@voicelayer/sdk 0.6.2 → 0.6.3

package/dist/index.js

@@ -14,11 +14,11 @@ import { LoggerProvider, BatchLogRecordProcessor } from '@opentelemetry/sdk-logs
 import { PeriodicExportingMetricReader } from '@opentelemetry/sdk-metrics';
 import { NodeSDK } from '@opentelemetry/sdk-node';
 import { ATTR_SERVICE_VERSION, ATTR_SERVICE_NAME } from '@opentelemetry/semantic-conventions';
+import { llm, DEFAULT_API_CONNECT_OPTIONS, intervalForRetry, voice as voice$1, APIStatusError, APITimeoutError, APIConnectionError } from '@livekit/agents';
 import http from 'http';
 import https from 'https';
 import { Readable } from 'stream';
 import { lookup } from 'dns/promises';
-import { voice as voice$1, llm } from '@livekit/agents';
 import { getQuickJS, shouldInterruptAfterDeadline } from 'quickjs-emscripten';
 import { randomUUID, createHash } from 'crypto';
 import { setup, createActor } from 'xstate';
@@ -4596,32 +4596,116 @@ function acceptsReasoningEffort(model2) {
   const id = baseModelId(model2);
   return isReasoningModel(model2) && !/^o1-(mini|preview)/.test(id) && !/-chat(-|$)/.test(id);
 }
+function acceptsNoReasoningEffort(model2) {
+  if (!acceptsReasoningEffort(model2))
+    return false;
+  const id = baseModelId(model2);
+  if (/-pro\b/.test(id))
+    return false;
+  const gpt = /^gpt-(\d+)(?:\.(\d+))?/.exec(id);
+  if (!gpt)
+    return false;
+  const major = Number(gpt[1]);
+  const minor = gpt[2] !== void 0 ? Number(gpt[2]) : 0;
+  return major > 5 || major === 5 && minor >= 1;
+}
+function forgetLearnedReasoningEfforts() {
+  learnedEfforts.clear();
+}
+function chatReasoningEffort(model2, opts) {
+  const learned = learnedEfforts.get(learnedKey(model2, opts.tools === true));
+  if (learned !== void 0)
+    return learned === "omit" ? void 0 : learned;
+  if (!acceptsReasoningEffort(model2))
+    return void 0;
+  if (opts.tools === true && acceptsNoReasoningEffort(model2))
+    return "none";
+  return opts.requested;
+}
 function chatCompletionParams(model2, input) {
+  const effort = chatReasoningEffort(model2, {
+    ...input.tools !== void 0 ? { tools: input.tools } : {},
+    ...input.reasoningEffort !== void 0 ? { requested: input.reasoningEffort } : {}
+  });
   if (isReasoningModel(model2)) {
     return {
       ...input.maxTokens !== void 0 ? { max_completion_tokens: Math.max(input.maxTokens, REASONING_MIN_COMPLETION_TOKENS) } : {},
-      ...input.reasoningEffort !== void 0 && acceptsReasoningEffort(model2) ? { reasoning_effort: input.reasoningEffort } : {}
+      ...effort !== void 0 ? { reasoning_effort: effort } : {}
     };
   }
   return {
+    // only when a provider taught the process that this model (one we don't know as a reasoning model) takes an effort
+    ...effort !== void 0 ? { reasoning_effort: effort } : {},
     ...input.maxTokens !== void 0 ? { max_tokens: input.maxTokens } : {},
     ...input.temperature !== void 0 ? { temperature: input.temperature } : {},
     ...input.topP !== void 0 ? { top_p: input.topP } : {}
   };
 }
-function modelRejectionOf(err) {
+function acceptsSamplingParams(model2) {
+  return !isReasoningModel(model2);
+}
+function providerErrorOf(err) {
+  if (err === null || typeof err !== "object")
+    return null;
   const e = err;
-  const status = typeof e?.status === "number" ? e.status : null;
-  if (status !== 400 && status !== 403 && status !== 404)
+  const body = e["body"] !== null && typeof e["body"] === "object" ? e["body"] : {};
+  const pick = (k) => e[k] !== void 0 && e[k] !== null ? e[k] : body[k];
+  const statusRaw = typeof e["status"] === "number" ? e["status"] : e["statusCode"];
+  const status = typeof statusRaw === "number" && statusRaw > 0 ? statusRaw : null;
+  const code = pick("code");
+  const type = pick("type");
+  const param = pick("param");
+  const message = typeof body["message"] === "string" ? body["message"] : typeof e["message"] === "string" ? e["message"] : "";
+  return {
+    status,
+    code: typeof code === "string" ? code : typeof type === "string" ? type : "",
+    param: typeof param === "string" ? param : null,
+    message
+  };
+}
+function reasoningEffortRejectionOf(err) {
+  const f = providerErrorOf(err);
+  if (!f || f.status !== 400)
+    return null;
+  if (f.param !== "reasoning_effort" && !/reasoning[_ ]effort/i.test(f.message))
+    return null;
+  return { status: f.status, message: f.message, suggestsNone: /reasoning[_ ]effort[^.]*\bnone\b/i.test(f.message) };
+}
+function learnReasoningEffort(model2, shape, err) {
+  const rejection = reasoningEffortRejectionOf(err);
+  if (!rejection)
+    return false;
+  const next = shape.sent !== "none" && (rejection.suggestsNone || shape.sent === void 0) ? "none" : "omit";
+  if ((next === "omit" ? void 0 : next) === shape.sent)
+    return false;
+  learnedEfforts.set(learnedKey(model2, shape.tools), next);
+  return true;
+}
+async function withReasoningEffortRetry(shape, call) {
+  const sent = chatReasoningEffort(shape.model, {
+    tools: shape.tools,
+    ...shape.requested !== void 0 ? { requested: shape.requested } : {}
+  });
+  try {
+    return await call();
+  } catch (err) {
+    if (!learnReasoningEffort(shape.model, { tools: shape.tools, sent }, err))
+      throw err;
+    return call();
+  }
+}
+function modelRejectionOf(err) {
+  const f = providerErrorOf(err);
+  const status = f?.status ?? null;
+  if (!f || status === null || status !== 400 && status !== 403 && status !== 404)
     return null;
-  const code = typeof e?.code === "string" ? e.code : typeof e?.type === "string" ? e.type : "";
-  const param = typeof e?.param === "string" ? e.param : null;
-  const rejected = REJECTION_CODES.has(code) || param !== null && MODEL_PARAMS.has(param);
+  const { code, param } = f;
+  const rejected = REJECTION_CODES.has(code) || param !== null && MODEL_PARAMS.has(param) || reasoningEffortRejectionOf(err) !== null;
   if (!rejected)
     return null;
   if (status === 403 && code !== "model_not_found")
     return null;
-  return { status, code: code || "invalid_request_error", message: typeof e?.message === "string" ? e.message : "" };
+  return { status, code: code || "invalid_request_error", message: f.message };
 }
 async function withModelFallback(opts) {
   const fallbackModel = opts.fallbackModel ?? PLATFORM_DEFAULT_CHAT_MODEL;
@@ -4642,11 +4726,13 @@ async function withModelFallback(opts) {
     return opts.call(fallbackModel);
   }
 }
-var PLATFORM_DEFAULT_CHAT_MODEL, REASONING_MIN_COMPLETION_TOKENS, REJECTION_CODES, MODEL_PARAMS, MODEL_REJECTION_TTL_MS, ModelRejectionCache;
+var PLATFORM_DEFAULT_CHAT_MODEL, REASONING_MIN_COMPLETION_TOKENS, learnedEfforts, learnedKey, REJECTION_CODES, MODEL_PARAMS, MODEL_REJECTION_TTL_MS, ModelRejectionCache;
 var init_chat_params = __esm({
   "../llm-client/dist/chat-params.js"() {
     PLATFORM_DEFAULT_CHAT_MODEL = "gpt-4o-mini";
     REASONING_MIN_COMPLETION_TOKENS = 2048;
+    learnedEfforts = /* @__PURE__ */ new Map();
+    learnedKey = (model2, tools) => `${baseModelId(model2)}|${tools ? "tools" : "plain"}`;
     REJECTION_CODES = /* @__PURE__ */ new Set(["unsupported_parameter", "unsupported_value", "model_not_found"]);
     MODEL_PARAMS = /* @__PURE__ */ new Set(["model", "max_tokens", "max_completion_tokens", "temperature", "top_p", "reasoning_effort"]);
     MODEL_REJECTION_TTL_MS = 5 * 60 * 1e3;
@@ -4674,6 +4760,30 @@ var init_chat_params = __esm({
     };
   }
 });
+
+// ../llm-client/dist/index.js
+var dist_exports = {};
+__export(dist_exports, {
+  MODEL_REJECTION_TTL_MS: () => MODEL_REJECTION_TTL_MS,
+  ModelRejectionCache: () => ModelRejectionCache,
+  PLATFORM_DEFAULT_CHAT_MODEL: () => PLATFORM_DEFAULT_CHAT_MODEL,
+  REASONING_MIN_COMPLETION_TOKENS: () => REASONING_MIN_COMPLETION_TOKENS,
+  acceptsNoReasoningEffort: () => acceptsNoReasoningEffort,
+  acceptsReasoningEffort: () => acceptsReasoningEffort,
+  acceptsSamplingParams: () => acceptsSamplingParams,
+  chatCompletionParams: () => chatCompletionParams,
+  chatReasoningEffort: () => chatReasoningEffort,
+  createChatClient: () => createChatClient,
+  forgetLearnedReasoningEfforts: () => forgetLearnedReasoningEfforts,
+  hasChatKey: () => hasChatKey,
+  isReasoningModel: () => isReasoningModel,
+  learnReasoningEffort: () => learnReasoningEffort,
+  modelRejectionOf: () => modelRejectionOf,
+  providerErrorOf: () => providerErrorOf,
+  reasoningEffortRejectionOf: () => reasoningEffortRejectionOf,
+  withModelFallback: () => withModelFallback,
+  withReasoningEffortRetry: () => withReasoningEffortRetry
+});
 function createChatClient(config = {}) {
   return new OpenAI({
     ...config.apiKey ? { apiKey: config.apiKey } : {},
@@ -4681,6 +4791,9 @@ function createChatClient(config = {}) {
     ...config.timeoutMs ? { timeout: config.timeoutMs } : {}
   });
 }
+function hasChatKey(config = {}) {
+  return Boolean(config.apiKey || process.env["OPENAI_API_KEY"]);
+}
 var init_dist = __esm({
   "../llm-client/dist/index.js"() {
     init_chat_params();
@@ -5139,512 +5252,213 @@ var init_wrap = __esm({
   "src/providers/wrap.ts"() {
   }
 });
-
-// src/providers/index.ts
-async function importOptional(spec, hint) {
-  try {
-    return await import(spec);
-  } catch {
-    throw new Error(`${hint}: "${spec}" is not installed. Add it with: pnpm add ${spec}`);
+function metricExportIntervalMs() {
+  const raw = Number(process.env.OTEL_METRIC_EXPORT_INTERVAL);
+  return Number.isFinite(raw) && raw > 0 ? raw : DEFAULT_METRIC_EXPORT_INTERVAL_MS;
+}
+function start(opts) {
+  if (started)
+    return;
+  started = true;
+  if (opts.debug) {
+    diag.setLogger(new DiagConsoleLogger(), DiagLogLevel.INFO);
+  }
+  const authKey = process.env.VOICELAYER_TELEMETRY_KEY || process.env.VOICELAYER_API_KEY;
+  if (!authKey) {
+    return;
+  }
+  if (!process.env.OTEL_EXPORTER_OTLP_ENDPOINT) {
+    process.env.OTEL_EXPORTER_OTLP_ENDPOINT = DEFAULT_OTLP_ENDPOINT;
   }
+  process.env.OTEL_EXPORTER_OTLP_HEADERS = `Authorization=Bearer ${authKey}`;
+  const resource = resourceFromAttributes({
+    [ATTR_SERVICE_NAME]: opts.service,
+    ...opts.version ? { [ATTR_SERVICE_VERSION]: opts.version } : {}
+  });
+  loggerProvider = new LoggerProvider({
+    resource,
+    processors: [new BatchLogRecordProcessor(new OTLPLogExporter())]
+  });
+  logs.setGlobalLoggerProvider(loggerProvider);
+  metricReader = new PeriodicExportingMetricReader({
+    exportIntervalMillis: metricExportIntervalMs(),
+    // DELTA temporality: each export carries only the increment since the
+    // previous one. The read side (apps/api openobserve-repository) totals a
+    // metric with SUM(value) over the query window, which is correct only for
+    // deltas. The OTLP exporter defaults to CUMULATIVE — it re-reports the
+    // running total every export, so SUM double-counts every call that lives
+    // longer than one export interval (inflating LLM/TTS/STT usage + cost).
+    // Setting DELTA makes the existing SUM reads correct without touching them.
+    exporter: new OTLPMetricExporter({
+      temporalityPreference: AggregationTemporalityPreference.DELTA
+    })
+  });
+  sdk = new NodeSDK({
+    resource,
+    traceExporter: new OTLPTraceExporter(),
+    metricReader,
+    instrumentations: [
+      getNodeAutoInstrumentations({
+        // fs instrumentation is extremely noisy and rarely useful for app traces.
+        "@opentelemetry/instrumentation-fs": { enabled: false }
+      })
+    ]
+  });
+  sdk.start();
+  const shutdown = async () => {
+    try {
+      await sdk?.shutdown();
+      await loggerProvider?.shutdown();
+    } catch {
+    }
+  };
+  process.once("SIGTERM", () => void shutdown().then(() => process.exit(0)));
+  process.once("SIGINT", () => void shutdown().then(() => process.exit(0)));
 }
-function withCreds(opts, creds2) {
-  if (creds2?.apiKey) opts["apiKey"] = creds2.apiKey;
-  if (creds2?.baseURL) opts["baseURL"] = creds2.baseURL;
-  return opts;
+async function flushTelemetry() {
+  const flushes = [];
+  if (metricReader)
+    flushes.push(metricReader.forceFlush());
+  if (loggerProvider)
+    flushes.push(loggerProvider.forceFlush());
+  const proxied = trace.getTracerProvider();
+  const tracerProvider = proxied.getDelegate?.() ?? proxied;
+  if (typeof tracerProvider.forceFlush === "function") {
+    flushes.push(tracerProvider.forceFlush());
+  }
+  await Promise.allSettled(flushes);
 }
-var deepgram, openai, anthropic, cartesia, elevenlabs, assemblyai, google, silero, livekitTurn, connector;
-var init_providers2 = __esm({
-  "src/providers/index.ts"() {
-    init_ssrf();
-    init_transport_callback();
-    init_connector_llm();
-    init_wrap();
-    deepgram = {
-      stt(options = {}) {
-        return async () => {
-          const dg = await importOptional("@voicelayer/agents-plugin-deepgram", "deepgram.stt");
-          return new dg.STT(
-            withCreds(
-              { model: options.model ?? "nova-3", language: options.language ?? "en-US" },
-              options
-            )
-          );
-        };
-      },
-      tts(options = {}) {
-        return async () => {
-          const dg = await importOptional("@voicelayer/agents-plugin-deepgram", "deepgram.tts");
-          return new dg.TTS(
-            withCreds({ model: options.model ?? "aura-2-harmonia-en" }, options)
-          );
-        };
-      }
-    };
-    openai = {
-      llm(options = {}) {
-        return async () => {
-          const oa = await importOptional("@voicelayer/agents-plugin-openai", "openai.llm");
-          return new oa.LLM(
-            withCreds({ model: options.model ?? "gpt-4o-mini" }, options)
-          );
-        };
-      },
-      tts(options = {}) {
-        return async () => {
-          const oa = await importOptional("@voicelayer/agents-plugin-openai", "openai.tts");
-          const opts = { model: options.model ?? "tts-1" };
-          if (options.voice) opts["voice"] = options.voice;
-          if (options.instructions) opts["instructions"] = options.instructions;
-          return new oa.TTS(withCreds(opts, options));
-        };
-      },
-      // Speech-to-speech via the OpenAI Realtime API. Returns a RealtimeProvider
-      // that the SDK uses in place of the stt/llm/tts pipeline.
-      realtime(options = {}) {
-        return async () => {
-          const oa = await importOptional("@voicelayer/agents-plugin-openai", "openai.realtime");
-          const opts = {};
-          if (options.model) opts["model"] = options.model;
-          if (options.voice) opts["voice"] = options.voice;
-          return new oa.realtime.RealtimeModel(withCreds(opts, options));
-        };
-      }
-    };
-    anthropic = {
-      llm(options = {}) {
-        return async () => {
-          const an = await importOptional("@livekit/agents-plugin-anthropic", "anthropic.llm");
-          return new an.LLM(
-            withCreds({ model: options.model ?? "claude-sonnet-5-5" }, options)
-          );
-        };
-      }
-    };
-    cartesia = {
-      tts(options = {}) {
-        return async () => {
-          const ct = await importOptional("@voicelayer/agents-plugin-cartesia", "cartesia.tts");
-          return new ct.TTS(options);
-        };
-      },
-      stt(options = {}) {
-        return async () => {
-          const ct = await importOptional("@voicelayer/agents-plugin-cartesia", "cartesia.stt");
-          const opts = {};
-          if (options.model) opts["model"] = options.model;
-          if (options.language) opts["language"] = options.language;
-          return new ct.STT(withCreds(opts, options));
-        };
-      }
-    };
-    elevenlabs = {
-      tts(options = {}) {
-        return async () => {
-          const el = await importOptional("@voicelayer/agents-plugin-elevenlabs", "elevenlabs.tts");
-          const opts = {};
-          if (options.voice) opts["voiceId"] = options.voice;
-          if (options.model) opts["model"] = options.model;
-          return new el.TTS(withCreds(opts, options));
-        };
-      }
-    };
-    assemblyai = {
-      stt(options = {}) {
-        return async () => {
-          const aai = await importOptional("@voicelayer/agents-plugin-assemblyai", "assemblyai.stt");
-          const opts = {};
-          if (options.language) opts["language"] = options.language;
-          return new aai.STT(withCreds(opts, options));
-        };
-      }
-    };
-    google = {
-      llm(options = {}) {
-        return async () => {
-          const g = await importOptional("@voicelayer/agents-plugin-google", "google.llm");
-          return new g.LLM(
-            withCreds({ model: options.model ?? "gemini-3.8-flash" }, options)
-          );
-        };
-      },
-      realtime(options = {}) {
-        return async () => {
-          const g = await importOptional("@voicelayer/agents-plugin-google", "google.realtime");
-          const opts = {};
-          if (options.model) opts["model"] = options.model;
-          if (options.voice) opts["voice"] = options.voice;
-          return new g.beta.realtime.RealtimeModel(withCreds(opts, options));
-        };
-      }
-    };
-    silero = {
-      vad() {
-        return async () => {
-          const sil = await importOptional("@voicelayer/agents-plugin-silero", "silero.vad");
-          return await sil.VAD.load();
-        };
-      }
-    };
-    livekitTurn = {
-      english() {
-        return async () => {
-          const lk = await importOptional("@voicelayer/agents-plugin-livekit", "livekitTurn.english");
-          return new lk.turnDetector.EnglishModel();
-        };
-      },
-      multilingual() {
-        return async () => {
-          const lk = await importOptional("@voicelayer/agents-plugin-livekit", "livekitTurn.multilingual");
-          return new lk.turnDetector.MultilingualModel();
-        };
-      }
-    };
-    connector = {
-      llm(config = {}) {
-        return async (call) => {
-          if (config.url) {
-            await assertPublicHttpsUrl(
-              config.url,
-              config.allowHosts ? { allowHosts: config.allowHosts } : {}
-            );
-            return openai.llm({
-              ...config.model ? { model: config.model } : {},
-              ...config.apiKey ? { apiKey: config.apiKey } : {},
-              baseURL: config.url
-            })(call);
-          }
-          const transport = config.transport ?? (config.onQuery ? callbackTransport(config.onQuery) : void 0);
-          if (!transport) {
-            throw new Error("connector.llm requires one of: url, onQuery, or transport");
-          }
-          const projectId = typeof call.metadata["projectId"] === "string" ? call.metadata["projectId"] : void 0;
-          const llmOptions = {
-            ...config.model ? { model: config.model } : {},
-            ...config.temperature !== void 0 ? { temperature: config.temperature } : {},
-            ...config.fallbackText ? { fallbackText: config.fallbackText } : {},
-            ...call.callId ? { callId: call.callId } : {},
-            ...projectId ? { projectId } : {}
-          };
-          return await createConnectorLLM(transport, llmOptions);
-        };
-      }
+function getLogger(name) {
+  return logs.getLogger(name);
+}
+var DEFAULT_OTLP_ENDPOINT, DEFAULT_METRIC_EXPORT_INTERVAL_MS, started, sdk, loggerProvider, metricReader;
+var init_start = __esm({
+  "../observability/dist/start.js"() {
+    DEFAULT_OTLP_ENDPOINT = "https://otel.vlayers.ai/v1/otlp";
+    DEFAULT_METRIC_EXPORT_INTERVAL_MS = 15e3;
+    started = false;
+    sdk = null;
+    loggerProvider = null;
+    metricReader = null;
+  }
+});
+
+// ../observability/dist/attributes.js
+function callContextToAttributes(ctx) {
+  const out = {};
+  if (ctx.projectId)
+    out[ATTR.projectId] = ctx.projectId;
+  if (ctx.callId)
+    out[ATTR.callId] = ctx.callId;
+  if (ctx.campaignId)
+    out[ATTR.campaignId] = ctx.campaignId;
+  if (ctx.room)
+    out[ATTR.room] = ctx.room;
+  if (ctx.agentId)
+    out[ATTR.agentId] = ctx.agentId;
+  if (ctx.phoneNumberId)
+    out[ATTR.phoneNumberId] = ctx.phoneNumberId;
+  if (ctx.bindingId)
+    out[ATTR.bindingId] = ctx.bindingId;
+  return out;
+}
+var ATTR;
+var init_attributes = __esm({
+  "../observability/dist/attributes.js"() {
+    ATTR = {
+      projectId: "vl.project_id",
+      callId: "vl.call_id",
+      campaignId: "vl.campaign_id",
+      room: "vl.room",
+      agentId: "vl.agent_id",
+      phoneNumberId: "vl.phone_number_id",
+      bindingId: "vl.binding_id",
+      source: "vl.source",
+      kind: "vl.kind"
     };
   }
 });
-
-// src/providers/registry.ts
-function entry(id, bits) {
-  const identity = providerIdentity(id);
-  if (!identity) throw new Error(`PROVIDER_REGISTRY: no provider identity in contracts for '${id}'`);
-  return { name: identity.id, capabilities: identity.capabilities, ...bits };
+async function withCallContext(opts, fn) {
+  const tracer = trace.getTracer(TRACER_NAME);
+  const ctx = {
+    ...opts.projectId ? { projectId: opts.projectId } : {},
+    ...opts.callId ? { callId: opts.callId } : {},
+    ...opts.room ? { room: opts.room } : {},
+    ...opts.agentId ? { agentId: opts.agentId } : {},
+    ...opts.phoneNumberId ? { phoneNumberId: opts.phoneNumberId } : {},
+    ...opts.bindingId ? { bindingId: opts.bindingId } : {}
+  };
+  return callContextStore.run(ctx, () => tracer.startActiveSpan(opts.kind, async (span) => {
+    span.setAttributes(callContextToAttributes(opts));
+    span.setAttribute(ATTR.kind, opts.kind);
+    if (opts.source)
+      span.setAttribute(ATTR.source, opts.source);
+    if (opts.attributes)
+      span.setAttributes(opts.attributes);
+    try {
+      return await fn(span);
+    } catch (err) {
+      span.recordException(err);
+      span.setStatus({ code: SpanStatusCode.ERROR, message: err.message });
+      throw err;
+    } finally {
+      span.end();
+    }
+  }));
 }
-function providerEntry(name) {
-  if (!name) return void 0;
-  return PROVIDER_REGISTRY[PROVIDER_ALIASES[name] ?? name];
+function getCurrentCallContext() {
+  return callContextStore.getStore();
 }
-function resolveStt(name, o) {
-  return (providerEntry(name)?.stt ?? PROVIDER_REGISTRY["deepgram"].stt)(o);
+var TRACER_NAME, callContextStore;
+var init_call_context = __esm({
+  "../observability/dist/call-context.js"() {
+    init_attributes();
+    TRACER_NAME = "@voicelayer/observability";
+    callContextStore = new AsyncLocalStorage();
+  }
+});
+var init_trace_propagation = __esm({
+  "../observability/dist/trace-propagation.js"() {
+  }
+});
+function meter() {
+  return metrics.getMeter(METER_NAME, METER_VERSION);
 }
-function resolveTts(name, o) {
-  return (providerEntry(name)?.tts ?? PROVIDER_REGISTRY["deepgram"].tts)(o);
+function attrs(a) {
+  const out = {
+    [ATTR.projectId]: a.projectId,
+    [ATTR.callId]: a.callId,
+    "vl.provider": a.provider
+  };
+  if (a.campaignId)
+    out[ATTR.campaignId] = a.campaignId;
+  if (a.model)
+    out["vl.model"] = a.model;
+  return out;
 }
-function resolveLlm(name, o) {
-  return (providerEntry(name)?.llm ?? PROVIDER_REGISTRY["openai"].llm)(withCurrentModel(name, o));
+function llmInputTokens() {
+  return _llmInputTokens ??= meter().createCounter(USAGE_METRIC_NAMES.llmInputTokens, {
+    description: "Input tokens consumed by an LLM call",
+    unit: "{tokens}"
+  });
 }
-function resolveRealtime(name, o) {
-  return (providerEntry(name)?.realtime ?? PROVIDER_REGISTRY["openai"].realtime)(withCurrentModel(name, o));
+function llmOutputTokens() {
+  return _llmOutputTokens ??= meter().createCounter(USAGE_METRIC_NAMES.llmOutputTokens, {
+    description: "Output tokens produced by an LLM call",
+    unit: "{tokens}"
+  });
 }
-function withCurrentModel(name, o) {
-  if (!name || !o.model) return o;
-  const current = currentModelFor(name, o.model);
-  if (!current.retired) return o;
-  console.warn("[agent] the configured model is retired; running its replacement", {
-    provider: name,
-    model: current.retired,
-    replacement: current.model
+function ttsChars() {
+  return _ttsChars ??= meter().createCounter(USAGE_METRIC_NAMES.ttsChars, {
+    description: "Characters synthesized by TTS",
+    unit: "{chars}"
   });
-  return { ...o, model: current.model };
 }
-var model, lang, voice, creds, PROVIDER_REGISTRY;
-var init_registry = __esm({
-  "src/providers/registry.ts"() {
-    init_providers2();
-    init_src();
-    model = (o) => o.model ? { model: o.model } : {};
-    lang = (o) => o.language ? { language: o.language } : {};
-    voice = (o) => o.voice ? { voice: o.voice } : {};
-    creds = (o) => o.creds ?? {};
-    PROVIDER_REGISTRY = {
-      deepgram: entry("deepgram", {
-        failureClass: "vendor-api",
-        engines: ["livekit"],
-        stt: (o) => deepgram.stt({ ...model(o), ...lang(o), ...creds(o) }),
-        // Deepgram folds the voice into the model id (aura-2-<name>-en): wire config
-        // passes a voice, code config passes a model — both land as `model`.
-        tts: (o) => {
-          const m = o.voice ?? o.model;
-          return deepgram.tts({ ...m ? { model: m } : {}, ...creds(o) });
-        }
-      }),
-      openai: entry("openai", {
-        failureClass: "vendor-api",
-        engines: ["livekit", "text"],
-        llm: (o) => openai.llm({ ...model(o), ...creds(o) }),
-        tts: (o) => openai.tts({ ...voice(o), ...model(o), ...creds(o) }),
-        realtime: (o) => openai.realtime({ ...model(o), ...voice(o), ...creds(o) })
-      }),
-      anthropic: entry("anthropic", {
-        failureClass: "vendor-api",
-        engines: ["livekit", "text"],
-        llm: (o) => anthropic.llm({ ...model(o), ...creds(o) })
-      }),
-      google: entry("google", {
-        failureClass: "vendor-api",
-        engines: ["livekit", "text"],
-        llm: (o) => google.llm({ ...model(o), ...creds(o) }),
-        realtime: (o) => google.realtime({ ...model(o), ...voice(o), ...creds(o) })
-      }),
-      cartesia: entry("cartesia", {
-        failureClass: "vendor-api",
-        engines: ["livekit"],
-        stt: (o) => cartesia.stt({ ...model(o), ...lang(o), ...creds(o) }),
-        tts: (o) => cartesia.tts({ ...voice(o), ...model(o), ...creds(o) })
-      }),
-      elevenlabs: entry("elevenlabs", {
-        failureClass: "vendor-api",
-        engines: ["livekit"],
-        tts: (o) => elevenlabs.tts({ ...voice(o), ...model(o), ...creds(o) })
-      }),
-      assemblyai: entry("assemblyai", {
-        failureClass: "vendor-api",
-        engines: ["livekit"],
-        stt: (o) => assemblyai.stt({ ...lang(o), ...creds(o) })
-        // no model knob
-      }),
-      silero: entry("silero", {
-        failureClass: "local",
-        engines: ["livekit"],
-        vad: () => silero.vad()
-      })
-    };
-  }
-});
-
-// src/providers/llm.ts
-var init_llm = __esm({
-  "src/providers/llm.ts"() {
-    init_openai_default();
-    init_registry();
-  }
-});
-function metricExportIntervalMs() {
-  const raw = Number(process.env.OTEL_METRIC_EXPORT_INTERVAL);
-  return Number.isFinite(raw) && raw > 0 ? raw : DEFAULT_METRIC_EXPORT_INTERVAL_MS;
-}
-function start(opts) {
-  if (started)
-    return;
-  started = true;
-  if (opts.debug) {
-    diag.setLogger(new DiagConsoleLogger(), DiagLogLevel.INFO);
-  }
-  const authKey = process.env.VOICELAYER_TELEMETRY_KEY || process.env.VOICELAYER_API_KEY;
-  if (!authKey) {
-    return;
-  }
-  if (!process.env.OTEL_EXPORTER_OTLP_ENDPOINT) {
-    process.env.OTEL_EXPORTER_OTLP_ENDPOINT = DEFAULT_OTLP_ENDPOINT;
-  }
-  process.env.OTEL_EXPORTER_OTLP_HEADERS = `Authorization=Bearer ${authKey}`;
-  const resource = resourceFromAttributes({
-    [ATTR_SERVICE_NAME]: opts.service,
-    ...opts.version ? { [ATTR_SERVICE_VERSION]: opts.version } : {}
-  });
-  loggerProvider = new LoggerProvider({
-    resource,
-    processors: [new BatchLogRecordProcessor(new OTLPLogExporter())]
-  });
-  logs.setGlobalLoggerProvider(loggerProvider);
-  metricReader = new PeriodicExportingMetricReader({
-    exportIntervalMillis: metricExportIntervalMs(),
-    // DELTA temporality: each export carries only the increment since the
-    // previous one. The read side (apps/api openobserve-repository) totals a
-    // metric with SUM(value) over the query window, which is correct only for
-    // deltas. The OTLP exporter defaults to CUMULATIVE — it re-reports the
-    // running total every export, so SUM double-counts every call that lives
-    // longer than one export interval (inflating LLM/TTS/STT usage + cost).
-    // Setting DELTA makes the existing SUM reads correct without touching them.
-    exporter: new OTLPMetricExporter({
-      temporalityPreference: AggregationTemporalityPreference.DELTA
-    })
-  });
-  sdk = new NodeSDK({
-    resource,
-    traceExporter: new OTLPTraceExporter(),
-    metricReader,
-    instrumentations: [
-      getNodeAutoInstrumentations({
-        // fs instrumentation is extremely noisy and rarely useful for app traces.
-        "@opentelemetry/instrumentation-fs": { enabled: false }
-      })
-    ]
-  });
-  sdk.start();
-  const shutdown = async () => {
-    try {
-      await sdk?.shutdown();
-      await loggerProvider?.shutdown();
-    } catch {
-    }
-  };
-  process.once("SIGTERM", () => void shutdown().then(() => process.exit(0)));
-  process.once("SIGINT", () => void shutdown().then(() => process.exit(0)));
-}
-async function flushTelemetry() {
-  const flushes = [];
-  if (metricReader)
-    flushes.push(metricReader.forceFlush());
-  if (loggerProvider)
-    flushes.push(loggerProvider.forceFlush());
-  const proxied = trace.getTracerProvider();
-  const tracerProvider = proxied.getDelegate?.() ?? proxied;
-  if (typeof tracerProvider.forceFlush === "function") {
-    flushes.push(tracerProvider.forceFlush());
-  }
-  await Promise.allSettled(flushes);
-}
-function getLogger(name) {
-  return logs.getLogger(name);
-}
-var DEFAULT_OTLP_ENDPOINT, DEFAULT_METRIC_EXPORT_INTERVAL_MS, started, sdk, loggerProvider, metricReader;
-var init_start = __esm({
-  "../observability/dist/start.js"() {
-    DEFAULT_OTLP_ENDPOINT = "https://otel.vlayers.ai/v1/otlp";
-    DEFAULT_METRIC_EXPORT_INTERVAL_MS = 15e3;
-    started = false;
-    sdk = null;
-    loggerProvider = null;
-    metricReader = null;
-  }
-});
-
-// ../observability/dist/attributes.js
-function callContextToAttributes(ctx) {
-  const out = {};
-  if (ctx.projectId)
-    out[ATTR.projectId] = ctx.projectId;
-  if (ctx.callId)
-    out[ATTR.callId] = ctx.callId;
-  if (ctx.campaignId)
-    out[ATTR.campaignId] = ctx.campaignId;
-  if (ctx.room)
-    out[ATTR.room] = ctx.room;
-  if (ctx.agentId)
-    out[ATTR.agentId] = ctx.agentId;
-  if (ctx.phoneNumberId)
-    out[ATTR.phoneNumberId] = ctx.phoneNumberId;
-  if (ctx.bindingId)
-    out[ATTR.bindingId] = ctx.bindingId;
-  return out;
-}
-var ATTR;
-var init_attributes = __esm({
-  "../observability/dist/attributes.js"() {
-    ATTR = {
-      projectId: "vl.project_id",
-      callId: "vl.call_id",
-      campaignId: "vl.campaign_id",
-      room: "vl.room",
-      agentId: "vl.agent_id",
-      phoneNumberId: "vl.phone_number_id",
-      bindingId: "vl.binding_id",
-      source: "vl.source",
-      kind: "vl.kind"
-    };
-  }
-});
-async function withCallContext(opts, fn) {
-  const tracer = trace.getTracer(TRACER_NAME);
-  const ctx = {
-    ...opts.projectId ? { projectId: opts.projectId } : {},
-    ...opts.callId ? { callId: opts.callId } : {},
-    ...opts.room ? { room: opts.room } : {},
-    ...opts.agentId ? { agentId: opts.agentId } : {},
-    ...opts.phoneNumberId ? { phoneNumberId: opts.phoneNumberId } : {},
-    ...opts.bindingId ? { bindingId: opts.bindingId } : {}
-  };
-  return callContextStore.run(ctx, () => tracer.startActiveSpan(opts.kind, async (span) => {
-    span.setAttributes(callContextToAttributes(opts));
-    span.setAttribute(ATTR.kind, opts.kind);
-    if (opts.source)
-      span.setAttribute(ATTR.source, opts.source);
-    if (opts.attributes)
-      span.setAttributes(opts.attributes);
-    try {
-      return await fn(span);
-    } catch (err) {
-      span.recordException(err);
-      span.setStatus({ code: SpanStatusCode.ERROR, message: err.message });
-      throw err;
-    } finally {
-      span.end();
-    }
-  }));
-}
-function getCurrentCallContext() {
-  return callContextStore.getStore();
-}
-var TRACER_NAME, callContextStore;
-var init_call_context = __esm({
-  "../observability/dist/call-context.js"() {
-    init_attributes();
-    TRACER_NAME = "@voicelayer/observability";
-    callContextStore = new AsyncLocalStorage();
-  }
-});
-var init_trace_propagation = __esm({
-  "../observability/dist/trace-propagation.js"() {
-  }
-});
-function meter() {
-  return metrics.getMeter(METER_NAME, METER_VERSION);
-}
-function attrs(a) {
-  const out = {
-    [ATTR.projectId]: a.projectId,
-    [ATTR.callId]: a.callId,
-    "vl.provider": a.provider
-  };
-  if (a.campaignId)
-    out[ATTR.campaignId] = a.campaignId;
-  if (a.model)
-    out["vl.model"] = a.model;
-  return out;
-}
-function llmInputTokens() {
-  return _llmInputTokens ??= meter().createCounter(USAGE_METRIC_NAMES.llmInputTokens, {
-    description: "Input tokens consumed by an LLM call",
-    unit: "{tokens}"
-  });
-}
-function llmOutputTokens() {
-  return _llmOutputTokens ??= meter().createCounter(USAGE_METRIC_NAMES.llmOutputTokens, {
-    description: "Output tokens produced by an LLM call",
-    unit: "{tokens}"
-  });
-}
-function ttsChars() {
-  return _ttsChars ??= meter().createCounter(USAGE_METRIC_NAMES.ttsChars, {
-    description: "Characters synthesized by TTS",
-    unit: "{chars}"
-  });
-}
-function sttSeconds() {
-  return _sttSeconds ??= meter().createCounter(USAGE_METRIC_NAMES.sttSeconds, {
-    description: "Audio seconds transcribed by STT",
-    unit: "s"
-  });
+function sttSeconds() {
+  return _sttSeconds ??= meter().createCounter(USAGE_METRIC_NAMES.sttSeconds, {
+    description: "Audio seconds transcribed by STT",
+    unit: "s"
+  });
 }
 function firstTokenMs() {
   return _firstTokenMs ??= meter().createHistogram("vl.turn.first_token_ms", {
@@ -5719,157 +5533,710 @@ function recordEouLatency(args) {
     eouOnTurnCompletedDelayMs().record(args.onUserTurnCompletedDelayMs, base);
   }
 }
-function interruptionTotalMs() {
-  return _interruptionTotalMs ??= meter().createHistogram("vl.interruption.total_ms", {
-    description: "RTT for interruption detection inference",
-    unit: "ms"
-  });
-}
-function interruptionPredictionMs() {
-  return _interruptionPredictionMs ??= meter().createHistogram("vl.interruption.prediction_ms", {
-    description: "Model-side time for interruption prediction",
-    unit: "ms"
-  });
-}
-function interruptionDetectionDelayMs() {
-  return _interruptionDetectionDelayMs ??= meter().createHistogram("vl.interruption.detection_delay_ms", {
-    description: "ms from speech onset to final interruption prediction",
-    unit: "ms"
-  });
-}
-function interruptionCountCounter() {
-  return _interruptionCount ??= meter().createCounter("vl.interruption.count", {
-    description: "Interruptions detected",
-    unit: "{interruptions}"
-  });
+function interruptionTotalMs() {
+  return _interruptionTotalMs ??= meter().createHistogram("vl.interruption.total_ms", {
+    description: "RTT for interruption detection inference",
+    unit: "ms"
+  });
+}
+function interruptionPredictionMs() {
+  return _interruptionPredictionMs ??= meter().createHistogram("vl.interruption.prediction_ms", {
+    description: "Model-side time for interruption prediction",
+    unit: "ms"
+  });
+}
+function interruptionDetectionDelayMs() {
+  return _interruptionDetectionDelayMs ??= meter().createHistogram("vl.interruption.detection_delay_ms", {
+    description: "ms from speech onset to final interruption prediction",
+    unit: "ms"
+  });
+}
+function interruptionCountCounter() {
+  return _interruptionCount ??= meter().createCounter("vl.interruption.count", {
+    description: "Interruptions detected",
+    unit: "{interruptions}"
+  });
+}
+function backchannelCountCounter() {
+  return _backchannelCount ??= meter().createCounter("vl.interruption.backchannel_count", {
+    description: "Backchannel utterances detected",
+    unit: "{backchannels}"
+  });
+}
+function recordInterruptionMetrics(args) {
+  const base = {
+    [ATTR.projectId]: args.projectId,
+    [ATTR.callId]: args.callId
+  };
+  if (args.campaignId)
+    base[ATTR.campaignId] = args.campaignId;
+  if (args.totalDurationMs !== void 0 && args.totalDurationMs > 0) {
+    interruptionTotalMs().record(args.totalDurationMs, base);
+  }
+  if (args.predictionDurationMs !== void 0 && args.predictionDurationMs > 0) {
+    interruptionPredictionMs().record(args.predictionDurationMs, base);
+  }
+  if (args.detectionDelayMs !== void 0 && args.detectionDelayMs > 0) {
+    interruptionDetectionDelayMs().record(args.detectionDelayMs, base);
+  }
+  if (args.numInterruptions !== void 0 && args.numInterruptions > 0) {
+    interruptionCountCounter().add(args.numInterruptions, base);
+  }
+  if (args.numBackchannels !== void 0 && args.numBackchannels > 0) {
+    backchannelCountCounter().add(args.numBackchannels, base);
+  }
+}
+function realtimeSessionDurationMs() {
+  return _realtimeSessionDurationMs ??= meter().createCounter("vl.realtime.session_duration_ms", {
+    description: "Realtime model session duration for session-billed providers",
+    unit: "ms"
+  });
+}
+function recordRealtimeSessionDuration(args) {
+  if (args.sessionDurationMs > 0) {
+    realtimeSessionDurationMs().add(args.sessionDurationMs, attrs(args));
+  }
+}
+function modelFallbacks() {
+  return _modelFallbacks ??= meter().createCounter("vl.llm.model_fallback", {
+    description: "Model calls the provider rejected that fell back to the platform default model",
+    unit: "{fallbacks}"
+  });
+}
+function recordModelFallback(args) {
+  const out = {
+    "vl.model": args.model,
+    "vl.fallback_model": args.fallbackModel,
+    "vl.error_code": args.code
+  };
+  if (args.projectId)
+    out[ATTR.projectId] = args.projectId;
+  if (args.agentId)
+    out[ATTR.agentId] = args.agentId;
+  if (args.surface)
+    out["vl.surface"] = args.surface;
+  modelFallbacks().add(1, out);
+}
+var METER_NAME, METER_VERSION, USAGE_METRIC_NAMES, _llmInputTokens, _llmOutputTokens, _ttsChars, _sttSeconds, _firstTokenMs, _ttsStartMs, _eouDelayMs, _eouTranscriptionDelayMs, _eouOnTurnCompletedDelayMs, _interruptionTotalMs, _interruptionPredictionMs, _interruptionDetectionDelayMs, _interruptionCount, _backchannelCount, _realtimeSessionDurationMs, _modelFallbacks;
+var init_metrics = __esm({
+  "../observability/dist/metrics.js"() {
+    init_attributes();
+    METER_NAME = "voicelayer";
+    METER_VERSION = "0.1.0";
+    USAGE_METRIC_NAMES = {
+      llmInputTokens: "vl.llm.input_tokens",
+      llmOutputTokens: "vl.llm.output_tokens",
+      ttsChars: "vl.tts.chars",
+      sttSeconds: "vl.stt.seconds"
+    };
+    _llmInputTokens = null;
+    _llmOutputTokens = null;
+    _ttsChars = null;
+    _sttSeconds = null;
+    _firstTokenMs = null;
+    _ttsStartMs = null;
+    _eouDelayMs = null;
+    _eouTranscriptionDelayMs = null;
+    _eouOnTurnCompletedDelayMs = null;
+    _interruptionTotalMs = null;
+    _interruptionPredictionMs = null;
+    _interruptionDetectionDelayMs = null;
+    _interruptionCount = null;
+    _backchannelCount = null;
+    _realtimeSessionDurationMs = null;
+    _modelFallbacks = null;
+  }
+});
+function recordTurnLatencySpan(args) {
+  if (!Number.isFinite(args.latencyMs) || args.latencyMs < 0)
+    return;
+  if (args.latencyMs === 0 && !args.allowZero)
+    return;
+  const attributes = {
+    [ATTR.projectId]: args.projectId,
+    [ATTR.callId]: args.callId,
+    [LATENCY_STAGE_ATTR]: args.stage,
+    [LATENCY_MS_ATTR]: args.latencyMs,
+    [LATENCY_REALTIME_ATTR]: args.realtime ? 1 : 0
+  };
+  if (args.campaignId)
+    attributes[ATTR.campaignId] = args.campaignId;
+  trace.getTracer(TRACER_NAME2).startSpan(LATENCY_SPAN_NAME, { attributes }).end();
+}
+var TRACER_NAME2, LATENCY_SPAN_NAME, LATENCY_STAGE_ATTR, LATENCY_MS_ATTR, LATENCY_REALTIME_ATTR;
+var init_latency_span = __esm({
+  "../observability/dist/latency-span.js"() {
+    init_attributes();
+    TRACER_NAME2 = "@voicelayer/observability";
+    LATENCY_SPAN_NAME = "vl.turn_latency";
+    LATENCY_STAGE_ATTR = "vl.stage";
+    LATENCY_MS_ATTR = "vl.latency_ms";
+    LATENCY_REALTIME_ATTR = "vl.realtime";
+  }
+});
+
+// ../observability/dist/index.js
+var init_dist2 = __esm({
+  "../observability/dist/index.js"() {
+    init_start();
+    init_call_context();
+    init_trace_propagation();
+    init_attributes();
+    init_metrics();
+    init_latency_span();
+  }
+});
+
+// src/providers/resilient-llm.ts
+var resilient_llm_exports = {};
+__export(resilient_llm_exports, {
+  LLM_APOLOGY: () => LLM_APOLOGY,
+  ResilientLLM: () => ResilientLLM,
+  isResilientLLM: () => isResilientLLM
+});
+function isResilientLLM(value) {
+  return typeof value === "object" && value !== null && value[RESILIENT] === true;
+}
+function transient(error) {
+  if (error instanceof APIStatusError) {
+    const s = error.statusCode;
+    return s === 408 || s === 429 || s < 0 || s >= 500;
+  }
+  return error instanceof APITimeoutError || error instanceof APIConnectionError;
+}
+var LLM_APOLOGY, RESILIENT, ResilientLLM, sameModel, ServedLLM, ResilientLLMStream;
+var init_resilient_llm = __esm({
+  "src/providers/resilient-llm.ts"() {
+    init_dist();
+    init_dist2();
+    LLM_APOLOGY = "Sorry, I'm having trouble right now. Could you give me a moment and say that again?";
+    RESILIENT = /* @__PURE__ */ Symbol.for("voicelayer.resilientLLM");
+    ResilientLLM = class extends llm.LLM {
+      [RESILIENT] = true;
+      #opts;
+      #fallback = null;
+      #listeners = /* @__PURE__ */ new Set();
+      /** Models the provider rejected on this call (a call's pipeline is built per call): later turns go straight to the fallback. */
+      rejections = new ModelRejectionCache();
+      constructor(opts) {
+        super();
+        this.#opts = opts;
+      }
+      label() {
+        return this.#opts.primary.label();
+      }
+      get model() {
+        return this.#opts.primary.model;
+      }
+      get provider() {
+        return this.#opts.primary.provider;
+      }
+      get apology() {
+        return this.#opts.apology ?? LLM_APOLOGY;
+      }
+      get reasoningParams() {
+        return this.#opts.reasoningParams === true;
+      }
+      /** The fallback LLM, built on first need; null when there is none or it is the agent's own model. */
+      fallbackLLM() {
+        if (!this.#opts.fallback) return null;
+        this.#fallback ??= this.#opts.fallback();
+        return sameModel(this.#fallback.model, this.model) ? null : this.#fallback;
+      }
+      /** Hear about every recovery (and every apology). Returns the unsubscribe. */
+      onIncident(listener) {
+        this.#listeners.add(listener);
+        return () => this.#listeners.delete(listener);
+      }
+      /** @internal */
+      report(incident) {
+        const log = incident.kind === "apology" ? console.error : console.warn;
+        log(`[voicelayer] llm ${incident.kind.replace(/_/g, " ")}`, incident);
+        if (incident.kind === "model_fallback") {
+          const call = getCurrentCallContext();
+          recordModelFallback({
+            model: incident.model,
+            fallbackModel: incident.fallbackModel,
+            code: incident.code,
+            surface: "voice",
+            ...call?.projectId ? { projectId: call.projectId } : {},
+            ...call?.agentId ? { agentId: call.agentId } : {}
+          });
+        }
+        for (const listener of this.#listeners) {
+          try {
+            listener(incident);
+          } catch {
+          }
+        }
+      }
+      chat(args) {
+        return new ResilientLLMStream(this, args);
+      }
+      prewarm() {
+        this.#opts.primary.prewarm();
+      }
+      async aclose() {
+        await Promise.all([this.#opts.primary.aclose(), this.#fallback?.aclose()]);
+      }
+      /** @internal */
+      get primary() {
+        return this.#opts.primary;
+      }
+    };
+    sameModel = (a, b) => a.trim().toLowerCase() === b.trim().toLowerCase();
+    ServedLLM = class extends llm.LLM {
+      constructor(owner, served) {
+        super();
+        this.owner = owner;
+        this.served = served;
+      }
+      owner;
+      served;
+      label() {
+        return this.owner.label();
+      }
+      get model() {
+        return this.served();
+      }
+      get provider() {
+        return this.owner.provider;
+      }
+      chat(args) {
+        return this.owner.chat(args);
+      }
+      emit(event, ...args) {
+        return this.owner.emit(event, ...args);
+      }
+    };
+    ResilientLLMStream = class extends llm.LLMStream {
+      #owner;
+      #args;
+      #conn;
+      /** The model answering this stream — what its metrics report. */
+      #served;
+      #current = null;
+      constructor(owner, args) {
+        const conn = args.connOptions ?? DEFAULT_API_CONNECT_OPTIONS;
+        const served = { model: owner.model };
+        super(new ServedLLM(owner, () => served.model), {
+          chatCtx: args.chatCtx,
+          ...args.toolCtx ? { toolCtx: args.toolCtx } : {},
+          connOptions: { ...conn, maxRetry: 0 }
+        });
+        this.#owner = owner;
+        this.#args = args;
+        this.#conn = conn;
+        this.#served = served;
+        this.abortController.signal.addEventListener("abort", () => this.#current?.close());
+      }
+      get #hasTools() {
+        return this.#args.toolCtx !== void 0 && Object.keys(this.#args.toolCtx).length > 0;
+      }
+      #extraKwargs(model2) {
+        const base = this.#args.extraKwargs;
+        if (!this.#owner.reasoningParams) return base;
+        const effort = chatReasoningEffort(model2, { tools: this.#hasTools });
+        if (effort === void 0) {
+          if (!base || !("reasoning_effort" in base)) return base;
+          const { reasoning_effort: _dropped, ...rest } = base;
+          return rest;
+        }
+        return { ...base, reasoning_effort: effort };
+      }
+      /** One request on `target`, forwarding its chunks. Its failure is the inner LLM's 'error' event. */
+      async #attempt(target) {
+        let failure2 = null;
+        const onError = (ev) => {
+          failure2 ??= ev.error;
+        };
+        target.on("error", onError);
+        let started2 = false;
+        try {
+          const extraKwargs = this.#extraKwargs(target.model);
+          const stream = target.chat({
+            ...this.#args,
+            connOptions: { ...this.#conn, maxRetry: 0 },
+            ...extraKwargs !== void 0 ? { extraKwargs } : {}
+          });
+          this.#current = stream;
+          for await (const chunk of stream) {
+            if (this.abortController.signal.aborted) break;
+            started2 = true;
+            this.queue.put(chunk);
+          }
+        } catch (err) {
+          failure2 ??= err instanceof Error ? err : new Error(String(err));
+        } finally {
+          target.off("error", onError);
+          this.#current = null;
+        }
+        return failure2 ? { ok: false, error: failure2, started: started2 } : { ok: true };
+      }
+      /** Run `target` until it answers, or a failure that retrying it won't fix. */
+      async #run(target) {
+        let effortRetried = false;
+        for (let retries = 0; ; ) {
+          const sent = this.#owner.reasoningParams ? chatReasoningEffort(target.model, { tools: this.#hasTools }) : void 0;
+          const result = await this.#attempt(target);
+          if (result.ok || result.started || this.abortController.signal.aborted) return result;
+          if (this.#owner.reasoningParams && !effortRetried && learnReasoningEffort(target.model, { tools: this.#hasTools, sent }, result.error)) {
+            effortRetried = true;
+            this.#owner.report({
+              kind: "reasoning_effort_adapted",
+              model: target.model,
+              message: providerErrorOf(result.error)?.message ?? result.error.message
+            });
+            continue;
+          }
+          if (!transient(result.error) || retries >= this.#conn.maxRetry) return result;
+          const wait = intervalForRetry(this.#conn, retries);
+          retries += 1;
+          if (wait > 0) await new Promise((r) => setTimeout(r, wait));
+          if (this.abortController.signal.aborted) return result;
+        }
+      }
+      async run() {
+        const owner = this.#owner;
+        const model2 = owner.model;
+        const fallback = owner.fallbackLLM();
+        const known = fallback ? owner.rejections.get(model2) : null;
+        let failure2;
+        if (known && fallback) {
+          owner.report({ kind: "model_fallback", model: model2, fallbackModel: fallback.model, code: known.code, message: known.message, cached: true });
+          failure2 = new Error(known.message);
+        } else {
+          const first = await this.#run(owner.primary);
+          if (first.ok || first.started || this.abortController.signal.aborted) return;
+          failure2 = first.error;
+          const rejection = modelRejectionOf(first.error);
+          if (rejection) owner.rejections.set(model2, rejection);
+          if (fallback) {
+            const f = providerErrorOf(first.error);
+            owner.report({
+              kind: "model_fallback",
+              model: model2,
+              fallbackModel: fallback.model,
+              code: rejection?.code ?? (f?.status ? String(f.status) : "error"),
+              message: f?.message || first.error.message,
+              cached: false
+            });
+          }
+        }
+        if (fallback) {
+          this.#served.model = fallback.model;
+          const second = await this.#run(fallback);
+          if (second.ok || second.started || this.abortController.signal.aborted) return;
+          failure2 = second.error;
+        }
+        owner.report({ kind: "apology", model: this.#served.model, message: providerErrorOf(failure2)?.message || failure2.message });
+        this.queue.put({ id: `vl-apology-${Date.now()}`, delta: { role: "assistant", content: owner.apology } });
+      }
+    };
+  }
+});
+
+// src/providers/index.ts
+async function importOptional(spec, hint) {
+  try {
+    return await import(spec);
+  } catch {
+    throw new Error(`${hint}: "${spec}" is not installed. Add it with: pnpm add ${spec}`);
+  }
+}
+function withCreds(opts, creds2) {
+  if (creds2?.apiKey) opts["apiKey"] = creds2.apiKey;
+  if (creds2?.baseURL) opts["baseURL"] = creds2.baseURL;
+  return opts;
+}
+var deepgram, openai, anthropic, cartesia, elevenlabs, assemblyai, google, silero, livekitTurn, connector;
+var init_providers2 = __esm({
+  "src/providers/index.ts"() {
+    init_ssrf();
+    init_transport_callback();
+    init_connector_llm();
+    init_wrap();
+    deepgram = {
+      stt(options = {}) {
+        return async () => {
+          const dg = await importOptional("@voicelayer/agents-plugin-deepgram", "deepgram.stt");
+          return new dg.STT(
+            withCreds(
+              { model: options.model ?? "nova-3", language: options.language ?? "en-US" },
+              options
+            )
+          );
+        };
+      },
+      tts(options = {}) {
+        return async () => {
+          const dg = await importOptional("@voicelayer/agents-plugin-deepgram", "deepgram.tts");
+          return new dg.TTS(
+            withCreds({ model: options.model ?? "aura-2-harmonia-en" }, options)
+          );
+        };
+      }
+    };
+    openai = {
+      llm(options = {}) {
+        return async () => {
+          const oa = await importOptional("@voicelayer/agents-plugin-openai", "openai.llm");
+          const { ResilientLLM: ResilientLLM2 } = await Promise.resolve().then(() => (init_resilient_llm(), resilient_llm_exports));
+          const { PLATFORM_DEFAULT_CHAT_MODEL: PLATFORM_DEFAULT_CHAT_MODEL2 } = await Promise.resolve().then(() => (init_dist(), dist_exports));
+          return new ResilientLLM2({
+            primary: new oa.LLM(withCreds({ model: options.model ?? "gpt-4o-mini" }, options)),
+            fallback: () => new oa.LLM(withCreds({ model: PLATFORM_DEFAULT_CHAT_MODEL2 }, options)),
+            reasoningParams: true
+          });
+        };
+      },
+      tts(options = {}) {
+        return async () => {
+          const oa = await importOptional("@voicelayer/agents-plugin-openai", "openai.tts");
+          const opts = { model: options.model ?? "tts-1" };
+          if (options.voice) opts["voice"] = options.voice;
+          if (options.instructions) opts["instructions"] = options.instructions;
+          return new oa.TTS(withCreds(opts, options));
+        };
+      },
+      // Speech-to-speech via the OpenAI Realtime API. Returns a RealtimeProvider
+      // that the SDK uses in place of the stt/llm/tts pipeline.
+      realtime(options = {}) {
+        return async () => {
+          const oa = await importOptional("@voicelayer/agents-plugin-openai", "openai.realtime");
+          const opts = {};
+          if (options.model) opts["model"] = options.model;
+          if (options.voice) opts["voice"] = options.voice;
+          return new oa.realtime.RealtimeModel(withCreds(opts, options));
+        };
+      }
+    };
+    anthropic = {
+      llm(options = {}) {
+        return async () => {
+          const an = await importOptional("@livekit/agents-plugin-anthropic", "anthropic.llm");
+          return new an.LLM(
+            withCreds({ model: options.model ?? "claude-sonnet-5-5" }, options)
+          );
+        };
+      }
+    };
+    cartesia = {
+      tts(options = {}) {
+        return async () => {
+          const ct = await importOptional("@voicelayer/agents-plugin-cartesia", "cartesia.tts");
+          return new ct.TTS(options);
+        };
+      },
+      stt(options = {}) {
+        return async () => {
+          const ct = await importOptional("@voicelayer/agents-plugin-cartesia", "cartesia.stt");
+          const opts = {};
+          if (options.model) opts["model"] = options.model;
+          if (options.language) opts["language"] = options.language;
+          return new ct.STT(withCreds(opts, options));
+        };
+      }
+    };
+    elevenlabs = {
+      tts(options = {}) {
+        return async () => {
+          const el = await importOptional("@voicelayer/agents-plugin-elevenlabs", "elevenlabs.tts");
+          const opts = {};
+          if (options.voice) opts["voiceId"] = options.voice;
+          if (options.model) opts["model"] = options.model;
+          return new el.TTS(withCreds(opts, options));
+        };
+      }
+    };
+    assemblyai = {
+      stt(options = {}) {
+        return async () => {
+          const aai = await importOptional("@voicelayer/agents-plugin-assemblyai", "assemblyai.stt");
+          const opts = {};
+          if (options.language) opts["language"] = options.language;
+          return new aai.STT(withCreds(opts, options));
+        };
+      }
+    };
+    google = {
+      llm(options = {}) {
+        return async () => {
+          const g = await importOptional("@voicelayer/agents-plugin-google", "google.llm");
+          const { ResilientLLM: ResilientLLM2 } = await Promise.resolve().then(() => (init_resilient_llm(), resilient_llm_exports));
+          return new ResilientLLM2({
+            primary: new g.LLM(withCreds({ model: options.model ?? "gemini-3.8-flash" }, options))
+          });
+        };
+      },
+      realtime(options = {}) {
+        return async () => {
+          const g = await importOptional("@voicelayer/agents-plugin-google", "google.realtime");
+          const opts = {};
+          if (options.model) opts["model"] = options.model;
+          if (options.voice) opts["voice"] = options.voice;
+          return new g.beta.realtime.RealtimeModel(withCreds(opts, options));
+        };
+      }
+    };
+    silero = {
+      vad() {
+        return async () => {
+          const sil = await importOptional("@voicelayer/agents-plugin-silero", "silero.vad");
+          return await sil.VAD.load();
+        };
+      }
+    };
+    livekitTurn = {
+      english() {
+        return async () => {
+          const lk = await importOptional("@voicelayer/agents-plugin-livekit", "livekitTurn.english");
+          return new lk.turnDetector.EnglishModel();
+        };
+      },
+      multilingual() {
+        return async () => {
+          const lk = await importOptional("@voicelayer/agents-plugin-livekit", "livekitTurn.multilingual");
+          return new lk.turnDetector.MultilingualModel();
+        };
+      }
+    };
+    connector = {
+      llm(config = {}) {
+        return async (call) => {
+          if (config.url) {
+            await assertPublicHttpsUrl(
+              config.url,
+              config.allowHosts ? { allowHosts: config.allowHosts } : {}
+            );
+            return openai.llm({
+              ...config.model ? { model: config.model } : {},
+              ...config.apiKey ? { apiKey: config.apiKey } : {},
+              baseURL: config.url
+            })(call);
+          }
+          const transport = config.transport ?? (config.onQuery ? callbackTransport(config.onQuery) : void 0);
+          if (!transport) {
+            throw new Error("connector.llm requires one of: url, onQuery, or transport");
+          }
+          const projectId = typeof call.metadata["projectId"] === "string" ? call.metadata["projectId"] : void 0;
+          const llmOptions = {
+            ...config.model ? { model: config.model } : {},
+            ...config.temperature !== void 0 ? { temperature: config.temperature } : {},
+            ...config.fallbackText ? { fallbackText: config.fallbackText } : {},
+            ...call.callId ? { callId: call.callId } : {},
+            ...projectId ? { projectId } : {}
+          };
+          return await createConnectorLLM(transport, llmOptions);
+        };
+      }
+    };
+  }
+});
+
+// src/providers/registry.ts
+function entry(id, bits) {
+  const identity = providerIdentity(id);
+  if (!identity) throw new Error(`PROVIDER_REGISTRY: no provider identity in contracts for '${id}'`);
+  return { name: identity.id, capabilities: identity.capabilities, ...bits };
+}
+function providerEntry(name) {
+  if (!name) return void 0;
+  return PROVIDER_REGISTRY[PROVIDER_ALIASES[name] ?? name];
 }
-function backchannelCountCounter() {
-  return _backchannelCount ??= meter().createCounter("vl.interruption.backchannel_count", {
-    description: "Backchannel utterances detected",
-    unit: "{backchannels}"
-  });
+function resolveStt(name, o) {
+  return (providerEntry(name)?.stt ?? PROVIDER_REGISTRY["deepgram"].stt)(o);
 }
-function recordInterruptionMetrics(args) {
-  const base = {
-    [ATTR.projectId]: args.projectId,
-    [ATTR.callId]: args.callId
-  };
-  if (args.campaignId)
-    base[ATTR.campaignId] = args.campaignId;
-  if (args.totalDurationMs !== void 0 && args.totalDurationMs > 0) {
-    interruptionTotalMs().record(args.totalDurationMs, base);
-  }
-  if (args.predictionDurationMs !== void 0 && args.predictionDurationMs > 0) {
-    interruptionPredictionMs().record(args.predictionDurationMs, base);
-  }
-  if (args.detectionDelayMs !== void 0 && args.detectionDelayMs > 0) {
-    interruptionDetectionDelayMs().record(args.detectionDelayMs, base);
-  }
-  if (args.numInterruptions !== void 0 && args.numInterruptions > 0) {
-    interruptionCountCounter().add(args.numInterruptions, base);
-  }
-  if (args.numBackchannels !== void 0 && args.numBackchannels > 0) {
-    backchannelCountCounter().add(args.numBackchannels, base);
-  }
+function resolveTts(name, o) {
+  return (providerEntry(name)?.tts ?? PROVIDER_REGISTRY["deepgram"].tts)(o);
 }
-function realtimeSessionDurationMs() {
-  return _realtimeSessionDurationMs ??= meter().createCounter("vl.realtime.session_duration_ms", {
-    description: "Realtime model session duration for session-billed providers",
-    unit: "ms"
-  });
+function resolveLlm(name, o) {
+  return (providerEntry(name)?.llm ?? PROVIDER_REGISTRY["openai"].llm)(withCurrentModel(name, o));
 }
-function recordRealtimeSessionDuration(args) {
-  if (args.sessionDurationMs > 0) {
-    realtimeSessionDurationMs().add(args.sessionDurationMs, attrs(args));
-  }
+function resolveRealtime(name, o) {
+  return (providerEntry(name)?.realtime ?? PROVIDER_REGISTRY["openai"].realtime)(withCurrentModel(name, o));
 }
-function modelFallbacks() {
-  return _modelFallbacks ??= meter().createCounter("vl.llm.model_fallback", {
-    description: "Model calls the provider rejected that fell back to the platform default model",
-    unit: "{fallbacks}"
+function withCurrentModel(name, o) {
+  if (!name || !o.model) return o;
+  const current = currentModelFor(name, o.model);
+  if (!current.retired) return o;
+  console.warn("[agent] the configured model is retired; running its replacement", {
+    provider: name,
+    model: current.retired,
+    replacement: current.model
   });
+  return { ...o, model: current.model };
 }
-function recordModelFallback(args) {
-  const out = {
-    "vl.model": args.model,
-    "vl.fallback_model": args.fallbackModel,
-    "vl.error_code": args.code
-  };
-  if (args.projectId)
-    out[ATTR.projectId] = args.projectId;
-  if (args.agentId)
-    out[ATTR.agentId] = args.agentId;
-  if (args.surface)
-    out["vl.surface"] = args.surface;
-  modelFallbacks().add(1, out);
-}
-var METER_NAME, METER_VERSION, USAGE_METRIC_NAMES, _llmInputTokens, _llmOutputTokens, _ttsChars, _sttSeconds, _firstTokenMs, _ttsStartMs, _eouDelayMs, _eouTranscriptionDelayMs, _eouOnTurnCompletedDelayMs, _interruptionTotalMs, _interruptionPredictionMs, _interruptionDetectionDelayMs, _interruptionCount, _backchannelCount, _realtimeSessionDurationMs, _modelFallbacks;
-var init_metrics = __esm({
-  "../observability/dist/metrics.js"() {
-    init_attributes();
-    METER_NAME = "voicelayer";
-    METER_VERSION = "0.1.0";
-    USAGE_METRIC_NAMES = {
-      llmInputTokens: "vl.llm.input_tokens",
-      llmOutputTokens: "vl.llm.output_tokens",
-      ttsChars: "vl.tts.chars",
-      sttSeconds: "vl.stt.seconds"
+var model, lang, voice, creds, PROVIDER_REGISTRY;
+var init_registry = __esm({
+  "src/providers/registry.ts"() {
+    init_providers2();
+    init_src();
+    model = (o) => o.model ? { model: o.model } : {};
+    lang = (o) => o.language ? { language: o.language } : {};
+    voice = (o) => o.voice ? { voice: o.voice } : {};
+    creds = (o) => o.creds ?? {};
+    PROVIDER_REGISTRY = {
+      deepgram: entry("deepgram", {
+        failureClass: "vendor-api",
+        engines: ["livekit"],
+        stt: (o) => deepgram.stt({ ...model(o), ...lang(o), ...creds(o) }),
+        // Deepgram folds the voice into the model id (aura-2-<name>-en): wire config
+        // passes a voice, code config passes a model — both land as `model`.
+        tts: (o) => {
+          const m = o.voice ?? o.model;
+          return deepgram.tts({ ...m ? { model: m } : {}, ...creds(o) });
+        }
+      }),
+      openai: entry("openai", {
+        failureClass: "vendor-api",
+        engines: ["livekit", "text"],
+        llm: (o) => openai.llm({ ...model(o), ...creds(o) }),
+        tts: (o) => openai.tts({ ...voice(o), ...model(o), ...creds(o) }),
+        realtime: (o) => openai.realtime({ ...model(o), ...voice(o), ...creds(o) })
+      }),
+      anthropic: entry("anthropic", {
+        failureClass: "vendor-api",
+        engines: ["livekit", "text"],
+        llm: (o) => anthropic.llm({ ...model(o), ...creds(o) })
+      }),
+      google: entry("google", {
+        failureClass: "vendor-api",
+        engines: ["livekit", "text"],
+        llm: (o) => google.llm({ ...model(o), ...creds(o) }),
+        realtime: (o) => google.realtime({ ...model(o), ...voice(o), ...creds(o) })
+      }),
+      cartesia: entry("cartesia", {
+        failureClass: "vendor-api",
+        engines: ["livekit"],
+        stt: (o) => cartesia.stt({ ...model(o), ...lang(o), ...creds(o) }),
+        tts: (o) => cartesia.tts({ ...voice(o), ...model(o), ...creds(o) })
+      }),
+      elevenlabs: entry("elevenlabs", {
+        failureClass: "vendor-api",
+        engines: ["livekit"],
+        tts: (o) => elevenlabs.tts({ ...voice(o), ...model(o), ...creds(o) })
+      }),
+      assemblyai: entry("assemblyai", {
+        failureClass: "vendor-api",
+        engines: ["livekit"],
+        stt: (o) => assemblyai.stt({ ...lang(o), ...creds(o) })
+        // no model knob
+      }),
+      silero: entry("silero", {
+        failureClass: "local",
+        engines: ["livekit"],
+        vad: () => silero.vad()
+      })
     };
-    _llmInputTokens = null;
-    _llmOutputTokens = null;
-    _ttsChars = null;
-    _sttSeconds = null;
-    _firstTokenMs = null;
-    _ttsStartMs = null;
-    _eouDelayMs = null;
-    _eouTranscriptionDelayMs = null;
-    _eouOnTurnCompletedDelayMs = null;
-    _interruptionTotalMs = null;
-    _interruptionPredictionMs = null;
-    _interruptionDetectionDelayMs = null;
-    _interruptionCount = null;
-    _backchannelCount = null;
-    _realtimeSessionDurationMs = null;
-    _modelFallbacks = null;
-  }
-});
-function recordTurnLatencySpan(args) {
-  if (!Number.isFinite(args.latencyMs) || args.latencyMs < 0)
-    return;
-  if (args.latencyMs === 0 && !args.allowZero)
-    return;
-  const attributes = {
-    [ATTR.projectId]: args.projectId,
-    [ATTR.callId]: args.callId,
-    [LATENCY_STAGE_ATTR]: args.stage,
-    [LATENCY_MS_ATTR]: args.latencyMs,
-    [LATENCY_REALTIME_ATTR]: args.realtime ? 1 : 0
-  };
-  if (args.campaignId)
-    attributes[ATTR.campaignId] = args.campaignId;
-  trace.getTracer(TRACER_NAME2).startSpan(LATENCY_SPAN_NAME, { attributes }).end();
-}
-var TRACER_NAME2, LATENCY_SPAN_NAME, LATENCY_STAGE_ATTR, LATENCY_MS_ATTR, LATENCY_REALTIME_ATTR;
-var init_latency_span = __esm({
-  "../observability/dist/latency-span.js"() {
-    init_attributes();
-    TRACER_NAME2 = "@voicelayer/observability";
-    LATENCY_SPAN_NAME = "vl.turn_latency";
-    LATENCY_STAGE_ATTR = "vl.stage";
-    LATENCY_MS_ATTR = "vl.latency_ms";
-    LATENCY_REALTIME_ATTR = "vl.realtime";
   }
 });
 
-// ../observability/dist/index.js
-var init_dist2 = __esm({
-  "../observability/dist/index.js"() {
-    init_start();
-    init_call_context();
-    init_trace_propagation();
-    init_attributes();
-    init_metrics();
-    init_latency_span();
+// src/providers/llm.ts
+var init_llm = __esm({
+  "src/providers/llm.ts"() {
+    init_openai_default();
+    init_registry();
   }
 });
 
@@ -9237,14 +9604,23 @@ function defaultComplete() {
       model: model2,
       fallbackModel: RUNNER_MODEL,
       helper: "agent runner",
-      call: (m) => client.chat.completions.create(
-        {
-          model: m,
-          ...chatCompletionParams(m, { ...temperature !== void 0 ? { temperature } : {}, reasoningEffort: "low" }),
-          messages,
-          ...tools.length ? { tools } : {}
-        },
-        { timeout: RUNNER_TIMEOUT_MS }
+      // With tools, a model that takes effort 'none' gets 'none' (OpenAI refuses tools at any other effort on chat
+      // completions for GPT-5.2 and later), and a reasoning_effort 400 is learned and retried once (burn-down G-45).
+      call: (m) => withReasoningEffortRetry(
+        { model: m, tools: tools.length > 0, requested: "low" },
+        () => client.chat.completions.create(
+          {
+            model: m,
+            ...chatCompletionParams(m, {
+              ...temperature !== void 0 ? { temperature } : {},
+              reasoningEffort: "low",
+              tools: tools.length > 0
+            }),
+            messages,
+            ...tools.length ? { tools } : {}
+          },
+          { timeout: RUNNER_TIMEOUT_MS }
+        )
       ),
       warn: graphWarn
     });
@@ -12897,7 +13273,7 @@ async function resolvePipeline(cfg, call, preloadedVad) {
     () => language.startsWith("en") ? livekitTurn.english() : livekitTurn.multilingual(),
     () => language.startsWith("en") ? livekitTurn.english() : livekitTurn.multilingual()
   ) : null;
-  const [stt, llm3, tts, vad] = await Promise.all([
+  const [stt, llm4, tts, vad] = await Promise.all([
     sttFactory2(call),
     llmFactory2(call),
     ttsFactory2(call),
@@ -12913,7 +13289,7 @@ async function resolvePipeline(cfg, call, preloadedVad) {
       });
     }
   }
-  return { stt, llm: llm3, tts, vad, turnDetector };
+  return { stt, llm: llm4, tts, vad, turnDetector };
 }
 var DEFAULT_REQUIRED_ENV = [
   "LIVEKIT_URL",
@@ -16417,7 +16793,7 @@ async function answerTextTurn(deps) {
   const { turn, config } = deps;
   const agentId = turn.agentId;
   const now = deps.now ?? Date.now;
-  const { voice: voice3, llm: llm3, initializeLogger, loggerOptions } = await import('@livekit/agents');
+  const { voice: voice3, llm: llm4, initializeLogger, loggerOptions } = await import('@livekit/agents');
   if (loggerOptions() === void 0) initializeLogger({ pretty: false, level: "warn" });
   let effective = config;
   let pipeline = null;
@@ -16459,7 +16835,7 @@ async function answerTextTurn(deps) {
     getCtx
   });
   const instructions = processRt.augmentPrompt(composeSystemPrompt(effective.prompt, effective.routingInstructions));
-  const chatCtx = llm3.ChatContext.empty();
+  const chatCtx = llm4.ChatContext.empty();
   let skipReply = false;
   for (const h of turn.history) {
     if (h.role === "user" && security && security.inputGuard.check(h.content).action === "block") {
@@ -19040,6 +19416,7 @@ function toolAuditObserver(emit) {
 }
 
 // src/agent.ts
+init_resilient_llm();
 init_helper_models();
 var FLOW_HANGUP_GRACE_MS = Number(process.env["VL_FLOW_HANGUP_GRACE_MS"] ?? "1200");
 function isLiveKitChildProcess() {
@@ -19372,6 +19749,9 @@ var Agent = class {
       void trackWrite(sdkClient.calls.appendEvent(opEventCallId, { kind, payload })).catch(() => {
       });
     };
+    if (isResilientLLM(pipeline.llm)) {
+      pipeline.llm.onIncident((incident) => emitOpEvent(`engine.llm.${incident.kind}`, { ...incident }));
+    }
     const graphEvents = graphMode ? withOpEventFanout(createGraphEvents(processRt), emitOpEvent) : null;
     const isOutbound = stringValue(dispatchMdEarly["direction"]) === "outbound";
     const perCallAmd = stringValue(dispatchMdEarly["amd"]);