tuna-agent 0.1.148 → 0.1.150

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

@@ -29,6 +29,8 @@ const RATES = {
     'gpt-4o': { in: 2.50, out: 10.0 },
     // Gemini 3 Flash preview: text+image input share one rate, output 6x.
     'gemini-3-flash-preview': { in: 0.50, out: 3.0 },
+    // Gemini 2.5 Flash: cheaper image-heavy reads (used by Phase-1).
+    'gemini-2.5-flash': { in: 0.30, out: 2.50 },
 };
 // Per-run cost+token accumulator. Single-threaded JS → plain mutation is safe
 // even across the parallel visionDescribe calls.
@@ -50,12 +52,12 @@ class CostTracker {
     }
     // Gemini reports usageMetadata.{promptTokenCount,candidatesTokenCount}
     // instead of OpenAI's prompt_tokens/completion_tokens.
-    geminiVision(bucket, usage) {
+    geminiVision(bucket, usage, model = 'gemini-3-flash-preview') {
         if (!usage) {
             this.add(bucket, 0);
             return;
         }
-        const r = RATES['gemini-3-flash-preview'];
+        const r = RATES[model];
         const cost = ((usage.promptTokenCount || 0) / 1e6) * r.in + ((usage.candidatesTokenCount || 0) / 1e6) * r.out;
         this.add(bucket, cost);
     }
@@ -174,12 +176,20 @@ ${rawText}`,
 // One Gemini generateContent call with key rotation + exponential backoff on
 // 429/5xx. A single free-tier key under the 5-way concurrent batch WILL
 // rate-limit; retrying (slower) beats dropping the scene description.
-async function geminiGenerate(parts, maxOutputTokens) {
+async function geminiGenerate(parts, maxOutputTokens, model = GEMINI_MODEL) {
     if (!GEMINI_KEYS.length)
         return { text: '' };
+    const generationConfig = { maxOutputTokens, temperature: 0.4 };
+    // Gemini 2.5 Flash is a thinking model: with thinking ON it burns the whole
+    // maxOutputTokens budget on internal reasoning and returns EMPTY text. These
+    // are extraction/description tasks (no CoT needed) → disable thinking so the
+    // budget goes to the actual answer.
+    if (model.includes('2.5-flash')) {
+        generationConfig.thinkingConfig = { thinkingBudget: 0 };
+    }
     const body = JSON.stringify({
         contents: [{ parts }],
-        generationConfig: { maxOutputTokens, temperature: 0.4 },
+        generationConfig,
     });
     const MAX_ATTEMPTS = 6;
     let keyIdx = 0;
@@ -187,7 +197,7 @@ async function geminiGenerate(parts, maxOutputTokens) {
     for (let attempt = 0; attempt < MAX_ATTEMPTS; attempt++) {
         const key = GEMINI_KEYS[keyIdx % GEMINI_KEYS.length];
         try {
-            const res = await fetch(`https://generativelanguage.googleapis.com/v1beta/models/${encodeURIComponent(GEMINI_MODEL)}:generateContent?key=${key}`, { method: 'POST', headers: { 'Content-Type': 'application/json' }, body });
+            const res = await fetch(`https://generativelanguage.googleapis.com/v1beta/models/${encodeURIComponent(model)}:generateContent?key=${key}`, { method: 'POST', headers: { 'Content-Type': 'application/json' }, body });
             if (res.status === 429 || res.status >= 500) {
                 lastErr = `Gemini ${res.status}`;
                 keyIdx++; // rotate to the next key before backing off
@@ -259,13 +269,10 @@ async function visionExtractPhase1(frames, transcript, cost) {
         master_cast_prompt: '',
         characters: [],
     };
-    if (!OPENAI_KEY || frames.length === 0)
+    if (!GEMINI_KEYS.length || 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, in order. Use them + the transcript to analyse the whole piece.
+        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.
 
 Transcript context: "${(transcript || '').slice(0, 4000)}"
 
@@ -282,22 +289,18 @@ Rules:
 - characters.name: a stable short uppercase label reused for this subject (e.g. "THE BISHOP", "U-94 SUBMARINE"). Max 4 words.
 - RECALL (CRITICAL): list EVERY distinct recurring subject SEPARATELY. If a family or group recurs, include EACH member as its own entry (e.g. adult man, adult woman, older boy, younger girl) — never merge them into one. Skip only true one-off background extras. Be COMPLETE: missing a recurring character is worse than one extra. Up to 8.
 - 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.`,
-            },
+- 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 } })),
         ];
-        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)
+        const { text: rawTxt, usage } = await geminiGenerate(parts, 1600, 'gemini-2.5-flash');
+        if (!rawTxt)
             return empty;
-        const data = await res.json();
-        cost?.chat('phase1', 'gpt-4o', data.usage);
-        const rawTxt = (data.choices?.[0]?.message?.content || '').trim();
+        cost?.geminiVision('phase1', usage, 'gemini-2.5-flash');
         let parsed = {};
         try {
             const m = rawTxt.match(/\{[\s\S]*\}/);
@@ -599,10 +602,12 @@ 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 30 frames evenly across the whole video — denser sampling
-        // is critical for cast RECALL on sparse-narration (ASMR) videos where
-        // Phase-1 relies almost entirely on frames (matches AI_Video_Clone).
-        const p1SampleCount = Math.min(30, frameBuffers.length);
+        // 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);
         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": "0.1.148",
+  "version": "0.1.150",
   "description": "Tuna Agent - Run AI coding tasks on your machine",
   "bin": {
     "tuna-agent": "dist/cli/index.js"