@livx.cc/agentx 0.99.102 → 0.99.104

package/dist/cli.js

@@ -9331,7 +9331,14 @@ var OverlapJudge = class {
 };
 function overlapModelFor(ai, env = process.env) {
   const model = env.AGENTX_CHANNEL_DECIDER_MODEL || void 0;
-  return model || env.GROQ_API_KEY ? fastOverlapModel(ai, model) : void 0;
+  if (!model && !env.GROQ_API_KEY) return void 0;
+  const m = fastOverlapModel(ai, model);
+  const t0 = performance.now();
+  void m("yeah", AbortSignal.timeout(1e4)).then(
+    () => log19.debug(`fast model warm in ${Math.round(performance.now() - t0)}ms`),
+    (e) => log19.warn(`fast model warm-up failed (the first overlap verdict may be late): ${e?.message ?? e}`)
+  );
+  return m;
 }
 var softBarge = (mode) => mode !== "turn";
 
@@ -13606,14 +13613,8 @@ function bindAckInstinct(v, instinct) {
 }
 function bindSecondChannel(v, sc, live) {
   const o = v.options;
-  const { onPartial, onUtterance, onState, onBargeIn, onBackchannel } = o;
-  const { show } = sc.options;
+  const { onPartial, onUtterance, onState, onBargeIn } = o;
   let own = false;
-  let blip = "";
-  o.onBackchannel = (p) => {
-    blip = p;
-    onBackchannel(p);
-  };
   const take = v.takeInterruptedReply.bind(v);
   v.takeInterruptedReply = () => {
     const r = take();
@@ -13621,7 +13622,7 @@ function bindSecondChannel(v, sc, live) {
     return r.queued.length ? { ...r, full: "", heard: "" } : null;
   };
   o.onPartial = (t) => {
-    if (live() && !v.audible) sc.onInterim(t);
+    if (live() && (own || !v.audible)) sc.onInterim(t);
     onPartial(t);
   };
   o.onUtterance = ((t, meta) => {
@@ -13644,25 +13645,34 @@ function bindSecondChannel(v, sc, live) {
     if (live()) sc.noteReply(t);
     return speakDelta(t, ...rest);
   });
+  Object.defineProperty(sc.options, "aec", { get: () => v.usingAec, set: () => {
+  }, enumerable: true, configurable: true });
   Object.assign(sc.options, {
-    aec: v.usingAec,
     speak: (text, kind) => {
       if (!live() || v.audible) return false;
-      if (kind === "backchannel") {
-        if (sc.language !== "en") return false;
-        blip = "";
-        v.fireBackchannel();
-        return;
-      }
+      if (kind === "backchannel") return speakBlip(v, text);
       own = true;
       v.speakFiller(text);
       return v.awaitIdle().finally(() => {
         own = false;
       });
-    },
-    show: (text, kind) => show(kind === "backchannel" && blip ? blip : text, kind)
+    }
   });
 }
+function speakBlip(v, text) {
+  const E = VoiceEngine;
+  const pool = E.BACKCHANNELS;
+  E.BACKCHANNELS = [text];
+  try {
+    v.fireBackchannel();
+    return true;
+  } catch (e) {
+    log28.warn(`backchannel blip failed \u2014 line dropped: ${e.message}`);
+    return false;
+  } finally {
+    E.BACKCHANNELS = pool;
+  }
+}
 function channelInput(sc, m) {
   if (m.type === "interim") {
     if (typeof m.aec === "boolean") sc.options.aec = m.aec;
@@ -13917,8 +13927,10 @@ function requestOpener(text) {
   return ASK.includes(w[0]) || WANT.test(w.slice(0, 4).join(" ")) || new RegExp("^\u05EA\\p{Script=Hebrew}{2,}$", "u").test(w[0]);
 }
 function pendingQuestion(text, lastWords = 6) {
-  const t = text.toLowerCase().replace(/[^\p{L}\p{N}'\s]+/gu, " ").split(/\s+/).filter(Boolean).slice(-lastWords).join(" ");
-  return /\b(can|could|would|will|do|does|did|is|are|should)\s+(you|we|i|it|there|they|he|she)\b/.test(t) || /(^|\s)(האם|אתה יכול|את יכולה)(\s|$)/u.test(t);
+  const all = text.toLowerCase().replace(/[^\p{L}\p{N}'\s.!?]+/gu, " ");
+  const sentence = all.split(/[.!?]/).at(-1).split(/\s+/).filter(Boolean);
+  const t = sentence.slice(-lastWords).join(" ");
+  return /\b(can|could|would|will|do|does|did|is|are|should)\s+(you|we|i|it|there|they|he|she)\b/.test(t) || /(^|\s)(האם|אתה יכול|את יכולה)(\s|$)/u.test(t) || /\b(what|why|how|where|when|who|which)\s+(do|does|did|is|are|should|would|can|could|will)\s+(you|we|i|it|they|he|she)\b/.test(sentence.slice(-4 * lastWords).join(" "));
 }
 var SCRIPTS = [
   [new RegExp("\\p{Script=Hebrew}", "gu"), "he"],
@@ -14194,10 +14206,13 @@ var CHANNEL_MODES = {
     generateFor: ["backchannel", "bridge"],
     reactionGapMs: 6e3,
     generateBudgetMs: 400,
+    // A breath must outlast the word cadence: at ~2.5 words/s interims arrive every ~400ms, so a 250ms "breath" fired after
+    // every word and only the 1.2s gaps held it (~25 lines per 30s, live probe). Lines ≥2.5s apart (light after light 3s), ≤6 in any 30s.
     backchannelMinWords: 5,
-    backchannelGapMs: 1200,
-    minGapMs: 1200,
-    backchannelPauseMs: 250
+    backchannelGapMs: 3e3,
+    minGapMs: 2500,
+    backchannelPauseMs: 500,
+    maxPerWindow: 6
   }
 };
 var CHANNEL_MODE_NAMES = Object.keys(CHANNEL_MODES);
@@ -14246,6 +14261,9 @@ var SecondChannelOptions = class {
   /** Backchannel after backchannel needs this gap (~4-6 cues in a 30s monologue — Para phone call 2026-10-13: 5s left
    *  one cue in a long statement). */
   backchannelGapMs = 3e3;
+  /** Soft cap: at most this many backchannels in any `windowMs` (0 = none). */
+  maxPerWindow = 0;
+  windowMs = 3e4;
   /** The user resumes within this long of their last words after a line of theirs was dispatched: the same statement goes
    *  on (an STT endpoint mid-monologue) — its words still count toward the backchannel guards, and it is no fresh request. */
   continueMs = 3e3;
@@ -14390,12 +14408,7 @@ var SecondChannel = class {
   }
   get o() {
     const o = this.options;
-    if (o.mode !== "chatty" || o.aec) return o;
-    if (!this.degradedLogged) {
-      this.degradedLogged = true;
-      log32.info("chatty mode without echo cancellation \u2014 normal timing (our lines would re-trigger the mic)");
-    }
-    return { ...o, ...this.normalTiming };
+    return o.mode !== "chatty" || o.aec ? o : { ...o, ...this.normalTiming };
   }
   /** The user's live partial for the current turn (the whole partial, not a delta). */
   onInterim(text) {
@@ -14533,10 +14546,17 @@ var SecondChannel = class {
   allowed(k) {
     if (!this.o.kinds.includes(k)) return false;
     if (this.audioLocked || this.inflight || this.o.now() - this.lastAt < this.o.minGapMs || k === this.lastKind && (k !== "backchannel" || this.o.now() - this.lastAt < this.o.backchannelGapMs)) return false;
-    if (k === "backchannel") return this.o.aec && this.userSpeaking && this.main === "idle";
+    if (k === "backchannel") return this.o.aec && this.userSpeaking && this.main === "idle" && !this.capped();
     if (k === "filler") return !this.userSpeaking && this.main === "idle";
     return (this.main === "queued" || this.main === "thinking") && !this.bridged;
   }
+  bcTimes = [];
+  capped() {
+    if (!this.o.maxPerWindow) return false;
+    const from = this.o.now() - this.o.windowMs;
+    this.bcTimes = this.bcTimes.filter((t) => t > from);
+    return this.bcTimes.length >= this.o.maxPerWindow;
+  }
   /** 0ms rule: the candidate line for this trigger, or silence. */
   rule(trigger, pauseMs = 0, breath = false, cue = null) {
     const w = words2(this.heard);
@@ -14579,6 +14599,11 @@ var SecondChannel = class {
     const n0 = words2(this.heard).length;
     this.decidedAt = n0;
     if (!(this.o.anticipate && this.plan && trigger === "interim")) cue = null;
+    const opts = this.options;
+    if (opts.mode === "chatty" && !opts.aec && !this.degradedLogged) {
+      this.degradedLogged = true;
+      log32.info("chatty mode without echo cancellation \u2014 normal timing (our lines would re-trigger the mic)");
+    }
     const rule = this.rule(trigger, pauseMs, breath, cue);
     if (rule === "silence") return;
     this.busy = true;
@@ -14658,6 +14683,7 @@ var SecondChannel = class {
     if (r === false) return false;
     this.lastAt = this.o.now();
     this.lastKind = kind;
+    if (kind === "backchannel") this.bcTimes.push(this.lastAt);
     if (kind === "bridge") this.bridged = true;
     this.recent = [...this.recent, text].slice(-this.o.avoidLines);
     this.live = text.trim();