Patchwork

A local, self-healing code-auditing agent. It reads a Python file, patches bugs, writes tests, and runs them in an isolated sandbox — using a small quantized LLM (Qwen 2.5 Coder 3B) that fits on a 4GB laptop GPU.

Status

Day 1 of 5 complete: the deterministic tools the agent will call are built and tested. The LangGraph reflection loop that connects them to the model doesn't exist yet.

What works right now

Three standalone Python tools, no LLM involved:

Tool	File	What it does
Sandbox runner	patchwork/tools/sandbox.py	Runs a pytest suite against source code in a disposable temp directory, with a hard timeout so a hung or infinite-looping test can't freeze the process
Linter wrapper	patchwork/tools/linter.py	Runs ruff check and returns structured, truncated diagnostics
AST inspector	patchwork/tools/ast_inspector.py	Parses source into an AST, extracts functions/classes/docstring coverage, and reports syntax errors without crashing

Each tool returns a typed Pydantic model, never raises on expected failure modes (timeouts, bad syntax, missing binaries), and has its own test file mixing mocked unit tests with real subprocess/parse integration tests.

Requirements

• Python 3.10+
• Ollama, running locally
• qwen2.5-coder:3b pulled via ollama pull qwen2.5-coder:3b
• An NVIDIA GPU with 4GB+ VRAM (developed against an RTX 3050 Mobile)

Setup

python -m venv .venv
source .venv/bin/activate  # .venv\Scripts\activate on Windows
pip install -r requirements.txt

Confirm Ollama can serve the model:

ollama run qwen2.5-coder:3b

Running the tests

pytest tests/ -v
ruff check patchwork/ tests/

54 tests currently pass across the three tools.

Project structure

patchwork/
├── patchwork/
│   ├── cli.py              # not yet implemented
│   ├── graph.py             # not yet implemented — LangGraph state machine
│   ├── state.py              # not yet implemented — shared schemas
│   ├── tools/
│   │   ├── ast_inspector.py
│   │   ├── linter.py
│   │   └── sandbox.py
│   └── telemetry/            # not yet implemented
├── benchmarks/                # not yet populated
└── tests/
    ├── test_ast_inspector.py
    ├── test_linter.py
    └── test_sandbox.py

Roadmap

• Day 1 — deterministic tools (sandbox, linter, AST inspector)
• Day 2 — Pydantic schemas, LangGraph node wiring, single-pass audit
• Day 3 — reflection loop with retry/reflect cycle
• Day 4 — VRAM/throughput profiling, 25-defect benchmark dataset
• Day 5 — benchmark results, comparison charts, final documentation

Design notes

Full architecture, hardware constraints, and per-tool design decisions are documented as the project progresses — see commit history and PR descriptions for the reasoning behind specific choices (timeout values, memory limits, truncation strategy).