@viralspin/mcp 0.2.0 → 0.2.2

package/engine/gemini.mjs

@@ -1,164 +0,0 @@
-// Node port of the Python Gemini scripts (_gem/transcribe/analyze/edit/style/albums) so the
-// engine needs only Node — no Python. All calls are Gemini REST (generateContent), thinking off.
-import fs from "node:fs";
-import path from "node:path";
-import { prompt as storedPrompt, fill } from "./prompts.mjs";
-
-const MODEL = "gemini-2.5-flash";
-
-function getKey() {
-  if (process.env.VITE_GEMINI_API_KEY) return process.env.VITE_GEMINI_API_KEY;
-  if (process.env.GEMINI_API_KEY) return process.env.GEMINI_API_KEY;
-  // local fallback: parse a .env (the repo's, or VIRALSPIN_ENV)
-  for (const env of [process.env.VIRALSPIN_ENV, "E:/AI test/github/viralspin_new/.env"]) {
-    if (env && fs.existsSync(env)) {
-      for (const ln of fs.readFileSync(env, "utf8").split(/\r?\n/)) {
-        if (ln.trim().startsWith("VITE_GEMINI_API_KEY")) return ln.split("=").slice(1).join("=").trim().replace(/^["']|["']$/g, "");
-      }
-    }
-  }
-  throw new Error("No Gemini API key (set GEMINI_API_KEY or VITE_GEMINI_API_KEY).");
-}
-const KEY = getKey();
-const URL = `https://generativelanguage.googleapis.com/v1beta/models/${MODEL}:generateContent?key=${KEY}`;
-
-const mimeOf = (p) => {
-  const e = path.extname(p).toLowerCase();
-  return e === ".png" ? "image/png" : e === ".wav" ? "audio/wav" : e === ".mp3" ? "audio/mpeg" : "image/jpeg";
-};
-export const inlinePart = (p) => ({ inline_data: { mime_type: mimeOf(p), data: fs.readFileSync(p).toString("base64") } });
-
-export async function call(parts, maxTokens = 600) {
-  const body = {
-    contents: [{ parts }],
-    generationConfig: { temperature: 0.2, maxOutputTokens: maxTokens, thinkingConfig: { thinkingBudget: 0 }, responseMimeType: "application/json" },
-  };
-  const r = await fetch(URL, { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify(body) });
-  const resp = await r.json();
-  const cand = resp?.candidates?.[0];
-  const txt = (cand?.content?.parts || []).map((p) => p.text || "").join("");
-  if (!txt.trim()) throw new Error(`empty Gemini response (finish=${cand?.finishReason} status=${r.status})`);
-  const m = txt.match(/\{[\s\S]*\}|\[[\s\S]*\]/);
-  return JSON.parse(m ? m[0] : txt);
-}
-
-export async function transcribe(wavPath) {
-  const d = await call([inlinePart(wavPath), { text: storedPrompt("local_assembly.transcribe") }], 1200);
-  return d.segments || (Array.isArray(d) ? d : []);
-}
-
-export async function analyzeFrame(jpgPath, kind) {
-  const d = await call([inlinePart(jpgPath), { text: storedPrompt("local_assembly.frame_analysis") }], 300);
-  const sx = d.secondary_x, sy = d.secondary_y;
-  return {
-    kind,
-    rotate: parseInt(d.rotate || 0, 10) || 0,
-    focal: [+(+d.focal_x || 0.5).toFixed(4), +(+d.focal_y || 0.45).toFixed(4)],
-    secondary: sx != null && sy != null ? [+(+sx).toFixed(4), +(+sy).toFixed(4)] : null,
-    subject: d.subject || "scenery",
-    desc: d.desc || "",
-    keywords: d.keywords || [],
-  };
-}
-
-export async function rankAlbums(transcriptText, folders) {
-  const prompt = fill(storedPrompt("local_assembly.album_ranking"), {
-    transcript_text: transcriptText,
-    albums_json: JSON.stringify(folders),
-  });
-  try {
-    const d = await call([{ text: prompt }], 700);
-    const w = {}; for (const [k, v] of Object.entries(d.weights || {})) w[k] = +v;
-    return { weights: w, note: d.note || "" };
-  } catch {
-    return { weights: Object.fromEntries(folders.map((f) => [f.name, 0.5])), note: "(fallback)" };
-  }
-}
-
-const VALID_CAPTION_STYLES = ["default", "hormozi", "glow_neon", "minimalist", "whiteboard"];
-
-export async function inferStyle(framePaths, meta) {
-  const prompt = fill(storedPrompt("local_assembly.style_inference"), {
-    frame_count: framePaths.length,
-    format: meta.fmt,
-    duration: meta.dur,
-    cuts_per_sec: meta.cps,
-  });
-  const d = await call([...framePaths.map(inlinePart), { text: prompt }], 400);
-  return {
-    format: meta.fmt,
-    captions: d.captions !== false,
-    hook: d.hook !== false,
-    pops: d.pops !== false,
-    pace: d.pace || "fast",
-    coverage: +(d.coverage ?? 0.5) || 0.5,
-    caption_style: VALID_CAPTION_STYLES.includes(d.caption_style) ? d.caption_style : null,
-    note: d.note || "",
-  };
-}
-
-// Infer caption style from transcript tone when no style_ref is available.
-export async function inferCaptionStyle(transcriptText) {
-  const prompt = `You are a social video editor. Given this A-roll transcript, pick the best caption style for a short-form social video.
-
-Styles:
-- "hormozi": high-energy, motivational, business/hustle content, talking-head, bold claims
-- "glow_neon": dark aesthetic, music, nightlife, gaming, cyberpunk, late-night vibes
-- "minimalist": calm, educational, soft aesthetic, wellness, thoughtful narration
-- "whiteboard": explainer, how-to, tutorial, step-by-step, educational breakdown
-- "default": neutral, general purpose, anything else
-
-Transcript: ${transcriptText.slice(0, 2000)}
-
-Return JSON: {"caption_style": "<one of the five options>", "reason": "<one sentence>"}`;
-  try {
-    const d = await call([{ text: prompt }], 200);
-    return VALID_CAPTION_STYLES.includes(d.caption_style) ? d.caption_style : "default";
-  } catch {
-    return "default";
-  }
-}
-
-export async function getDirectorDecisions(transcriptSegments) {
-  if (!transcriptSegments || transcriptSegments.length === 0) return [];
-  const transcriptText = transcriptSegments.map((s) => s.text || "").join(" ");
-  const prompt = `Analyze this A-roll transcript. Extract entities that a viewer would expect to SEE visualized on screen — named people, specific products, UI tools, app interfaces, or branded concepts.
-
-For each entity return:
-- name: entity name
-- search_query: a 5-10 word Twelve Labs video search query (e.g. "Jensen Huang talking about Blackwell GPU")
-- layout: "pip" if it is a named person (keep the speaker visible alongside them), "interrupt" if it is a product/UI/demo (full-screen cutaway)
-- first_mention_ms: estimated millisecond offset of the first mention (base on word position in transcript)
-
-Transcript: ${transcriptText.slice(0, 3000)}
-
-Return JSON: {"entities": [{"name": "...", "search_query": "...", "layout": "pip", "first_mention_ms": 0}]}`;
-  try {
-    const d = await call([{ text: prompt }], 800);
-    return (d.entities || []).filter((e) => e.name && e.search_query);
-  } catch (err) {
-    return [];
-  }
-}
-
-export async function editTimeline(transcript, analysis, totalMs, userPrompt, style) {
-  const pace = style?.pace || "fast";
-  const PACE = {
-    fast: "Energetic: short cutaways 600-1100 ms and a 3-4 shot montage on numbers/claims. RELEVANCE WINS over coverage — only cut where a clip genuinely fits; otherwise stay on the speaker. Do NOT chase a coverage %.",
-    medium: "Cutaways 1000-1800 ms, only where a clip genuinely fits; otherwise stay on the speaker.",
-    slow: "Few, longer cutaways 1800-3000 ms; relevance first.",
-  }[pace] || "";
-  const incHook = style ? style.hook !== false : true;
-  const incPops = style ? style.pops !== false : true;
-  const candidates = Object.entries(analysis).map(([id, a]) => ({ id, kind: a.kind, subject: a.subject, desc: a.desc, keywords: a.keywords }));
-  const sys = fill(storedPrompt("local_assembly.timeline_edit"), {
-    pacing: PACE,
-    transcript_json: JSON.stringify(transcript),
-    candidates_json: JSON.stringify(candidates),
-    total_ms: totalMs,
-    user_direction: userPrompt || "(none)",
-  });
-  const d = await call([{ text: sys }], 2000);
-  const cuts = (d.cuts || []).filter((c) => analysis[c.broll_id]).sort((a, b) => (a.from_ms || 0) - (b.from_ms || 0));
-  return { cuts, pops: incPops ? (d.pops || []).filter((p) => p.text) : [], hook: incHook ? d.hook || null : null };
-}

package/engine/kinetic.mjs

@@ -1,84 +0,0 @@
-#!/usr/bin/env node
-/**
- * Kinetic text renderer — no A-roll, no B-roll, no transcription.
- *   node kinetic.mjs --text "..." [--effect typing|billboard_flip|isometric]
- *                    [--style terminal|cinematic|minimal] [--format 9x16|16x9|1x1]
- *                    [--font-size <px>] [--duration-sec <n>]
- *                    [--camera-rx <deg>] [--camera-ry <deg>] [--camera-rz <deg>] [--camera-tz <px>]
- *                    [--out <path>]
- * Writes props JSON, calls Remotion render for the KineticText composition, prints RESULT:<path>.
- */
-import { fileURLToPath } from "node:url";
-import fs from "node:fs";
-import path from "node:path";
-import { spawnSync } from "node:child_process";
-
-const ENGINE = path.dirname(fileURLToPath(import.meta.url));
-const PUBLIC = path.join(ENGINE, "public");
-const OUT_DIR = path.join(ENGINE, "out");
-
-fs.mkdirSync(PUBLIC, { recursive: true });
-fs.mkdirSync(OUT_DIR, { recursive: true });
-
-const argv = process.argv.slice(2);
-const one = (flag, def) => { const i = argv.indexOf(flag); return i >= 0 && argv[i + 1] ? argv[i + 1] : def; };
-
-const text = one("--text", "");
-if (!text) { console.error("ERROR: --text is required"); process.exit(1); }
-
-const effect    = one("--effect",      "typing");
-const style     = one("--style",       "terminal");
-const format    = one("--format",      "9x16");
-const fontSizeArg = one("--font-size", null);
-const durationSecArg = one("--duration-sec", null);
-const cameraRx  = parseFloat(one("--camera-rx", null) ?? "NaN");
-const cameraRy  = parseFloat(one("--camera-ry", null) ?? "NaN");
-const cameraRz  = parseFloat(one("--camera-rz", null) ?? "NaN");
-const cameraTz  = parseFloat(one("--camera-tz", null) ?? "NaN");
-const outArg    = one("--out", null);
-
-const fps = 30;
-
-// Auto-size: ~3 frames per char + 45 frame tail (1.5s hold at end)
-const totalFrames = durationSecArg
-  ? Math.round(parseFloat(durationSecArg) * fps)
-  : Math.max(60, Math.ceil(text.length * 3 + 45));
-
-// Default font size by format
-const defaultFontSize = format === "16x9" ? 96 : 72;
-const fontSize = fontSizeArg ? parseInt(fontSizeArg, 10) : defaultFontSize;
-
-// Per-effect camera defaults
-const CAMERA_DEFAULTS = {
-  typing:         { rotateX: 8,  rotateY: 0,   rotateZ: 0,   translateZ: -120 },
-  billboard_flip: { rotateX: 0,  rotateY: 0,   rotateZ: 0,   translateZ: 0    },
-  isometric:      { rotateX: 55, rotateY: 0,   rotateZ: -45, translateZ: -80  },
-};
-const defaults = CAMERA_DEFAULTS[effect] ?? CAMERA_DEFAULTS.typing;
-
-const camera = {
-  rotateX:   isNaN(cameraRx) ? defaults.rotateX   : cameraRx,
-  rotateY:   isNaN(cameraRy) ? defaults.rotateY   : cameraRy,
-  rotateZ:   isNaN(cameraRz) ? defaults.rotateZ   : cameraRz,
-  translateZ: isNaN(cameraTz) ? defaults.translateZ : cameraTz,
-};
-
-const timestamp = Date.now();
-const out = path.resolve(outArg ?? path.join(OUT_DIR, `kinetic_${timestamp}.mp4`));
-fs.mkdirSync(path.dirname(out), { recursive: true });
-
-const props = { text, effect, style, fontSize, format, fps, totalFrames, camera };
-const propsFile = path.join(PUBLIC, "kineticProps.json");
-fs.writeFileSync(propsFile, JSON.stringify(props, null, 2));
-
-process.stderr.write(`[kinetic] text="${text.slice(0, 60)}" effect=${effect} style=${style} format=${format} frames=${totalFrames}\n`);
-
-const entry = path.join(ENGINE, "src", "index.ts");
-const cmd = `npx remotion render "${entry}" KineticText "${out}" "--props=${propsFile}" "--public-dir=${PUBLIC}"`;
-const r = spawnSync(cmd, { cwd: ENGINE, shell: true, stdio: "inherit" });
-if (r.status !== 0) {
-  console.error(`[kinetic] remotion render failed (exit ${r.status})`);
-  process.exit(1);
-}
-
-console.log("RESULT:" + out);

package/engine/prompts.mjs

@@ -1,94 +0,0 @@
-// "Secret sauce" prompts for the local video-assembly engine.
-//
-// These BUILTIN strings are the editable source of truth surfaced in the admin
-// dashboard LLM Prompt Store under the "LOCAL VIDEO ASSEMBLY" section
-// (keys local_assembly.*). At runtime the MCP tool fetches the active rows from
-// the backend and hands them to the engine in the VIRALSPIN_ASSEMBLY_PROMPTS env
-// var (a JSON map key→value_text). Anything missing, blank, or malformed falls
-// back to the BUILTIN below, so an offline run — or a misnamed/deactivated row —
-// still produces a video.
-//
-// Templated prompts use {{double_brace}} tokens (filled by fill() in the engine)
-// so the literal single-brace JSON in their output spec is never disturbed.
-
-export const BUILTIN = {
-  "local_assembly.transcribe":
-    "Transcribe this short spoken clip accurately (a creator talking to camera). " +
-    'Return ONLY JSON: {"segments":[{"start_ms":int,"end_ms":int,"text":str}]}. ' +
-    "Break into natural phrases. Best-guess unclear words; do not invent content. " +
-    "ALSO include significant NON-SPEECH audio as their own segments, with the sound in parentheses — " +
-    "e.g. (coughing), (laughing), (sighing), (music), (silence). These give the editor context.",
-
-  "local_assembly.frame_analysis":
-    "Framing a vertical 9:16 social video from this (often horizontal) source. Return ONLY JSON: " +
-    '{"subject":"person|product|scenery","rotate":0,"focal_x":0-1,"focal_y":0-1,' +
-    '"secondary_x":0-1|null,"secondary_y":0-1|null,"desc":"<=8 words","keywords":["..."]}. ' +
-    "rotate = CLOCKWISE degrees (0/90/180/270) to make it UPRIGHT (0 if already upright); give focal/secondary " +
-    "in coords AFTER that rotation. focal_x/y = center of the MAIN subject (a person's FACE) to keep in a tall 9:16 " +
-    "crop. secondary_x/y = a second notable element to reveal with a pan (sign/text/landmark/product) or null. " +
-    "keywords = 2-4 concrete nouns/themes.",
-
-  "local_assembly.album_ranking":
-    `A short video's narration says:\n"{{transcript_text}}"\n\nWe have these B-roll ALBUMS (folder name + file count):\n` +
-    `{{albums_json}}\n\nRank how relevant each album is as B-roll for this narration, judging mainly by NAME. ` +
-    'Return ONLY JSON: {"weights":{"<album>":0.1-1.0},"note":"<one sentence>"}. Every album gets a weight.',
-
-  "local_assembly.style_inference":
-    `These are {{frame_count}} sampled frames from a short-form social video (aspect {{format}}, {{duration}}s, ` +
-    `~{{cuts_per_sec}} hard scene-cuts/sec). Infer its EDITING STYLE. Return ONLY JSON: ` +
-    '{"captions":bool,"hook":bool,"pops":bool,"pace":"fast|medium|slow","coverage":0-1,"note":"..."}. ' +
-    "captions=bold spoken-word subtitles burned in? hook=persistent headline/title overlay near top? " +
-    "pops=big animated keyword/number callouts? pace=overall energy. coverage=fraction showing B-roll vs talking head.",
-
-  "local_assembly.timeline_edit":
-    `You are a senior short-form editor cutting a HIGH-ENERGY vertical 9:16 Reel (bold captions, hook, pops, quick cuts).\n` +
-    `The A-roll is a person talking (their voice is the soundtrack). B-roll is a MIX of (a) GENERIC ` +
-    `lifestyle/aspirational footage (beach/travel/product) that does NOT literally depict the topic — ` +
-    `energetic punctuation; and (b) the CREATOR'S OWN PERSONAL photos that may genuinely depict the real ` +
-    `moments she describes (sick, recovering, hospital, pregnancy/childbirth, family). Match by content.\n\n` +
-    `TRANSCRIPT (ms):\n{{transcript_json}}\n\nB-ROLL CANDIDATES:\n{{candidates_json}}\n\n` +
-    `TOTAL: {{total_ms}} ms. USER DIRECTION: {{user_direction}}\n\n` +
-    `Produce THREE things:\n` +
-    `1. "hook": a punchy 4-8 word headline for the first ~3.5s + ONE "keyword" within it to highlight.\n` +
-    `2. "pops": 2-4 big keyword/number callouts that punch in ON the beat said. Each {"from_ms":int,"text":"SHORT UPPERCASE"}.\n` +
-    `3. "cuts": B-roll cutaways. RULES:\n` +
-    `   - PACING: {{pacing}}\n` +
-    `   - STRONGLY prefer video-kind assets; alternate dir 0/1. A cutaway must GENUINELY relate to the beat.\n` +
-    `   - TWO kinds of B-roll, match each candidate's CONTENT (desc/keywords) to the beat:\n` +
-    `       • GENERIC lifestyle/beach/travel/product → ONLY on posting/content/success/going-viral/lifestyle/aspiration lines. NEVER a generic clip on a specific physical-state line (sick, recovering, surgery, giving birth, coughing) — looks broken.\n` +
-    `       • The creator's OWN PERSONAL photos that GENUINELY depict the moment (visibly sick/in bed, hospital, pregnant/newborn, recovering, real family) → DO use these ON the matching personal line (a real sick photo on "you're sick, recovering" is GREAT; a beach photo is not).\n` +
-    `   - KEEP THE SPEAKER (no cutaway) on: non-speech beats like "(coughing)", the final line, abstract lines, and ANY beat where NO candidate (generic OR personal) genuinely fits. Staying on the speaker beats an irrelevant clip; don't force a cutaway to fill time.\n` +
-    `   - The clip opens with a COUGH before she speaks — keep the speaker there.\n` +
-    `   - Each broll_id at most twice; no overlaps; within total duration.\n` +
-    `   Each cut: {"broll_id":str,"from_ms":int,"dur_ms":int,"dir":0,"video":bool}.\n\n` +
-    `Return ONLY JSON: {"hook":{"text":str,"keyword":str},"pops":[...],"cuts":[...]}`,
-};
-
-let _overrides = null;
-/** Parse VIRALSPIN_ASSEMBLY_PROMPTS once. Any failure → empty (use builtins). */
-function overrides() {
-  if (_overrides) return _overrides;
-  _overrides = {};
-  const raw = process.env.VIRALSPIN_ASSEMBLY_PROMPTS;
-  if (raw) {
-    try {
-      const m = JSON.parse(raw);
-      for (const [k, v] of Object.entries(m)) {
-        if (typeof v === "string" && v.trim()) _overrides[k] = v;
-      }
-    } catch { /* malformed override blob → keep builtins */ }
-  }
-  return _overrides;
-}
-
-/** Admin-store override for `key` if present and non-empty, else the builtin. */
-export function prompt(key) {
-  return overrides()[key] ?? BUILTIN[key];
-}
-
-/** Replace {{token}} placeholders. Missing keys are left untouched. */
-export function fill(template, vars) {
-  let out = template;
-  for (const [k, v] of Object.entries(vars)) out = out.split(`{{${k}}}`).join(String(v));
-  return out;
-}

package/engine/remotion.config.ts

@@ -1,4 +0,0 @@
-import { Config } from "@remotion/cli/config";
-
-Config.setVideoImageFormat("jpeg");
-Config.setOverwriteOutput(true);