OpenCode setup: subagents, skills and GitHub reviews

The speaker walks through an OpenCode coding assistant setup for project work and pull request reviews. Configuration can be global or specific to a project, with environment variable overrides. The walkthrough skips installation and focuses on daily use, including model selection through Zen and resuming sessions stored in SQLite.

Agent permissions shape the workflow. The speaker describes Plan as restricted to reading and planning, while Build can execute tasks. Custom subagents handle code review and security work. A larger experimental team adds a lead who delegates requirements, implementation and testing. The speaker finds this structure useful for large features, but reports mistakes, higher token use and extra effort tracking separate threads. Written instructions alone do not enforce boundaries; permissions provide the actual restrictions.

Skills supply instructions and supporting scripts on demand. The tutorial covers installing them globally or per project, then editing prompts in Neovim through a key binding.

For remote access, the speaker exposes OpenCode's local web port through ngrok and continues sessions on an iPad. GitHub comments can trigger reviews through an installed action. The walkthrough also covers a native GUI described as beta and image input in the terminal. These examples use hosted model services; the video does not establish an offline inference setup.