AI-Dev-Tracker

AI-Assisted Software Development Traceability Tool

---

📌 Project Overview

AI-Dev-Tracker is a CLI-based tool designed to track, analyze, and measure the impact of AI-assisted development in software projects.

As developers increasingly rely on AI tools (e.g., ChatGPT, Claude), there is no systematic way to:

• Track prompt history
• Link AI responses to code changes
• Measure AI contribution to final software
• Identify areas where developers struggled
• Maintain structured, goal-driven development sessions

This tool provides a structured solution to record, analyze, and report AI-assisted development activity.

---

🎯 Problem Statement

Modern software development increasingly integrates AI-generated code. However, there is no standardized mechanism to:

• Maintain prompt history
• Link AI interactions to version control
• Identify which parts of code were AI-generated
• Analyze developer struggle through AI usage patterns
• Ensure session coherence when developers ask unrelated questions mid-session

AI-Dev-Tracker addresses this by creating a Git-linked AI interaction logging and analysis system with structured session flow enforcement.

---

🚀 Key Features

• Multi-LLM support — works with any OpenAI-compatible provider (Gemini, OpenAI, Groq, Mistral, etc.)
• Interactive model switching — select a provider at startup or switch anytime via CLI
• Structured session flow — goal-driven sessions with relevance enforcement
  • Session goals — attach a purpose/topic to each session
  • Soft warning — warns when a prompt appears off-topic, asks to confirm
  • Guard mode — hard-blocks off-topic prompts entirely (opt-in)
  • AI system prompt scoping — the AI is instructed to stay on-topic and redirect unrelated questions
  • Session summary — view a structured overview of the current session at any time
• CLI-based prompt logging
• AI response tracking with rich markdown rendering in the terminal
• Session-based project management
• Git commit hash integration
• File-wise prompt mapping
• Multi-source relevance detection (keyword overlap against first prompt, session goal, and recent conversation)
• AI contribution reporting
• Struggle detection (rapid-prompt, sustained, escalating, long-session)

---

⚙️ Setup & Installation

Prerequisites

• Python 3.8 or higher
• Git (for commit hash integration)
• An API key from any OpenAI-compatible provider (Gemini, OpenAI, Groq, etc.)

1. Clone the Repository

git clone https://github.com/your-repo/AIDevTracker.git
cd AIDevTracker

2. Install as a Package (Recommended)

Installing via setup.py registers the aidt command globally so you can run it from anywhere:

pip install -e .

This installs all dependencies (openai, python-dotenv, rich) automatically and creates the aidt CLI entry point.

Verify the installation:

aidt

You should see the usage/help output.

3. Configure Your LLM Provider

Option A — Interactive (on first run):

Run any command and you'll be prompted to select your provider:

Select your LLM provider:
  1. Gemini
  2. OpenAI
  3. Groq
  4. Other (custom base URL)

Option B — Manual .env file:

Create a .env file in the project root (copy from .env.example):

LLM_API_KEY=your_api_key_here
LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/
LLM_MODEL=gemini-2.5-flash
RELEVANCE_THRESHOLD=0.4

Variable	Description	Example
LLM_API_KEY	Your provider's API key	sk-...
LLM_BASE_URL	OpenAI-compatible base URL	See provider docs
LLM_MODEL	Model name to use	gemini-2.5-flash
RELEVANCE_THRESHOLD	Off-topic detection cutoff (0.0–1.0)	0.4

Common provider URLs:

• Gemini: https://generativelanguage.googleapis.com/v1beta/openai/
• OpenAI: https://api.openai.com/v1/
• Groq: https://api.groq.com/openai/v1/

---

🔧 Running the Application

All commands use the aidt entry point after installation. If you have not installed via pip install -e ., you can substitute aidt with python -m aidevtracker.main.

Core Commands

aidt ask "your prompt here" [file.py]       # Ask the AI (with optional file context)
aidt analyze [file.py]                       # Run analysis (repo-wide or per-file)
aidt report                                  # Generate detailed AI contribution report
aidt visualize                               # Show visual summary in terminal
aidt export [output.csv]                     # Export all interactions to CSV

Session Management

aidt session new "Project Name" [--goal "Session goal"]   # Create a new session
aidt session list                                          # List all sessions
aidt session use <id>                                      # Switch active session
aidt session guard <on|off>                                # Toggle hard-block guard mode
aidt session summary                                       # Show active session overview

Model Configuration

aidt model                                                 # Switch LLM provider (interactive)
aidt model <API_KEY> --base-url <URL> --model <MODEL>      # Switch LLM provider (direct)

Global Flags

aidt --threshold <0.0-1.0> <command>                       # Override relevance threshold for one command

Quick Start Example

# 1. Install the package
pip install -e .

# 2. Create a new session with a goal
aidt session new "My Web App" --goal "Building the authentication module"

# 3. Ask the AI a question scoped to a file
aidt ask "How should I structure JWT token validation?" auth.py

# 4. Run an analysis on your repo
aidt analyze

# 5. Generate a contribution report
aidt report

---

🛡️ Structured Session Flow

Sessions can be created with a goal that defines the session's purpose. This enables:

1. AI System Prompt Scoping — the AI is primed to stay focused on the session goal and redirect off-topic questions.
2. Pre-flight Relevance Check — before each AI call, the prompt is compared against three context sources using keyword overlap:
   • First prompt for the file (topical anchor)
   • Session goal (overall scope)
   • Most recent prompt + response (conversational continuity)
   • The max of all scores is used — if relevant to any source, the prompt passes.
3. Soft Warning (default) — if the prompt scores below the threshold, the developer is warned and asked Proceed anyway? [y/N].
4. Guard Mode (opt-in) — when enabled via aidt session guard on, off-topic prompts are hard-blocked with no AI call made.

Developer: aidt ask "..." file.py
              │
              ▼
     Session active AND file has a prior prompt?
              │
        Yes   ├── max(keyword_overlap scores) < threshold?
              │         │
              │    Yes   ├── guard_mode ON  → 🚫 HARD BLOCK
              │          └── guard_mode OFF → ⚠  Soft warn → y/N
              │    No    └── Proceed
              ▼
     AI responds (system prompt scoped to session goal)
              │
              ▼
     Relevance scored → Interaction saved to DB

---

🏗️ Project Structure

AIDevTracker/
├── aidevtracker/            # Main Python package
│   ├── __init__.py
│   ├── main.py              # Entry point (CLI argument parsing)
│   ├── cli.py               # Command implementations
│   ├── ai_client.py         # LLM provider abstraction
│   ├── analyzer.py          # Relevance & contribution analysis
│   ├── db.py                # SQLite database layer
│   ├── env_utils.py         # .env loading & model configuration
│   ├── git_utils.py         # Git commit hash integration
│   └── visualizer.py        # Terminal visualizations
├── setup.py                 # Package install config (creates `aidt` command)
├── .env.example             # Environment variable template
├── metadata.md              # Full metadata field reference
└── README.md

---

🏗️ System Architecture

---

🔄 Workflow Chart

---

📊 AI Contribution Evaluation Methodology

AI contribution is measured using the following metrics:

1. Prompt Count per File

   • Number of AI prompts linked to a specific file.

2. Git Commit Linkage

   • Each prompt is associated with a commit hash.

3. Hybrid Relevance Detection

   • Two methods are used and the higher score is taken:
     • SequenceMatcher — catches verbatim text/code reuse
     • Keyword Overlap — catches topical relevance (shared terms)
   • Default threshold = 0.4 (configurable via .env or --threshold flag)

4. Relevance Classification

   • If score ≥ threshold → Marked as AI-contributed.
   • Otherwise → Marked as non-contributing interaction.

5. Struggle Detection

   • Rapid-prompt struggle (< 5 min between prompts)
   • Sustained struggle (3+ prompts within a 30-min window)
   • Escalating dependency (prompt frequency accelerating over time)
   • Long session (continuous AI usage over 2+ hours on a file)

These metrics help identify:

• AI-dependent modules
• Developer struggle areas
• AI usage intensity across project timeline

---

📋 Metadata Captured

Per-interaction: prompt, response, file_path, commit_hash, timestamp, prompt_length, response_length, model_used, response_time, relevance, session_id

Per-session: project_name, created_at, goal, guard_mode

See metadata.md for the full reference.

---

⚠️ Limitations

• Uses text-based similarity (not semantic embeddings)
• Mock AI engine used for offline testing
• Similarity threshold may require tuning
• No deep architectural inference implemented

---

🔮 Future Enhancements

• Embedding-based similarity scoring
• Web dashboard visualization
• Developer struggle heatmap
• Architectural drift detection
• Export reports to PDF

---

👥 Team Members

• Shivam – System Architecture & Core Implementation
• Suhail – API Testing & Database Design
• Sreeja – Documentation,analyser.py, System Architecture & workflow chart
• Anil – Testing & CLI Validation
• Harsith – Research & Architectural Drift Study

---

📚 Academic Purpose

This project is developed as part of a Software Engineering Lab to study:

• AI-assisted development traceability
• Version control integration
• AI contribution quantification
• Software maintainability analysis

---