OpenCode configuration, commands, skills, plugins, and model conventions

Treat https://opencode.ai/config.json as authoritative and preserve $schema plus unrelated fields. OpenCode discovers project config upward to the worktree root; plugins auto-load only from immediate entries under discovered .opencode/plugins/ or global config plugins, not a root-level sibling plugins/. Prefer conventional separate agent, command, skill, and plugin files. Commands use the Markdown body as their template, with $ARGUMENTS or positional variables. Skills require a matching lowercase-hyphenated folder/name and a description. Use provider-prefixed model references. Strict shapes matter: skills, references, agent, and command are keyed objects; plugin is an array; MCP commands are string arrays and require type. SessionPrompt.prompt() tools overwrite session permissions, so configure tools through the agent. For opencode-pty, blanket bash permission avoids the ask→deny spawn trap; restart after configuration changes.