tuna-agent 0.1.151 → 0.1.152

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

@@ -269,7 +269,7 @@ async function visionExtractPhase1(frames, transcript, cost) {
         master_cast_prompt: '',
         characters: [],
     };
-    if (!GEMINI_KEYS.length || frames.length === 0)
+    if (!OPENAI_KEY || frames.length === 0)
         return empty;
     try {
         const promptText = `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.
@@ -291,17 +291,25 @@ Rules:
 - NO GROUP ENTRIES (CRITICAL): NEVER output a collective/crowd label as a single entry — forbidden: "VILLAGERS", "LADIES GROUP", "KNITTING GROUP", "CROWD", "GROUP OF ...", any "*_GROUP". If 2+ similar secondary people RECUR across scenes, list them as SEPARATE numbered individuals (e.g. WOMAN_1, WOMAN_2, WOMAN_3), each with its OWN distinct face/hair/body/age. Only a truly anonymous one-off background that never recurs may be omitted entirely.
 - characters.description: ENGLISH only, factual, no camera/action words.
 - DISTINCT FACES (CRITICAL): every character MUST have a HIGHLY UNIQUE facial structure, a distinct hairstyle, a specific body type and a clearly different age. NEVER reuse the same or a similar facial description for two characters — they must look completely different from one another.`;
-        // Phase-1 on Gemini 2.5 Flash: image-heavy read is far cheaper than gpt-4o,
-        // and cast recall is backstopped by the post-Phase-2 reconcile pass, so a
-        // small frame sample suffices here.
-        const parts = [
-            { text: promptText },
-            ...frames.map(b64 => ({ inlineData: { mimeType: 'image/jpeg', data: b64 } })),
-        ];
-        const { text: rawTxt, usage } = await geminiGenerate(parts, 1600, 'gemini-2.5-flash');
+        // Phase-1 stays on gpt-4o (best cinematic summary + style — the narrative
+        // spine for the cloned script) with a dense 30-frame seed. Only 1
+        // call/video; final cast recall is double-covered by the reconcile pass.
+        const content = [{ type: 'text', text: promptText }];
+        for (const b64 of frames) {
+            content.push({ type: 'image_url', image_url: { url: `data:image/jpeg;base64,${b64}` } });
+        }
+        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: 1600, messages: [{ role: 'user', content }] }),
+        });
+        if (!res.ok)
+            return empty;
+        const data = await res.json();
+        cost?.chat('phase1', 'gpt-4o', data.usage);
+        const rawTxt = (data.choices?.[0]?.message?.content || '').trim();
         if (!rawTxt)
             return empty;
-        cost?.geminiVision('phase1', usage, 'gemini-2.5-flash');
         let parsed = {};
         try {
             const m = rawTxt.match(/\{[\s\S]*\}/);
@@ -604,12 +612,12 @@ export async function analyzeVideo(url, onProgress) {
         // 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 10 frames evenly — enough for summary + style + a naming
-        // seed. Cast RECALL no longer depends on this sample: the post-Phase-2
-        // reconcile pass derives the definitive cast from every per-scene
-        // description, so a small sample keeps the (now Gemini 2.5 Flash) Phase-1
-        // call cheap.
-        const p1SampleCount = Math.min(10, frameBuffers.length);
+        // Sample up to 30 frames evenly (matches AI_Video_Clone). Final cast
+        // RECALL is owned by the post-Phase-2 reconcile pass (reads EVERY scene),
+        // but a dense 30-frame seed gives the per-scene pass a consistent naming
+        // vocabulary up-front → cleaner reconcile + safer on hard cases. Only 1
+        // call/video so the richer sample is worth it.
+        const p1SampleCount = Math.min(30, frameBuffers.length);
         const p1Step = Math.max(1, Math.floor(frameBuffers.length / p1SampleCount));
         const p1Samples = frameBuffers
             .filter((_, i) => i % p1Step === 0)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tuna-agent",
-  "version": "0.1.151",
+  "version": "0.1.152",
   "description": "Tuna Agent - Run AI coding tasks on your machine",
   "bin": {
     "tuna-agent": "dist/cli/index.js"