@pyai/sdk 0.6.1 → 0.7.1

package/AGENT_GUIDE.md

@@ -11,7 +11,7 @@ Markdown and does not require the marketing website renderer.
 ## Availability and sources of truth
 
 The CLI is included in `@pyai/sdk` 0.6.1. Install it with
-`npm install -g @pyai/sdk@0.6.1`. Browser login supports existing PyAI accounts;
+`npm install -g @pyai/sdk@0.7.1`. Browser login supports existing PyAI accounts;
 unattended agents should use an authorized environment key. Inspect the
 installed command schema before acting.
 
@@ -35,7 +35,7 @@ do not invent an endpoint or present the feature as deployed.
 Install the published CLI with Node.js 22 or newer recommended:
 
 ```bash
-npm install -g @pyai/sdk@0.6.1
+npm install -g @pyai/sdk@0.7.1
 pyai --version
 pyai schema --json
 ```
@@ -242,6 +242,7 @@ Treat that file as sensitive; never commit it or paste it into a model prompt.
 
 ```bash
 pyai init voice-project
+pyai init voice-omni --template omni
 pyai init voice-ts --template typescript
 pyai init voice-python --template python
 pyai init voice-project --template agent --dry-run --json
@@ -252,6 +253,8 @@ present. Scaffolding is offline; it creates `PYAI.md` and starter assets without
 installing packages, minting credentials, or making API calls. Every template includes `README.md`, `PYAI.md`, `.env.example`, and
 `.gitignore`. The agent template adds `agent.json`, `speech.json`, and
 `job.json`; TypeScript adds `main.ts` and `package.json`; Python adds `main.py`.
+Omni adds a complete Node 20.19+/22 ESM example with `ws`, one paced audio
+stream, a read-only tool, captured WAVs, and simulated interruption handling.
 Follow the created README. JSON output includes `directory`, `template`, `files`,
 `created`, and `next_steps`.
 

package/CLI.md

@@ -7,7 +7,7 @@ and Codex. `--json` provides structured results; `--help` and `schema` describe
 the installed command surface.
 
 **Version:** the expanded CLI is included in `@pyai/sdk` 0.6.1. Install it from
-npm with `npm install -g @pyai/sdk@0.6.1`, then run `pyai login` for browser
+npm with `npm install -g @pyai/sdk@0.7.1`, then run `pyai login` for browser
 sign-in or provide `PYAI_API_KEY` for unattended automation.
 
 ## Contents
@@ -31,7 +31,7 @@ sign-in or provide `PYAI_API_KEY` for unattended automation.
 Install the published package (Node.js 22 or newer recommended):
 
 ```bash
-npm install -g @pyai/sdk@0.6.1
+npm install -g @pyai/sdk@0.7.1
 pyai --version
 pyai login
 pyai speak "Hello from PyAI." -o hello.wav
@@ -436,7 +436,7 @@ The offline schema is authoritative for the installed build.
 | Hear vocabulary | `vocabulary get`, `set` |
 | AMD | `amd calls`, `get ID`, `config`, `configure` |
 | Discovery | `schema`, `schema --openapi`, `recipes [NAME]`, `help` |
-| Project setup | `init DIRECTORY --template agent\|typescript\|python` |
+| Project setup | `init DIRECTORY --template agent\|typescript\|python\|omni` |
 | Generic API | `request METHOD /PATH` |
 | Diagnostics | `doctor`, `smoke` |
 
@@ -483,6 +483,7 @@ Create a starter directory without credentials or network calls:
 
 ```bash
 pyai init voice-project
+pyai init voice-omni --template omni
 pyai init voice-ts --template typescript
 pyai init voice-python --template python
 pyai init voice-project --template agent --dry-run -j
@@ -493,6 +494,8 @@ existing parent; scaffolding
 does not merge into or overwrite an existing project. Every template includes `README.md`, `PYAI.md`, `.env.example`, and
 `.gitignore`. The agent template adds `agent.json`, `speech.json`, and
 `job.json`; TypeScript adds `main.ts` and `package.json`; Python adds `main.py`.
+Omni adds a complete Node 20.19+/22 ESM example with `ws`, one paced audio
+stream, a read-only tool, captured WAVs, and simulated interruption handling.
 `PYAI.md` provides an integration handoff for an engineer or coding agent. Follow the
 starter's README for its files, runtime prerequisites, and next commands.
 Scaffolding does not install dependencies, create an account, or make paid API

package/CLI.schema.json

@@ -1307,6 +1307,21 @@
         ".env.example",
         ".gitignore"
       ]
+    },
+    {
+      "id": "omni",
+      "description": "Runnable Node 20/22 Omni audio, read tool, capture and simulated interruption",
+      "files": [
+        "README.md",
+        "package.json",
+        "timing.mjs",
+        "main.mjs",
+        "session.mjs",
+        "audio.mjs",
+        ".env.example",
+        ".gitignore",
+        "PYAI.md"
+      ]
     }
   ],
   "output": {

package/README.md

@@ -78,29 +78,47 @@ the instant the socket opens and routes server frames to typed callbacks, so you
 **can't** trip the #1 Omni integration bug (a hand-rolled `{"event":"configure"}`
 is acked but silently dropped, giving you a connected session with zero turns):
 
-For a Node 20 server, install `npm install ws` and inject its WebSocket constructor
-as shown below. Use ESM imports. Node 22+ can omit the `ws` import and `webSocket`
-option to use its global WebSocket. This SDK transport option does not change the
-MCP server's Node 22+ requirement.
+<!-- omni-starter:start -->
+Requires SDK 0.7.1 for `--template omni`. The runnable starter uses ESM,
+Node 20.19+ with an explicit `ws` transport, or Node 22+. MCP 0.5.0
+requires Node 22+. The SDK's REST client continues to support Node 18+.
+
+```sh
+npm install @pyai/sdk@0.7.1
+npx pyai init voice-demo --template omni
+cd voice-demo
+npm install
+```
 
-```js
-import WebSocket from "ws";
-
-// Omni is zero-state: the key's org authorizes the session, nothing to create.
-const omni = pyai.omni.connect({
-  webSocket: WebSocket, // Supplies the transport on Node 20.
-  rate: 16000, // 24000 browser · 16000 wideband telephony · 8000 G.711/Twilio
-  configure: { voice_id: "stock_emma_en_gb", persona: "You are a receptionist." },
-  onAudio: (chunk) => speaker.write(chunk), // binary agent audio, play it out
-  onTranscript: (f) => console.log(f.text),
-  onError: (e) => console.error(e),
-});
+Initialization is offline; installing dependencies and running `main.mjs` are
+separate steps. For an existing checkout of this example, run `npm install`
+in its directory instead.
 
-omni.sendAudio(pcm16Chunk); // stream caller audio continuously (server-side VAD)
-omni.sendDtmf("5");
-omni.close();
+Inject `PYAI_API_KEY` through your environment, then supply a 24 kHz PCM16 mono
+WAV (at most 20 seconds) saying “Please look up the office opening time.”
+
+```sh
+node main.mjs caller.wav
+# Optional interruption clip, sent while reply audio is queued:
+node main.mjs caller.wav interruption.wav
 ```
 
+The generated project contains the complete runnable source, including its
+Node 20 WebSocket import, WAV reader, one paced input stream, read-only tool,
+and bounded capture. It waits for configuration and greeting playback to drain.
+It sends caller PCM or silence in each slot, never overlapping silence timers.
+
+Running it consumes Omni and Hear usage under the injected key. It saves a
+private report and WAV files. The report separates received audio from an
+answer recovered by Hear from captured bytes; synthesis text alone cannot pass.
+The playback sink is simulated. Interruption clears that queue; physical speaker
+playback remains a separate test. Capture uses two seconds of quiet after the
+queue drains, because the protocol has no reply-end marker.
+
+This section and the CLI project are generated from the same tested example.
+Release readiness: https://pyai.com/agents/bot-release.json
+<!-- omni-starter:end -->
+
 Live `0x02` transcript bodies are plain UTF-8 caller-text deltas, not JSON.
 `onTranscript` receives the normalized
 `{ event:"transcript", role:"user", text, final:false, mode:"delta" }` shape.
@@ -116,6 +134,9 @@ Caller transcripts continue to require the `0x02` carrier.
 `rate` configures caller input. A `rate: 16000` session still receives 24 kHz
 agent audio; `rate: 8000` receives 8 kHz. Read `hello.audio_out` before playback.
 Omni has no commit frame—keep streaming silence during caller pauses.
+Use one paced input stream: send caller PCM when available, otherwise silence.
+Pause any separate silence timer while microphone or fixture frames are being
+sent; never interleave extra silence with active input.
 
 For a client-executed lookup, declare `side_effect: "read"` in its tool definition:
 

package/dist/cli-dx.js

@@ -34,5 +34,5 @@ export const recipes = [
         notes: ["Local schema and recipes work without credentials. Live OpenAPI needs a network connection but no API key.", "The CLI manages Agent profiles; use the SDK or WebSocket API for a live Omni audio session."] },
     { name: "ci", title: "Use PyAI in a build", description: "Check access and produce an audio artifact with machine-readable output.",
         commands: ["pyai whoami --json", 'pyai speak "Your build is ready" -o build.wav --json', "pyai doctor --json"],
-        notes: ["Set PYAI_API_KEY through the CI secret store; do not put it in source or shell arguments.", "doctor synthesizes and transcribes a sample, which consumes usage. Use whoami for an authentication-only check.", "Install the CLI with npm install -g @pyai/sdk@0.6.1 and pin that version in repeatable builds."] },
+        notes: ["Set PYAI_API_KEY through the CI secret store; do not put it in source or shell arguments.", "doctor synthesizes and transcribes a sample, which consumes usage. Use whoami for an authentication-only check.", "Install the CLI with npm install -g @pyai/sdk@0.7.1 and pin that version in repeatable builds."] },
 ].map(recipe => ({ ...recipe, docs_url: "https://pyai.com/agents/speech-calling.md" }));

package/dist/cli-init.d.ts

@@ -1,4 +1,4 @@
-export type CliTemplate = "agent" | "typescript" | "python";
+export type CliTemplate = "agent" | "typescript" | "python" | "omni";
 export interface CliTemplateManifest {
     id: CliTemplate;
     description: string;

package/dist/cli-init.js

@@ -1,6 +1,7 @@
 import { constants } from "node:fs";
 import fs from "node:fs/promises";
 import { basename, dirname, isAbsolute, join, parse, resolve } from "node:path";
+import { omniTemplate } from "./cli-omni-template.js";
 export class CliInitError extends Error {
     code;
     constructor(code, message) {
@@ -127,7 +128,7 @@ const install = `## Install the CLI
 Install the published CLI with Node.js 22 or newer recommended:
 
 \`\`\`bash
-npm install -g @pyai/sdk@0.6.1
+npm install -g @pyai/sdk@0.7.1
 pyai --version
 \`\`\`
 
@@ -220,7 +221,7 @@ and product calls below are separate explicit steps. Read \`PYAI.md\` first.
 Install the official SDK:
 
 \`\`\`bash
-npm install @pyai/sdk@0.6.1
+npm install @pyai/sdk@0.7.1
 cp .env.example .env
 \`\`\`
 
@@ -321,6 +322,7 @@ Check the live contract before expanding the example.
 
 ${install}`;
 const content = Object.freeze({
+    omni: Object.freeze({ ...omniTemplate, "PYAI.md": context }),
     agent: Object.freeze({ "README.md": agentReadme, "PYAI.md": context, "speech.json": speech, "job.json": job, "agent.json": agent, ".env.example": env, ".gitignore": ignore }),
     typescript: Object.freeze({ "README.md": tsReadme, "PYAI.md": context, "main.ts": tsMain, "package.json": JSON.stringify({ name: "pyai-starter", version: "0.0.0", private: true, type: "module", engines: { node: ">=22.6.0" }, scripts: { speak: "node --experimental-strip-types --env-file=.env main.ts speak", hear: "node --experimental-strip-types --env-file=.env main.ts hear" } }, null, 2) + "\n", ".env.example": env, ".gitignore": ignore }),
     python: Object.freeze({ "README.md": pythonReadme, "PYAI.md": context, "main.py": pythonMain, ".env.example": env, ".gitignore": ignore }),
@@ -330,6 +332,7 @@ export const CLI_TEMPLATES = Object.freeze([
     { id: "agent", description: "API context and JSON requests for coding agents" },
     { id: "typescript", description: "Runnable Node.js Speak and Hear examples using @pyai/sdk" },
     { id: "python", description: "Runnable Python Speak and Hear examples using pyai-sdk" },
+    { id: "omni", description: "Runnable Node 20/22 Omni audio, read tool, capture and simulated interruption" },
 ].map((entry) => Object.freeze({ ...entry, files: Object.freeze(Object.keys(content[entry.id])) })));
 function hasCode(error, code) {
     return typeof error === "object" && error !== null && "code" in error && error.code === code;
@@ -350,7 +353,7 @@ async function exists(path) {
 export async function scaffoldProject(options) {
     const template = options.template ?? "agent";
     if (!Object.hasOwn(content, template))
-        throw new CliInitError("invalid_template", "Choose agent, typescript, or python with --template.");
+        throw new CliInitError("invalid_template", "Choose agent, typescript, python, or omni with --template.");
     if (typeof options.directory !== "string" || !options.directory.trim() || /[\p{Cc}]/u.test(options.directory)) {
         throw new CliInitError("invalid_directory", "Provide a new project directory without control characters.");
     }

package/dist/cli-omni-template.d.ts

@@ -0,0 +1,10 @@
+export declare const omniTemplate: Readonly<{
+    "README.md": "# Runnable Omni starter\n\nRequires Node 20.19+ and ESM. Node 20 uses the included `ws` transport;\nNode 22 also works. The separate MCP server requires Node 22+.\nThis server-side example uses a synthetic read-only office-hours tool. Read\n`PYAI.md` when generating the project through the CLI.\n\n<!-- omni-install:start -->\n```sh\nnpm install @pyai/sdk@0.7.1\nnpx pyai init voice-demo --template omni\ncd voice-demo\nnpm install\n```\n\nInitialization is offline; installing dependencies and running `main.mjs` are\nseparate steps. For an existing checkout of this example, run `npm install`\nin its directory instead.\n<!-- omni-install:end -->\n\nInject `PYAI_API_KEY` through your secret manager or shell environment. Do not\npaste a real key into source, a prompt, or shell history. `.env.example` documents\nthe variable; this program does not load `.env` or saved CLI profiles.\n\nPrepare a 24 kHz PCM16 mono WAV, at most 20 seconds, saying “Please look up the\noffice opening time.” Then run:\n\n```sh\nnode main.mjs caller.wav\n```\n\nRunning the program explicitly opens one bounded Omni session and uploads its\ncaptured reply to Hear. Both consume your key's allowed usage. Automatic retries\nare disabled. No managed number is bought and no phone call is placed.\n\n`session.mjs` waits for configuration and the greeting to drain, sends exactly\none paced input stream (caller PCM or silence), returns the read tool's result,\nand captures output using the rate from `hello.audio_out`. Never run a separate\nsilence timer alongside caller audio. `timing.mjs` is generated from the existing\nOmni evaluation harness's timing utilities.\n\nThe new private `omni-run-*` folder contains `session.wav`, `reply.wav`, and\n`report.json`. The report separates received audio, an office-hours answer\nrecovered by Hear from captured audio, and physical playback (not tested).\nThe office-hours check accepts formatting such as `opens@9a.m`, `9 a.m.` and\n`nine in the morning` in a complete affirmative opening-time sentence. Wrong\ntimes, negation, conflicting or truncated text remain unverified. This is a\nconservative check for the example fact, not a general answer-quality score.\nA synthesis transcript is advisory; it is not evidence that speech arrived.\nCapture ends after the simulated queue drains and two seconds of quiet. Omni\nhas no protocol reply-end marker, so this is a bounded capture heuristic.\nListen to the saved file when qualifying actual sound quality or completeness.\n\nTo exercise interruption, supply a second WAV saying “Stop speaking now”:\n\n```sh\nnode main.mjs caller.wav interruption.wav\n```\n\nThe second clip starts while reply audio is queued. `onBargeIn` clears that\nsimulated queue, and capture observes at least five seconds after that clip ends.\nAn interrupted answer can correctly fail the full-answer check;\ninspect the separate interruption evidence. This is not a physical speaker test.\nFor a real playback adapter, cancel both queued and currently playing audio.\nReplace the example lookup with authorized application data before product use.\n\nOn a session error, the program exits with JSON containing a recognized machine\n`code` (for example `media_dead` or `invalid_configure`). Unknown server codes\nbecome `unrecognized_server_error`; transport or decoding errors without a\nrecognized transport code become `transport_or_protocol_error`. Raw error text,\ncall IDs and unknown code values are omitted. No automatic retry is attempted;\ninspect the code and your authorized diagnostics before deciding to retry.\n";
+    "package.json": "{\n  \"name\": \"pyai-omni-starter\",\n  \"version\": \"0.0.0\",\n  \"private\": true,\n  \"type\": \"module\",\n  \"engines\": {\n    \"node\": \">=20.19.0\"\n  },\n  \"scripts\": {\n    \"start\": \"node main.mjs\"\n  },\n  \"dependencies\": {\n    \"@pyai/sdk\": \"0.7.1\",\n    \"ws\": \"^8.18.3\"\n  }\n}\n";
+    "timing.mjs": "// Monotonic capture timing. Arrival time is distinct from queued playback:\n// a single packet may contain seconds of audio and must finish before settling.\nexport const FRAME_MS = 20;\nexport const REALTIME_GAP_LIMIT_MS = 100;\n// Interruption tests need an observation window after caller transmission,\n// even when an earlier cue/response already finished. These are test bounds,\n// not an inferred engine reply-end or a naturalness threshold.\nexport const INTERRUPTION_OBSERVATION = Object.freeze({\n  minimumAfterCallerStreamMs: 5000, quietMs: 2000, maximumAfterCallerStreamMs: 25000,\n});\n\n/** Approximate audible bounds; this is energy detection, not a speech/VAD claim. */\nexport function audibleBounds(pcm, rate, rmsFloor = 160) {\n  const window = Math.max(1, Math.round(rate / 100)); // 10 ms windows\n  let first = null;\n  let last = null;\n  for (let off = 0; off < pcm.length; off += window) {\n    const end = Math.min(pcm.length, off + window);\n    let energy = 0;\n    for (let i = off; i < end; i++) energy += pcm[i] * pcm[i];\n    if (Math.sqrt(energy / (end - off)) >= rmsFloor) {\n      first ??= off;\n      last = end;\n    }\n  }\n  return { first, last };\n}\n\nexport function newAudioCapture(started) {\n  return {\n    started, firstPacketAt: null, lastPacketAt: null,\n    firstAudioAt: null, lastAudioAt: null, playbackEndAt: null,\n    samples: 0, pcm: [], turnBeginAt: null, eouMs: null,\n    latestTurnBeginAt: null, postTurnBeginFirstPacketAt: null,\n    postTurnBeginFirstAudioAt: null, lastAssistantTranscriptAt: null, lastCallerTranscriptAt: null,\n    callerTranscriptEvents: [], assistantTranscriptEvents: [],\n    turnBegins: [], tools: [], toolResults: [], kb: null, events: [],\n  };\n}\n\nexport function recordTurnBegin(ctx, at, turn, origin = 0) {\n  ctx.turnBeginAt ??= at;\n  ctx.latestTurnBeginAt = at;\n  // Earlier PCM can be a listening cue. A new reply is still pending even\n  // when that cue has already finished playing; retain all of its PCM.\n  ctx.postTurnBeginFirstPacketAt = null;\n  ctx.postTurnBeginFirstAudioAt = null;\n  ctx.turnBegins.push({ atMs: Math.round(at - origin), turn: turn ?? null });\n}\n\nexport function recordAudio(ctx, pcm, at, rate) {\n  if (!(pcm instanceof Int16Array) || !pcm.length) return null;\n  const copy = pcm.slice(); // Some transports reuse callback buffers.\n  const playbackAt = Math.max(at, ctx.playbackEndAt ?? at);\n  const endAt = playbackAt + (copy.length / rate) * 1000;\n  const bounds = audibleBounds(copy, rate);\n  ctx.firstPacketAt ??= at;\n  ctx.lastPacketAt = at;\n  ctx.playbackEndAt = endAt;\n  const afterTurnBegin = ctx.latestTurnBeginAt != null && at >= ctx.latestTurnBeginAt;\n  if (afterTurnBegin) ctx.postTurnBeginFirstPacketAt ??= at;\n  if (bounds.first != null) {\n    ctx.firstAudioAt ??= playbackAt + (bounds.first / rate) * 1000;\n    ctx.lastAudioAt = playbackAt + (bounds.last / rate) * 1000;\n    if (afterTurnBegin) ctx.postTurnBeginFirstAudioAt ??= playbackAt + (bounds.first / rate) * 1000;\n  }\n  ctx.samples += copy.length;\n  ctx.pcm.push(copy);\n  return { playbackAt, endAt, pcm: copy };\n}\n\n/** Preserve negative gaps: they are overlaps, never missing measurements. */\nexport function captureTiming(ctx, caller, origin = 0) {\n  const relative = (at) => at == null ? null : Math.round(at - origin);\n  const gap = (at) => at == null || caller.speechOffsetAt == null\n    ? null : Math.round(at - caller.speechOffsetAt);\n  return {\n    ttfbMs: gap(ctx.firstAudioAt),\n    anyAudioTtfbMs: gap(ctx.firstAudioAt),\n    postTurnBeginTtfbMs: gap(ctx.postTurnBeginFirstAudioAt),\n    turnMs: gap(ctx.lastAudioAt),\n    packetTtfbMs: gap(ctx.firstPacketAt),\n    callerStartMs: relative(caller.startedAt),\n    callerSpeechOnsetMs: relative(caller.speechOnsetAt),\n    callerSpeechOffsetMs: relative(caller.speechOffsetAt),\n    callerStreamEndMs: relative(caller.streamEndAt),\n    agentFirstPacketMs: relative(ctx.firstPacketAt),\n    agentLastPacketMs: relative(ctx.lastPacketAt),\n    agentSpeechOnsetMs: relative(ctx.firstAudioAt),\n    agentSpeechOffsetMs: relative(ctx.lastAudioAt),\n    agentPlaybackEndMs: relative(ctx.playbackEndAt),\n    latestTurnBeginMs: relative(ctx.latestTurnBeginAt),\n    postTurnBeginFirstPacketMs: relative(ctx.postTurnBeginFirstPacketAt),\n    postTurnBeginSpeechOnsetMs: relative(ctx.postTurnBeginFirstAudioAt),\n    lastAssistantTranscriptMs: relative(ctx.lastAssistantTranscriptAt),\n    callerMaxFrameGapMs: Math.round(caller.maxFrameGapMs ?? 0),\n    method: \"client-monotonic-queued-playback-energy-bounds\",\n  };\n}\n\nexport async function streamPcmRealtime(omni, pcm, rate, clock, onFrame = () => {}) {\n  const frameSize = Math.max(1, Math.round((rate * FRAME_MS) / 1000));\n  const bounds = audibleBounds(pcm, rate);\n  const startedAt = clock.now();\n  let speechOnsetAt = null;\n  let speechOffsetAt = null;\n  let previousFrameEnd = startedAt;\n  let maxFrameGapMs = 0;\n  for (let off = 0; off < pcm.length; off += frameSize) {\n    const frame = pcm.subarray(off, Math.min(off + frameSize, pcm.length));\n    const at = clock.now();\n    maxFrameGapMs = Math.max(maxFrameGapMs, at - previousFrameEnd);\n    const durationMs = (frame.length / rate) * 1000;\n    omni.sendAudio(frame);\n    onFrame(frame, at);\n    if (bounds.first != null && bounds.first >= off && bounds.first < off + frame.length) {\n      speechOnsetAt = at + ((bounds.first - off) / rate) * 1000;\n    }\n    if (bounds.last != null && bounds.last > off && bounds.last <= off + frame.length) {\n      speechOffsetAt = at + ((bounds.last - off) / rate) * 1000;\n    }\n    previousFrameEnd = at + durationMs;\n    // Pace from the actual send. Never burst late frames to \"catch up\".\n    await clock.sleep(Math.max(0, previousFrameEnd - clock.now()));\n  }\n  return { startedAt, speechOnsetAt, speechOffsetAt,\n    streamEndAt: clock.now(), maxFrameGapMs };\n}\n\nexport async function streamSilenceWhile(omni, rate, shouldContinue, clock, onFrame = () => {}) {\n  const silence = new Int16Array(Math.max(1, Math.round((rate * FRAME_MS) / 1000)));\n  while (shouldContinue()) {\n    const at = clock.now();\n    omni.sendAudio(silence);\n    onFrame(silence, at);\n    await clock.sleep(FRAME_MS);\n  }\n}\n\nexport async function waitForAgentSettle(getCtx, clock, opts = {}) {\n  const started = clock.now();\n  const observation = opts.callerStreamEndAt != null;\n  if (observation && (!Number.isFinite(opts.callerStreamEndAt)\n      || opts.callerStreamEndAt < 0 || opts.callerStreamEndAt > started)) {\n    throw new Error(\"Invalid caller observation boundary\");\n  }\n  const timeoutMs = observation ? INTERRUPTION_OBSERVATION.maximumAfterCallerStreamMs : opts.timeoutMs ?? 25000;\n  const settleMs = observation ? INTERRUPTION_OBSERVATION.quietMs : opts.settleMs ?? 2000;\n  const deadlineAt = (observation ? opts.callerStreamEndAt : started) + timeoutMs;\n  const minimumAt = observation ? opts.callerStreamEndAt + INTERRUPTION_OBSERVATION.minimumAfterCallerStreamMs : -Infinity;\n  while (true) {\n    const ctx = getCtx();\n    const at = clock.now();\n    if (opts.isClosed?.()) return { reason: \"closed\", at };\n    if (observation && at >= deadlineAt) return { reason: \"timeout\", at };\n    const hasResponseAudio = !opts.requireTurnBegin || (ctx.latestTurnBeginAt != null\n      && ctx.postTurnBeginFirstAudioAt != null);\n    // Synthesis text is progress, not a completion event. It can arrive after\n    // early/final audio, so extend the quiet floor without demanding another\n    // packet after every advisory. The protocol still has no reply-end marker.\n    const activityEnd = Math.max(ctx.playbackEndAt ?? -Infinity,\n      ctx.latestTurnBeginAt ?? -Infinity, ctx.lastAssistantTranscriptAt ?? -Infinity,\n      observation ? ctx.lastCallerTranscriptAt ?? -Infinity : -Infinity);\n    // Do not demand a new post-caller turn_begin: an earlier turn can be\n    // coalesced and speak later. This records a bounded quiet observation;\n    // downstream response-timing checks can impose a stricter begin boundary.\n    if (ctx.firstPacketAt != null && hasResponseAudio && at >= minimumAt && at >= activityEnd + settleMs) {\n      return { reason: \"settled\", at };\n    }\n    if (opts.allowEmpty && ctx.firstPacketAt == null && at >= minimumAt && at - started >= (opts.emptyWaitMs ?? 800)) {\n      return { reason: \"empty\", at };\n    }\n    if (at - started >= timeoutMs) return { reason: \"timeout\", at };\n    await clock.sleep(observation ? Math.min(20, deadlineAt - at) : 20);\n  }\n}\n\nexport async function withTimeout(promise, ms, message, clock) {\n  let timer;\n  try {\n    return await Promise.race([\n      promise,\n      new Promise((_, reject) => {\n        timer = clock.setTimeout(() => reject(new Error(message)), ms);\n      }),\n    ]);\n  } finally {\n    clock.clearTimeout(timer);\n  }\n}\n";
+    "main.mjs": "import PyAI from \"@pyai/sdk\";\nimport WebSocket from \"ws\"; // Explicit transport makes this work on Node 20 too.\nimport { readFile, writeFile, mkdtemp } from \"node:fs/promises\";\nimport { join } from \"node:path\";\nimport { runSession } from \"./session.mjs\";\nimport { readWav, writeWav, assessAnswer } from \"./audio.mjs\";\n\nasync function main() {\n  const [callerPath, interruptionPath, extra] = process.argv.slice(2);\n  if (!callerPath || extra) throw new Error(\"Usage: node main.mjs caller.wav [interruption.wav]\");\n  if (!process.env.PYAI_API_KEY) throw new Error(\"Inject PYAI_API_KEY through your environment first\");\n  const caller = readWav(await readFile(callerPath));\n  const interruption = interruptionPath ? readWav(await readFile(interruptionPath)) : undefined;\n  const pyai = new PyAI({ apiKey: process.env.PYAI_API_KEY, maxRetries: 0,\n    fetch: (url, options) => fetch(url, { ...options, signal: AbortSignal.timeout(30000) }),\n  });\n  const { report, audio, replyAudio } = await runSession({ pyai, webSocket: WebSocket, caller, interruption });\n  const directory = await mkdtemp(join(process.cwd(), \"omni-run-\"));\n  const save = (name, data) => writeFile(join(directory, name), data, { mode: 0o600, flag: \"wx\" });\n  await save(\"session.wav\", writeWav(audio, report.output_rate ?? 24000));\n  const reply = writeWav(replyAudio, report.output_rate ?? 24000);\n  await save(\"reply.wav\", reply);\n  // Hear the captured bytes, rather than trusting synthesis-advisory text.\n  let capturedTranscript = \"\";\n  if (report.reply_audio_bytes) {\n    try {\n      const heard = await pyai.audio.transcriptions.create({ file: new File([reply], \"reply.wav\", { type: \"audio/wav\" }), model: \"pyai-hear\" });\n      capturedTranscript = heard.text;\n    } catch { report.hear_error = \"Captured audio saved; Hear verification failed\"; }\n  }\n  const result = { ...report, captured_transcript: capturedTranscript, evidence: assessAnswer(report, capturedTranscript) };\n  await save(\"report.json\", JSON.stringify(result, null, 2) + \"\\n\");\n  console.log(JSON.stringify({ directory, ...result }, null, 2));\n  if (result.evidence.captured_answer !== \"verified_by_hear\" || (interruption && result.evidence.interruption !== \"simulated_queue_cleared\")) process.exitCode = 1;\n}\n\nmain().catch(error => {\n  const message = String(error.message).replaceAll(process.env.PYAI_API_KEY || \"\\0\", \"[REDACTED]\");\n  console.error(JSON.stringify({ error: message, ...(error.omni_code ? { code: error.omni_code } : {}) })); process.exitCode = 1;\n});\n";
+    "session.mjs": "import { newAudioCapture, recordAudio, recordTurnBegin, streamPcmRealtime, INTERRUPTION_OBSERVATION } from \"./timing.mjs\";\n\nexport const realClock = {\n  now: () => performance.now(),\n  sleep: ms => new Promise(resolve => setTimeout(resolve, ms)),\n};\n\n// Only recognized machine codes may leave the example. Error messages, call\n// IDs and arbitrary code values can contain credentials or customer content.\nconst publicErrorCodes = new Set([\n  \"invalid_configure\", \"unsupported_tool_transport\", \"tool_configuration_locked\",\n  \"media_dead\", \"unauthorized\", \"insufficient_scope\", \"rate_limit_exceeded\",\n  \"concurrency_limit_exceeded\", \"daily_cap_exceeded\", \"credit_exhausted\",\n]);\nconst transportErrorCodes = new Set([\"ECONNRESET\", \"ECONNREFUSED\", \"ETIMEDOUT\", \"ENOTFOUND\", \"EAI_AGAIN\"]);\nexport function omniFailure(error) {\n  const code = error instanceof Error\n    ? (transportErrorCodes.has(error.code) ? error.code : \"transport_or_protocol_error\")\n    : (publicErrorCodes.has(error?.code) ? error.code : \"unrecognized_server_error\");\n  return Object.assign(new Error(`Omni failed (${code}); no automatic retry was attempted`), { omni_code: code });\n}\n\n/** One bounded, server-side example. Its playback queue is simulated. */\nexport async function runSession({ pyai, webSocket, caller, interruption, clock = realClock, timeoutMs = 60000 }) {\n  const rate = 24000, frameSize = 480, silence = new Int16Array(frameSize);\n  if (!(caller instanceof Int16Array) || !caller.length || caller.length > rate * 20) throw new Error(\"Provide 1–20 seconds of caller PCM16 at 24 kHz\");\n  if (interruption && (!(interruption instanceof Int16Array) || !interruption.length || interruption.length > rate * 20)) throw new Error(\"Invalid interruption PCM\");\n  const started = clock.now(), all = newAudioCapture(started), reply = newAudioCapture(started);\n  const report = { input_rate: rate, output_rate: null, playback_sink: \"simulated\", tool_calls: 0, tool_executions: 0, duplicate_tool_calls: 0, flushes_after_interruption: 0, cleared_queue_ms: 0, interruption_requested: !!interruption, caller_started_ms: null, interruption_started_ms: null, advisory_transcripts: [], end_reason: \"timeout\" };\n  let session, configuredAt = null, closed = false, failure = null, phase = \"greeting\", offset = 0, queueEnd = started, interruptionSent = false, inputEndAt = started;\n  const handled = new Set();\n  const fail = message => { failure ??= new Error(message); };\n  try {\n    session = pyai.omni.connect({\n      webSocket, rate,\n      configure: {\n        voice_id: \"stock_amos_en_us\", language: \"en\", greeting: \"Hello. How can I help?\",\n        persona: \"When asked about office hours, call lookup_office_hours. Never guess. After the result, say one short sentence giving its opening_time.\",\n        tools: [{ name: \"lookup_office_hours\", description: \"Read the synthetic example office opening time.\", side_effect: \"read\", parameters: { type: \"object\", properties: {}, required: [], additionalProperties: false } }],\n      },\n      onHello: frame => {\n        const match = /^pcm16@(8000|24000)$/.exec(frame.audio_out ?? \"\");\n        if (!match) return fail(\"Unsupported Omni output format\");\n        report.output_rate = Number(match[1]);\n      },\n      onConfigured: () => { configuredAt ??= clock.now(); },\n      onAudio: chunk => {\n        if (!report.output_rate) return fail(\"Audio arrived before its output format\");\n        const bytes = chunk instanceof ArrayBuffer ? new Uint8Array(chunk) : new Uint8Array(chunk.buffer, chunk.byteOffset, chunk.byteLength);\n        if (bytes.byteLength % 2) return fail(\"Invalid PCM16 output\");\n        const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);\n        const pcm = Int16Array.from({ length: bytes.length / 2 }, (_, i) => view.getInt16(i * 2, true));\n        const now = clock.now();\n        recordAudio(all, pcm, now, report.output_rate);\n        queueEnd = Math.max(now, queueEnd) + pcm.length / report.output_rate * 1000;\n        if (report.tool_executions) recordAudio(reply, pcm, now, report.output_rate);\n      },\n      onEvent: frame => {\n        if (frame.event === \"turn_begin\") recordTurnBegin(all, clock.now(), frame.turn);\n      },\n      onTranscript: frame => {\n        // An assistant advisory is progress; never treat it as received speech.\n        if (report.advisory_transcripts.length < 100) report.advisory_transcripts.push({ role: frame.role, text: frame.text });\n        if (frame.role === \"assistant\") all.lastAssistantTranscriptAt = clock.now();\n      },\n      onToolCall: frame => {\n        report.tool_calls++;\n        if (handled.has(frame.call_id)) { report.duplicate_tool_calls++; return; }\n        handled.add(frame.call_id);\n        if (frame.name !== \"lookup_office_hours\") { session.toolResult(frame.call_id, { error: \"Unknown example tool\" }); return; }\n        // Replace this read-only fixture with your authorized data lookup.\n        report.tool_executions++;\n        session.toolResult(frame.call_id, { result: { opening_time: \"9 a.m.\", source: \"synthetic example office\" } });\n      },\n      onBargeIn: () => {\n        const now = clock.now(), queued = Math.max(0, queueEnd - now);\n        if (report.interruption_started_ms != null) { report.flushes_after_interruption++; report.cleared_queue_ms += queued; }\n        // A real speaker adapter must cancel scheduled and currently playing audio.\n        queueEnd = now; all.playbackEndAt = now; reply.playbackEndAt = now;\n      },\n      onError: error => { failure ??= omniFailure(error); },\n      onClose: () => { closed = true; },\n    });\n    while (clock.now() - started < timeoutMs) {\n      if (failure) throw failure;\n      if (closed) { report.end_reason = \"closed\"; break; }\n      const now = clock.now();\n      if (configuredAt == null || !report.output_rate) {\n        if (now - started >= 10000) throw new Error(\"Omni connection/configuration timed out\");\n        await clock.sleep(20); continue;\n      }\n      const quietAt = Math.max(queueEnd, all.lastPacketAt ?? now, all.lastAssistantTranscriptAt ?? 0, inputEndAt);\n      if (phase === \"greeting\") {\n        if (all.firstPacketAt != null && now - configuredAt >= 2000 && now >= quietAt + 800) {\n          phase = \"caller\"; report.caller_started_ms = now - started;\n        } else if (now - configuredAt > 15000) throw new Error(\"Greeting did not drain within 15 seconds\");\n      }\n      if (phase === \"reply\" && interruption && !interruptionSent && reply.samples && queueEnd - now >= 200) {\n        phase = \"interruption\"; offset = 0; interruptionSent = true; report.interruption_started_ms = now - started;\n      }\n      const observationEnd = inputEndAt + (interruptionSent ? INTERRUPTION_OBSERVATION.minimumAfterCallerStreamMs : 0);\n      if (phase === \"reply\" && reply.samples && now >= quietAt + 2000 && now >= observationEnd && (!interruption || interruptionSent)) {\n        report.end_reason = \"quiet_window\"; break;\n      }\n      // Exactly one producer: each 20 ms slot contains caller PCM OR silence.\n      let pcm = silence;\n      if (phase === \"caller\" || phase === \"interruption\") {\n        const source = phase === \"caller\" ? caller : interruption;\n        pcm = source.subarray(offset, offset + frameSize); offset += pcm.length;\n        if (offset >= source.length) {\n          inputEndAt = now + pcm.length / rate * 1000;\n          report[phase === \"caller\" ? \"caller_completed_ms\" : \"interruption_completed_ms\"] = inputEndAt - started;\n          phase = \"reply\"; offset = 0;\n        }\n      }\n      await streamPcmRealtime(session, pcm, rate, clock);\n    }\n    report.audio_bytes = all.samples * 2;\n    report.reply_audio_bytes = reply.samples * 2;\n    report.elapsed_ms = Math.round(clock.now() - started);\n    return { report, audio: all.pcm, replyAudio: reply.pcm };\n  } finally { session?.close(); }\n}\n";
+    "audio.mjs": "// Small, strict WAV boundary: the wire receives samples, never WAV headers.\nexport function readWav(bytes) {\n  const b = Buffer.from(bytes);\n  if (b.length < 44 || b.toString(\"ascii\", 0, 4) !== \"RIFF\" || b.toString(\"ascii\", 8, 12) !== \"WAVE\") throw new Error(\"Expected a PCM16 mono WAV file\");\n  const end = b.readUInt32LE(4) + 8;\n  if (end > b.length) throw new Error(\"Truncated WAV file\");\n  let rate, data;\n  for (let at = 12; at + 8 <= end;) {\n    const name = b.toString(\"ascii\", at, at + 4), size = b.readUInt32LE(at + 4);\n    const start = at + 8;\n    if (start + size > end) throw new Error(\"Truncated WAV chunk\");\n    if (name === \"fmt \") {\n      if (size < 16 || b.readUInt16LE(start) !== 1 || b.readUInt16LE(start + 2) !== 1 || b.readUInt16LE(start + 14) !== 16) throw new Error(\"Convert input to PCM16 mono WAV first\");\n      rate = b.readUInt32LE(start + 4);\n    } else if (name === \"data\") data = b.subarray(start, start + size);\n    at = start + size + (size % 2);\n  }\n  if (rate !== 24000 || !data?.length || data.length % 2 || data.length > rate * 2 * 20) throw new Error(\"Use a non-empty 24 kHz PCM16 mono WAV, at most 20 seconds\");\n  return Int16Array.from({ length: data.length / 2 }, (_, i) => data.readInt16LE(i * 2));\n}\n\nexport function writeWav(chunks, rate) {\n  const samples = chunks.reduce((n, pcm) => n + pcm.length, 0);\n  const b = Buffer.alloc(44 + samples * 2);\n  b.write(\"RIFF\", 0); b.writeUInt32LE(b.length - 8, 4); b.write(\"WAVEfmt \", 8);\n  b.writeUInt32LE(16, 16); b.writeUInt16LE(1, 20); b.writeUInt16LE(1, 22);\n  b.writeUInt32LE(rate, 24); b.writeUInt32LE(rate * 2, 28);\n  b.writeUInt16LE(2, 32); b.writeUInt16LE(16, 34); b.write(\"data\", 36);\n  b.writeUInt32LE(samples * 2, 40);\n  let at = 44;\n  for (const pcm of chunks) for (const value of pcm) { b.writeInt16LE(value, at); at += 2; }\n  return b;\n}\n\nexport function assessAnswer(report, capturedTranscript) {\n  // Conservative fixture check, not a general semantic evaluator. Normalize\n  // Hear's formatting without accepting negated, conflicting or partial claims.\n  const text = String(capturedTranscript ?? \"\").normalize(\"NFKC\").toLowerCase()\n    .replace(/@/g, \" at \")\n    .replace(/(?<![a-z])([ap])\\s*\\.?\\s*m\\b\\.?/g, \" $1m \")\n    .replace(/[.,!…]/g, \" \").replace(/\\s+/g, \" \").trim();\n  const expected = /^(?:(?:sure|yes|okay|ok|certainly) )?(?:(?:the|our) )?office (?:opens|(?:is )?opening|will open) (?:at )?(?:nine|9(?::00)?) (?:am|in the morning)(?: today)?$/.test(text);\n  return {\n    audio_received: report.reply_audio_bytes > 0,\n    captured_answer: report.reply_audio_bytes > 0 && report.end_reason === \"quiet_window\" && report.tool_executions === 1 && expected ? \"verified_by_hear\" : \"not_verified\",\n    capture_boundary: \"bounded_quiet_window_no_protocol_reply_end\",\n    physical_playback: \"not_tested\",\n    interruption: report.interruption_requested ? (report.flushes_after_interruption > 0 && report.cleared_queue_ms > 0 ? \"simulated_queue_cleared\" : \"not_verified\") : \"not_tested\",\n  };\n}\n";
+    ".env.example": "PYAI_API_KEY=\n";
+    ".gitignore": "node_modules/\n.env\nomni-run-*/\n*.wav\n";
+}>;

package/dist/cli-omni-template.js

@@ -0,0 +1,11 @@
+// Generated by scripts/generate-omni-starter.mjs; edit examples/omni-starter instead.
+export const omniTemplate = Object.freeze({
+    "README.md": "# Runnable Omni starter\n\nRequires Node 20.19+ and ESM. Node 20 uses the included `ws` transport;\nNode 22 also works. The separate MCP server requires Node 22+.\nThis server-side example uses a synthetic read-only office-hours tool. Read\n`PYAI.md` when generating the project through the CLI.\n\n<!-- omni-install:start -->\n```sh\nnpm install @pyai/sdk@0.7.1\nnpx pyai init voice-demo --template omni\ncd voice-demo\nnpm install\n```\n\nInitialization is offline; installing dependencies and running `main.mjs` are\nseparate steps. For an existing checkout of this example, run `npm install`\nin its directory instead.\n<!-- omni-install:end -->\n\nInject `PYAI_API_KEY` through your secret manager or shell environment. Do not\npaste a real key into source, a prompt, or shell history. `.env.example` documents\nthe variable; this program does not load `.env` or saved CLI profiles.\n\nPrepare a 24 kHz PCM16 mono WAV, at most 20 seconds, saying “Please look up the\noffice opening time.” Then run:\n\n```sh\nnode main.mjs caller.wav\n```\n\nRunning the program explicitly opens one bounded Omni session and uploads its\ncaptured reply to Hear. Both consume your key's allowed usage. Automatic retries\nare disabled. No managed number is bought and no phone call is placed.\n\n`session.mjs` waits for configuration and the greeting to drain, sends exactly\none paced input stream (caller PCM or silence), returns the read tool's result,\nand captures output using the rate from `hello.audio_out`. Never run a separate\nsilence timer alongside caller audio. `timing.mjs` is generated from the existing\nOmni evaluation harness's timing utilities.\n\nThe new private `omni-run-*` folder contains `session.wav`, `reply.wav`, and\n`report.json`. The report separates received audio, an office-hours answer\nrecovered by Hear from captured audio, and physical playback (not tested).\nThe office-hours check accepts formatting such as `opens@9a.m`, `9 a.m.` and\n`nine in the morning` in a complete affirmative opening-time sentence. Wrong\ntimes, negation, conflicting or truncated text remain unverified. This is a\nconservative check for the example fact, not a general answer-quality score.\nA synthesis transcript is advisory; it is not evidence that speech arrived.\nCapture ends after the simulated queue drains and two seconds of quiet. Omni\nhas no protocol reply-end marker, so this is a bounded capture heuristic.\nListen to the saved file when qualifying actual sound quality or completeness.\n\nTo exercise interruption, supply a second WAV saying “Stop speaking now”:\n\n```sh\nnode main.mjs caller.wav interruption.wav\n```\n\nThe second clip starts while reply audio is queued. `onBargeIn` clears that\nsimulated queue, and capture observes at least five seconds after that clip ends.\nAn interrupted answer can correctly fail the full-answer check;\ninspect the separate interruption evidence. This is not a physical speaker test.\nFor a real playback adapter, cancel both queued and currently playing audio.\nReplace the example lookup with authorized application data before product use.\n\nOn a session error, the program exits with JSON containing a recognized machine\n`code` (for example `media_dead` or `invalid_configure`). Unknown server codes\nbecome `unrecognized_server_error`; transport or decoding errors without a\nrecognized transport code become `transport_or_protocol_error`. Raw error text,\ncall IDs and unknown code values are omitted. No automatic retry is attempted;\ninspect the code and your authorized diagnostics before deciding to retry.\n",
+    "package.json": "{\n  \"name\": \"pyai-omni-starter\",\n  \"version\": \"0.0.0\",\n  \"private\": true,\n  \"type\": \"module\",\n  \"engines\": {\n    \"node\": \">=20.19.0\"\n  },\n  \"scripts\": {\n    \"start\": \"node main.mjs\"\n  },\n  \"dependencies\": {\n    \"@pyai/sdk\": \"0.7.1\",\n    \"ws\": \"^8.18.3\"\n  }\n}\n",
+    "timing.mjs": "// Monotonic capture timing. Arrival time is distinct from queued playback:\n// a single packet may contain seconds of audio and must finish before settling.\nexport const FRAME_MS = 20;\nexport const REALTIME_GAP_LIMIT_MS = 100;\n// Interruption tests need an observation window after caller transmission,\n// even when an earlier cue/response already finished. These are test bounds,\n// not an inferred engine reply-end or a naturalness threshold.\nexport const INTERRUPTION_OBSERVATION = Object.freeze({\n  minimumAfterCallerStreamMs: 5000, quietMs: 2000, maximumAfterCallerStreamMs: 25000,\n});\n\n/** Approximate audible bounds; this is energy detection, not a speech/VAD claim. */\nexport function audibleBounds(pcm, rate, rmsFloor = 160) {\n  const window = Math.max(1, Math.round(rate / 100)); // 10 ms windows\n  let first = null;\n  let last = null;\n  for (let off = 0; off < pcm.length; off += window) {\n    const end = Math.min(pcm.length, off + window);\n    let energy = 0;\n    for (let i = off; i < end; i++) energy += pcm[i] * pcm[i];\n    if (Math.sqrt(energy / (end - off)) >= rmsFloor) {\n      first ??= off;\n      last = end;\n    }\n  }\n  return { first, last };\n}\n\nexport function newAudioCapture(started) {\n  return {\n    started, firstPacketAt: null, lastPacketAt: null,\n    firstAudioAt: null, lastAudioAt: null, playbackEndAt: null,\n    samples: 0, pcm: [], turnBeginAt: null, eouMs: null,\n    latestTurnBeginAt: null, postTurnBeginFirstPacketAt: null,\n    postTurnBeginFirstAudioAt: null, lastAssistantTranscriptAt: null, lastCallerTranscriptAt: null,\n    callerTranscriptEvents: [], assistantTranscriptEvents: [],\n    turnBegins: [], tools: [], toolResults: [], kb: null, events: [],\n  };\n}\n\nexport function recordTurnBegin(ctx, at, turn, origin = 0) {\n  ctx.turnBeginAt ??= at;\n  ctx.latestTurnBeginAt = at;\n  // Earlier PCM can be a listening cue. A new reply is still pending even\n  // when that cue has already finished playing; retain all of its PCM.\n  ctx.postTurnBeginFirstPacketAt = null;\n  ctx.postTurnBeginFirstAudioAt = null;\n  ctx.turnBegins.push({ atMs: Math.round(at - origin), turn: turn ?? null });\n}\n\nexport function recordAudio(ctx, pcm, at, rate) {\n  if (!(pcm instanceof Int16Array) || !pcm.length) return null;\n  const copy = pcm.slice(); // Some transports reuse callback buffers.\n  const playbackAt = Math.max(at, ctx.playbackEndAt ?? at);\n  const endAt = playbackAt + (copy.length / rate) * 1000;\n  const bounds = audibleBounds(copy, rate);\n  ctx.firstPacketAt ??= at;\n  ctx.lastPacketAt = at;\n  ctx.playbackEndAt = endAt;\n  const afterTurnBegin = ctx.latestTurnBeginAt != null && at >= ctx.latestTurnBeginAt;\n  if (afterTurnBegin) ctx.postTurnBeginFirstPacketAt ??= at;\n  if (bounds.first != null) {\n    ctx.firstAudioAt ??= playbackAt + (bounds.first / rate) * 1000;\n    ctx.lastAudioAt = playbackAt + (bounds.last / rate) * 1000;\n    if (afterTurnBegin) ctx.postTurnBeginFirstAudioAt ??= playbackAt + (bounds.first / rate) * 1000;\n  }\n  ctx.samples += copy.length;\n  ctx.pcm.push(copy);\n  return { playbackAt, endAt, pcm: copy };\n}\n\n/** Preserve negative gaps: they are overlaps, never missing measurements. */\nexport function captureTiming(ctx, caller, origin = 0) {\n  const relative = (at) => at == null ? null : Math.round(at - origin);\n  const gap = (at) => at == null || caller.speechOffsetAt == null\n    ? null : Math.round(at - caller.speechOffsetAt);\n  return {\n    ttfbMs: gap(ctx.firstAudioAt),\n    anyAudioTtfbMs: gap(ctx.firstAudioAt),\n    postTurnBeginTtfbMs: gap(ctx.postTurnBeginFirstAudioAt),\n    turnMs: gap(ctx.lastAudioAt),\n    packetTtfbMs: gap(ctx.firstPacketAt),\n    callerStartMs: relative(caller.startedAt),\n    callerSpeechOnsetMs: relative(caller.speechOnsetAt),\n    callerSpeechOffsetMs: relative(caller.speechOffsetAt),\n    callerStreamEndMs: relative(caller.streamEndAt),\n    agentFirstPacketMs: relative(ctx.firstPacketAt),\n    agentLastPacketMs: relative(ctx.lastPacketAt),\n    agentSpeechOnsetMs: relative(ctx.firstAudioAt),\n    agentSpeechOffsetMs: relative(ctx.lastAudioAt),\n    agentPlaybackEndMs: relative(ctx.playbackEndAt),\n    latestTurnBeginMs: relative(ctx.latestTurnBeginAt),\n    postTurnBeginFirstPacketMs: relative(ctx.postTurnBeginFirstPacketAt),\n    postTurnBeginSpeechOnsetMs: relative(ctx.postTurnBeginFirstAudioAt),\n    lastAssistantTranscriptMs: relative(ctx.lastAssistantTranscriptAt),\n    callerMaxFrameGapMs: Math.round(caller.maxFrameGapMs ?? 0),\n    method: \"client-monotonic-queued-playback-energy-bounds\",\n  };\n}\n\nexport async function streamPcmRealtime(omni, pcm, rate, clock, onFrame = () => {}) {\n  const frameSize = Math.max(1, Math.round((rate * FRAME_MS) / 1000));\n  const bounds = audibleBounds(pcm, rate);\n  const startedAt = clock.now();\n  let speechOnsetAt = null;\n  let speechOffsetAt = null;\n  let previousFrameEnd = startedAt;\n  let maxFrameGapMs = 0;\n  for (let off = 0; off < pcm.length; off += frameSize) {\n    const frame = pcm.subarray(off, Math.min(off + frameSize, pcm.length));\n    const at = clock.now();\n    maxFrameGapMs = Math.max(maxFrameGapMs, at - previousFrameEnd);\n    const durationMs = (frame.length / rate) * 1000;\n    omni.sendAudio(frame);\n    onFrame(frame, at);\n    if (bounds.first != null && bounds.first >= off && bounds.first < off + frame.length) {\n      speechOnsetAt = at + ((bounds.first - off) / rate) * 1000;\n    }\n    if (bounds.last != null && bounds.last > off && bounds.last <= off + frame.length) {\n      speechOffsetAt = at + ((bounds.last - off) / rate) * 1000;\n    }\n    previousFrameEnd = at + durationMs;\n    // Pace from the actual send. Never burst late frames to \"catch up\".\n    await clock.sleep(Math.max(0, previousFrameEnd - clock.now()));\n  }\n  return { startedAt, speechOnsetAt, speechOffsetAt,\n    streamEndAt: clock.now(), maxFrameGapMs };\n}\n\nexport async function streamSilenceWhile(omni, rate, shouldContinue, clock, onFrame = () => {}) {\n  const silence = new Int16Array(Math.max(1, Math.round((rate * FRAME_MS) / 1000)));\n  while (shouldContinue()) {\n    const at = clock.now();\n    omni.sendAudio(silence);\n    onFrame(silence, at);\n    await clock.sleep(FRAME_MS);\n  }\n}\n\nexport async function waitForAgentSettle(getCtx, clock, opts = {}) {\n  const started = clock.now();\n  const observation = opts.callerStreamEndAt != null;\n  if (observation && (!Number.isFinite(opts.callerStreamEndAt)\n      || opts.callerStreamEndAt < 0 || opts.callerStreamEndAt > started)) {\n    throw new Error(\"Invalid caller observation boundary\");\n  }\n  const timeoutMs = observation ? INTERRUPTION_OBSERVATION.maximumAfterCallerStreamMs : opts.timeoutMs ?? 25000;\n  const settleMs = observation ? INTERRUPTION_OBSERVATION.quietMs : opts.settleMs ?? 2000;\n  const deadlineAt = (observation ? opts.callerStreamEndAt : started) + timeoutMs;\n  const minimumAt = observation ? opts.callerStreamEndAt + INTERRUPTION_OBSERVATION.minimumAfterCallerStreamMs : -Infinity;\n  while (true) {\n    const ctx = getCtx();\n    const at = clock.now();\n    if (opts.isClosed?.()) return { reason: \"closed\", at };\n    if (observation && at >= deadlineAt) return { reason: \"timeout\", at };\n    const hasResponseAudio = !opts.requireTurnBegin || (ctx.latestTurnBeginAt != null\n      && ctx.postTurnBeginFirstAudioAt != null);\n    // Synthesis text is progress, not a completion event. It can arrive after\n    // early/final audio, so extend the quiet floor without demanding another\n    // packet after every advisory. The protocol still has no reply-end marker.\n    const activityEnd = Math.max(ctx.playbackEndAt ?? -Infinity,\n      ctx.latestTurnBeginAt ?? -Infinity, ctx.lastAssistantTranscriptAt ?? -Infinity,\n      observation ? ctx.lastCallerTranscriptAt ?? -Infinity : -Infinity);\n    // Do not demand a new post-caller turn_begin: an earlier turn can be\n    // coalesced and speak later. This records a bounded quiet observation;\n    // downstream response-timing checks can impose a stricter begin boundary.\n    if (ctx.firstPacketAt != null && hasResponseAudio && at >= minimumAt && at >= activityEnd + settleMs) {\n      return { reason: \"settled\", at };\n    }\n    if (opts.allowEmpty && ctx.firstPacketAt == null && at >= minimumAt && at - started >= (opts.emptyWaitMs ?? 800)) {\n      return { reason: \"empty\", at };\n    }\n    if (at - started >= timeoutMs) return { reason: \"timeout\", at };\n    await clock.sleep(observation ? Math.min(20, deadlineAt - at) : 20);\n  }\n}\n\nexport async function withTimeout(promise, ms, message, clock) {\n  let timer;\n  try {\n    return await Promise.race([\n      promise,\n      new Promise((_, reject) => {\n        timer = clock.setTimeout(() => reject(new Error(message)), ms);\n      }),\n    ]);\n  } finally {\n    clock.clearTimeout(timer);\n  }\n}\n",
+    "main.mjs": "import PyAI from \"@pyai/sdk\";\nimport WebSocket from \"ws\"; // Explicit transport makes this work on Node 20 too.\nimport { readFile, writeFile, mkdtemp } from \"node:fs/promises\";\nimport { join } from \"node:path\";\nimport { runSession } from \"./session.mjs\";\nimport { readWav, writeWav, assessAnswer } from \"./audio.mjs\";\n\nasync function main() {\n  const [callerPath, interruptionPath, extra] = process.argv.slice(2);\n  if (!callerPath || extra) throw new Error(\"Usage: node main.mjs caller.wav [interruption.wav]\");\n  if (!process.env.PYAI_API_KEY) throw new Error(\"Inject PYAI_API_KEY through your environment first\");\n  const caller = readWav(await readFile(callerPath));\n  const interruption = interruptionPath ? readWav(await readFile(interruptionPath)) : undefined;\n  const pyai = new PyAI({ apiKey: process.env.PYAI_API_KEY, maxRetries: 0,\n    fetch: (url, options) => fetch(url, { ...options, signal: AbortSignal.timeout(30000) }),\n  });\n  const { report, audio, replyAudio } = await runSession({ pyai, webSocket: WebSocket, caller, interruption });\n  const directory = await mkdtemp(join(process.cwd(), \"omni-run-\"));\n  const save = (name, data) => writeFile(join(directory, name), data, { mode: 0o600, flag: \"wx\" });\n  await save(\"session.wav\", writeWav(audio, report.output_rate ?? 24000));\n  const reply = writeWav(replyAudio, report.output_rate ?? 24000);\n  await save(\"reply.wav\", reply);\n  // Hear the captured bytes, rather than trusting synthesis-advisory text.\n  let capturedTranscript = \"\";\n  if (report.reply_audio_bytes) {\n    try {\n      const heard = await pyai.audio.transcriptions.create({ file: new File([reply], \"reply.wav\", { type: \"audio/wav\" }), model: \"pyai-hear\" });\n      capturedTranscript = heard.text;\n    } catch { report.hear_error = \"Captured audio saved; Hear verification failed\"; }\n  }\n  const result = { ...report, captured_transcript: capturedTranscript, evidence: assessAnswer(report, capturedTranscript) };\n  await save(\"report.json\", JSON.stringify(result, null, 2) + \"\\n\");\n  console.log(JSON.stringify({ directory, ...result }, null, 2));\n  if (result.evidence.captured_answer !== \"verified_by_hear\" || (interruption && result.evidence.interruption !== \"simulated_queue_cleared\")) process.exitCode = 1;\n}\n\nmain().catch(error => {\n  const message = String(error.message).replaceAll(process.env.PYAI_API_KEY || \"\\0\", \"[REDACTED]\");\n  console.error(JSON.stringify({ error: message, ...(error.omni_code ? { code: error.omni_code } : {}) })); process.exitCode = 1;\n});\n",
+    "session.mjs": "import { newAudioCapture, recordAudio, recordTurnBegin, streamPcmRealtime, INTERRUPTION_OBSERVATION } from \"./timing.mjs\";\n\nexport const realClock = {\n  now: () => performance.now(),\n  sleep: ms => new Promise(resolve => setTimeout(resolve, ms)),\n};\n\n// Only recognized machine codes may leave the example. Error messages, call\n// IDs and arbitrary code values can contain credentials or customer content.\nconst publicErrorCodes = new Set([\n  \"invalid_configure\", \"unsupported_tool_transport\", \"tool_configuration_locked\",\n  \"media_dead\", \"unauthorized\", \"insufficient_scope\", \"rate_limit_exceeded\",\n  \"concurrency_limit_exceeded\", \"daily_cap_exceeded\", \"credit_exhausted\",\n]);\nconst transportErrorCodes = new Set([\"ECONNRESET\", \"ECONNREFUSED\", \"ETIMEDOUT\", \"ENOTFOUND\", \"EAI_AGAIN\"]);\nexport function omniFailure(error) {\n  const code = error instanceof Error\n    ? (transportErrorCodes.has(error.code) ? error.code : \"transport_or_protocol_error\")\n    : (publicErrorCodes.has(error?.code) ? error.code : \"unrecognized_server_error\");\n  return Object.assign(new Error(`Omni failed (${code}); no automatic retry was attempted`), { omni_code: code });\n}\n\n/** One bounded, server-side example. Its playback queue is simulated. */\nexport async function runSession({ pyai, webSocket, caller, interruption, clock = realClock, timeoutMs = 60000 }) {\n  const rate = 24000, frameSize = 480, silence = new Int16Array(frameSize);\n  if (!(caller instanceof Int16Array) || !caller.length || caller.length > rate * 20) throw new Error(\"Provide 1–20 seconds of caller PCM16 at 24 kHz\");\n  if (interruption && (!(interruption instanceof Int16Array) || !interruption.length || interruption.length > rate * 20)) throw new Error(\"Invalid interruption PCM\");\n  const started = clock.now(), all = newAudioCapture(started), reply = newAudioCapture(started);\n  const report = { input_rate: rate, output_rate: null, playback_sink: \"simulated\", tool_calls: 0, tool_executions: 0, duplicate_tool_calls: 0, flushes_after_interruption: 0, cleared_queue_ms: 0, interruption_requested: !!interruption, caller_started_ms: null, interruption_started_ms: null, advisory_transcripts: [], end_reason: \"timeout\" };\n  let session, configuredAt = null, closed = false, failure = null, phase = \"greeting\", offset = 0, queueEnd = started, interruptionSent = false, inputEndAt = started;\n  const handled = new Set();\n  const fail = message => { failure ??= new Error(message); };\n  try {\n    session = pyai.omni.connect({\n      webSocket, rate,\n      configure: {\n        voice_id: \"stock_amos_en_us\", language: \"en\", greeting: \"Hello. How can I help?\",\n        persona: \"When asked about office hours, call lookup_office_hours. Never guess. After the result, say one short sentence giving its opening_time.\",\n        tools: [{ name: \"lookup_office_hours\", description: \"Read the synthetic example office opening time.\", side_effect: \"read\", parameters: { type: \"object\", properties: {}, required: [], additionalProperties: false } }],\n      },\n      onHello: frame => {\n        const match = /^pcm16@(8000|24000)$/.exec(frame.audio_out ?? \"\");\n        if (!match) return fail(\"Unsupported Omni output format\");\n        report.output_rate = Number(match[1]);\n      },\n      onConfigured: () => { configuredAt ??= clock.now(); },\n      onAudio: chunk => {\n        if (!report.output_rate) return fail(\"Audio arrived before its output format\");\n        const bytes = chunk instanceof ArrayBuffer ? new Uint8Array(chunk) : new Uint8Array(chunk.buffer, chunk.byteOffset, chunk.byteLength);\n        if (bytes.byteLength % 2) return fail(\"Invalid PCM16 output\");\n        const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);\n        const pcm = Int16Array.from({ length: bytes.length / 2 }, (_, i) => view.getInt16(i * 2, true));\n        const now = clock.now();\n        recordAudio(all, pcm, now, report.output_rate);\n        queueEnd = Math.max(now, queueEnd) + pcm.length / report.output_rate * 1000;\n        if (report.tool_executions) recordAudio(reply, pcm, now, report.output_rate);\n      },\n      onEvent: frame => {\n        if (frame.event === \"turn_begin\") recordTurnBegin(all, clock.now(), frame.turn);\n      },\n      onTranscript: frame => {\n        // An assistant advisory is progress; never treat it as received speech.\n        if (report.advisory_transcripts.length < 100) report.advisory_transcripts.push({ role: frame.role, text: frame.text });\n        if (frame.role === \"assistant\") all.lastAssistantTranscriptAt = clock.now();\n      },\n      onToolCall: frame => {\n        report.tool_calls++;\n        if (handled.has(frame.call_id)) { report.duplicate_tool_calls++; return; }\n        handled.add(frame.call_id);\n        if (frame.name !== \"lookup_office_hours\") { session.toolResult(frame.call_id, { error: \"Unknown example tool\" }); return; }\n        // Replace this read-only fixture with your authorized data lookup.\n        report.tool_executions++;\n        session.toolResult(frame.call_id, { result: { opening_time: \"9 a.m.\", source: \"synthetic example office\" } });\n      },\n      onBargeIn: () => {\n        const now = clock.now(), queued = Math.max(0, queueEnd - now);\n        if (report.interruption_started_ms != null) { report.flushes_after_interruption++; report.cleared_queue_ms += queued; }\n        // A real speaker adapter must cancel scheduled and currently playing audio.\n        queueEnd = now; all.playbackEndAt = now; reply.playbackEndAt = now;\n      },\n      onError: error => { failure ??= omniFailure(error); },\n      onClose: () => { closed = true; },\n    });\n    while (clock.now() - started < timeoutMs) {\n      if (failure) throw failure;\n      if (closed) { report.end_reason = \"closed\"; break; }\n      const now = clock.now();\n      if (configuredAt == null || !report.output_rate) {\n        if (now - started >= 10000) throw new Error(\"Omni connection/configuration timed out\");\n        await clock.sleep(20); continue;\n      }\n      const quietAt = Math.max(queueEnd, all.lastPacketAt ?? now, all.lastAssistantTranscriptAt ?? 0, inputEndAt);\n      if (phase === \"greeting\") {\n        if (all.firstPacketAt != null && now - configuredAt >= 2000 && now >= quietAt + 800) {\n          phase = \"caller\"; report.caller_started_ms = now - started;\n        } else if (now - configuredAt > 15000) throw new Error(\"Greeting did not drain within 15 seconds\");\n      }\n      if (phase === \"reply\" && interruption && !interruptionSent && reply.samples && queueEnd - now >= 200) {\n        phase = \"interruption\"; offset = 0; interruptionSent = true; report.interruption_started_ms = now - started;\n      }\n      const observationEnd = inputEndAt + (interruptionSent ? INTERRUPTION_OBSERVATION.minimumAfterCallerStreamMs : 0);\n      if (phase === \"reply\" && reply.samples && now >= quietAt + 2000 && now >= observationEnd && (!interruption || interruptionSent)) {\n        report.end_reason = \"quiet_window\"; break;\n      }\n      // Exactly one producer: each 20 ms slot contains caller PCM OR silence.\n      let pcm = silence;\n      if (phase === \"caller\" || phase === \"interruption\") {\n        const source = phase === \"caller\" ? caller : interruption;\n        pcm = source.subarray(offset, offset + frameSize); offset += pcm.length;\n        if (offset >= source.length) {\n          inputEndAt = now + pcm.length / rate * 1000;\n          report[phase === \"caller\" ? \"caller_completed_ms\" : \"interruption_completed_ms\"] = inputEndAt - started;\n          phase = \"reply\"; offset = 0;\n        }\n      }\n      await streamPcmRealtime(session, pcm, rate, clock);\n    }\n    report.audio_bytes = all.samples * 2;\n    report.reply_audio_bytes = reply.samples * 2;\n    report.elapsed_ms = Math.round(clock.now() - started);\n    return { report, audio: all.pcm, replyAudio: reply.pcm };\n  } finally { session?.close(); }\n}\n",
+    "audio.mjs": "// Small, strict WAV boundary: the wire receives samples, never WAV headers.\nexport function readWav(bytes) {\n  const b = Buffer.from(bytes);\n  if (b.length < 44 || b.toString(\"ascii\", 0, 4) !== \"RIFF\" || b.toString(\"ascii\", 8, 12) !== \"WAVE\") throw new Error(\"Expected a PCM16 mono WAV file\");\n  const end = b.readUInt32LE(4) + 8;\n  if (end > b.length) throw new Error(\"Truncated WAV file\");\n  let rate, data;\n  for (let at = 12; at + 8 <= end;) {\n    const name = b.toString(\"ascii\", at, at + 4), size = b.readUInt32LE(at + 4);\n    const start = at + 8;\n    if (start + size > end) throw new Error(\"Truncated WAV chunk\");\n    if (name === \"fmt \") {\n      if (size < 16 || b.readUInt16LE(start) !== 1 || b.readUInt16LE(start + 2) !== 1 || b.readUInt16LE(start + 14) !== 16) throw new Error(\"Convert input to PCM16 mono WAV first\");\n      rate = b.readUInt32LE(start + 4);\n    } else if (name === \"data\") data = b.subarray(start, start + size);\n    at = start + size + (size % 2);\n  }\n  if (rate !== 24000 || !data?.length || data.length % 2 || data.length > rate * 2 * 20) throw new Error(\"Use a non-empty 24 kHz PCM16 mono WAV, at most 20 seconds\");\n  return Int16Array.from({ length: data.length / 2 }, (_, i) => data.readInt16LE(i * 2));\n}\n\nexport function writeWav(chunks, rate) {\n  const samples = chunks.reduce((n, pcm) => n + pcm.length, 0);\n  const b = Buffer.alloc(44 + samples * 2);\n  b.write(\"RIFF\", 0); b.writeUInt32LE(b.length - 8, 4); b.write(\"WAVEfmt \", 8);\n  b.writeUInt32LE(16, 16); b.writeUInt16LE(1, 20); b.writeUInt16LE(1, 22);\n  b.writeUInt32LE(rate, 24); b.writeUInt32LE(rate * 2, 28);\n  b.writeUInt16LE(2, 32); b.writeUInt16LE(16, 34); b.write(\"data\", 36);\n  b.writeUInt32LE(samples * 2, 40);\n  let at = 44;\n  for (const pcm of chunks) for (const value of pcm) { b.writeInt16LE(value, at); at += 2; }\n  return b;\n}\n\nexport function assessAnswer(report, capturedTranscript) {\n  // Conservative fixture check, not a general semantic evaluator. Normalize\n  // Hear's formatting without accepting negated, conflicting or partial claims.\n  const text = String(capturedTranscript ?? \"\").normalize(\"NFKC\").toLowerCase()\n    .replace(/@/g, \" at \")\n    .replace(/(?<![a-z])([ap])\\s*\\.?\\s*m\\b\\.?/g, \" $1m \")\n    .replace(/[.,!…]/g, \" \").replace(/\\s+/g, \" \").trim();\n  const expected = /^(?:(?:sure|yes|okay|ok|certainly) )?(?:(?:the|our) )?office (?:opens|(?:is )?opening|will open) (?:at )?(?:nine|9(?::00)?) (?:am|in the morning)(?: today)?$/.test(text);\n  return {\n    audio_received: report.reply_audio_bytes > 0,\n    captured_answer: report.reply_audio_bytes > 0 && report.end_reason === \"quiet_window\" && report.tool_executions === 1 && expected ? \"verified_by_hear\" : \"not_verified\",\n    capture_boundary: \"bounded_quiet_window_no_protocol_reply_end\",\n    physical_playback: \"not_tested\",\n    interruption: report.interruption_requested ? (report.flushes_after_interruption > 0 && report.cleared_queue_ms > 0 ? \"simulated_queue_cleared\" : \"not_verified\") : \"not_tested\",\n  };\n}\n",
+    ".env.example": "PYAI_API_KEY=\n",
+    ".gitignore": "node_modules/\n.env\nomni-run-*/\n*.wav\n"
+});

package/dist/cli.js

@@ -405,7 +405,7 @@ async function main() {
         if (flags.json)
             output({ name: "pyai", globals, aliases, short_flags: shortFlags, commands: visible }, flags, undefined, false);
         else if (!spec && !group)
-            process.stdout.write(`PyAI CLI — speech, agents, and automation\n\nStart here\n  pyai login                                 Sign in through your browser\n  pyai speak "Hello from PyAI" -o hello.wav    Turn text into audio\n  pyai hear call.wav --text-only              Print a transcript\n  pyai dub call.wav --from en --to hi -o hi.wav\n                                             Submit, wait, and save dubbed audio\n  pyai init voice-demo --template typescript  Create a starter project offline\n\nCalling\n  pyai numbers list                         List owned phone numbers\n  pyai numbers search --country US           Find available numbers\n  pyai calls list                           Inspect outbound calls\n  pyai recipes calling --json                Setup and preview examples\n  pyai calls create --data @call.json --idempotency-key REQUEST_KEY --dry-run\n                                             Preview a call without submitting it\n  --dry-run previews a change without sending the request.\n  Number purchases and outbound calls require --confirm and --idempotency-key.\n  Use --confirm only after that purchase or call is authorized.\n\nDiscover\n  pyai recipes [name]        Copyable workflows\n  pyai help <command>        Command options and examples\n  pyai help all              Every command\n  pyai schema [command] -j   Machine-readable command definitions\n\nGroups\n  auth  profiles  models  voices  agents  jobs  clones  design\n  cast  dub  recap  trace  tools  vocabulary  amd  numbers  calls  omni\n\nShortcuts: login, logout, whoami, use, say (speak), hear (transcribe)\nOptions: -o FILE  -f FILE  -t TEXT  -p PROFILE  -j (JSON)\nGlobal: --base-url URL --api-key KEY --timeout SECONDS --retries N --dry-run\nCredentials: pyai login or PYAI_API_KEY. Run pyai doctor for a usage-consuming diagnostic.\n`);
+            process.stdout.write(`PyAI CLI — speech, agents, and automation\n\nStart here\n  pyai login                                 Sign in through your browser\n  pyai speak "Hello from PyAI" -o hello.wav    Turn text into audio\n  pyai hear call.wav --text-only              Print a transcript\n  pyai dub call.wav --from en --to hi -o hi.wav\n                                             Submit, wait, and save dubbed audio\n  pyai init voice-demo --template omni        Create a runnable Omni starter offline\n\nCalling\n  pyai numbers list                         List owned phone numbers\n  pyai numbers search --country US           Find available numbers\n  pyai calls list                           Inspect outbound calls\n  pyai recipes calling --json                Setup and preview examples\n  pyai calls create --data @call.json --idempotency-key REQUEST_KEY --dry-run\n                                             Preview a call without submitting it\n  --dry-run previews a change without sending the request.\n  Number purchases and outbound calls require --confirm and --idempotency-key.\n  Use --confirm only after that purchase or call is authorized.\n\nDiscover\n  pyai recipes [name]        Copyable workflows\n  pyai help <command>        Command options and examples\n  pyai help all              Every command\n  pyai schema [command] -j   Machine-readable command definitions\n\nGroups\n  auth  profiles  models  voices  agents  jobs  clones  design\n  cast  dub  recap  trace  tools  vocabulary  amd  numbers  calls  omni\n\nShortcuts: login, logout, whoami, use, say (speak), hear (transcribe)\nOptions: -o FILE  -f FILE  -t TEXT  -p PROFILE  -j (JSON)\nGlobal: --base-url URL --api-key KEY --timeout SECONDS --retries N --dry-run\nCredentials: pyai login or PYAI_API_KEY. Run pyai doctor for a usage-consuming diagnostic.\n`);
         else
             process.stdout.write(`PyAI CLI — speech, agents, and automation\n\n${visible.map(c => `  pyai ${c.command}${(c.args ?? []).map(a => c.optionalArgs ? ` [${a}${c.variadic ? "..." : ""}]` : ` <${a}>`).join("")}\n    ${c.description}${c.flags ? `\n    Options: ${Object.keys(c.flags).map(k => `--${k}${c.flags[k] === "boolean" ? "" : " VALUE"}`).join(" ")}` : ""}${c.example ? `\n    ${c.example}` : ""}`).join("\n")}\n\nGlobal: --profile NAME --json --timeout SECONDS --retries N --dry-run\n        --base-url URL --api-key KEY --help --version\nShort flags: -o FILE  -f FILE  -t TEXT  -p PROFILE  -j (JSON)\nAuth: pyai login or PYAI_API_KEY. Discover: pyai schema --json\n`);
         return;

package/dist/index.d.ts

@@ -3,7 +3,8 @@
  *
  * Thin, dependency-free wrapper over the public OpenAI-compatible surface at
  * https://api.pyai.com (contract: https://api.pyai.com/openapi.json). Runs in
- * the browser and Node 22+ (uses global fetch + WebSocket). Keys are opaque, * never parsed.
+ * the browser and Node 18+ for REST. Realtime needs an injected WebSocket
+ * on Node 20; Node 22+ provides a global WebSocket. Keys are opaque, never parsed.
  */
 export interface PyAIOptions {
     /** A pyai_live_ or pyai_test_ key. */

package/dist/index.js

@@ -3,7 +3,8 @@
  *
  * Thin, dependency-free wrapper over the public OpenAI-compatible surface at
  * https://api.pyai.com (contract: https://api.pyai.com/openapi.json). Runs in
- * the browser and Node 22+ (uses global fetch + WebSocket). Keys are opaque, * never parsed.
+ * the browser and Node 18+ for REST. Realtime needs an injected WebSocket
+ * on Node 20; Node 22+ provides a global WebSocket. Keys are opaque, never parsed.
  */
 /** Stable, machine-readable error. Branch on `code`, not `message`. */
 export class PyAIError extends Error {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pyai/sdk",
-  "version": "0.6.1",
+  "version": "0.7.1",
   "description": "Official TypeScript/JavaScript SDK for PyAI, speech-to-text (Hear), text-to-speech (Speak), realtime voice agents (Omni), and call compliance (Trace).",
   "type": "module",
   "main": "dist/index.js",
@@ -31,8 +31,8 @@
   "bugs": "https://github.com/atomsai/pyai-platform-backend/issues",
   "author": "PyAI",
   "scripts": {
-    "gen:cli-assets": "node scripts/generate-cli-assets.mjs",
-    "check:cli-assets": "node scripts/generate-cli-assets.mjs --check",
+    "gen:cli-assets": "node ../../scripts/generate-omni-starter.mjs && node scripts/generate-cli-assets.mjs",
+    "check:cli-assets": "node ../../scripts/generate-omni-starter.mjs --check && node scripts/generate-cli-assets.mjs --check",
     "build": "tsc -p tsconfig.build.json",
     "test": "node --test test/**/*.test.ts",
     "typecheck": "tsc --noEmit"

package/src/cli-dx.ts

@@ -35,5 +35,5 @@ export const recipes: Recipe[] = [
     notes: ["Local schema and recipes work without credentials. Live OpenAPI needs a network connection but no API key.", "The CLI manages Agent profiles; use the SDK or WebSocket API for a live Omni audio session."] },
   { name: "ci", title: "Use PyAI in a build", description: "Check access and produce an audio artifact with machine-readable output.",
     commands: ["pyai whoami --json", 'pyai speak "Your build is ready" -o build.wav --json', "pyai doctor --json"],
-    notes: ["Set PYAI_API_KEY through the CI secret store; do not put it in source or shell arguments.", "doctor synthesizes and transcribes a sample, which consumes usage. Use whoami for an authentication-only check.", "Install the CLI with npm install -g @pyai/sdk@0.6.1 and pin that version in repeatable builds."] },
+    notes: ["Set PYAI_API_KEY through the CI secret store; do not put it in source or shell arguments.", "doctor synthesizes and transcribes a sample, which consumes usage. Use whoami for an authentication-only check.", "Install the CLI with npm install -g @pyai/sdk@0.7.1 and pin that version in repeatable builds."] },
 ].map(recipe => ({ ...recipe, docs_url: "https://pyai.com/agents/speech-calling.md" }));

package/src/cli-init.ts

@@ -2,7 +2,9 @@ import { constants, type Stats } from "node:fs";
 import fs from "node:fs/promises";
 import { basename, dirname, isAbsolute, join, parse, resolve } from "node:path";
 
-export type CliTemplate = "agent" | "typescript" | "python";
+import { omniTemplate } from "./cli-omni-template.ts";
+
+export type CliTemplate = "agent" | "typescript" | "python" | "omni";
 export interface CliTemplateManifest {
   id: CliTemplate;
   description: string;
@@ -149,7 +151,7 @@ const install = `## Install the CLI
 Install the published CLI with Node.js 22 or newer recommended:
 
 \`\`\`bash
-npm install -g @pyai/sdk@0.6.1
+npm install -g @pyai/sdk@0.7.1
 pyai --version
 \`\`\`
 
@@ -242,7 +244,7 @@ and product calls below are separate explicit steps. Read \`PYAI.md\` first.
 Install the official SDK:
 
 \`\`\`bash
-npm install @pyai/sdk@0.6.1
+npm install @pyai/sdk@0.7.1
 cp .env.example .env
 \`\`\`
 
@@ -344,6 +346,7 @@ Check the live contract before expanding the example.
 ${install}`;
 
 const content: Readonly<Record<CliTemplate, Readonly<Record<string, string>>>> = Object.freeze({
+  omni: Object.freeze({ ...omniTemplate, "PYAI.md": context }),
   agent: Object.freeze({ "README.md": agentReadme, "PYAI.md": context, "speech.json": speech, "job.json": job, "agent.json": agent, ".env.example": env, ".gitignore": ignore }),
   typescript: Object.freeze({ "README.md": tsReadme, "PYAI.md": context, "main.ts": tsMain, "package.json": JSON.stringify({ name: "pyai-starter", version: "0.0.0", private: true, type: "module", engines: { node: ">=22.6.0" }, scripts: { speak: "node --experimental-strip-types --env-file=.env main.ts speak", hear: "node --experimental-strip-types --env-file=.env main.ts hear" } }, null, 2) + "\n", ".env.example": env, ".gitignore": ignore }),
   python: Object.freeze({ "README.md": pythonReadme, "PYAI.md": context, "main.py": pythonMain, ".env.example": env, ".gitignore": ignore }),
@@ -354,6 +357,7 @@ export const CLI_TEMPLATES: readonly CliTemplateManifest[] = Object.freeze(([
   { id: "agent", description: "API context and JSON requests for coding agents" },
   { id: "typescript", description: "Runnable Node.js Speak and Hear examples using @pyai/sdk" },
   { id: "python", description: "Runnable Python Speak and Hear examples using pyai-sdk" },
+  { id: "omni", description: "Runnable Node 20/22 Omni audio, read tool, capture and simulated interruption" },
 ] as const).map((entry) => Object.freeze({ ...entry, files: Object.freeze(Object.keys(content[entry.id])) })));
 
 function hasCode(error: unknown, code: string): boolean {
@@ -368,7 +372,7 @@ async function exists(path: string): Promise<boolean> {
 /** Create a new project only. Never authenticates, installs packages, or runs a shell. */
 export async function scaffoldProject(options: ScaffoldOptions): Promise<ScaffoldResult> {
   const template = options.template ?? "agent";
-  if (!Object.hasOwn(content, template)) throw new CliInitError("invalid_template", "Choose agent, typescript, or python with --template.");
+  if (!Object.hasOwn(content, template)) throw new CliInitError("invalid_template", "Choose agent, typescript, python, or omni with --template.");
   if (typeof options.directory !== "string" || !options.directory.trim() || /[\p{Cc}]/u.test(options.directory)) {
     throw new CliInitError("invalid_directory", "Provide a new project directory without control characters.");
   }

package/src/cli-omni-template.ts

@@ -0,0 +1,11 @@
+// Generated by scripts/generate-omni-starter.mjs; edit examples/omni-starter instead.
+export const omniTemplate = Object.freeze({
+  "README.md": "# Runnable Omni starter\n\nRequires Node 20.19+ and ESM. Node 20 uses the included `ws` transport;\nNode 22 also works. The separate MCP server requires Node 22+.\nThis server-side example uses a synthetic read-only office-hours tool. Read\n`PYAI.md` when generating the project through the CLI.\n\n<!-- omni-install:start -->\n```sh\nnpm install @pyai/sdk@0.7.1\nnpx pyai init voice-demo --template omni\ncd voice-demo\nnpm install\n```\n\nInitialization is offline; installing dependencies and running `main.mjs` are\nseparate steps. For an existing checkout of this example, run `npm install`\nin its directory instead.\n<!-- omni-install:end -->\n\nInject `PYAI_API_KEY` through your secret manager or shell environment. Do not\npaste a real key into source, a prompt, or shell history. `.env.example` documents\nthe variable; this program does not load `.env` or saved CLI profiles.\n\nPrepare a 24 kHz PCM16 mono WAV, at most 20 seconds, saying “Please look up the\noffice opening time.” Then run:\n\n```sh\nnode main.mjs caller.wav\n```\n\nRunning the program explicitly opens one bounded Omni session and uploads its\ncaptured reply to Hear. Both consume your key's allowed usage. Automatic retries\nare disabled. No managed number is bought and no phone call is placed.\n\n`session.mjs` waits for configuration and the greeting to drain, sends exactly\none paced input stream (caller PCM or silence), returns the read tool's result,\nand captures output using the rate from `hello.audio_out`. Never run a separate\nsilence timer alongside caller audio. `timing.mjs` is generated from the existing\nOmni evaluation harness's timing utilities.\n\nThe new private `omni-run-*` folder contains `session.wav`, `reply.wav`, and\n`report.json`. The report separates received audio, an office-hours answer\nrecovered by Hear from captured audio, and physical playback (not tested).\nThe office-hours check accepts formatting such as `opens@9a.m`, `9 a.m.` and\n`nine in the morning` in a complete affirmative opening-time sentence. Wrong\ntimes, negation, conflicting or truncated text remain unverified. This is a\nconservative check for the example fact, not a general answer-quality score.\nA synthesis transcript is advisory; it is not evidence that speech arrived.\nCapture ends after the simulated queue drains and two seconds of quiet. Omni\nhas no protocol reply-end marker, so this is a bounded capture heuristic.\nListen to the saved file when qualifying actual sound quality or completeness.\n\nTo exercise interruption, supply a second WAV saying “Stop speaking now”:\n\n```sh\nnode main.mjs caller.wav interruption.wav\n```\n\nThe second clip starts while reply audio is queued. `onBargeIn` clears that\nsimulated queue, and capture observes at least five seconds after that clip ends.\nAn interrupted answer can correctly fail the full-answer check;\ninspect the separate interruption evidence. This is not a physical speaker test.\nFor a real playback adapter, cancel both queued and currently playing audio.\nReplace the example lookup with authorized application data before product use.\n\nOn a session error, the program exits with JSON containing a recognized machine\n`code` (for example `media_dead` or `invalid_configure`). Unknown server codes\nbecome `unrecognized_server_error`; transport or decoding errors without a\nrecognized transport code become `transport_or_protocol_error`. Raw error text,\ncall IDs and unknown code values are omitted. No automatic retry is attempted;\ninspect the code and your authorized diagnostics before deciding to retry.\n",
+  "package.json": "{\n  \"name\": \"pyai-omni-starter\",\n  \"version\": \"0.0.0\",\n  \"private\": true,\n  \"type\": \"module\",\n  \"engines\": {\n    \"node\": \">=20.19.0\"\n  },\n  \"scripts\": {\n    \"start\": \"node main.mjs\"\n  },\n  \"dependencies\": {\n    \"@pyai/sdk\": \"0.7.1\",\n    \"ws\": \"^8.18.3\"\n  }\n}\n",
+  "timing.mjs": "// Monotonic capture timing. Arrival time is distinct from queued playback:\n// a single packet may contain seconds of audio and must finish before settling.\nexport const FRAME_MS = 20;\nexport const REALTIME_GAP_LIMIT_MS = 100;\n// Interruption tests need an observation window after caller transmission,\n// even when an earlier cue/response already finished. These are test bounds,\n// not an inferred engine reply-end or a naturalness threshold.\nexport const INTERRUPTION_OBSERVATION = Object.freeze({\n  minimumAfterCallerStreamMs: 5000, quietMs: 2000, maximumAfterCallerStreamMs: 25000,\n});\n\n/** Approximate audible bounds; this is energy detection, not a speech/VAD claim. */\nexport function audibleBounds(pcm, rate, rmsFloor = 160) {\n  const window = Math.max(1, Math.round(rate / 100)); // 10 ms windows\n  let first = null;\n  let last = null;\n  for (let off = 0; off < pcm.length; off += window) {\n    const end = Math.min(pcm.length, off + window);\n    let energy = 0;\n    for (let i = off; i < end; i++) energy += pcm[i] * pcm[i];\n    if (Math.sqrt(energy / (end - off)) >= rmsFloor) {\n      first ??= off;\n      last = end;\n    }\n  }\n  return { first, last };\n}\n\nexport function newAudioCapture(started) {\n  return {\n    started, firstPacketAt: null, lastPacketAt: null,\n    firstAudioAt: null, lastAudioAt: null, playbackEndAt: null,\n    samples: 0, pcm: [], turnBeginAt: null, eouMs: null,\n    latestTurnBeginAt: null, postTurnBeginFirstPacketAt: null,\n    postTurnBeginFirstAudioAt: null, lastAssistantTranscriptAt: null, lastCallerTranscriptAt: null,\n    callerTranscriptEvents: [], assistantTranscriptEvents: [],\n    turnBegins: [], tools: [], toolResults: [], kb: null, events: [],\n  };\n}\n\nexport function recordTurnBegin(ctx, at, turn, origin = 0) {\n  ctx.turnBeginAt ??= at;\n  ctx.latestTurnBeginAt = at;\n  // Earlier PCM can be a listening cue. A new reply is still pending even\n  // when that cue has already finished playing; retain all of its PCM.\n  ctx.postTurnBeginFirstPacketAt = null;\n  ctx.postTurnBeginFirstAudioAt = null;\n  ctx.turnBegins.push({ atMs: Math.round(at - origin), turn: turn ?? null });\n}\n\nexport function recordAudio(ctx, pcm, at, rate) {\n  if (!(pcm instanceof Int16Array) || !pcm.length) return null;\n  const copy = pcm.slice(); // Some transports reuse callback buffers.\n  const playbackAt = Math.max(at, ctx.playbackEndAt ?? at);\n  const endAt = playbackAt + (copy.length / rate) * 1000;\n  const bounds = audibleBounds(copy, rate);\n  ctx.firstPacketAt ??= at;\n  ctx.lastPacketAt = at;\n  ctx.playbackEndAt = endAt;\n  const afterTurnBegin = ctx.latestTurnBeginAt != null && at >= ctx.latestTurnBeginAt;\n  if (afterTurnBegin) ctx.postTurnBeginFirstPacketAt ??= at;\n  if (bounds.first != null) {\n    ctx.firstAudioAt ??= playbackAt + (bounds.first / rate) * 1000;\n    ctx.lastAudioAt = playbackAt + (bounds.last / rate) * 1000;\n    if (afterTurnBegin) ctx.postTurnBeginFirstAudioAt ??= playbackAt + (bounds.first / rate) * 1000;\n  }\n  ctx.samples += copy.length;\n  ctx.pcm.push(copy);\n  return { playbackAt, endAt, pcm: copy };\n}\n\n/** Preserve negative gaps: they are overlaps, never missing measurements. */\nexport function captureTiming(ctx, caller, origin = 0) {\n  const relative = (at) => at == null ? null : Math.round(at - origin);\n  const gap = (at) => at == null || caller.speechOffsetAt == null\n    ? null : Math.round(at - caller.speechOffsetAt);\n  return {\n    ttfbMs: gap(ctx.firstAudioAt),\n    anyAudioTtfbMs: gap(ctx.firstAudioAt),\n    postTurnBeginTtfbMs: gap(ctx.postTurnBeginFirstAudioAt),\n    turnMs: gap(ctx.lastAudioAt),\n    packetTtfbMs: gap(ctx.firstPacketAt),\n    callerStartMs: relative(caller.startedAt),\n    callerSpeechOnsetMs: relative(caller.speechOnsetAt),\n    callerSpeechOffsetMs: relative(caller.speechOffsetAt),\n    callerStreamEndMs: relative(caller.streamEndAt),\n    agentFirstPacketMs: relative(ctx.firstPacketAt),\n    agentLastPacketMs: relative(ctx.lastPacketAt),\n    agentSpeechOnsetMs: relative(ctx.firstAudioAt),\n    agentSpeechOffsetMs: relative(ctx.lastAudioAt),\n    agentPlaybackEndMs: relative(ctx.playbackEndAt),\n    latestTurnBeginMs: relative(ctx.latestTurnBeginAt),\n    postTurnBeginFirstPacketMs: relative(ctx.postTurnBeginFirstPacketAt),\n    postTurnBeginSpeechOnsetMs: relative(ctx.postTurnBeginFirstAudioAt),\n    lastAssistantTranscriptMs: relative(ctx.lastAssistantTranscriptAt),\n    callerMaxFrameGapMs: Math.round(caller.maxFrameGapMs ?? 0),\n    method: \"client-monotonic-queued-playback-energy-bounds\",\n  };\n}\n\nexport async function streamPcmRealtime(omni, pcm, rate, clock, onFrame = () => {}) {\n  const frameSize = Math.max(1, Math.round((rate * FRAME_MS) / 1000));\n  const bounds = audibleBounds(pcm, rate);\n  const startedAt = clock.now();\n  let speechOnsetAt = null;\n  let speechOffsetAt = null;\n  let previousFrameEnd = startedAt;\n  let maxFrameGapMs = 0;\n  for (let off = 0; off < pcm.length; off += frameSize) {\n    const frame = pcm.subarray(off, Math.min(off + frameSize, pcm.length));\n    const at = clock.now();\n    maxFrameGapMs = Math.max(maxFrameGapMs, at - previousFrameEnd);\n    const durationMs = (frame.length / rate) * 1000;\n    omni.sendAudio(frame);\n    onFrame(frame, at);\n    if (bounds.first != null && bounds.first >= off && bounds.first < off + frame.length) {\n      speechOnsetAt = at + ((bounds.first - off) / rate) * 1000;\n    }\n    if (bounds.last != null && bounds.last > off && bounds.last <= off + frame.length) {\n      speechOffsetAt = at + ((bounds.last - off) / rate) * 1000;\n    }\n    previousFrameEnd = at + durationMs;\n    // Pace from the actual send. Never burst late frames to \"catch up\".\n    await clock.sleep(Math.max(0, previousFrameEnd - clock.now()));\n  }\n  return { startedAt, speechOnsetAt, speechOffsetAt,\n    streamEndAt: clock.now(), maxFrameGapMs };\n}\n\nexport async function streamSilenceWhile(omni, rate, shouldContinue, clock, onFrame = () => {}) {\n  const silence = new Int16Array(Math.max(1, Math.round((rate * FRAME_MS) / 1000)));\n  while (shouldContinue()) {\n    const at = clock.now();\n    omni.sendAudio(silence);\n    onFrame(silence, at);\n    await clock.sleep(FRAME_MS);\n  }\n}\n\nexport async function waitForAgentSettle(getCtx, clock, opts = {}) {\n  const started = clock.now();\n  const observation = opts.callerStreamEndAt != null;\n  if (observation && (!Number.isFinite(opts.callerStreamEndAt)\n      || opts.callerStreamEndAt < 0 || opts.callerStreamEndAt > started)) {\n    throw new Error(\"Invalid caller observation boundary\");\n  }\n  const timeoutMs = observation ? INTERRUPTION_OBSERVATION.maximumAfterCallerStreamMs : opts.timeoutMs ?? 25000;\n  const settleMs = observation ? INTERRUPTION_OBSERVATION.quietMs : opts.settleMs ?? 2000;\n  const deadlineAt = (observation ? opts.callerStreamEndAt : started) + timeoutMs;\n  const minimumAt = observation ? opts.callerStreamEndAt + INTERRUPTION_OBSERVATION.minimumAfterCallerStreamMs : -Infinity;\n  while (true) {\n    const ctx = getCtx();\n    const at = clock.now();\n    if (opts.isClosed?.()) return { reason: \"closed\", at };\n    if (observation && at >= deadlineAt) return { reason: \"timeout\", at };\n    const hasResponseAudio = !opts.requireTurnBegin || (ctx.latestTurnBeginAt != null\n      && ctx.postTurnBeginFirstAudioAt != null);\n    // Synthesis text is progress, not a completion event. It can arrive after\n    // early/final audio, so extend the quiet floor without demanding another\n    // packet after every advisory. The protocol still has no reply-end marker.\n    const activityEnd = Math.max(ctx.playbackEndAt ?? -Infinity,\n      ctx.latestTurnBeginAt ?? -Infinity, ctx.lastAssistantTranscriptAt ?? -Infinity,\n      observation ? ctx.lastCallerTranscriptAt ?? -Infinity : -Infinity);\n    // Do not demand a new post-caller turn_begin: an earlier turn can be\n    // coalesced and speak later. This records a bounded quiet observation;\n    // downstream response-timing checks can impose a stricter begin boundary.\n    if (ctx.firstPacketAt != null && hasResponseAudio && at >= minimumAt && at >= activityEnd + settleMs) {\n      return { reason: \"settled\", at };\n    }\n    if (opts.allowEmpty && ctx.firstPacketAt == null && at >= minimumAt && at - started >= (opts.emptyWaitMs ?? 800)) {\n      return { reason: \"empty\", at };\n    }\n    if (at - started >= timeoutMs) return { reason: \"timeout\", at };\n    await clock.sleep(observation ? Math.min(20, deadlineAt - at) : 20);\n  }\n}\n\nexport async function withTimeout(promise, ms, message, clock) {\n  let timer;\n  try {\n    return await Promise.race([\n      promise,\n      new Promise((_, reject) => {\n        timer = clock.setTimeout(() => reject(new Error(message)), ms);\n      }),\n    ]);\n  } finally {\n    clock.clearTimeout(timer);\n  }\n}\n",
+  "main.mjs": "import PyAI from \"@pyai/sdk\";\nimport WebSocket from \"ws\"; // Explicit transport makes this work on Node 20 too.\nimport { readFile, writeFile, mkdtemp } from \"node:fs/promises\";\nimport { join } from \"node:path\";\nimport { runSession } from \"./session.mjs\";\nimport { readWav, writeWav, assessAnswer } from \"./audio.mjs\";\n\nasync function main() {\n  const [callerPath, interruptionPath, extra] = process.argv.slice(2);\n  if (!callerPath || extra) throw new Error(\"Usage: node main.mjs caller.wav [interruption.wav]\");\n  if (!process.env.PYAI_API_KEY) throw new Error(\"Inject PYAI_API_KEY through your environment first\");\n  const caller = readWav(await readFile(callerPath));\n  const interruption = interruptionPath ? readWav(await readFile(interruptionPath)) : undefined;\n  const pyai = new PyAI({ apiKey: process.env.PYAI_API_KEY, maxRetries: 0,\n    fetch: (url, options) => fetch(url, { ...options, signal: AbortSignal.timeout(30000) }),\n  });\n  const { report, audio, replyAudio } = await runSession({ pyai, webSocket: WebSocket, caller, interruption });\n  const directory = await mkdtemp(join(process.cwd(), \"omni-run-\"));\n  const save = (name, data) => writeFile(join(directory, name), data, { mode: 0o600, flag: \"wx\" });\n  await save(\"session.wav\", writeWav(audio, report.output_rate ?? 24000));\n  const reply = writeWav(replyAudio, report.output_rate ?? 24000);\n  await save(\"reply.wav\", reply);\n  // Hear the captured bytes, rather than trusting synthesis-advisory text.\n  let capturedTranscript = \"\";\n  if (report.reply_audio_bytes) {\n    try {\n      const heard = await pyai.audio.transcriptions.create({ file: new File([reply], \"reply.wav\", { type: \"audio/wav\" }), model: \"pyai-hear\" });\n      capturedTranscript = heard.text;\n    } catch { report.hear_error = \"Captured audio saved; Hear verification failed\"; }\n  }\n  const result = { ...report, captured_transcript: capturedTranscript, evidence: assessAnswer(report, capturedTranscript) };\n  await save(\"report.json\", JSON.stringify(result, null, 2) + \"\\n\");\n  console.log(JSON.stringify({ directory, ...result }, null, 2));\n  if (result.evidence.captured_answer !== \"verified_by_hear\" || (interruption && result.evidence.interruption !== \"simulated_queue_cleared\")) process.exitCode = 1;\n}\n\nmain().catch(error => {\n  const message = String(error.message).replaceAll(process.env.PYAI_API_KEY || \"\\0\", \"[REDACTED]\");\n  console.error(JSON.stringify({ error: message, ...(error.omni_code ? { code: error.omni_code } : {}) })); process.exitCode = 1;\n});\n",
+  "session.mjs": "import { newAudioCapture, recordAudio, recordTurnBegin, streamPcmRealtime, INTERRUPTION_OBSERVATION } from \"./timing.mjs\";\n\nexport const realClock = {\n  now: () => performance.now(),\n  sleep: ms => new Promise(resolve => setTimeout(resolve, ms)),\n};\n\n// Only recognized machine codes may leave the example. Error messages, call\n// IDs and arbitrary code values can contain credentials or customer content.\nconst publicErrorCodes = new Set([\n  \"invalid_configure\", \"unsupported_tool_transport\", \"tool_configuration_locked\",\n  \"media_dead\", \"unauthorized\", \"insufficient_scope\", \"rate_limit_exceeded\",\n  \"concurrency_limit_exceeded\", \"daily_cap_exceeded\", \"credit_exhausted\",\n]);\nconst transportErrorCodes = new Set([\"ECONNRESET\", \"ECONNREFUSED\", \"ETIMEDOUT\", \"ENOTFOUND\", \"EAI_AGAIN\"]);\nexport function omniFailure(error) {\n  const code = error instanceof Error\n    ? (transportErrorCodes.has(error.code) ? error.code : \"transport_or_protocol_error\")\n    : (publicErrorCodes.has(error?.code) ? error.code : \"unrecognized_server_error\");\n  return Object.assign(new Error(`Omni failed (${code}); no automatic retry was attempted`), { omni_code: code });\n}\n\n/** One bounded, server-side example. Its playback queue is simulated. */\nexport async function runSession({ pyai, webSocket, caller, interruption, clock = realClock, timeoutMs = 60000 }) {\n  const rate = 24000, frameSize = 480, silence = new Int16Array(frameSize);\n  if (!(caller instanceof Int16Array) || !caller.length || caller.length > rate * 20) throw new Error(\"Provide 1–20 seconds of caller PCM16 at 24 kHz\");\n  if (interruption && (!(interruption instanceof Int16Array) || !interruption.length || interruption.length > rate * 20)) throw new Error(\"Invalid interruption PCM\");\n  const started = clock.now(), all = newAudioCapture(started), reply = newAudioCapture(started);\n  const report = { input_rate: rate, output_rate: null, playback_sink: \"simulated\", tool_calls: 0, tool_executions: 0, duplicate_tool_calls: 0, flushes_after_interruption: 0, cleared_queue_ms: 0, interruption_requested: !!interruption, caller_started_ms: null, interruption_started_ms: null, advisory_transcripts: [], end_reason: \"timeout\" };\n  let session, configuredAt = null, closed = false, failure = null, phase = \"greeting\", offset = 0, queueEnd = started, interruptionSent = false, inputEndAt = started;\n  const handled = new Set();\n  const fail = message => { failure ??= new Error(message); };\n  try {\n    session = pyai.omni.connect({\n      webSocket, rate,\n      configure: {\n        voice_id: \"stock_amos_en_us\", language: \"en\", greeting: \"Hello. How can I help?\",\n        persona: \"When asked about office hours, call lookup_office_hours. Never guess. After the result, say one short sentence giving its opening_time.\",\n        tools: [{ name: \"lookup_office_hours\", description: \"Read the synthetic example office opening time.\", side_effect: \"read\", parameters: { type: \"object\", properties: {}, required: [], additionalProperties: false } }],\n      },\n      onHello: frame => {\n        const match = /^pcm16@(8000|24000)$/.exec(frame.audio_out ?? \"\");\n        if (!match) return fail(\"Unsupported Omni output format\");\n        report.output_rate = Number(match[1]);\n      },\n      onConfigured: () => { configuredAt ??= clock.now(); },\n      onAudio: chunk => {\n        if (!report.output_rate) return fail(\"Audio arrived before its output format\");\n        const bytes = chunk instanceof ArrayBuffer ? new Uint8Array(chunk) : new Uint8Array(chunk.buffer, chunk.byteOffset, chunk.byteLength);\n        if (bytes.byteLength % 2) return fail(\"Invalid PCM16 output\");\n        const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);\n        const pcm = Int16Array.from({ length: bytes.length / 2 }, (_, i) => view.getInt16(i * 2, true));\n        const now = clock.now();\n        recordAudio(all, pcm, now, report.output_rate);\n        queueEnd = Math.max(now, queueEnd) + pcm.length / report.output_rate * 1000;\n        if (report.tool_executions) recordAudio(reply, pcm, now, report.output_rate);\n      },\n      onEvent: frame => {\n        if (frame.event === \"turn_begin\") recordTurnBegin(all, clock.now(), frame.turn);\n      },\n      onTranscript: frame => {\n        // An assistant advisory is progress; never treat it as received speech.\n        if (report.advisory_transcripts.length < 100) report.advisory_transcripts.push({ role: frame.role, text: frame.text });\n        if (frame.role === \"assistant\") all.lastAssistantTranscriptAt = clock.now();\n      },\n      onToolCall: frame => {\n        report.tool_calls++;\n        if (handled.has(frame.call_id)) { report.duplicate_tool_calls++; return; }\n        handled.add(frame.call_id);\n        if (frame.name !== \"lookup_office_hours\") { session.toolResult(frame.call_id, { error: \"Unknown example tool\" }); return; }\n        // Replace this read-only fixture with your authorized data lookup.\n        report.tool_executions++;\n        session.toolResult(frame.call_id, { result: { opening_time: \"9 a.m.\", source: \"synthetic example office\" } });\n      },\n      onBargeIn: () => {\n        const now = clock.now(), queued = Math.max(0, queueEnd - now);\n        if (report.interruption_started_ms != null) { report.flushes_after_interruption++; report.cleared_queue_ms += queued; }\n        // A real speaker adapter must cancel scheduled and currently playing audio.\n        queueEnd = now; all.playbackEndAt = now; reply.playbackEndAt = now;\n      },\n      onError: error => { failure ??= omniFailure(error); },\n      onClose: () => { closed = true; },\n    });\n    while (clock.now() - started < timeoutMs) {\n      if (failure) throw failure;\n      if (closed) { report.end_reason = \"closed\"; break; }\n      const now = clock.now();\n      if (configuredAt == null || !report.output_rate) {\n        if (now - started >= 10000) throw new Error(\"Omni connection/configuration timed out\");\n        await clock.sleep(20); continue;\n      }\n      const quietAt = Math.max(queueEnd, all.lastPacketAt ?? now, all.lastAssistantTranscriptAt ?? 0, inputEndAt);\n      if (phase === \"greeting\") {\n        if (all.firstPacketAt != null && now - configuredAt >= 2000 && now >= quietAt + 800) {\n          phase = \"caller\"; report.caller_started_ms = now - started;\n        } else if (now - configuredAt > 15000) throw new Error(\"Greeting did not drain within 15 seconds\");\n      }\n      if (phase === \"reply\" && interruption && !interruptionSent && reply.samples && queueEnd - now >= 200) {\n        phase = \"interruption\"; offset = 0; interruptionSent = true; report.interruption_started_ms = now - started;\n      }\n      const observationEnd = inputEndAt + (interruptionSent ? INTERRUPTION_OBSERVATION.minimumAfterCallerStreamMs : 0);\n      if (phase === \"reply\" && reply.samples && now >= quietAt + 2000 && now >= observationEnd && (!interruption || interruptionSent)) {\n        report.end_reason = \"quiet_window\"; break;\n      }\n      // Exactly one producer: each 20 ms slot contains caller PCM OR silence.\n      let pcm = silence;\n      if (phase === \"caller\" || phase === \"interruption\") {\n        const source = phase === \"caller\" ? caller : interruption;\n        pcm = source.subarray(offset, offset + frameSize); offset += pcm.length;\n        if (offset >= source.length) {\n          inputEndAt = now + pcm.length / rate * 1000;\n          report[phase === \"caller\" ? \"caller_completed_ms\" : \"interruption_completed_ms\"] = inputEndAt - started;\n          phase = \"reply\"; offset = 0;\n        }\n      }\n      await streamPcmRealtime(session, pcm, rate, clock);\n    }\n    report.audio_bytes = all.samples * 2;\n    report.reply_audio_bytes = reply.samples * 2;\n    report.elapsed_ms = Math.round(clock.now() - started);\n    return { report, audio: all.pcm, replyAudio: reply.pcm };\n  } finally { session?.close(); }\n}\n",
+  "audio.mjs": "// Small, strict WAV boundary: the wire receives samples, never WAV headers.\nexport function readWav(bytes) {\n  const b = Buffer.from(bytes);\n  if (b.length < 44 || b.toString(\"ascii\", 0, 4) !== \"RIFF\" || b.toString(\"ascii\", 8, 12) !== \"WAVE\") throw new Error(\"Expected a PCM16 mono WAV file\");\n  const end = b.readUInt32LE(4) + 8;\n  if (end > b.length) throw new Error(\"Truncated WAV file\");\n  let rate, data;\n  for (let at = 12; at + 8 <= end;) {\n    const name = b.toString(\"ascii\", at, at + 4), size = b.readUInt32LE(at + 4);\n    const start = at + 8;\n    if (start + size > end) throw new Error(\"Truncated WAV chunk\");\n    if (name === \"fmt \") {\n      if (size < 16 || b.readUInt16LE(start) !== 1 || b.readUInt16LE(start + 2) !== 1 || b.readUInt16LE(start + 14) !== 16) throw new Error(\"Convert input to PCM16 mono WAV first\");\n      rate = b.readUInt32LE(start + 4);\n    } else if (name === \"data\") data = b.subarray(start, start + size);\n    at = start + size + (size % 2);\n  }\n  if (rate !== 24000 || !data?.length || data.length % 2 || data.length > rate * 2 * 20) throw new Error(\"Use a non-empty 24 kHz PCM16 mono WAV, at most 20 seconds\");\n  return Int16Array.from({ length: data.length / 2 }, (_, i) => data.readInt16LE(i * 2));\n}\n\nexport function writeWav(chunks, rate) {\n  const samples = chunks.reduce((n, pcm) => n + pcm.length, 0);\n  const b = Buffer.alloc(44 + samples * 2);\n  b.write(\"RIFF\", 0); b.writeUInt32LE(b.length - 8, 4); b.write(\"WAVEfmt \", 8);\n  b.writeUInt32LE(16, 16); b.writeUInt16LE(1, 20); b.writeUInt16LE(1, 22);\n  b.writeUInt32LE(rate, 24); b.writeUInt32LE(rate * 2, 28);\n  b.writeUInt16LE(2, 32); b.writeUInt16LE(16, 34); b.write(\"data\", 36);\n  b.writeUInt32LE(samples * 2, 40);\n  let at = 44;\n  for (const pcm of chunks) for (const value of pcm) { b.writeInt16LE(value, at); at += 2; }\n  return b;\n}\n\nexport function assessAnswer(report, capturedTranscript) {\n  // Conservative fixture check, not a general semantic evaluator. Normalize\n  // Hear's formatting without accepting negated, conflicting or partial claims.\n  const text = String(capturedTranscript ?? \"\").normalize(\"NFKC\").toLowerCase()\n    .replace(/@/g, \" at \")\n    .replace(/(?<![a-z])([ap])\\s*\\.?\\s*m\\b\\.?/g, \" $1m \")\n    .replace(/[.,!…]/g, \" \").replace(/\\s+/g, \" \").trim();\n  const expected = /^(?:(?:sure|yes|okay|ok|certainly) )?(?:(?:the|our) )?office (?:opens|(?:is )?opening|will open) (?:at )?(?:nine|9(?::00)?) (?:am|in the morning)(?: today)?$/.test(text);\n  return {\n    audio_received: report.reply_audio_bytes > 0,\n    captured_answer: report.reply_audio_bytes > 0 && report.end_reason === \"quiet_window\" && report.tool_executions === 1 && expected ? \"verified_by_hear\" : \"not_verified\",\n    capture_boundary: \"bounded_quiet_window_no_protocol_reply_end\",\n    physical_playback: \"not_tested\",\n    interruption: report.interruption_requested ? (report.flushes_after_interruption > 0 && report.cleared_queue_ms > 0 ? \"simulated_queue_cleared\" : \"not_verified\") : \"not_tested\",\n  };\n}\n",
+  ".env.example": "PYAI_API_KEY=\n",
+  ".gitignore": "node_modules/\n.env\nomni-run-*/\n*.wav\n"
+});

package/src/cli.ts

@@ -270,7 +270,7 @@ async function main(): Promise<void> {
     const grouped = group ? commands.filter(c => c.command.startsWith(`${group} `)) : commands;
     const visible = spec ? [spec, ...commands.filter(c => c.command.startsWith(`${spec.command} `))] : grouped.length ? grouped : commands;
     if (flags.json) output({ name: "pyai", globals, aliases, short_flags: shortFlags, commands: visible }, flags, undefined, false);
-    else if (!spec && !group) process.stdout.write(`PyAI CLI — speech, agents, and automation\n\nStart here\n  pyai login                                 Sign in through your browser\n  pyai speak "Hello from PyAI" -o hello.wav    Turn text into audio\n  pyai hear call.wav --text-only              Print a transcript\n  pyai dub call.wav --from en --to hi -o hi.wav\n                                             Submit, wait, and save dubbed audio\n  pyai init voice-demo --template typescript  Create a starter project offline\n\nCalling\n  pyai numbers list                         List owned phone numbers\n  pyai numbers search --country US           Find available numbers\n  pyai calls list                           Inspect outbound calls\n  pyai recipes calling --json                Setup and preview examples\n  pyai calls create --data @call.json --idempotency-key REQUEST_KEY --dry-run\n                                             Preview a call without submitting it\n  --dry-run previews a change without sending the request.\n  Number purchases and outbound calls require --confirm and --idempotency-key.\n  Use --confirm only after that purchase or call is authorized.\n\nDiscover\n  pyai recipes [name]        Copyable workflows\n  pyai help <command>        Command options and examples\n  pyai help all              Every command\n  pyai schema [command] -j   Machine-readable command definitions\n\nGroups\n  auth  profiles  models  voices  agents  jobs  clones  design\n  cast  dub  recap  trace  tools  vocabulary  amd  numbers  calls  omni\n\nShortcuts: login, logout, whoami, use, say (speak), hear (transcribe)\nOptions: -o FILE  -f FILE  -t TEXT  -p PROFILE  -j (JSON)\nGlobal: --base-url URL --api-key KEY --timeout SECONDS --retries N --dry-run\nCredentials: pyai login or PYAI_API_KEY. Run pyai doctor for a usage-consuming diagnostic.\n`);
+    else if (!spec && !group) process.stdout.write(`PyAI CLI — speech, agents, and automation\n\nStart here\n  pyai login                                 Sign in through your browser\n  pyai speak "Hello from PyAI" -o hello.wav    Turn text into audio\n  pyai hear call.wav --text-only              Print a transcript\n  pyai dub call.wav --from en --to hi -o hi.wav\n                                             Submit, wait, and save dubbed audio\n  pyai init voice-demo --template omni        Create a runnable Omni starter offline\n\nCalling\n  pyai numbers list                         List owned phone numbers\n  pyai numbers search --country US           Find available numbers\n  pyai calls list                           Inspect outbound calls\n  pyai recipes calling --json                Setup and preview examples\n  pyai calls create --data @call.json --idempotency-key REQUEST_KEY --dry-run\n                                             Preview a call without submitting it\n  --dry-run previews a change without sending the request.\n  Number purchases and outbound calls require --confirm and --idempotency-key.\n  Use --confirm only after that purchase or call is authorized.\n\nDiscover\n  pyai recipes [name]        Copyable workflows\n  pyai help <command>        Command options and examples\n  pyai help all              Every command\n  pyai schema [command] -j   Machine-readable command definitions\n\nGroups\n  auth  profiles  models  voices  agents  jobs  clones  design\n  cast  dub  recap  trace  tools  vocabulary  amd  numbers  calls  omni\n\nShortcuts: login, logout, whoami, use, say (speak), hear (transcribe)\nOptions: -o FILE  -f FILE  -t TEXT  -p PROFILE  -j (JSON)\nGlobal: --base-url URL --api-key KEY --timeout SECONDS --retries N --dry-run\nCredentials: pyai login or PYAI_API_KEY. Run pyai doctor for a usage-consuming diagnostic.\n`);
     else process.stdout.write(`PyAI CLI — speech, agents, and automation\n\n${visible.map(c => `  pyai ${c.command}${(c.args ?? []).map(a => c.optionalArgs ? ` [${a}${c.variadic ? "..." : ""}]` : ` <${a}>`).join("")}\n    ${c.description}${c.flags ? `\n    Options: ${Object.keys(c.flags).map(k => `--${k}${c.flags![k] === "boolean" ? "" : " VALUE"}`).join(" ")}` : ""}${c.example ? `\n    ${c.example}` : ""}`).join("\n")}\n\nGlobal: --profile NAME --json --timeout SECONDS --retries N --dry-run\n        --base-url URL --api-key KEY --help --version\nShort flags: -o FILE  -f FILE  -t TEXT  -p PROFILE  -j (JSON)\nAuth: pyai login or PYAI_API_KEY. Discover: pyai schema --json\n`);
     return;
   }

package/src/index.ts

@@ -3,7 +3,8 @@
  *
  * Thin, dependency-free wrapper over the public OpenAI-compatible surface at
  * https://api.pyai.com (contract: https://api.pyai.com/openapi.json). Runs in
- * the browser and Node 22+ (uses global fetch + WebSocket). Keys are opaque, * never parsed.
+ * the browser and Node 18+ for REST. Realtime needs an injected WebSocket
+ * on Node 20; Node 22+ provides a global WebSocket. Keys are opaque, never parsed.
  */
 
 export interface PyAIOptions {