Developer Builds Local LLM Agent to Auto-Generate and Run Solidity Fuzz Tests

A developer has shared a workflow that uses qwen2.5-coder, a 7-billion-parameter local language model run via Ollama, to autonomously write and iterate on fuzz tests for Solidity smart contracts in the Foundry framework. The setup gives the model exactly three tools: reading a contract file, writing a fuzz test, and running the test suite, keeping the toolset minimal to reduce model confusion. The model reads a target contract, identifies invariants, writes property tests, and iterates on failures in a loop without any data leaving the developer's machine. Key implementation choices include forcing the model to always write a single complete test file rather than diffs, and aggressively truncating forge output to prevent large traces from overwhelming the model's context window. While the approach has clear limitations with small models, the developer found it practically useful for surfacing broken invariants with minimal manual effort.