Claude Code /goal Command Ships: Set Completion Conditions, Walk Away — Haiku Evaluator Loops Until Tests Pass, Runs Across Multiple Turns for Hours or Days

Claude Code Docs·high signal

Claude Code v2.1.139 (May 12) introduced /goal, an outcome-based autonomous mode where developers define a completion condition (e.g., 'all tests in test/auth pass and lint is clean') and Claude works across multiple turns until met. After each step, an evaluator model (Haiku by default) checks whether the condition is satisfied; if not, the agent keeps running. The system tracks elapsed time, turns, and tokens. This separates the working agent from the evaluation agent — a meaningful architectural pattern for long-running autonomous work. VentureBeat called it 'the most underrated AI feature of 2026' for eliminating manual iteration in multi-step tasks.