OpenCode

AO resolves your local opencode executable and reuses its authentication.

ao agent ls --refresh
ao spawn --project my-app --name opencode-task --agent opencode --mode chat

OpenCode can run as its native Terminal UI or through its native ACP interface in structured Chat. Chat history is persisted by AO as a durable, paged conversation projection while provider conversation identity is retained for continuity.

AO's approval choices map onto what OpenCode itself does, so a session behaves the way the same session would outside AO. Default approvals adds no rules and leaves your configuration to decide. Accept edits approves file changes, the one behaviour OpenCode has no mode for. Auto-approve is OpenCode's --auto: everything that is not explicitly denied runs, including the out-of-worktree and repeat-loop guards it otherwise asks about. Bypass permissions is permission: "allow", its own full-access setting, and the only mode that also passes explicit denies.

Your own policy stays authoritative in every mode but Bypass, wherever you wrote it. AO never rewrites your permission rules for Accept edits or Auto: in the Terminal UI, Auto is the provider's --auto flag, and in Chat, where there are no flags, AO answers each permission request the way that flag does. A tool you denied never raises a request, so it stays denied whether the rule lives in your global config, the repository, a .opencode directory, or settings your organisation manages.

The Chat picker switches between AO's agents, so a change applies immediately in either direction, including back to a stricter mode.

OpenCode is also a shipped reviewer harness. Reviewer configuration is independent from the worker harness; inspect project settings and ao review --help for the current review workflow.