@stackbone/cli 0.7.1 → 0.8.0

package/voice-worker/main.js

@@ -6,7 +6,7 @@ import {
   defineAgent,
   inference,
   telemetry as telemetry2,
-  voice as voice4
+  voice as voice5
 } from "@livekit/agents";
 import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-http";
 
@@ -24,11 +24,11 @@ var PATTERN = [
   [0, 900]
 ];
 var renderPattern = () => {
-  const total = PATTERN.reduce((n, [, ms]) => n + SAMPLE_RATE * ms / 1e3, 0);
+  const total = PATTERN.reduce((n, [, ms2]) => n + SAMPLE_RATE * ms2 / 1e3, 0);
   const pcm = new Int16Array(total);
   let offset = 0;
-  for (const [hz, ms] of PATTERN) {
-    const length = SAMPLE_RATE * ms / 1e3;
+  for (const [hz, ms2] of PATTERN) {
+    const length = SAMPLE_RATE * ms2 / 1e3;
     for (let i = 0; hz > 0 && i < length; i++) {
       const envelope = 0.5 - 0.5 * Math.cos(2 * Math.PI * i / length);
       const sample = Math.sin(2 * Math.PI * hz * i / SAMPLE_RATE) * envelope * VOLUME;
@@ -101,8 +101,11 @@ var loadVad = async () => {
 };
 
 // libs/runtime/voice-worker/src/worker.ts
-import { log as log4, voice as voice3 } from "@livekit/agents";
-import { BackgroundVoiceCancellation } from "@livekit/noise-cancellation-node";
+import { log as log4, voice as voice4 } from "@livekit/agents";
+import {
+  BackgroundVoiceCancellation,
+  TelephonyBackgroundVoiceCancellation
+} from "@livekit/noise-cancellation-node";
 import { ParticipantKind } from "@livekit/rtc-node";
 
 // libs/shared/validators/src/lib/common/json-object.ts
@@ -558,9 +561,22 @@ var ragCollectionNameParamSchema = z7.string().min(1).max(128).regex(STUDIO_RAG_
 var ragDistanceSchema = z7.enum(["cosine", "l2", "ip"]);
 var ragIndexKindSchema = z7.enum(["hnsw", "exact"]);
 var ragNumberFormatSchema = z7.enum(["eu", "us"]);
+var ragRangeFieldSchema = z7.string().min(1).refine((field) => !field.includes("."), {
+  message: "field must be a literal top-level metadata key without dots"
+});
+var ragExtractFieldSchema = z7.object({
+  name: z7.string().trim().pipe(ragRangeFieldSchema),
+  type: z7.enum(["number", "date", "text"]),
+  removed: z7.boolean().optional()
+});
+var uniqueFieldNames = (fields) => new Set(fields.map((field) => field.name)).size === fields.length;
+var ragExtractFieldsRequestSchema = z7.array(ragExtractFieldSchema).refine(uniqueFieldNames, {
+  message: "Field names must be unique"
+});
 var ragCollectionSchema = z7.object({
   name: z7.string(),
   dimensions: z7.number().int().min(1),
+  embedding_state: z7.enum(["ready", "rebuilding", "failed"]).optional(),
   // Optional for boxes that predate collection-owned embeddings.
   embedding_provider: z7.string().min(1).optional(),
   embedding_model: z7.string().min(1).optional(),
@@ -574,7 +590,9 @@ var ragCollectionSchema = z7.object({
   last_ingested_at: z7.iso.datetime().nullable(),
   // STA3-57. Optional so a list from a box older than hybrid search still parses.
   hybrid_enabled: z7.boolean().optional(),
-  number_format: ragNumberFormatSchema.optional()
+  number_format: ragNumberFormatSchema.optional(),
+  extract_enabled: z7.boolean().optional(),
+  extract_fields: z7.array(ragExtractFieldSchema).optional()
 });
 var ragCollectionListResponseSchema = z7.object({
   items: z7.array(ragCollectionSchema)
@@ -621,7 +639,7 @@ var ragDocumentSchema = z7.object({
   pending_version: z7.number().int().min(1).nullable().optional()
 });
 var ragFilterSchema = jsonObject();
-var ragDateBoundSchema = z7.string().length(10).regex(/^[0-9]{4}-[0-9]{2}-[0-9]{2}$/).refine((value) => {
+var ragDateValueSchema = z7.string().length(10).regex(/^[0-9]{4}-[0-9]{2}-[0-9]{2}$/).refine((value) => {
   const year = Number(value.slice(0, 4));
   const month = Number(value.slice(5, 7));
   const day = Number(value.slice(8, 10));
@@ -629,9 +647,6 @@ var ragDateBoundSchema = z7.string().length(10).regex(/^[0-9]{4}-[0-9]{2}-[0-9]{
   date.setUTCFullYear(year, month - 1, day);
   return year >= 1 && year <= 9999 && date.getUTCFullYear() === year && date.getUTCMonth() === month - 1 && date.getUTCDate() === day;
 }, "date must be a real YYYY-MM-DD day in years 0001\u20139999");
-var ragRangeFieldSchema = z7.string().min(1).refine((field) => !field.includes("."), {
-  message: "field must be a literal top-level metadata key without dots"
-});
 var ragRangeSchema = z7.discriminatedUnion("type", [
   z7.strictObject({
     field: ragRangeFieldSchema,
@@ -644,10 +659,10 @@ var ragRangeSchema = z7.discriminatedUnion("type", [
   z7.strictObject({
     field: ragRangeFieldSchema,
     type: z7.literal("date"),
-    gt: ragDateBoundSchema.optional(),
-    gte: ragDateBoundSchema.optional(),
-    lt: ragDateBoundSchema.optional(),
-    lte: ragDateBoundSchema.optional()
+    gt: ragDateValueSchema.optional(),
+    gte: ragDateValueSchema.optional(),
+    lt: ragDateValueSchema.optional(),
+    lte: ragDateValueSchema.optional()
   })
 ]).superRefine((range, ctx) => {
   const lower = range.gt ?? range.gte;
@@ -690,6 +705,32 @@ var ragDocumentVersionListResponseSchema = z7.object({
   items: z7.array(ragDocumentVersionSchema)
 });
 var ragDocumentVersionNumberSchema = z7.number().int().min(1).max(2147483647);
+var updateRagDocumentValuesRequestSchema = z7.strictObject({
+  version: ragDocumentVersionNumberSchema.optional(),
+  values: jsonObject(),
+  add_fields: z7.array(ragExtractFieldSchema.omit({ removed: true })).refine(uniqueFieldNames, {
+    message: "Field names must be unique"
+  }).optional()
+});
+var updateRagDocumentValuesResponseSchema = z7.object({
+  id: z7.string(),
+  collection: z7.string(),
+  version: ragDocumentVersionNumberSchema,
+  metadata: jsonObject()
+});
+var extractRagDocumentRequestSchema = z7.strictObject({
+  version: ragDocumentVersionNumberSchema.optional()
+});
+var extractRagDocumentResponseSchema = z7.object({
+  proposal: z7.array(
+    ragExtractFieldSchema.omit({ removed: true }).extend({
+      value: z7.union([z7.number().finite(), z7.string()]),
+      new: z7.boolean()
+    })
+  ),
+  failed: z7.boolean(),
+  truncated: z7.boolean()
+});
 var ragDocumentVersionParamSchema = z7.string().regex(/^\d+$/, { message: "version must be a positive integer" }).transform((value) => Number.parseInt(value, 10)).pipe(ragDocumentVersionNumberSchema);
 var activateRagDocumentVersionResponseSchema = z7.object({
   id: z7.string(),
@@ -721,6 +762,8 @@ var ragQueryHitSchema = z7.object({
   // The reciprocal rank fusion score a hybrid search ordered by (STA3-57 D10).
   // Absent when only the search by meaning ran, and from a box older than it.
   mixed_score: z7.number().optional(),
+  dense_rank: z7.number().int().positive().optional(),
+  lexical_rank: z7.number().int().positive().optional(),
   content: z7.string(),
   metadata: jsonObject()
 });
@@ -767,7 +810,9 @@ var createRagCollectionRequestSchema = z7.object({
   embedding_model: z7.string().trim().min(1).optional(),
   // Omitted, the box creates the collection with hybrid search off and `eu`.
   hybrid_enabled: z7.boolean().optional(),
-  number_format: ragNumberFormatSchema.optional()
+  number_format: ragNumberFormatSchema.optional(),
+  extract_enabled: z7.boolean().optional(),
+  extract_fields: ragExtractFieldsRequestSchema.optional()
 }).refine(
   (value) => value.embedding_provider === void 0 === (value.embedding_model === void 0),
   {
@@ -777,14 +822,19 @@ var createRagCollectionRequestSchema = z7.object({
 );
 var updateRagCollectionRequestSchema = z7.strictObject({
   hybrid_enabled: z7.boolean().optional(),
-  number_format: ragNumberFormatSchema.optional()
-}).refine((value) => value.hybrid_enabled !== void 0 || value.number_format !== void 0, {
-  message: "Send hybrid_enabled, number_format or both"
+  number_format: ragNumberFormatSchema.optional(),
+  extract_enabled: z7.boolean().optional(),
+  extract_fields: ragExtractFieldsRequestSchema.optional()
+}).refine((value) => Object.values(value).some((setting) => setting !== void 0), {
+  message: "Send at least one collection setting"
 });
 var updateRagCollectionResponseSchema = z7.object({
   name: z7.string(),
   hybrid_enabled: z7.boolean(),
-  number_format: ragNumberFormatSchema
+  number_format: ragNumberFormatSchema,
+  extract_enabled: z7.boolean().optional(),
+  extract_fields: z7.array(ragExtractFieldSchema).optional(),
+  unconverted: z7.number().int().nonnegative().optional()
 });
 var deleteRagCollectionResponseSchema = z7.object({
   name: z7.string(),
@@ -881,7 +931,8 @@ var ragSchema = z8.strictObject({
   minScore: z8.number().min(0).max(1),
   minWords: z8.int().min(0).max(20),
   topK: z8.int().min(1).max(10),
-  timeoutMs: z8.int().min(200).max(1e4)
+  timeoutMs: z8.int().min(200).max(1e4),
+  previousExchanges: z8.int().min(0).max(5)
 });
 var effectiveVoiceSettingsObjectSchema = z8.strictObject({
   /** BCP-47-ish tag the STT is hinted with and the TTS speaks (`es`, `en-US`), or `multi`. */
@@ -915,20 +966,11 @@ var effectiveVoiceSettingsObjectSchema = z8.strictObject({
    * dispatch to the worker leaves it out (`callVoiceSettingsSchema`).
    */
   recording: z8.boolean(),
-  /**
-   * Ask OpenRouter to route this agent's model calls to the provider with the
-   * highest throughput (`provider: { sort: 'throughput' }`) instead of its default
-   * price routing. Gemma 4 31B on price routing landed on a 13 tok/s provider (5 s
-   * a turn); throughput routing answered in about 1 s. A request field, not a
-   * `:nitro` model suffix, so the model id the box checks and prices stays the
-   * one written. The box strips it off any other gateway. The SDK applies it when
-   * it builds the graph, so the worker never reads it.
-   */
-  fastModelRouting: z8.boolean(),
   /** The box runs the search before the model's turn, so the worker never reads it. */
   rag: ragSchema,
   /**
-   * What the agent says when an input guardrail stops the caller's sentence (FDE-66):
+   * What the agent says when a guardrail stops the caller's sentence (FDE-66) or the
+   * agent's reply (STA3-59):
    * a policy message written for a chat reader would leak the rule if a TTS read it.
    * The box answers the turn itself, so the worker never reads it. Not nullable: a
    * blocked turn always says something.
@@ -963,7 +1005,6 @@ var checkCrossFieldRules = (value, ctx) => {
 var effectiveVoiceSettingsSchema = effectiveVoiceSettingsObjectSchema.superRefine(checkCrossFieldRules);
 var callVoiceSettingsObjectSchema = effectiveVoiceSettingsObjectSchema.omit({
   recording: true,
-  fastModelRouting: true,
   rag: true,
   refusalPhrase: true
 });
@@ -989,12 +1030,12 @@ var VOICE_SETTINGS_LEAF_PATHS = [
   "phoneLine",
   "expressive",
   "recording",
-  "fastModelRouting",
   "rag.collection",
   "rag.minScore",
   "rag.minWords",
   "rag.topK",
   "rag.timeoutMs",
+  "rag.previousExchanges",
   "refusalPhrase"
 ];
 var voiceSettingOriginSchema = z8.enum(["studio", "code", "default"]);
@@ -1058,6 +1099,7 @@ var VOICE_TRACES_PATH = "/voice/v1/traces";
 var VOICE_WORKER_AGENT_NAME = "stackbone-voice";
 var VOICE_WORKER_AGENT_NAME_ENV = "STACKBONE_VOICE_AGENT_NAME";
 var VOICE_WORKER_PORT_ENV = "STACKBONE_VOICE_WORKER_PORT";
+var VOICE_TURN_METRICS_MESSAGE = "voice turn metrics";
 var VOICE_CALL_TOKEN_MAX = 2048;
 var voiceDispatchSchema = z8.strictObject({
   agent: deepAgentNameSchema,
@@ -1122,15 +1164,34 @@ var voiceCallTranscriptSchema = z8.array(
 ).max(1e3).refine((items) => items.reduce((size, item) => size + item.text.length, 0) <= 2e5, {
   message: "A call transcript may contain at most 200000 characters."
 });
+var VOICE_CALL_TURNS_MAX = 1e3;
+var AGUI_RUN_ID_MAX = 200;
+var VOICE_TURN_LATENCY_MAX_MS = 36e5;
+var turnLatencyMs = z8.number().int().min(0).max(VOICE_TURN_LATENCY_MAX_MS).optional();
+var voiceCallTurnMetricsSchema = z8.strictObject({
+  /** The AG-UI `runId` of the box run that answered; absent on a reply the box did not run. */
+  runId: z8.string().min(1).max(AGUI_RUN_ID_MAX).optional(),
+  endOfTurnMs: turnLatencyMs,
+  transcriptionMs: turnLatencyMs,
+  llmTtftMs: turnLatencyMs,
+  ttsTtfbMs: turnLatencyMs,
+  e2eMs: turnLatencyMs
+});
+var voiceCallTurnsSchema = z8.array(voiceCallTurnMetricsSchema).max(VOICE_CALL_TURNS_MAX);
 var voiceCallCompleteSchema = z8.strictObject({
   transcript: voiceCallTranscriptSchema,
-  outcome: z8.enum(["completed", "disconnected", "failed"])
+  outcome: z8.enum(["completed", "disconnected", "failed"]),
+  // The transcript, the memory and the `onEnd` start this body carries must not be
+  // refused over a bad latency, so `turns` is any value here and read apart through
+  // `readVoiceCallTurns`. Not `.catch()` on the array: this schema is a documented
+  // route's body, and the OpenAPI generator cannot describe a catch (STA3-56).
+  turns: z8.unknown().optional()
 });
 var voiceCallLifecycleResponseSchema = z8.strictObject({ ok: z8.literal(true) });
 
 // libs/runtime/voice-worker/src/box-llm.ts
 import { randomUUID } from "node:crypto";
-import { DEFAULT_API_CONNECT_OPTIONS, llm, log } from "@livekit/agents";
+import { DEFAULT_API_CONNECT_OPTIONS, llm as llm2, log } from "@livekit/agents";
 
 // libs/shared/llm-wire/src/lib/sse/read-sse-data.ts
 async function* readSseData(body) {
@@ -1175,8 +1236,9 @@ var BoxTurnError = class extends Error {
   }
 };
 var boxEndpoint = (boxUrl, path) => `${boxUrl.replace(/\/+$/, "")}${path}`;
+var mintAguiRunId = () => `run-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
 async function* aguiTurn(opts) {
-  const runId = `run-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+  const runId = opts.runId ?? mintAguiRunId();
   const heard = "toolResult" in opts ? void 0 : opts.heard;
   const message = "toolResult" in opts ? {
     id: `tool-${opts.toolResult.toolCallId}`,
@@ -1456,8 +1518,45 @@ var lastNewAssistantMessage = (chatCtx, seen) => {
   return void 0;
 };
 
+// libs/runtime/voice-worker/src/turn-metrics.ts
+import { llm, voice as voice2 } from "@livekit/agents";
+var AGUI_RUN_ID_EXTRA = "stackboneAguiRunId";
+var ms = (seconds) => {
+  if (typeof seconds !== "number" || !Number.isFinite(seconds)) return void 0;
+  const value = Math.round(seconds * 1e3);
+  return value >= 0 && value <= VOICE_TURN_LATENCY_MAX_MS ? value : void 0;
+};
+var defined = (metrics) => Object.fromEntries(Object.entries(metrics).filter(([, value]) => value !== void 0));
+var watchTurnMetrics = (session, write, options = {}) => {
+  let caller;
+  let delivered;
+  session.on(voice2.AgentSessionEventTypes.ConversationItemAdded, ({ item }) => {
+    if (!(item instanceof llm.ChatMessage)) return;
+    if (item.role === "user") {
+      caller = item.metrics;
+      delivered = options.endOfTurnMs?.();
+      return;
+    }
+    if (item.role !== "assistant") return;
+    const reply = item.metrics;
+    const runId = item.extra[AGUI_RUN_ID_EXTRA];
+    const e2eMs = ms(reply.e2eLatency);
+    const metrics = defined({
+      endOfTurnMs: delivered ?? ms(caller?.endOfTurnDelay),
+      transcriptionMs: options.endOfTurnMs ? void 0 : ms(caller?.transcriptionDelay),
+      llmTtftMs: ms(reply.llmNodeTtft),
+      ttsTtfbMs: ms(reply.ttsNodeTtfb),
+      e2eMs: e2eMs === void 0 || delivered === void 0 ? e2eMs : e2eMs + delivered
+    });
+    caller = void 0;
+    delivered = void 0;
+    if (Object.keys(metrics).length === 0) return;
+    write(typeof runId === "string" ? { ...metrics, runId } : metrics);
+  });
+};
+
 // libs/runtime/voice-worker/src/box-llm.ts
-var BoxLLM = class extends llm.LLM {
+var BoxLLM = class extends llm2.LLM {
   #opts;
   #replies = new HeardTracker();
   #pending = /* @__PURE__ */ new Set();
@@ -1554,7 +1653,7 @@ var lastUserText = (chatCtx) => {
   }
   return "";
 };
-var BoxLLMStream = class extends llm.LLMStream {
+var BoxLLMStream = class extends llm2.LLMStream {
   #opts;
   #replies;
   #settle;
@@ -1584,6 +1683,7 @@ var BoxLLMStream = class extends llm.LLMStream {
       ...turn
     } = this.#opts;
     const id = randomUUID();
+    const runId = mintAguiRunId();
     const { signal } = this.abortController;
     await pendingResume?.();
     if (signal.aborted || isEnding?.()) return;
@@ -1608,7 +1708,8 @@ var BoxLLMStream = class extends llm.LLMStream {
         tools: clientTools,
         text: lastUserText(this.chatCtx),
         ...heard ? { heard } : {},
-        signal: turnSignal
+        signal: turnSignal,
+        runId
       })) {
         restartIdle();
         reply.observe(ev);
@@ -1619,7 +1720,7 @@ var BoxLLMStream = class extends llm.LLMStream {
               delta: {
                 role: "assistant",
                 content: ev.text,
-                extra: { [VOICE_RESPONSE_MARKER]: id }
+                extra: { [VOICE_RESPONSE_MARKER]: id, [AGUI_RUN_ID_EXTRA]: runId }
               }
             });
             break;
@@ -1666,6 +1767,49 @@ var BoxLLMStream = class extends llm.LLMStream {
   }
 };
 
+// libs/runtime/voice-worker/src/stt-delivery.ts
+import { stt } from "@livekit/agents";
+import { TransformStream } from "node:stream/web";
+var finalDeliveryMs = (event, audioSentSec) => {
+  const alt = event.alternatives?.[0];
+  if (alt === void 0 || !Number.isFinite(alt.endTime) || alt.endTime <= 0) return void 0;
+  const segmentEndSec = alt.endTime - (alt.words?.[0]?.startTimeOffset ?? 0);
+  const ms2 = Math.round((audioSentSec - segmentEndSec) * 1e3);
+  return ms2 >= 0 ? ms2 : void 0;
+};
+var createSttDeliveryMeter = () => {
+  let last;
+  return {
+    wrap: async (audio, inner) => {
+      let audioSentSec = 0;
+      async function* counted() {
+        for await (const frame of audio) {
+          audioSentSec += frame.samplesPerChannel / frame.sampleRate;
+          yield frame;
+        }
+      }
+      const events = await inner(counted());
+      if (events === null) return null;
+      return events.pipeThrough(
+        new TransformStream({
+          transform(event, controller) {
+            if (typeof event !== "string" && event.type === stt.SpeechEventType.FINAL_TRANSCRIPT) {
+              last = finalDeliveryMs(event, audioSentSec) ?? last;
+            }
+            controller.enqueue(event);
+          }
+        })
+      );
+    },
+    take: () => {
+      const taken = last;
+      last = void 0;
+      return taken;
+    }
+  };
+};
+var deliversItsOwnEndOfTurn = (settings) => settings.turn.mode === "stt" && settings.sttModel.startsWith("deepgram/flux");
+
 // libs/runtime/voice-worker/src/dispatch.ts
 var readDispatch = (metadata) => {
   if (!metadata) return null;
@@ -1742,17 +1886,17 @@ var createHangUp = ({
 };
 
 // libs/runtime/voice-worker/src/session-options.ts
-import { tts, voice as voice2 } from "@livekit/agents";
+import { tts, voice as voice3 } from "@livekit/agents";
 
 // libs/runtime/voice-worker/src/number-text.ts
-import { TransformStream } from "node:stream/web";
+import { TransformStream as TransformStream2 } from "node:stream/web";
 var THOUSANDS_COMMA = /(\d),(?=\d{3}(?!\d))/g;
 var TRAILING_NUMBER = /[\d.,]+$/;
 var rewrite = (text) => text.replace(THOUSANDS_COMMA, "$1");
 var dropThousandsComma = (input) => {
   let pending = "";
   return input.pipeThrough(
-    new TransformStream({
+    new TransformStream2({
       transform(chunk, controller) {
         const text = pending + chunk;
         const tail = TRAILING_NUMBER.exec(text)?.[0] ?? "";
@@ -1769,7 +1913,7 @@ var dropThousandsComma = (input) => {
 
 // libs/runtime/voice-worker/src/phone-voice.ts
 import { AudioFrame as AudioFrame2 } from "@livekit/rtc-node";
-import { TransformStream as TransformStream2 } from "node:stream/web";
+import { TransformStream as TransformStream3 } from "node:stream/web";
 var LOW_CUT_HZ = 300;
 var HIGH_CUT_HZ = 3400;
 var Q = Math.SQRT1_2;
@@ -1805,7 +1949,7 @@ var phoneLine = (audio) => {
   let chainRate = 0;
   const norm = Math.tanh(DRIVE);
   return audio.pipeThrough(
-    new TransformStream2({
+    new TransformStream3({
       transform(frame, controller) {
         if (!chain || chainRate !== frame.sampleRate) {
           chainRate = frame.sampleRate;
@@ -1858,7 +2002,16 @@ var markupContext = (expressive, speech) => {
   });
   return guide === void 0 ? [] : [{ description: VOICE_TTS_MARKUP_CONTEXT, value: guide }];
 };
-var PhoneLineAgent = class extends voice2.Agent {
+var VoiceCallAgent = class extends voice3.Agent {
+  constructor() {
+    super(...arguments);
+    this.sttDelivery = createSttDeliveryMeter();
+  }
+  async sttNode(audio, modelSettings) {
+    return this.sttDelivery.wrap(audio, (counted) => super.sttNode(counted, modelSettings));
+  }
+};
+var PhoneLineAgent = class extends VoiceCallAgent {
   async ttsNode(text, modelSettings) {
     const audio = await super.ttsNode(text, modelSettings);
     return audio && phoneLine(audio);
@@ -1912,7 +2065,7 @@ var voiceSessionOptions = (settings, pieces) => {
       }
     },
     // The box holds the prompt: LiveKit's own chat context stays prompt-less.
-    agent: settings.phoneLine ? new PhoneLineAgent({ instructions: "" }) : new voice2.Agent({ instructions: "" }),
+    agent: settings.phoneLine ? new PhoneLineAgent({ instructions: "" }) : new VoiceCallAgent({ instructions: "" }),
     boxContext: markupContext(settings.expressive, speech)
   };
 };
@@ -2011,9 +2164,9 @@ var toolCallCue = () => {
   };
 };
 var messageOf = (err) => err instanceof Error ? err.message : String(err);
-var callerNoiseCancellation = () => {
+var callerNoiseCancellation = (caller) => {
   try {
-    return BackgroundVoiceCancellation();
+    return caller.kind === ParticipantKind.SIP ? TelephonyBackgroundVoiceCancellation() : BackgroundVoiceCancellation();
   } catch (err) {
     log4().warn({ error: messageOf(err) }, "voice worker: no noise cancellation on this call");
     return void 0;
@@ -2147,6 +2300,7 @@ var runVoiceJob = async (ctx, deps) => {
     onRunFinished
   });
   const session = deps.createSession({ ...plan.session, llm: brain });
+  const turns = [];
   let completion;
   const finishCall = (outcome = "disconnected") => {
     if (completion) return completion;
@@ -2155,14 +2309,20 @@ var runVoiceJob = async (ctx, deps) => {
       await hangUp.pendingResume();
       await session.close();
       await brain.settled();
-      await voiceCallRequest(
-        wire,
-        "complete",
-        voiceCallCompleteSchema.parse({
-          transcript: brain.heardTranscript(session.history),
-          outcome
-        })
-      );
+      const transcript = brain.heardTranscript(session.history);
+      const complete = (body) => voiceCallRequest(wire, "complete", voiceCallCompleteSchema.parse(body));
+      try {
+        await complete({ transcript, outcome, ...turns.length > 0 ? { turns } : {} });
+      } catch (error) {
+        if (!(turns.length > 0 && error instanceof BoxTurnError && error.status === 400)) {
+          throw error;
+        }
+        log4().warn(
+          { agent: dispatch.agent, sessionKey: dispatch.sessionKey, status: error.status },
+          "voice worker: the box refused the turn metrics, completing without them"
+        );
+        await complete({ transcript, outcome });
+      }
     }).catch((error) => {
       log4().warn(
         { agent: dispatch.agent, sessionKey: dispatch.sessionKey, reason: "completion_failed" },
@@ -2173,11 +2333,22 @@ var runVoiceJob = async (ctx, deps) => {
     return completion;
   };
   ctx.addShutdownCallback(() => finishCall());
-  session.on(voice3.AgentSessionEventTypes.Close, () => {
+  session.on(voice4.AgentSessionEventTypes.Close, () => {
     if (completion || ending) return;
     void finishCall().catch(() => void 0).finally(() => ctx.shutdown("caller left"));
   });
-  const noiseCancellation = callerNoiseCancellation();
+  watchTurnMetrics(
+    session,
+    (metrics) => {
+      if (turns.length < VOICE_CALL_TURNS_MAX) turns.push(metrics);
+      log4().info(
+        { agent: dispatch.agent, sessionKey: dispatch.sessionKey, room: ctx.room.name, ...metrics },
+        VOICE_TURN_METRICS_MESSAGE
+      );
+    },
+    deliversItsOwnEndOfTurn(settings) ? { endOfTurnMs: () => plan.agent.sttDelivery.take() } : {}
+  );
+  const noiseCancellation = callerNoiseCancellation(participant);
   await session.start({
     agent: plan.agent,
     room: ctx.room,
@@ -2191,7 +2362,7 @@ var runVoiceJob = async (ctx, deps) => {
 
 // libs/runtime/voice-worker/src/main.ts
 var liveDeps = () => ({
-  createSession: (options) => new voice4.AgentSession(options),
+  createSession: (options) => new voice5.AgentSession(options),
   createStt: (options) => new inference.STT(options),
   createTts: (options) => new inference.TTS(options),
   createTurnDetector: (options) => new inference.TurnDetector(options),