tuna-agent 0.1.142 → 0.1.143

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

@@ -31,6 +31,14 @@ export interface AnalyzeVideoResult {
         voiceover: string;
         visual_description: string;
     }>;
+    stats?: {
+        duration_ms: number;
+        cost_usd: number;
+        breakdown: Record<string, {
+            calls: number;
+            cost: number;
+        }>;
+    };
     isError?: boolean;
     error?: string;
 }

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

@@ -14,6 +14,33 @@ const OPENAI_KEY = process.env.OPENAI_API_KEY || '';
 const YT_DLP = process.env.YT_DLP_BIN || '/home/gatoasang94/.local/bin/yt-dlp';
 const FFMPEG = process.env.FFMPEG_BIN || '/usr/bin/ffmpeg';
 const FFPROBE = process.env.FFPROBE_BIN || '/usr/bin/ffprobe';
+// USD rates (update if OpenAI pricing changes). Chat = per 1M tokens.
+const RATES = {
+    whisperPerMin: 0.006,
+    'gpt-4o-mini': { in: 0.15, out: 0.60 },
+    'gpt-4o': { in: 2.50, out: 10.0 },
+};
+// Per-run cost+token accumulator. Single-threaded JS → plain mutation is safe
+// even across the parallel visionDescribe calls.
+class CostTracker {
+    breakdown = {};
+    add(bucket, cost) {
+        const b = this.breakdown[bucket] || (this.breakdown[bucket] = { calls: 0, cost: 0 });
+        b.calls++;
+        b.cost += cost;
+    }
+    chat(bucket, model, usage) {
+        if (!usage) {
+            this.add(bucket, 0);
+            return;
+        }
+        const r = RATES[model];
+        const cost = ((usage.prompt_tokens || 0) / 1e6) * r.in + ((usage.completion_tokens || 0) / 1e6) * r.out;
+        this.add(bucket, cost);
+    }
+    whisper(audioSec) { this.add('whisper', (audioSec / 60) * RATES.whisperPerMin); }
+    total() { return Object.values(this.breakdown).reduce((s, b) => s + b.cost, 0); }
+}
 // Downloaded source videos are cached by URL hash so re-analyze doesn't
 // re-download (saves bandwidth + time on long clips). relabs01 shares disk
 // with Demucs + the local media server, so the cache is bounded: drop files
@@ -85,7 +112,7 @@ async function whisperTranscribe(audioPath) {
         throw new Error(`whisper ${res.status}: ${(await res.text()).slice(0, 300)}`);
     return res.json();
 }
-async function correctTranscript(rawText, language) {
+async function correctTranscript(rawText, language, cost) {
     if (!OPENAI_KEY || !rawText || rawText.length < 20)
         return rawText;
     try {
@@ -111,6 +138,7 @@ ${rawText}`,
         if (!res.ok)
             return rawText;
         const data = await res.json();
+        cost?.chat('correction', 'gpt-4o-mini', data.usage);
         return data.choices?.[0]?.message?.content?.trim() || rawText;
     }
     catch {
@@ -122,7 +150,7 @@ ${rawText}`,
 // turns") instead of guessing from a single frozen midpoint. The model is
 // told the frames are chronological so it describes the action arc, not 3
 // separate moments.
-async function visionDescribe(frameB64s, voiceoverText, castContext = '') {
+async function visionDescribe(frameB64s, voiceoverText, castContext = '', cost) {
     if (!OPENAI_KEY)
         return '';
     const frames = frameB64s.filter(Boolean);
@@ -164,6 +192,7 @@ Voiceover during this scene: "${voiceoverText || '(none)'}"` },
     if (!res.ok)
         return '';
     const data = await res.json();
+    cost?.chat('vision', 'gpt-4o-mini', data.usage);
     return data.choices?.[0]?.message?.content?.trim() || '';
 }
 // Phase 1 (the strong part of AI_Video_Clone, ported): ONE gpt-4o call over
@@ -177,7 +206,7 @@ Voiceover during this scene: "${voiceoverText || '(none)'}"` },
 // 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) {
+async function visionExtractPhase1(frames, transcript, cost) {
     const empty = {
         video_summary: '',
         video_style: '',
@@ -220,6 +249,7 @@ Rules:
         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();
         let parsed = {};
         try {
@@ -253,6 +283,8 @@ Rules:
 }
 export async function analyzeVideo(url, onProgress) {
     const progress = onProgress || (() => { });
+    const cost = new CostTracker();
+    const t0 = Date.now();
     const tmpDir = path.join(os.tmpdir(), 'tuna-analyze-' + crypto.randomBytes(6).toString('hex'));
     await fs.mkdir(tmpDir, { recursive: true });
     // Video lives in the persistent URL-keyed cache (NOT tmpDir) so re-analyze
@@ -314,9 +346,10 @@ export async function analyzeVideo(url, onProgress) {
         progress('Đang transcribe bằng Whisper...');
         console.log('[analyze_video] Transcribing via Whisper');
         const rawTranscript = await whisperTranscribe(audioPath);
+        cost.whisper(durationSec);
         progress('Đang sửa lỗi transcript...');
         console.log('[analyze_video] AI correcting transcript');
-        const correctedText = await correctTranscript(rawTranscript.text, rawTranscript.language);
+        const correctedText = await correctTranscript(rawTranscript.text, rawTranscript.language, cost);
         const transcript = { ...rawTranscript, text: correctedText };
         const segments = transcript.segments || [];
         const sceneSlots = [];
@@ -445,7 +478,7 @@ export async function analyzeVideo(url, onProgress) {
             .filter((_, i) => i % p1Step === 0)
             .slice(0, p1SampleCount)
             .map(f => f.thumb.toString('base64'));
-        const { video_summary, video_style, master_cast_prompt, characters } = await visionExtractPhase1(p1Samples, transcript.text || '');
+        const { video_summary, video_style, master_cast_prompt, characters } = await visionExtractPhase1(p1Samples, transcript.text || '', cost);
         const castContext = characters.length
             ? characters.map(c => `- ${c.name}: ${c.description}`).join('\n')
             : '';
@@ -460,7 +493,7 @@ export async function analyzeVideo(url, onProgress) {
             progress(`Đang phân tích scene ${b + 1}-${Math.min(b + BATCH_SIZE, frameBuffers.length)}/${frameBuffers.length}...`);
             const results = await Promise.all(batch.map(async ({ idx, frames, thumb, slot }) => {
                 try {
-                    const visual_description = await visionDescribe(frames.map(f => f.toString('base64')), slot.voiceover, castContext);
+                    const visual_description = await visionDescribe(frames.map(f => f.toString('base64')), slot.voiceover, castContext, cost);
                     return {
                         scene_number: idx + 1,
                         timestamp_start: Math.round(slot.start * 10) / 10,
@@ -490,6 +523,11 @@ export async function analyzeVideo(url, onProgress) {
             characters,
             segments: segments.map((s) => ({ start: s.start, end: s.end, text: s.text })),
             scenes,
+            stats: {
+                duration_ms: Date.now() - t0,
+                cost_usd: +cost.total().toFixed(4),
+                breakdown: cost.breakdown,
+            },
         };
     }
     finally {

package/dist/daemon/extension-handlers.js

@@ -85,7 +85,15 @@ export async function handleClaudePrompt(ws, code, taskId, prompt, systemPrompt,
             timeoutMs: timeoutMs && timeoutMs > 0 ? timeoutMs : 60000,
         });
         const text = typeof result === 'string' ? result : result.result || JSON.stringify(result);
-        console.log(`[claude_prompt] Result: ${text.substring(0, 200)}`);
+        // claude-cli reports the run's real cost/duration in --output-format json.
+        const meta = (result && typeof result === 'object')
+            ? {
+                costUsd: result.costUsd,
+                durationMs: result.durationMs,
+                numTurns: result.numTurns,
+            }
+            : undefined;
+        console.log(`[claude_prompt] Result: ${text.substring(0, 200)} | cost=$${meta?.costUsd ?? '?'} dur=${meta?.durationMs ?? '?'}ms`);
         // Try to parse JSON from response
         const cleaned = text.replace(/```json?\s*/g, '').replace(/```/g, '').trim();
         let parsed = null;
@@ -100,8 +108,8 @@ export async function handleClaudePrompt(ws, code, taskId, prompt, systemPrompt,
                 }
                 catch { }
         }
-        ws.sendExtensionEvent(code, { type: 'prompt_result', taskId, result: parsed, raw: text });
-        ws.sendExtensionDone(code, taskId, { result: parsed, raw: text });
+        ws.sendExtensionEvent(code, { type: 'prompt_result', taskId, result: parsed, raw: text, meta });
+        ws.sendExtensionDone(code, taskId, { result: parsed, raw: text, meta });
     }
     catch (err) {
         console.error(`[claude_prompt] Error: ${err.message}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tuna-agent",
-  "version": "0.1.142",
+  "version": "0.1.143",
   "description": "Tuna Agent - Run AI coding tasks on your machine",
   "bin": {
     "tuna-agent": "dist/cli/index.js"