ATTest

A controllable CLI agent for generating and running tests for Python operators and APIs (such as PyTorch and TensorFlow) through a fixed, human-approved workflow.

• 🔧 Tool Mode (Chat): Ad hoc queries and file operations
• 🔄 Workflow Mode: A 7-stage intelligent test generation pipeline

• 7-stage pipeline: Function understanding → requirements generation → test plan design → code generation → test execution → result analysis → report generation
• Smart feedback: Supports natural-language feedback and special commands (/regenerate, /goto, /retry)
• Persistent state: Supports interruption recovery, with versioned artifact management
• Custom execution: Configurable pytest commands, environment variables, and dependency installation

cd ATTest
pip install -e .

attest chat --workspace ~/my-project

# Interactive mode (recommended)
attest run -f torch.nn.functional.relu --workspace ./my-project/torch.nn.functional.relu

# Fully automatic mode
attest run -f torch.add --mode full-auto --workspace ./my-project/torch.add

# Fully automatic mode with multi-round iteration (3 rounds)
attest run -f torch.add --mode full-auto --epoch 3

# Specify the project root when it differs from workspace
attest run -f torch.add --workspace ./my-project --project-root ./my-project/src

# Resume an interrupted workflow
attest run -f torch.add --resume

• WORKFLOW_GUIDE.md - Complete usage guide (recommended)

  • Detailed workflow mode walkthrough
  • Custom build command configuration
  • Testing and debugging
  • Extension and customization

• QUICK_REFERENCE.md - Quick reference card

  • Common command cheat sheet
  • Workflow interactive commands
  • Configuration examples

• USAGE.md - Chat mode usage guide

╔═══════════════════════════════════════════════════╗
║   ATTest Workflow - torch.add (python)           ║
╠═══════════════════════════════════════════════════╣
║ [●] 1. Understand Function                       ║
║ [●] 2. Generate Requirements                     ║
║ [▶] 3. Design Test Plan        ← Current Stage   ║
║ [ ] 4. Generate Code                             ║
║ [ ] 5. Execute Tests                             ║
║ [ ] 6. Analyze Results                           ║
║ [ ] 7. Generate Report                           ║
╚═══════════════════════════════════════════════════╝

After each stage completes, you can:

• Press Enter to continue
• Use /regenerate to regenerate the current stage
• Use /goto <stage> to jump to another stage
• Provide natural-language feedback for smart interpretation

When you need to adjust the pytest command or environment variables, edit ~/.attest_cli/config.json:

{
  "api": {
    "model": "deepseek-chat",
    "base_url": "https://api.deepseek.com/v1",
    "api_key": "your-api-key",
    "temperature": 0.2,
    "max_tokens": 4096
  },
  "preferences": {
    "auto_approve": false
  },
  "project": {
    "root": ".",
    "test_file_template": "tests/test_{target_slug}.py",
    "build_dir": "",
    "output_binary_template": ""
  },
  "commands": {
    "compile": "",
    "install": "",
    "run_test": "PYTHONPATH={project_root}:$PYTHONPATH pytest -q {test_file_path}"
  }
}

Available variables:

• {target} / {target_slug} - Target function FQN and its slug
• {project_root} - Project root directory
• {test_file_path} - Generated pytest file path

See WORKFLOW_GUIDE.md - Custom Build Commands for details.

attest config list              # Show all configuration values
attest config set KEY VALUE     # Set a configuration value
attest config get KEY           # Get a configuration value

attest sessions list            # List all sessions
attest sessions clear ID        # Clear a specific session

attest chat [--workspace DIR] [--auto-approve]

attest run -f package.module:function \
  [--workspace DIR] \
  [--project-root DIR] \
  [--mode interactive|full-auto] \
  [--epoch N] \
  [--resume]

Arguments:

• -f, --func: Fully qualified target function name (required)
• --workspace: Working directory, defaults to the current directory
• --project-root: Project root directory, defaults to the same value as workspace
• --mode: Run mode, either interactive or full-auto, defaults to interactive
• --epoch: Number of iteration rounds in full-auto mode; after analyze_results, the workflow loops back to generate_code; defaults to 1
• --resume: Resume an interrupted workflow

1. Create a tool class in src/attest_cli/tools/builtin.py
2. Register it in src/attest_cli/tools/runner.py
3. Add it to a stage's tools list

Edit the corresponding stage file, for example src/attest_cli/workflow/stages/requirements.py, and update _get_prompt_template().

1. Create src/attest_cli/workflow/stages/your_stage.py
2. Register it in stages/__init__.py
3. Add it to STAGE_NAMES in workflow/engine.py

See WORKFLOW_GUIDE.md - Extension and Customization for details.

ATTest/
├── src/attest_cli/
│   ├── cli.py                    # CLI entry point
│   ├── config.py                 # Configuration management
│   ├── llm.py                    # LLM client
│   ├── session.py                # Session management
│   ├── chat.py                   # Chat mode
│   ├── tools/                    # Tool system
│   │   ├── base.py
│   │   ├── builtin.py
│   │   └── runner.py
│   └── workflow/                 # Workflow engine
│       ├── engine.py             # Main controller
│       ├── state.py              # State management
│       ├── stage.py              # Stage base class
│       ├── supervisor.py         # Feedback interpretation
│       ├── display.py            # Progress display
│       └── stages/               # Individual stages
│           ├── understand.py
│           ├── requirements.py
│           ├── planning.py
│           ├── codegen.py
│           ├── execution.py
│           ├── analysis.py
│           └── report.py
├── WORKFLOW_GUIDE.md             # Complete usage guide
├── QUICK_REFERENCE.md            # Quick reference
└── USAGE.md                      # Chat mode guide

# Test the core framework
python test_milestone1.py

# Test custom commands
python test_custom_commands.py

attest run -f torch.add --workspace ~/my-project --mode full-auto

Automatically generates pytest cases, executes them, and produces a test report.

attest config set commands.run_test "PYTHONPATH={project_root}:$PYTHONPATH pytest -q {test_file_path} -k gpu"
attest run -f torch.add --workspace ~/my-project --mode full-auto

Uses a custom command to run the generated tests, which can be reused in a virtual environment or CI pipeline.

attest chat --workspace ~/my-project

Temporarily inspect or analyze files, or generate small testing snippets.

Contributions are welcome. Main extension points include:

• New tools (file operations, analysis utilities, and so on)
• New workflow stages (performance testing, coverage analysis, and so on)
• Improved prompt templates
• Better error handling

MIT

Inspired by:

• CliAgent - Flow-first design
• Cougar-CLI - Config and session management
• mini-kode - Permission-aware tool execution

Quick start: See WORKFLOW_GUIDE.md for detailed usage.