Interrupt speech on reader navigation

Co-authored-by: Codex <codex@openai.com>

.gitignore

@@ -5,6 +5,7 @@ venv/
 outputs/*.wav
 outputs/*.mp3
 outputs/*.flac
+outputs/*.log
 .gradio/
 .env
 server.pid

FIELD_NOTES.md

@@ -53,6 +53,8 @@ Reader settings are manifest-backed: the app exposes known Kokoro voices and spe
 
 Auto-advance is intentionally opt-in. This keeps the default reader mode user-controlled while still supporting a continuous-reading demo path.
 
+Reader navigation interrupts active playback before requesting the next narration. That keeps fast keyboard movement predictable and avoids overlapping speech during demos.
+
 Every model-facing backend path reports `elapsed_ms`. The frontend surfaces the latest end-to-end reader latency and stores per-item timing in the transcript, giving the build report concrete evidence for the small-model responsiveness claim.
 
 The app exposes `/api/award-evidence` and renders that data in the session panel, so the live demo can point directly to the hackathon bonus targets it is designed to satisfy.

README.md

@@ -94,6 +94,8 @@ The reader bar exposes Kokoro voice selection and speaking speed controls. Defau
 
 Auto-advance is available as an opt-in reader control. It stays off by default so users keep manual control unless they choose continuous reading.
 
+Navigation commands interrupt current speech before starting the next request, matching the expectation that a screen reader responds immediately when the user moves.
+
 Reader-brain, image-description, speech, and image-generation responses include `elapsed_ms`. The session panel and transcript show recent latency so the Field Notes can discuss responsiveness with concrete numbers.
 
 The session panel also renders a manifest-backed demo checklist for Tiny Titan, Llama Champion, Off-Brand, and Field Notes evidence.

scripts/verify.py

@@ -30,6 +30,13 @@ def verify_static_assets() -> None:
     for path in required:
         assert_true(path.exists(), f"Missing required asset: {path}")
 
+    app_js = (ROOT / "static" / "app.js").read_text(encoding="utf-8")
+    assert_true("function haltPlayback" in app_js, "Reader controls should expose a shared playback halt helper")
+    assert_true(
+        "haltPlayback({ clearAutoAdvance: false });" in app_js,
+        "New narration commands should interrupt current speech immediately",
+    )
+
 
 def verify_core_fallbacks() -> None:
     narration = app.reader_brain_core(

static/app.js

@@ -220,16 +220,22 @@ function setActive(index) {
   currentStatus.textContent = `${node.type}, item ${currentIndex + 1} of ${nodes.length}`;
 }
 
-function stopAudio() {
+function haltPlayback({ clearAutoAdvance = true } = {}) {
   audio.pause();
   audio.removeAttribute("src");
   window.speechSynthesis?.cancel();
-  activeAutoAdvance = false;
+  if (clearAutoAdvance) {
+    activeAutoAdvance = false;
+  }
   activeSpeechSerial = 0;
   playing = false;
   controls.play.textContent = "Play";
 }
 
+function stopAudio() {
+  haltPlayback();
+}
+
 function maybeAutoAdvance(serial) {
   if (!enabled || !autoAdvanceControl.checked || !activeAutoAdvance || serial !== requestSerial) {
     return;
@@ -314,6 +320,7 @@ async function postJson(url, payload) {
 async function narrate(index) {
   if (!enabled || !nodes[index]) return;
   const serial = ++requestSerial;
+  haltPlayback({ clearAutoAdvance: false });
   setActive(index);
   const node = nodes[index];
   runtimeStatus.textContent = "Thinking";
@@ -364,6 +371,7 @@ async function narrate(index) {
 async function summarizeCurrentSection() {
   if (!enabled) return;
   const serial = ++requestSerial;
+  haltPlayback({ clearAutoAdvance: false });
   const section = currentSectionPayload();
   runtimeStatus.textContent = "Summarizing";
   controls.summary.disabled = true;