A hands-on guide to using OpenCode (a coding agent), llama.cpp, and Qwen3.6 models to automatically find bugs in your own codebase. Covers building llama.cpp with ROCm/AMD GPU support, downloading and quantizing models, configuring llama-server, setting up OpenCode with a dedicated isolated account, and crafting effective prompts for targeted bug-finding sessions. Includes important safety warnings about running LLM agents without proper OS-level sandboxing, practical tips for improving efficiency (handling false positives, typo fixes, session continuation), and honest limitations about false positive rates and responsible use of AI-generated bug reports.
Table of contents
BackgroundSafety firstArchitectureBuilding Llama.cppDownloading a modelStarting llama-serverSetting up OpenCodeSetting up the projectStart looking for suspicious patternsImproving efficiencyVariationsLimitations90 Impressions