iterate-untillisted

# Iterate Until ## Purpose Formalize the "loop work-then-verify until done" pattern. A specialist proposes a fix, a verifier checks it, the lead decides whether to accept, retry, or escalate. The loop has a hard iteration cap; the verifier is always human-checkable; every iteration is logged. Adapted from oh-my-openagent's "Ralph Loop" pattern. yakOS-flavored: the verifier is **never** the agent's own judgement that work is complete. It's a test command, a hook exit code, a `yakos validate` result, a file-presence check, or a human-readable check the lead can audit after the fact. This is a soft control. The hard control is the iteration cap — the loop refuses to continue past N iterations and forces escalation to the human. ## Scope The pattern handles tasks of the shape: > Apply a fix; if the verifier passes, done. If the verifier fails, > read the verifier's output, propose a refined fix, retry. Cap at N. It does NOT handle: - **Open-ended exploration.** "Make the app faster" without a concrete perf target isn't a job for iterate-until — it's a job for the planner to decompose into iterate-until-able tasks. - **Multi-domain rework.** If the verifier failure surfaces a problem in a different specialist's domain, escalate; don't loop on it. - **Verifier-not-yet-defined tasks.** If you can't write the verifier command before you start, you don't have a clear enough target. Push back on the ask. ## When to use - A flaky test the test-runner says is real-no