@livx.cc/agentx 0.99.92 → 0.99.93

package/dist/cli.d.ts

@@ -134,6 +134,7 @@ interface Args {
     version: boolean;
     duplex: boolean;
     duplexV2?: boolean;
+    channelOnly?: boolean;
     voiceModel?: string;
     thinkModel?: string | false;
     voice: boolean;

package/dist/cli.js

@@ -9757,6 +9757,10 @@ var DuplexAgentOptions = class {
   host;
   /** How many recent transcript messages are rendered into a worker's brief. */
   excerptTurns = 6;
+  /** A Hold turn whose user says nothing more for this long is answered after all (a re-voice turn told they stopped):
+   *  without it a wrong Hold left the turn as a lone "mm-hm" forever. 0 = off. Only for a host that reports the user's
+   *  live speech via userActive() — else a slow talker's next words would be spoken over. */
+  holdResumeMs = 0;
   /** Voice register: 'neutral' = clean spoken style; 'conversational' = human-like — fillers,
    *  backchannels, impulsive first reactions before content (mimics real duplex conversation). */
   voiceStyle = "neutral";
@@ -9883,7 +9887,7 @@ function isOfferAcceptance(text) {
 }
 var isQuestion = (s) => /\?["'”’)\]]*$/.test(s.trim());
 var OFFER_OVERRUN = 150;
-var VOICE_SYSTEM_PROMPT = 'You are a spoken voice assistant \u2014 the user HEARS everything you say. Use short sentences. One idea per sentence. No markdown, no bullet lists, no code blocks, no headings, no emoji. Never emit stage directions or parenthetical asides about your own process \u2014 nothing like "(waiting for the result...)" or "(checking)"; while work runs, either say it as plain speech or end your turn.\nThis holds even when asked to "print", "list", "show", or "make a table" \u2014 there is no screen for the spoken channel. Speak it as flowing prose ("Tuesday is half a meter, Wednesday a bit less\u2026"), or if they truly need it on screen, route it to Act to render. Never emit dashes or pipes into speech.\nKeep turns SHORT \u2014 one to three sentences, then stop. Never lecture, enumerate cases, or add caveats unprompted. Conversation is a fast exchange: give the one thing asked, and let the user pull more if they want it.\nYou have three cognitive tiers \u2014 like a human brain:\n\u2022 YOU (reflex) \u2014 instant, lightweight. Handle greetings, simple questions, status checks, QuickLook.\n\u2022 `Act` \u2014 your hands. A background worker with its own configured tools and access to the user\'s environment (files and shell{{WORKER_WEB}}). Use for reading, editing, searching, running tasks, building \u2014 any real work.\n{{THINK_SLOT}}\nWhen you are unsure whether you can do or access something, do NOT assume and do NOT claim a capability you have not confirmed. To check what you can do, QuickLook `capabilities` (instant \u2014 it lists your worker\'s real tools) and answer from that. Never promise an ability that is not in your capabilities; if it is not there, tell the user plainly you can\'t. To actually DO real work, call `Act`. When the user mentions their project, folder, files, or environment ("this project", "the current folder", "my code"), call `Act` IMMEDIATELY \u2014 do not ask for paths or details the worker can discover itself. Never pretend to have done the work or invent results \u2014 the worker\'s report is your only source.\nYou cannot mute the microphone or stop voice capture yourself \u2014 no tool does it. If the user asks you to stop listening or turn the voice off, never claim you did: tell them to say exactly "voice off" (handled by the app directly), or type /voice.\nYou are NOT a knowledge base. For any question whose answer needs SPECIFIC verifiable facts you do not already have in hand \u2014 how to build/configure/implement something, exact API, library, entitlement, command or option names, current events, or particular numbers, dates, or names \u2014 do NOT answer from your own memory: you will confidently make things up (a fake API, a wrong entitlement, an event that did not happen). Route it to `Act`, which can search and verify, and speak only what its report says. DELEGATION RULE \u2014 decide for yourself, the user never has to push: if you cannot answer confidently from the conversation plus trivial well-known knowledge, do NOT refuse and do NOT guess \u2014 dispatch `Act` immediately with a clear brief and say you are checking. Anything needing CURRENT data (weather, news, prices, dates, sky/astronomy, "right now"), real computation, or verification is an automatic dispatch \u2014 never a refusal. The user should never need to say "search the web" or "think harder" to make you act; needing fresh or verified information IS the trigger. Refuse only what your worker genuinely cannot do (check `capabilities`), and say why. Answer inline ONLY for general conversation, chit-chat, and trivia you are sure of, or facts you can see via QuickLook. When elaborating on a completed task ("tell me more", "the gist"), stay strictly within what that result actually said \u2014 if the user asks for something the result did not cover, that is NEW information: dispatch `Act`, do not improvise.\nALWAYS react before you work: the FIRST thing in your turn is a brief spoken acknowledgement of what you heard and what you are about to do ("got it \u2014 opening that now", "sure, let me pull it up", "okay, checking"). NEVER call a tool (Act, Think, QuickLook) silently \u2014 the user must hear you react before you go quiet to work. After dispatching Act or Think, that same one short sentence IS your turn \u2014 end it and do not wait for the result. Exactly ONE short line: never stack a second acknowledgement, and never narrate your own presence or status while waiting ("I\'m here", "let me see", "still checking" right after you already acked). One clean ack then silence reads as competent; repeated check-ins read as nervous.\nA completed task speaks its OWN result to the user (the worker voices what matters as it finishes) \u2014 you do NOT re-voice clean task results. A FAILED or INCOMPLETE task still arrives as a "[task t1 failed] \u2026" event for you to handle. The completed result stays in YOUR context \u2014 it is yours to draw on. When the user follows up ("tell me more", "what else", "and?"), answer FROM that result first: you already have the detail, so elaborate on what you have. Do NOT spawn a fresh worker to re-search or re-gather what you were just handed. Re-dispatch ONLY when genuinely new information is needed \u2014 e.g. the user wants the full contents of a SPECIFIC source, which is one WebFetch of that URL, not a brand-new search. "[task t1 progress] \u2026" events are interim status, NOT results \u2014 on a genuinely LONG wait you MAY give one brief half-sentence aside, but only occasionally; silence while working is normal and fine (the user knows you are on it). Do NOT narrate every step or re-announce yourself. Never present progress as a finished result.\nCRITICAL: while a task is still running you have NO answer yet \u2014 never state a specific result of any kind (a number, size, count, name, path, or value). The real answer exists ONLY once the task finishes (the worker speaks it, and `TaskStatus` then returns it); inventing one meanwhile (a made-up disk size, commit count, etc.) is a serious error. Until then, only acknowledge and wait.\nNever read raw file paths, diffs, or code aloud verbatim.\nDo NOT end every turn with the same canned offer ("want a rundown?", "want the steps?"). Offer once at most; if the user pushes back, repeats themselves, or sounds unsatisfied ("you know what I mean?", "think deeper", "are you sure?"), do NOT re-offer the same thing \u2014 change approach: dispatch `Act`/`Think` to actually dig in, or ask one concrete clarifying question. Repeating a non-answer is worse than silence.\n"[task t1 asks] \u2026" events are QUESTIONS from a background task \u2014 relay to the user in your own words, short, then end your turn. When the user answers, call `AnswerTask` with that id and their answer. NEVER answer on the user\'s behalf for permissions or risky operations; if their reply is ambiguous, confirm first.\nIf the user\'s message sounds INCOMPLETE \u2014 trailing off mid-sentence, a fragment that needs more context ("and then we", "but the problem is"), hesitation fillers ("uh", "um") \u2014 call `Hold` instead of answering. This keeps listening for the rest of their thought. Only respond with substance when you have a complete question or request.\nDispatch discipline: send ONE self-contained task per request \u2014 a single worker with the full brief beats several workers with fragments (each worker starts fresh and re-discovers context). NEVER dispatch a worker just to read files or gather information \u2014 workers explore and discover context themselves; pass on what you already know and let one worker do the whole job. Split into parallel tasks only when the user asks for genuinely independent things. When a task completes, report its result and stop \u2014 do NOT dispatch follow-up work (verification, polish, extras) the user did not ask for, unless the report itself signals failure or doubt.\nDo not fire a second Act/Think for work already in flight, and NEVER spawn a second task to re-count, cross-check, or verify a result a worker already gave you \u2014 trust its answer; a single question gets ONE task. Call `TaskStatus` at most ONCE per turn; if a task is still running, give one brief status aside and end the turn \u2014 never poll it again and again in a loop. "Still on it" is ONLY for a repeat check on a task you already acknowledged earlier \u2014 NEVER as the ack when you first dispatch work ("still" is wrong when nothing was running yet; use a fresh "on it"/"got it, checking" instead). Use `CancelTask` when the user asks to stop something \u2014 to stop everything ("stop all", "cancel everything") pass id "all". When the user only ADDS to or CORRECTS a request whose task is still running (a detail, a path, a name, a spelling), call `Act` with `refines` set to that task id \u2014 it updates the running worker in place; never cancel and re-dispatch for a refinement.\nPRIORITY: when the user says goodbye or wants to end/finish/wrap up the session ("ok bye", "that\'s all", "let\'s finish", "let\'s end", "goodnight", "exit", "wrap up"), call `ExitSession` IMMEDIATELY \u2014 do not act, do not check status, just exit.\nFor TRIVIAL instant lookups only \u2014 current time, git branch, listing a folder, peeking at a small file, or checking your own `capabilities`/tools \u2014 use `QuickLook` (instant, no task). Whenever the user asks what you can do or whether you have some ability, QuickLook `capabilities` and answer from that \u2014 never guess. Anything requiring searching, reasoning, running commands, or editing goes through `Act`.\n{{MEMORY_SLOT}}\nUser messages may arrive via speech-to-text (audio transcribed to text) and can carry transcription errors \u2014 odd words, cut-offs, homophones ("for you" vs "folder"), and especially MISHEARD NAMES: a surname, company, or project name often comes out as ordinary sound-alike words. Read for INTENT, not surface text, and do not silently rewrite what they said. If a message seems garbled, surprising, or only half-parses, or a word sounds like a name you already know (from memory or this conversation) but is spelled differently, treat it with suspicion \u2014 do NOT act on the odd word blindly or improvise content from it; briefly confirm what they meant ("did you mean\u2026?") and wait. A one-line confirm beats a confident wrong answer. NEVER save to memory a name, spelling, or term you only HEARD (it may be a transcription error) \u2014 when the user spells something aloud, read the spelling back and get a yes before storing it.';
+var VOICE_SYSTEM_PROMPT = 'You are a spoken voice assistant \u2014 the user HEARS everything you say. Use short sentences. One idea per sentence. No markdown, no bullet lists, no code blocks, no headings, no emoji. Never emit stage directions or parenthetical asides about your own process \u2014 nothing like "(waiting for the result...)" or "(checking)"; while work runs, either say it as plain speech or end your turn.\nThis holds even when asked to "print", "list", "show", or "make a table" \u2014 there is no screen for the spoken channel. Speak it as flowing prose ("Tuesday is half a meter, Wednesday a bit less\u2026"), or if they truly need it on screen, route it to Act to render. Never emit dashes or pipes into speech.\nKeep turns SHORT \u2014 one to three sentences, then stop. Never lecture, enumerate cases, or add caveats unprompted. Conversation is a fast exchange: give the one thing asked, and let the user pull more if they want it.\nYou have three cognitive tiers \u2014 like a human brain:\n\u2022 YOU (reflex) \u2014 instant, lightweight. Handle greetings, simple questions, status checks, QuickLook.\n\u2022 `Act` \u2014 your hands. A background worker with its own configured tools and access to the user\'s environment (files and shell{{WORKER_WEB}}). Use for reading, editing, searching, running tasks, building \u2014 any real work.\n{{THINK_SLOT}}\nWhen you are unsure whether you can do or access something, do NOT assume and do NOT claim a capability you have not confirmed. To check what you can do, QuickLook `capabilities` (instant \u2014 it lists your worker\'s real tools) and answer from that. Never promise an ability that is not in your capabilities; if it is not there, tell the user plainly you can\'t. To actually DO real work, call `Act`. When the user mentions their project, folder, files, or environment ("this project", "the current folder", "my code"), call `Act` IMMEDIATELY \u2014 do not ask for paths or details the worker can discover itself. Never pretend to have done the work or invent results \u2014 the worker\'s report is your only source.\nYou cannot mute the microphone or stop voice capture yourself \u2014 no tool does it. If the user asks you to stop listening or turn the voice off, never claim you did: tell them to say exactly "voice off" (handled by the app directly), or type /voice.\nYou are NOT a knowledge base. For any question whose answer needs SPECIFIC verifiable facts you do not already have in hand \u2014 how to build/configure/implement something, exact API, library, entitlement, command or option names, current events, or particular numbers, dates, or names \u2014 do NOT answer from your own memory: you will confidently make things up (a fake API, a wrong entitlement, an event that did not happen). Route it to `Act`, which can search and verify, and speak only what its report says. DELEGATION RULE \u2014 decide for yourself, the user never has to push: if you cannot answer confidently from the conversation plus trivial well-known knowledge, do NOT refuse and do NOT guess \u2014 dispatch `Act` immediately with a clear brief and say you are checking. Anything needing CURRENT data (weather, news, prices, dates, sky/astronomy, "right now"), real computation, or verification is an automatic dispatch \u2014 never a refusal. The user should never need to say "search the web" or "think harder" to make you act; needing fresh or verified information IS the trigger. Refuse only what your worker genuinely cannot do (check `capabilities`), and say why. Answer inline ONLY for general conversation, chit-chat, and trivia you are sure of, or facts you can see via QuickLook. When elaborating on a completed task ("tell me more", "the gist"), stay strictly within what that result actually said \u2014 if the user asks for something the result did not cover, that is NEW information: dispatch `Act`, do not improvise.\nALWAYS react before you work: the FIRST thing in your turn is a brief spoken acknowledgement of what you heard and what you are about to do ("got it \u2014 opening that now", "sure, let me pull it up", "okay, checking"). NEVER call a tool (Act, Think, QuickLook) silently \u2014 the user must hear you react before you go quiet to work. After dispatching Act or Think, that same one short sentence IS your turn \u2014 end it and do not wait for the result. Exactly ONE short line: never stack a second acknowledgement, and never narrate your own presence or status while waiting ("I\'m here", "let me see", "still checking" right after you already acked). One clean ack then silence reads as competent; repeated check-ins read as nervous.\nA completed task speaks its OWN result to the user (the worker voices what matters as it finishes) \u2014 you do NOT re-voice clean task results. A FAILED or INCOMPLETE task still arrives as a "[task t1 failed] \u2026" event for you to handle. The completed result stays in YOUR context \u2014 it is yours to draw on. When the user follows up ("tell me more", "what else", "and?"), answer FROM that result first: you already have the detail, so elaborate on what you have. Do NOT spawn a fresh worker to re-search or re-gather what you were just handed. Re-dispatch ONLY when genuinely new information is needed \u2014 e.g. the user wants the full contents of a SPECIFIC source, which is one WebFetch of that URL, not a brand-new search. "[task t1 progress] \u2026" events are interim status, NOT results \u2014 on a genuinely LONG wait you MAY give one brief half-sentence aside, but only occasionally; silence while working is normal and fine (the user knows you are on it). Do NOT narrate every step or re-announce yourself. Never present progress as a finished result.\nCRITICAL: while a task is still running you have NO answer yet \u2014 never state a specific result of any kind (a number, size, count, name, path, or value). The real answer exists ONLY once the task finishes (the worker speaks it, and `TaskStatus` then returns it); inventing one meanwhile (a made-up disk size, commit count, etc.) is a serious error. Until then, only acknowledge and wait.\nNever read raw file paths, diffs, or code aloud verbatim.\nDo NOT end every turn with the same canned offer ("want a rundown?", "want the steps?"). Offer once at most; if the user pushes back, repeats themselves, or sounds unsatisfied ("you know what I mean?", "think deeper", "are you sure?"), do NOT re-offer the same thing \u2014 change approach: dispatch `Act`/`Think` to actually dig in, or ask one concrete clarifying question. Repeating a non-answer is worse than silence.\n"[task t1 asks] \u2026" events are QUESTIONS from a background task \u2014 relay to the user in your own words, short, then end your turn. When the user answers, call `AnswerTask` with that id and their answer. NEVER answer on the user\'s behalf for permissions or risky operations; if their reply is ambiguous, confirm first.\nIf the user\'s message sounds INCOMPLETE \u2014 it ENDS mid-sentence or on a hesitation ("and then we", "but the problem is", "so I was, um") \u2014 call `Hold` instead of answering. A hesitation INSIDE a finished message ("can you, um, open it?") does not make it incomplete, and an unclear or odd but finished message is not a fragment: answer it, or ask one short question \u2014 never Hold on it. This keeps listening for the rest of their thought. Only respond with substance when you have a complete question or request.\nDispatch discipline: send ONE self-contained task per request \u2014 a single worker with the full brief beats several workers with fragments (each worker starts fresh and re-discovers context). NEVER dispatch a worker just to read files or gather information \u2014 workers explore and discover context themselves; pass on what you already know and let one worker do the whole job. Split into parallel tasks only when the user asks for genuinely independent things. When a task completes, report its result and stop \u2014 do NOT dispatch follow-up work (verification, polish, extras) the user did not ask for, unless the report itself signals failure or doubt.\nDo not fire a second Act/Think for work already in flight, and NEVER spawn a second task to re-count, cross-check, or verify a result a worker already gave you \u2014 trust its answer; a single question gets ONE task. Call `TaskStatus` at most ONCE per turn; if a task is still running, give one brief status aside and end the turn \u2014 never poll it again and again in a loop. "Still on it" is ONLY for a repeat check on a task you already acknowledged earlier \u2014 NEVER as the ack when you first dispatch work ("still" is wrong when nothing was running yet; use a fresh "on it"/"got it, checking" instead). Use `CancelTask` when the user asks to stop something \u2014 to stop everything ("stop all", "cancel everything") pass id "all". When the user only ADDS to or CORRECTS a request whose task is still running (a detail, a path, a name, a spelling), call `Act` with `refines` set to that task id \u2014 it updates the running worker in place; never cancel and re-dispatch for a refinement.\nPRIORITY: when the user says goodbye or wants to end/finish/wrap up the session ("ok bye", "that\'s all", "let\'s finish", "let\'s end", "goodnight", "exit", "wrap up"), call `ExitSession` IMMEDIATELY \u2014 do not act, do not check status, just exit.\nFor TRIVIAL instant lookups only \u2014 current time, git branch, listing a folder, peeking at a small file, or checking your own `capabilities`/tools \u2014 use `QuickLook` (instant, no task). Whenever the user asks what you can do or whether you have some ability, QuickLook `capabilities` and answer from that \u2014 never guess. Anything requiring searching, reasoning, running commands, or editing goes through `Act`.\n{{MEMORY_SLOT}}\nUser messages may arrive via speech-to-text (audio transcribed to text) and can carry transcription errors \u2014 odd words, cut-offs, homophones ("for you" vs "folder"), and especially MISHEARD NAMES: a surname, company, or project name often comes out as ordinary sound-alike words. Read for INTENT, not surface text, and do not silently rewrite what they said. If a message seems garbled, surprising, or only half-parses, or a word sounds like a name you already know (from memory or this conversation) but is spelled differently, treat it with suspicion \u2014 do NOT act on the odd word blindly or improvise content from it; briefly confirm what they meant ("did you mean\u2026?") and wait. A one-line confirm beats a confident wrong answer. NEVER save to memory a name, spelling, or term you only HEARD (it may be a transcription error) \u2014 when the user spells something aloud, read the spelling back and get a yes before storing it.';
 var THINK_GUIDANCE = "\u2022 `Think` \u2014 your brain. A premium reasoning model, FAR more expensive than Act. Reserve it for open-ended architecture/design questions, or a problem Act already FAILED at. ALL implementation work \u2014 coding, refactoring, debugging, edge cases, tests \u2014 goes to Act; Act is highly capable. Never send the same work to both.";
 var THINK_DISABLED_GUIDANCE = "(Think tier is not available \u2014 use Act for all escalations.)";
 var VOICE_STYLE_CONVERSATIONAL = `Speak like a person in a live conversation, not an assistant reading a script. React first, then deliver: a quick impulsive beat ("oh nice", "hmm, hold on", "ah, got it") before the substance. Use contractions always. Vary sentence length \u2014 some very short. Light fillers and backchannels are fine ("mm-hm", "right", "let's see") but at most one per reply \u2014 never stack them. When you escalate to Act or Think, say it like a human would ("hang on, let me actually dig into that \u2014 gimme a minute") instead of announcing a task. When a result comes back, react to it like you just found out ("okay so \u2014 turns out\u2026"). Match the user's energy: a quick question gets a quick answer \u2014 a few words is a perfectly good turn. Prefer a short answer plus an offer ("want the details?") over covering everything. Never narrate your own mechanics (no "I will now act", no task ids out loud).`;
@@ -9938,6 +9942,10 @@ var DuplexAgent = class _DuplexAgent {
   // host spoke on our behalf (adaptive micro-ack) — not reflex output
   heldThisTurn = false;
   // Hold called this turn → turn is INTENTIONALLY silent (suppress reflex text + no dead-air ack)
+  holdTimer = null;
+  // armed by a Hold turn's end (holdResumeMs)
+  holdResumed = false;
+  // the hold expired: the resume turn must answer (Hold refused) — until the user speaks
   nudging = false;
   // re-ack pass in flight: block ALL tools, prevent recursion
   reflexBuf = "";
@@ -10293,6 +10301,10 @@ ${nowLine()}. Anchor every relative or time-sensitive reference \u2014 "today",
     await this.drainReflex();
     this.scrubLeakFromTranscript();
     this.scrubHeldText();
+    if (this.heldThisTurn) {
+      this.armHoldResume();
+      return;
+    }
     if (this.silentTurn && res?.finishReason === "error" && refusedCall(res.error)) this.emitHost({ kind: "text_delta", message: fallback ?? _DuplexAgent.FALLBACK_ERROR });
     else if (this.silentTurn) await this.ackIfSilent(fallback);
   }
@@ -10375,6 +10387,27 @@ ${nowLine()}. Anchor every relative or time-sensitive reference \u2014 "today",
   noteExternalSpeech() {
     this.externalSpeech = true;
   }
+  /** The user is speaking (a live partial, a barge): a pending Hold expiry waits for them again. */
+  userActive() {
+    if (this.holdTimer) this.armHoldResume();
+  }
+  clearHoldResume() {
+    if (this.holdTimer) clearTimeout(this.holdTimer);
+    this.holdTimer = null;
+  }
+  armHoldResume() {
+    this.clearHoldResume();
+    const ms = this.options.holdResumeMs;
+    if (!ms) return;
+    this.holdTimer = setTimeout(() => {
+      this.holdTimer = null;
+      this.holdResumed = true;
+      this.notify("diag", "hold_expired", { ms });
+      this.queueRevoice(_DuplexAgent.HOLD_EXPIRED);
+    }, ms);
+    this.holdTimer.unref?.();
+  }
+  static HOLD_EXPIRED = "[the user stopped there and did not go on \u2014 respond to what they said now; if it is unclear, ask one short question. Do not Hold.]";
   /** True when the just-finished turn voiced NOTHING — dead air to repair. Two ways this happens:
    *  (a) a dispatch with no spoken ack, and (b) an INLINE turn whose `final` channel came back empty —
    *  gpt-oss harmony sometimes puts the whole reply in `analysis` (→ thinking_delta, suppressed in
@@ -10437,6 +10470,8 @@ ${nowLine()}. Anchor every relative or time-sensitive reference \u2014 "today",
    *  second Act). The host gets `turn_replace` {id, replaces, rolledBack, followed?}: only a rollback drops the replaced
    *  line; `followed` = the kept turn whose follow-up ran (absent: the merged text ran). */
   send(content, opts) {
+    this.clearHoldResume();
+    this.holdResumed = false;
     if (opts?.replaces) this.replaceTurn(opts.replaces);
     const cur = this.voice.options.signal;
     const ext = cur && cur === this.turnSignal?.own ? this.turnSignal.ext : cur;
@@ -11465,6 +11500,8 @@ Another agent just implemented the above. Independently check the CURRENT state
         }
       },
       run: async ({ filler }) => {
+        if (this.heldThisTurn) return "Already holding \u2014 stay silent.";
+        if (this.holdResumed) return "Not now \u2014 the user already stopped after those words. Answer them now in one or two short sentences, or ask one short question.";
         this.heldThisTurn = true;
         this.notify("hold_filler", filler ? String(filler) : "");
         return "Holding \u2014 listening for the rest of the user's thought. Do not respond further this turn.";
@@ -13657,7 +13694,7 @@ var scriptOf = (w) => SCRIPTS.find(([re]) => {
   re.lastIndex = 0;
   return re.test(w);
 })?.[1] ?? null;
-function detectLanguage(text) {
+function detectLanguage(text, switching = false) {
   const ws = text.toLowerCase().match(new RegExp("\\p{L}+", "gu")) ?? [];
   const n = /* @__PURE__ */ new Map();
   for (const w of ws) {
@@ -13665,9 +13702,10 @@ function detectLanguage(text) {
     if (s) n.set(s, (n.get(s) ?? 0) + 1);
   }
   const [best, c] = [...n].sort((a2, b2) => b2[1] - a2[1])[0] ?? [null, 0];
-  if (best !== "latin") return c >= 1 && (best === "zh" || best === "ja" || c >= 2 || ws.length === 1) ? best : null;
+  if (best !== "latin") return c >= 1 && (best === "zh" || best === "ja" || c >= 2 || ws.length === 1 && !switching) ? best : null;
   const score = Object.entries(STOP3).map(([l2, s]) => [l2, ws.filter((w) => s.has(w)).length]).sort((a2, b2) => b2[1] - a2[1]);
   const [[l, a], [, b]] = score;
+  if (switching) return a - b >= 2 && ws.length >= 4 ? l : null;
   return a > b && (a >= 2 || a === 1 && ws.length <= 3) ? l : null;
 }
 var NAMES = new Intl.DisplayNames(["en"], { type: "language" });
@@ -13716,8 +13754,8 @@ function validateBank(raw, lang, minPerKind = 3) {
 }
 var KIND_HINT = {
   backchannel: 'tiny listener cues while they keep talking (like "mm-hm", "right") \u2014 never takes the turn',
-  filler: 'soft thinking sounds at their pause (like "hmm\u2026")',
-  bridge: 'quick lead-ins before the real answer while it is prepared (like "let me check\u2014")',
+  filler: 'soft LISTENER sounds while they pause mid-thought (like "mm\u2026", "mhm", "yeah\u2026") \u2014 never a word that continues or completes their sentence ("kind of", "basically", "you know") and never the start of yours',
+  bridge: `neutral one-moment cues while the answer is prepared (like "let me see\u2026", "one sec\u2026") \u2014 never the start of a statement or opinion ("the thing is", "well actually", "here's the deal")`,
   phatic: 'greetings, thanks and yes/no openers a person starts a turn with (like "hi", "thanks", "how are you")'
 };
 function fastBankGenerator(ai, model, perKind = 12) {
@@ -13750,6 +13788,7 @@ var ChannelBanksOptions = class {
   fallback = { en: EN_BANK };
   now = () => Date.now();
 };
+var BANK_VERSION = 2;
 var ChannelBanks = class {
   constructor(options) {
     this.options = options;
@@ -13780,6 +13819,10 @@ var ChannelBanks = class {
       if (e.code !== "ENOENT") log29.warn(`bank cache ${f} unreadable \u2014 regenerating: ${e.message}`);
       return null;
     });
+    if (cached && cached.v !== BANK_VERSION) {
+      log29.debug(`bank ${lang}: cache v${cached.v ?? 1} predates v${BANK_VERSION} \u2014 regenerating`);
+      cached = null;
+    }
     const valid = cached && validateBank(cached.bank, lang, 0);
     const complete = !!valid && BANK_KEYS.every((k) => valid[k].length);
     if (valid) {
@@ -13806,7 +13849,7 @@ var ChannelBanks = class {
       this.banks.set(lang, bank);
       log29.debug(`bank ${lang}: ${have ? "refreshed" : "generated"} (${BANK_KEYS.map((k) => bank[k].length).join("/")}) in ${Math.round(this.o.now() - t0)}ms`);
       const f = this.file(lang);
-      if (f) await mkdir(this.o.cacheDir, { recursive: true }).then(() => writeFile(f, JSON.stringify({ at: this.o.now(), bank }))).catch((e) => log29.warn(`bank cache write ${f} failed: ${e.message}`));
+      if (f) await mkdir(this.o.cacheDir, { recursive: true }).then(() => writeFile(f, JSON.stringify({ at: this.o.now(), v: BANK_VERSION, bank }))).catch((e) => log29.warn(`bank cache write ${f} failed: ${e.message}`));
     } catch (e) {
       log29.warn(`bank ${lang}: generation failed \u2014 ${have ? "keeping the cached bank" : lang in this.o.fallback ? "fallback bank" : "silent"}: ${e.message}`);
     } finally {
@@ -13829,7 +13872,8 @@ var SecondChannelOptions = class {
    *  tell it confidently; the channel is silent until then (never English by default). */
   lang;
   /** Kinds worth generating (content-conditioned); the rest are canned sounds. */
-  generateFor = ["backchannel", "bridge"];
+  generateFor = ["bridge"];
+  // a backchannel is timing, not content: canned at 0ms (a generated one arrived after the micro-pause)
   /** Host TTS on the second channel; resolves when the line finished playing. `false` = the host could not play it. */
   speak = () => {
   };
@@ -13855,7 +13899,12 @@ var SecondChannelOptions = class {
   avoidLines = 4;
   /** Canned picks exclude the last N picks of the same kind. */
   bankExclude = 2;
-  backchannelMinWords = 8;
+  backchannelMinWords = 6;
+  /** No new interim for this long mid-utterance (a breath / clause gap, shorter than a pause) → a backchannel moment. */
+  backchannelPauseMs = 350;
+  /** A bridge waits this long after the user's turn is dispatched, and plays only if channel-1 has said nothing yet
+   *  (a fast reply/ack needs no bridge — one before it reads as the agent starting a different sentence). 0 = at once. */
+  bridgeDelayMs = 1200;
   /** No new interim for this long → a pause (hosts may also call onPause). */
   fillerPauseMs = 700;
   fillerMinWords = 4;
@@ -13888,13 +13937,14 @@ var SecondChannel = class {
   heard = "";
   userSpeaking = false;
   main = "idle";
-  lead = "";
   gen = 0;
   // bumped by barge-in / a new user turn → in-flight decisions drop
   decidedAt = 0;
   // word count at the last decision
   busy = false;
   pauseTimer = null;
+  bcTimer = null;
+  bridgeTimer = null;
   lastAt = -Infinity;
   lastKind = null;
   live = null;
@@ -13926,9 +13976,11 @@ var SecondChannel = class {
     this.picks.clear();
     void this.o.banks?.ensure(lang);
   }
+  /** Once a language is known, switching needs strong evidence (detectLanguage's `switching`): an English "Um," read as
+   *  Portuguese flipped a live session to pt (silence until a pt bank was generated — it then spoke Portuguese). */
   detect(text) {
     if (this.pinned) return;
-    const l = detectLanguage(text);
+    const l = detectLanguage(text, !!this.lang);
     if (l) this.setLanguage(l);
   }
   get bank() {
@@ -13950,6 +14002,11 @@ var SecondChannel = class {
       this.pauseTimer = null;
       this.onPause(this.o.fillerPauseMs);
     }, this.o.fillerPauseMs);
+    if (this.bcTimer) clearTimeout(this.bcTimer);
+    this.bcTimer = this.o.aec ? setTimeout(() => {
+      this.bcTimer = null;
+      if (this.userSpeaking && this.heard === t) void this.consider("interim", 0, true);
+    }, this.o.backchannelPauseMs) : null;
     if (words2(t).length - this.decidedAt >= this.o.decideEveryWords) void this.consider("interim");
   }
   /** The user went quiet for `ms` (host VAD/STT pause, or the internal interim timer). */
@@ -13958,6 +14015,10 @@ var SecondChannel = class {
       clearTimeout(this.pauseTimer);
       this.pauseTimer = null;
     }
+    if (this.bcTimer) {
+      clearTimeout(this.bcTimer);
+      this.bcTimer = null;
+    }
     this.userSpeaking = false;
     void this.consider("pause", ms);
   }
@@ -13966,22 +14027,30 @@ var SecondChannel = class {
     if (s === this.main) return;
     this.main = s;
     if (s === "first_audio" || s === "speaking") this.live = null;
+    if (s !== "queued" && s !== "thinking" && this.bridgeTimer) {
+      clearTimeout(this.bridgeTimer);
+      this.bridgeTimer = null;
+    }
     if (s === "idle") {
       this.heard = "";
       this.decidedAt = 0;
-      this.lead = "";
       this.bridged = false;
       return;
     }
     if (s === "queued") {
       this.gen++;
       this.userSpeaking = false;
-      this.lead = "";
       this.bridged = false;
-      if (this.pauseTimer) clearTimeout(this.pauseTimer);
-      this.pauseTimer = null;
+      for (const k of ["pauseTimer", "bcTimer", "bridgeTimer"]) {
+        if (this[k]) clearTimeout(this[k]);
+        this[k] = null;
+      }
+      if (this.o.bridgeDelayMs > 0) this.bridgeTimer = setTimeout(() => {
+        this.bridgeTimer = null;
+        void this.consider("working");
+      }, this.o.bridgeDelayMs);
+      else void this.consider("working");
     }
-    if (s === "queued" || s === "thinking" || s === "streaming_text") void this.consider("working");
   }
   /** The user's final line was dispatched to the main agent (its words are the bridge's micro-context). */
   onUserTurn(text) {
@@ -13990,10 +14059,9 @@ var SecondChannel = class {
     this.main = "idle";
     this.onAgentState("queued");
   }
-  /** The main answer's streamed text (before its audio): a bridge can lead into it. */
+  /** The main answer's streamed text (before its audio): channel-1 is responding — no bridge from here on. */
   onMainTokens(text) {
-    if (this.lead.length < 120) this.lead += text;
-    if (this.main === "queued" || this.main === "thinking") this.onAgentState("streaming_text");
+    if (text.trim() && (this.main === "queued" || this.main === "thinking")) this.onAgentState("streaming_text");
   }
   /** The user barged in: cut the in-flight line now and drop every pending decision. */
   onBargeIn() {
@@ -14022,9 +14090,6 @@ var SecondChannel = class {
   get audioLocked() {
     return this.main === "first_audio" || this.main === "speaking";
   }
-  get working() {
-    return this.main === "queued" || this.main === "thinking" || this.main === "streaming_text";
-  }
   tail() {
     return words2(this.heard).slice(-this.o.microWords).join(" ");
   }
@@ -14033,21 +14098,29 @@ var SecondChannel = class {
     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 === "filler") return !this.userSpeaking && this.main === "idle";
-    return this.working && !this.bridged;
+    return (this.main === "queued" || this.main === "thinking") && !this.bridged;
   }
   /** 0ms rule: the candidate line for this trigger, or silence. */
-  rule(trigger, pauseMs = 0) {
+  rule(trigger, pauseMs = 0, breath = false) {
     const w = words2(this.heard);
     if (!this.lang) return "silence";
     if (w.length <= 3 || w.length <= 6 && isPhatic(this.heard, this.bank)) return "silence";
-    const k = trigger === "interim" ? w.length >= this.o.backchannelMinWords && /[,.;:!?…]$/.test(this.heard) ? "backchannel" : null : trigger === "pause" ? pauseMs >= this.o.fillerPauseMs && w.length >= this.o.fillerMinWords ? "filler" : null : w.length >= this.o.bridgeMinWords ? "bridge" : null;
+    const k = trigger === "interim" ? this.backchannelMoment(w.length, breath) ? "backchannel" : null : trigger === "pause" ? pauseMs >= this.o.fillerPauseMs && w.length >= this.o.fillerMinWords ? "filler" : null : w.length >= this.o.bridgeMinWords ? "bridge" : null;
     return k && this.allowed(k) ? k : "silence";
   }
-  async consider(trigger, pauseMs) {
+  /** A clause boundary mid-floor: a breath or a comma-like close. Not a question (it wants the answer, not "mm-hm"), and a
+   *  sentence end only deep into a story — a short request's end is the end of their turn (replay: "mm-hm" right after
+   *  "Check hacker news for recent news.", just before the final). */
+  backchannelMoment(n, breath) {
+    if (n < this.o.backchannelMinWords || /\?$/.test(this.heard)) return false;
+    if (/[.!]$/.test(this.heard)) return n >= 2 * this.o.backchannelMinWords;
+    return breath || /[,;:…]$/.test(this.heard);
+  }
+  async consider(trigger, pauseMs, breath = false) {
     if (this.busy) return;
     const n0 = words2(this.heard).length;
     this.decidedAt = n0;
-    const rule = this.rule(trigger, pauseMs);
+    const rule = this.rule(trigger, pauseMs, breath);
     if (rule === "silence") return;
     this.busy = true;
     const gen = this.gen, t0 = this.o.now();
@@ -14087,7 +14160,7 @@ var SecondChannel = class {
   async line(kind) {
     if (this.o.generate && this.o.generateFor.includes(kind)) {
       const ac = new AbortController();
-      const ctx = { kind, heard: this.tail(), state: this.main, avoid: [...this.recent], lang: this.lang, ...this.lead ? { lead: this.lead } : {} };
+      const ctx = { kind, heard: this.tail(), state: this.main, avoid: [...this.recent], lang: this.lang };
       const raw = await withBudget(this.o.generate(ctx, ac.signal), this.o.generateBudgetMs, `${kind} generate`);
       if (raw === null) ac.abort();
       const text = cleanLine(raw ?? "", this.lang);
@@ -14135,17 +14208,16 @@ var SecondChannel = class {
   }
 };
 var KIND_HINT2 = {
-  backchannel: "a tiny listener cue while they keep talking \u2014 shows you follow THEIR point, does not take the turn",
-  filler: "a soft thinking sound at their pause",
-  bridge: "a quick lead-in before the real answer, while it is being prepared"
+  backchannel: 'a tiny listener cue while they keep talking ("mm-hm", "right", "yeah") \u2014 shows you follow, never takes the turn',
+  filler: 'a soft listener sound while THEY search for words ("mm\u2026", "mhm", "yeah\u2026") \u2014 never a word that continues their sentence or starts yours',
+  bridge: 'a neutral one-moment cue while the answer is prepared ("let me see\u2026", "one sec\u2026") \u2014 never the start of a statement or opinion ("the thing is", "well actually")'
 };
 function fastLineGenerator(ai, model) {
   return async (c, signal) => {
     const prompt = `You voice the tiny side-sounds of a live spoken conversation. Write ONE line of 1-5 words: ${KIND_HINT2[c.kind]}.
 The user just said: "${c.heard}"
 The assistant is: ${c.state}.
-` + (c.lead ? `The real answer begins: "${c.lead.trim()}" \u2014 lead naturally into it; do not repeat it.
-` : "") + (c.avoid.length ? `Do not reuse any of: ${c.avoid.map((a) => `"${a}"`).join(", ")}.
+` + (c.avoid.length ? `Do not reuse any of: ${c.avoid.map((a) => `"${a}"`).join(", ")}.
 ` : "") + `React to their actual words, casually, in ${languageName(c.lang)}. Output only the line.`;
     const reasoning = /gpt-oss/.test(model) ? { providerOptions: { reasoning_effort: "low" } } : {};
     const r = await ai.chat({ model, messages: [{ role: "user", content: prompt }], stream: false, signal, maxTokens: 150, ...reasoning });
@@ -17283,7 +17355,8 @@ function parseArgs(argv) {
     else if (x === "--duplex-v2") {
       a.duplex = true;
       a.duplexV2 = true;
-    } else if (x === "--conversational" || x === "--convo" || x === "--voice") {
+    } else if (x === "--channel-only") a.channelOnly = true;
+    else if (x === "--conversational" || x === "--convo" || x === "--voice") {
       a.voice = true;
       a.duplex = true;
     } else if (x === "--voice-model") a.voiceModel = val(++i, x);
@@ -17322,10 +17395,11 @@ function parseArgs(argv) {
   if (a.seed && !a.boddb) throw new Error("--seed only applies with --boddb (it seeds the database from cwd on first run)");
   if (a.worktree && (a.vfs || a.boddb)) throw new Error("--worktree is a real-disk concept \u2014 incompatible with --vfs/--boddb");
   if (a.duplex && (a.task || a.print)) throw new Error("--duplex is interactive-only (a conversational mode) \u2014 drop the task/-p");
-  if (a.inputFormat && (!a.duplex && !a.print || a.outputFormat !== "stream-json")) throw new Error("--input-format stream-json needs --output-format stream-json and --duplex (or --voice) or -p");
+  if (a.inputFormat && (!a.duplex && !a.print && !a.channelOnly || a.outputFormat !== "stream-json")) throw new Error("--input-format stream-json needs --output-format stream-json and --duplex (or --voice), --channel-only or -p");
+  if (a.channelOnly && !a.inputFormat) throw new Error("--channel-only needs --input-format stream-json --output-format stream-json");
   if (a.duplex && a.plan) throw new Error("--plan is not supported in --duplex (workers are non-interactive; a plan could never be approved)");
   if (a.voiceSystemPrompt && !a.duplex) throw new Error("--voice-system-prompt only applies with --duplex");
-  if (a.voiceModel && !a.duplex) throw new Error("--voice-model only applies with --duplex");
+  if (a.voiceModel && !a.duplex && !a.channelOnly) throw new Error("--voice-model only applies with --duplex or --channel-only");
   if (a.thinkModel !== void 0 && !a.duplex) throw new Error("--think-model/--no-think only apply with --duplex");
   if (!canPrompt && (a.ask || a.plan))
     throw new Error(`${a.ask ? "--ask" : "--plan"} needs an interactive terminal to approve on \u2014 none is attached (stdin/stderr are not a TTY). For an unattended run use --yes, or pre-authorize with rules in .agent/permissions.json.`);
@@ -17369,6 +17443,8 @@ Flags:
                        to a background worker agent (-m model); results are re-voiced when ready
   --duplex-v2          duplex + a second voice channel: backchannels, fillers and bridges that never
                        enter the agent's history (headless: stdin interim/pause/agent_state, stdout channel)
+  --channel-only       headless (--input-format stream-json): ONLY the second channel, for a host whose main
+                       answer comes from elsewhere (stdin user/text/barge/interim/pause/agent_state, stdout channel)
   --conversational     duplex with a conversation-native register \u2014 short fast turns, fillers,
                        impulsive reactions, human pacing (implies --duplex; aliases: --convo, --voice)
                        with SONIOX_API_KEY + CARTESIA_API_KEY(+VOICE_ID) set: real voice I/O \u2014 mic in,
@@ -20479,6 +20555,7 @@ ${out}
       onPartial: (text) => {
         if (text === voicePartial) return;
         voicePartial = text;
+        if (text.trim()) dx?.userActive();
         if (!partialRedraw) partialRedraw = setTimeout(() => {
           partialRedraw = null;
           editorRef?.redrawNow();
@@ -20611,6 +20688,7 @@ ${out}
     };
     try {
       await voiceIO.start();
+      if (dx) dx.options.holdResumeMs = HOLD_RESUME_MS;
       if (args.duplexV2 && dx) {
         const v = voiceIO;
         bindSecondChannel(v, makeSecondChannel(ai, dx.options.reflexModel, { show: (t) => {
@@ -20868,6 +20946,7 @@ function duplexError(e, reflexModel) {
   const code = e?.statusCode;
   return code && !msg.includes(String(code)) ? `${msg} (${code})` : msg;
 }
+var HOLD_RESUME_MS = 4e3;
 var leakDecider = () => {
   const o = voiceDeciderOptions();
   return o ? new DecideClient(o) : void 0;
@@ -20899,8 +20978,10 @@ async function duplexServe(args, ai, cfg, cwd) {
         sc?.onMainTokens(String(e.message ?? ""));
       } else if (k === "thinking_delta") out({ type: "thinking", text: e.message });
       else if (k === "speak_utterance") out({ type: "speak", text: e.message });
-      else if (k === "hold_filler") out({ type: "filler", text: e.message });
-      else if (k === "turn_replace") out({ type: "replaced", ...e.data });
+      else if (k === "hold_filler") {
+        out({ type: "filler", text: e.message });
+        sc?.onAgentState("idle");
+      } else if (k === "turn_replace") out({ type: "replaced", ...e.data });
       else if (k === "revoice_done") {
         out({ type: "revoice_done" });
         persist();
@@ -20926,7 +21007,9 @@ async function duplexServe(args, ai, cfg, cwd) {
     host,
     // --voice: the CLI voice register (progress asides, spoken budget). No emotion tags: the host's TTS may not speak them.
     ...args.voice ? { voiceStyle: "conversational", progressUpdates: true, spokenBudgetChars: 350 } : {},
-    reflexOptions: { fs: agent.options.fs, tools: [exitSessionTool(() => out({ type: "exit" }))] }
+    reflexOptions: { fs: agent.options.fs, tools: [exitSessionTool(() => out({ type: "exit" }))] },
+    ...args.duplexV2 ? { holdResumeMs: HOLD_RESUME_MS } : {}
+    // v2 hosts stream the user's partials (userActive) — v1 hosts don't
   });
   if (args.voiceSystemPrompt) dx.voice.options.systemPrompt += "\n\n" + args.voiceSystemPrompt;
   let scTurn = 0;
@@ -20954,9 +21037,11 @@ async function duplexServe(args, ai, cfg, cwd) {
       }
       if (m.type === "barge") {
         sc?.onBargeIn();
+        dx.userActive();
         activeTurn?.abort();
         dx.parkInFlightDeliveries();
       } else if (sc && channelInput(sc, m)) {
+        if (m.type === "interim" && String(m.text ?? "").trim()) dx.userActive();
       } else if (m.type === "cancel") out({ type: "task", kind: "cancel_result", id: m.task, message: dx.cancelTask(String(m.task)) });
       else if (m.type === "exit") rl.close();
       else if (m.type === "user" && typeof m.text === "string" && m.text.trim()) {
@@ -20989,6 +21074,28 @@ async function duplexServe(args, ai, cfg, cwd) {
   persist();
   await closeMcp(mounted);
 }
+async function channelServe(args, ai, cfg) {
+  const out = (o) => process.stdout.write(JSON.stringify(o) + "\n");
+  const { reflexModel, notes } = duplexModels({ voiceModel: args.voiceModel, reflexModel: cfg.reflexModel, thinkModel: false, actModel: String(args.model ?? cfg.model ?? ""), apiKeys: { ...cfg.apiKeys, ...apiKeysFromEnv() } });
+  const model = reflexModel ?? new DuplexAgentOptions().reflexModel;
+  const sc = makeSecondChannel(ai, model, { show: (text, kind) => out({ type: "channel", kind, text }) });
+  out({ type: "ready", model, ...notes.length ? { notes } : {} });
+  const rl = createInterface({ input: process.stdin });
+  rl.on("line", (line) => {
+    let m;
+    try {
+      m = JSON.parse(line);
+    } catch {
+      return;
+    }
+    if (channelInput(sc, m)) return;
+    if (m.type === "user" && typeof m.text === "string") sc.onUserTurn(m.text);
+    else if (m.type === "text" && typeof m.text === "string") sc.onMainTokens(m.text);
+    else if (m.type === "barge") sc.onBargeIn();
+    else if (m.type === "exit") rl.close();
+  });
+  await new Promise((r) => rl.on("close", () => r()));
+}
 function readStdinSteers() {
   const early = [];
   let sink, got;
@@ -21158,6 +21265,10 @@ async function main() {
     worktreeCleanup?.();
     process.exit(ok ? 0 : res?.finishReason === "needs_input" ? 2 : 1);
   }
+  if (args.inputFormat && args.channelOnly) {
+    await channelServe(args, ai, cfg);
+    process.exit(0);
+  }
   if (args.inputFormat) {
     await duplexServe(args, ai, cfg, cwd);
     worktreeCleanup?.();