@haposoft/cafekit 0.7.9 → 0.7.11

package/bin/install.js

@@ -471,7 +471,7 @@ function copyPlatformFiles(platformKey, results, options = {}) {
       requiredSkills = CLAUDE_MIGRATION_MANIFEST?.skills?.required || [];
     } else if (platformKey === 'antigravity') {
       // Antigravity also needs shared investigation and impact-analysis skills
-      requiredSkills = ['impact-analysis', 'debug', 'llm-moe'];
+      requiredSkills = ['impact-analysis', 'debug', 'ai-multimodal'];
     }
 
     requiredSkills
@@ -939,7 +939,7 @@ function configureGeminiKey(apiKey) {
     }
 
     // Luôn ghi trực tiếp key vào rốn của não bộ AI
-    fs.writeFileSync(localEnvFile, `GEMINI_API_KEY=${apiKey}\nVISUAL_MODEL=gemma-4-31b-it\n`, { mode: 0o600 });
+    fs.writeFileSync(localEnvFile, `GEMINI_API_KEY=${apiKey}\nVISUAL_MODEL=gemma-4-31b-it\nSEARCH_MODEL=gemini-2.5-pro\n`, { mode: 0o600 });
     console.log('  ✓ Gemini API key configured securely in project (.claude/.env)');
 
     return true;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@haposoft/cafekit",
-  "version": "0.7.9",
+  "version": "0.7.11",
   "description": "Spec-Driven Development workflow for AI coding assistants. Supports Claude Code and Antigravity with spec-first workflows plus Claude Code hapo: skills.",
   "author": "Haposoft <nghialt@haposoft.com>",
   "license": "MIT",

package/src/claude/CLAUDE.md

@@ -70,11 +70,11 @@ When you need to search the internet for information (research, docs lookup, tro
 
 | Priority | Tool | Command | When to use |
 |----------|------|---------|-------------|
-| 🥇 **P1** | `web-search.cjs` | `node .claude/scripts/web-search.cjs "query"` | **ALWAYS try first.** Works on ALL models via Gemini Google Search Grounding. Supports `--multi "q1" "q2"` for batch. Returns JSON with answer + sources. |
-| 🥈 **P2** | `WebSearch` (native) | Use WebSearch tool directly | Secondary verification, or when P1 fails/unavailable. |
-| 🥉 **P3** | `docs-fetch.js` | `node .claude/scripts/docs-fetch.js "library"` | Only when you already know a specific library and need its raw documentation. |
+| 🥇 **P1** | `web-search.cjs` | `node .claude/scripts/web-search.cjs "[query]"` | **EXCLUSIVE PRIMARY.** Works via Gemini Grounding. Supports `--multi`. **TRUST THE SYNTHESIZED ANSWER** — do NOT manually scrape source URLs. |
+| 🥈 **P2** | `WebSearch` (native) | Use WebSearch tool directly | Secondary verification, or when P1 fails. |
+| 🥉 **P3** | `docs-fetch.js` | `node .claude/scripts/docs-fetch.js "library"` | Only for fetching raw documentation when synthesis is insufficient. |
 
-**IMPORTANT**: When the user asks you to find information, research a topic, look up documentation, or investigate anything that requires internet access, you MUST use the Web Search Protocol above. Do NOT reply with "I cannot search the web" — you have `web-search.cjs` available via Bash.
+**IMPORTANT**: When the user asks you to find information, research a topic, or investigate anything that requires internet access, you MUST use the Web Search Protocol above. **NEVER** reply with "I cannot search the web". **NEVER** attempt manual `Fetch` or Python-based scraping for search results if `web-search.cjs` provides an answer. Trust the grounding.
 
 ## Code Refactoring Triggers
 

package/src/claude/agents/brainstormer.md

@@ -50,7 +50,6 @@ Before concluding any brainstorm session, verify each measurement metric:
 1. **Engineering Trinity:** YAGNI, KISS, and DRY.
 2. **Brutal Honesty:** Interrogate assumptions. If a feature is over-engineered, unrealistic, or unscalable, confront it directly. Your value lies in preventing costly mistakes.
 3. **Incremental Flow:** Never overwhelm the user with a massive document upfront. Proceed step by step, section by section.
-4. **Web Search Protocol:** When needing to search the internet for references, benchmarks, or latest practices, ALWAYS use `node .claude/scripts/web-search.cjs "query"` first (Gemini Grounding). Use native WebSearch as secondary. Use `docs-fetch.js` only for known library docs.
 
 ## Ecosystem Alliances (Collaboration Tools)
 

package/src/claude/agents/code-auditor.md

@@ -11,7 +11,7 @@ Goal: Catch the mistakes AI-written code commonly makes — logic errors, securi
 
 You DO NOT fix code. You only READ, SCORE, and REPORT.
 
-**Web Search Protocol:** When needing to verify security best practices or lookup CVE databases, ALWAYS use `node .claude/scripts/web-search.cjs "query"` first (Gemini Grounding). Use native WebSearch as secondary.
+
 
 ## Pre-Review: Blast Radius Check (MANDATORY)
 

package/src/claude/agents/god-developer.md

@@ -19,7 +19,7 @@ Any logic gaps must be clarified BEFORE typing, not discovered after bugs ship.
 - **Token efficiency**: Write concisely, report briefly, no prose.
 - **Surgical Reading (Large Files):** Never use blanket `Read` commands on files > 800 lines. Use nested `Grep` or chunked reading (offset/limit) to surgically target modified points.
 - **Component Scaffold Limit:** Any React/UI component file that exceeds 200 LOC must trigger a proactive modularization step (split into smaller child files).
-- **Web Search Protocol:** When needing to search the internet, ALWAYS use `node .claude/scripts/web-search.cjs "query"` first (Gemini Grounding). Use native WebSearch as secondary. Use `docs-fetch.js` only for known library docs.
+
 
 ## Self-Check Checklist (Before Reporting Complete)
 

package/src/claude/agents/project-manager.md

@@ -15,7 +15,7 @@ Unlike typical managers who report on "feelings" or conversational summaries, yo
 1. **Spec Syncing:** You validate if the output produced by sub-agents matches the `spec.json` requirements and the `design.md` architectural constraints.
 2. **Blocker Assassination:** You identify task stagnation (e.g., a spec stuck in 'in-progress' across multiple sessions) and force the immediate assignment of next-step actions.
 3. **Agile Aggregation:** When parallel sub-agents (like `god-developer` and `test-runner`) report completion, you sweep their logs, consolidate the facts, and generate a single authoritative **Feature Release Report**.
-4. **Web Search Protocol:** When needing to search for project management best practices, dependency updates, or changelog references, ALWAYS use `node .claude/scripts/web-search.cjs "query"` first (Gemini Grounding). Use native WebSearch as secondary.
+
 
 ## Execution Constraints
 

package/src/claude/agents/researcher.md

@@ -41,9 +41,10 @@ You possess extreme proficiency in:
 - Segregating Stable Production Practices away from Toxic Experimental Paradigms.
 - Sniffing out valid Adoption Patterns and real-world implementation trending.
 - Forgiving nothing when crafting Trade-off computational matrices for thousands of competing libraries.
-- **[PRIORITY 1]** Deploying `scripts/web-search.cjs` as the **PRIMARY search tool** for all web queries. Usage: `node .claude/scripts/web-search.cjs "query"` or `node .claude/scripts/web-search.cjs --multi "q1" "q2"`. Returns JSON with answer, sources, and citations via Gemini Google Search Grounding. ALWAYS attempt this first before any other search method.
-- **[PRIORITY 2]** If WebSearch native tool is available, use it as secondary verification or when `web-search.cjs` fails.
-- **[PRIORITY 3]** Deploying `scripts/docs-fetch.js` only when official Github/Doc URLs are already identified and you need to pull raw documentation content.
+- **[PRIORITY 1]** Deploying `node .claude/scripts/web-search.cjs "[query]"` as the **EXCLUSIVE PRIMARY search tool**. This tool uses Gemini Grounding to return a synthesized **answer** plus cited sources. **STOP SEARCHING** once you have a sufficient answer from this script. Do NOT manually crawl source URLs if the provided synthesis is clear.
+- **[PRIORITY 2]** Trust the script's output directly. READ the JSON and extract the `answer` field. **STRICTLY FORBIDDEN**: Writing Python scripts to parse this JSON or manually `Fetch` every URL listed in the sources unless the user explicitly demands a deep-dive implementation detail only found in a raw document.
+- **[PRIORITY 3]** If `web-search.cjs` fails or returns no results, use native `WebSearch` tool (if available) as a backup.
+- **[PRIORITY 4]** Deploying `scripts/docs-fetch.js` ONLY for raw documents where the direct URL is already known and synthesis is insufficient.
 - Deploying Bash and raw Grep utilities to surgically dissect embedded Document architectures and internal file payloads to evaluate raw insights.
 
 **ABSOLUTE IMMOVEABLE DIRECTIVE**: You are **STRICTLY PROHIBITED** from generating executable endpoint "Implementation Code". You exist ONLY to maneuver data streams, render synthesis Summary text, and return comprehensive Markdown documentation pathways to the main caller Agent.

package/src/claude/agents/spec-maker.md

@@ -19,7 +19,7 @@ You DO NOT write implementation code. You produce Specifications that downstream
 - **The 5 Whys:** Dig past the surface request to find the REAL problem.
 - **80/20 MVP:** Identify the 20% of features that deliver 80% of value.
 - **Systems Thinking:** How does this feature connect to (or break) existing systems?
-- **Web Search Protocol:** When needing to search the internet, ALWAYS use `node .claude/scripts/web-search.cjs "query"` first (Gemini Grounding). Use native WebSearch as secondary. Use `docs-fetch.js` only for known library docs.
+
 
 ## Pre-Completion Checklist
 

package/src/claude/agents/ui-ux-designer.md

@@ -18,7 +18,7 @@ You are an award-caliber UI/UX designer. You merge aesthetic excellence with eng
 - **Micro-interactions:** Purposeful animations that enhance UX without performance cost.
 - **Accessibility:** WCAG 2.1 AA compliance as a baseline, not an afterthought.
 - **3D/WebGL:** Three.js scene composition, shader development (when appropriate).
-- **Web Search Protocol:** When needing to search the internet for design trends, component libraries, or accessibility guides, ALWAYS use `node .claude/scripts/web-search.cjs "query"` first (Gemini Grounding). Use native WebSearch as secondary. Use `docs-fetch.js` only for known library docs.
+
 
 ## Design Workflow
 

package/src/claude/migration-manifest.json

@@ -21,7 +21,7 @@
       "hotfix",
       "impact-analysis",
       "inspect",
-      "llm-moe",
+      "ai-multimodal",
       "mobile-development",
       "react-best-practices",
       "research",

package/src/claude/scripts/web-search.cjs

@@ -19,11 +19,7 @@ const fs = require('fs');
 // ---------------------------------------------------------------------------
 // ENV Resolution: .claude/.env → process.env
 // ---------------------------------------------------------------------------
-function resolveApiKey() {
-  // Priority 1: Already in environment
-  if (process.env.GEMINI_API_KEY) return process.env.GEMINI_API_KEY;
-
-  // Priority 2: Project-local .claude/.env
+function loadEnv() {
   const envPaths = [
     path.join(process.cwd(), '.claude', '.env'),
     path.join(process.cwd(), '..', '.claude', '.env'),
@@ -33,13 +29,21 @@ function resolveApiKey() {
     try {
       if (fs.existsSync(envPath)) {
         const content = fs.readFileSync(envPath, 'utf8');
-        const match = content.match(/^GEMINI_API_KEY=(.+)$/m);
-        if (match) return match[1].trim().replace(/^["']|["']$/g, '');
+        content.split(/\r?\n/).forEach(line => {
+          const match = line.match(/^([^=]+)=(.*)$/);
+          if (match) {
+            const key = match[1].trim();
+            const val = match[2].trim().replace(/^["']|["']$/g, '');
+            // Only set if not already present in environment
+            if (process.env[key] === undefined) {
+              process.env[key] = val;
+            }
+          }
+        });
+        return; // Loaded successfully, no need to check other paths
       }
     } catch { /* skip */ }
   }
-
-  return null;
 }
 
 // ---------------------------------------------------------------------------
@@ -89,10 +93,38 @@ function callGemini(apiKey, query, model) {
   });
 }
 
+// ---------------------------------------------------------------------------
+// Resolve Vertex AI grounding redirect URLs to real URLs
+// ---------------------------------------------------------------------------
+function resolveRedirectUrl(url) {
+  return new Promise((resolve) => {
+    if (!url || !url.includes('grounding-api-redirect')) {
+      resolve(url);
+      return;
+    }
+
+    const protocol = url.startsWith('https') ? https : require('http');
+    const req = protocol.request(url, { method: 'HEAD', timeout: 5000 }, (res) => {
+      // Follow redirect chain - Location header has the real URL
+      resolve(res.headers.location || url);
+    });
+    req.on('error', () => resolve(url));
+    req.on('timeout', () => { req.destroy(); resolve(url); });
+    req.end();
+  });
+}
+
+async function resolveAllUrls(sources) {
+  return Promise.all(sources.map(async (src) => {
+    const realUrl = await resolveRedirectUrl(src.url);
+    return { ...src, url: realUrl };
+  }));
+}
+
 // ---------------------------------------------------------------------------
 // Parse Grounding Metadata → Structured Output
 // ---------------------------------------------------------------------------
-function parseResponse(geminiResponse, query) {
+async function parseResponse(geminiResponse, query) {
   const candidate = geminiResponse.candidates?.[0];
   if (!candidate) return { query, error: 'No candidates returned' };
 
@@ -100,11 +132,22 @@ function parseResponse(geminiResponse, query) {
   const meta = candidate.groundingMetadata || {};
 
   // Extract source URLs from groundingChunks
-  const sources = (meta.groundingChunks || []).map(chunk => ({
+  let sources = (meta.groundingChunks || []).map(chunk => ({
     title: chunk.web?.title || 'Unknown',
     url: chunk.web?.uri || '',
   }));
 
+  // Resolve redirect URLs to real URLs
+  sources = await resolveAllUrls(sources);
+
+  // Deduplicate by resolved URL
+  const seen = new Set();
+  sources = sources.filter(s => {
+    if (seen.has(s.url)) return false;
+    seen.add(s.url);
+    return true;
+  });
+
   // Extract search queries used by the model
   const searchQueries = meta.webSearchQueries || [];
 
@@ -132,7 +175,8 @@ async function main() {
     process.exit(1);
   }
 
-  const apiKey = resolveApiKey();
+  loadEnv();
+  const apiKey = process.env.GEMINI_API_KEY;
   if (!apiKey) {
     console.error(JSON.stringify({
       error: 'GEMINI_API_KEY not found. Set it in .claude/.env or environment variable.'
@@ -140,15 +184,20 @@ async function main() {
     process.exit(1);
   }
 
-  // Use model from env or default to gemini-2.5-flash
-  const model = process.env.SEARCH_MODEL || 'gemini-2.5-flash';
+  // Determine which model to use. User might configure MODEL or VISUAL_MODEL in their .env
+  let model = process.env.SEARCH_MODEL || process.env.MODEL || process.env.VISUAL_MODEL || 'gemini-2.5-flash';
+
+  // Google Search Grounding ONLY supports Gemini models (not Claude, not Gemma)
+  if (!model.toLowerCase().includes('gemini') && !model.toLowerCase().includes('learnlm')) {
+    model = 'gemini-2.5-flash'; // Fallback to safe search model
+  }
 
   const results = [];
 
   for (const query of queries) {
     try {
       const raw = await callGemini(apiKey, query, model);
-      results.push(parseResponse(raw, query));
+      results.push(await parseResponse(raw, query));
     } catch (err) {
       results.push({ query, error: err.message });
     }

package/src/claude/skills/research/SKILL.md

@@ -23,11 +23,11 @@ Call the `TaskCreate` tool to spin up the `researcher` subagent.
 **Instructions to pass to Researcher:**
 ```text
 Conduct comprehensive research on: [topic]
-Constraint 1: ALWAYS use `node .claude/scripts/web-search.cjs "query"` as PRIMARY search method (supports --multi for batch). This uses Gemini Google Search Grounding and returns JSON with answer + sources.
-Constraint 2: Use native WebSearch tool as secondary verification or when web-search.cjs fails.
-Constraint 3: Use scripts/docs-fetch.js ONLY when official Github/Doc URLs are already identified.
-Constraint 4: Limit total search calls to a maximum of 5 distinct queries to conserve context.
-Constraint 5: Validate information via cross-referencing capabilities.
+Constraint 1: ALWAYS use `node .claude/scripts/web-search.cjs "[query]"` as the EXCLUSIVE primary search method. This tool uses Gemini Grounding and returns a synthesized answer + cited sources. Do NOT manually crawl source URLs if the script provides a sufficient answer.
+Constraint 2: TRUST THE SYNTHESIS. The output contains the research results. Read the JSON and use the `answer` field directly. Do NOT write Python scripts to re-parse it or manually `Fetch` sources unless deep implementation details are missing.
+Constraint 3: Use native WebSearch or manual Fetch ONLY if the script fails or returns no results.
+Constraint 4: Limit total search calls to a maximum of 5 distinct queries.
+Constraint 5: Stop excessive "chain-searching". Use the grounding answer as the definitive summary.
 Output Format: Must strictly follow the 'Standard Research Report' layout.
 ```
 

package/src/claude/skills/test/references/execution-strategy.md

@@ -237,7 +237,7 @@ Collects Core Web Vitals: `LCP`, `FID`, `CLS`, `FCP`, `TTFB`, `JSHeapUsedSize`.
 
 ---
 
-### Phase C-5: Responsive Screenshots (screenshot.js & visual-analyze.js)
+### Phase C-5: Responsive Screenshots (screenshot.js & gemini_batch_process.py)
 
 ```bash
 # Capture screenshots

package/src/claude/skills/test/references/failure-triage.md

@@ -23,7 +23,7 @@ Referenced by `SKILL.md` Phase 3 and by the `test-runner` agent.
 | 12 | **SEO Error** | Missing meta tags, H1 issues, broken canonical | `evaluate.js` |
 | 13 | **Security Warning** | Missing HTTP security headers, exposed secrets | `network.js` / source |
 | 14 | **Broken Links** | Internal crawler found 404s on discovered pages | Phase 0.5 Crawler |
-| 15 | **UI Visual Error** | AI visual analysis detected layout break, overlap | `visual-analyze.js` |
+| 15 | **UI Visual Error** | AI visual analysis detected layout break, overlap | `gemini_batch_process.py` |
 
 ---
 

package/src/claude/skills/llm-moe/SKILL.md

@@ -1,62 +0,0 @@
-# `hapo:llm-moe` Skill
-
-The **LLM Mixture-of-Experts (MoE) Hub** serves as a centralized gateway for executing advanced contextual tasks (Visual Analysis, Document Understanding, Data Extraction) across various models (Gemini, Gemma) using API scripts.
-
-By decoupling LLM functionality into this standalone skill, `hapo:test` or any other workflow can simply offload complex multimodal tasks to `hapo:llm-moe` scripts rather than bundling their own integration.
-
-## Core Capabilities (Hiện tại & Mở rộng)
-
-Được thiết kế theo chuẩn MoE (Mixture of Experts), Skill này không chỉ giới hạn ở một Model duy nhất mà tự động định tuyến (route) các tác vụ chuyên biệt tới các model phù hợp (như *gemma-4-31b-it* cho logic, *gemini-2.5-flash* cho đa phương tiện siêu tốc, hoặc *gemini-2.5-pro* cho suy luận y tế/toán học).
-
-### 1. Visual Understanding (Đã triển khai)
-- **UI/UX Regression Analysis:** Phân tích ảnh chụp màn hình tự động, dò tìm CSS layout vỡ, button đè text, responsive rác.
-- **Visual Q&A / Object Detection:** Truy vấn vị trí phần tử HTML, đếm số lượng form inputs trong màn hình.
-- *(Sử dụng script: `scripts/visual-analyze.js`)*
-
-### 2. Optical Character Recognition - OCR (Sắp triển khai)
-- **Data Extraction:** Trích xuất mảng JSON từ ảnh chụp Hóa đơn (Invoices), Căn cước (ID Cards), hay Bảng biểu (Tables).
-- **Handwriting Parsing:** Dịch chữ viết tay trên biểu mẫu thành text markdown.
-
-### 3. Document AI & Parsing (Sắp triển khai)
-- **PDF Comprehension:** Đọc và phân mảnh (chunking) file tài liệu PDF (>1000 trang) trả về các luồng tài liệu trích xuất nội dung liên quan (RAG base).
-- **Codebase Indexing:** Hỗ trợ nhai các file log, file báo cáo (.csv, .xlsx) để báo cáo phân tích rủi ro hệ thống.
-
-### 4. Generative Engineering (Sắp triển khai)
-- **Code & Scaffold Generation:** Sinh cấu trúc thư mục, Boilerplate code dựa trên bản vẽ Design UI (Image-to-Code).
-- **Audio/Video Transcribing:** Cắt âm thanh từ luồng test tích hợp (nếu có WebRTC/Media tests) và phân tích lỗi thoại.
-
-## Usage Guide for Agents
-
-Other agents (like `test-runner` or `reviewer`) should call `hapo:llm-moe`'s tools by launching its scripts securely via `bash`.
-
-### 1. Visual Analysis (`visual-analyze.js`)
-
-Used to interpret screenshot logic, UI regressions, or visual QA.
-
-**Caller requirements:**
-- Requires Node.js.
-- Execution directory must be relative to the caller or via `{{SKILLS_DIR}}/llm-moe/scripts/...`
-
-**Command format:**
-```bash
-node <path-to-skills>/llm-moe/scripts/visual-analyze.js \
-  --image "path/to/screenshot.png" \
-  --prompt "Check if the button overlaps the text."
-```
-
-**JSON Output:**
-```json
-{
-  "success": true,
-  "file": "path/to/screenshot.png",
-  "analysis": "The red submit button overlaps the footer text by 15px. Layout is broken."
-}
-```
-
-## Model Configuration
-
-The API Key is globally seeded during the `cafekit` setup into `~/.gemini/.env` and the local `.env` of your workspace.
-
-Fallback override environment variables:
-- `GEMINI_API_KEY`: The authentication key.
-- `VISUAL_MODEL` (Optional): Specify the underlying Google/Gemma model (default: `gemma-4-31b-it`).

package/src/claude/skills/llm-moe/references/vision-understanding.md

@@ -1,36 +0,0 @@
-# Vision Understanding
-
-The `hapo:llm-moe` skill leverages the `visual-analyze.js` agent-compatible script to convert unstructured screenshots and pictures into structured JSON insights.
-
-## Quick Start (Node.js)
-
-To parse an image, run the script from bash:
-
-```bash
-cd <skills_dir>/llm-moe/scripts
-npm install  # (only required on first setup to load @google/generative-ai)
-
-node visual-analyze.js \
-  --image "../../test/screenshots/ui-error.png" \
-  --prompt "Find the login button and describe its visual state."
-```
-
-## Prompt Engineering for Vision
-
-To get the most accurate layout results from `gemma-4-31b-it` or `gemini-2.5-flash`, structure your `--prompt` effectively:
-
-### Regression Testing (UI Checks)
-Instead of asking "Is it broken?", ask for explicit state analysis:
-> `"List all elements that are overlapping. Determine if the main text is cut off by the border constraints. Return only the issues found."`
-
-### Component Identification
-> `"Identify the CSS Hex color of the top primary button and describe its corner radius visually."`
-
-### Accessibility Visual Check
-> `"Look at the form inputs in this image. Do their text contrasts against the background seem accessible? Are the labels visually aligned with the inputs?"`
-
-## Technical Details
-
-- **Supported Inputs:** PNG, JPEG, WEBP.
-- **Size Limitation:** Try to keep screenshots under 4MB to ensure fast tokenization.
-- **Model Loading:** The script creates an inline base64 string and embeds it directly into the Prompt Part payload of the `@google/generative-ai` request matrix.

package/src/claude/skills/llm-moe/scripts/package.json

@@ -1,10 +0,0 @@
-{
-  "name": "hapo-llm-moe-scripts",
-  "version": "1.0.0",
-  "description": "LLM scripts for visual and text analysis using Gemini/Gemma models.",
-  "main": "visual-analyze.js",
-  "type": "module",
-  "dependencies": {
-    "@google/generative-ai": "^0.21.0"
-  }
-}

package/src/claude/skills/llm-moe/scripts/visual-analyze.js

@@ -1,67 +0,0 @@
-import { parseArgs } from 'util';
-import fs from 'fs';
-import { GoogleGenerativeAI } from '@google/generative-ai';
-
-const options = {
-  image: { type: 'string' },
-  prompt: { type: 'string', default: 'Identify any UI bugs, visual overlaps, missing images, or cut-off text in this screenshot. Return a concise analysis.' },
-};
-
-const args = parseArgs({ args: process.argv.slice(2), options }).values;
-
-if (!args.image || !fs.existsSync(args.image)) {
-  console.error("Error: --image path is required and must exist.");
-  process.exit(1);
-}
-
-// Ensure the API key is set
-const apiKey = process.env.GEMINI_API_KEY;
-if (!apiKey) {
-  console.error(JSON.stringify({
-    success: false,
-    error: "GEMINI_API_KEY environment variable is missing."
-  }));
-  process.exit(1);
-}
-
-async function analyze() {
-  try {
-    const genAI = new GoogleGenerativeAI(apiKey);
-    const modelName = process.env.VISUAL_MODEL || "gemma-4-31b-it";
-    const model = genAI.getGenerativeModel({ model: modelName });
-
-    const imageBytes = fs.readFileSync(args.image).toString("base64");
-    
-    // Validate image format based on extension (simple approximation)
-    const ext = args.image.split('.').pop().toLowerCase();
-    const mimeType = ext === 'png' ? 'image/png' : (ext === 'webp' ? 'image/webp' : 'image/jpeg');
-
-    const result = await model.generateContent([
-      { text: args.prompt },
-      {
-        inlineData: {
-          data: imageBytes,
-          mimeType: mimeType
-        }
-      }
-    ]);
-
-    const analysisText = result.response.text();
-    
-    console.log(JSON.stringify({
-      success: true,
-      file: args.image,
-      analysis: analysisText
-    }, null, 2));
-
-  } catch (error) {
-    console.error(JSON.stringify({
-      success: false,
-      file: args.image,
-      error: error.message
-    }, null, 2));
-    process.exit(1);
-  }
-}
-
-analyze();