Merge pull request #3 from togethercomputer/fede/more-tests

tests

Author: vinid

.github/workflows/ci.yml

@@ -20,7 +20,7 @@ jobs:
     - name: Install test dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install pytest httpx requests
+        pip install pytest pytest-asyncio httpx requests
         pip install -e .
 
     - name: Build Docker image
@@ -42,8 +42,9 @@ jobs:
         TOGETHER_API_KEY: ${{ secrets.TOGETHER_API_KEY }}
       run: |
         python -m pytest tests/test_functional.py -v
-        python -m pytest tests/test_executors.py -v
         python -m pytest tests/test_agent_data_analysis.py -v
+        python -m pytest tests/test_variable_isolation.py -v
+        python -m pytest tests/test_executors.py -v
 
     - name: Stop container
       if: always()

README.md

@@ -2,6 +2,17 @@
 
 An AI-powered data analysis assistant that follows the ReAct (Reasoning + Acting) framework to perform comprehensive data science tasks. The agent can execute Python code either locally via Docker or in the cloud using Together's Code Interpreter.
 
+## ⚠️ Experimental Software Notice
+
+**This is an experimental tool** powered by large language models. Please be aware of the following limitations:
+
+- **AI-Generated Code**: All analysis and code is generated by AI and may contain errors, bugs, or suboptimal approaches
+- **No Guarantee of Accuracy**: Results should be carefully reviewed and validated before making important decisions
+- **Learning Tool**: Best suited for exploration, learning, and initial analysis rather than production use
+- **Human Oversight Required**: Always verify outputs, especially for critical business or research applications
+- **Evolving Technology**: Capabilities and reliability may vary as the underlying models are updated
+
+
 ## 🚀 Installation
 
 ### Prerequisites
@@ -50,7 +61,7 @@ The ReAct agent supports two execution modes for running Python code:
 | **Setup Required** | API key only | Docker + docker-compose |
 | **File Handling** | ☁️ Files uploaded to cloud | 🏠 Files stay local |
 | **Session Persistence** | ✅ Managed by Together | ✅ Local session management |
-| **Session Isolation** | ✅ Independent isolated sessions | ⚠️ Shared filesystem across sessions |
+| **Session Isolation** | ✅ Independent isolated sessions | ⚠️ Limited isolation (see below) |
 | **Concurrent Usage** | ✅ Multiple users/processes safely | ⚠️ File conflicts possible |
 | **Dependencies** | Pre-installed environment | Custom Docker environment |
 | **Plot Saving** | ✅ Can save created plots to disk | ❌ Plots not saved to disk |
@@ -61,6 +72,44 @@ The ReAct agent supports two execution modes for running Python code:
 
 **Docker Mode**: All code execution and file processing happens locally in your Docker container.
 
+## ⚠️ Docker Mode Session Isolation Limitations
+
+**Important**: While Docker mode provides basic session isolation for variables, it has significant limitations:
+
+### ✅ What IS Isolated:
+- **User variables**: `x = 1` in one session won't affect another session
+- **Session state**: Each session maintains its own execution context
+
+### ❌ What is NOT Isolated:
+- **Module modifications**: Changes to imported libraries affect ALL sessions
+- **Global state changes**: Modifications to `sys.path`, `os.environ`, etc. are shared
+- **Library monkey-patching**: Modifying `json.dumps`, `numpy` settings, etc. corrupts other sessions
+
+### Examples of Problematic Code:
+```python
+# These operations will affect ALL sessions:
+import json
+json.dumps = custom_function  # ❌ Breaks all sessions
+
+import sys
+sys.path.append('/custom/path')  # ❌ Affects all sessions
+
+import os
+os.environ['KEY'] = 'value'  # ❌ Global environment change
+```
+
+### Docker Mode is OK For:
+- **Data analysis workflows**: Reading CSV/JSON files, pandas operations, statistical analysis
+- **Machine learning**: Training models, feature engineering, model evaluation
+- **Visualization**: Creating plots with matplotlib, seaborn, plotly
+- **Standard data science**: EDA, data cleaning, hypothesis testing
+- **Single-user development** and **testing environments**
+
+### When to Use TCI Mode Instead:
+- **Multi-user environments** where sessions must be completely isolated
+- **Production applications** with concurrent users
+- **Workflows that modify global state** (if unavoidable)
+
 ## 🛠️ Usage
 
 ### 🖥️ Command Line Interface
@@ -234,7 +283,6 @@ Both modes support automatic file discovery and upload:
 # - Text files (.txt) 
 # - JSON files (.json)
 # - Python files (.py)
-# - Excel files (.xlsx, .xls) - handled by pandas
 
 agent = ReActDataScienceAgent(
     executor="tci",  # or "internal"