tuna-agent 0.1.138 → 0.1.139

package/dist/daemon/analyze-video-handler.d.ts

@@ -11,6 +11,7 @@ export interface AnalyzeVideoResult {
     duration_sec: number;
     language: string;
     transcript: string;
+    summary: string;
     video_style: string;
     master_cast_prompt: string;
     characters: Array<{

package/dist/daemon/analyze-video-handler.js

@@ -166,38 +166,47 @@ Voiceover during this scene: "${voiceoverText || '(none)'}"` },
     const data = await res.json();
     return data.choices?.[0]?.message?.content?.trim() || '';
 }
-// Phase 1 (borrowed from AI_Video_Clone): extract the recurring character
-// cast ONCE from frames sampled across the whole video + the transcript.
-// Returns a master-cast prompt block in the exact [AESTHETIC & STYLE] /
-// [CHARACTER CAST LIST] format that channel-manager's ScriptImporter parses,
-// plus a structured characters[] list. Doing this upfront (a) populates
-// idea.master_cast_prompt so FlowKit has a reference sheet to generate, and
-// (b) gives every per-scene describe call a consistent naming vocabulary so
-// scene 1 and scene 50 refer to "THE BISHOP" instead of "a man in a suit".
-async function visionExtractMasterCast(frames, transcript, videoStyle) {
-    const empty = { master_cast_prompt: '', characters: [] };
+// Phase 1 (the strong part of AI_Video_Clone, ported): ONE gpt-4o call over
+// frames sampled across the whole video + transcript that returns, together:
+//   - video_summary: a cinematic paragraph of the whole story (drives the
+//     downstream script-generation prompt — the tool's biggest edge)
+//   - video_style:   a rich 3-4 sentence aesthetic analysis (medium, palette,
+//     lighting, camera language) — replaces the old terse 1-2 sentence
+//     visionExtractStyle gpt-4o-mini call entirely
+//   - characters[]:  the recurring cast for the [CHARACTER CAST LIST] block
+// Folding all three into one call is cheaper than the previous two calls
+// (style + cast) AND uses gpt-4o for style (was gpt-4o-mini). The master-cast
+// prompt is assembled here in the exact format ScriptImporter parses.
+async function visionExtractPhase1(frames, transcript) {
+    const empty = {
+        video_summary: '',
+        video_style: '',
+        master_cast_prompt: '',
+        characters: [],
+    };
     if (!OPENAI_KEY || frames.length === 0)
         return empty;
     try {
         const content = [
             {
                 type: 'text',
-                text: `Act as a Master Film Director. These frames are sampled across an entire video. Identify EVERY recurring character/subject (people, anthropomorphic objects, animals, mascots).
+                text: `Act as a Master Film Director. These frames are sampled across an ENTIRE video, in order. Use them + the transcript to analyse the whole piece.
 
-Transcript context (may name characters): "${(transcript || '').slice(0, 1500)}"
+Transcript context: "${(transcript || '').slice(0, 4000)}"
 
 Return ONLY a JSON object, no markdown fences:
 {
+  "video_summary": "One detailed cinematic paragraph (5-8 sentences, English) telling the WHOLE story start to finish: setup, key beats, climax, resolution. This is the narrative spine — be specific about what happens.",
+  "video_style": "3-4 sentences (English): artistic medium (2D/3D/live-action/CGI), color palette, lighting, camera language, overall aesthetic vibe. Cinematic, specific.",
   "characters": [
     { "name": "SHORT_UPPERCASE_LABEL", "description": "one-line English visual description: age/build, face, hair, outfit, colors, distinguishing features" }
   ]
 }
 
 Rules:
-- name: a stable short uppercase label you will reuse for this subject (e.g. "THE BISHOP", "RED CAR", "NARRATOR DOG"). Max 4 words.
-- Only RECURRING subjects worth a reference sheet. Skip one-off background extras.
-- description: ENGLISH only, factual, no camera/action words.
-- Max 6 characters.`,
+- characters.name: a stable short uppercase label reused for this subject (e.g. "THE BISHOP", "U-94 SUBMARINE"). Max 4 words.
+- Only RECURRING subjects worth a reference sheet. Skip one-off extras. Max 6.
+- characters.description: ENGLISH only, factual, no camera/action words.`,
             },
         ];
         for (const b64 of frames) {
@@ -206,16 +215,16 @@ Rules:
         const res = await fetch('https://api.openai.com/v1/chat/completions', {
             method: 'POST',
             headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${OPENAI_KEY}` },
-            body: JSON.stringify({ model: 'gpt-4o', max_tokens: 1200, messages: [{ role: 'user', content }] }),
+            body: JSON.stringify({ model: 'gpt-4o', max_tokens: 1600, messages: [{ role: 'user', content }] }),
         });
         if (!res.ok)
             return empty;
         const data = await res.json();
-        const raw = (data.choices?.[0]?.message?.content || '').trim();
+        const rawTxt = (data.choices?.[0]?.message?.content || '').trim();
         let parsed = {};
         try {
-            const m = raw.match(/\{[\s\S]*\}/);
-            parsed = JSON.parse(m ? m[0] : raw);
+            const m = rawTxt.match(/\{[\s\S]*\}/);
+            parsed = JSON.parse(m ? m[0] : rawTxt);
         }
         catch {
             return empty;
@@ -224,43 +233,22 @@ Rules:
             .filter(c => c && c.name && c.description)
             .slice(0, 6)
             .map(c => ({ name: String(c.name).trim(), description: String(c.description).trim() }));
-        if (characters.length === 0)
-            return empty;
-        // Assemble the verbatim-style master cast block ScriptImporter expects.
-        const styleLine = (videoStyle || '').trim() || 'Keep the original video’s visual style, color grading, and lighting.';
-        const castList = characters.map(c => `- ${c.name}: ${c.description}`).join('\n');
-        const master_cast_prompt = `[AESTHETIC & STYLE]\n${styleLine}\n` +
-            `[COMPOSITION & LAYOUT]\nCharacter Reference Sheet. Full-body side-by-side.\n` +
-            `[CHARACTER CAST LIST]\n${castList}\n` +
-            `[TECHNICAL SPECIFICATIONS]\nHigh detail, 8k resolution, consistent facial structures across all frames.`;
-        return { master_cast_prompt, characters };
-    }
-    catch {
-        return empty;
-    }
-}
-async function visionExtractStyle(frames) {
-    if (!OPENAI_KEY || frames.length === 0)
-        return '';
-    try {
-        const content = [
-            { type: 'text', text: 'Analyze these frames from a video and extract a concise visual style description (1-2 sentences). Focus on: animation style (cartoon, realistic, anime, etc.), color palette, lighting, character design approach (anthropomorphized objects, real people, etc.), and overall aesthetic.\n\nReturn ONLY the style description, nothing else.' },
-        ];
-        for (const b64 of frames) {
-            content.push({ type: 'image_url', image_url: { url: `data:image/jpeg;base64,${b64}` } });
+        const video_summary = (parsed.video_summary || '').trim();
+        const video_style = (parsed.video_style || '').trim();
+        let master_cast_prompt = '';
+        if (characters.length > 0) {
+            const styleLine = video_style || 'Keep the original video’s visual style, color grading, and lighting.';
+            const castList = characters.map(c => `- ${c.name}: ${c.description}`).join('\n');
+            master_cast_prompt =
+                `[AESTHETIC & STYLE]\n${styleLine}\n` +
+                    `[COMPOSITION & LAYOUT]\nCharacter Reference Sheet. Full-body side-by-side.\n` +
+                    `[CHARACTER CAST LIST]\n${castList}\n` +
+                    `[TECHNICAL SPECIFICATIONS]\nHigh detail, 8k resolution, consistent facial structures across all frames.`;
         }
-        const res = await fetch('https://api.openai.com/v1/chat/completions', {
-            method: 'POST',
-            headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${OPENAI_KEY}` },
-            body: JSON.stringify({ model: 'gpt-4o-mini', max_tokens: 200, messages: [{ role: 'user', content }] }),
-        });
-        if (!res.ok)
-            return '';
-        const data = await res.json();
-        return (data.choices?.[0]?.message?.content || '').trim().replace(/\*\*/g, '');
+        return { video_summary, video_style, master_cast_prompt, characters };
     }
     catch {
-        return '';
+        return empty;
     }
 }
 export async function analyzeVideo(url, onProgress) {
@@ -446,25 +434,22 @@ export async function analyzeVideo(url, onProgress) {
                 frameBuffers.push({ idx: i, frames: buffers, thumb: thumb || buffers[0], slot });
             }
         }
-        // Step 2: Style + Master Cast FIRST (Phase 1), so per-scene describe can
-        // reuse consistent character labels (the AI_Video_Clone lesson).
-        progress('Đang phân tích video style...');
-        const styleSamples = frameBuffers.slice(0, 3).map(f => f.thumb.toString('base64'));
-        const video_style = await visionExtractStyle(styleSamples);
-        console.log('[analyze_video] Video style:', video_style.substring(0, 100));
-        progress('Đang trích xuất dàn nhân vật (Master Cast)...');
-        // Sample up to 12 frames evenly across the whole video for cast detection.
-        const castSampleCount = Math.min(12, frameBuffers.length);
-        const castStep = Math.max(1, Math.floor(frameBuffers.length / castSampleCount));
-        const castSamples = frameBuffers
-            .filter((_, i) => i % castStep === 0)
-            .slice(0, castSampleCount)
+        // Step 2: Phase 1 — ONE gpt-4o call returning summary + rich style +
+        // master cast + characters. Runs before per-scene describe so the cast
+        // context keeps naming consistent across the whole timeline.
+        progress('Đang phân tích tổng thể (summary + style + master cast)...');
+        // Sample up to 12 frames evenly across the whole video.
+        const p1SampleCount = Math.min(12, frameBuffers.length);
+        const p1Step = Math.max(1, Math.floor(frameBuffers.length / p1SampleCount));
+        const p1Samples = frameBuffers
+            .filter((_, i) => i % p1Step === 0)
+            .slice(0, p1SampleCount)
             .map(f => f.thumb.toString('base64'));
-        const { master_cast_prompt, characters } = await visionExtractMasterCast(castSamples, transcript.text || '', video_style);
+        const { video_summary, video_style, master_cast_prompt, characters } = await visionExtractPhase1(p1Samples, transcript.text || '');
         const castContext = characters.length
             ? characters.map(c => `- ${c.name}: ${c.description}`).join('\n')
             : '';
-        console.log('[analyze_video] Master cast:', characters.map(c => c.name).join(', ') || '(none)');
+        console.log('[analyze_video] Phase1 — style:', video_style.slice(0, 80), '| cast:', characters.map(c => c.name).join(', ') || '(none)', '| summary:', video_summary.length, 'chars');
         // Step 3: Vision describe all frames in parallel (batch of 5), passing the
         // cast context so naming stays consistent across the whole timeline.
         progress(`Đang phân tích ${frameBuffers.length} scenes song song...`);
@@ -499,6 +484,7 @@ export async function analyzeVideo(url, onProgress) {
             duration_sec: Math.round(durationSec),
             language: transcript.language || 'unknown',
             transcript: transcript.text || '',
+            summary: video_summary,
             video_style,
             master_cast_prompt,
             characters,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tuna-agent",
-  "version": "0.1.138",
+  "version": "0.1.139",
   "description": "Tuna Agent - Run AI coding tasks on your machine",
   "bin": {
     "tuna-agent": "dist/cli/index.js"