long-horizon

Break the work into verified rounds where an executor agent performs each step and an independent auditor checks the result. Persist progress in state and workspace files so later rounds resume without relying on conversation history.

Store task state in files rather than chat history, then spawn fresh agents with standalone briefs pointing to those state and workspace paths. This lets work continue after compaction or failed retries without re-explaining the task.

By default executor and auditor rounds spawn agents with fork_turns set to none, receiving only a standalone brief plus state paths. A positive turn count is an exception only when required data cannot be recovered from files, and must be recorded in the audit log.

No. Skill invocation alone does not authorize commits, pushes, pull requests, merges, deploys, migrations, or dependency installation. Those outward actions require an explicit user request.

Avoid it for short single-step tasks that fit comfortably in one context window, since round-based execution and auditing add overhead. It is designed for work spanning hours, sessions, or many dependent steps.