shep-daemon 0.1.26

The supervision engine behind shep: process spawning, restart policy, the log plane, watch and cron restarts, and the RPC server
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
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
//! The connection layer: peer auth, handshake, subscriptions
//!
//! [`RpcServer`] owns the bound [`Listener`] and accepts connections until
//! told to stop. Every accepted connection runs `handle_conn` (private — the
//! connection state machine is an implementation detail, not public API) in
//! its own task: a same-uid check ([`check_peer`], unix only), a version
//! handshake, then a read loop that decodes envelopes and hands them to
//! [`rpc::dispatch`](crate::rpc::dispatch) — the portable dispatcher Task 4
//! built, which never sees a socket or a byte.
//!
//! The OS transport lives one crate down in [`shep_core::transport`], so the
//! accept loop, the handshake and the connection state machine here are one
//! implementation over a unix socket and a Windows named pipe alike. The
//! single genuine platform difference left in this file is [`check_peer`].
//!
//! # Security
//!
//! See [`RpcServer`]'s doc for the canonical writeup; everything
//! security-relevant in this crate anchor-links there rather than repeating it.

use core::fmt;
use core::time::Duration;

use bytes::Bytes;
use futures_util::{SinkExt, StreamExt};
use serde::Serialize;
use tokio::sync::{mpsc, watch};
use tokio::task::JoinHandle;
use tokio_util::codec::{FramedRead, FramedWrite, LengthDelimitedCodec};

use shep_core::protocol::{
    Envelope, Hello, HelloAck, HelloReply, PROTOCOL_VERSION, RpcError, RpcErrorCode, WireError,
    codec, decode_frame, encode_frame,
};
use shep_core::transport::{Listener, ServerReadHalf, ServerStream, ServerWriteHalf};

use crate::bus::spawn_forwarder;
use crate::rpc::{Outcome, RpcContext, dispatch};
use crate::supervisor::ConnId;

/// Frames queued toward one client before the connection back-pressures.
pub const CONN_QUEUE: usize = 64;
/// How long a connected peer has to send its `Hello` before the daemon closes.
pub const HANDSHAKE_TIMEOUT_MS: u64 = 5_000;

type Frames = FramedRead<ServerReadHalf, LengthDelimitedCodec>;

/// The control socket — shep's privilege boundary
///
/// Wraps an already-bound [`Listener`] with the daemon-wide [`RpcContext`]
/// and drives the accept loop.
///
/// # Security
///
/// This is the canonical security writeup for the whole daemon (IR-29):
/// every other module that touches something security-relevant links back
/// here instead of re-arguing it.
///
/// Design criteria: `$SHEP_HOME/run` is *intended* to be created `0700` so
/// no other user can reach the socket path at all — that creation (and any
/// permission check of an already-existing `run` dir) is the daemon's boot
/// path's responsibility ([`crate::boot::init_dirs`], Task 6), not this
/// module's: nothing in [`RpcServer`] creates, checks, or enforces that
/// mode, it only accepts on whatever listener it is handed. Every accepted
/// connection is checked with `SO_PEERCRED`/`getpeereid` ([`check_peer`])
/// and refused unless the peer's uid equals the daemon's; this fails
/// CLOSED — a connection whose credentials the OS will not report
/// ([`AuthError::NoCredentials`]) is refused, not admitted, exactly like a
/// confirmed uid mismatch.
///
/// **Both of those sentences describe the unix tier.** Windows reaches the
/// same posture by a different route and one step earlier: there is no
/// `0700` directory and no post-accept credential check, because the control
/// pipe's own ACL denies a foreign local user the open-for-write that
/// speaking this protocol requires, and the OS enforces that at `CreateFile`
/// time before any byte reaches this module.
/// [`shep_core::transport`]'s module doc is the canonical account of that
/// difference, including what it does and does not cover.
///
/// The handshake refuses protocol skew with a typed
/// error ([`RpcErrorCode::ProtocolMismatch`]) rather than silence; frames
/// are capped at [`shep_core::protocol::MAX_FRAME_BYTES`]. Every
/// peer-supplied *pattern* is bounded before it can cost the daemon
/// unbounded work compiling it: a `Subscribe`'s topic-glob *count* (not the
/// byte length of any individual pattern, which is unbounded short of
/// `MAX_FRAME_BYTES`) is capped at [`crate::bus::MAX_TOPIC_PATTERNS`], and a
/// `/regex/` [`shep_core::protocol::SelectorSpec`] — on every verb that
/// carries one, without exception — is capped at a 1 MiB compiled size by
/// [`shep_core::selector::ProcessSelector`]'s `TryFrom<SelectorSpec>` impl,
/// which is the single door from the wire type to the matcher.
/// Every call carries
/// a clamped deadline ([`crate::rpc::budget`]). The one place this daemon
/// writes secrets to disk — an app's `env`, verbatim, so a muster restore
/// can reproduce it (spec §10 redacts them everywhere else) — is
/// [`crate::snapshot::write_atomic`]'s `flock.json`, created owner-only
/// (`0600`, Task 3) and kept there across its atomic rename.
///
/// A separate, install-time trust boundary: the CLI that daemonizes this
/// process hands it a readiness-pipe descriptor over `SHEP_READY_FD` (spec
/// §3), adopted through [`crate::sys::adopt_fd`] — this crate's only
/// `unsafe fn`, and (as of Decision 1, 2026-08-08) its ONLY unsafe surface,
/// full stop: [`crate::boot::boot`] receives the already-adopted
/// [`std::fs::File`] via [`crate::boot::BootOptions::ready_fd`] and never
/// touches a raw descriptor itself (`sys.rs`'s own doc has the full
/// test-call-site accounting). That boundary sits between this process and
/// its own parent, not between this process and an RPC peer — nothing
/// arriving over the control socket can reach it — and a hostile or stale
/// descriptor is refused (below fd 3, or not currently open) rather than
/// adopted; see [`crate::sys`]'s own rationale essay for the full threat
/// model.
///
/// Explicit non-goals: root can always read daemon memory; a peer with the
/// same uid is fully trusted (it could simply run the binary itself); there
/// is no post-handshake idle timeout (a same-uid peer that completes the
/// handshake and then never sends another frame holds its connection, and
/// this module's queue/task resources, open indefinitely); and there is no
/// cap on the number of concurrent connections one uid can hold open
/// ([`RpcServer::serve`] spawns and detaches unconditionally on every
/// accepted connection).
#[derive(Debug)]
pub struct RpcServer {
    listener: Listener,
    ctx: RpcContext,
}

impl RpcServer {
    /// Wraps an already-bound listener with the request-handling context.
    #[must_use]
    pub fn new(listener: Listener, ctx: RpcContext) -> Self {
        Self { listener, ctx }
    }

    /// Accepts connections, each on its own task, until `shutdown` flips to
    /// `true` or its sender drops.
    ///
    /// The accept loop is a `select!` over `listener.accept()` and
    /// `shutdown.changed()` — both cancel-safe, so neither branch loses
    /// state when the other resolves first. A transient accept error (e.g.
    /// `EMFILE`) is logged and the loop continues: one bad `accept()` must
    /// not take the whole daemon down.
    ///
    /// Each accepted connection's task is spawned and detached — this loop
    /// does not track or await it. That is fine for `serve` itself (it never
    /// blocks a still-running connection), but it means `serve` returning is
    /// not a guarantee that every in-flight connection has finished; a
    /// future daemon-shutdown sequence that needs to *drain* live
    /// connections before exiting will need its own seam here (a
    /// `tokio::task::JoinSet` in place of the bare `tokio::spawn`), not
    /// yet built.
    pub async fn serve(self, mut shutdown: watch::Receiver<bool>) {
        // `mut` because `Listener::accept` needs `&mut self` on both
        // platforms: a Windows named pipe server instance is consumed by
        // whoever connects to it, so accepting means handing that instance
        // out and creating the next one. See `shep_core::transport::Listener`.
        let Self { mut listener, ctx } = self;
        // A shutdown signal that was ALREADY `true` before this loop's first
        // `changed()` call would otherwise never be observed: `changed()`
        // only resolves on a value newer than the one this receiver has
        // last seen, and a receiver in fresh-from-the-daemon condition
        // hasn't seen anything past the value it was constructed with.
        if *shutdown.borrow() {
            return;
        }
        loop {
            tokio::select! {
                accepted = listener.accept() => {
                    match accepted {
                        Ok(stream) => {
                            let ctx = ctx.clone();
                            tokio::spawn(async move {
                                if let Err(err) = handle_conn(stream, ctx).await {
                                    tracing::debug!(%err, "connection ended");
                                }
                            });
                        }
                        Err(err) => tracing::warn!(%err, "accept failed; continuing"),
                    }
                }
                changed = shutdown.changed() => {
                    // An `Err` here means the sender dropped — treat that the
                    // same as an explicit `true`: either way, stop serving.
                    if changed.is_err() || *shutdown.borrow() {
                        break;
                    }
                }
            }
        }
    }
}

/// Checks that a connected peer runs as the daemon's own user.
///
// Peer-credential decision (deviation, deliberate): this uses
// `tokio::net::UnixStream::peer_cred()`, not
// `nix::sys::socket::getsockopt(PeerCredentials)`. nix 0.29 gates
// `PeerCredentials` behind `#[cfg(linux_android)]` — on macOS, a tier-1
// platform (spec §11), that sockopt does not exist and the daemon would not
// compile. tokio's `UCred` already dispatches to `SO_PEERCRED` on Linux,
// `getpeereid` on macOS/BSD, and `LOCAL_PEERCRED`/`getpeerucred` elsewhere,
// needs no new dependency, and adds no unsafe. nix is still used for
// `geteuid()` below, which has no such split.
///
/// # Errors
/// - [`AuthError::NoCredentials`] — the OS refused to report peer credentials.
/// - [`AuthError::ForeignUid`] — the peer's uid is not the daemon's.
///
/// # Platform
///
/// Unix only. There is no Windows counterpart, and there is not meant to be
/// one: see [`handle_conn`]'s own comment and
/// [`shep_core::transport`]'s module doc for why the pipe's ACL answers this
/// question earlier than any post-accept check could.
#[cfg(unix)]
pub fn check_peer(stream: &tokio::net::UnixStream, daemon_uid: u32) -> Result<u32, AuthError> {
    let cred = stream
        .peer_cred()
        .map_err(|err| AuthError::NoCredentials(err.to_string()))?;
    let peer = cred.uid();
    if peer == daemon_uid {
        Ok(peer)
    } else {
        Err(AuthError::ForeignUid {
            peer,
            daemon: daemon_uid,
        })
    }
}

/// The daemon's effective uid.
///
/// # Platform
///
/// Unix only, alongside [`check_peer`], its only caller.
#[cfg(unix)]
#[must_use]
pub fn daemon_uid() -> u32 {
    nix::unistd::geteuid().as_raw()
}

/// Why [`check_peer`] refused a connection.
///
/// `#[non_exhaustive]`: today's two variants cover a credentials read that
/// failed outright and one that succeeded but named the wrong uid, and a
/// future check — a group-membership or TLS peer-certificate requirement —
/// would need its own variant rather than stretching [`Self::ForeignUid`] to
/// mean something it does not, and shep-daemon is a published library an
/// out-of-tree matcher should not break for (IR-20).
#[non_exhaustive]
#[cfg_attr(windows, allow(dead_code))]
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum AuthError {
    /// The OS would not report peer credentials on this socket (carries the
    /// OS error message).
    NoCredentials(String),
    /// The peer runs as another user (carries both uids).
    ForeignUid {
        /// The connecting peer's uid.
        peer: u32,
        /// The daemon's own uid.
        daemon: u32,
    },
}

impl fmt::Display for AuthError {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::NoCredentials(msg) => write!(f, "could not read peer credentials: {msg}"),
            Self::ForeignUid { peer, daemon } => {
                write!(f, "peer uid {peer} does not match daemon uid {daemon}")
            }
        }
    }
}

impl core::error::Error for AuthError {}

/// Error type ending one connection.
///
/// Every variant is terminal: the connection layer logs it (via
/// [`RpcServer::serve`]'s spawn) and closes the socket. None of these panic
/// the daemon — a malformed or hostile peer can only ever cost itself its
/// own connection.
///
/// `#[non_exhaustive]`: the connection layer already distinguishes seven
/// failure points across auth, framing, encode/decode, and handshake
/// timing, and a future one — a TLS handshake failure, or a rate-limit
/// refusal — would add its own variant rather than overloading
/// [`Self::Auth`], which is specifically [`check_peer`]'s verdict, and
/// shep-daemon is a published library an out-of-tree matcher should not
/// break for (IR-20).
#[non_exhaustive]
#[derive(Debug)]
pub enum ConnError {
    /// [`check_peer`] refused the connection.
    Auth(AuthError),
    /// The framed transport failed reading or writing a length-delimited frame.
    Frame(std::io::Error),
    /// A frame's payload failed to decode as the expected type.
    Decode(WireError),
    /// A reply or event failed to encode onto the wire.
    Encode(WireError),
    /// The peer's `Hello.protocol` did not match [`PROTOCOL_VERSION`] (carries
    /// the client's claimed version; the refusal is written before this is
    /// returned).
    ProtocolMismatch {
        /// The protocol version the client sent.
        client: u32,
    },
    /// The peer did not send `Hello` within [`HANDSHAKE_TIMEOUT_MS`].
    HandshakeTimeout,
    /// The peer closed the connection before sending `Hello`.
    NoHandshake,
    /// The connection's write queue is gone — the writer task exited.
    PeerGone,
}

impl fmt::Display for ConnError {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::Auth(err) => write!(f, "peer-credential check failed: {err}"),
            Self::Frame(err) => write!(f, "frame transport error: {err}"),
            Self::Decode(err) => write!(f, "frame decode error: {err}"),
            Self::Encode(err) => write!(f, "frame encode error: {err}"),
            Self::ProtocolMismatch { client } => write!(
                f,
                "client sent protocol {client}, daemon speaks {PROTOCOL_VERSION}"
            ),
            Self::HandshakeTimeout => write!(
                f,
                "peer did not send Hello within {HANDSHAKE_TIMEOUT_MS} ms"
            ),
            Self::NoHandshake => f.write_str("peer closed before sending Hello"),
            Self::PeerGone => f.write_str("connection's write queue is gone"),
        }
    }
}

impl core::error::Error for ConnError {
    fn source(&self) -> Option<&(dyn core::error::Error + 'static)> {
        match self {
            Self::Auth(err) => Some(err),
            Self::Frame(err) => Some(err),
            Self::Decode(err) | Self::Encode(err) => Some(err),
            Self::ProtocolMismatch { .. }
            | Self::HandshakeTimeout
            | Self::NoHandshake
            | Self::PeerGone => None,
        }
    }
}

impl From<AuthError> for ConnError {
    fn from(source: AuthError) -> Self {
        Self::Auth(source)
    }
}

// `Decode` and `Encode` both wrap `WireError`, so only one of them could
// ever claim `impl From<WireError> for ConnError`: the compiler forbids a
// second one for the same source type, and picking one anyway would make a
// bare `?` silently mislabel the other direction. Both stay explicit
// `map_err` calls; see this task's own report.

impl From<std::io::Error> for ConnError {
    fn from(source: std::io::Error) -> Self {
        Self::Frame(source)
    }
}

// The ordering here is load-bearing (see `handshake` and `converse` below):
// auth before a single byte is read from the peer, the handshake before any
// request, and the writer task joined on every exit path so a protocol-skew
// refusal is guaranteed to reach the wire before the socket closes.
async fn handle_conn(stream: ServerStream, ctx: RpcContext) -> Result<(), ConnError> {
    // Unix only, and its absence on Windows is a deliberate design decision
    // rather than a gap — `shep_core::transport`'s module doc is the
    // canonical writeup. In short: on unix the peer is admitted by the
    // filesystem (`0700` on `$SHEP_HOME/run`) and this is the second layer
    // behind that; on Windows the pipe's own ACL refuses a foreign user's
    // open-for-write before a byte reaches this function, so the equivalent
    // check has already happened, earlier and in the kernel. Adding a
    // post-accept check there would need `ImpersonateNamedPipeClient` and
    // raw FFI to re-answer a question the OS already answered.
    #[cfg(unix)]
    check_peer(&stream, daemon_uid())?;
    // Minted after the peer check, not before: a connection refused for its
    // uid never reaches a handler, so it has nothing to scope.
    let conn = ConnId::next();
    let (read_half, write_half) = shep_core::transport::split(stream);
    let mut frames = FramedRead::new(read_half, codec());
    let (out_tx, out_rx) = mpsc::channel::<Bytes>(CONN_QUEUE);
    let writer = tokio::spawn(write_loop(FramedWrite::new(write_half, codec()), out_rx));

    let outcome = converse(&mut frames, &out_tx, conn, &ctx).await;

    // Drop the queue's sender and JOIN the writer before returning, on EVERY
    // path: a protocol-skew refusal is written by that task, so returning the
    // error early would close the socket before the client ever saw why.
    drop(out_tx);
    let _ = writer.await;
    // Beside the two lines above for the same reason their comment gives:
    // this block is on EVERY path out. A smit belongs to the connection that
    // painted it, and this is the one place that is true of. After the writer
    // join, so a client that painted and immediately read still sees its own
    // mark in the reply it was already sent.
    ctx.supervisor.forget_smits(conn).await;
    outcome
}

async fn converse(
    frames: &mut Frames,
    out: &mpsc::Sender<Bytes>,
    conn: ConnId,
    ctx: &RpcContext,
) -> Result<(), ConnError> {
    handshake(frames, out, ctx).await?;
    let mut forwarder: Option<JoinHandle<()>> = None;
    let outcome = read_loop(frames, out, conn, ctx, &mut forwarder).await;
    // EVERY path out of read_loop — Ok or any `?`-propagated Err — lands
    // here: a live forwarder MUST be aborted, not just dropped. Dropping a
    // JoinHandle detaches the task rather than stopping it, and a detached
    // forwarder keeps holding its own clone of `out` forever, which keeps
    // write_loop's `rx.recv()` from ever observing every sender gone —
    // which hangs `handle_conn`'s `writer.await` forever, which leaks the
    // task and the connection's fds and never actually closes the socket.
    // (Regression: see `a_garbage_frame_after_subscribing_still_closes_the_connection`.)
    if let Some(forwarder) = forwarder {
        forwarder.abort();
    }
    outcome
}

async fn read_loop(
    frames: &mut Frames,
    out: &mpsc::Sender<Bytes>,
    conn: ConnId,
    ctx: &RpcContext,
    forwarder: &mut Option<JoinHandle<()>>,
) -> Result<(), ConnError> {
    while let Some(frame) = frames.next().await {
        let frame = frame?; // oversize/short frame ends the connection
        let envelope: Envelope = decode_frame(&frame).map_err(ConnError::Decode)?;
        match dispatch(envelope, conn, ctx).await {
            Outcome::Reply(reply) => send(out, &reply).await?,
            Outcome::Subscribe { reply, filter } => {
                send(out, &reply).await?; // ordered ahead of any event by the queue
                // A second Subscribe REPLACES the first: spec §6 gives a
                // connection one topic list, not a growing union.
                if let Some(old) =
                    forwarder.replace(spawn_forwarder(&ctx.events, filter, out.clone()))
                {
                    old.abort();
                }
            }
            Outcome::Shutdown(reply) => {
                send(out, &reply).await?;
                ctx.shutdown();
                break;
            }
        }
    }
    Ok(())
}

async fn handshake(
    frames: &mut Frames,
    out: &mpsc::Sender<Bytes>,
    ctx: &RpcContext,
) -> Result<(), ConnError> {
    let frame = tokio::time::timeout(Duration::from_millis(HANDSHAKE_TIMEOUT_MS), frames.next())
        .await
        .map_err(|_| ConnError::HandshakeTimeout)?
        .ok_or(ConnError::NoHandshake)??;
    let hello: Hello = decode_frame(&frame).map_err(ConnError::Decode)?;
    if hello.protocol != PROTOCOL_VERSION {
        // Version skew is a typed error, not silence (spec §6).
        let refusal: HelloReply = Err(RpcError {
            code: RpcErrorCode::ProtocolMismatch,
            message: format!(
                "daemon speaks protocol {PROTOCOL_VERSION}, client sent {}",
                hello.protocol
            ),
            // The refusal names our PROTOCOL, which does not say which shep
            // is running. `shep daemon reload` chooses its mechanism by
            // version, and this is the one path where the ack that would
            // have carried it never arrives. Same field the ack uses, so a
            // client cannot learn two versions for one daemon.
            daemon_version: Some(ctx.daemon_version.clone()),
        });
        send(out, &refusal).await?;
        // The refusal is on the writer's queue before anything else
        // happens, so a peer that is about to be restarted still learns
        // why. Then: which dog was that, and what does this daemon owe it?
        //
        // `None` is every client that is not a dog, and the CLI is nearly
        // all of them. A `shep` verb has no way to name a dog here — the
        // name travels only on `ReconnectingClient`, the type dogs use and
        // no verb does — so the branch below cannot be reached by an
        // operator with a stray `$SHEP_DOG_NAME` in their environment.
        match &hello.dog_name {
            Some(dog) => {
                crate::dogs::record_refused_dog(
                    dog,
                    &hello.client_version,
                    &ctx.dog_refusals,
                    &ctx.supervisor,
                );
            }
            // A refused client this daemon cannot restart and would not
            // want to: an operator running an older `shep` against a newer
            // daemon, who is already reading the skew in plain English from
            // their own CLI. `debug!` rather than `warn!`, which this was
            // until a real reload across a protocol bump measured it: the
            // CLI polls the socket while it waits for a successor, so ONE
            // `shep daemon reload` produced 442 of these in 9.8 seconds,
            // and the daemon's default level is `warn`. The only fact this
            // adds over `handle_conn`'s own "connection ended" line is the
            // client's crate version, which is worth carrying and is not
            // worth waking anybody for.
            None => tracing::debug!(
                client_protocol = hello.protocol,
                client_version = %hello.client_version,
                "refused a client on protocol skew"
            ),
        }
        return Err(ConnError::ProtocolMismatch {
            client: hello.protocol,
        });
    }
    // A dog that got in is not stale, whatever this daemon held against it
    // before — including the restart it was just given, which is exactly
    // the case that has to clear.
    if let Some(dog) = &hello.dog_name {
        ctx.dog_refusals.handshook(dog);
    }
    let ack: HelloReply = Ok(HelloAck {
        daemon_version: ctx.daemon_version.clone(),
        protocol: PROTOCOL_VERSION,
        pid: ctx.pid,
    });
    send(out, &ack).await
}

async fn write_loop(
    mut sink: FramedWrite<ServerWriteHalf, LengthDelimitedCodec>,
    mut rx: mpsc::Receiver<Bytes>,
) {
    while let Some(bytes) = rx.recv().await {
        if sink.send(bytes).await.is_err() {
            break; // peer gone; nothing left to drain the queue
        }
    }
}

async fn send<T: Serialize>(out: &mpsc::Sender<Bytes>, value: &T) -> Result<(), ConnError> {
    let bytes = encode_frame(value).map_err(ConnError::Encode)?;
    out.send(bytes).await.map_err(|_| ConnError::PeerGone)
}

#[cfg(test)]
mod tests {
    // Real time: every test here drives a real UnixStream. Under a paused
    // clock the runtime auto-advances whenever it goes idle, which can expire
    // HANDSHAKE_TIMEOUT_MS before the peer's bytes are delivered.
    use super::*;
    use crate::bus::SharedEvent;
    use crate::fake::{FIRST_SCRIPTED_PID, ProcScript};
    use crate::testing::harness;
    use futures_util::{SinkExt, StreamExt};
    use serde::Serialize;
    use serde::de::DeserializeOwned;
    use shep_core::protocol::{
        BusEvent, DogSource, Envelope, Hello, HelloReply, PROTOCOL_VERSION, ProcessEventKind,
        ProcessInfo, Request, Response, RpcErrorCode, ServerFrame, codec, decode_frame,
        encode_frame,
    };
    use shep_core::status::ProcStatus;
    use tokio_util::codec::Framed;

    const RECV_TIMEOUT: Duration = Duration::from_secs(5);

    struct Client {
        frames: Framed<shep_core::transport::ClientStream, tokio_util::codec::LengthDelimitedCodec>,
    }

    impl Client {
        async fn send<T: Serialize>(&mut self, value: &T) {
            self.frames
                .send(encode_frame(value).unwrap())
                .await
                .unwrap();
        }

        async fn recv<T: DeserializeOwned>(&mut self) -> T {
            let frame = tokio::time::timeout(RECV_TIMEOUT, self.frames.next())
                .await
                .expect("timed out waiting for a frame")
                .expect("connection closed early")
                .unwrap();
            decode_frame(&frame).unwrap()
        }

        async fn closed(&mut self) -> bool {
            tokio::time::timeout(RECV_TIMEOUT, self.frames.next())
                .await
                .expect("timed out waiting for close")
                .is_none()
        }
    }

    /// Spawns `handle_conn` over a real connected pair and hands back the
    /// client end.
    ///
    /// A real transport on both platforms — a socketpair on unix, an actual
    /// named pipe on Windows — rather than an in-memory duplex, because
    /// several tests below turn on what a peer sees when the other side
    /// closes, which only a real transport reproduces. `async` (it was not,
    /// when it could call the synchronous `UnixStream::pair`) because
    /// creating a pipe pair means connecting one, and every caller is
    /// already inside a `#[tokio::test]`.
    async fn connected(ctx: RpcContext) -> Client {
        let (server, client) = shep_core::transport::connected_pair().await.unwrap();
        tokio::spawn(async move {
            let _ = handle_conn(server, ctx).await;
        });
        Client {
            frames: Framed::new(client, codec()),
        }
    }

    #[tokio::test]
    async fn handshake_acks_a_matching_protocol() {
        let h = harness(vec![]); // same helper shape as rpc.rs's tests
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.0".to_string(),
                protocol: PROTOCOL_VERSION,
                dog_name: None,
            })
            .await;
        let ack: HelloReply = client.recv().await;
        let ack = ack.expect("a matching protocol must be acked");
        assert_eq!(ack.protocol, PROTOCOL_VERSION);
        assert_eq!(ack.pid, h.ctx.pid);
        assert_eq!(ack.daemon_version, h.ctx.daemon_version);
    }

    #[tokio::test]
    async fn handshake_refuses_protocol_skew_before_closing() {
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "9.9.9".to_string(),
                protocol: PROTOCOL_VERSION + 1,
                dog_name: None,
            })
            .await;
        let refusal: HelloReply = client.recv().await;
        let err = refusal.expect_err("skew must be refused");
        assert_eq!(err.code, RpcErrorCode::ProtocolMismatch);
        assert!(
            client.closed().await,
            "the daemon must close after refusing"
        );
    }

    #[tokio::test]
    async fn a_protocol_refusal_carries_the_daemon_version() {
        // The refusal reports the daemon's PROTOCOL, which says nothing
        // about which crate version is running. `shep daemon reload` picks
        // between a handover and a stop-and-start by version, and a protocol
        // bump is exactly when it is needed, so the refusal has to name it.
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "9.9.9".to_string(),
                protocol: PROTOCOL_VERSION + 1,
                dog_name: None,
            })
            .await;
        let refusal: HelloReply = client.recv().await;
        let err = refusal.expect_err("skew must be refused");
        assert_eq!(err.code, RpcErrorCode::ProtocolMismatch);
        // The same string the ack would have carried, from the same field:
        // a client must never learn two different versions for one daemon
        // depending on whether its handshake was accepted.
        assert_eq!(err.daemon_version.as_deref(), Some(&*h.ctx.daemon_version));
    }

    #[tokio::test]
    async fn a_request_before_the_handshake_ends_the_connection() {
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Envelope {
                id: 1,
                deadline_ms: None,
                body: Request::Ping,
            })
            .await;
        assert!(client.closed().await);
    }

    #[tokio::test]
    async fn ping_round_trips_over_the_socket() {
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.0".to_string(),
                protocol: PROTOCOL_VERSION,
                dog_name: None,
            })
            .await;
        let _: HelloReply = client.recv().await;
        client
            .send(&Envelope {
                id: 11,
                deadline_ms: Some(1000),
                body: Request::Ping,
            })
            .await;
        let frame: ServerFrame = client.recv().await;
        let ServerFrame::Reply(reply) = frame else {
            panic!("expected a reply frame")
        };
        assert_eq!(reply.id, 11);
        assert_eq!(reply.result.unwrap(), Response::Pong);
    }

    #[tokio::test]
    async fn subscribe_streams_only_matching_events() {
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.0".to_string(),
                protocol: PROTOCOL_VERSION,
                dog_name: None,
            })
            .await;
        let _: HelloReply = client.recv().await;
        client
            .send(&Envelope {
                id: 1,
                deadline_ms: None,
                body: Request::Subscribe {
                    topics: vec!["process.*".to_string()],
                },
            })
            .await;
        let frame: ServerFrame = client.recv().await;
        assert!(matches!(frame, ServerFrame::Reply(ref r) if r.result == Ok(Response::Subscribed)));

        let event = |kind| -> SharedEvent {
            BusEvent::Process {
                event: kind,
                info: ProcessInfo::builder(0, "web", shep_core::status::ProcStatus::Online)
                    .pid(Some(1000))
                    .out_file(Some("/logs/web-0-out.log".to_string()))
                    .err_file(Some("/logs/web-0-err.log".to_string()))
                    .build(),
                manually: false,
                at_ms: 0,
            }
            .into()
        };
        h.ctx.events.send(event(ProcessEventKind::Start)).unwrap();
        h.ctx
            .events
            .send(
                BusEvent::LogOut {
                    id: 0,
                    line: "filtered".to_string(),
                }
                .into(),
            )
            .unwrap();
        h.ctx.events.send(event(ProcessEventKind::Online)).unwrap();

        // Back-to-back arrival is the filtering assertion — no negative wait.
        let first: ServerFrame = client.recv().await;
        let second: ServerFrame = client.recv().await;
        assert!(matches!(
            first,
            ServerFrame::Event(BusEvent::Process {
                event: ProcessEventKind::Start,
                ..
            })
        ));
        assert!(matches!(
            second,
            ServerFrame::Event(BusEvent::Process {
                event: ProcessEventKind::Online,
                ..
            })
        ));
    }

    #[tokio::test]
    async fn a_garbage_frame_ends_the_connection_without_panicking() {
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.0".to_string(),
                protocol: PROTOCOL_VERSION,
                dog_name: None,
            })
            .await;
        let _: HelloReply = client.recv().await;
        client
            .frames
            .send(bytes::Bytes::from_static(b"not json"))
            .await
            .unwrap();
        assert!(client.closed().await);
    }

    #[tokio::test]
    async fn a_garbage_frame_after_subscribing_still_closes_the_connection() {
        // Regression test (Opus security review, post-Task-5): a live
        // forwarder holds its own clone of `out`. If a connection error
        // (garbage frame, oversize frame, ...) after Subscribe skipped
        // aborting that forwarder, `out_tx`'s drop in `handle_conn` would NOT
        // be the last sender — write_loop's `rx.recv()` would never see
        // every sender go away, `handle_conn`'s `writer.await` would hang
        // forever, and the socket would never actually close. Subscribing
        // first is what makes that path reachable; the plain
        // `a_garbage_frame_ends_the_connection_without_panicking` test above
        // never subscribes, so it could not have caught this.
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.0".to_string(),
                protocol: PROTOCOL_VERSION,
                dog_name: None,
            })
            .await;
        let _: HelloReply = client.recv().await;
        client
            .send(&Envelope {
                id: 1,
                deadline_ms: None,
                body: Request::Subscribe {
                    topics: vec!["*".to_string()],
                },
            })
            .await;
        let _: ServerFrame = client.recv().await; // the Subscribed reply
        client
            .frames
            .send(bytes::Bytes::from_static(b"not json"))
            .await
            .unwrap();
        assert!(
            client.closed().await,
            "a live forwarder must not keep the connection open past a decode error"
        );
    }

    /// `cfg(unix)`, like [`check_peer`] itself. Windows has no counterpart
    /// to gate: the pipe's ACL refuses a foreign user's open before
    /// `handle_conn` is reached at all, so there is no post-accept decision
    /// here for a test to exercise. See `shep_core::transport`'s module doc.
    #[cfg(unix)]
    #[tokio::test]
    async fn peer_credentials_gate_on_uid() {
        // What this DOES prove: check_peer's own uid-comparison and error
        // construction are correct — same-uid accepts and hands back the
        // uid, a mismatched daemon_uid is refused as ForeignUid with both
        // uids recorded.
        //
        // What this does NOT prove: that a connection from an actual
        // different-uid OS process is rejected. `UnixStream::pair()` always
        // reports both ends as owned by this test process's own uid (there is
        // no way to fake the peer side of `SO_PEERCRED`/`getpeereid` — the
        // kernel derives it from the real socket, and `tokio::net::unix::UCred`
        // has no public constructor for a synthetic one), so this test can
        // only vary the `daemon_uid` argument, never the peer's true uid.
        // Exercising the real cross-uid path needs a second OS user (root in
        // CI, or two accounts) actually connecting — out of reach for this
        // crate's test harness; see the report's security-review note.
        let (a, _b) = tokio::net::UnixStream::pair().unwrap();
        let me = daemon_uid();
        assert_eq!(check_peer(&a, me).unwrap(), me);
        assert_eq!(
            check_peer(&a, me + 1).unwrap_err(),
            AuthError::ForeignUid {
                peer: me,
                daemon: me + 1
            }
        );
    }

    #[tokio::test]
    async fn a_slow_subscriber_gets_a_dropped_notice_instead_of_hanging_the_bus() {
        // Adversarial finding #3: bus.rs's `step()` unit test proves the
        // Lagged->Dropped translation in isolation, but nothing before this
        // exercised it through the REAL connection stack — CONN_QUEUE
        // filling, the forwarder parking on `out.send`, and the broadcast
        // ring (bus.rs's BUS_CAPACITY) actually overflowing for a subscriber
        // that truly never reads. Real time: real socket, matching this
        // whole test mod's rule.
        let h = harness(vec![]);
        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.0".to_string(),
                protocol: PROTOCOL_VERSION,
                dog_name: None,
            })
            .await;
        let _: HelloReply = client.recv().await;
        client
            .send(&Envelope {
                id: 1,
                deadline_ms: None,
                body: Request::Subscribe {
                    topics: vec!["log.*".to_string()],
                },
            })
            .await;
        let _: ServerFrame = client.recv().await; // the Subscribed reply

        // Never call client.recv() again until AFTER the flood: CONN_QUEUE
        // fills, the forwarder blocks on `out.send`, and the broadcast ring
        // (BUS_CAPACITY) takes the overflow from there — the exact
        // back-pressure chain bus.rs's module comment documents.
        let flood = crate::bus::BUS_CAPACITY + CONN_QUEUE + 16;
        for i in 0..flood {
            h.ctx
                .events
                .send(
                    BusEvent::LogOut {
                        id: 0,
                        line: format!("line-{i}"),
                    }
                    .into(),
                )
                .unwrap();
        }

        // Resume reading. The count comes from tokio's own Lagged(n) inside
        // the forwarder, never hand-computed here (no-hand-computed-sequences
        // rule) — this only asserts a Dropped notice arrives and is nonzero.
        let dropped = loop {
            match client.recv::<ServerFrame>().await {
                ServerFrame::Event(BusEvent::Dropped { count }) => break count,
                ServerFrame::Event(_) => continue,
                other => panic!("expected eventually a Dropped notice, got {other:?}"),
            }
        };
        assert!(
            dropped > 0,
            "a flood past CONN_QUEUE + BUS_CAPACITY must report a real lag"
        );
    }

    // --- G8: what a refused DOG's handshake costs it ------------------
    //
    // A refused handshake never reaches a request, so `Request::DogConfig`'s
    // name — the one place a dog otherwise identifies itself — is
    // unreachable on exactly the path where the daemon has to know which
    // dog it just refused. `Hello.dog_name` is what closes that, and these
    // four tests are what the daemon does with it.

    /// Registers `name` as a built-in dog and returns the row it produced.
    ///
    /// Straight through [`crate::supervisor::SupervisorHandle::start_dog`]
    /// rather than through `Request::EnableDog`, because the request path
    /// would need a handshaken connection of its own and the point here is
    /// what a REFUSED one does.
    async fn start_dog(ctx: &RpcContext, name: &str) -> ProcessInfo {
        let spec = crate::dogs::DogSpec {
            name: name.to_string(),
            source: DogSource::BuiltIn,
        };
        let app = crate::dogs::dog_app(&spec, &ctx.paths).expect("the dog fixture must assemble");
        ctx.supervisor
            .start_dog(app, DogSource::BuiltIn)
            .await
            .expect("the dog fixture must start")
    }

    /// One refused handshake, announcing `dog` (or nothing, for a client
    /// that is not a dog), returning once the daemon has closed on it.
    ///
    /// The close is the forcing mechanism for everything synchronous
    /// (IR-46): the daemon records the refusal and decides what it owes the
    /// dog BEFORE it returns the error that closes the socket, so a caller
    /// that has seen the close can read the verdict without racing it. The
    /// restart itself runs on its own task and needs [`await_dog`].
    async fn refuse_as(ctx: &RpcContext, dog: Option<&str>) {
        let mut client = connected(ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.14".to_string(),
                protocol: PROTOCOL_VERSION + 1,
                dog_name: dog.map(str::to_owned),
            })
            .await;
        let refusal: HelloReply = client.recv().await;
        refusal.expect_err("a skewed protocol must be refused");
        assert!(
            client.closed().await,
            "the daemon must close after refusing"
        );
    }

    /// The flock row named `name`, or a panic naming what was there.
    async fn dog_row(ctx: &RpcContext, name: &str) -> ProcessInfo {
        ctx.supervisor
            .list()
            .await
            .into_iter()
            .find(|info| info.name == name)
            .unwrap_or_else(|| panic!("no row named {name}"))
    }

    /// Waits until `name` is running as `pid`, or fails inside
    /// [`RECV_TIMEOUT`], returning how long it took.
    ///
    /// The restart a refusal triggers runs on its own task, so there is no
    /// handle to await; polling a real condition against a hard ceiling is
    /// the forcing mechanism (IR-46). The elapsed time is returned because
    /// the never-restart-twice test below sizes its negative window against
    /// it rather than against a guess.
    async fn await_dog(ctx: &RpcContext, name: &str, pid: u32) -> Duration {
        let began = tokio::time::Instant::now();
        let seen = tokio::time::timeout(RECV_TIMEOUT, async {
            loop {
                if dog_row(ctx, name).await.pid == Some(pid) {
                    return;
                }
                tokio::task::yield_now().await;
            }
        })
        .await;
        assert!(
            seen.is_ok(),
            "{name} never reached pid {pid} within {RECV_TIMEOUT:?}: {:?}",
            ctx.supervisor.list().await
        );
        began.elapsed()
    }

    /// fails if a refused dog is left mute. This is G8's step 2, and the
    /// daemon is the only party that can take it: the dog's own client has
    /// stopped rather than spinning (that is `ReconnectingClient`'s half),
    /// and a dog carried across a handover is a live process holding a dead
    /// socket that nothing else will replace.
    ///
    /// The pid moving is the assertion, not the restart count: a restart
    /// that re-registered the row without re-spawning would leave a dog
    /// exactly as mute as it was.
    #[tokio::test]
    async fn a_refused_dog_is_restarted_once_from_the_binary_on_disk() {
        let h = harness(vec![ProcScript::never_exits(), ProcScript::never_exits()]);
        let started = start_dog(&h.ctx, "metrics").await;
        assert_eq!(started.pid, Some(FIRST_SCRIPTED_PID));

        refuse_as(&h.ctx, Some("metrics")).await;

        await_dog(&h.ctx, "metrics", FIRST_SCRIPTED_PID + 1).await;
        assert!(
            h.ctx.dog_refusals.stale().is_empty(),
            "one refusal buys a restart; it does not condemn the dog"
        );
    }

    /// fails if the daemon restarts a dog it has already restarted — the
    /// spin G8 exists to forbid. A second refusal proves the binary on disk
    /// cannot satisfy this daemon either, because the restart already ran
    /// it, so a third attempt is not optimism.
    ///
    /// Three independent things have to hold, and each catches a different
    /// mutation. The dog must be REPORTED stale, which a daemon that
    /// silently stopped restarting would fail. Its pid must not move inside
    /// a window sized against the restart that really happened earlier in
    /// this same test — a negative assertion is only as good as its window,
    /// and a measured one scales with a loaded runner where a fixed number
    /// would not. And the harness is scripted with exactly the two spawns
    /// G8 permits, so a third would fail to spawn and take the dog out of
    /// `Online` even if it somehow beat the window.
    #[tokio::test]
    async fn a_twice_refused_dog_is_reported_stale_and_never_restarted_again() {
        let h = harness(vec![ProcScript::never_exits(), ProcScript::never_exits()]);
        start_dog(&h.ctx, "metrics").await;

        refuse_as(&h.ctx, Some("metrics")).await;
        let restart_took = await_dog(&h.ctx, "metrics", FIRST_SCRIPTED_PID + 1).await;

        refuse_as(&h.ctx, Some("metrics")).await;
        assert_eq!(
            h.ctx.dog_refusals.stale(),
            vec!["metrics".to_string()],
            "the second refusal must be reported, not swallowed"
        );

        // Ten times the restart that DID happen, floored so a machine fast
        // enough to make the measurement meaningless still watches for a
        // real interval.
        let window = (restart_took * 10).max(Duration::from_millis(200));
        tokio::time::sleep(window).await;

        let after = dog_row(&h.ctx, "metrics").await;
        assert_eq!(
            after.pid,
            Some(FIRST_SCRIPTED_PID + 1),
            "a second restart within {window:?} is the spin G8 forbids"
        );
        assert_eq!(
            after.status,
            ProcStatus::Online,
            "a third spawn would exhaust the script and error the dog"
        );
    }

    /// fails if a dog that got back in stays condemned. The restart G8 owes
    /// a refused dog is worth nothing if the successful handshake it
    /// produces does not clear the mark: the dog would be reported stale
    /// forever while answering perfectly, and a LATER daemon that refused
    /// it would skip straight past its one restart.
    #[tokio::test]
    async fn a_dog_that_handshakes_is_no_longer_stale() {
        let h = harness(vec![]);
        refuse_as(&h.ctx, Some("metrics")).await;
        refuse_as(&h.ctx, Some("metrics")).await;
        assert_eq!(h.ctx.dog_refusals.stale(), vec!["metrics".to_string()]);

        let mut client = connected(h.ctx.clone()).await;
        client
            .send(&Hello {
                client_version: "0.1.22".to_string(),
                protocol: PROTOCOL_VERSION,
                dog_name: Some("metrics".to_string()),
            })
            .await;
        let ack: HelloReply = client.recv().await;
        ack.expect("a matching protocol must be acked");

        assert!(
            h.ctx.dog_refusals.stale().is_empty(),
            "a dog talking to this daemon is not stale by any definition it can apply"
        );
    }

    /// fails if an operator running an older `shep` has a dog restarted
    /// under them. The CLI cannot name a dog — the name travels only on
    /// `ReconnectingClient`, which no verb uses — so a refusal carrying no
    /// name must leave the flock exactly as it was, however many of them
    /// arrive.
    #[tokio::test]
    async fn a_refused_client_that_is_not_a_dog_touches_nothing() {
        let h = harness(vec![ProcScript::never_exits()]);
        let started = start_dog(&h.ctx, "metrics").await;

        for _ in 0..3 {
            refuse_as(&h.ctx, None).await;
        }

        assert!(h.ctx.dog_refusals.stale().is_empty());
        let after = dog_row(&h.ctx, "metrics").await;
        assert_eq!(
            after.pid, started.pid,
            "a nameless refusal must not restart anything"
        );
        assert_eq!(after.status, ProcStatus::Online);
    }
}