askrobots / dbbasic-face
Public
Puppet Stage: a browser puppet that lip-syncs to local text-to-speech and takes stage directions, live or from a screenplay.
server.js · 51.7 KB · 1251 lines
| 1 | #!/usr/bin/env node |
| 2 | /** |
| 3 | * Puppet stage server. |
| 4 | * |
| 5 | * - Renders TTS to WAV via macOS `say` or `espeak`, extracts a viseme |
| 6 | * timeline with Rhubarb Lip Sync, and serves both to the browser stage. |
| 7 | * - Broadcasts direction cues (speech, actions, movement) to all connected |
| 8 | * stages over Server-Sent Events. |
| 9 | * - Runs "screenplay" scripts: plain lines are spoken, [bracketed] lines are |
| 10 | * stage directions, (action) prefixes run concurrently with speech. |
| 11 | * |
| 12 | * Zero npm dependencies. |
| 13 | */ |
| 14 | 'use strict'; |
| 15 | |
| 16 | const http = require('http'); |
| 17 | const fs = require('fs'); |
| 18 | const path = require('path'); |
| 19 | const crypto = require('crypto'); |
| 20 | const { execFile } = require('child_process'); |
| 21 | |
| 22 | const ROOT = __dirname; |
| 23 | const CACHE = path.join(ROOT, 'cache'); |
| 24 | const PUBLIC = path.join(ROOT, 'public'); |
| 25 | const CHARACTERS = path.join(ROOT, 'characters'); |
| 26 | const ASSETS = path.join(ROOT, 'assets'); |
| 27 | const EXAMPLES = path.join(ROOT, 'examples'); |
| 28 | const RHUBARB = findRhubarb(); |
| 29 | const PORT = process.env.PORT || 3123; |
| 30 | |
| 31 | // Format-agnostic asset resolution: an asset id resolves to the first |
| 32 | // existing file among a known extension set (SVG first, then raster), so |
| 33 | // PNG/JPG/WebP/GIF drop into the same slots as SVG with no code change. |
| 34 | const ASSET_EXTS = ['.svg', '.png', '.jpg', '.jpeg', '.webp', '.gif']; |
| 35 | |
| 36 | // music/sfx assets are JSON note patterns (see docs/design/actors-wearables- |
| 37 | // music.md), not images, so they resolve against their own extension set. |
| 38 | const AUDIO_KINDS = ['music', 'sfx']; |
| 39 | const AUDIO_EXTS = ['.json']; |
| 40 | |
| 41 | function resolveAsset(kind, id) { |
| 42 | const exts = AUDIO_KINDS.includes(kind) ? AUDIO_EXTS : ASSET_EXTS; |
| 43 | for (const ext of exts) { |
| 44 | const file = path.join(ASSETS, kind, id + ext); |
| 45 | if (fs.existsSync(file)) return file; |
| 46 | } |
| 47 | return null; |
| 48 | } |
| 49 | |
| 50 | function findRhubarb() { |
| 51 | if (process.env.RHUBARB_PATH) return process.env.RHUBARB_PATH; |
| 52 | const toolsDir = path.join(ROOT, 'tools'); |
| 53 | try { |
| 54 | for (const entry of fs.readdirSync(toolsDir)) { |
| 55 | const candidate = path.join(toolsDir, entry, 'rhubarb'); |
| 56 | if (fs.existsSync(candidate)) return candidate; |
| 57 | } |
| 58 | } catch { /* no tools dir yet */ } |
| 59 | return 'rhubarb'; // hope it's on PATH; probeRhubarb() handles absence |
| 60 | } |
| 61 | |
| 62 | fs.mkdirSync(CACHE, { recursive: true }); |
| 63 | |
| 64 | // ---------------------------------------------------------------- helpers |
| 65 | |
| 66 | function run(cmd, args, opts = {}) { |
| 67 | return new Promise((resolve, reject) => { |
| 68 | execFile(cmd, args, { maxBuffer: 16 * 1024 * 1024, ...opts }, (err, stdout, stderr) => { |
| 69 | if (err) reject(new Error(`${cmd} failed: ${stderr || err.message}`)); |
| 70 | else resolve(stdout); |
| 71 | }); |
| 72 | }); |
| 73 | } |
| 74 | |
| 75 | function sha1(s) { |
| 76 | return crypto.createHash('sha1').update(s).digest('hex').slice(0, 16); |
| 77 | } |
| 78 | |
| 79 | function readBody(req) { |
| 80 | return new Promise((resolve, reject) => { |
| 81 | let data = ''; |
| 82 | req.on('data', (c) => { data += c; if (data.length > 1e6) req.destroy(); }); |
| 83 | req.on('end', () => { |
| 84 | try { resolve(data ? JSON.parse(data) : {}); } catch (e) { reject(e); } |
| 85 | }); |
| 86 | req.on('error', reject); |
| 87 | }); |
| 88 | } |
| 89 | |
| 90 | function readRaw(req, limit = 32 * 1024 * 1024) { |
| 91 | return new Promise((resolve, reject) => { |
| 92 | const chunks = []; |
| 93 | let size = 0; |
| 94 | req.on('data', (c) => { |
| 95 | size += c.length; |
| 96 | if (size > limit) { req.destroy(); reject(new Error('body too large')); return; } |
| 97 | chunks.push(c); |
| 98 | }); |
| 99 | req.on('end', () => resolve(Buffer.concat(chunks))); |
| 100 | req.on('error', reject); |
| 101 | }); |
| 102 | } |
| 103 | |
| 104 | function json(res, code, obj) { |
| 105 | const body = JSON.stringify(obj); |
| 106 | res.writeHead(code, { 'Content-Type': 'application/json', 'Content-Length': Buffer.byteLength(body) }); |
| 107 | res.end(body); |
| 108 | } |
| 109 | |
| 110 | // ------------------------------------------------------------ SSE clients |
| 111 | |
| 112 | const clients = new Set(); |
| 113 | |
| 114 | function broadcast(cue) { |
| 115 | const msg = `data: ${JSON.stringify(cue)}\n\n`; |
| 116 | for (const res of clients) res.write(msg); |
| 117 | } |
| 118 | |
| 119 | setInterval(() => { for (const res of clients) res.write(': ping\n\n'); }, 25000); |
| 120 | |
| 121 | // ------------------------------------------------------------- TTS + sync |
| 122 | |
| 123 | let rhubarbOk = null; // lazily probed |
| 124 | let lastSlowpx = 0; // Date.now() at the start of the most recent /slowpx request (nav marker for screenshot timing) |
| 125 | |
| 126 | async function probeRhubarb() { |
| 127 | if (rhubarbOk !== null) return rhubarbOk; |
| 128 | try { await run(RHUBARB, ['--version']); rhubarbOk = true; } |
| 129 | catch { rhubarbOk = false; console.warn('rhubarb unavailable; falling back to amplitude lip-sync'); } |
| 130 | return rhubarbOk; |
| 131 | } |
| 132 | |
| 133 | // Per-key in-flight render dedup: two concurrent callers for the same cache |
| 134 | // key (e.g. a script re-performed while its render-ahead window is still |
| 135 | // in flight, or /api/say racing that same render-ahead) must not both render |
| 136 | // to the same raw temp path — the second render's ffmpeg step can lose the |
| 137 | // race against the first's cleanup unlink ("ffmpeg failed"). Instead, the |
| 138 | // second caller awaits the first's promise and both read the same result. |
| 139 | const speechInFlight = new Map(); // key -> Promise<void> |
| 140 | const externalAudioInFlight = new Map(); // key -> Promise<void> |
| 141 | |
| 142 | /** |
| 143 | * Render `text` with the given engine and produce {audio, timeline, duration}. |
| 144 | * Results are cached by content hash in cache/. |
| 145 | */ |
| 146 | async function prepareSpeech({ text, engine = 'say', voice = '', rate = 0 }) { |
| 147 | engine = engine === 'espeak' ? 'espeak' : 'say'; |
| 148 | const key = sha1([engine, voice, rate, text].join('|')); |
| 149 | const wav = path.join(CACHE, `${key}.wav`); |
| 150 | const meta = path.join(CACHE, `${key}.meta.json`); |
| 151 | |
| 152 | if (!fs.existsSync(wav) || !fs.existsSync(meta)) { |
| 153 | let p = speechInFlight.get(key); |
| 154 | if (!p) { |
| 155 | p = renderSpeech(key, wav, meta, { text, engine, voice, rate }); |
| 156 | speechInFlight.set(key, p); |
| 157 | p.finally(() => { |
| 158 | if (speechInFlight.get(key) === p) speechInFlight.delete(key); |
| 159 | }); |
| 160 | } |
| 161 | await p; |
| 162 | } |
| 163 | |
| 164 | const { timeline, duration } = JSON.parse(fs.readFileSync(meta, 'utf8')); |
| 165 | return { audio: `/audio/${key}.wav`, timeline, duration, text }; |
| 166 | } |
| 167 | |
| 168 | // Does the actual say/espeak -> ffmpeg -> rhubarb render for prepareSpeech. |
| 169 | // Split out so prepareSpeech can dedup concurrent callers on `key` instead |
| 170 | // of each starting its own render into the same raw temp path. |
| 171 | async function renderSpeech(key, wav, meta, { text, engine, voice, rate }) { |
| 172 | const raw = path.join(CACHE, `${key}.raw.${engine === 'say' ? 'aiff' : 'wav'}`); |
| 173 | if (engine === 'say') { |
| 174 | const args = ['-o', raw]; |
| 175 | if (voice) args.push('-v', voice); |
| 176 | if (rate) args.push('-r', String(rate)); |
| 177 | args.push(text); |
| 178 | await run('say', args); |
| 179 | } else { |
| 180 | const args = ['-w', raw]; |
| 181 | if (voice) args.push('-v', voice); |
| 182 | if (rate) args.push('-s', String(rate)); |
| 183 | args.push(text); |
| 184 | await run('espeak', args); |
| 185 | } |
| 186 | await run('ffmpeg', ['-y', '-loglevel', 'error', '-i', raw, |
| 187 | '-ar', '16000', '-ac', '1', '-c:a', 'pcm_s16le', wav]); |
| 188 | fs.unlinkSync(raw); |
| 189 | |
| 190 | let timeline = null; |
| 191 | let duration = 0; |
| 192 | if (await probeRhubarb()) { |
| 193 | try { |
| 194 | const dialog = path.join(CACHE, `${key}.txt`); |
| 195 | fs.writeFileSync(dialog, text); |
| 196 | const out = await run(RHUBARB, ['-f', 'json', '--extendedShapes', 'GHX', |
| 197 | '--dialogFile', dialog, wav]); |
| 198 | const parsed = JSON.parse(out); |
| 199 | timeline = parsed.mouthCues; |
| 200 | duration = parsed.metadata.duration; |
| 201 | fs.unlinkSync(dialog); |
| 202 | } catch (e) { |
| 203 | console.warn('rhubarb failed for this clip:', e.message); |
| 204 | } |
| 205 | } |
| 206 | if (!duration) { |
| 207 | const probe = await run('ffprobe', ['-v', 'error', '-show_entries', |
| 208 | 'format=duration', '-of', 'csv=p=0', wav]); |
| 209 | duration = parseFloat(probe) || 0; |
| 210 | } |
| 211 | fs.writeFileSync(meta, JSON.stringify({ timeline, duration })); |
| 212 | } |
| 213 | |
| 214 | /** |
| 215 | * Lip-sync audio rendered elsewhere (e.g. an external TTS service): |
| 216 | * normalize it, run Rhubarb (using `text` as a dialog hint if given), |
| 217 | * and return the same {audio, timeline, duration} shape as prepareSpeech. |
| 218 | */ |
| 219 | async function prepareExternalAudio(bytes, text = '') { |
| 220 | const key = sha1('ext|' + text + '|' + crypto.createHash('sha1').update(bytes).digest('hex')); |
| 221 | const wav = path.join(CACHE, `${key}.wav`); |
| 222 | const meta = path.join(CACHE, `${key}.meta.json`); |
| 223 | |
| 224 | if (!fs.existsSync(wav) || !fs.existsSync(meta)) { |
| 225 | let p = externalAudioInFlight.get(key); |
| 226 | if (!p) { |
| 227 | p = renderExternalAudio(key, wav, meta, bytes, text); |
| 228 | externalAudioInFlight.set(key, p); |
| 229 | p.finally(() => { |
| 230 | if (externalAudioInFlight.get(key) === p) externalAudioInFlight.delete(key); |
| 231 | }); |
| 232 | } |
| 233 | await p; |
| 234 | } |
| 235 | |
| 236 | const { timeline, duration } = JSON.parse(fs.readFileSync(meta, 'utf8')); |
| 237 | return { audio: `/audio/${key}.wav`, timeline, duration, text }; |
| 238 | } |
| 239 | |
| 240 | // Does the actual ffmpeg -> rhubarb render for prepareExternalAudio. Split |
| 241 | // out so prepareExternalAudio can dedup concurrent callers on `key` instead |
| 242 | // of each starting its own render into the same raw temp path. |
| 243 | async function renderExternalAudio(key, wav, meta, bytes, text) { |
| 244 | const raw = path.join(CACHE, `${key}.raw`); |
| 245 | fs.writeFileSync(raw, bytes); |
| 246 | await run('ffmpeg', ['-y', '-loglevel', 'error', '-i', raw, |
| 247 | '-ar', '16000', '-ac', '1', '-c:a', 'pcm_s16le', wav]); |
| 248 | fs.unlinkSync(raw); |
| 249 | |
| 250 | let timeline = null; |
| 251 | let duration = 0; |
| 252 | if (await probeRhubarb()) { |
| 253 | try { |
| 254 | const args = ['-f', 'json', '--extendedShapes', 'GHX']; |
| 255 | if (text) { |
| 256 | const dialog = path.join(CACHE, `${key}.txt`); |
| 257 | fs.writeFileSync(dialog, text); |
| 258 | args.push('--dialogFile', dialog); |
| 259 | } |
| 260 | const out = await run(RHUBARB, [...args, wav]); |
| 261 | const parsed = JSON.parse(out); |
| 262 | timeline = parsed.mouthCues; |
| 263 | duration = parsed.metadata.duration; |
| 264 | } catch (e) { |
| 265 | console.warn('rhubarb failed for external audio:', e.message); |
| 266 | } |
| 267 | } |
| 268 | if (!duration) { |
| 269 | const probe = await run('ffprobe', ['-v', 'error', '-show_entries', |
| 270 | 'format=duration', '-of', 'csv=p=0', wav]); |
| 271 | duration = parseFloat(probe) || 0; |
| 272 | } |
| 273 | fs.writeFileSync(meta, JSON.stringify({ timeline, duration })); |
| 274 | } |
| 275 | |
| 276 | // ---------------------------------------------------------- script runner |
| 277 | |
| 278 | const sleep = (ms) => new Promise((r) => setTimeout(r, ms)); |
| 279 | |
| 280 | // rough pacing estimates for sequencing; the browser animates the real thing |
| 281 | const ACTION_SECONDS = { wave: 1.6, jump: 0.9, nod: 1.2, shake: 1.2, bow: 1.8, dance: 2.4 }; |
| 282 | |
| 283 | // Layout presets: macros that clear existing frames and create the named |
| 284 | // ones. Only the frame-id bookkeeping needed while parsing lives here — the |
| 285 | // actual rect math is the client's job (public/stage.js). |
| 286 | const LAYOUT_FRAMES = { |
| 287 | single: ['main'], |
| 288 | split: ['left', 'right'], |
| 289 | stack: ['top', 'bottom'], |
| 290 | thirds: ['third-l', 'third-c', 'third-r'], |
| 291 | 'pip-tr': ['main', 'pip'], 'pip-tl': ['main', 'pip'], |
| 292 | 'pip-br': ['main', 'pip'], 'pip-bl': ['main', 'pip'], |
| 293 | }; |
| 294 | |
| 295 | // Tokenize `key:value key:"quoted value"` pairs, e.g. `character:ava bg:desk`. |
| 296 | function parseKV(str) { |
| 297 | const out = {}; |
| 298 | const re = /([a-z][\w-]*):("([^"]*)"|\S+)/gi; |
| 299 | let m; |
| 300 | while ((m = re.exec(str))) out[m[1].toLowerCase()] = m[3] !== undefined ? m[3] : m[2]; |
| 301 | return out; |
| 302 | } |
| 303 | |
| 304 | // left/center|middle/right resolve to stage-percent positions (25/50/75, |
| 305 | // case-insensitive); anything else parses as a literal percent number. |
| 306 | // Used wherever a direction's x might be authored as a word instead of a |
| 307 | // number: [place ... at <pos>], [walk to <pos>], [<id> walk <pos>], |
| 308 | // [<id> move <pos>]. Does NOT apply to [enter/exit] left|right, which mean |
| 309 | // off-stage sides, not on-stage positions. |
| 310 | function parsePosWord(v) { |
| 311 | const w = String(v == null ? '' : v).toLowerCase(); |
| 312 | if (w === 'left') return 25; |
| 313 | if (w === 'center' || w === 'middle') return 50; |
| 314 | if (w === 'right') return 75; |
| 315 | return parseFloat(v); |
| 316 | } |
| 317 | |
| 318 | // Resolves a bare id/name to the frame or actor it refers to: a declared |
| 319 | // frame id or actor id wins as-is; otherwise, if `id` names a character |
| 320 | // who's been framed/placed (tracked in charLocs, see parseScript), resolve |
| 321 | // through that mapping. Falls back to `id` itself (unresolved) so callers |
| 322 | // keep their existing "unknown target" handling for genuinely bad input. |
| 323 | function resolveTarget(id, frameIds, actorIds, charLocs) { |
| 324 | if (frameIds.has(id) || actorIds.has(id)) return id; |
| 325 | const loc = charLocs && charLocs.get(id); |
| 326 | return loc ? loc.id : id; |
| 327 | } |
| 328 | |
| 329 | // Speaker-prefix resolution shared by plain "<id>: text" lines and the |
| 330 | // bracket-line tolerance in parseScript: frame id -> actor id -> character |
| 331 | // name (via charLocs). Returns {frame|actor, text} or null if the label |
| 332 | // isn't a known target, in which case the string is left alone by the |
| 333 | // caller (spoken/directed verbatim, prefix and all). |
| 334 | function matchSpeaker(text, frameIds, actorIds, charLocs) { |
| 335 | const m = text.match(/^([A-Za-z][\w-]*):\s*(.*)$/); |
| 336 | if (!m) return null; |
| 337 | if (frameIds.has(m[1])) return { frame: m[1], text: m[2] }; |
| 338 | if (actorIds.has(m[1])) return { actor: m[1], text: m[2] }; |
| 339 | const loc = charLocs && charLocs.get(m[1]); |
| 340 | if (loc) return loc.kind === 'frame' ? { frame: loc.id, text: m[2] } : { actor: loc.id, text: m[2] }; |
| 341 | return null; |
| 342 | } |
| 343 | |
| 344 | /** |
| 345 | * Parse a screenplay into a cue list. |
| 346 | * [wave] → action cue |
| 347 | * [walk to 70] → movement (percent of stage width; also accepts |
| 348 | * left|center|middle|right as 25/50/75) |
| 349 | * [look left|right|front], [emote happy], [wait 1.5] |
| 350 | * [engine espeak], [voice Samantha], [rate 180] → speech settings |
| 351 | * [layout split] → frame layout preset |
| 352 | * [frame left character:ava bg:desk view:face] → create/update a frame |
| 353 | * [frame right clear] → remove a frame |
| 354 | * [scene desk] → set the active frame's background |
| 355 | * [show image:chart fit:contain] → content tile in the active frame |
| 356 | * [lower-third "Ava Reyes" "Host"] → overlay, auto-dismissed after a |
| 357 | * 6000ms hold by default; add a trailing |
| 358 | * `hold:<ms>` to override (`hold:0` persists — |
| 359 | * no auto-dismiss — matching the old behavior) |
| 360 | * [lower-third clear] → remove the current lower-third early |
| 361 | * [clear] → wipe every placed actor (and its worn props), |
| 362 | * every frame's content tile, and every overlay |
| 363 | * across the whole stage — the same cleanup a |
| 364 | * new script runs automatically on script-start. |
| 365 | * Music, captions on/off, backgrounds, and frame |
| 366 | * layout are left alone. |
| 367 | * [captions on|off] → toggle the caption bar (default on) |
| 368 | * [iris out|in 900], [fade out|in 900] → fullscreen transition |
| 369 | * (ms optional, default 700) |
| 370 | * left: Good evening. → speak cue targeting frame "left" (id must be |
| 371 | * a frame declared earlier in the script); once |
| 372 | * a character has been framed/placed, its NAME |
| 373 | * works the same way — `rex: Woof!` resolves to |
| 374 | * whichever frame/actor rex currently occupies |
| 375 | * (resolution order: frame id, actor id, name) |
| 376 | * [left: Good evening.] → same as the line above, tolerated: a bracketed |
| 377 | * line whose inner text is itself a speaker |
| 378 | * prefix is unwrapped to a spoken line rather |
| 379 | * than treated as an unknown direction |
| 380 | * [left wave] → direction targeting frame "left"; a placed/ |
| 381 | * framed character's name also routes here, |
| 382 | * e.g. [rex sit] once rex occupies a frame/actor |
| 383 | * (wave) Hello! → action fired concurrently with the line |
| 384 | * Hello! → speech |
| 385 | * |
| 386 | * [place hoop at 70 scale:0.5] → place a prop/character actor |
| 387 | * ("at" required — a percent number or |
| 388 | * left|center|middle|right (25/50/75); "id:" |
| 389 | * defaults to the name, "scale:" defaults 1 for |
| 390 | * characters/0.4 for props, bare "behind" flag |
| 391 | * draws it behind the frame's primary character) |
| 392 | * [place rex at 15 scale:0.6 id:sidekick] → `what` names a folder in |
| 393 | * characters/ → a full rig; otherwise it |
| 394 | * resolves as assets/props/<what> |
| 395 | * [remove sidekick] → destroy a placed actor |
| 396 | * [<id> move 40], [<id> scale 0.8], [<id> spin], [<id> bounce] → prop |
| 397 | * actor directions (glide/resize/spin/hop) |
| 398 | * sidekick: Woof! → speak cue targeting an actor id (same |
| 399 | * resolution order as frame ids: frame first, |
| 400 | * then actor) |
| 401 | * [sidekick leap] → direction on a character actor — the full |
| 402 | * direction set applies (walk/emote/actions/ |
| 403 | * speak), same grammar as a frame id |
| 404 | * [wear right tophat], [wear right tophat scale:1.2 anchor:head], |
| 405 | * [unwear right] → pin/remove a prop asset at a character's |
| 406 | * named anchor (`target` is a frame id — its |
| 407 | * primary character — an actor id, or (once |
| 408 | * framed/placed) a character name; anchor |
| 409 | * defaults to "head") |
| 410 | * [music circus] → cross-fade to assets/music/circus.json |
| 411 | * (a looping JSON note pattern), stage-global |
| 412 | * [music off] → fade out and stop the current music |
| 413 | * [sfx tada] → one-shot assets/sfx/tada.json, not looped, |
| 414 | * not ducked |
| 415 | */ |
| 416 | function parseScript(src) { |
| 417 | const cues = []; |
| 418 | const frameIds = new Set(['main']); // tracks declared frame ids as we go |
| 419 | const actorIds = new Set(); // tracks ids declared by [place ...] as we go |
| 420 | // character name -> {kind:'frame'|'actor', id} it currently occupies, so |
| 421 | // writers can address characters by name once they've been framed/placed |
| 422 | // (`[rex sit]`, `rex: Woof!`, `[wear bo tophat]`); later assignments |
| 423 | // override earlier ones. Populated by directionCue's frame/place handling. |
| 424 | const charLocs = new Map(); |
| 425 | for (const rawLine of src.split('\n')) { |
| 426 | const line = rawLine.trim(); |
| 427 | if (!line || line.startsWith('#')) continue; |
| 428 | |
| 429 | if (/^(\[[^\]]+\]\s*)+$/.test(line)) { |
| 430 | for (const m of line.matchAll(/\[([^\]]+)\]/g)) { |
| 431 | const inner = m[1]; |
| 432 | // Tolerance: `[left: some words]` / `[<name>: words]` is a spoken |
| 433 | // line wearing brackets by mistake (common in LLM-authored |
| 434 | // screenplays) — unwrap it instead of treating it as an unknown |
| 435 | // direction, but only when the label actually resolves. |
| 436 | const sp = matchSpeaker(inner, frameIds, actorIds, charLocs); |
| 437 | if (sp) { |
| 438 | const cue = { type: 'speak', text: sp.text, concurrent: null, source: line }; |
| 439 | if (sp.frame) cue.frame = sp.frame; |
| 440 | if (sp.actor) cue.actor = sp.actor; |
| 441 | cues.push(cue); |
| 442 | } else { |
| 443 | cues.push(directionCue(inner, line, frameIds, actorIds, charLocs)); |
| 444 | } |
| 445 | } |
| 446 | continue; |
| 447 | } |
| 448 | |
| 449 | let text = line; |
| 450 | let frame = null; |
| 451 | let actor = null; |
| 452 | const sp = matchSpeaker(text, frameIds, actorIds, charLocs); |
| 453 | if (sp) { frame = sp.frame || null; actor = sp.actor || null; text = sp.text; } |
| 454 | |
| 455 | let concurrent = null; |
| 456 | const inline = text.match(/^\(([^)]+)\)\s*(.*)$/); |
| 457 | if (inline) { |
| 458 | concurrent = directionCue(inline[1], line, frameIds, actorIds, charLocs); |
| 459 | text = inline[2]; |
| 460 | if (frame) concurrent.frame = frame; |
| 461 | if (actor) concurrent.actor = actor; |
| 462 | } |
| 463 | const cue = { type: 'speak', text, concurrent, source: line }; |
| 464 | if (frame) cue.frame = frame; |
| 465 | if (actor) cue.actor = actor; |
| 466 | cues.push(cue); |
| 467 | } |
| 468 | return cues; |
| 469 | } |
| 470 | |
| 471 | function directionCue(body, source, frameIds, actorIds, charLocs) { |
| 472 | frameIds = frameIds || new Set(['main']); |
| 473 | actorIds = actorIds || new Set(); |
| 474 | charLocs = charLocs || new Map(); |
| 475 | const parts = body.trim().split(/\s+/); |
| 476 | const head = parts[0].toLowerCase(); |
| 477 | |
| 478 | if (head === 'layout') { |
| 479 | const preset = (parts[1] || 'single').toLowerCase(); |
| 480 | const ids = LAYOUT_FRAMES[preset] || ['main']; |
| 481 | frameIds.clear(); |
| 482 | for (const id of ids) frameIds.add(id); |
| 483 | return { type: 'layout', preset, source }; |
| 484 | } |
| 485 | |
| 486 | if (head === 'frame') { |
| 487 | const id = parts[1]; |
| 488 | const restStr = parts.slice(2).join(' '); |
| 489 | if (/^clear\b/i.test(restStr)) { |
| 490 | frameIds.delete(id); |
| 491 | return { type: 'frame-clear', id, source }; |
| 492 | } |
| 493 | const kv = parseKV(restStr); |
| 494 | const cue = { type: 'frame', id, source }; |
| 495 | if (kv.slot) cue.slot = kv.slot; |
| 496 | if (kv.bg) cue.bg = kv.bg; |
| 497 | if (kv.character) cue.character = kv.character; |
| 498 | if (kv.view) cue.view = kv.view; |
| 499 | if (kv.facing !== undefined) cue.facing = parseInt(kv.facing, 10); |
| 500 | if (kv.rect) cue.rect = kv.rect.split(',').map(Number); |
| 501 | frameIds.add(id); |
| 502 | if (kv.character) charLocs.set(kv.character, { kind: 'frame', id }); |
| 503 | return cue; |
| 504 | } |
| 505 | |
| 506 | if (head === 'scene') return { type: 'scene', bg: parts.slice(1).join(' '), source }; |
| 507 | |
| 508 | if (head === 'show') { |
| 509 | const kv = parseKV(parts.slice(1).join(' ')); |
| 510 | let kind = 'text', value = ''; |
| 511 | for (const k of ['text', 'image', 'video']) { |
| 512 | if (kv[k] !== undefined) { kind = k; value = kv[k]; break; } |
| 513 | } |
| 514 | const cue = { type: 'content', kind, value, source }; |
| 515 | if (kv.fit) cue.fit = kv.fit; |
| 516 | return cue; |
| 517 | } |
| 518 | |
| 519 | if (head === 'lower-third') { |
| 520 | // stable id so a new lower-third replaces the current one instead of |
| 521 | // stacking; `[lower-third clear]` removes it. |
| 522 | if ((parts[1] || '').toLowerCase() === 'clear' && !body.includes('"')) { |
| 523 | return { type: 'overlay-clear', id: 'lower-third', source }; |
| 524 | } |
| 525 | const strs = [...body.matchAll(/"([^"]*)"/g)].map((m) => m[1]); |
| 526 | const cue = { |
| 527 | type: 'overlay', template: 'lower-third', id: 'lower-third', |
| 528 | slots: { title: strs[0] || '', subtitle: strs[1] || '' }, source, |
| 529 | }; |
| 530 | // Optional trailing `hold:<ms>` key: default 6000 (auto-dismiss after |
| 531 | // 6s) when omitted; `hold:0` means persist (omit hold entirely, same as |
| 532 | // the old un-timed behavior); any other value is that hold in ms. |
| 533 | const holdMatch = body.match(/\bhold:(\d+)\b/i); |
| 534 | const holdMs = holdMatch ? parseInt(holdMatch[1], 10) : 6000; |
| 535 | if (holdMs > 0) cue.hold = holdMs; |
| 536 | return cue; |
| 537 | } |
| 538 | |
| 539 | if (head === 'walk' || head === 'enter' || head === 'exit') { |
| 540 | // [walk to 70] [walk to center] [enter from left] [exit right] |
| 541 | const arg = parts[parts.length - 1].toLowerCase(); |
| 542 | // walk's left/right are on-stage position words (25/75); enter/exit's |
| 543 | // left/right stay off-stage sides — handled separately below. |
| 544 | let x = head === 'walk' ? parsePosWord(arg) : parseFloat(arg); |
| 545 | if (head === 'enter') x = 50; |
| 546 | else if (head === 'exit' && isNaN(x)) x = arg === 'left' ? -15 : 115; |
| 547 | const from = parts.includes('from') ? (parts[parts.indexOf('from') + 1] || '') : null; |
| 548 | return { type: 'walk', x, from, jump: head === 'enter' ? from : null, source }; |
| 549 | } |
| 550 | if (head === 'wait') return { type: 'wait', seconds: parseFloat(parts[1]) || 1, source }; |
| 551 | if (head === 'view' || head === 'zoom') { |
| 552 | return { type: 'view', mode: (parts[1] || 'body').toLowerCase(), source }; |
| 553 | } |
| 554 | if (head === 'look') return { type: 'look', dir: parts[1] || 'front', source }; |
| 555 | if (head === 'emote') return { type: 'action', name: parts[1] || 'neutral', source }; |
| 556 | if (head === 'engine' || head === 'voice' || head === 'rate') { |
| 557 | return { type: 'setting', key: head, value: parts.slice(1).join(' '), source }; |
| 558 | } |
| 559 | if (head === 'captions') { |
| 560 | const arg = (parts[1] || 'on').toLowerCase(); |
| 561 | return { type: 'captions', on: arg !== 'off', source }; |
| 562 | } |
| 563 | if (head === 'iris' || head === 'fade') { |
| 564 | const ms = parseInt(parts[2], 10); |
| 565 | return { type: 'transition', name: head, dir: (parts[1] || 'out').toLowerCase(), ms: isNaN(ms) ? 700 : ms, source }; |
| 566 | } |
| 567 | |
| 568 | if (head === 'place') { |
| 569 | // [place <what> at <x> [id:<id>] [scale:<n>] [behind]] — <x> accepts a |
| 570 | // percent number or left|center|middle|right (25/50/75). |
| 571 | const rest = parts.slice(1); |
| 572 | const what = rest[0]; |
| 573 | const atIdx = rest.indexOf('at'); |
| 574 | const x = atIdx >= 0 ? parsePosWord(rest[atIdx + 1]) : NaN; |
| 575 | const restStr = parts.slice(1).join(' '); |
| 576 | const kv = parseKV(restStr); |
| 577 | const id = kv.id || what; |
| 578 | const cue = { type: 'place', id, what, x, source }; |
| 579 | if (kv.scale !== undefined) cue.scale = parseFloat(kv.scale); |
| 580 | if (/\bbehind\b/.test(restStr)) cue.behind = true; |
| 581 | actorIds.add(id); |
| 582 | // `what` names a character iff it's a folder in characters/ — track |
| 583 | // where it now lives so later `[<what> ...]`/`what: text` addresses it. |
| 584 | if (fs.existsSync(path.join(CHARACTERS, what))) charLocs.set(what, { kind: 'actor', id }); |
| 585 | return cue; |
| 586 | } |
| 587 | |
| 588 | if (head === 'remove') { |
| 589 | const id = parts[1]; |
| 590 | actorIds.delete(id); |
| 591 | return { type: 'remove', id, source }; |
| 592 | } |
| 593 | |
| 594 | if (head === 'wear') { |
| 595 | // [wear <target> <prop> [scale:<n>] [anchor:<name>]] — target resolves |
| 596 | // frame id -> actor id -> character name (see resolveTarget). |
| 597 | const target = resolveTarget(parts[1], frameIds, actorIds, charLocs); |
| 598 | const prop = parts[2]; |
| 599 | const kv = parseKV(parts.slice(3).join(' ')); |
| 600 | const cue = { type: 'wear', target, prop, source }; |
| 601 | if (kv.scale !== undefined) cue.scale = parseFloat(kv.scale); |
| 602 | if (kv.anchor !== undefined) cue.anchor = kv.anchor; |
| 603 | return cue; |
| 604 | } |
| 605 | |
| 606 | if (head === 'unwear') { |
| 607 | const target = resolveTarget(parts[1], frameIds, actorIds, charLocs); |
| 608 | const cue = { type: 'unwear', target, source }; |
| 609 | if (parts[2]) cue.anchor = parts[2]; |
| 610 | return cue; |
| 611 | } |
| 612 | |
| 613 | if (head === 'music') { |
| 614 | const arg = (parts[1] || '').toLowerCase(); |
| 615 | if (arg === 'off') return { type: 'music', off: true, source }; |
| 616 | return { type: 'music', id: parts[1], source }; |
| 617 | } |
| 618 | |
| 619 | if (head === 'sfx') return { type: 'sfx', id: parts[1], source }; |
| 620 | |
| 621 | // Prop-actor directions — only meaningful when the head token that routed |
| 622 | // here was an actor id (see the actorIds fallback below), e.g. |
| 623 | // [hoop1 move 40] recurses into directionCue("move 40", ...). |
| 624 | if (head === 'move') return { type: 'move', x: parsePosWord(parts[1]), source }; |
| 625 | if (head === 'scale') return { type: 'scale', value: parseFloat(parts[1]), source }; |
| 626 | if (head === 'spin') return { type: 'spin', source }; |
| 627 | if (head === 'bounce') return { type: 'bounce', source }; |
| 628 | |
| 629 | // [<id> <direction...>] — first token is a known frame id: apply the rest |
| 630 | // of the direction to that frame. |
| 631 | if (frameIds.has(head) && parts.length > 1) { |
| 632 | const sub = directionCue(parts.slice(1).join(' '), source, frameIds, actorIds, charLocs); |
| 633 | sub.frame = head; |
| 634 | return sub; |
| 635 | } |
| 636 | |
| 637 | // Same, for a placed actor id (checked after frame ids, per spec). |
| 638 | if (actorIds.has(head) && parts.length > 1) { |
| 639 | const sub = directionCue(parts.slice(1).join(' '), source, frameIds, actorIds, charLocs); |
| 640 | sub.actor = head; |
| 641 | return sub; |
| 642 | } |
| 643 | |
| 644 | // Same, for a character NAME that's been framed/placed (checked after |
| 645 | // frame ids and actor ids, per spec) — e.g. [rex sit] once rex occupies a |
| 646 | // frame or actor slot; resolves to whichever one it currently occupies. |
| 647 | { |
| 648 | const loc = charLocs.get(head); |
| 649 | if (loc && parts.length > 1) { |
| 650 | const sub = directionCue(parts.slice(1).join(' '), source, frameIds, actorIds, charLocs); |
| 651 | if (loc.kind === 'frame') sub.frame = loc.id; else sub.actor = loc.id; |
| 652 | return sub; |
| 653 | } |
| 654 | } |
| 655 | |
| 656 | // [clear] — wipe placed actors/worn props/content tiles/overlays across |
| 657 | // every frame, on demand mid-script (same cleanup `script-start` runs |
| 658 | // automatically). Music, captions on/off, backgrounds, and frame layout |
| 659 | // are untouched. |
| 660 | if (head === 'clear') return { type: 'clear', source }; |
| 661 | |
| 662 | return { type: 'action', name: head, source }; |
| 663 | } |
| 664 | |
| 665 | let scriptToken = 0; |
| 666 | |
| 667 | // Per-character declared voice, e.g. characters/ava/manifest.json's |
| 668 | // `"voice": {"engine":"say","voice":"Samantha"}`. Cached by character name |
| 669 | // so screenplays with many lines don't re-read manifests per cue. |
| 670 | const characterVoiceCache = {}; |
| 671 | function characterVoice(name) { |
| 672 | if (!name) return null; |
| 673 | if (Object.prototype.hasOwnProperty.call(characterVoiceCache, name)) return characterVoiceCache[name]; |
| 674 | let voice = null; |
| 675 | try { |
| 676 | const manifest = JSON.parse(fs.readFileSync(path.join(CHARACTERS, name, 'manifest.json'), 'utf8')); |
| 677 | if (manifest && manifest.voice) voice = manifest.voice; |
| 678 | } catch { /* missing/unreadable manifest; fall back to default voice */ } |
| 679 | characterVoiceCache[name] = voice; |
| 680 | return voice; |
| 681 | } |
| 682 | |
| 683 | // Single ordered pass over a parsed cue list that replays the same |
| 684 | // settings/frameChar/override-flag state the (formerly inline) speak handler |
| 685 | // used, and attaches the fully-resolved TTS params to every speak cue as |
| 686 | // `cue.speech`. Pulling this out lets runScript render speech ahead of the |
| 687 | // playback loop instead of only when a line is about to play. |
| 688 | function resolveSpeech(cues, mainCharacter) { |
| 689 | const settings = { engine: 'say', voice: '', rate: 0 }; |
| 690 | let engineOverridden = false; |
| 691 | let voiceOverridden = false; |
| 692 | // frame id -> character name, so each spoken line can resolve the voice of |
| 693 | // whoever is actually in that frame. Seeded with the client's notion of |
| 694 | // who's in the main frame, so solo screenplays pick up that character's |
| 695 | // declared voice too. |
| 696 | const frameChar = {}; |
| 697 | if (mainCharacter) frameChar.main = mainCharacter; |
| 698 | // actor id -> character name, for actor-targeted speak cues (mirrors |
| 699 | // frameChar but seeded from [place] cues instead of [frame]/[character]). |
| 700 | const actorChar = {}; |
| 701 | |
| 702 | for (const cue of cues) { |
| 703 | if (cue.type === 'setting') { |
| 704 | settings[cue.key] = cue.value; |
| 705 | if (cue.key === 'engine') engineOverridden = true; |
| 706 | if (cue.key === 'voice') voiceOverridden = true; |
| 707 | continue; |
| 708 | } |
| 709 | if (cue.type === 'frame' && cue.character) frameChar[cue.id] = cue.character; |
| 710 | if (cue.type === 'character') frameChar[cue.frame || 'main'] = cue.name; |
| 711 | if (cue.type === 'place') { |
| 712 | // `what` names a character iff it's a folder in characters/; a prop |
| 713 | // actor placed over a stale id clears any earlier character mapping. |
| 714 | if (fs.existsSync(path.join(CHARACTERS, cue.what))) actorChar[cue.id] = cue.what; |
| 715 | else delete actorChar[cue.id]; |
| 716 | } |
| 717 | if (cue.type === 'remove') delete actorChar[cue.id]; |
| 718 | |
| 719 | if (cue.type === 'speak') { |
| 720 | // Resolution: an explicit [engine]/[voice] direction always wins; |
| 721 | // otherwise use the declared voice of the character in this line's |
| 722 | // actor (if actor-targeted) or frame; otherwise fall back to the |
| 723 | // (default) settings. |
| 724 | const charName = cue.actor ? actorChar[cue.actor] : frameChar[cue.frame || 'main']; |
| 725 | const cv = characterVoice(charName); |
| 726 | const engine = engineOverridden ? settings.engine : (cv && cv.engine) || settings.engine; |
| 727 | const voice = voiceOverridden ? settings.voice : (cv && cv.voice !== undefined ? cv.voice : settings.voice); |
| 728 | cue.speech = { text: cue.text, engine, voice, rate: settings.rate }; |
| 729 | } |
| 730 | } |
| 731 | } |
| 732 | |
| 733 | const RENDER_WINDOW = 4; |
| 734 | |
| 735 | async function runScript(src, mainCharacter) { |
| 736 | const token = ++scriptToken; |
| 737 | const cues = parseScript(src); |
| 738 | resolveSpeech(cues, mainCharacter); |
| 739 | const speakCues = cues.filter((c) => c.type === 'speak'); |
| 740 | |
| 741 | // Bounded render-ahead: kick off prepareSpeech for the k-th speak cue at |
| 742 | // most once, and stash the promise on the cue itself. Calling this |
| 743 | // repeatedly (e.g. re-priming a window that's already full) is a no-op. |
| 744 | function ensureRender(k) { |
| 745 | if (k < 0 || k >= speakCues.length) return; |
| 746 | const c = speakCues[k]; |
| 747 | if (!c._renderP) { |
| 748 | c._renderP = prepareSpeech(c.speech); |
| 749 | c._renderP.catch(() => {}); // avoid unhandled-rejection noise; real errors surface at the await below |
| 750 | } |
| 751 | } |
| 752 | |
| 753 | for (let k = 0; k < RENDER_WINDOW; k++) ensureRender(k); |
| 754 | let speakIndex = 0; |
| 755 | |
| 756 | if (speakCues.length > 0) { |
| 757 | try { |
| 758 | await speakCues[0]._renderP; |
| 759 | } catch (e) { |
| 760 | // ignore here; the playback loop's per-line error handling will surface |
| 761 | // this once it reaches line 1 |
| 762 | } |
| 763 | if (token !== scriptToken) return; // replaced while pre-rendering; don't start |
| 764 | } |
| 765 | |
| 766 | broadcast({ type: 'script-start', lines: cues.map((c) => c.source) }); |
| 767 | |
| 768 | for (let i = 0; i < cues.length; i++) { |
| 769 | if (token !== scriptToken) break; // stopped or replaced |
| 770 | const cue = cues[i]; |
| 771 | broadcast({ type: 'script-line', index: i }); |
| 772 | |
| 773 | if (cue.type === 'setting' || cue.type === 'wait') { |
| 774 | if (cue.type === 'wait') await sleep(cue.seconds * 1000); |
| 775 | continue; |
| 776 | } |
| 777 | |
| 778 | if (cue.type === 'speak') { |
| 779 | const k = speakIndex++; |
| 780 | ensureRender(k); // in case it wasn't primed (e.g. window > remaining lines) |
| 781 | let prepared; |
| 782 | try { |
| 783 | prepared = await cue._renderP; |
| 784 | } catch (e) { |
| 785 | broadcast({ type: 'error', message: e.message }); |
| 786 | ensureRender(k + RENDER_WINDOW); |
| 787 | continue; |
| 788 | } |
| 789 | if (token !== scriptToken) break; |
| 790 | if (cue.concurrent) broadcast(cue.concurrent); |
| 791 | const speakCue = { type: 'speak', ...prepared }; |
| 792 | if (cue.frame) speakCue.frame = cue.frame; |
| 793 | if (cue.actor) speakCue.actor = cue.actor; |
| 794 | broadcast(speakCue); |
| 795 | await sleep(prepared.duration * 1000 + 300); |
| 796 | ensureRender(k + RENDER_WINDOW); |
| 797 | continue; |
| 798 | } |
| 799 | |
| 800 | broadcast(cue); |
| 801 | if (cue.type === 'action') await sleep((ACTION_SECONDS[cue.name] || 1) * 1000); |
| 802 | else if (cue.type === 'walk') await sleep(1200); |
| 803 | else if (cue.type === 'transition') await sleep(cue.ms + 100); |
| 804 | else await sleep(400); |
| 805 | } |
| 806 | |
| 807 | if (token === scriptToken) broadcast({ type: 'script-end' }); |
| 808 | } |
| 809 | |
| 810 | // ----------------------------------------------------------------- voices |
| 811 | |
| 812 | async function listVoices(engine) { |
| 813 | if (engine === 'espeak') { |
| 814 | const out = await run('espeak', ['--voices']); |
| 815 | return out.split('\n').slice(1).map((l) => l.trim().split(/\s+/)[3]).filter(Boolean); |
| 816 | } |
| 817 | const out = await run('say', ['-v', '?']); |
| 818 | return out.split('\n') |
| 819 | .map((l) => l.match(/^(.+?)\s{2,}([a-z]{2}[_-]\w+)/)) |
| 820 | .filter((m) => m && m[2].startsWith('en')) |
| 821 | .map((m) => m[1].trim()); |
| 822 | } |
| 823 | |
| 824 | // ------------------------------------------------------ conversation mode |
| 825 | // |
| 826 | // CONVERSATION MODE: push-to-talk speech -> OpenAI transcription -> an |
| 827 | // OpenAI chat completion replying in character -> spoken back through the |
| 828 | // existing prepareSpeech/broadcast pipeline. Key loading mirrors the |
| 829 | // integrations/dbbasic/puppet-stage director.py convention (OPENAI_API_KEY |
| 830 | // env, else an OPENAI_KEY_FILE .env-style file) so both surfaces agree on |
| 831 | // how a deployment supplies the key. The key lives only in this |
| 832 | // module-level variable — never logged, never echoed back in a response. |
| 833 | |
| 834 | let openaiKey = ''; |
| 835 | (function loadOpenAIKey() { |
| 836 | const env = (process.env.OPENAI_API_KEY || '').trim(); |
| 837 | if (env) { openaiKey = env; return; } |
| 838 | const file = (process.env.OPENAI_KEY_FILE || '').trim(); |
| 839 | if (!file) return; |
| 840 | try { |
| 841 | const text = fs.readFileSync(file, 'utf8'); |
| 842 | for (const line of text.split('\n')) { |
| 843 | const l = line.trim(); |
| 844 | if (l.startsWith('OPENAI_API_KEY=')) { |
| 845 | openaiKey = l.slice('OPENAI_API_KEY='.length).trim().replace(/^["']|["']$/g, ''); |
| 846 | break; |
| 847 | } |
| 848 | } |
| 849 | } catch { /* missing/unreadable key file; converse endpoints report not-configured */ } |
| 850 | })(); |
| 851 | |
| 852 | const NO_KEY_ERROR = 'no OpenAI key configured (set OPENAI_API_KEY or OPENAI_KEY_FILE)'; |
| 853 | const TRANSCRIBE_MODEL = process.env.TRANSCRIBE_MODEL || 'gpt-4o-mini-transcribe'; |
| 854 | const CONVERSE_MODEL = process.env.CONVERSE_MODEL || 'gpt-5.4-mini'; |
| 855 | |
| 856 | // Content-Type -> filename extension, so the multipart upload's filename |
| 857 | // matches the container the browser actually recorded (webm/ogg/mp4/wav). |
| 858 | function audioExtFromContentType(ct) { |
| 859 | ct = (ct || '').toLowerCase(); |
| 860 | if (ct.includes('ogg')) return 'ogg'; |
| 861 | if (ct.includes('mp4') || ct.includes('m4a') || ct.includes('aac')) return 'mp4'; |
| 862 | if (ct.includes('wav')) return 'wav'; |
| 863 | return 'webm'; |
| 864 | } |
| 865 | |
| 866 | // Zero-dependency multipart/form-data POST to OpenAI's transcription |
| 867 | // endpoint: two fields (model, file) built by hand via Buffer.concat. |
| 868 | async function openaiTranscribe(bytes, contentType) { |
| 869 | if (!openaiKey) throw new Error(NO_KEY_ERROR); |
| 870 | const boundary = `----puppetstage${crypto.randomBytes(16).toString('hex')}`; |
| 871 | const ext = audioExtFromContentType(contentType); |
| 872 | const head = Buffer.from( |
| 873 | `--${boundary}\r\nContent-Disposition: form-data; name="model"\r\n\r\n${TRANSCRIBE_MODEL}\r\n` + |
| 874 | `--${boundary}\r\nContent-Disposition: form-data; name="file"; filename="audio.${ext}"\r\n` + |
| 875 | `Content-Type: ${contentType || 'application/octet-stream'}\r\n\r\n`); |
| 876 | const tail = Buffer.from(`\r\n--${boundary}--\r\n`); |
| 877 | const body = Buffer.concat([head, bytes, tail]); |
| 878 | |
| 879 | const controller = new AbortController(); |
| 880 | const timer = setTimeout(() => controller.abort(), 60000); |
| 881 | let resp, text; |
| 882 | try { |
| 883 | resp = await fetch('https://api.openai.com/v1/audio/transcriptions', { |
| 884 | method: 'POST', |
| 885 | headers: { Authorization: `Bearer ${openaiKey}`, 'Content-Type': `multipart/form-data; boundary=${boundary}` }, |
| 886 | body, |
| 887 | signal: controller.signal, |
| 888 | }); |
| 889 | text = await resp.text(); |
| 890 | } finally { |
| 891 | clearTimeout(timer); |
| 892 | } |
| 893 | if (!resp.ok) throw new Error(`OpenAI ${resp.status}: ${text.slice(0, 300)}`); |
| 894 | let parsed; |
| 895 | try { parsed = JSON.parse(text); } catch { throw new Error('unexpected OpenAI response (transcribe)'); } |
| 896 | return parsed.text || ''; |
| 897 | } |
| 898 | |
| 899 | // Zero-dependency chat completion call for the in-character reply. |
| 900 | async function openaiChat(messages) { |
| 901 | if (!openaiKey) throw new Error(NO_KEY_ERROR); |
| 902 | const controller = new AbortController(); |
| 903 | const timer = setTimeout(() => controller.abort(), 60000); |
| 904 | let resp, text; |
| 905 | try { |
| 906 | resp = await fetch('https://api.openai.com/v1/chat/completions', { |
| 907 | method: 'POST', |
| 908 | headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${openaiKey}` }, |
| 909 | body: JSON.stringify({ model: CONVERSE_MODEL, messages, temperature: 0.8 }), |
| 910 | signal: controller.signal, |
| 911 | }); |
| 912 | text = await resp.text(); |
| 913 | } finally { |
| 914 | clearTimeout(timer); |
| 915 | } |
| 916 | if (!resp.ok) throw new Error(`OpenAI ${resp.status}: ${text.slice(0, 300)}`); |
| 917 | let parsed; |
| 918 | try { parsed = JSON.parse(text); } catch { throw new Error('unexpected OpenAI response (chat)'); } |
| 919 | const content = parsed.choices && parsed.choices[0] && parsed.choices[0].message && parsed.choices[0].message.content; |
| 920 | if (!content) throw new Error('unexpected OpenAI response shape (chat)'); |
| 921 | return content.trim(); |
| 922 | } |
| 923 | |
| 924 | // Per-session rolling chat history: session id -> {messages:[{role,content}], |
| 925 | // lastActive}. Capped at the last 20 messages; idle sessions (30 min) are |
| 926 | // pruned opportunistically whenever /api/converse is next called. |
| 927 | const CONVERSE_HISTORY_CAP = 20; |
| 928 | const CONVERSE_IDLE_MS = 30 * 60 * 1000; |
| 929 | const conversations = new Map(); |
| 930 | |
| 931 | function conversationFor(id) { |
| 932 | const now = Date.now(); |
| 933 | for (const [k, v] of conversations) { |
| 934 | if (now - v.lastActive > CONVERSE_IDLE_MS) conversations.delete(k); |
| 935 | } |
| 936 | let s = conversations.get(id); |
| 937 | if (!s) { s = { messages: [], lastActive: now }; conversations.set(id, s); } |
| 938 | s.lastActive = now; |
| 939 | return s; |
| 940 | } |
| 941 | |
| 942 | function pushCapped(list, entry) { |
| 943 | list.push(entry); |
| 944 | if (list.length > CONVERSE_HISTORY_CAP) list.splice(0, list.length - CONVERSE_HISTORY_CAP); |
| 945 | } |
| 946 | |
| 947 | // Builds the system prompt for a character: manifest.persona if the |
| 948 | // manifest declares one (another agent is adding that field in parallel — |
| 949 | // this only ever READS it), else a generic fallback. Also returns the |
| 950 | // character's action names so the reply can be parsed against them. |
| 951 | function buildConverseContext(character) { |
| 952 | let persona = null; |
| 953 | let actionNames = []; |
| 954 | try { |
| 955 | const manifest = JSON.parse(fs.readFileSync(path.join(CHARACTERS, character, 'manifest.json'), 'utf8')); |
| 956 | if (manifest && manifest.persona) persona = manifest.persona; |
| 957 | actionNames = Object.keys((manifest && manifest.actions) || {}); |
| 958 | } catch { /* missing/unreadable manifest; fall back below */ } |
| 959 | const displayName = character ? character.charAt(0).toUpperCase() + character.slice(1) : 'the puppet'; |
| 960 | const base = persona || `You are ${displayName}, a puppet character on a small stage.`; |
| 961 | const actionsList = actionNames.length ? actionNames.join(', ') : 'none'; |
| 962 | const prompt = `${base}\n` + |
| 963 | 'You are a live animated puppet on a real stage, and you have a body: you can ' + |
| 964 | `physically perform these actions right now: ${actionsList}. ` + |
| 965 | 'To perform one, write its name in parentheses inline with your words, e.g. "(nod) Absolutely." ' + |
| 966 | 'Use one when it fits what you are saying — agree with a (nod), refuse with a (shake), ' + |
| 967 | 'delight with a (happy). If someone asks whether you can do something on that list, ' + |
| 968 | 'say yes and DO it in the same reply. At most two actions per reply; only names from ' + |
| 969 | 'the list; nothing else in parentheses.\n' + |
| 970 | 'Reply in character, conversationally, 1-3 short sentences (spoken aloud). No emoji, no markdown.'; |
| 971 | return { prompt, actionNames }; |
| 972 | } |
| 973 | |
| 974 | // Extracts (action) tokens that match the character's known action names out |
| 975 | // of a reply, in order, and strips them (and the whitespace they leave |
| 976 | // behind) from the spoken text. A parenthetical that ISN'T a known action |
| 977 | // name is left in place rather than silently eaten — the persona prompt |
| 978 | // asks the model not to use parens for anything else, but a stray one |
| 979 | // shouldn't corrupt the spoken line. |
| 980 | function extractActions(text, actionNames) { |
| 981 | const known = new Set(actionNames); |
| 982 | const actions = []; |
| 983 | const stripped = text.replace(/\(([a-zA-Z][\w-]*)\)/g, (whole, name) => { |
| 984 | if (known.has(name)) { actions.push(name); return ' '; } |
| 985 | return whole; |
| 986 | }); |
| 987 | return { text: stripped.replace(/\s+/g, ' ').trim(), actions }; |
| 988 | } |
| 989 | |
| 990 | // ----------------------------------------------------------------- static |
| 991 | |
| 992 | const MIME = { |
| 993 | '.html': 'text/html', '.js': 'text/javascript', '.css': 'text/css', |
| 994 | '.json': 'application/json', '.svg': 'image/svg+xml', '.wav': 'audio/wav', |
| 995 | '.png': 'image/png', '.jpg': 'image/jpeg', '.jpeg': 'image/jpeg', |
| 996 | '.webp': 'image/webp', '.gif': 'image/gif', |
| 997 | '.mp4': 'video/mp4', '.webm': 'video/webm', |
| 998 | }; |
| 999 | |
| 1000 | function serveFile(res, filePath) { |
| 1001 | fs.readFile(filePath, (err, data) => { |
| 1002 | if (err) { res.writeHead(404); res.end('not found'); return; } |
| 1003 | res.writeHead(200, { 'Content-Type': MIME[path.extname(filePath)] || 'application/octet-stream' }); |
| 1004 | res.end(data); |
| 1005 | }); |
| 1006 | } |
| 1007 | |
| 1008 | // ----------------------------------------------------------------- server |
| 1009 | |
| 1010 | const server = http.createServer(async (req, res) => { |
| 1011 | const url = new URL(req.url, `http://${req.headers.host}`); |
| 1012 | const p = url.pathname; |
| 1013 | |
| 1014 | try { |
| 1015 | if (p === '/api/events') { |
| 1016 | res.writeHead(200, { |
| 1017 | 'Content-Type': 'text/event-stream', |
| 1018 | 'Cache-Control': 'no-cache', |
| 1019 | Connection: 'keep-alive', |
| 1020 | }); |
| 1021 | res.write(': connected\n\n'); |
| 1022 | clients.add(res); |
| 1023 | req.on('close', () => clients.delete(res)); |
| 1024 | return; |
| 1025 | } |
| 1026 | |
| 1027 | if (p === '/api/say' && req.method === 'POST') { |
| 1028 | const body = await readBody(req); |
| 1029 | if (!body.text) return json(res, 400, { error: 'text required' }); |
| 1030 | const prepared = await prepareSpeech(body); |
| 1031 | const cue = { type: 'speak', ...prepared }; |
| 1032 | if (body.frame) cue.frame = body.frame; |
| 1033 | if (body.actor) cue.actor = body.actor; |
| 1034 | broadcast(cue); |
| 1035 | return json(res, 200, prepared); |
| 1036 | } |
| 1037 | |
| 1038 | // external-TTS entry point: POST raw audio bytes (any ffmpeg-readable |
| 1039 | // format), optional ?text= transcript improves lip-sync accuracy |
| 1040 | if (p === '/api/speak-audio' && req.method === 'POST') { |
| 1041 | const bytes = await readRaw(req); |
| 1042 | if (!bytes.length) return json(res, 400, { error: 'audio body required' }); |
| 1043 | const prepared = await prepareExternalAudio(bytes, url.searchParams.get('text') || ''); |
| 1044 | broadcast({ type: 'speak', ...prepared }); |
| 1045 | return json(res, 200, prepared); |
| 1046 | } |
| 1047 | |
| 1048 | // CONVERSATION MODE: readiness probe — the UI hides the Converse panel |
| 1049 | // when this reports not-ready (no key configured). |
| 1050 | if (p === '/api/converse' && req.method === 'GET') { |
| 1051 | return json(res, 200, { ready: !!openaiKey }); |
| 1052 | } |
| 1053 | |
| 1054 | // raw audio bytes body (webm/ogg/mp4/wav — whatever the browser's |
| 1055 | // MediaRecorder produced; Content-Type says which), forwarded to |
| 1056 | // OpenAI's transcription endpoint. |
| 1057 | if (p === '/api/transcribe' && req.method === 'POST') { |
| 1058 | if (!openaiKey) return json(res, 400, { error: NO_KEY_ERROR }); |
| 1059 | const bytes = await readRaw(req); |
| 1060 | if (!bytes.length) return json(res, 400, { error: 'audio body required' }); |
| 1061 | try { |
| 1062 | const text = await openaiTranscribe(bytes, req.headers['content-type']); |
| 1063 | return json(res, 200, { text }); |
| 1064 | } catch (e) { |
| 1065 | return json(res, 502, { error: e.message }); |
| 1066 | } |
| 1067 | } |
| 1068 | |
| 1069 | // {text, character, session, frame?} -> an in-character OpenAI reply, |
| 1070 | // spoken on stage. Broadcasts (in order) a user-said caption cue, an |
| 1071 | // optional concurrent action cue (the reply's first inline action, like |
| 1072 | // a screenplay's (action) prefix), the speak cue itself, then any |
| 1073 | // further inline actions as trailing action cues. |
| 1074 | if (p === '/api/converse' && req.method === 'POST') { |
| 1075 | if (!openaiKey) return json(res, 400, { error: NO_KEY_ERROR }); |
| 1076 | const body = await readBody(req); |
| 1077 | const userText = (body.text || '').trim(); |
| 1078 | const character = body.character; |
| 1079 | if (!userText) return json(res, 400, { error: 'text required' }); |
| 1080 | if (!character) return json(res, 400, { error: 'character required' }); |
| 1081 | |
| 1082 | const session = conversationFor(body.session || 'default'); |
| 1083 | const { prompt, actionNames } = buildConverseContext(character); |
| 1084 | pushCapped(session.messages, { role: 'user', content: userText }); |
| 1085 | |
| 1086 | let reply; |
| 1087 | try { |
| 1088 | reply = await openaiChat([{ role: 'system', content: prompt }, ...session.messages]); |
| 1089 | } catch (e) { |
| 1090 | session.messages.pop(); // don't keep a user turn that never got a reply |
| 1091 | return json(res, 502, { error: e.message }); |
| 1092 | } |
| 1093 | pushCapped(session.messages, { role: 'assistant', content: reply }); |
| 1094 | |
| 1095 | const { text: spokenText, actions } = extractActions(reply, actionNames); |
| 1096 | |
| 1097 | broadcast({ type: 'user-said', text: userText }); |
| 1098 | if (actions[0]) { |
| 1099 | const actionCue = { type: 'action', name: actions[0] }; |
| 1100 | if (body.frame) actionCue.frame = body.frame; |
| 1101 | broadcast(actionCue); |
| 1102 | } |
| 1103 | |
| 1104 | let prepared; |
| 1105 | try { |
| 1106 | const voice = characterVoice(character) || {}; |
| 1107 | prepared = await prepareSpeech({ text: spokenText || reply, engine: voice.engine, voice: voice.voice, rate: voice.rate }); |
| 1108 | } catch (e) { |
| 1109 | return json(res, 502, { error: e.message }); |
| 1110 | } |
| 1111 | const speakCue = { type: 'speak', ...prepared }; |
| 1112 | if (body.frame) speakCue.frame = body.frame; |
| 1113 | broadcast(speakCue); |
| 1114 | |
| 1115 | for (const name of actions.slice(1)) { |
| 1116 | const cue = { type: 'action', name }; |
| 1117 | if (body.frame) cue.frame = body.frame; |
| 1118 | broadcast(cue); |
| 1119 | } |
| 1120 | |
| 1121 | return json(res, 200, { reply: spokenText, actions, performed: true }); |
| 1122 | } |
| 1123 | |
| 1124 | if (p === '/api/cue' && req.method === 'POST') { |
| 1125 | const body = await readBody(req); |
| 1126 | if (!body.type) return json(res, 400, { error: 'type required' }); |
| 1127 | broadcast(body); |
| 1128 | return json(res, 200, { ok: true }); |
| 1129 | } |
| 1130 | |
| 1131 | if (p === '/api/script' && req.method === 'POST') { |
| 1132 | const body = await readBody(req); |
| 1133 | if (!body.script) return json(res, 400, { error: 'script required' }); |
| 1134 | runScript(body.script, body.mainCharacter); // runs in background |
| 1135 | return json(res, 200, { ok: true }); |
| 1136 | } |
| 1137 | |
| 1138 | if (p === '/api/script/stop' && req.method === 'POST') { |
| 1139 | scriptToken++; |
| 1140 | broadcast({ type: 'script-end' }); |
| 1141 | return json(res, 200, { ok: true }); |
| 1142 | } |
| 1143 | |
| 1144 | if (p === '/api/voices') { |
| 1145 | return json(res, 200, { voices: await listVoices(url.searchParams.get('engine')) }); |
| 1146 | } |
| 1147 | |
| 1148 | if (p === '/api/characters') { |
| 1149 | const names = fs.readdirSync(CHARACTERS, { withFileTypes: true }) |
| 1150 | .filter((d) => d.isDirectory()).map((d) => d.name); |
| 1151 | return json(res, 200, { characters: names }); |
| 1152 | } |
| 1153 | |
| 1154 | // GET /api/examples → basenames (no .txt) of examples/*.txt, sorted. |
| 1155 | if (p === '/api/examples') { |
| 1156 | let names = []; |
| 1157 | try { |
| 1158 | names = fs.readdirSync(EXAMPLES) |
| 1159 | .filter((f) => f.endsWith('.txt')) |
| 1160 | .map((f) => f.slice(0, -'.txt'.length)) |
| 1161 | .sort(); |
| 1162 | } catch { /* no examples dir yet */ } |
| 1163 | return json(res, 200, { examples: names }); |
| 1164 | } |
| 1165 | |
| 1166 | // GET /api/examples/<name> → {name, script} of examples/<name>.txt. |
| 1167 | if (p.startsWith('/api/examples/')) { |
| 1168 | const rel = path.normalize(p.slice('/api/examples/'.length)); |
| 1169 | if (rel.startsWith('..')) { res.writeHead(403); return res.end(); } |
| 1170 | const file = path.join(EXAMPLES, rel + '.txt'); |
| 1171 | if (!fs.existsSync(file)) return json(res, 404, { error: 'example not found' }); |
| 1172 | return json(res, 200, { name: rel, script: fs.readFileSync(file, 'utf8') }); |
| 1173 | } |
| 1174 | |
| 1175 | // GET /api/assets → {backgrounds, props, overlays, music, sfx}: ids |
| 1176 | // (filenames minus extension) available in assets/<kind>/, deduped |
| 1177 | // across the relevant extension set (svg/raster for most kinds, .json |
| 1178 | // for the audio kinds). |
| 1179 | if (p === '/api/assets') { |
| 1180 | const listKind = (kind) => { |
| 1181 | let files = []; |
| 1182 | try { files = fs.readdirSync(path.join(ASSETS, kind)); } catch { return []; } |
| 1183 | const exts = AUDIO_KINDS.includes(kind) ? AUDIO_EXTS : ASSET_EXTS; |
| 1184 | const ids = new Set(); |
| 1185 | for (const f of files) { |
| 1186 | const ext = path.extname(f); |
| 1187 | if (exts.includes(ext)) ids.add(f.slice(0, -ext.length)); |
| 1188 | } |
| 1189 | return [...ids].sort(); |
| 1190 | }; |
| 1191 | return json(res, 200, { |
| 1192 | backgrounds: listKind('backgrounds'), |
| 1193 | props: listKind('props'), |
| 1194 | overlays: listKind('overlays'), |
| 1195 | music: listKind('music'), |
| 1196 | sfx: listKind('sfx'), |
| 1197 | }); |
| 1198 | } |
| 1199 | |
| 1200 | // 1x1 gif served after a delay — lets headless screenshots wait for async boot |
| 1201 | if (p === '/slowpx') { |
| 1202 | lastSlowpx = Date.now(); // marks navigation: the page requests this immediately on load |
| 1203 | const gif = Buffer.from('R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7', 'base64'); |
| 1204 | setTimeout(() => { res.writeHead(200, { 'Content-Type': 'image/gif' }); res.end(gif); }, |
| 1205 | parseInt(url.searchParams.get('d'), 10) || 2500); |
| 1206 | return; |
| 1207 | } |
| 1208 | |
| 1209 | // GET /api/last-shot → {t}: Date.now() (server clock) at the start of the |
| 1210 | // most recent /slowpx request, i.e. the navigation timestamp of the most |
| 1211 | // recent headless screenshot page load. Used by the screenshot harness to |
| 1212 | // anchor action cues to actual page navigation instead of a fixed sleep |
| 1213 | // after spawning the browser process (browser startup jitter otherwise |
| 1214 | // made cue timing nondeterministic). 0 if no /slowpx request has landed yet. |
| 1215 | if (p === '/api/last-shot') { |
| 1216 | return json(res, 200, { t: lastSlowpx || 0 }); |
| 1217 | } |
| 1218 | |
| 1219 | // GET /api/asset/<kind>/<id> → the resolved file (first existing |
| 1220 | // extension among svg/png/jpg/jpeg/webp/gif), so the client can learn |
| 1221 | // which format exists without guessing. |
| 1222 | if (p.startsWith('/api/asset/')) { |
| 1223 | const [kind, id] = p.slice('/api/asset/'.length).split('/'); |
| 1224 | const file = kind && id ? resolveAsset(kind, id) : null; |
| 1225 | if (!file) return json(res, 404, { error: 'asset not found' }); |
| 1226 | return serveFile(res, file); |
| 1227 | } |
| 1228 | |
| 1229 | // static routes |
| 1230 | if (p.startsWith('/audio/')) return serveFile(res, path.join(CACHE, path.basename(p))); |
| 1231 | if (p.startsWith('/characters/')) { |
| 1232 | const rel = path.normalize(p.slice('/characters/'.length)); |
| 1233 | if (rel.startsWith('..')) { res.writeHead(403); return res.end(); } |
| 1234 | return serveFile(res, path.join(CHARACTERS, rel)); |
| 1235 | } |
| 1236 | if (p.startsWith('/assets/')) { |
| 1237 | const rel = path.normalize(p.slice('/assets/'.length)); |
| 1238 | if (rel.startsWith('..')) { res.writeHead(403); return res.end(); } |
| 1239 | return serveFile(res, path.join(ASSETS, rel)); |
| 1240 | } |
| 1241 | const file = p === '/' ? 'index.html' : path.normalize(p).replace(/^\/+/, ''); |
| 1242 | if (file.startsWith('..')) { res.writeHead(403); return res.end(); } |
| 1243 | return serveFile(res, path.join(PUBLIC, file)); |
| 1244 | } catch (e) { |
| 1245 | return json(res, 500, { error: e.message }); |
| 1246 | } |
| 1247 | }); |
| 1248 | |
| 1249 | server.listen(PORT, () => { |
| 1250 | console.log(`Puppet stage: http://localhost:${PORT}`); |
| 1251 | }); |