Add settings.json for permissions and create QUICK_REFERENCE.md for SmolVM usage

.qwen/settings.json.orig

@@ -0,0 +1,7 @@
+{
+  "permissions": {
+    "allow": [
+      "WebSearch"
+    ]
+  }
+}

QUICK_REFERENCE.md

@@ -0,0 +1,129 @@
+# 🚀 SmolVM Quick Reference Card
+
+## ✅ Installation Status
+- **Original Error:** `ModuleNotFoundError: No module named 'pwd'` ❌
+- **Current Status:** Resolved - Running in WSL2 Ubuntu ✅
+- **Access Method:** Windows script (`smolvm.cmd`) ✅
+
+---
+
+## 💻 Using SmolVM from Windows
+
+### From CMD:
+```cmd
+smolvm.cmd doctor
+smolvm.cmd --help
+smolvm.cmd list
+smolvm.cmd create <config>
+```
+
+### From PowerShell:
+```powershell
+.\smolvm.ps1 doctor
+.\smolvm.ps1 --help
+.\smolvm.ps1 list
+.\smolvm.ps1 create <config>
+```
+
+---
+
+## 📋 Common Commands
+
+| Command | Description |
+|---------|-------------|
+| `smolvm.cmd --help` | Show all available commands |
+| `smolvm.cmd doctor` | Check installation status |
+| `smolvm.cmd setup` | Configure network permissions |
+| `smolvm.cmd list` | List all sandboxes |
+| `smolvm.cmd create <config>` | Create a new sandbox |
+| `smolvm.cmd ssh <name>` | SSH into a sandbox |
+| `smolvm.cmd stop <name>` | Stop a running sandbox |
+| `smolvm.cmd browser` | Manage browser sessions |
+| `smolvm.cmd ui` | Start dashboard UI |
+
+---
+
+## 📂 Files Created
+
+### Scripts (Ready to Use):
+- ✅ `smolvm.cmd` - Windows CMD launcher
+- ✅ `smolvm.ps1` - PowerShell launcher
+
+### Docker Setup (Alternative):
+- 📁 `smolvm-docker/` - Docker-based setup (not needed now)
+  - `Dockerfile`
+  - `docker-compose.yml`
+  - `README.md`
+
+### Documentation:
+- 📄 `SCRIPTS_README.md` - Detailed setup guide
+- 📄 `QUICK_REFERENCE.md` - This file
+
+---
+
+## ⚠️ Known Limitations
+
+### KVM Not Available
+- **Issue:** `/dev/kvm does not exist` in WSL2
+- **Impact:** Firecracker micro-VMs may not work
+- **Reason:** WSL2 doesn't expose KVM by default
+- **Status:** Basic installation works, but VM execution may fail
+
+### What Works:
+✅ SmolVM installation  
+✅ Command-line interface  
+✅ Configuration management  
+✅ Help and doctor commands  
+
+### What May Not Work:
+❌ Running actual Firecracker micro-VMs (requires KVM)  
+❌ Network setup commands (requires sudo)  
+
+---
+
+## 🔧 Making `smolvm.cmd` Available Everywhere
+
+### Option 1: Copy to Windows directory
+```cmd
+copy "E:\Git\Final_Assignment_Template\smolvm.cmd" C:\Windows\
+```
+
+### Option 2: Add to PATH
+Add `E:\Git\Final_Assignment_Template` to your system PATH environment variable.
+
+### Option 3: Create an alias (PowerShell)
+Add to your `$PROFILE`:
+```powershell
+function smolvm { & "E:\Git\Final_Assignment_Template\smolvm.cmd" @args }
+```
+
+---
+
+## 🎯 Next Steps
+
+1. **Test basic commands:**
+   ```cmd
+   smolvm.cmd --version
+   smolvm.cmd list
+   ```
+
+2. **Try creating a sandbox** (may fail without KVM):
+   ```cmd
+   smolvm.cmd create --help
+   ```
+
+3. **(Optional) Enable KVM in WSL2:**
+   - Requires Windows 11 Build 26100+
+   - Enable in `.wslconfig`: 
+     ```ini
+     [wsl2]
+     nestedVirtualization=true
+     ```
+
+---
+
+## 📞 Need Help?
+
+- Check `SCRIPTS_README.md` for detailed troubleshooting
+- Visit: https://github.com/CelestoAI/SmolVM
+- Run: `smolvm.cmd --help`

agent.py

@@ -12,6 +12,7 @@ from dotenv import load_dotenv
 from langchain_core.messages import HumanMessage, AIMessage, SystemMessage, ToolMessage
 from langchain_core.tools import tool
 from langchain_groq import ChatGroq
+from langchain_google_genai import ChatGoogleGenerativeAI
 from langgraph.graph import StateGraph, START, END
 from langchain_community.document_loaders import WikipediaLoader, UnstructuredFileLoader
 from langchain_community.document_loaders.image import UnstructuredImageLoader
@@ -171,29 +172,33 @@ class AgentState(TypedDict):
 
 # --- LLM Invocation with Fallback ---
 def _invoke_llm_with_tools(messages, fallback_count=0):
-    """Invoke LLM with tool binding and rate limit handling."""
-    model_name = os.getenv("MODEL_NAME")
-    prefer_free = os.getenv("PREFER_FREE_MODELS", "0") == "1"
-    if not model_name:
-        if prefer_free:
-            # Prefer free/open-source model; set MODEL_NAME env to a usable local model name if available
-            model_name = "open-source-local"
-        else:
-            model_name = "llama-3.3-70b-versatile" if fallback_count == 0 else "llama-3.1-8b-instant"
+    """Invoke LLM with tool binding and rate limit handling. 
+    Primary: Gemini 1.5 Flash (Multimodal, Free Tier).
+    Fallback: Groq (Llama 3.3).
+    """
     try:
-        model = ChatGroq(model=model_name, temperature=0)
+        # Primary: Gemini 1.5 Flash
+        model = ChatGoogleGenerativeAI(model="gemini-1.5-flash", temperature=0)
         model_with_tools = model.bind_tools(tools)
         return model_with_tools.invoke(messages)
     except Exception as e:
-        err_msg = str(e).lower()
-        if ("rate limit" in err_msg or "429" in err_msg) and fallback_count < 2:
-            import time
-            wait_time = 10 * (fallback_count + 1)
-            print(f"Rate limit hit. Waiting {wait_time}s...")
-            time.sleep(wait_time)
-            return _invoke_llm_with_tools(messages, fallback_count + 1)
-        print(f"LLM Error: {e}")
-        return AIMessage(content=f"ERROR: LLM invocation failed: {e}")
+        print(f"Gemini Error: {e}. Falling back to Groq...")
+        try:
+            # Fallback: Groq
+            groq_model = "llama-3.3-70b-versatile" if fallback_count == 0 else "llama-3.1-8b-instant"
+            model = ChatGroq(model=groq_model, temperature=0)
+            model_with_tools = model.bind_tools(tools)
+            return model_with_tools.invoke(messages)
+        except Exception as groq_e:
+            err_msg = str(groq_e).lower()
+            if ("rate limit" in err_msg or "429" in err_msg) and fallback_count < 2:
+                import time
+                wait_time = 10 * (fallback_count + 1)
+                print(f"Groq Rate limit hit. Waiting {wait_time}s...")
+                time.sleep(wait_time)
+                return _invoke_llm_with_tools(messages, fallback_count + 1)
+            print(f"Critical LLM Error: {groq_e}")
+            return AIMessage(content=f"ERROR: All LLM invocations failed: {groq_e}")
 
 # --- Helper Functions ---
 def is_reversed_text(question: str) -> bool:
@@ -222,28 +227,26 @@ def call_model(state: AgentState):
     
     # Add System Message if not present
     if not any(isinstance(m, SystemMessage) for m in messages):
-        system_prompt = """You are a highly capable General AI Assistant (GAIA). Your goal is to solve complex, multi-step tasks using your tools.
+        system_prompt = """You are a highly capable General AI Assistant (GAIA). Your goal is to solve complex, multi-step tasks.
 
 Your thought process MUST be methodical:
 1. THINK: 
-    - Analyze the question deeply. Identify the core goal and any constraints (e.g., specific units, date formats, or required precision).
-    - Review all available information (including attached files).
+    - Analyze the question deeply. Identify the core goal and ALL constraints (units, date formats, precision, etc.).
+    - If the task involves an image or video, describe the visual elements before attempting to solve.
     - Plan your steps. Break the problem into smaller sub-problems.
-    - Consider potential pitfalls or alternative interpretations of the question.
-2. ACT: Call tools as needed. Use `python_repl` for any math, counting, data analysis, or file processing to avoid manual errors. Use `web_search` for quick facts and `browse_url` for in-depth reading.
-3. OBSERVE: Carefully review tool outputs. If an error occurs, diagnose it and adapt your plan.
-4. REFINE: If the answer is not yet clear, iterate. Question your assumptions.
-5. VERIFY: Before providing the final answer, double-check:
-    - Does the answer directly address all parts of the question?
-    - Are the units correct? (e.g., if it asks for 'meters', don't give 'kilometers').
-    - Is the precision correct? (e.g., if it asks for 'two decimal places', ensure it has exactly two).
-    - Is the format exactly as requested?
-6. FINALIZE: Once you are absolutely confident, provide the result in the exact format: FINAL ANSWER: <answer>.
+2. ACT (Python-First): 
+    - Use `python_repl` for ANY task involving: math, counting, data analysis, list filtering (e.g., botany), or verifying logic (e.g., commutativity). DO NOT do these manually.
+    - Use `web_search` for initial discovery and `browse_url` to verify details from the source.
+3. OBSERVE: Carefully review tool outputs. If a result is ambiguous, search for a second source to triangulate.
+4. REFINE: Question your assumptions. If the answer seems too simple for a complex GAIA task, you likely missed a constraint.
+5. VERIFY: Before finalizing, double-check units and precision.
+6. FINALIZE: Provide the result in the exact format: FINAL ANSWER: <answer>.
 
 Guidelines:
-- If you find an [Attached File Local Path: ...], *always* use `read_file` to access its content.
-- Be precise. Double-check year ranges, units, and specific formatting requirements.
-- Return ONLY the final answer in the requested format when done. Do not include any extra commentary once you provide the final answer.
+- [Attached Files]: Always use `read_file` for local files.
+- Research: Don't trust a single snippet; browse the full page if the answer is buried.
+- Constraints: If the question says 'alphabetize' or 'comma-separated', use Python to ensure it is perfect.
+- Final Output: Return ONLY the final answer in the requested format.
 """
         messages = [SystemMessage(content=system_prompt)] + messages