Long Run Harness

Use this skill when the user wants the agent to keep advancing a complex task until it is actually done, not merely improved.

Default to git-repository work. You may also apply the same operating model to documentation, research, or mixed projects when durable state and autonomous continuation matter.

Do not use this skill for one-shot edits, lightweight planning, or requests where the user wants synchronous oversight at each milestone.

Core Operating Rules

• Externalize state. Do not rely on conversation memory for long-running work.
• Work one task at a time. Do not keep multiple partial tasks open in parallel.
• Require validation before task completion.
• Require review before advancing to the next task.
• Verify state before trusting it and after each significant state transition.
• Use git history as a recovery aid when a repository is available.
• Continue automatically unless a real blocker or completion condition exists.

Required State Files