@livx.cc/agentx 0.99.98 → 0.99.100

package/dist/cli.js

@@ -3432,7 +3432,7 @@ function copyTextToClipboard(text, platform2 = process.platform) {
 
 // cli/cli.ts
 import { join as join16, resolve as resolve5, basename as basename4, extname, dirname as dirname6 } from "path";
-import { AIClient, DecideClient, listModels, listProviders, getProviderFromModel, getModelInfo, resolveModel, isModelSupported, disposeCursorSessions, disposeClaudeCodeSessions } from "ai.libx.js";
+import { AIClient as AIClient2, listModels, listProviders, getProviderFromModel, getModelInfo, resolveModel, isModelSupported, disposeCursorSessions, disposeClaudeCodeSessions } from "ai.libx.js";
 
 // src/ccBridge.ts
 var CC_BRIDGED = /* @__PURE__ */ new Set(["Recall", "Remember", "MemorySearch"]);
@@ -9251,12 +9251,20 @@ configureLogging((name) => {
 
 // src/voice/instinct.ts
 init_logging();
+import { AIClient, DecideClient, JevAdapter, LlmDecideAdapter } from "ai.libx.js";
 var log19 = forComponent("Instinct");
 function voiceDeciderOptions(env = process.env) {
   if (env.TYPESAFE_AI_KEY) return { apiKey: env.TYPESAFE_AI_KEY };
   if (env.GROQ_API_KEY) return { provider: "llm", model: "groq/openai/gpt-oss-120b", apiKey: env.GROQ_API_KEY };
   return void 0;
 }
+function voiceDecider(env = process.env, timeoutMs = 1500) {
+  const o = voiceDeciderOptions(env);
+  if (!o) return void 0;
+  const ai = new AIClient();
+  const llm = new LlmDecideAdapter({ client: { chat: (c) => ai.chat({ reasoningEffort: "low", ...c }) } });
+  return new DecideClient({ ...o, timeoutMs, adapters: { typesafe: new JevAdapter(), llm } });
+}
 var InstinctOptions = class {
   decider;
   /** Partial unchanged this long (a micro-pause) → consult. */
@@ -14065,7 +14073,7 @@ Reply JSON only: {"turn":"...","density":"...","cues":[{"when":["phrase",...],"l
 var log31 = forComponent("SecondChannel");
 var SecondChannelOptions = class {
   /** Refiner (Jev/fast LLM): the choice over CHANNEL_ACTIONS, or null on failure. Absent → the rule alone decides.
-   *  `state.rule` is the kind the rule proposed (fastBackchannelDecider answers only 'backchannel' moments). */
+   *  `state.rule` is the kind the rule proposed (fastChannelDecider answers yes/no on it). `signal` aborts a late call. */
   decide;
   /** Fast-model line writer (see fastLineGenerator). Absent or late → the canned bank. */
   generate;
@@ -14094,8 +14102,12 @@ var SecondChannelOptions = class {
   decisionBudgetMs = 300;
   generateBudgetMs = 500;
   minGapMs = 2500;
-  /** Backchannel after backchannel needs this gap (v1 Instinct's cadence, ≤12/min). */
-  backchannelGapMs = 5e3;
+  /** 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;
+  /** 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;
   /** A decision whose partial grew by this many words while in flight is stale → dropped. */
   staleWords = 3;
   microWords = 30;
@@ -14137,11 +14149,12 @@ var SecondChannelOptions = class {
   random = Math.random;
 };
 var words2 = (s) => s.trim() ? s.trim().split(/\s+/) : [];
-function withBudget(p, ms, what) {
+function withBudget(p, ms, what, ac) {
   let t;
   const late = new Promise((r) => {
     t = setTimeout(() => {
       log31.debug(`${what}: over the ${ms}ms budget \u2014 dropped`);
+      ac?.abort();
       r(null);
     }, ms);
   });
@@ -14182,6 +14195,11 @@ var SecondChannel = class {
   inflight = null;
   startedAt;
   lang = null;
+  carried = 0;
+  // words of this statement already dispatched (a monologue split by STT endpoints)
+  floorWords = 0;
+  // carried + the last dispatched line's words
+  lastHeardAt = -Infinity;
   pinned = false;
   // anticipate: the plan for the user's current/next turn (turnNo bumps per dispatched turn → a late plan is dropped)
   turnNo = 0;
@@ -14229,6 +14247,8 @@ var SecondChannel = class {
   onInterim(text) {
     const t = text.trim();
     if (!t || t === this.heard) return;
+    if (this.main !== "idle" || !this.heard) this.carried = this.o.now() - this.lastHeardAt < this.o.continueMs ? this.floorWords : 0;
+    this.lastHeardAt = this.o.now();
     if (this.main !== "idle") this.onAgentState("idle");
     this.heard = t;
     this.detect(t);
@@ -14310,6 +14330,7 @@ var SecondChannel = class {
   /** The user's final line was dispatched to the main agent (its words are the bridge's micro-context). */
   onUserTurn(text) {
     this.heard = text.trim();
+    this.floorWords = this.carried + words2(this.heard).length;
     this.lastUser = this.heard;
     this.detect(this.heard);
     this.main = "idle";
@@ -14391,8 +14412,10 @@ var SecondChannel = class {
    *  sentence end only deep into a story — a short request's end is the end of their turn (replay: "mm-hm" right after
    *  "Check hacker news for recent news.", just before the final). */
   backchannelMoment(n, breath) {
-    if (n < this.o.backchannelMinWords || /\?$/.test(this.heard) || pendingQuestion(this.heard)) return false;
-    if (n < this.o.requestHoldWords && requestOpener(this.heard)) return false;
+    if (/\?$/.test(this.heard) || pendingQuestion(this.heard)) return false;
+    if (this.carried) n += this.carried;
+    else if (n < this.o.requestHoldWords && requestOpener(this.heard)) return false;
+    if (n < this.o.backchannelMinWords) return false;
     if (/[.!]$/.test(this.heard)) return n >= 2 * this.o.backchannelMinWords;
     return breath || /[,;:…]$/.test(this.heard);
   }
@@ -14411,7 +14434,8 @@ var SecondChannel = class {
       let kind = rule;
       let choice;
       if (this.o.decide && !cue) {
-        const d = await withBudget(this.o.decide(this.tail(), { ...this.state(pauseMs), rule }), this.o.decisionBudgetMs, "channel decision");
+        const ac = new AbortController();
+        const d = await withBudget(this.o.decide(this.tail(), { ...this.state(pauseMs), rule }, ac.signal), this.o.decisionBudgetMs, "channel decision", ac);
         choice = d?.choice;
         if (choice && choice in CHANNEL_ACTIONS && (choice === "silence" || this.allowed(choice))) kind = choice;
         else if (choice && choice in CHANNEL_ACTIONS) log31.debug(`decider chose ${choice} for a ${rule} moment \u2014 the rule stands`);
@@ -14491,22 +14515,28 @@ var SecondChannel = class {
     return true;
   }
 };
-function fastBackchannelDecider(ai, model = "groq/qwen/qwen3.8-27b") {
-  return async (heard, state) => {
-    if (state.rule !== "backchannel") return null;
-    const prompt = `Live voice call. The user is still talking and just took a breath. So far: "${heard}"
-Would an attentive human listener say a tiny "mm-hm" now? Yes ONLY if they are telling a story/explaining across several clauses. No if it is a request, command, question, or they are mid-request. Answer yes or no.`;
+var FAST_ASK = {
+  backchannel: 'The user is still talking and just took a breath. Would an attentive human listener say a tiny "mm-hm" now? Yes ONLY if they are telling a story/explaining across several clauses. No if it is a request, command, question, or they are mid-request.',
+  filler: 'The user went quiet mid-turn. Would a human listener make a soft "hmm\u2026" now? Yes ONLY if their sentence is unfinished and they are searching for words. No if they said a complete request, question, answer, thanks or remark (the reply comes next).',
+  bridge: `The user finished and the assistant's answer is taking a moment. Would a human say a short "let me check\u2014" now? Yes if they asked for real work or a substantial/hard question. No for small talk, a greeting, thanks, an ok/yes/agreement.`
+};
+function fastChannelDecider(ai, model = "groq/qwen/qwen3.8-27b") {
+  return async (heard, state, signal) => {
+    const kind = state.rule;
+    if (!(kind in FAST_ASK)) return null;
+    const prompt = `Live voice call. ${FAST_ASK[kind]}
+So far the user said: "${heard}"
+Answer yes or no.`;
     const reasoning = /qwen/.test(model) ? { providerOptions: { reasoning_effort: "none" } } : /gpt-oss/.test(model) ? { providerOptions: { reasoning_effort: "low" } } : {};
-    const r = await ai.chat({ model, messages: [{ role: "user", content: prompt }], stream: false, maxTokens: /gpt-oss/.test(model) ? 200 : 3, ...reasoning });
+    const r = await ai.chat({ model, messages: [{ role: "user", content: prompt }], stream: false, signal, temperature: 0, maxTokens: /gpt-oss/.test(model) ? 200 : 3, ...reasoning });
     const a = (r.content ?? "").trim().toLowerCase();
-    return /^yes/.test(a) ? { choice: "backchannel" } : /^no/.test(a) ? { choice: "silence" } : null;
+    return /^yes/.test(a) ? { choice: kind } : /^no/.test(a) ? { choice: "silence" } : null;
   };
 }
 function channelDecide(ai, instinct, env = process.env) {
   const model = env.AGENTX_CHANNEL_DECIDER_MODEL || void 0;
-  const bc = model || env.GROQ_API_KEY ? fastBackchannelDecider(ai, model) : void 0;
-  if (!bc && !instinct) return void 0;
-  return async (heard, state) => bc && state.rule === "backchannel" ? bc(heard, state) : instinct ? instinct.channel(heard, state) : null;
+  if (model || env.GROQ_API_KEY) return fastChannelDecider(ai, model);
+  return instinct ? (heard, state) => instinct.channel(heard, state) : void 0;
 }
 var KIND_HINT2 = {
   backchannel: 'a tiny listener cue while they keep talking ("mm-hm", "right", "yeah") \u2014 shows you follow, never takes the turn',
@@ -21266,10 +21296,7 @@ function duplexError(e, reflexModel) {
   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;
-};
+var leakDecider = () => voiceDecider();
 function makeSecondChannel(ai, reflexModel, opts) {
   const decider = leakDecider();
   const instinct = decider ? new Instinct({ decider }) : void 0;
@@ -21536,7 +21563,7 @@ async function main() {
     console.error(red("No provider key found. Set ANTHROPIC_API_KEY (or OPENAI_API_KEY / GOOGLE_API_KEY / GROQ_API_KEY) in ~/.agent/.env, or set BODIFY_API_KEY + BODIFY_APP_ID to pull from Bodify."));
     process.exit(1);
   }
-  const ai = new AIClient({
+  const ai = new AIClient2({
     apiKeys,
     ...cfg.baseUrls ? { baseUrls: cfg.baseUrls } : {},
     retry: {