Merge pull request #33 from inkbox-ai/fix/hosted-voice-delivery

Stabilize hosted voice delivery and update dependencies

Author: alex-w-99

.github/workflows/live-voice.yml

@@ -94,6 +94,7 @@ jobs:
             HOSTED_MARKER="$(node scripts/nato-marker.mjs "$GITHUB_RUN_ID" "$GITHUB_RUN_ATTEMPT")"
             echo "HOSTED_POST_CALL_MARKER=$HOSTED_MARKER" >> "$GITHUB_ENV"
             export VOICE_DRIVER_LINE="After we hang up, send me one SMS containing exactly these three words: $HOSTED_MARKER. Create one post-call action now. Set both the action title and the action details to this exact five-word phrase: Send SMS $HOSTED_MARKER. Wait for the action tool to succeed, then read the exact three-word SMS body back to me. Do not paraphrase, omit a word, or send the SMS during the call."
+            export VOICE_DRIVER_QUIET_AFTER_TRANSCRIPT=2
             export VOICE_DRIVER_LISTEN=180
             export VOICE_DRIVER_AUTO_STOP=false
           fi

CHANGELOG.md

@@ -2,6 +2,8 @@
 
 ## Unreleased
 
+- Stabilizes hosted voice checks by waiting for a quiet greeting and verifying the caller request on both call legs.
+- Updates transitive HTTP and development dependencies to patched versions.
 - Runs live test files sequentially to avoid interference between channel checks.
 - Adds negotiated 16 kHz PCM call audio with streaming resampling and legacy call compatibility.
 - Uses the published, pinned SDK in CI and gives periodic progress delivery time to finish before the live task ends.

package-lock.json

@@ -1,6 +1,6 @@
 {
   "name": "@inkbox/opencode-plugin",
-  "version": "0.2.12",
+  "version": "0.2.13",
   "private": true,
   "description": "Inkbox for opencode \u2014 give your agent an email address, a phone number (SMS/MMS + voice), iMessage, contacts, notes, and an encrypted credential vault.",
   "license": "MIT",

package.json

@@ -4,6 +4,7 @@ import { describe, expect, it } from "vitest";
 const liveAut = readFileSync("scripts/live-aut.sh", "utf8");
 const liveChannels = readFileSync(".github/workflows/live-channels.yml", "utf8");
 const liveVoice = readFileSync(".github/workflows/live-voice.yml", "utf8");
+const voiceDriver = readFileSync("tests/live/voice-driver.mjs", "utf8");
 const liveStack = readFileSync(".github/workflows/live-stack.yml", "utf8");
 
 function shellCommands(source: string): string[] {
@@ -63,6 +64,12 @@ describe("live harness readiness bounds", () => {
     );
   });
 
+  it("waits for the hosted greeting to go quiet before speaking", () => {
+    expect(liveVoice).toContain("export VOICE_DRIVER_QUIET_AFTER_TRANSCRIPT=2");
+    expect(voiceDriver).toContain("lastTranscriptAt + QUIET_AFTER_TRANSCRIPT_MS");
+    expect(voiceDriver).not.toContain("speak now if the greeting beat our timer");
+  });
+
   it("keeps failure diagnostics content-free and out of public artifacts", () => {
     for (const workflow of [liveChannels, liveVoice]) {
       expect(workflow).toContain("Report content-free failure state");

tests/contract/live-harness.test.ts

@@ -11,7 +11,8 @@
 // file (ws url + phone-number id) the test reads.
 //
 // Env: REMOTE_INKBOX_API_KEY, INKBOX_BASE_URL, VOICE_DRIVER_STATE,
-//      VOICE_DRIVER_LINE, VOICE_DRIVER_SPEAK_AFTER (s), VOICE_DRIVER_LISTEN (s),
+//      VOICE_DRIVER_LINE, VOICE_DRIVER_SPEAK_AFTER (s),
+//      VOICE_DRIVER_QUIET_AFTER_TRANSCRIPT (s), VOICE_DRIVER_LISTEN (s),
 //      VOICE_DRIVER_AUTO_STOP (false lets the test own hangup timing)
 import { writeFileSync } from "node:fs";
 import { Inkbox } from "@inkbox/sdk";
@@ -32,6 +33,8 @@ const GREETING = process.env.VOICE_DRIVER_GREETING || "Hello?";
 // then give the agent a turn and hang up (a dropped WS does NOT end the call — an
 // explicit stop is required or the leg lingers to the server max-duration cap).
 const SPEAK_AFTER_MS = Number(process.env.VOICE_DRIVER_SPEAK_AFTER || "5") * 1000;
+const QUIET_AFTER_TRANSCRIPT_MS =
+  Number(process.env.VOICE_DRIVER_QUIET_AFTER_TRANSCRIPT || "0") * 1000;
 const LISTEN_MS = Number(process.env.VOICE_DRIVER_LISTEN || "12") * 1000;
 const AUTO_STOP = process.env.VOICE_DRIVER_AUTO_STOP !== "false";
 
@@ -58,6 +61,7 @@ async function callWsHandler(ws) {
   });
   console.log("call WS accepted");
   let spoke = false;
+  let lastTranscriptAt = 0;
   const say = async (text) => {
     await ws.send(JSON.stringify({ event: "text", delta: text }));
     await ws.send(JSON.stringify({ event: "text", done: true }));
@@ -70,7 +74,13 @@ async function callWsHandler(ws) {
   };
   const runTurn = async () => {
     await say(GREETING);
-    await sleep(SPEAK_AFTER_MS);
+    const earliestSpeechAt = Date.now() + SPEAK_AFTER_MS;
+    while (!spoke) {
+      const speechAt = Math.max(earliestSpeechAt, lastTranscriptAt + QUIET_AFTER_TRANSCRIPT_MS);
+      const delay = speechAt - Date.now();
+      if (delay <= 0) break;
+      await sleep(delay);
+    }
     await speak(LINE);
     await sleep(LISTEN_MS);
     if (!AUTO_STOP) return;
@@ -92,9 +102,9 @@ async function callWsHandler(ws) {
       if (ev.event === "start") {
         console.log("call start");
         void runTurn();
-      } else if (ev.event === "transcript" && ev.is_final) {
-        console.log("heard (final):", ev.text);
-        await speak(LINE); // speak now if the greeting beat our timer
+      } else if (ev.event === "transcript") {
+        lastTranscriptAt = Date.now();
+        if (ev.is_final) console.log("heard (final):", ev.text);
       } else if (ev.event === "stop") {
         console.log("call stop");
         break;

tests/live/voice-driver.mjs

@@ -452,14 +452,21 @@ describe.skipIf(!LIVE || !REAL_MODEL)("live voice", () => {
             .join(" ")
             .toLowerCase()
             .replace(/[^a-z0-9]+/g, " ");
+          const autCaller = autSegments.remote
+            .join(" ")
+            .toLowerCase()
+            .replace(/[^a-z0-9]+/g, " ");
           const openActions = (currentAutCall.postCallActionItems ?? []).filter(
             (item: any) => String(item.status ?? "").toLowerCase() === "open",
           );
           const actionEvidence = openActions.map((item: any) =>
             [item.action, item.details].filter(Boolean).join(" "),
           );
-          const callerReady =
+          const driverCallerReady =
             hasAfterCallSmsIntent(caller) && containsVoiceMarker(caller, HOSTED_MARKER);
+          const autCallerReady =
+            hasAfterCallSmsIntent(autCaller) && containsVoiceMarker(autCaller, HOSTED_MARKER);
+          const callerReady = driverCallerReady && autCallerReady;
           const matchingActions = actionEvidence.filter(
             (value: string) => hasSmsIntent(value) && containsVoiceMarker(value, HOSTED_MARKER),
           );
@@ -477,7 +484,8 @@ describe.skipIf(!LIVE || !REAL_MODEL)("live voice", () => {
             markerActionCount === 1;
           progress.last =
             `agent_segments=${autSegments.local.length} two_way_ready=${twoWayReady} ` +
-            `caller_ready=${callerReady} ` +
+            `caller_ready=${callerReady} driver_caller_ready=${driverCallerReady} ` +
+            `aut_caller_ready=${autCallerReady} ` +
             `action_ready=${actionReady} open_actions=${openActions.length} ` +
             `sms_actions=${smsActionCount} marker_actions=${markerActionCount}`;
           if (twoWayReady && callerReady && actionReady) break;