@absolutejs/voice 0.0.22-beta.615 → 0.0.22-beta.617

package/dist/core/costAccounting.d.ts

@@ -39,30 +39,40 @@ export type VoiceCostTelephonyRecord = {
     minutes: number;
     provider?: string;
 };
+export type VoiceCostProviderSlice = {
+    provider: string;
+    usd: number;
+    cachedInputTokens?: number;
+    inputTokens?: number;
+    outputTokens?: number;
+    characters?: number;
+    audioMs?: number;
+    minutes?: number;
+};
 export type VoiceCostBreakdown = {
     llm: {
         cachedInputTokens: number;
         inputTokens: number;
         outputTokens: number;
-        provider?: string;
+        byProvider: VoiceCostProviderSlice[];
         usd: number;
     };
     sessionId?: string;
     stt: {
         audioMs: number;
-        provider?: string;
+        byProvider: VoiceCostProviderSlice[];
         usd: number;
     };
     telephony: {
         minutes: number;
-        provider?: string;
+        byProvider: VoiceCostProviderSlice[];
         usd: number;
     };
     totalUsd: number;
     tts: {
         audioMs: number;
         characters: number;
-        provider?: string;
+        byProvider: VoiceCostProviderSlice[];
         usd: number;
     };
 };

package/dist/core/modelAdapters.d.ts

@@ -46,6 +46,10 @@ export type VoiceProviderRouterEvent<TProvider extends string = string> = {
     at: number;
     attempt: number;
     elapsedMs: number;
+    /** Set when this event marks a provider returning an empty completion (no text
+     *  and no tool call) that triggered a same-provider retry — so consumers can
+     *  track how often the model glitches empty, and whether the retry recovered. */
+    emptyCompletion?: boolean;
     error?: string;
     fallbackProvider?: TProvider;
     latencyBudgetMs?: number;
@@ -139,6 +143,14 @@ export type VoiceProviderRouterOptions<TContext = unknown, TSession extends Voic
     allowProviders?: readonly TProvider[] | ((input: VoiceAgentModelInput<TContext, TSession>) => readonly TProvider[] | Promise<readonly TProvider[]>);
     fallback?: TProvider[] | ((input: VoiceAgentModelInput<TContext, TSession>) => readonly TProvider[] | Promise<readonly TProvider[]>);
     fallbackMode?: VoiceProviderRouterFallbackMode;
+    /** Retry the SAME provider this many times when it returns a non-error but
+     *  EMPTY completion (no assistant text AND no tool call) before accepting it.
+     *  An empty completion is a transient model glitch (observed on OpenAI
+     *  /responses returning a 200 with zero output items mid-conversation), NOT a
+     *  provider error — so it bypasses fallbackMode and never suppresses provider
+     *  health. Without this a one-off empty surfaces to the app as a dead turn
+     *  (the caller hears a re-engage / "go on, I'm listening" loop). Default 1. */
+    emptyCompletionRetries?: number;
     isProviderError?: (error: unknown, provider: TProvider) => boolean;
     isRateLimitError?: (error: unknown, provider: TProvider) => boolean;
     isTimeoutError?: (error: unknown, provider: TProvider) => boolean;

package/dist/index.js

@@ -41866,30 +41866,20 @@ var createVoiceCostAccountant = (options = {}) => {
   let sttUsd = 0;
   let telephonyMinutes = 0;
   let telephonyUsd = 0;
-  const llmByProvider = new Map;
-  const ttsByProvider = new Map;
-  const sttByProvider = new Map;
-  const telephonyByProvider = new Map;
-  let lastLlmProvider;
-  let lastTtsProvider;
-  let lastSttProvider;
-  let lastTelephonyProvider;
-  const addProvider = (byProvider, provider, usd) => {
-    if (!provider)
-      return;
-    byProvider.set(provider, (byProvider.get(provider) ?? 0) + usd);
-  };
-  const dominant = (byProvider, fallback) => {
-    let best;
-    let bestUsd = -1;
-    for (const [provider, usd] of byProvider) {
-      if (usd > bestUsd) {
-        bestUsd = usd;
-        best = provider;
-      }
+  const llmSlices = new Map;
+  const ttsSlices = new Map;
+  const sttSlices = new Map;
+  const telephonySlices = new Map;
+  const sliceFor = (slices, provider) => {
+    let slice = slices.get(provider);
+    if (!slice) {
+      slice = { provider, usd: 0 };
+      slices.set(provider, slice);
     }
-    return best ?? fallback;
+    return slice;
   };
+  const round6 = (value) => Math.round(value * 1e6) / 1e6;
+  const finalizeSlices = (slices) => [...slices.values()].map((slice) => ({ ...slice, usd: round6(slice.usd) }));
   return {
     recordLLM: (usage) => {
       const input = usage.inputTokens ?? 0;
@@ -41898,86 +41888,98 @@ var createVoiceCostAccountant = (options = {}) => {
       llmInput += input;
       llmCachedInput += cached;
       llmOutput += output;
-      if (usage.provider)
-        lastLlmProvider = usage.provider;
       const rates = lookupRates(priceBook, usage.provider, usage.model)?.llm;
-      if (!rates) {
-        return;
+      let delta = 0;
+      if (rates) {
+        const cachedRate = rates.cachedInputPerMillionTokensUsd ?? rates.inputPerMillionTokensUsd;
+        delta = Math.max(0, input - cached) * rates.inputPerMillionTokensUsd / 1e6 + cached * cachedRate / 1e6 + output * rates.outputPerMillionTokensUsd / 1e6;
+        llmUsd += delta;
+      }
+      if (usage.provider) {
+        const slice = sliceFor(llmSlices, usage.provider);
+        slice.usd += delta;
+        slice.inputTokens = (slice.inputTokens ?? 0) + input;
+        slice.outputTokens = (slice.outputTokens ?? 0) + output;
+        slice.cachedInputTokens = (slice.cachedInputTokens ?? 0) + cached;
       }
-      const cachedRate = rates.cachedInputPerMillionTokensUsd ?? rates.inputPerMillionTokensUsd;
-      const delta = Math.max(0, input - cached) * rates.inputPerMillionTokensUsd / 1e6 + cached * cachedRate / 1e6 + output * rates.outputPerMillionTokensUsd / 1e6;
-      llmUsd += delta;
-      addProvider(llmByProvider, usage.provider, delta);
     },
     recordSTT: (input) => {
-      sttAudioMs += Math.max(0, input.audioMs);
-      if (input.provider)
-        lastSttProvider = input.provider;
+      const audioMs = Math.max(0, input.audioMs);
+      sttAudioMs += audioMs;
       const rates = lookupRates(priceBook, input.provider, input.model)?.stt;
-      if (!rates) {
-        return;
+      let delta = 0;
+      if (rates) {
+        delta = audioMs / 1000 * rates.perSecondUsd;
+        sttUsd += delta;
+      }
+      if (input.provider) {
+        const slice = sliceFor(sttSlices, input.provider);
+        slice.usd += delta;
+        slice.audioMs = (slice.audioMs ?? 0) + audioMs;
       }
-      const delta = Math.max(0, input.audioMs) / 1000 * rates.perSecondUsd;
-      sttUsd += delta;
-      addProvider(sttByProvider, input.provider, delta);
     },
     recordTelephony: (input) => {
-      telephonyMinutes += Math.max(0, input.minutes);
-      if (input.provider)
-        lastTelephonyProvider = input.provider;
+      const minutes = Math.max(0, input.minutes);
+      telephonyMinutes += minutes;
       const rates = lookupRates(priceBook, input.provider)?.telephony;
-      if (!rates) {
-        return;
+      let delta = 0;
+      if (rates) {
+        delta = minutes * rates.perMinuteUsd;
+        telephonyUsd += delta;
+      }
+      if (input.provider) {
+        const slice = sliceFor(telephonySlices, input.provider);
+        slice.usd += delta;
+        slice.minutes = (slice.minutes ?? 0) + minutes;
       }
-      const delta = Math.max(0, input.minutes) * rates.perMinuteUsd;
-      telephonyUsd += delta;
-      addProvider(telephonyByProvider, input.provider, delta);
     },
     recordTTS: (input) => {
       const chars = input.characters ?? 0;
       const audioMs = input.audioMs ?? 0;
       ttsCharacters += chars;
       ttsAudioMs += audioMs;
-      if (input.provider)
-        lastTtsProvider = input.provider;
       const rates = lookupRates(priceBook, input.provider, input.voice)?.tts;
-      if (!rates) {
-        return;
-      }
       let delta = 0;
-      if (rates.perMillionCharactersUsd !== undefined && chars > 0) {
-        delta = chars * rates.perMillionCharactersUsd / 1e6;
-      } else if (rates.perSecondUsd !== undefined && audioMs > 0) {
-        delta = audioMs / 1000 * rates.perSecondUsd;
+      if (rates) {
+        if (rates.perMillionCharactersUsd !== undefined && chars > 0) {
+          delta = chars * rates.perMillionCharactersUsd / 1e6;
+        } else if (rates.perSecondUsd !== undefined && audioMs > 0) {
+          delta = audioMs / 1000 * rates.perSecondUsd;
+        }
+        ttsUsd += delta;
+      }
+      if (input.provider) {
+        const slice = sliceFor(ttsSlices, input.provider);
+        slice.usd += delta;
+        slice.characters = (slice.characters ?? 0) + chars;
+        slice.audioMs = (slice.audioMs ?? 0) + audioMs;
       }
-      ttsUsd += delta;
-      addProvider(ttsByProvider, input.provider, delta);
     },
     snapshot: () => ({
       llm: {
+        byProvider: finalizeSlices(llmSlices),
         cachedInputTokens: llmCachedInput,
         inputTokens: llmInput,
         outputTokens: llmOutput,
-        provider: dominant(llmByProvider, lastLlmProvider),
-        usd: Math.round(llmUsd * 1e6) / 1e6
+        usd: round6(llmUsd)
       },
       sessionId: options.sessionId,
       stt: {
         audioMs: sttAudioMs,
-        provider: dominant(sttByProvider, lastSttProvider),
-        usd: Math.round(sttUsd * 1e6) / 1e6
+        byProvider: finalizeSlices(sttSlices),
+        usd: round6(sttUsd)
       },
       telephony: {
+        byProvider: finalizeSlices(telephonySlices),
         minutes: telephonyMinutes,
-        provider: dominant(telephonyByProvider, lastTelephonyProvider),
-        usd: Math.round(telephonyUsd * 1e6) / 1e6
+        usd: round6(telephonyUsd)
       },
-      totalUsd: Math.round((llmUsd + ttsUsd + sttUsd + telephonyUsd) * 1e6) / 1e6,
+      totalUsd: round6(llmUsd + ttsUsd + sttUsd + telephonyUsd),
       tts: {
         audioMs: ttsAudioMs,
+        byProvider: finalizeSlices(ttsSlices),
         characters: ttsCharacters,
-        provider: dominant(ttsByProvider, lastTtsProvider),
-        usd: Math.round(ttsUsd * 1e6) / 1e6
+        usd: round6(ttsUsd)
       }
     })
   };
@@ -45452,6 +45454,7 @@ var createJSONVoiceAssistantModel = (options) => ({
     return options.mapOutput?.(output) ?? normalizeRouteOutput(output);
   }
 });
+var isEmptyModelOutput = (output) => !output.assistantText?.trim() && !output.toolCalls?.length;
 var createVoiceProviderRouter = (options) => {
   const providerIds = Object.keys(options.providers);
   const firstProvider = providerIds[0];
@@ -45459,6 +45462,7 @@ var createVoiceProviderRouter = (options) => {
   const policy = resolveVoiceProviderRoutingPolicy(options.policy) ?? resolveVoiceProviderRoutingPolicy(orchestrationSurface?.policy);
   const strategy = policy?.strategy ?? "prefer-selected";
   const fallbackMode = policy?.fallbackMode ?? options.fallbackMode ?? orchestrationSurface?.fallbackMode ?? "provider-error";
+  const emptyCompletionRetries = Math.max(0, options.emptyCompletionRetries ?? 1);
   const providerProfiles = {
     ...orchestrationSurface?.providerProfiles ?? {},
     ...options.providerProfiles ?? {}
@@ -45633,6 +45637,24 @@ var createVoiceProviderRouter = (options) => {
       }
     }
   };
+  const runProviderWithEmptyRetry = async (provider, model, input, index, selectedProvider, startedAt) => {
+    let output = await runProvider(provider, model, input);
+    let emptyAttempt = 0;
+    while (emptyAttempt < emptyCompletionRetries && isEmptyModelOutput(output)) {
+      emptyAttempt += 1;
+      await emit({
+        at: Date.now(),
+        attempt: index + 1,
+        elapsedMs: Date.now() - startedAt,
+        emptyCompletion: true,
+        provider,
+        selectedProvider,
+        status: "success"
+      }, input);
+      output = await runProvider(provider, model, input);
+    }
+    return output;
+  };
   return {
     generate: async (input) => {
       const { order, selectedProvider } = await resolveOrder(input);
@@ -45647,7 +45669,7 @@ var createVoiceProviderRouter = (options) => {
         }
         const startedAt = Date.now();
         try {
-          const output = await runProvider(provider, model, input);
+          const output = await runProviderWithEmptyRetry(provider, model, input, index, selectedProvider, startedAt);
           const providerHealth = recordProviderSuccess(provider);
           await emit({
             at: Date.now(),
@@ -45932,10 +45954,21 @@ var finalizeToolCalls = (calls) => [...calls.values()].filter((call) => call.nam
 var consumeOpenAIResponsesStream = async (response, onTextDelta, abortOptions) => {
   let assistantText = "";
   let usage;
+  let status;
+  let incompleteReason;
+  let refusal = "";
+  let failureMessage;
+  let outputItems = 0;
   const calls = new Map;
   await readServerSentEvents(response, (event) => {
     const type = typeof event.type === "string" ? event.type : "";
     const item = event.item;
+    if (type === "response.output_item.added") {
+      outputItems += 1;
+    }
+    if ((type === "response.refusal.delta" || type === "response.refusal.done") && typeof event.delta === "string") {
+      refusal += event.delta;
+    }
     if (type === "response.output_text.delta" && typeof event.delta === "string") {
       assistantText += event.delta;
       onTextDelta?.(event.delta);
@@ -45953,14 +45986,34 @@ var consumeOpenAIResponsesStream = async (response, onTextDelta, abortOptions) =
       const entry = calls.get(String(item.id ?? item.call_id ?? ""));
       if (entry)
         entry.args = item.arguments;
-    } else if (type === "response.completed") {
+    } else if (type === "response.completed" || type === "response.incomplete" || type === "response.failed") {
       const completed = event.response;
       if (completed?.usage && typeof completed.usage === "object") {
         usage = completed.usage;
       }
+      if (typeof completed?.status === "string") {
+        status = completed.status;
+      }
+      const incomplete = completed?.incomplete_details;
+      if (typeof incomplete?.reason === "string") {
+        incompleteReason = incomplete.reason;
+      }
+      const failure = completed?.error;
+      if (typeof failure?.message === "string") {
+        failureMessage = failure.message;
+      }
     }
   }, abortOptions);
-  return { assistantText, toolCalls: finalizeToolCalls(calls), usage };
+  return {
+    assistantText,
+    failureMessage,
+    incompleteReason,
+    outputItems,
+    refusal: refusal || undefined,
+    status,
+    toolCalls: finalizeToolCalls(calls),
+    usage
+  };
 };
 var createOpenAIVoiceAssistantModel = (options) => {
   const fetchImpl = hardenFetch(options.fetch);
@@ -46021,6 +46074,11 @@ var createOpenAIVoiceAssistantModel = (options) => {
       let assistantText;
       let toolCalls;
       let usage;
+      let status;
+      let incompleteReason;
+      let refusal;
+      let failureMessage;
+      let outputItems = 0;
       let firstDeltaSeen = false;
       const onTextDelta = input.onTextDelta ? (delta) => {
         if (!firstDeltaSeen) {
@@ -46030,7 +46088,16 @@ var createOpenAIVoiceAssistantModel = (options) => {
         input.onTextDelta?.(delta);
       } : undefined;
       try {
-        ({ assistantText, toolCalls, usage } = await consumeOpenAIResponsesStream(response, onTextDelta, {
+        ({
+          assistantText,
+          toolCalls,
+          usage,
+          status,
+          incompleteReason,
+          refusal,
+          failureMessage,
+          outputItems
+        } = await consumeOpenAIResponsesStream(response, onTextDelta, {
           signal: ac.signal,
           inactivityMs: 1e4
         }));
@@ -46038,6 +46105,15 @@ var createOpenAIVoiceAssistantModel = (options) => {
           textChars: assistantText?.length ?? 0,
           toolCalls: toolCalls.length
         });
+        if (!assistantText?.trim() && !toolCalls.length) {
+          stamp("openai.empty-completion", {
+            failureMessage: failureMessage ?? "none",
+            incompleteReason: incompleteReason ?? "none",
+            outputItems,
+            refusal: refusal ? refusal.slice(0, 200) : "none",
+            status: status ?? "unknown"
+          });
+        }
       } finally {
         clearTimeout(timer);
       }

package/dist/testing/index.js

@@ -4477,6 +4477,7 @@ var createJSONVoiceAssistantModel = (options) => ({
     return options.mapOutput?.(output) ?? normalizeRouteOutput(output);
   }
 });
+var isEmptyModelOutput = (output) => !output.assistantText?.trim() && !output.toolCalls?.length;
 var createVoiceProviderRouter = (options) => {
   const providerIds = Object.keys(options.providers);
   const firstProvider = providerIds[0];
@@ -4484,6 +4485,7 @@ var createVoiceProviderRouter = (options) => {
   const policy = resolveVoiceProviderRoutingPolicy(options.policy) ?? resolveVoiceProviderRoutingPolicy(orchestrationSurface?.policy);
   const strategy = policy?.strategy ?? "prefer-selected";
   const fallbackMode = policy?.fallbackMode ?? options.fallbackMode ?? orchestrationSurface?.fallbackMode ?? "provider-error";
+  const emptyCompletionRetries = Math.max(0, options.emptyCompletionRetries ?? 1);
   const providerProfiles = {
     ...orchestrationSurface?.providerProfiles ?? {},
     ...options.providerProfiles ?? {}
@@ -4658,6 +4660,24 @@ var createVoiceProviderRouter = (options) => {
       }
     }
   };
+  const runProviderWithEmptyRetry = async (provider, model, input, index, selectedProvider, startedAt) => {
+    let output = await runProvider(provider, model, input);
+    let emptyAttempt = 0;
+    while (emptyAttempt < emptyCompletionRetries && isEmptyModelOutput(output)) {
+      emptyAttempt += 1;
+      await emit({
+        at: Date.now(),
+        attempt: index + 1,
+        elapsedMs: Date.now() - startedAt,
+        emptyCompletion: true,
+        provider,
+        selectedProvider,
+        status: "success"
+      }, input);
+      output = await runProvider(provider, model, input);
+    }
+    return output;
+  };
   return {
     generate: async (input) => {
       const { order, selectedProvider } = await resolveOrder(input);
@@ -4672,7 +4692,7 @@ var createVoiceProviderRouter = (options) => {
         }
         const startedAt = Date.now();
         try {
-          const output = await runProvider(provider, model, input);
+          const output = await runProviderWithEmptyRetry(provider, model, input, index, selectedProvider, startedAt);
           const providerHealth = recordProviderSuccess(provider);
           await emit({
             at: Date.now(),
@@ -4957,10 +4977,21 @@ var finalizeToolCalls = (calls) => [...calls.values()].filter((call) => call.nam
 var consumeOpenAIResponsesStream = async (response, onTextDelta, abortOptions) => {
   let assistantText = "";
   let usage;
+  let status;
+  let incompleteReason;
+  let refusal = "";
+  let failureMessage;
+  let outputItems = 0;
   const calls = new Map;
   await readServerSentEvents(response, (event) => {
     const type = typeof event.type === "string" ? event.type : "";
     const item = event.item;
+    if (type === "response.output_item.added") {
+      outputItems += 1;
+    }
+    if ((type === "response.refusal.delta" || type === "response.refusal.done") && typeof event.delta === "string") {
+      refusal += event.delta;
+    }
     if (type === "response.output_text.delta" && typeof event.delta === "string") {
       assistantText += event.delta;
       onTextDelta?.(event.delta);
@@ -4978,14 +5009,34 @@ var consumeOpenAIResponsesStream = async (response, onTextDelta, abortOptions) =
       const entry = calls.get(String(item.id ?? item.call_id ?? ""));
       if (entry)
         entry.args = item.arguments;
-    } else if (type === "response.completed") {
+    } else if (type === "response.completed" || type === "response.incomplete" || type === "response.failed") {
       const completed = event.response;
       if (completed?.usage && typeof completed.usage === "object") {
         usage = completed.usage;
       }
+      if (typeof completed?.status === "string") {
+        status = completed.status;
+      }
+      const incomplete = completed?.incomplete_details;
+      if (typeof incomplete?.reason === "string") {
+        incompleteReason = incomplete.reason;
+      }
+      const failure = completed?.error;
+      if (typeof failure?.message === "string") {
+        failureMessage = failure.message;
+      }
     }
   }, abortOptions);
-  return { assistantText, toolCalls: finalizeToolCalls(calls), usage };
+  return {
+    assistantText,
+    failureMessage,
+    incompleteReason,
+    outputItems,
+    refusal: refusal || undefined,
+    status,
+    toolCalls: finalizeToolCalls(calls),
+    usage
+  };
 };
 var createOpenAIVoiceAssistantModel = (options) => {
   const fetchImpl = hardenFetch(options.fetch);
@@ -5046,6 +5097,11 @@ var createOpenAIVoiceAssistantModel = (options) => {
       let assistantText;
       let toolCalls;
       let usage;
+      let status;
+      let incompleteReason;
+      let refusal;
+      let failureMessage;
+      let outputItems = 0;
       let firstDeltaSeen = false;
       const onTextDelta = input.onTextDelta ? (delta) => {
         if (!firstDeltaSeen) {
@@ -5055,7 +5111,16 @@ var createOpenAIVoiceAssistantModel = (options) => {
         input.onTextDelta?.(delta);
       } : undefined;
       try {
-        ({ assistantText, toolCalls, usage } = await consumeOpenAIResponsesStream(response, onTextDelta, {
+        ({
+          assistantText,
+          toolCalls,
+          usage,
+          status,
+          incompleteReason,
+          refusal,
+          failureMessage,
+          outputItems
+        } = await consumeOpenAIResponsesStream(response, onTextDelta, {
           signal: ac.signal,
           inactivityMs: 1e4
         }));
@@ -5063,6 +5128,15 @@ var createOpenAIVoiceAssistantModel = (options) => {
           textChars: assistantText?.length ?? 0,
           toolCalls: toolCalls.length
         });
+        if (!assistantText?.trim() && !toolCalls.length) {
+          stamp("openai.empty-completion", {
+            failureMessage: failureMessage ?? "none",
+            incompleteReason: incompleteReason ?? "none",
+            outputItems,
+            refusal: refusal ? refusal.slice(0, 200) : "none",
+            status: status ?? "unknown"
+          });
+        }
       } finally {
         clearTimeout(timer);
       }

package/dist/vue/VoiceProviderSimulationControls.d.ts

@@ -79,10 +79,10 @@ export declare const VoiceProviderSimulationControls: import("vue").DefineCompon
 }>> & Readonly<{}>, {
     kind: string;
     title: string;
+    failureMessage: string;
     fallbackRequiredProvider: string;
     fallbackRequiredMessage: string;
     failureProviders: readonly string[] | undefined;
-    failureMessage: string;
     pathPrefix: string;
     class: string;
 }, {}, {}, {}, string, import("vue").ComponentProvideOptions, true, {}, any>;

package/package.json

@@ -1,11 +1,15 @@
 {
   "name": "@absolutejs/voice",
-  "version": "0.0.22-beta.615",
+  "version": "0.0.22-beta.617",
   "description": "Voice primitives and Elysia plugin for AbsoluteJS",
   "repository": {
     "type": "git",
     "url": "https://github.com/absolutejs/voice.git"
   },
+  "homepage": "https://github.com/absolutejs/voice",
+  "bugs": {
+    "url": "https://github.com/absolutejs/voice/issues"
+  },
   "files": [
     "dist",
     "README.md"