@absolutejs/voice 0.0.22-beta.582 → 0.0.22-beta.584

package/dist/core/debugTiming.d.ts

@@ -0,0 +1,11 @@
+export declare const logVoiceTiming: (sessionId: string, stage: string, elapsedMs: number, detail?: Record<string, unknown>) => void;
+/**
+ * A per-turn stopwatch. `stamp(stage, detail?)` logs the elapsed since the timer
+ * was created, so one timer per turn lays every stage out on a single timeline:
+ *
+ *   const stamp = startVoiceTimer(session.id);
+ *   stamp("agent.system-resolved", { chars });
+ *   stamp("agent.round0.generate-done", { ms });
+ */
+export declare const startVoiceTimer: (sessionId: string) => (stage: string, detail?: Record<string, unknown>) => void;
+export declare const voiceTimingEnabled: () => boolean;

package/dist/core/hardenedFetch.d.ts

@@ -0,0 +1,3 @@
+export declare const hardenFetch: (baseFetch?: typeof fetch) => ((input: Parameters<typeof fetch>[0], init: Parameters<typeof fetch>[1]) => Promise<Response>) & {
+    preconnect: typeof fetch.preconnect;
+};

package/dist/index.d.ts

@@ -71,6 +71,8 @@ export { createVoiceSessionListRoutes, createVoiceSessionReplayHTMLHandler, crea
 export { createVoiceAgent, createVoiceAgentSquad, createVoiceAgentTool, } from "./core/agent";
 export { createPersonaVoiceCaller, createScriptedVoiceCaller, renderVoiceSimulationTranscript, runVoiceConversationSimulation, } from "./core/conversationSimulator";
 export type { RunVoiceConversationSimulationInput, VoiceConversationSimulationEndedReason, VoiceConversationSimulationResult, VoicePersonaCallerCompletion, VoiceScriptedCallerStep, VoiceSimulatedSpeaker, VoiceSimulatedTurn, VoiceSimulatorCaller, VoiceSimulatorCallerModel, VoiceSimulatorCallerReply, } from "./core/conversationSimulator";
+export { logVoiceTiming, startVoiceTimer, voiceTimingEnabled, } from "./core/debugTiming";
+export { hardenFetch } from "./core/hardenedFetch";
 export { createVoiceMCPToolset } from "./core/mcpToolset";
 export type { CreateVoiceMCPToolsetOptions, MCPClientLike, MCPToolCallResult, MCPToolContentBlock, MCPToolDefinition, VoiceMCPToolResult, } from "./core/mcpToolset";
 export { createAIVoiceModel } from "./core/aiVoiceModel";

package/dist/index.js

@@ -3091,6 +3091,21 @@ var toVoiceSessionSummary = (session) => ({
 // src/core/session.ts
 import { Buffer as Buffer2 } from "buffer";
 
+// src/core/debugTiming.ts
+var timingEnabled = () => process.env.ABSOLUTEJS_VOICE_TIMING === "1" || process.env.ABSOLUTEJS_VOICE_TIMING === "true";
+var emitTiming = (sessionId, stage, elapsedMs, detail) => {
+  if (!timingEnabled())
+    return;
+  const extra = detail ? ` ${JSON.stringify(detail)}` : "";
+  console.info(`[voice][timing] session=${sessionId} ${stage} +${Math.round(elapsedMs)}ms${extra}`);
+};
+var logVoiceTiming = (sessionId, stage, elapsedMs, detail) => emitTiming(sessionId, stage, elapsedMs, detail);
+var startVoiceTimer = (sessionId) => {
+  const startedAt = Date.now();
+  return (stage, detail) => emitTiming(sessionId, stage, Date.now() - startedAt, detail);
+};
+var voiceTimingEnabled = () => timingEnabled();
+
 // src/core/backchannel.ts
 var DEFAULT_CUES = [
   { text: "mm-hmm" },
@@ -5534,6 +5549,7 @@ var createVoiceSession = (options) => {
     const onTurnTimeoutMs = options.routeOnTurnTimeoutMs ?? 45000;
     let committedOutput;
     const onTurnStartedAt = Date.now();
+    logVoiceTiming(session.id, "session.commit-to-onturn", onTurnStartedAt - (turn.committedAt || onTurnStartedAt), { fillerScheduled: fillerTimer !== null });
     try {
       const onTurnPromise = options.route.onTurn({
         api,
@@ -7446,7 +7462,9 @@ var createVoiceAgent = (options) => {
   const maxToolRounds = Math.max(0, options.maxToolRounds ?? 2);
   const audit = resolveVoiceAgentAuditLogger(options.audit);
   const run = async (input) => {
+    const stamp = startVoiceTimer(input.session.id);
     const messages = input.messages ?? createHistoryMessages(input.session, input.turn);
+    stamp("agent.history-built", { messages: messages.length });
     const toolResults = [];
     const baseSystem = typeof options.system === "function" ? await options.system({
       context: input.context,
@@ -7456,9 +7474,11 @@ var createVoiceAgent = (options) => {
     const system = [baseSystem, input.system].filter((value) => Boolean(value?.trim())).join(`
 
 `) || undefined;
+    stamp("agent.system-resolved", { systemChars: system?.length ?? 0 });
     let output = {};
     for (let round = 0;round <= maxToolRounds; round += 1) {
       const modelStartedAt = Date.now();
+      stamp(`agent.round${round}.generate-start`);
       try {
         output = await options.model.generate({
           agentId: options.id,
@@ -7474,6 +7494,11 @@ var createVoiceAgent = (options) => {
           })),
           turn: input.turn
         });
+        stamp(`agent.round${round}.generate-done`, {
+          ms: Date.now() - modelStartedAt,
+          textChars: output.assistantText?.length ?? 0,
+          toolCalls: output.toolCalls?.length ?? 0
+        });
         await audit?.providerCall({
           actor: {
             id: options.id,
@@ -7487,6 +7512,9 @@ var createVoiceAgent = (options) => {
           sessionId: input.session.id
         });
       } catch (error) {
+        stamp(`agent.round${round}.generate-error`, {
+          ms: Date.now() - modelStartedAt
+        });
         await audit?.providerCall({
           actor: {
             id: options.id,
@@ -40814,6 +40842,44 @@ Respond with only your spoken line. When your goal is met or you want to hang up
     persona: options.persona
   };
 };
+// src/core/hardenedFetch.ts
+var ATTEMPT_TIMEOUT_MS = 6000;
+var isBun = "Bun" in globalThis;
+var oneAttempt = async (baseFetch, input, init) => {
+  const controller = new AbortController;
+  const callerSignal = init?.signal ?? undefined;
+  const onCallerAbort = () => controller.abort(callerSignal?.reason);
+  if (callerSignal?.aborted)
+    controller.abort(callerSignal.reason);
+  else
+    callerSignal?.addEventListener("abort", onCallerAbort, { once: true });
+  const timer = setTimeout(() => {
+    controller.abort(new Error(`fetch exceeded ${ATTEMPT_TIMEOUT_MS}ms before response headers (stale Bun keep-alive socket?)`));
+  }, ATTEMPT_TIMEOUT_MS);
+  const headers = new Headers(init?.headers);
+  if (isBun)
+    headers.set("Connection", "close");
+  try {
+    return await baseFetch(input, {
+      ...init,
+      headers,
+      signal: controller.signal
+    });
+  } finally {
+    clearTimeout(timer);
+    callerSignal?.removeEventListener("abort", onCallerAbort);
+  }
+};
+var hardenFetch = (baseFetch = globalThis.fetch) => Object.assign(async (input, init) => {
+  try {
+    return await oneAttempt(baseFetch, input, init);
+  } catch (error) {
+    if (init?.signal?.aborted)
+      throw error;
+    console.warn(`[voice] hardened fetch retrying on a fresh connection: ${error instanceof Error ? error.message : String(error)}`);
+    return oneAttempt(baseFetch, input, init);
+  }
+}, { preconnect: baseFetch.preconnect.bind(baseFetch) });
 // src/core/mcpToolset.ts
 var flattenContent = (result) => {
   const blocks = result.content ?? [];
@@ -45362,12 +45428,13 @@ var consumeOpenAIResponsesStream = async (response, onTextDelta, abortOptions) =
   return { assistantText, toolCalls: finalizeToolCalls(calls), usage };
 };
 var createOpenAIVoiceAssistantModel = (options) => {
-  const fetchImpl = options.fetch ?? globalThis.fetch;
+  const fetchImpl = hardenFetch(options.fetch);
   const baseUrl = options.baseUrl ?? "https://api.openai.com/v1";
   const model = options.model ?? "gpt-4.1-mini";
   const timeoutMs = options.timeoutMs ?? 60000;
   return {
     generate: async (input) => {
+      const stamp = startVoiceTimer(input.session.id);
       const ac = new AbortController;
       const timer = setTimeout(() => {
         ac.abort(new Error(`OpenAI /responses timed out after ${timeoutMs}ms (no completion event received)`));
@@ -45408,6 +45475,10 @@ var createOpenAIVoiceAssistantModel = (options) => {
         clearTimeout(timer);
         throw error;
       }
+      stamp("openai.fetch-returned", {
+        messages: input.messages.length,
+        status: response.status
+      });
       if (!response.ok) {
         clearTimeout(timer);
         throw createHTTPError("OpenAI", response);
@@ -45415,11 +45486,23 @@ var createOpenAIVoiceAssistantModel = (options) => {
       let assistantText;
       let toolCalls;
       let usage;
+      let firstDeltaSeen = false;
+      const onTextDelta = input.onTextDelta ? (delta) => {
+        if (!firstDeltaSeen) {
+          firstDeltaSeen = true;
+          stamp("openai.first-delta");
+        }
+        input.onTextDelta?.(delta);
+      } : undefined;
       try {
-        ({ assistantText, toolCalls, usage } = await consumeOpenAIResponsesStream(response, input.onTextDelta, {
+        ({ assistantText, toolCalls, usage } = await consumeOpenAIResponsesStream(response, onTextDelta, {
           signal: ac.signal,
           inactivityMs: 1e4
         }));
+        stamp("openai.stream-done", {
+          textChars: assistantText?.length ?? 0,
+          toolCalls: toolCalls.length
+        });
       } finally {
         clearTimeout(timer);
       }
@@ -45470,7 +45553,7 @@ var consumeAnthropicStream = async (response, onTextDelta) => {
   return { assistantText, toolCalls: finalizeToolCalls(calls), usage };
 };
 var createAnthropicVoiceAssistantModel = (options) => {
-  const fetchImpl = options.fetch ?? globalThis.fetch;
+  const fetchImpl = hardenFetch(options.fetch);
   const baseUrl = options.baseUrl ?? "https://api.anthropic.com/v1";
   const model = options.model ?? "claude-sonnet-4-5";
   return {
@@ -45556,7 +45639,7 @@ var consumeGeminiStream = async (response, onTextDelta) => {
   return { assistantText, toolCalls, usage };
 };
 var createGeminiVoiceAssistantModel = (options) => {
-  const fetchImpl = options.fetch ?? globalThis.fetch;
+  const fetchImpl = hardenFetch(options.fetch);
   const baseUrl = options.baseUrl ?? "https://generativelanguage.googleapis.com/v1beta";
   const model = options.model ?? "gemini-2.5-flash";
   const maxRetries = Math.max(0, options.maxRetries ?? 2);
@@ -52391,6 +52474,7 @@ export {
   wrapVoiceHTMLInHTMXContainer,
   withVoiceOpsTaskId,
   withVoiceIntegrationEventId,
+  voiceTimingEnabled,
   voiceTelephonyOutcomeToRouteResult,
   voiceObservabilityExportSchemaVersion,
   voiceObservabilityExportSchemaId,
@@ -52448,6 +52532,7 @@ export {
   summarizeVoiceAuditSinkDeliveries,
   summarizeVoiceAssistantRuns,
   summarizeVoiceAssistantHealth,
+  startVoiceTimer,
   startVoiceOpsTask,
   signVoiceWebhookBody,
   signVoiceTwilioWebhook,
@@ -52648,6 +52733,7 @@ export {
   matchesVoiceOpsTaskAssignmentRule,
   markVoiceOpsTaskSLABreached,
   mapVoiceProofTargetsWithConcurrency,
+  logVoiceTiming,
   loadVoiceRealCallProfileEvidenceFromTraceStore,
   loadVoiceRealCallProfileEvidenceFromStore,
   loadVoiceObservabilityExportReplaySource,
@@ -52663,6 +52749,7 @@ export {
   importVoiceCampaignRecipients,
   heartbeatVoiceOpsTask,
   hasVoiceOpsTaskSLABreach,
+  hardenFetch,
   getVoiceProofTargetLogicalFailure,
   getVoiceLiveOpsControlStatus,
   getVoiceCampaignDialerProofStatus,

package/dist/testing/index.js

@@ -4195,6 +4195,60 @@ var createVoiceIOProviderFailureSimulator = (options) => {
     run
   };
 };
+// src/core/debugTiming.ts
+var timingEnabled = () => process.env.ABSOLUTEJS_VOICE_TIMING === "1" || process.env.ABSOLUTEJS_VOICE_TIMING === "true";
+var emitTiming = (sessionId, stage, elapsedMs, detail) => {
+  if (!timingEnabled())
+    return;
+  const extra = detail ? ` ${JSON.stringify(detail)}` : "";
+  console.info(`[voice][timing] session=${sessionId} ${stage} +${Math.round(elapsedMs)}ms${extra}`);
+};
+var logVoiceTiming = (sessionId, stage, elapsedMs, detail) => emitTiming(sessionId, stage, elapsedMs, detail);
+var startVoiceTimer = (sessionId) => {
+  const startedAt = Date.now();
+  return (stage, detail) => emitTiming(sessionId, stage, Date.now() - startedAt, detail);
+};
+var voiceTimingEnabled = () => timingEnabled();
+
+// src/core/hardenedFetch.ts
+var ATTEMPT_TIMEOUT_MS = 6000;
+var isBun = "Bun" in globalThis;
+var oneAttempt = async (baseFetch, input, init) => {
+  const controller = new AbortController;
+  const callerSignal = init?.signal ?? undefined;
+  const onCallerAbort = () => controller.abort(callerSignal?.reason);
+  if (callerSignal?.aborted)
+    controller.abort(callerSignal.reason);
+  else
+    callerSignal?.addEventListener("abort", onCallerAbort, { once: true });
+  const timer = setTimeout(() => {
+    controller.abort(new Error(`fetch exceeded ${ATTEMPT_TIMEOUT_MS}ms before response headers (stale Bun keep-alive socket?)`));
+  }, ATTEMPT_TIMEOUT_MS);
+  const headers = new Headers(init?.headers);
+  if (isBun)
+    headers.set("Connection", "close");
+  try {
+    return await baseFetch(input, {
+      ...init,
+      headers,
+      signal: controller.signal
+    });
+  } finally {
+    clearTimeout(timer);
+    callerSignal?.removeEventListener("abort", onCallerAbort);
+  }
+};
+var hardenFetch = (baseFetch = globalThis.fetch) => Object.assign(async (input, init) => {
+  try {
+    return await oneAttempt(baseFetch, input, init);
+  } catch (error) {
+    if (init?.signal?.aborted)
+      throw error;
+    console.warn(`[voice] hardened fetch retrying on a fresh connection: ${error instanceof Error ? error.message : String(error)}`);
+    return oneAttempt(baseFetch, input, init);
+  }
+}, { preconnect: baseFetch.preconnect.bind(baseFetch) });
+
 // src/core/modelAdapters.ts
 var isVoiceProviderRoutingPolicyPreset = (value) => value === "balanced" || value === "cost-cap" || value === "cost-first" || value === "latency-first" || value === "quality-first";
 var resolveVoiceProviderRoutingPolicyPreset = (preset, options = {}) => {
@@ -4899,12 +4953,13 @@ var consumeOpenAIResponsesStream = async (response, onTextDelta, abortOptions) =
   return { assistantText, toolCalls: finalizeToolCalls(calls), usage };
 };
 var createOpenAIVoiceAssistantModel = (options) => {
-  const fetchImpl = options.fetch ?? globalThis.fetch;
+  const fetchImpl = hardenFetch(options.fetch);
   const baseUrl = options.baseUrl ?? "https://api.openai.com/v1";
   const model = options.model ?? "gpt-4.1-mini";
   const timeoutMs = options.timeoutMs ?? 60000;
   return {
     generate: async (input) => {
+      const stamp = startVoiceTimer(input.session.id);
       const ac = new AbortController;
       const timer = setTimeout(() => {
         ac.abort(new Error(`OpenAI /responses timed out after ${timeoutMs}ms (no completion event received)`));
@@ -4945,6 +5000,10 @@ var createOpenAIVoiceAssistantModel = (options) => {
         clearTimeout(timer);
         throw error;
       }
+      stamp("openai.fetch-returned", {
+        messages: input.messages.length,
+        status: response.status
+      });
       if (!response.ok) {
         clearTimeout(timer);
         throw createHTTPError("OpenAI", response);
@@ -4952,11 +5011,23 @@ var createOpenAIVoiceAssistantModel = (options) => {
       let assistantText;
       let toolCalls;
       let usage;
+      let firstDeltaSeen = false;
+      const onTextDelta = input.onTextDelta ? (delta) => {
+        if (!firstDeltaSeen) {
+          firstDeltaSeen = true;
+          stamp("openai.first-delta");
+        }
+        input.onTextDelta?.(delta);
+      } : undefined;
       try {
-        ({ assistantText, toolCalls, usage } = await consumeOpenAIResponsesStream(response, input.onTextDelta, {
+        ({ assistantText, toolCalls, usage } = await consumeOpenAIResponsesStream(response, onTextDelta, {
           signal: ac.signal,
           inactivityMs: 1e4
         }));
+        stamp("openai.stream-done", {
+          textChars: assistantText?.length ?? 0,
+          toolCalls: toolCalls.length
+        });
       } finally {
         clearTimeout(timer);
       }
@@ -5007,7 +5078,7 @@ var consumeAnthropicStream = async (response, onTextDelta) => {
   return { assistantText, toolCalls: finalizeToolCalls(calls), usage };
 };
 var createAnthropicVoiceAssistantModel = (options) => {
-  const fetchImpl = options.fetch ?? globalThis.fetch;
+  const fetchImpl = hardenFetch(options.fetch);
   const baseUrl = options.baseUrl ?? "https://api.anthropic.com/v1";
   const model = options.model ?? "claude-sonnet-4-5";
   return {
@@ -5093,7 +5164,7 @@ var consumeGeminiStream = async (response, onTextDelta) => {
   return { assistantText, toolCalls, usage };
 };
 var createGeminiVoiceAssistantModel = (options) => {
-  const fetchImpl = options.fetch ?? globalThis.fetch;
+  const fetchImpl = hardenFetch(options.fetch);
   const baseUrl = options.baseUrl ?? "https://generativelanguage.googleapis.com/v1beta";
   const model = options.model ?? "gemini-2.5-flash";
   const maxRetries = Math.max(0, options.maxRetries ?? 2);
@@ -7676,6 +7747,7 @@ var createVoiceSession = (options) => {
     const onTurnTimeoutMs = options.routeOnTurnTimeoutMs ?? 45000;
     let committedOutput;
     const onTurnStartedAt = Date.now();
+    logVoiceTiming(session.id, "session.commit-to-onturn", onTurnStartedAt - (turn.committedAt || onTurnStartedAt), { fillerScheduled: fillerTimer !== null });
     try {
       const onTurnPromise = options.route.onTurn({
         api,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@absolutejs/voice",
-  "version": "0.0.22-beta.582",
+  "version": "0.0.22-beta.584",
   "description": "Voice primitives and Elysia plugin for AbsoluteJS",
   "repository": {
     "type": "git",