Agentic C / C++ Unit Tester

An agentic tool that analyzes C and C++ codebases, identifies test coverage gaps, and generates Google Test (gtest/gmock) code with an LLM-assisted workflow. It supports free functions, member methods, static methods, constructors, destructors, and private/protected members — with automatic boundary, equivalence partition, and MCDC test case planning.

---

🏗️ Design & Architecture

The tool follows a modular, multi-stage pipeline designed for precision and extensibility:

1. Analysis (libclang):
   • Extracts deep AST information: Enums, Structs, Unions, Classes, and Macros.
   • Identifies logic-heavy structures like Switch-Case blocks (with enum label mapping).
   • Captures the Source Body of each function to provide implementation context for the LLM.
2. Strategy Planning:
   • Unifies boundary analysis, equivalence partitioning, and MCDC coverage into a single "Master Strategy".
   • Analyzes existing unit tests to "reverse engineer" current coverage strategies and identify gaps.
3. LLM-Assisted Generation:
   • Uses a Context-Aware Prompting system: provides the LLM with the actual source code, type definitions, and macro context.
   • Balanced Brace Parser: A robust Python-based parser ensures that even complex, nested GTest bodies are extracted correctly from LLM responses.
   • Isolated Scoping: Each generated scenario is wrapped in its own scope { ... } to allow multiple verification attempts in one test without variable name collisions (e.g., redundant obj declarations).
4. Verification Pipeline:
   • Automatically generates CMakeLists.txt, compiles with coverage flags, runs tests, and generates GCOV coverage summaries.

---

🚀 Key Features

• Private & Protected Acccess: Uses a #define private public trick in templates to allow direct testing of internal class members.
• Switch-Case Awareness: Fully maps enum labels to switch cases for branch-perfect strategies.
• Coverage Integration: Direct integration with gcov for real-time line-coverage feedback.
• Provider Agnostic: Supports Ollama (Local), Gemini, OpenAI, and Azure.

---

📖 Usage

1. Full Build & Coverage (Recommended)

Analyze, generate strategy, generate code, compile, and run coverage in one command:

python3 main.py build <source_file> --coverage

2. Manual Workflow

• Analyze: python3 main.py analyze <source_file>
• Generate Test: python3 main.py generate <strategy_yaml>
• Clean Artifacts: python3 main.py clean

---

📜 Change History & Improvements

Phase 2: Robust Extraction & Context (Latest)

• Implemented Balanced Brace Parser: Resolved issues with truncated or malformed test bodies from LLMs.
• Implementation-Aware Prompts: Added function source code to the LLM prompt, resulting in a 90% reduction in compilation errors (invalid signatures/missing arguments).
• Scoped Scenario Merging: Allowed multiple LLM-generated scenarios to be merged into a single TEST_F block without "redeclaration" errors.
• GCOV Fixes: Resolved pathing issues with CMake-generated .gcda files to ensure accurate coverage reporting.

Phase 1: Core Architecture

• Initial libclang integration and modular LLM provider support.
• Basic Google Test template with private access bypass.
• Strategy generation for free functions and simple classes.

---

📊 Final Verification Results (complex_cases.cpp)

The pipeline was stress-tested against a complex C++ class featuring private state and switch logic:

• Test Generation Success: 100% (Zero compilation errors).
• Functional Pass Rate: 75% (3/4 tests passed).
• Line Coverage: 41.67%.

---

🤖 LLM Configuration

Configure via .env:

LLM_PROVIDER=ollama
OLLAMA_MODEL=qwen2.5-coder:3b

Supported providers: ollama, gemini, openai, azure, mock.