dqz00116/opencode-solo

This project is a configuration package for opencode, a coding assistant tool, that sets up a team of specialized AI agents working together instead of relying on one general purpose agent to do everything. It adds a main orchestrator agent called Solo along with six supporting agents, each handling a different part of the coding workflow. Solo works in a closed loop: it directly runs tests using bash commands, reads the raw results, and decides what to do next, rather than just making changes and hoping they work. It repeats an edit, test, and decide cycle until the target tests pass, then stops immediately. To keep this efficient, most of the heavy work, like reading through large files or exploring an unfamiliar codebase, happens inside separate helper agents so that Solo's own working memory only holds short summaries and decisions rather than huge amounts of raw text. A faster, cheaper model handles routine exploring and editing tasks, while a more capable and expensive model is reserved for planning and for double checking risky changes, which the project's authors say keeps quality high while lowering overall cost. The seven agents included are solo, the orchestrator, explore, which maps the codebase once and finds the relevant tests and root cause, editor, which makes small focused code changes, verify, which does an adversarial double check only for large or risky changes, reviewer, for on demand code quality checks, observer, for looking at screenshots and diagrams, and general, a fallback agent for anything else. The project includes benchmark results comparing this setup against a single all in one agent, showing similar overall token usage but a better balance between cheap and expensive model usage. Setting it up means cloning the repository, copying the provided agent configuration files into opencode's configuration folder, and editing a settings file to choose which AI models power each agent. The project is released under the MIT license.