::<HashSet<_>>();
            handle.set_turn_overlay(allow, deny)?;
        } else {
            handle.clear_turn_overlay();
        }
        Ok(())
    }

    pub fn set_runtime_execution_kind(
        &mut self,
        execution_kind: Option<crate::lifecycle::RuntimeExecutionKind>,
    ) {
        self.runtime_execution_kind = execution_kind;
    }

    fn clear_runtime_execution_kind(&mut self) {
        self.runtime_execution_kind = None;
    }

    fn require_runtime_execution_kind(&self) -> Result<(), AgentError> {
        if self.runtime_execution_kind_required && self.runtime_execution_kind.is_none() {
            return Err(AgentError::InternalError(
                "runtime_execution_kind not set: turn-state handle is attached but \
                 the runtime did not stamp RuntimeTurnMetadata.execution_kind"
                    .to_string(),
            ));
        }
        Ok(())
    }

    /// Apply accumulated session effects from tool dispatch.
    ///
    /// Called by the agent loop after each parallel tool batch completes.
    /// State-mutating effects are applied before `Message::ToolResults`; effects
    /// that append assistant transcript blocks are applied after tool results so
    /// provider tool-call adjacency remains intact.
    ///
    /// The session's `build_state` is the source of truth. The shared
    /// `mob_authority_handle` (if present) is updated as a derived projection
    /// after the canonical write succeeds.
    pub(crate) fn apply_session_effects(
        &mut self,
        effects: &[crate::ops::SessionEffect],
    ) -> Result<(), crate::error::AgentError> {
        use crate::error::AgentError;

        let mut build_state = self.session.build_state().unwrap_or_default();
        let mut build_state_changed = false;
        let mut visibility_changed = false;

        for effect in effects {
            match effect {
                crate::ops::SessionEffect::GrantManageMob { mob_id } => {
                    let authority =
                        build_state
                            .mob_tool_authority_context
                            .as_mut()
                            .ok_or_else(|| {
                                AgentError::InternalError(
                                "mob authority effect applied without canonical authority context"
                                    .into(),
                            )
                            })?;
                    authority.grant_manage_mob_in_place(mob_id.clone());
                    build_state_changed = true;
                }
                crate::ops::SessionEffect::RequestDeferredTools { authorities } => {
                    self.tool_scope
                        .add_requested_deferred_authorities(authorities)
                        .map_err(|err| {
                            AgentError::InternalError(format!(
                                "failed to record requested deferred tool authorities: {err}"
                            ))
                        })?;
                    visibility_changed = true;
                }
                crate::ops::SessionEffect::AppendAssistantBlocks { blocks } => {
                    self.session.push(crate::types::Message::BlockAssistant(
                        crate::types::BlockAssistantMessage::new(
                            blocks.clone(),
                            crate::types::StopReason::EndTurn,
                        ),
                    ));
                }
            }
        }

        if build_state_changed {
            self.session.set_build_state(build_state).map_err(|e| {
                AgentError::InternalError(format!(
                    "failed to persist session effects into build state: {e}"
                ))
            })?;
        }

        if visibility_changed && let Err(err) = self.publish_committed_visible_set() {
            return Err(AgentError::InternalError(format!(
                "failed to persist session effects into tool visibility state: {err}"
            )));
        }

        // Update the shared effective-authority handle so mob tools in
        // subsequent batches see the widened scope. The handle is a derived
        // projection of the canonical session build_state — it is never
        // treated as an independent truth source.
        if build_state_changed && let Some(ref handle) = self.mob_authority_handle {
            let updated = self
                .session
                .build_state()
                .and_then(|bs| bs.mob_tool_authority_context);
            if let Some(authority) = updated {
                *handle
                    .write()
                    .unwrap_or_else(std::sync::PoisonError::into_inner) = authority;
            }
        }

        Ok(())
    }

    /// Set the shared mob authority handle for session-effect application.
    ///
    /// The agent updates this handle after merging `SessionEffect`s from tool
    /// dispatch. Mob tools read from it for authorization checks.
    pub fn set_mob_authority_handle(
        &mut self,
        handle: Arc<std::sync::RwLock<crate::service::MobToolAuthorityContext>>,
    ) {
        self.mob_authority_handle = Some(handle);
    }

    /// Replace the LLM client for subsequent turns.
    ///
    /// Enables hot-swapping the model/provider on a live session without
    /// rebuilding the agent. The new client takes effect on the next
    /// `run()` / `run_with_events()` call.
    pub fn replace_client(&mut self, client: Arc<C>) {
        self.client = client;
    }

    /// Apply the live LLM request policy paired with an identity hot-swap.
    pub fn apply_llm_request_policy(&mut self, policy: crate::SessionLlmRequestPolicy) {
        self.config.model = policy.model;
        self.config.provider_params = policy.provider_params;
        self.config.provider_tool_defaults = policy.provider_tool_defaults;
    }

    /// Replace the LLM client and its next-turn request policy together.
    pub fn replace_client_with_request_policy(
        &mut self,
        client: Arc<C>,
        policy: crate::SessionLlmRequestPolicy,
    ) {
        self.replace_client(client);
        self.apply_llm_request_policy(policy);
    }

    /// Rotate runtime auth-lease tracking alongside a live LLM identity swap.
    pub fn rotate_auth_lease_auth_binding(
        &self,
        previous: Option<&crate::AuthBindingRef>,
        target: Option<&crate::AuthBindingRef>,
    ) -> Result<(), AgentError> {
        let Some(handle) = self.auth_lease_handle.as_deref() else {
            return Ok(());
        };
        if previous == target {
            return Ok(());
        }
        if let Some(previous) = previous {
            let previous_key = crate::handles::LeaseKey::from_auth_binding(previous);
            let _ = handle.release_lease(&previous_key);
        }
        if let Some(target) = target {
            let target_key = crate::handles::LeaseKey::from_auth_binding(target);
            let target_snapshot = handle.snapshot(&target_key);
            if target_snapshot.credential_present && target_snapshot.phase.is_some() {
                return Ok(());
            }
            handle.acquire_lease(&target_key, u64::MAX).map_err(|err| {
                AgentError::ConfigError(format!(
                    "failed to rotate auth lease to auth_binding {target_key}: {err}"
                ))
            })?;
        }
        Ok(())
    }

    /// Shared live control flag for boundary-only cancellation requests.
    pub fn cancel_after_boundary_handle(&self) -> Arc<std::sync::atomic::AtomicBool> {
        Arc::clone(&self.cancel_after_boundary_requested)
    }

    /// Persist the currently committed visible tool set into canonical session metadata.
    pub(crate) fn publish_committed_visible_set(&mut self) -> Result<(), AgentError> {
        // Session metadata is a durable projection/export of the canonical
        // visibility owner state so checkpoint/recovery stays aligned.
        let visibility_state = self.tool_scope.visibility_state().map_err(|err| {
            AgentError::InternalError(format!(
                "failed to snapshot canonical tool visibility state: {err}"
            ))
        })?;
        self.session
            .set_tool_visibility_state(visibility_state)
            .map_err(|err| {
                AgentError::InternalError(format!(
                    "failed to persist canonical tool visibility state: {err}"
                ))
            })
    }

    /// Dispatch one external tool call through the canonical tool dispatcher.
    ///
    /// This reuses the same visibility owner and session-effect application path
    /// as ordinary LLM-driven tool batches, but without synthesizing a full turn.
    pub async fn dispatch_external_tool_call(
        &mut self,
        call: crate::types::ToolCall,
    ) -> Result<ToolDispatchOutcome, AgentError> {
        self.dispatch_external_tool_call_with_timeout_policy(
            call,
            ToolDispatchTimeoutPolicy::Disabled,
        )
        .await
    }

    /// Dispatch an external product/runtime tool call with an optional
    /// caller-owned timeout. Timeout terminalization is still canonical:
    /// timeout expiry becomes `ToolError::Timeout`, then flows through
    /// `terminal_tool_outcome_for_error` like normal tool execution failures.
    pub async fn dispatch_external_tool_call_with_timeout_policy(
        &mut self,
        call: crate::types::ToolCall,
        timeout_policy: ToolDispatchTimeoutPolicy,
    ) -> Result<ToolDispatchOutcome, AgentError> {
        let visible_tool_names = self
            .tool_scope
            .visible_tool_names()
            .map_err(|err| AgentError::InternalError(err.to_string()))?
            .into_iter()
            .collect::<ToolNameSet>();
        if let Err(error) =
            precheck_visible_tool_call(self.tools.as_ref(), &visible_tool_names, call.name.as_str())
        {
            return Ok(crate::ops::terminal_tool_outcome_for_error(call.id, error));
        }
        let args = to_raw_value(&call.args).map_err(|err| {
            AgentError::InternalError(format!(
                "failed to serialize external tool-call arguments: {err}"
            ))
        })?;
        let view = ToolCallView {
            id: &call.id,
            name: &call.name,
            args: args.as_ref(),
        };
        let dispatch_result = match timeout_policy.timeout() {
            Some(timeout) => match tokio::time::timeout(timeout, self.tools.dispatch(view)).await {
                Ok(result) => result,
                Err(_) => Err(crate::error::ToolError::timeout(
                    call.name.clone(),
                    timeout_policy.timeout_ms().unwrap_or(u64::MAX),
                )),
            },
            None => self.tools.dispatch(view).await,
        };

        match dispatch_result {
            Ok(mut outcome) => {
                outcome.clear_terminal_cause();
                if outcome.result.tool_use_id.is_empty() {
                    outcome.result.tool_use_id = call.id;
                }
                if !outcome.session_effects.is_empty() {
                    self.apply_session_effects(&outcome.session_effects)?;
                }
                Ok(outcome)
            }
            Err(crate::error::ToolError::CallbackPending { tool_name, args }) => {
                Err(AgentError::CallbackPending { tool_name, args })
            }
            Err(error) => Ok(crate::ops::terminal_tool_outcome_for_error(call.id, error)),
        }
    }

    #[cfg(test)]
    pub(crate) fn inject_tool_scope_boundary_failure_once_for_test(&self) {
        self.tool_scope.inject_boundary_failure_once_for_test();
    }
}

impl<C, T, S> Agent<C, T, S>
where
    C: AgentLlmClient + ?Sized + 'static,
    T: AgentToolDispatcher + ?Sized + 'static,
    S: AgentSessionStore + ?Sized + 'static,
{
    /// Create a new agent builder
    pub fn builder() -> AgentBuilder {
        AgentBuilder::new()
    }

    /// Get the current session
    pub fn session(&self) -> &Session {
        &self.session
    }

    /// Get mutable access to the session (for setting metadata)
    pub fn session_mut(&mut self) -> &mut Session {
        &mut self.session
    }

    /// Get the current budget
    pub fn budget(&self) -> &Budget {
        &self.budget
    }

    /// Get the current state
    pub fn state(&self) -> LoopState {
        self.execution_snapshot()
            .map(|snapshot| snapshot.loop_state)
            .unwrap_or(LoopState::CallingLlm)
    }

    /// Snapshot the agent's live execution state for diagnostics and mapping.
    ///
    /// Returns `None` when the agent has no runtime-backed turn-state
    /// handle attached (standalone/ephemeral execution paths).
    pub fn execution_snapshot(&self) -> Option<crate::AgentExecutionSnapshot> {
        let handle = self.turn_state_handle.as_deref()?;
        runtime_execution_snapshot(handle, self.applied_cursor)
    }

    /// Snapshot the agent's live tool-scope state for diagnostics and mapping.
    pub fn tool_scope_snapshot(&self) -> Option<crate::ToolScopeSnapshot> {
        self.tool_scope.snapshot()
    }

    /// Snapshot the live external tool-surface state, if supported by the dispatcher chain.
    pub fn external_tool_surface_snapshot(&self) -> Option<crate::ExternalToolSurfaceSnapshot> {
        if let Some(handle) = self.external_tool_surface_handle.as_deref() {
            if let Some(snapshot) = runtime_external_tool_surface_snapshot(handle) {
                return Some(snapshot);
            }
            tracing::warn!(
                "failed to convert runtime external-tool-surface snapshot; falling back to dispatcher snapshot"
            );
        }
        self.tools.external_tool_surface_snapshot()
    }

    /// Get the retry policy
    pub fn retry_policy(&self) -> &RetryPolicy {
        &self.retry_policy
    }

    /// Get the current nesting depth
    pub fn depth(&self) -> u32 {
        self.depth
    }

    /// Get the event tap for interaction-scoped streaming.
    pub fn event_tap(&self) -> &crate::event_tap::EventTap {
        &self.event_tap
    }

    /// Access the live tool-scope projection bridge.
    pub fn tool_scope(&self) -> &crate::ToolScope {
        &self.tool_scope
    }

    /// Get shared runtime system-context control state.
    pub fn system_context_state(
        &self,
    ) -> Arc<std::sync::Mutex<crate::session::SessionSystemContextState>> {
        Arc::clone(&self.system_context_state)
    }

    /// Clone the current session with the latest shared system-context state merged into metadata.
    pub fn session_with_system_context_state(&self) -> Session {
        let mut session = self.session.clone();
        let state = match self.system_context_state.lock() {
            Ok(guard) => guard.clone(),
            Err(poisoned) => {
                tracing::warn!("system-context state lock poisoned while cloning session");
                poisoned.into_inner().clone()
            }
        };
        if let Err(err) = session.set_system_context_state(state) {
            tracing::warn!(error = %err, "failed to serialize system-context state into session");
        }
        if let Ok(visibility_state) = self.tool_scope.visibility_state()
            && let Err(err) = session.set_tool_visibility_state(visibility_state)
        {
            tracing::warn!(error = %err, "failed to serialize tool visibility state into session");
        }
        session
    }

    /// Synchronize the shared system-context state into the in-memory session metadata.
    #[doc(hidden)]
    pub fn sync_system_context_state_to_session(&mut self) {
        let state = match self.system_context_state.lock() {
            Ok(guard) => guard.clone(),
            Err(poisoned) => {
                tracing::warn!("system-context state lock poisoned while syncing session");
                poisoned.into_inner().clone()
            }
        };
        if let Err(err) = self.session.set_system_context_state(state) {
            tracing::warn!(error = %err, "failed to serialize system-context state into session");
        }
    }

    /// Consume all pending system-context appends for the next LLM boundary.
    ///
    /// The returned appends are intended for transient request composition only;
    /// they must not be written back into the canonical session prompt.
    pub(crate) fn take_pending_system_context_boundary(
        &mut self,
    ) -> Vec<PendingSystemContextAppend> {
        let pending = {
            let mut state = match self.system_context_state.lock() {
                Ok(guard) => guard,
                Err(poisoned) => {
                    tracing::warn!("system-context state lock poisoned while applying boundary");
                    poisoned.into_inner()
                }
            };
            if state.pending.is_empty() {
                return Vec::new();
            }
            let pending = state.pending.clone();
            state.mark_pending_applied();
            pending
        };

        self.sync_system_context_state_to_session();
        pending
    }

    pub(crate) fn llm_messages_with_runtime_system_context(
        &self,
        appends: &[PendingSystemContextAppend],
    ) -> Vec<Message> {
        if appends.is_empty() {
            return self.session.messages().to_vec();
        }

        let mut session = self.session.clone();
        session.append_system_context_blocks(appends);
        session.messages().to_vec()
    }

    /// Persist the current session through the configured checkpointer after syncing control state.
    #[doc(hidden)]
    pub async fn checkpoint_current_session(&mut self) {
        self.sync_system_context_state_to_session();
        if let Some(ref cp) = self.checkpointer {
            cp.checkpoint(&self.session).await;
        }
    }

    async fn run_started_hooks(
        &self,
        prompt: &ContentInput,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) -> Result<(), AgentError> {
        let report = self
            .execute_hooks(
                HookInvocation::run_started(self.session.id().clone(), prompt.clone()),
                event_tx,
            )
            .await?;

        if let Some(error) = report.denial_error(HookPoint::RunStarted) {
            return Err(error);
        }
        Ok(())
    }

    pub(super) async fn run_completed_hooks(
        &mut self,
        result: &mut RunResult,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) -> Result<(), AgentError> {
        let report = self
            .execute_hooks(
                HookInvocation::run_completed(self.session.id().clone(), result.turns),
                event_tx,
            )
            .await?;

        if let Some(error) = report.denial_error(HookPoint::RunCompleted) {
            return Err(error);
        }

        self.run_completed_hooks_applied = true;
        Ok(())
    }

    pub(super) async fn emit_run_completed_event(
        &self,
        result: &RunResult,
        extraction_required: bool,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) {
        let _ = crate::event_tap::tap_emit(
            &self.event_tap,
            event_tx,
            AgentEvent::RunCompleted {
                session_id: self.session.id().clone(),
                result: result.text.clone(),
                structured_output: result.structured_output.clone(),
                extraction_required,
                usage: result.usage.clone(),
                terminal_cause_kind: result.terminal_cause_kind,
            },
        )
        .await;
    }

    pub(super) async fn emit_extraction_succeeded_event(
        &self,
        structured_output: serde_json::Value,
        schema_warnings: Option<Vec<crate::schema::SchemaWarning>>,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) {
        let _ = crate::event_tap::tap_emit(
            &self.event_tap,
            event_tx,
            AgentEvent::ExtractionSucceeded {
                session_id: self.session.id().clone(),
                structured_output,
                schema_warnings,
            },
        )
        .await;
    }

    pub(super) async fn emit_extraction_failed_event(
        &self,
        error: &crate::types::ExtractionError,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) {
        let _ = crate::event_tap::tap_emit(
            &self.event_tap,
            event_tx,
            AgentEvent::ExtractionFailed {
                session_id: self.session.id().clone(),
                last_output: error.last_output.clone(),
                attempts: error.attempts,
                reason: error.reason.clone(),
            },
        )
        .await;
    }

    async fn emit_run_started_event(
        &self,
        prompt: ContentInput,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) {
        let _ = crate::event_tap::tap_emit(
            &self.event_tap,
            event_tx,
            AgentEvent::RunStarted {
                session_id: self.session.id().clone(),
                prompt,
            },
        )
        .await;
    }

    async fn emit_run_failed_event(
        &self,
        error: &AgentError,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) {
        let error_report = crate::event::AgentErrorReport::from_agent_error(error);
        let terminal_cause_kind = match error {
            AgentError::TerminalFailure { cause_kind, .. }
                if cause_kind.is_specific_failure_cause() =>
            {
                Some(*cause_kind)
            }
            _ => self
                .execution_snapshot()
                .and_then(|snapshot| snapshot.terminal_cause_kind)
                .filter(|cause_kind| *cause_kind != TurnTerminalCauseKind::Unknown),
        };
        let _ = crate::event_tap::tap_emit(
            &self.event_tap,
            event_tx,
            AgentEvent::RunFailed {
                session_id: self.session.id().clone(),
                error_class: error_report.class,
                error: error_report.message.clone(),
                terminal_cause_kind,
                error_report: Some(error_report),
            },
        )
        .await;
    }

    async fn handle_run_failure(
        &self,
        error: &AgentError,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) {
        if let Err(hook_err) = self.run_failed_hooks(error, event_tx).await {
            tracing::warn!(?hook_err, "run_failed hook execution failed");
        }
        self.emit_run_failed_event(error, event_tx).await;
    }

    async fn run_failed_hooks(
        &self,
        error: &AgentError,
        event_tx: Option<&mpsc::Sender<AgentEvent>>,
    ) -> Result<(), AgentError> {
        let report = self
            .execute_hooks(
                HookInvocation::run_failed(self.session.id().clone(), error),
                event_tx,
            )
            .await?;

        if let Some(error) = report.denial_error(HookPoint::RunFailed) {
            return Err(error);
        }
        Ok(())
    }

    /// Run the agent with a user message.
    pub async fn run(&mut self, user_input: ContentInput) -> Result<RunResult, AgentError> {
        self.run_inner(user_input, None).await
    }

    /// Run the agent with events streamed to the provided channel.
    pub async fn run_with_events(
        &mut self,
        user_input: ContentInput,
        event_tx: mpsc::Sender<AgentEvent>,
    ) -> Result<RunResult, AgentError> {
        self.run_inner(user_input, Some(event_tx)).await
    }

    /// Run the agent using the pending continuation boundary already in the session.
    ///
    /// This is useful when the session has been pre-populated with a continuation
    /// boundary (for example a deferred first-turn user message or staged callback
    /// tool results). Unlike `run()`, this method does NOT add a new user message;
    /// it runs directly from the session's current state.
    ///
    /// Returns an error if the session doesn't end at a resumable continuation
    /// boundary (`User` or `ToolResults`).
    pub async fn run_pending(&mut self) -> Result<RunResult, AgentError> {
        self.run_pending_inner(None).await
    }

    /// Run the agent using the pending continuation boundary, with event streaming.
    ///
    /// Like `run_pending()`, but emits events to the provided channel.
    pub async fn run_pending_with_events(
        &mut self,
        event_tx: mpsc::Sender<AgentEvent>,
    ) -> Result<RunResult, AgentError> {
        self.run_pending_inner(Some(event_tx)).await
    }

    /// Core run implementation shared by `run()` and `run_with_events()`.
    ///
    /// Adds user_input as a user message, emits lifecycle events when `event_tx`
    /// is provided, and delegates to `run_loop`.
    async fn run_inner(
        &mut self,
        user_input: ContentInput,
        event_tx: Option<mpsc::Sender<AgentEvent>>,
    ) -> Result<RunResult, AgentError> {
        let event_tx = event_tx.or_else(|| self.default_event_tx.clone());

        self.require_runtime_execution_kind()?;

        // Reset state for new run (allows multi-turn on same agent).
        self.extraction_state.reset();
        self.run_completed_hooks_applied = false;
        self.run_completed_event_emitted = false;

        // Apply canonical per-turn skill references staged by the surface.
        // Skill refs are text-only so they operate on the text projection.
        let user_input = if user_input.has_non_text_content() {
            // For multimodal input, prepend skill text to the text blocks only.
            let skill_text = self.apply_skill_ref(String::new()).await;
            if skill_text.is_empty() {
                user_input
            } else {
                // Prepend skill text as a leading text block.
                let mut blocks = vec![crate::types::ContentBlock::Text { text: skill_text }];
                blocks.extend(user_input.into_blocks());
                ContentInput::Blocks(blocks)
            }
        } else {
            let text = self.apply_skill_ref(user_input.text_content()).await;
            ContentInput::Text(text)
        };

        // Hooks/events receive the typed content input; legacy hook fields
        // still include the text projection for compatibility.
        let run_prompt_input = user_input.clone();

        // Add user message — preserve image blocks when present.
        let user_message = if user_input.has_non_text_content() {
            crate::types::UserMessage::with_blocks(user_input.into_blocks())
        } else {
            crate::types::UserMessage::text(user_input.text_content())
        };
        self.session.push(Message::User(user_message));

        self.emit_run_started_event(run_prompt_input.clone(), event_tx.as_ref())
            .await;

        if let Err(err) = self
            .run_started_hooks(&run_prompt_input, event_tx.as_ref())
            .await
        {
            self.handle_run_failure(&err, event_tx.as_ref()).await;
            self.clear_runtime_execution_kind();
            return Err(err);
        }

        match self.run_loop(event_tx.clone()).await {
            Ok(mut result) => {
                if !self.run_completed_hooks_applied
                    && let Err(err) = self
                        .run_completed_hooks(&mut result, event_tx.as_ref())
                        .await
                {
                    self.handle_run_failure(&err, event_tx.as_ref()).await;
                    self.clear_runtime_execution_kind();
                    return Err(err);
                }
                if !self.run_completed_event_emitted {
                    self.emit_run_completed_event(&result, false, event_tx.as_ref())
                        .await;
                    self.run_completed_event_emitted = true;
                }
                self.checkpoint_current_session().await;
                self.clear_runtime_execution_kind();
                Ok(result)
            }
            Err(err) => {
                self.handle_run_failure(&err, event_tx.as_ref()).await;
                self.clear_runtime_execution_kind();
                Err(err)
            }
        }
    }

    /// Core run-pending implementation shared by `run_pending()` and
    /// `run_pending_with_events()`.
    ///
    /// Uses the existing pending continuation boundary in the session (does NOT
    /// push a new user message). Emits lifecycle events when `event_tx` is
    /// provided. Also used by continuation paths after response injection or
    /// staged callback tool-result admission.
    pub(super) async fn run_pending_inner(
        &mut self,
        event_tx: Option<mpsc::Sender<AgentEvent>>,
    ) -> Result<RunResult, AgentError> {
        let event_tx = event_tx.or_else(|| self.default_event_tx.clone());

        let pending_prompt = self.session.messages().last().and_then(|m| match m {
            Message::User(u) => Some(u.text_content()),
            Message::ToolResults { .. } => Some(String::new()),
            _ => None,
        });

        let Some(prompt) = pending_prompt else {
            self.clear_runtime_execution_kind();
            return Err(AgentError::ConfigError(
                "run_pending requires a pending user or tool-results continuation boundary in the session".to_string(),
            ));
        };

        self.require_runtime_execution_kind()?;

        // Reset state for new run (allows multi-turn on same agent).
        self.extraction_state.reset();
        self.run_completed_hooks_applied = false;
        self.run_completed_event_emitted = false;

        self.emit_run_started_event(ContentInput::Text(prompt.clone()), event_tx.as_ref())
            .await;

        if let Err(err) = self
            .run_started_hooks(&ContentInput::Text(prompt.clone()), event_tx.as_ref())
            .await
        {
            self.handle_run_failure(&err, event_tx.as_ref()).await;
            self.clear_runtime_execution_kind();
            return Err(err);
        }

        match self.run_loop(event_tx.clone()).await {
            Ok(mut result) => {
                if !self.run_completed_hooks_applied
                    && let Err(err) = self
                        .run_completed_hooks(&mut result, event_tx.as_ref())
                        .await
                {
                    self.handle_run_failure(&err, event_tx.as_ref()).await;
                    self.clear_runtime_execution_kind();
                    return Err(err);
                }
                if !self.run_completed_event_emitted {
                    self.emit_run_completed_event(&result, false, event_tx.as_ref())
                        .await;
                    self.run_completed_event_emitted = true;
                }
                self.checkpoint_current_session().await;
                self.clear_runtime_execution_kind();
                Ok(result)
            }
            Err(err) => {
                self.handle_run_failure(&err, event_tx.as_ref()).await;
                self.clear_runtime_execution_kind();
                Err(err)
            }
        }
    }

    /// Cancel the current run
    pub fn cancel(&mut self) {
        use crate::turn_execution_authority::TurnExecutionInput;

        self.clear_runtime_execution_kind();
        let snapshot = self
            .turn_state_handle
            .as_deref()
            .map(crate::handles::TurnStateHandle::snapshot);
        let input = match snapshot.and_then(|s| s.active_run_id) {
            Some(run_id) => TurnExecutionInput::CancelNow { run_id },
            None => TurnExecutionInput::ForceCancelNoRun,
        };
        let _ = self.apply_turn_input(input);
    }

    /// Consume canonical pending `skill_references` staged by the surface and
    /// prepend resolved skill bodies to the next user input.
    ///
    /// Compatibility slash refs are handled at transport/resolver boundaries;
    /// core runtime no longer parses slash refs directly.
    async fn apply_skill_ref(&mut self, user_input: String) -> String {
        let engine = match &self.skill_engine {
            Some(e) => e.clone(),
            None => return user_input,
        };

        let mut prefix_parts: Vec<String> = Vec::new();

        // 1. Consume pending_skill_references (from wire format / API)
        if let Some(refs) = self.pending_skill_references.take()
            && !refs.is_empty()
        {
            let canonical_keys: Vec<crate::skills::SkillKey> = refs.into_iter().collect();
            match engine.resolve_and_render(&canonical_keys).await {
                Ok(resolved) => {
                    for skill in &resolved {
                        tracing::info!(
                            skill_key = %skill.key,
                            "Per-turn skill activation via skill_references"
                        );
                        prefix_parts.push(skill.rendered_body.clone());
                    }
                }
                Err(e) => {
                    tracing::warn!(
                        error = %e,
                        "Failed to resolve source-pinned skill_references"
                    );
                }
            }
        }

        if prefix_parts.is_empty() {
            return user_input;
        }

        if user_input.is_empty() {
            prefix_parts.join("\n\n")
        } else {
            format!("{}\n\n{user_input}", prefix_parts.join("\n\n"))
        }
    }
}