rutster/docs/superpowers/specs/2026-07-01-slice-4-barge-in-design.md

Rutster slice 4 — Barge-in: VAD-driven playout kill on a dedicated media thread

• Status: Draft (pending review)
• Date: 2026-07-01
• Spearhead step: 4 of 6 (vision-revision §10 / PORT_PLAN "Phasing")
• Origin: brainstorming session 2026-07-01
• Depends on: slice 1 — WebRTC media loopback, slice 2 — The agent tap, and slice-3's OpenAI Realtime brain (merged as c30a452 — MockRealtimeBrain + translator + the speech_started / speech_stopped advisory events). All three must be landed and green.
• Related: ADR-0002 (fused vertical — the hot-path hop invariant this slice re-affirms), ADR-0008 (FOB/green-zone doctrine — the reflex is a FOB member: hot-path, differentiating), ARCHITECTURE.md §"Biggest technical risk" (the reflex loop is the remaining long pole), ARCHITECTURE.md §"Media plane" ("Dedicated timing threads for the 20ms loop, never the shared tokio pool" — this slice finally lands that mandate).

---

TL;DR

Stand up spearhead step 4: the FOB reflex loop. Slice 3 pre-paved the advisory signals (speech_started / speech_stopped from the brain) and locked the turn-ownership decision (OpenAI Realtime server-side VAD disabled; the FOB owns turn-taking). Slice 4 acts on the caller's speech with a local in-core VAD as the primary trigger — an RMS/energy detector running in on_pcm_frame on the dedicated thread, in the 20 ms loop, with zero brain round-trip between caller speech and playout kill. This is the property ARCHITECTURE.md:79-81 demands ("Local real-time reflexes... live in-core because the brain round-trip is too slow to enforce them") and the proof wedge #1 rests on ("VAD killing TTS the instant the caller speaks, without the brain" — README:98-100). Slice-3's speech_started/speech_stopped advisory becomes the secondary/confirmation signal — the brain's ASR-quality VAD confirms the local kill slightly later, but the kill itself fires from the FOB's own inspection of caller audio.

Slice 4 also graduates the media loop off the tokio pool: a single dedicated std::thread owns all RtcSessions exclusively and drives the 20 ms tick via Instant::sleep_until. This honors ARCHITECTURE.md's "never the shared tokio pool" mandate, which slice-1 explicitly deferred to "step 4 (barge-in)" (loop_driver.rs:18-23). The graduation is load-bearing: the reflex is the differentiator and the long pole, and its timing discipline demands a thread that doesn't compete with the axum runtime for scheduling.

The seam slice 1→3 preserved (loop_driver.rs + rtc_session.rs byte-identical) holds for slice-4 as well: the reflex is a pair of composing AudioPipe decorators (Reflex<P> + LocalVadReflex<P>) in rutster-media, invisibly to loop_driver::drive. Only the binary-side wiring (session_map.rs → media_thread.rs) changes shape; the media crate's hot path stays untouched.

---

1. Scope

1.1 In scope

• Implementation of spearhead step 4: barge-in / VAD-driven playout kill, driven by a local in-core VAD (RMS/energy detector in on_pcm_frame — the primary trigger, zero brain round-trip) with the brain's speech_started/speech_stopped advisory as the secondary/confirmation signal. This proves wedge #1 ("VAD killing TTS the instant the caller speaks, without the brain" — README:98-100, ARCHITECTURE.md:79-81).
• A new Reflex<P: AudioPipe> wrapper (rutster-media/src/reflex.rs) that decorates the pipe the RtcSession holds. The reflex owns the mute state machine, the advisory channel receiver, and the barge-in flush trigger. It is the concrete embodiment of ARCHITECTURE.md's "local real-time reflexes" row for the barge-in case. Fed by BOTH the local VAD (via the outer LocalVadReflex wrapper that injects AdvisoryEvents into the same mpsc) AND the brain's advisories.
• A new LocalVadReflex<P: AudioPipe> wrapper (rutster-media/src/reflex.rs) — the PRIMARY trigger. Decorates AudioPipe's on_pcm_frame path: inspects the caller's decoded PCM samples (computes RMS, compares to a const threshold, debounces N consecutive above-threshold frames), fires AdvisoryEvent::SpeechStarted { at: Instant::now() } into the Reflex's advisory mpsc when local VAD trips. Wraps Reflex<TapAudioPipe> (composition: LocalVadReflex<Reflex<TapAudioPipe>> — the decorator pattern from §6.4 pays off here). Forward-compatible: the threshold is a const for the MVP (no tuning framework — that's the only piece deferred, per §1.2).
• A new barge_in_flush method on the AudioPipe trait (default impl delegates to clear_playout_ring) — the seam object's "kill now" path: clear the playout ring AND drain the brain-bound rx_audio_out channel of any frames queued before the barge so the first audio_out observed post-barge is provably post-barge. TapAudioPipe overrides; EchoAudioPipe uses the default.
• A new AdvisoryEvent enum (SpeechStarted { at }, SpeechStopped { at }) flowing over a tokio mpsc from the TapEngine (tokio) to the Reflex (media thread). The engine pushes the events it already decodes from the brain (slice-3 wired these as log+count; slice-4 forwards them into the reflex).
• A new dedicated media thread (rutster/src/media_thread.rs) replacing the tokio spawn_poll_task. One std::thread::spawn at binary startup owns HashMap<ChannelId, RtcSession> exclusively; all access from axum is via a command channel (AcceptOffer, Delete, Shutdown). The 20 ms tick is std::thread::sleep.
• Rewired session_map.rs (binary): SessionEntry.rtc: Arc<Mutex<RtcSession>> → cmd_tx: mpsc::Sender<MediaCmd>. create_session, post_offer, close, spawn_poll_task all route through the command channel. The async handlers are cold-path; no cross-thread coordination happens on the 20 ms tick.
• MockRealtimeBrain extension (rutster-brain-realtime/src/mock.rs): gains the ability to emit speech_started / speech_stopped on a programmable schedule (e.g. "after N audio_in frames received, send speech_started; after M more, send speech_stopped").
• Barge-in e2e integration test (extends slice-3's crates/rutster/tests/realtime_integration.rs harness): synthetic WebRTC peer → MediaThread → TapEngine → MockRealtimeBrain; mock emits speech_started; assert playout goes silent within ≤1 tick (20 ms); mock emits fresh audio_out; assert playout resumes.
• New ReflexMetrics (barge_in_count, advisory_dropped, frames_suppressed) mirroring TapMetrics shape (atomics, snapshot fn). Threaded through the same TapConn.metrics surface where reasonable, or a new side-car.
• Thorough learner-facing comments on the new std-thread / channel-bridge / wrapper-decorator patterns (slice-1 §7 standard carries over).

1.2 Out of scope (with scheduled return)

Deferred item	Returns in	Why deferred
Local VAD tuning framework (configurable thresholds, per-environment calibration, adaptive noise floor)	post-spearhead refinement	The VAD itself (RMS/energy detector + debounce) IS in scope for slice-4 (the primary barge-in trigger, proving wedge #1). Only the tuning framework — configurable thresholds, calibration UI, adaptive noise floors — is deferred. The MVP ships with a single const threshold + N-frame debounce, exercised by the e2e test with a synthetic loud signal. Tuning to real-world noise conditions is post-spearhead.
Per-session media threads / threadpool shard	later rung	Single thread covers spearhead scale (loopback dev + low-concurrency PSTN via slice-5). The command-channel seam between axum and the thread makes the graduation to a threadpool shard localized.
Trickle ICE	later	Unchanged from slice-1 deferral.
Min-mute floor / inter-word-gap debouncing	post-spearhead	SpeechStopped is a no-op for mute; a floor timer on resume would protect against brain-yield races (brain emits fresh audio_out before the caller's inter-word gap ends). Defer until observed in practice.
Brain-side input_audio_buffer.interrupt / clear on barge	slice-5 or brain-side	Whether the brain should clear its own input buffer on speech_started is a brain-UX decision, not a FOB one; the FOB only kills playout (its half-duplex gate). The advisory already tells the brain what happened; the brain's response is its own concern.
Half-duplex gating beyond playout kill	later rung	Barge-in is the first half-duplex reflex; full HD gating (mixing, jitter buffer interaction, multi-party) arrives with conferencing.
TLS on HTTP / WSS	slice-5	Unchanged.
Authn / authz / multi-tenancy	slice-6	Unchanged.
Spend cap / abuse gate	slice-6	Unchanged.
Browser-based automated e2e (Playwright/Selenium)	post-spearhead	Unchanged. The synthetic-peer harness from slice-2/3 is the test vehicle.

---

2. Architecture delta

2.1 The reflex wrapper

Reflex<P: AudioPipe> is a zero-cost-style decorator around any AudioPipe. It sits between RtcSession.pipe (which loop_driver::drive calls via session.pipe.next_pcm_frame()) and the concrete pipe (TapAudioPipe in production, EchoAudioPipe in slice-1's unit tests). loop_driver is oblivious to the wrapper: it still calls session.pipe.next_pcm_frame(), the dynamic dispatch through Box<dyn AudioPipe> lands in Reflex::next_pcm_frame, which applies the state machine and delegates to inner.next_pcm_frame() per the table in §3.2.

The reflex owns three pieces of state:

• advisory_rx: mpsc::Receiver<AdvisoryEvent> — drained sync-non-blocking via try_recv on the 20 ms tick before delegating to inner. Fed by TWO senders on the same channel: the outer LocalVadReflex (local VAD, the primary trigger) and the TapEngine task (the brain's advisory, secondary/confirmation) — both over tokio mpsc.
• muted: bool — the kill state. next_pcm_frame returns None while muted, unless the inner returns Some (the resume condition — the first fresh audio_out clears mute).
• barge_epoch: u64 — incremented on every SpeechStarted. Load-bearing this slice: the local VAD (primary) and the brain's advisory (secondary) can both fire on the same barge, and the epoch distinguishes a genuine re-barge from the slower confirmation landing on the same event (§3.2). The flush + drain keeps resume race-free regardless.

2.2 The dedicated media thread

A single std::thread::spawn replaces the tokio spawn_poll_task. The thread owns HashMap<ChannelId, RtcSession> exclusively — no Arc<Mutex<RtcSession>> shared with axum. All access from the axum handlers is via a command channel:

enum MediaCmd {
    AcceptOffer { id: ChannelId, sdp: String, reply: oneshot::Sender<Result<String, String>> },
    Delete     { id: ChannelId, reply: oneshot::Sender<()> },
    Shutdown   { reply: oneshot::Sender<()> },
}

The thread loop per 10 ms meta-tick:

1. Drain cmd_rx via try_recv loop — handle all pending commands before ticking.
2. For each session in the map: drain the per-session flush_rx side-channel (slice-2's existing disconnect-flush signal) BEFORE run_poll_once, then call RtcSession::run_poll_once(now) (the unchanged loop_driver::drive).
3. After run_poll_once, observe channel.state:
   • Connected && tap.is_none() → spawn the TapEngine (tokio task via the tokio::runtime::Handle captured at thread-start) + wire Reflex<TapAudioPipe> as the session's pipe. Mirror of slice-2's spawn seam, relocated from session_map.rs::drive_all_sessions to here.
   • Closed → remove the entry + drop the session.
4. std::thread::sleep(Duration::from_millis(10)) — 10 ms meta-tick. (Stable API: std::thread::sleep_until is nightly-only; sleep(dur) is the stable path. The 20 ms outbound encode tick is driven inside loop_driver::drive (unchanged); the 10 ms meta-tick gives finer resolution so str0m's Timeout outputs are honored promptly.)

The tokio ↔ std-thread bridge: all channels are tokio mpsc/oneshot (constructable on tokio, drainable via try_recv/blocking_recv from any thread). The tokio::runtime::Handle captured at MediaThread::spawn time is used on the std thread to handle.spawn(...) the TapEngine when the Connected transition fires. No async code runs on the std thread itself — only sync channel ops + RtcSession::run_poll_once.

Why a single thread, not per-session: spearhead scale. One loopback peer at a time in dev; even at low PSTN concurrency (slice-5) one thread drives dozens of sessions in 10 ms. Per-session threads arrive when the threadpool shard model lands (deferred). The command-channel seam between axum and the thread makes that graduation localized.

2.3 The hot-path audit (ADR-0002 honored)

ADR-0002's load-bearing rule: "the control↔media gRPC hop on the per-call hot path is removed." Slice 4 does not re-introduce a hop:

• The reflex's kill decision happens inside Reflex::next_pcm_frame on the dedicated thread — no channel send, no cross-thread coordination on the 20 ms tick. The advisory arrives via a try_recv drain (sync, non-blocking).
• axum → media-thread is cold-path only (SDP accept, DELETE). None of it runs on the 20 ms tick.
• The brain WS ↔ TapEngine (tokio) path is unchanged from slice-3. The advisory channel is a third mpsc alongside the existing tx_pcm_in/rx_audio_out/flush_tx — same pattern, additive.

The fused vertical stays fused. ADR-0002 honored.

---

3. Component design

3.1 AdvisoryEvent enum

// crates/rutster-media/src/reflex.rs

/// A turn-event advisory from the brain. The brain decodes its own
/// speech-to-text / VAD results and forwards these; the FOB *owns*
/// turn-taking and acts on them (slice-3 §4.3 — OpenAI Realtime
/// server-side VAD is DISABLED; the FOB's reflex is authoritative).
///
/// Carried over a tokio mpsc from the TapEngine (tokio task) to the
/// `Reflex` wrapper (media thread). Drained sync via `try_recv` on the
/// 20 ms tick — the kill decision lives in the loop, not in a handler.
#[derive(Debug)]
pub enum AdvisoryEvent {
    /// The brain detected caller speech. Trigger barge-in: kill playout.
    SpeechStarted { at: Instant },
    /// The brain detected caller speech ended. Observed + counted; does
    /// NOT toggle mute (the resume condition is "first fresh audio_out
    /// after the barge", not "speech_stopped" — see §3.2 state table).
    SpeechStopped { at: Instant },
}

3.2 Reflex<P> state machine

Current state	Event	Action	New state
Playing	SpeechStarted	muted=true; epoch++; inner.barge_in_flush() (clear ring + drain rx_audio_out so stale brain frames queued pre-barge are dropped); metrics.barge_in_count++	Muted
Muted	SpeechStarted (duplicate/re-barge)	epoch++; barge_in_flush() again (fresh barge resets the "fresh audio" clock); barge_in_count++	Muted
Muted	SpeechStopped	increment advisory_observed_speech_stopped counter; no state change	Muted
Playing	SpeechStopped	increment counter; no state change	Playing
Muted	inner next_pcm_frame() returns Some(f) (fresh brain audio arrived post-barge)	muted=false; return Some(f)	Playing
Muted	inner next_pcm_frame() returns None	return None (silence); metrics.frames_suppressed++	Muted

Why SpeechStopped is a no-op for mute: per the resume-semantics decision (resume on first fresh audio_out). The brain's speech_stopped is observed (counter) but doesn't gate — this avoids the inter-word-gap problem (caller pauses, VAD fires stopped, brain un-mutes too early, brain's audio overlaps caller's next word). The resume condition is "the brain has yielded and started a new response," which is provably signaled by the first audio_out frame after the barge — not by the caller's silence.

Why epoch: with two concurrent trigger sources this slice — the local VAD (primary) and the brain's advisory (secondary/confirmation, §6.1) — the epoch disambiguates "is this barge a re-barge of the same event, or a new one" when both race into the same advisory channel. The local VAD trips first; the brain's slower ASR-grade advisory lands ~300 ms later on the same event and must not be counted as a fresh barge. The epoch is that disambiguator — load-bearing now, not a forward-compat seam.

3.3 AudioPipe trait extension

// crates/rutster-media/src/pcm.rs — additive method on `AudioPipe`

/// Barge-in flush: clear the playout ring AND drain the inbound brain
/// audio queue of any frames queued before the barge. Called by `Reflex`
/// on `SpeechStarted`. The drain of `rx_audio_out` is what makes the
/// resume condition race-free: the first `audio_out` observed post-barge
/// is provably post-barge (frames queued pre-barge are dropped here).
///
/// Default impl delegates to `clear_playout_ring` — sufficient for
/// pipes without an inbound queue to drain (like `EchoAudioPipe`).
fn barge_in_flush(&mut self) {
    self.clear_playout_ring();
}

TapAudioPipe overrides:

// crates/rutster-tap/src/tap_audio_pipe.rs

fn barge_in_flush(&mut self) {
    // Clear the playout ring (drops buffered brain-proposed frames).
    self.playout_ring.clear();
    // Drain rx_audio_out of any frames the engine task queued before
    // the barge. Without this, a stale frame in the mpsc would un-mute
    // immediately on the next tick — defeating the "first fresh audio_out"
    // resume condition. Hot-path: try_recv loop, bounded, no blocking.
    while self.rx_audio_out.try_recv().is_ok() {
        self.metrics.barge_drained_inflight.fetch_add(1, Ordering::Relaxed);
    }
}

3.4 LocalVadReflex<P> — the primary trigger (FOB-local, zero brain round-trip)

The decorator that proves wedge #1. Wraps any AudioPipe; inspects the caller's decoded PCM in on_pcm_frame, computes RMS energy, fires AdvisoryEvent::SpeechStarted into the inner Reflex's advisory channel when local VAD trips. Composes as LocalVadReflex<Reflex<TapAudioPipe>> — the outer wrapper does local-VAD; the inner wrapper applies the mute state machine to the advisory stream (which now has TWO sources: the local VAD + the brain's advisory, both feeding the same mpsc).

// crates/rutster-media/src/reflex.rs

/// RMS energy threshold for caller-speech detection. The MVP ships with
/// a single tuned-for-synthetic-loud-signal const; the tuning framework
/// (per-environment calibration, adaptive noise floor) is deferred per
/// slice-4 §1.2. The synthetic-peer e2e test sends a frame with samples
/// well above this threshold so the trip is deterministic.
///
/// i16 max is 32767; a |sample| average of ~500 (~1.5% of full scale) is
/// a quiet-but-unmistakable signal; room-tone background sits well below.
/// The debounce (N consecutive frames above threshold) filters transient
/// spikes (clicks, nose-breaths) without slowing the trip materially.
pub const VAD_RMS_THRESHOLD: f64 = 500.0;

/// Number of consecutive above-threshold frames required before the VAD
/// trips. At 20 ms/frame, N=3 = 60 ms of above-threshold audio — well
/// below the brain's ~300 ms ASR-VAD latency, comfortably instant for
/// the wedge-#1 demonstration. Tunable in a later slice; const for MVP.
pub const VAD_DEBOUNCE_FRAMES: u32 = 3;

pub struct LocalVadReflex<P: AudioPipe> {
    inner: P,
    /// The advisory_tx the LocalVadReflex pushes into when local VAD
    /// trips. This is the SAME channel the brain's advisories arrive
    /// on (the Reflex holds the paired rx) — so the Reflex's mute state
    /// machine sees both sources uniformly via `drain_advisories`.
    advisory_tx: mpsc::Sender<AdvisoryEvent>,
    /// Consecutive above-threshold frame count (the debounce counter).
    above_threshold_streak: u32,
    /// True once the VAD has tripped for the current caller-speech
    /// burst. Reset to false when the streak breaks (caller stopped)
    /// — so a new burst trips a fresh `SpeechStarted`.
    vad_armed: bool,
}

impl<P: AudioPipe> LocalVadReflex<P> {
    pub fn new(inner: P, advisory_tx: mpsc::Sender<AdvisoryEvent>) -> Self {
        Self {
            inner,
            advisory_tx,
            above_threshold_streak: 0,
            vad_armed: true, // armed on construction
        }
    }

    /// Compute RMS energy of a PCM frame. The frame is 480 i16 samples
    /// @ 24 kHz; RMS = sqrt(mean(sample²)). Cheap: ~480 multiplications,
    /// one division, one sqrt — well under the 20 ms tick budget.
    fn rms(frame: &PcmFrame) -> f64 {
        let sum_sq: u64 = frame.samples.iter()
            .map(|&s| (s as i64 * s as i64) as u64)
            .sum();
        (sum_sq as f64 / frame.samples.len() as f64).sqrt()
    }

    /// Inspect a caller PCM frame + apply the debounce state machine.
    /// Called from `on_pcm_frame` (the sink path). Returns true if the
    /// VAD tripped THIS call (so the caller can push the advisory).
    fn observe(&mut self, frame: &PcmFrame) -> bool {
        let energy = Self::rms(frame);
        if energy >= VAD_RMS_THRESHOLD {
            self.above_threshold_streak += 1;
            if self.above_threshold_streak >= VAD_DEBOUNCE_FRAMES && self.vad_armed {
                self.vad_armed = false; // disarm until caller stops
                return true; // trip!
            }
        } else {
            // Caller went quiet → re-arm for the next speech burst.
            self.above_threshold_streak = 0;
            self.vad_armed = true;
        }
        false
    }
}

impl<P: AudioPipe> AudioSource for LocalVadReflex<P> {
    fn next_pcm_frame(&mut self) -> Option<PcmFrame> {
        // Pure delegation — the VAD only observes the SINK path
        // (inbound caller audio); playout is the inner Reflex's concern.
        self.inner.next_pcm_frame()
    }
}

impl<P: AudioPipe> AudioSink for LocalVadReflex<P> {
    fn on_pcm_frame(&mut self, frame: PcmFrame) {
        // THE PRIMARY TRIGGER: inspect the caller's audio BEFORE delegating.
        // If the local VAD trips, push an advisory into the same channel the
        // brain's advisories arrive on — the inner Reflex drains it on the
        // next next_pcm_frame call + applies the kill. Zero brain round-trip.
        if self.observe(&frame) {
            let _ = self.advisory_tx.try_send(AdvisoryEvent::SpeechStarted {
                at: Instant::now(),
            });
            // try_send failure (channel full) → drop + observe (hot-path
            // policy). The brain's advisory path is the backstop; a missed
            // local-VAD trip here is not catastrophic — the brain will fire
            // its ASR-VAD ~300 ms later.
        }
        // Delegate to inner — the caller's audio still reaches the brain.
        self.inner.on_pcm_frame(frame)
    }
}

impl<P: AudioPipe> AudioPipe for LocalVadReflex<P> {
    fn clear_playout_ring(&mut self) { self.inner.clear_playout_ring() }
    fn barge_in_flush(&mut self) { self.inner.barge_in_flush() }
}

Why the VAD trips on_pcm_frame (sink), not next_pcm_frame (source): the caller's audio is what the VAD inspects — and caller audio arrives via the sink path (decoded from the peer's RTP by loop_driver::drive → session.pipe.on_pcm_frame(pcm)). The source path is the brain's audio_out — the playout being killed, not the signal being detected.

Why the SAME advisory channel as the brain: the inner Reflex drains all advisories uniformly — it doesn't care whether the source is local VAD or the brain's ASR. The brain's advisory arriving ~300 ms later either finds the kill already applied (no-op, the muted flag is already true) or confirms it. This is the composition pattern §6.4 anticipated; the Reflex<P> wrapper shape was designed for exactly this layering.

3.5 Reflex<P> struct + impl

// crates/rutster-media/src/reflex.rs

pub struct Reflex<P: AudioPipe> {
    inner: P,
    advisory_rx: mpsc::Receiver<AdvisoryEvent>,
    muted: bool,
    barge_epoch: u64,
    metrics: Arc<ReflexMetrics>,
}

impl<P: AudioPipe> Reflex<P> {
    pub fn new(inner: P, advisory_rx: mpsc::Receiver<AdvisoryEvent>, metrics: Arc<ReflexMetrics>) -> Self {
        Self { inner, advisory_rx, muted: false, barge_epoch: 0, metrics }
    }

    /// Drain all pending advisories + apply the state table. Called at
    /// the top of `next_pcm_frame`. Hot-path: try_recv loop, bounded.
    fn drain_advisories(&mut self) {
        while let Ok(ev) = self.advisory_rx.try_recv() {
            match ev {
                AdvisoryEvent::SpeechStarted { at } => {
                    self.muted = true;
                    self.barge_epoch = self.barge_epoch.wrapping_add(1);
                    self.inner.barge_in_flush();
                    self.metrics.barge_in_count.fetch_add(1, Ordering::Relaxed);
                    tracing::info!(epoch = self.barge_epoch, ?at, "barge-in");
                }
                AdvisoryEvent::SpeechStopped { at: _ } => {
                    self.metrics.advisory_observed_speech_stopped.fetch_add(1, Ordering::Relaxed);
                    // No state change — see §3.2.
                }
            }
        }
    }
}

impl<P: AudioPipe> AudioPipe for Reflex<P> {
    fn next_pcm_frame(&mut self) -> Option<PcmFrame> {
        self.drain_advisories();
        if self.muted {
            // Muted: pull from inner. Some(f) = fresh brain audio arrived
            // post-barge → un-mute + return. None = silence, stay muted.
            match self.inner.next_pcm_frame() {
                Some(f) => {
                    self.muted = false;
                    Some(f)
                }
                None => {
                    self.metrics.frames_suppressed.fetch_add(1, Ordering::Relaxed);
                    None
                }
            }
        } else {
            self.inner.next_pcm_frame()
        }
    }

    fn on_pcm_frame(&mut self, frame: PcmFrame) {
        // Inbound caller audio is NEVER gated by the reflex. The brain
        // still hears the caller during barge — that's the point (the
        // brain needs to know the caller interrupted; the FOB only kills
        // its OWN playout, not the caller's path to the brain).
        self.inner.on_pcm_frame(frame)
    }

    fn clear_playout_ring(&mut self) {
        // The reconnect-flush path (slice-2 §5.3) still works through the
        // wrapper. If it fires during mute, the ring stays empty and mute
        // clears on the next post-reconnect audio_out.
        self.inner.clear_playout_ring()
    }

    fn barge_in_flush(&mut self) {
        // Allow the outer `LocalVadReflex` (primary trigger) to barge the inner.
        self.inner.barge_in_flush()
    }
}

3.6 ReflexMetrics

Mirror of TapMetrics shape (atomics + snapshot struct):

// crates/rutster-media/src/reflex.rs

#[derive(Default)]
pub struct ReflexMetrics {
    pub barge_in_count: AtomicU64,
    pub advisory_dropped: AtomicU64,       // advisory channel full (e.g. 16-cap)
    pub frames_suppressed: AtomicU64,      // None returns while muted
    pub advisory_observed_speech_stopped: AtomicU64,
}

pub struct ReflexMetricsSnapshot {
    pub barge_in_count: u64,
    pub advisory_dropped: u64,
    pub frames_suppressed: u64,
    pub advisory_observed_speech_stopped: u64,
}

// `barge_drained_inflight` lives on `TapMetrics` (in `rutster-tap`), not
// `ReflexMetrics`, because the drain happens inside `TapAudioPipe::barge_in_flush`,
// not inside `Reflex`. The path: `Reflex::drain_advisories` calls
// `inner.barge_in_flush()` which is `TapAudioPipe::barge_in_flush`, which is
// where the `rx_audio_out` drain + the counter increment happen.

---

4. The dedicated media thread

4.1 MediaThread

// crates/rutster/src/media_thread.rs

pub struct MediaThread {
    cmd_tx: mpsc::Sender<MediaCmd>,
    join: Option<std::thread::JoinHandle<()>>,
}

enum MediaCmd {
    AcceptOffer { id: ChannelId, sdp: String, reply: oneshot::Sender<Result<String, String>> },
    Delete     { id: ChannelId, reply: oneshot::Sender<()> },
    Shutdown   { reply: oneshot::Sender<()> },
}

Spawned at binary startup (main.rs), before axum::serve. The thread captures a tokio::runtime::Handle (to spawn TapEngine tasks when Connected transitions fire) and owns HashMap<ChannelId, RtcSession> + (per-session, lazily) the TapConn / advisory_rx / Reflex wrapper.

4.2 Thread loop (per 10 ms meta-tick)

1. cmd_rx.try_recv() loop — handle ALL pending commands before ticking. AcceptOffer calls RtcSession::accept_offer(sdp) and replies via the oneshot. Delete fires close_tx + bounded-await the engine task (750 ms cap via tokio::runtime::Handle::block_on(timeout(...))) — the std thread briefly enters the tokio runtime to await; cold-path, not the 20 ms tick. Shutdown drains + replies.
2. For each RtcSession in the map:
   • Drain per-session flush_rx side-channel (slice-2's existing disconnect-flush) BEFORE run_poll_once.
   • Call RtcSession::run_poll_once(now) — the unchanged loop_driver::drive.
   • Observe channel.state:
     • Connected && tap.is_none() → handle.spawn(spawn_tap_engine(...)) to bring up the tokio task; construct Reflex::new(TapAudioPipe::new(...), advisory_rx, metrics); call RtcSession::set_pipe(reflex). Mirror of slice-2's spawn seam.
     • Closed → remove the entry (drops the RtcSession + its pipe + advisory ends).
3. std::thread::sleep(Duration::from_millis(10)) — 10 ms meta-tick.

4.3 session_map.rs rewire

SessionEntry loses rtc: Arc<Mutex<RtcSession>>, gains cmd_tx: mpsc::Sender<MediaCmd> (cloned per-entry; cheap). tap_url stays (the thread reads it when spawning the engine). tap_conn: Option<TapConn> moves onto the media thread (the thread owns it after spawn).

• AppState::create_session → sends a Register { tap_url, reply } command to the media thread; the thread constructs RtcSession::new() (saves a cross-thread move of the struct + keeps all RtcSession construction on the thread that owns it). The thread replies with (id, cmd_tx_for_this_session); axum stores SessionEntry { cmd_tx, tap_url, tap_conn: None }.
• AppState::get(id) (SDP path) → cmd_tx.send(AcceptOffer { ... }).await + reply.await. Cold-path; the axum handler is async.
• AppState::close(id) → cmd_tx.send(Delete { id, reply }).await + reply.await. The reply returns after the TapEngine teardown completes on the thread.
• spawn_poll_task → spawn_media_thread: constructs the channels, spawns the std thread, stores cmd_tx + join in AppState. Same idempotent-guard pattern.

4.4 TapEngine extension

spawn_tap_engine returns a third channel end: advisory_tx: mpsc::Sender<AdvisoryEvent>. The pump loop, on receiving speech_started / speech_stopped from the brain (slice-3 already decodes these in the tap protocol layer — protocol_events.rs), pushes the corresponding AdvisoryEvent into advisory_tx. If the channel is full, drop + count (hot-path "drop + observe" policy; an advisory is a hint, not a command). The Reflex wrapper holds advisory_rx.

4.5 MockRealtimeBrain extension

rutster-brain-realtime/src/mock.rs gains a programmable advisory schedule: the test can register "after N audio_in frames received, send speech_started" and "after M more, send speech_stopped". The mock already asserts turn_detection: null on session.update (slice-3's S4 lock); slice-4 keeps that assertion.

---

5. Data flow

5.1 Barge-in (the kill) — primary path: local VAD, zero brain round-trip

1. caller speaks into mic → peer RTP → str0m decode → on_pcm_frame
   → LocalVadReflex::on_pcm_frame INSPECTS the frame FIRST:
     • rms(frame) computed (~480 muls, < 1 µs)
     • if rms ≥ VAD_RMS_THRESHOLD: streak++; if streak ≥ VAD_DEBOUNCE_FRAMES (3 = 60 ms) → TRIP
     • on trip: try_send(AdvisoryEvent::SpeechStarted) into advisory_tx (same channel as
       brain's advisories) → disarm until caller goes quiet
   → THEN delegates to inner.on_pcm_frame(frame) → tx_pcm_in → TapClient → audio_in (WS) → brain
2. [THREE HOPS HAPPEN HERE, ON THE BRAIN SIDE — NOT on the kill path]:
   - brain's ASR-VAD fires (~300 ms later, slower but more accurate)
   - brain sends speech_started back over WS
   - TapEngine → advisory_tx (the SAME mpsc; both sources feed one channel)
3. media thread next 20 ms tick → Reflex::next_pcm_frame → drain_advisories:
   • The LOCAL VAD's SpeechStarted arrives FIRST (it was pushed in step 1, same thread,
     no WS hop) → muted=true; epoch++; inner.barge_in_flush() (ring cleared + rx_audio_out drained)
   • The brain's SpeechStarted arrives ~300 ms LATER → Reflex sees muted=true already
     → re-barge (epoch++, barge_in_flush again — harmless, mute stays)
   → returns None (silence) for this + subsequent ticks while muted
4. loop_driver::drive pulls None from pipe → encodes Opus silence → peer hears silence
   (the brain's in-flight audio_out frames are dropped; no overlap with caller's speech)

Latency budget: caller speaks → kill fires in ≤ 60 ms (3 debounce frames × 20 ms tick) + one tick to drain the advisory + apply the kill = ≤ 80 ms wallclock. Zero brain round-trip on the primary path. (The brain's ASR-VAD advisory arrives ~300 ms later — it confirms the kill but doesn't gate it.) This is what ARCHITECTURE.md:80 demands ("the brain round-trip is too slow to enforce them") and the proof wedge #1 rests on.

5.2 Resume (the un-mute)

1. brain decides to yield/respond → sends a fresh audio_out frame
   (provably post-barge: barge_in_flush drained rx_audio_out)
2. TapClient → audio_out (WS) → TapEngine → tx_audio_out → rx_audio_out → playout ring
3. media thread 20 ms tick → Reflex::next_pcm_frame → drain_advisories (empty)
   → muted=true → inner.next_pcm_frame() returns Some(f) (fresh brain audio)
   → muted=false; return Some(f)
4. loop_driver encodes + writes → peer hears the brain's new response

5.3 Cold-path (axum ↔ media thread)

- POST /v1/sessions → AppState::create_session → MediaCmd::Register → thread constructs RtcSession → reply(id)
- POST /v1/sessions/{id}/offer → AppState::get + cmd_tx.send(AcceptOffer) → thread.lock(session).accept_offer(sdp) → reply(answer)
- DELETE /v1/sessions/{id} → AppState::close → cmd_tx.send(Delete) → thread: fire close_tx, bounded-await engine task teardown → reply
- graceful shutdown → cmd_tx.send(Shutdown) → thread drains + drops → reply → join

---

6. Why these decisions

6.1 Why both, local VAD primary (revised after adversarial review)

The initial brainstorming landed on advisory-only for MVP (cheapest path to a working barge-in; the brain's VAD already runs for STT). The 2026-07-01 adversarial review surfaced the load-bearing problem with that choice: README:98-100 + ARCHITECTURE.md:79-81 rest the wedge on "local reflexes that don't need the brain — VAD killing TTS the instant the caller speaks." Advisory-only puts the brain round-trip in the trigger path (brain VAD → WS → TapEngine → mpsc → Reflex); the kill DECISION is in-core but the TRIGGER SOURCE crossed the brain. The spearhead's step 4 would prove a reflex that depends on the brain — the opposite of the property steps 1-4 exist to prove.

The revision: local VAD in on_pcm_frame is the primary trigger (RMS/energy detector on the dedicated thread, in the 20 ms loop, zero brain round-trip — the actual wedge-#1 proof). Slice-3's advisory becomes the secondary/confirmation signal — the brain's ASR-quality VAD confirms the local kill ~300 ms later (slower but more accurate — it knows words, not just energy). The two sources feed the SAME advisory mpsc; the Reflex wrapper drains them uniformly. The composition is LocalVadReflex<Reflex<TapAudioPipe>> — the decorator pattern from §6.4, which was originally specced as deferred to "when local VAD arrives." It arrived.

• The VAD itself (~25 lines: RMS, threshold, debounce) is in scope. The tuning framework (configurable thresholds, per-environment calibration, adaptive noise floor) is deferred — the MVP ships a single const threshold + N-frame debounce, exercised by the e2e test with a synthetic loud signal.
• Slice-3's advisory plumbing + MockRealtimeBrain schedule aren't wasted — they become the secondary path, still exercised by the e2e test (the brain's advisory fires SLIGHTLY after the local VAD; both feeds land in the Reflex's drain).
• The Reflex<P> wrapper shape already supported composition — no structural rework, just a new LocalVadReflex<P> decorator landing in scope (Task 2b in the plan).

6.2 Why resume on first fresh audio_out (not speech_stopped)

• The "the brain has yielded and started a new response" condition is provably signaled by the first audio_out frame after the barge — not by the caller's silence. speech_stopped fires between words; resuming on it un-mutes too early (inter-word-gap overlap).
• The barge_in_flush drain of rx_audio_out makes the resume race-free: the first audio_out observed post-barge is provably post-barge (frames queued pre-barge are dropped in the flush).

6.3 Why a single dedicated thread (not per-session)

• Spearhead scale: one loopback peer in dev; even at low PSTN concurrency (slice-5), one thread drives dozens of sessions in 10 ms.
• The command-channel seam between axum and the thread makes the graduation to a threadpool shard localized — when per-CPU-shard threading arrives, it's a fan-out of the cmd_rx/HashMap shape, not a redesign.
• Per-session threads arrive when load demands; the spearhead's "shortest blocking path" rule dislikes spawning work per session that may not need it (pre-ICE-connected sessions would redundantly spin).

6.4 Why Reflex<P> as a wrapper (not inline in TapAudioPipe)

• Composition: LocalVadReflex<P> composes outside the advisory Reflex<P> (§6.1), the same way Reflex<TapAudioPipe> composes. The pattern (decorator over AudioPipe) stacks the two trigger sources without restructuring either.
• The seam: loop_driver.rs byte-identical (still calls pipe.next_pcm_frame()). If the reflex lived inline in TapAudioPipe, the binary-side wiring would still change but the TapAudioPipe module itself would grow the reflex state — less isolated.
• The payoff is realized this slice, not deferred: two stacked reflexes (local-VAD primary + advisory secondary) live as independent, separately-testable decorators rather than one module's commingled state. Keeping them as wrappers is what makes that separation free.

6.5 Why barge_in_flush on AudioPipe (not just clear_playout_ring)

• clear_playout_ring (slice-2) clears the ring. barge_in_flush clears the ring AND drains the inbound brain queue (rx_audio_out). The distinction matters: on a brain disconnect (slice-2's case), the brain is gone — rx_audio_out will drain itself on the next Disconnected try_recv. On a barge-in, the brain is alive and may have queued frames pre-barge that would un-mute immediately if not drained here. Two different "clear the playout path" semantics, two methods.

---

7. Done-criteria

1. cargo test --all passes (stable + 1.85, the CI matrix).
2. cargo fmt --check + cargo clippy -- -D warnings clean.
3. loop_driver.rs + rtc_session.rs byte-identical to slice-3 — CI-asserted via git diff --exit-code main -- crates/rutster-media/src/loop_driver.rs crates/rutster-media/src/rtc_session.rs (the §8.5 #6 seam gate, restated for slice-4).
4. Dedicated media thread drives sessions off the tokio pool; MediaThread integration test passes (AcceptOffer / Delete / Shutdown).
5. Reflex state-machine unit tests all pass:
   • SpeechStarted → next next_pcm_frame returns None even if ring has frames.
   • SpeechStarted then inner.next_pcm_frame()=Some → un-mutes, returns the frame.
   • SpeechStopped during Muted → stays Muted.
   • SpeechStopped during Playing → no-op.
   • Duplicate SpeechStarted re-flushes + stays Muted.
   • Metrics counters (barge_in_count, frames_suppressed) increment correctly.
   • advisory_rx full → advisory_dropped increments, no panic.
6. LocalVadReflex unit tests pass:
   • RMS computation verified against a known-loud + known-quiet frame.
   • Debounce: N-1 above-threshold frames do NOT trip; the Nth does.
   • Re-arm: above-threshold → trip; below-threshold → re-arm; next streak trips again.
   • on_pcm_frame ALWAYS delegates to inner (caller audio reaches the brain even during barge).
7. barge_in_flush unit tests pass (ring + rx_audio_out drain).
8. Barge-in e2e (PRIMARY PATH, proves wedge #1): synthetic loud caller audio → playout killed within ≤4 ticks (≤80 ms wallclock: 3 debounce + 1 drain+apply) WITHOUT any brain advisory. The brain's speech_started advisory arrives later + is a no-op (mute already applied).
9. Barge-in e2e (SECONDARY PATH, exercises slice-3 advisory plumbing): MockRealtimeBrain emits speech_started on schedule; local VAD is NOT tripped (quiet synthetic caller audio); advisory → kill → fresh audio_out → resume. Proves the advisory→reflex→kill path still works.
10. S4 turn-ownership lock preserved: MockRealtimeBrain still asserts turn_detection: null on session.update (slice-3's #7, unchanged).
11. MockRealtimeBrain extended to emit speech_started/speech_stopped on schedule.
12. cargo doc --no-deps renders the new reflex.rs + media_thread.rs module/item docs cleanly (learner-facing comments present per AGENTS.md code style).

---

8. Open decisions

• Trigger source — decided: both, local VAD primary + advisory secondary (revised after 2026-07-01 adversarial review; initial brainstorming landed on advisory-only, the review surfaced that advisory-only contradicts wedge #1's "VAD killing TTS the instant the caller speaks, without the brain"). Local VAD in on_pcm_frame is the primary trigger; the brain's speech_started/speech_stopped advisory is the secondary/confirmation.
• Resume semantics — decided: first fresh audio_out post-barge; SpeechStopped observational only.
• Thread model — decided: single dedicated std::thread; per-session/threadpool deferred.
• MockRealtimeBrain advisory schedule API shape — landed in §4.5 as a programmable "after N audio_in frames" schedule. Could alternatively be a free-form Vec<(trigger_frame_count, AdvisoryEvent)> queue. The plan will pin the concrete API.
• Thread shutdown ordering vs TapEngine teardown — Delete command handler fires close_tx + bounded-await the engine task (750 ms cap via tokio::runtime::Handle::block_on(timeout(...))); the reply oneshot returns after teardown. Cold-path, std thread briefly enters the tokio runtime to await. Documented as an acceptable deviation (not the 20 ms tick).

---

9. Cross-references

• slice-1 spec — the media loop + the seam (AudioSource/AudioSink traits in rutster-media); slice-1 §8.5 #6 is the seam gate this slice re-affirms.
• slice-2 spec — the tap interface, the TapAudioPipe, the core-authoritative playout buffer (§4.1), the flush_tx side-channel pattern that the advisory_rx mirrors.
• slice-3 (merged c30a452) — MockRealtimeBrain, the translator, the speech_started/speech_stopped protocol events, the S4 turn-ownership lock.
• ADR-0002 — fused vertical; the hot-path hop invariant this slice re-affirms (§2.3 audit).
• ADR-0008 — FOB/green-zone doctrine; the reflex is a FOB member (hot-path, security-constitutive for turn-taking, differentiating).
• ARCHITECTURE.md — §"Media plane" ("Dedicated timing threads for the 20ms loop, never the shared tokio pool" — this slice lands it); §"Biggest technical risk" (the reflex loop is the remaining long pole).
• PORT_PLAN.md — §Phasing, step 4 = barge-in.