localharness 0.61.0

Agents that own themselves: one Rust crate that's both an agent SDK (streaming, tools, hooks, policies, triggers, MCP) and a wallet-owning, self-sovereign agent that runs in the browser.
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
//! The generic streaming TURN ENGINE (roadmap R7, phase 1) — ONE copy of the
//! turn-loop scaffold the streaming backends (gemini / anthropic / openai)
//! each re-implemented ~600 lines of.
//!
//! The engine owns everything that was byte-identical across the three loops:
//! the idle/cancel atomics, the pre-turn hook gate (deny BEFORE the prompt
//! enters history), the history push, the retry-wrapped stream open
//! ([`crate::backends::retry::open_stream_with_retry`]), the idle-stall arm
//! ([`crate::backends::stream_timeout`]), `MAX_TOOL_ROUNDS`, the `finish`-tool
//! special case + `finish_summary` capture, per-round usage `merge_round`, the
//! terminal `Step::turn_complete`, post-turn hooks, the compaction trigger,
//! the idle notify, and every error-exit path.
//!
//! Everything wire-specific lives behind [`TurnProvider`] — a STATIC-DISPATCH
//! trait on a zero-sized marker with associated `Message`/`Accum` types (the
//! proven [`crate::backends::compaction::CompactionModel`] pattern: no
//! async-trait objects, no `Send` bounds, wasm-safe by construction). The two
//! genuinely-async provider concerns — opening the model stream and running
//! compaction — are passed into [`run_turn`] as closures (exactly how the
//! compaction engine takes its `summarize` request), so the engine stays
//! client-agnostic and the spawned turn future's `Send`-ness is inferred at
//! monomorphization on native while staying `?Send`-clean on wasm.
//!
//! Two control-flow HOOKS (with engine defaults) cover the real divergence
//! found in the 3-way loop diff:
//! - [`TurnProvider::on_stream_end`] — anthropic's `pause_turn` resume
//!   (re-request against identical history, accumulators retained).
//! - [`TurnProvider::on_cancel_with_pending_calls`] — anthropic #82 / openai
//!   L22 tool-result balancing so a cancelled turn never leaves a dangling
//!   tool call that 400s the next request.
//!
//! Phase 1 migrated the openai loop; phase 2 migrated anthropic (proving
//! BOTH control-flow hooks: `pause_turn` resume + the #82 cancel balancing).
//! gemini follows in phase 3 (its loop is untouched — the seam above was
//! designed against all three).

use std::future::Future;
use std::sync::Arc;
use std::sync::atomic::Ordering;

use futures_core::Stream;
use serde_json::{json, Value};
use tracing::{debug, warn};
use uuid::Uuid;

use crate::backends::dispatch::{dispatch_post_turn, dispatch_tool_call, gate_pre_turn};
use crate::backends::gemini::tools::FINISH_TOOL_NAME;
use crate::backends::loop_util::extract_canonical_path;
use crate::backends::state::LoopState;
use crate::backends::stream_timeout::{idle_timeout_ms, next_with_idle_timeout, NextChunk};
use crate::content::Content;
use crate::error::{Error, Result};
use crate::hooks::{HookRunner, SessionContext};
use crate::tools::ToolRunner;
use crate::types::{
    Step, StepStatus, StreamChunk, ToolCall as NeutralToolCall, ToolResult, UsageMetadata,
};

/// Maximum dispatch rounds per turn — cap runaway tool loops. (Hoisted from
/// the identical per-backend consts.)
pub(crate) const MAX_TOOL_ROUNDS: u32 = 16;

/// One tool call resolved out of a round's stream accumulators, in wire
/// order. `id` is the provider's correlation id (`None` on Gemini, which
/// correlates by name); `parse_error` carries the malformed-streamed-args
/// message per the `loop_util::resolve_tool_args` convention.
pub(crate) struct ResolvedCall {
    pub id: Option<String>,
    pub name: String,
    pub args: Value,
    pub parse_error: Option<String>,
}

/// One dispatched call's outcome, handed back to the provider to shape into
/// wire tool-result message(s). `value` is the JSON surfaced to the model
/// (result / `{"ok":true}` finish sentinel / `{"error":..}` envelope);
/// `is_error` marks failures for wires that type them (anthropic `is_error`).
pub(crate) struct DispatchedResult {
    pub call: ResolvedCall,
    pub value: Value,
    /// Read by the anthropic provider (typed `tool_result.is_error`) —
    /// openai's wire has no error typing on tool messages, so this is dead
    /// on an openai-only build.
    #[allow(dead_code)]
    pub is_error: bool,
}

/// What to do when a round's stream ends (see [`TurnProvider::on_stream_end`]).
/// `Resume`/`ProceedAndEndTurn` are constructed by the anthropic provider's
/// `pause_turn` handling and the engine tests (dead on an openai-only build).
#[allow(dead_code)]
pub(crate) enum StreamEnd {
    /// The normal path: resolve calls, persist the assistant turn, dispatch.
    Proceed,
    /// Re-open the stream against IDENTICAL history, retaining the round's
    /// accumulators (anthropic `pause_turn`). Ignored while cancelled — a
    /// cancelled pause persists what streamed and ends the turn instead.
    Resume,
    /// Persist what streamed, then END the turn without dispatching tools
    /// (anthropic's pause-resume cap).
    ProceedAndEndTurn,
}

/// Per-round emit surface handed to [`TurnProvider::fold_event`]: the engine
/// owns the accumulated visible text (identical across backends) and the
/// `Step` delta plumbing; the provider just pushes what it decoded.
pub(crate) struct EmitCtx<'a, M> {
    state: &'a LoopState<M>,
    trajectory_id: &'a str,
    step_index: u32,
    text: String,
}

impl<M> EmitCtx<'_, M> {
    /// Append visible text and emit its streaming delta step.
    pub fn push_text(&mut self, t: &str) {
        if !t.is_empty() {
            self.text.push_str(t);
            self.state
                .emit(Step::text_delta(self.trajectory_id, self.step_index, t));
        }
    }

    /// Emit a thinking delta step (the provider keeps any thinking it must
    /// echo back — e.g. anthropic signed blocks — in its own accumulator).
    /// Consumed by the anthropic provider (gemini follows in R7 phase 3;
    /// dead on an openai-only build).
    #[allow(dead_code)]
    pub fn push_thought(&mut self, t: &str) {
        if !t.is_empty() {
            self.state
                .emit(Step::thought_delta(self.trajectory_id, self.step_index, t));
        }
    }
}

/// The per-backend seam: everything the engine needs to know about a wire.
/// Implemented on a zero-sized marker; the engine is monomorphized over it
/// (static dispatch — no trait objects, no async methods).
pub(crate) trait TurnProvider {
    /// The wire history message type (`openai::wire::Message`, …).
    type Message: Clone;
    /// The backend's `LoopConfig`.
    type Config;
    /// The request body sent per round.
    type Request: Clone;
    /// One decoded stream item (chunk/event).
    type Event;
    /// Per-round stream accumulator (tool-call fragments, finish reason,
    /// wire usage, provider-owned extras like anthropic thinking blocks).
    type Accum: Default;

    /// Build one round's request from config + a history snapshot.
    fn build_request(config: &Self::Config, history: &[Self::Message]) -> Self::Request;

    /// The compaction trigger threshold (`None` disables).
    fn compaction_threshold(config: &Self::Config) -> Option<u32>;

    /// Fold ONE stream event into the round accumulator, emitting deltas via
    /// `ctx`. An `Err` fails the turn through the shared stream-error path
    /// (anthropic's in-band `error` event).
    fn fold_event(
        acc: &mut Self::Accum,
        ctx: &mut EmitCtx<'_, Self::Message>,
        event: Self::Event,
    ) -> Result<()>;

    /// Drain the accumulated tool-call fragments into ordered, parsed calls
    /// (the `resolve_tool_args` convention: empty args are a valid no-arg
    /// call; a non-empty fragment that fails to parse carries `parse_error`).
    fn resolve_pending_calls(acc: &mut Self::Accum) -> Vec<ResolvedCall>;

    /// The round's usage in neutral form (default = "none reported").
    fn round_usage(acc: &Self::Accum) -> UsageMetadata;

    /// Map the round's provider finish/stop reason onto the terminal step's
    /// `(status, error message)`.
    fn map_finish_reason(acc: &Self::Accum) -> (StepStatus, &'static str);

    /// Assemble the assistant turn to persist (text + tool calls; anthropic
    /// leads with its signed thinking blocks out of `acc`). `None` skips the
    /// push (nothing streamed).
    fn assemble_assistant_message(
        acc: Self::Accum,
        text: &str,
        calls: &[ResolvedCall],
    ) -> Option<Self::Message>;

    /// Shape dispatched results into wire tool-result message(s) to append:
    /// openai returns one `role:"tool"` message PER call; gemini/anthropic
    /// return one batched user turn.
    fn tool_result_messages(results: Vec<DispatchedResult>) -> Vec<Self::Message>;

    /// HOOK: called when a round's stream ends, before resolving calls.
    /// Default: proceed. Anthropic overrides for `pause_turn` (`Resume` under
    /// its `MAX_PAUSE_RESUMES` cap via `pause_resumes`, else
    /// `ProceedAndEndTurn`).
    fn on_stream_end(_acc: &mut Self::Accum, _pause_resumes: u32) -> StreamEnd {
        StreamEnd::Proceed
    }

    /// HOOK: called when the turn is cancelled AFTER the assistant message
    /// carrying tool calls was persisted but BEFORE dispatch. Returns
    /// balancing messages to append so history stays wire-valid (openai L22 /
    /// anthropic #82 — a dangling tool call 400s the next request). Default:
    /// nothing (gemini's wire tolerates the dangle).
    fn on_cancel_with_pending_calls(_calls: &[ResolvedCall]) -> Vec<Self::Message> {
        Vec::new()
    }
}

/// Deps the engine needs per turn — the backend's `TurnDeps` minus the client
/// (the client rides inside the `open`/`compact` closures).
pub(crate) struct EngineDeps<P: TurnProvider> {
    pub config: P::Config,
    pub state: Arc<LoopState<P::Message>>,
    pub tool_runner: Option<Arc<ToolRunner>>,
    pub hook_runner: Option<Arc<HookRunner>>,
    pub session_ctx: Option<SessionContext>,
}

/// Emit the shared turn-failure shape (error step + idle release) and hand
/// the error back for the `return Err(..)` at the call site.
fn turn_fail<M>(state: &LoopState<M>, e: Error) -> Error {
    state.emit_error(e.to_string());
    state.idle.store(true, Ordering::Release);
    state.idle_notify.notify_waiters();
    e
}

/// Drive ONE user-initiated turn to completion: optionally many model ↔ tool
/// round-trips, terminating when the model stops with no tool calls (or calls
/// `finish`). `open` opens one model stream for a request (retried by the
/// shared policy); `compact` runs the provider's compaction fold (fired only
/// when the threshold trips).
pub(crate) async fn run_turn<P, St, Open, OFut, Compact, CFut>(
    deps: EngineDeps<P>,
    user: P::Message,
    prompt: Content,
    open: Open,
    compact: Compact,
) -> Result<()>
where
    P: TurnProvider,
    St: Stream<Item = Result<P::Event>> + Unpin,
    Open: Fn(P::Request) -> OFut,
    OFut: Future<Output = Result<St>>,
    Compact: FnOnce() -> CFut,
    CFut: Future<Output = ()>,
{
    deps.state.idle.store(false, Ordering::Release);
    // Fresh turn starts uncancelled — clear any stale stop from before.
    deps.state.cancel.store(false, Ordering::Release);

    // ONE turn context shared by the pre-turn gate, the per-call tool hooks,
    // and the post-turn hooks of this turn.
    let turn_ctx = deps
        .session_ctx
        .as_ref()
        .map(|s| s.child())
        .unwrap_or_default();

    // Pre-turn gate — BEFORE the prompt enters history, so a denied prompt
    // never pollutes context. On deny the model is never called; the
    // turn_error Step becomes a stream `Err` via `subscribe_step_stream`.
    if let Some(denied) = gate_pre_turn(deps.hook_runner.as_ref(), &turn_ctx, &prompt).await {
        return Err(turn_fail(&deps.state, Error::other(denied)));
    }

    deps.state.history.lock().push(user);
    *deps.state.last_turn_usage.lock() = Some(UsageMetadata::default());
    *deps.state.last_structured_output.lock() = None;

    let mut rounds = 0u32;
    let mut last_text = String::new();
    let mut last_status: (StepStatus, &'static str) = (StepStatus::Done, "");
    // The model called `finish` this turn — flags the terminal step as Finish.
    let mut finished_turn = false;
    // The closing `summary` arg from a `finish` call — painted on the terminal
    // step so a tool-only turn still ends with a reply, not an empty bubble.
    let mut finish_summary: Option<String> = None;
    let trajectory_id = Uuid::new_v4().to_string();

    loop {
        rounds += 1;
        if rounds > MAX_TOOL_ROUNDS {
            warn!(rounds, "exceeded MAX_TOOL_ROUNDS; forcing turn end");
            break;
        }
        // Stop requested before this round's model call — end the turn.
        if deps.state.cancel.load(Ordering::Acquire) {
            debug!("turn cancelled before model call");
            break;
        }

        let step_index = deps.state.alloc_step_index();
        let mut acc = P::Accum::default();
        let mut ctx = EmitCtx {
            state: &deps.state,
            trajectory_id: &trajectory_id,
            step_index,
            text: String::new(),
        };
        let mut pause_resumes = 0u32;

        // Stream(-resume) loop: normally one pass; `StreamEnd::Resume`
        // re-requests against identical history with accumulators retained.
        // Break value: end the turn after persisting (skip tool dispatch).
        let end_after_persist = 'request: loop {
            let request = P::build_request(&deps.config, &deps.state.history.lock());
            // Retry the stream OPEN on a transient transport/5xx/timeout (ONE
            // shared policy+wrapper, #29). A mid-stream error and
            // auth/credits/rate-limit still fail fast.
            let mut stream = match crate::backends::retry::open_stream_with_retry(|| {
                open(request.clone())
            })
            .await
            {
                Ok(s) => s,
                Err(e) => return Err(turn_fail(&deps.state, e)),
            };

            // Idle-stall guard: a fresh `idle_ms` timer is armed for EACH
            // event so a steadily streaming response never trips it — only
            // `idle_ms` of total silence does. On a stall we end the stream
            // with an Err so the turn returns via the normal error path and
            // the one-turn guard releases (vs. hanging on a dead socket the
            // cooperative cancel can't reach).
            let idle_ms = idle_timeout_ms();
            loop {
                let ev_res = match next_with_idle_timeout(&mut stream, idle_ms).await {
                    NextChunk::Item(item) => item,
                    NextChunk::End => break,
                    NextChunk::IdleTimeout => {
                        let e = Error::other(format!(
                            "model stream stalled — no data for {}s",
                            idle_ms / 1000
                        ));
                        return Err(turn_fail(&deps.state, e));
                    }
                };
                // Cooperative stop: drop the rest of this streamed response.
                if deps.state.cancel.load(Ordering::Acquire) {
                    break;
                }
                let ev = match ev_res {
                    Ok(c) => c,
                    Err(e) => return Err(turn_fail(&deps.state, e)),
                };
                if let Err(e) = P::fold_event(&mut acc, &mut ctx, ev) {
                    return Err(turn_fail(&deps.state, e));
                }
            }

            let cancelled = deps.state.cancel.load(Ordering::Acquire);
            match P::on_stream_end(&mut acc, pause_resumes) {
                StreamEnd::Resume if !cancelled => {
                    pause_resumes += 1;
                    debug!(pause_resumes, "provider resumed the stream");
                    continue 'request;
                }
                // A cancelled pause never resumes: persist what streamed and
                // end the turn (mirrors anthropic's `paused` break).
                StreamEnd::Resume | StreamEnd::ProceedAndEndTurn => break 'request true,
                StreamEnd::Proceed => break 'request false,
            }
        };

        // Resolve the accumulated tool calls, map the finish reason, and
        // persist the assistant turn.
        let pending_calls = P::resolve_pending_calls(&mut acc);
        last_status = P::map_finish_reason(&acc);
        let usage = P::round_usage(&acc);
        if let Some(msg) = P::assemble_assistant_message(acc, &ctx.text, &pending_calls) {
            deps.state.history.lock().push(msg);
        }

        // Accumulate usage across rounds.
        if usage != UsageMetadata::default() {
            let mut slot = deps.state.last_turn_usage.lock();
            match slot.as_mut() {
                Some(a) => a.merge_round(&usage),
                None => *slot = Some(usage),
            }
        }

        last_text = ctx.text;

        // No tool calls (or the provider ended the round) → turn over.
        if pending_calls.is_empty() || end_after_persist {
            break;
        }

        // Stop requested while streaming — end now instead of executing the
        // tools. The assistant message carrying these calls is already in
        // history; let the provider balance them so history stays wire-valid.
        if deps.state.cancel.load(Ordering::Acquire) {
            debug!("turn cancelled before tool dispatch");
            let balance = P::on_cancel_with_pending_calls(&pending_calls);
            deps.state.history.lock().extend(balance);
            break;
        }

        // Dispatch every tool call; results are shaped back onto the wire by
        // the provider at the end of the round.
        let mut results: Vec<DispatchedResult> = Vec::with_capacity(pending_calls.len());
        let mut saw_finish = false;
        for call in pending_calls {
            // Streamed args failed to parse — surface a clear tool error to
            // the model instead of running the tool with `{}`. Skip execution.
            if let Some(msg) = call.parse_error.clone() {
                let post_result = ToolResult {
                    name: call.name.clone(),
                    id: call.id.clone(),
                    result: Some(json!({ "error": msg.clone() })),
                    error: Some(msg.clone()),
                };
                deps.state
                    .emit_chunk_step(StreamChunk::ToolResult(post_result));
                results.push(DispatchedResult {
                    call,
                    value: json!({ "error": msg }),
                    is_error: true,
                });
                continue;
            }
            // `finish` is special: capture structured_output + the closing
            // summary, mark the turn complete, but still produce a result so
            // the model history stays well-formed.
            if call.name == FINISH_TOOL_NAME {
                if let Some(out) = call.args.get("output").cloned() {
                    *deps.state.last_structured_output.lock() = Some(out);
                }
                if let Some(sm) = call.args.get("summary").and_then(|v| v.as_str()) {
                    if !sm.is_empty() {
                        finish_summary = Some(sm.to_string());
                    }
                }
                saw_finish = true;
                results.push(DispatchedResult {
                    call,
                    value: json!({ "ok": true }),
                    is_error: false,
                });
                continue;
            }

            let tool_call = NeutralToolCall {
                name: call.name.clone(),
                args: call.args.clone(),
                id: call.id.clone(),
                canonical_path: extract_canonical_path(&call.args),
            };
            deps.state
                .emit_chunk_step(StreamChunk::ToolCall(tool_call.clone()));

            // The shared pipeline: pre-hooks → execute → error-lift →
            // post-hooks. `post_result.id` carries the tool-call id so
            // results correlate on id-keyed wires.
            let post_result = dispatch_tool_call(
                deps.tool_runner.as_ref(),
                deps.hook_runner.as_ref(),
                &turn_ctx,
                &tool_call,
            )
            .await;
            let value = post_result.result.clone().unwrap_or(Value::Null);
            let is_error = post_result.error.is_some();
            deps.state
                .emit_chunk_step(StreamChunk::ToolResult(post_result));

            results.push(DispatchedResult {
                call,
                value,
                is_error,
            });
        }

        // Push the tool results back into history in the provider's shape.
        deps.state
            .history
            .lock()
            .extend(P::tool_result_messages(results));

        if saw_finish {
            finished_turn = true;
            break;
        }
        // Otherwise loop and let the model react to the tool results.
    }

    let usage = deps.state.last_turn_usage.lock().clone().unwrap_or_default();
    let usage_opt = if usage == UsageMetadata::default() {
        None
    } else {
        Some(usage.clone())
    };

    let (status, error_msg) = last_status;
    let structured = deps.state.last_structured_output.lock().clone();
    let terminal = Step::turn_complete(
        trajectory_id,
        deps.state.alloc_step_index(),
        status,
        last_text.as_str(),
        error_msg,
        finished_turn,
        structured,
        usage_opt,
    )
    .with_finish_summary(finish_summary);
    deps.state.emit(terminal);

    // Post-turn hooks observe the completed turn's final text — fired after
    // the terminal step, never on denied or errored turns.
    dispatch_post_turn(deps.hook_runner.as_ref(), &turn_ctx, &last_text).await;

    // Compaction: if the turn pushed prompt tokens over the threshold,
    // summarize the old prefix before the next turn starts.
    let used = usage.prompt_token_count;
    if crate::backends::compaction::should_compact(
        used,
        P::compaction_threshold(&deps.config),
    ) {
        debug!(used, "compaction triggered");
        compact().await;
    }

    deps.state.idle.store(true, Ordering::Release);
    deps.state.idle_notify.notify_waiters();
    debug!(rounds, "turn complete");
    Ok(())
}

/// Test-only: fold a batch of events through a provider with a throwaway
/// [`EmitCtx`], so per-backend tests can exercise their fold seam directly
/// (the ctx fields are module-private by design).
#[cfg(test)]
pub(crate) fn test_fold_events<P: TurnProvider>(
    state: &LoopState<P::Message>,
    acc: &mut P::Accum,
    events: Vec<P::Event>,
) {
    let mut ctx = EmitCtx {
        state,
        trajectory_id: "test",
        step_index: 0,
        text: String::new(),
    };
    for ev in events {
        P::fold_event(acc, &mut ctx, ev).expect("fold_event ok");
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::hooks::TurnContext;
    use crate::types::{HookResult, StepSource, StepType};
    use parking_lot::Mutex;
    use std::collections::VecDeque;
    use tokio::sync::broadcast;

    /// A scripted stream event for the mock wire.
    #[derive(Clone)]
    enum Ev {
        Text(&'static str),
        Call {
            id: &'static str,
            name: &'static str,
            args: &'static str,
        },
        /// Ask `on_stream_end` to resume (the anthropic pause_turn shape).
        Resume,
        /// Flip the cancel atomic mid-stream (reachable via `ctx.state`).
        Cancel,
    }

    #[derive(Default)]
    struct Accum {
        calls: Vec<(String, String, String)>,
        resume: bool,
    }

    /// Mock wire: `Message = String` history entries with readable tags.
    struct MockProvider;
    impl TurnProvider for MockProvider {
        type Message = String;
        type Config = ();
        type Request = usize;
        type Event = Ev;
        type Accum = Accum;

        fn build_request(_c: &(), history: &[String]) -> usize {
            history.len()
        }
        fn compaction_threshold(_c: &()) -> Option<u32> {
            None
        }
        fn fold_event(acc: &mut Accum, ctx: &mut EmitCtx<'_, String>, ev: Ev) -> Result<()> {
            match ev {
                Ev::Text(t) => ctx.push_text(t),
                Ev::Call { id, name, args } => {
                    acc.calls.push((id.into(), name.into(), args.into()))
                }
                Ev::Resume => acc.resume = true,
                Ev::Cancel => ctx.state.cancel.store(true, Ordering::Release),
            }
            Ok(())
        }
        fn resolve_pending_calls(acc: &mut Accum) -> Vec<ResolvedCall> {
            std::mem::take(&mut acc.calls)
                .into_iter()
                .map(|(id, name, args)| {
                    let (args, parse_error) =
                        crate::backends::loop_util::resolve_tool_args(&name, &args);
                    ResolvedCall {
                        id: Some(id),
                        name,
                        args,
                        parse_error,
                    }
                })
                .collect()
        }
        fn round_usage(_acc: &Accum) -> UsageMetadata {
            UsageMetadata::default()
        }
        fn map_finish_reason(_acc: &Accum) -> (StepStatus, &'static str) {
            (StepStatus::Done, "")
        }
        fn assemble_assistant_message(
            _acc: Accum,
            text: &str,
            calls: &[ResolvedCall],
        ) -> Option<String> {
            (!text.is_empty() || !calls.is_empty())
                .then(|| format!("assistant:{text}:{}", calls.len()))
        }
        fn tool_result_messages(results: Vec<DispatchedResult>) -> Vec<String> {
            results
                .into_iter()
                .map(|r| format!("tool:{}:{}", r.call.id.unwrap_or_default(), r.value))
                .collect()
        }
        fn on_stream_end(acc: &mut Accum, pause_resumes: u32) -> StreamEnd {
            if std::mem::take(&mut acc.resume) {
                if pause_resumes < 2 {
                    StreamEnd::Resume
                } else {
                    StreamEnd::ProceedAndEndTurn
                }
            } else {
                StreamEnd::Proceed
            }
        }
        fn on_cancel_with_pending_calls(calls: &[ResolvedCall]) -> Vec<String> {
            calls
                .iter()
                .map(|c| format!("cancelled:{}", c.id.as_deref().unwrap_or_default()))
                .collect()
        }
    }

    type Steps = broadcast::Receiver<Step>;

    /// Run a turn over scripted streams: each inner Vec is one stream's
    /// events (`open` pops the next). Returns (state, steps rx, open count).
    async fn run(
        streams: Vec<Vec<Ev>>,
        hook_runner: Option<Arc<HookRunner>>,
    ) -> (Arc<LoopState<String>>, Steps, u32) {
        let (tx, rx) = broadcast::channel::<Step>(64);
        let state = Arc::new(LoopState::new(tx));
        let deps = EngineDeps::<MockProvider> {
            config: (),
            state: state.clone(),
            tool_runner: None,
            hook_runner,
            session_ctx: None,
        };
        let script = Mutex::new(streams.into_iter().collect::<VecDeque<_>>());
        let opens = std::sync::atomic::AtomicU32::new(0);
        let prompt = Content::text("hi");
        let res = run_turn::<MockProvider, _, _, _, _, _>(
            deps,
            "user:hi".to_string(),
            prompt,
            |_req| {
                opens.fetch_add(1, Ordering::SeqCst);
                let evs = script.lock().pop_front().unwrap_or_default();
                async move {
                    Ok(futures_util::stream::iter(
                        evs.into_iter().map(Ok::<_, Error>),
                    ))
                }
            },
            || async {},
        )
        .await;
        // Only the deny test expects Err; others assert on state.
        let _ = res;
        (state, rx, opens.load(Ordering::SeqCst))
    }

    struct DenyAllTurns;
    #[async_trait::async_trait]
    impl crate::hooks::PreTurnHook for DenyAllTurns {
        fn name(&self) -> &str {
            "test::deny_all_turns"
        }
        async fn run(&self, _ctx: &TurnContext, _prompt: &Content) -> Result<HookResult> {
            Ok(HookResult::deny("nope"))
        }
    }

    /// THE history-pollution invariant, pinned against the ENGINE's copy of
    /// the scaffold (mirrors the gemini loop's pinning test): a denied turn
    /// never pushes the prompt, never opens a stream, emits the System/Error
    /// turn_error shape, and releases the idle guard.
    #[tokio::test]
    async fn pre_turn_deny_keeps_prompt_out_of_history() {
        let hooks = Arc::new(HookRunner::new());
        hooks.register_pre_turn(Arc::new(DenyAllTurns));
        let (state, mut rx, opens) = run(vec![vec![Ev::Text("never")]], Some(hooks)).await;

        assert!(state.history.lock().is_empty(), "denied prompt must not enter history");
        assert_eq!(opens, 0, "the model must never be called on deny");
        assert!(state.idle.load(Ordering::Acquire), "idle guard must release");
        let step = rx.recv().await.expect("a step was broadcast");
        assert_eq!(step.source, StepSource::System);
        assert_eq!(step.status, StepStatus::Error);
        assert!(step.error.contains("turn denied by hook: nope"));
    }

    /// The finish-tool special case: structured output + summary captured,
    /// the tool result persisted through the provider, terminal step Finish.
    #[tokio::test]
    async fn finish_tool_ends_turn_with_summary_and_structured_output() {
        let (state, mut rx, opens) = run(
            vec![vec![
                Ev::Text("working"),
                Ev::Call {
                    id: "c1",
                    name: FINISH_TOOL_NAME,
                    args: r#"{"summary":"all done","output":{"x":1}}"#,
                },
            ]],
            None,
        )
        .await;

        assert_eq!(opens, 1);
        let hist = state.history.lock().clone();
        assert_eq!(
            hist,
            vec![
                "user:hi".to_string(),
                "assistant:working:1".to_string(),
                "tool:c1:{\"ok\":true}".to_string(),
            ],
            "user + assistant + finish result persisted in order"
        );
        // Drain to the terminal step.
        let mut terminal = None;
        while let Ok(s) = rx.try_recv() {
            if s.is_complete_response == Some(true) {
                terminal = Some(s);
            }
        }
        let t = terminal.expect("terminal step emitted");
        assert_eq!(t.kind, StepType::Finish);
        assert_eq!(t.finish_summary.as_deref(), Some("all done"));
        assert_eq!(t.structured_output, Some(json!({"x": 1})));
        assert_eq!(t.content, "working");
        assert!(state.idle.load(Ordering::Acquire));
    }

    /// `on_stream_end` Resume re-opens against identical history with the
    /// accumulators retained; the resume cap (`ProceedAndEndTurn`) persists
    /// what streamed and ends the turn WITHOUT dispatching the pending call.
    #[tokio::test]
    async fn stream_end_resume_reopens_then_end_turn_skips_dispatch() {
        let (state, _rx, opens) = run(
            vec![
                vec![Ev::Text("a"), Ev::Resume],
                vec![Ev::Text("b"), Ev::Resume],
                vec![
                    Ev::Call { id: "c9", name: "view_file", args: "{}" },
                    Ev::Resume, // cap (2 resumes done) → ProceedAndEndTurn
                ],
            ],
            None,
        )
        .await;

        assert_eq!(opens, 3, "one open + two resumes");
        let hist = state.history.lock().clone();
        assert_eq!(
            hist,
            vec!["user:hi".to_string(), "assistant:ab:1".to_string()],
            "text accumulates across resumes; the pending call persists but is NOT dispatched"
        );
    }

    /// Cancel between the assistant push and dispatch routes through
    /// `on_cancel_with_pending_calls` so the provider can balance history.
    #[tokio::test]
    async fn cancel_with_pending_calls_appends_the_providers_balance() {
        let (state, _rx, _opens) = run(
            vec![vec![
                Ev::Call { id: "c1", name: "view_file", args: r#"{"path":"a.rs"}"# },
                Ev::Cancel,
            ]],
            None,
        )
        .await;

        let hist = state.history.lock().clone();
        assert_eq!(
            hist,
            vec![
                "user:hi".to_string(),
                "assistant::1".to_string(),
                "cancelled:c1".to_string(),
            ],
            "the pending call is balanced, never dispatched"
        );
        assert!(state.idle.load(Ordering::Acquire));
    }
}