pua-loop

PUA Loop is an autonomous iterative-development mode that borrows a gated "Oracle" verification design from the autoresearch project. It solves the problem of an AI agent falsely declaring a task done: the agent may output a `<promise>LOOP_DONE</promise>` completion signal, but a Stop hook independently runs a user-set `verify_command` (embedded in a state-file frontmatter the agent cannot modify) and only accepts completion if that command exits zero, otherwise feeding the failure output back and continuing the loop.

Its main features are five design patterns: Oracle isolation (a verify command the agent can't edit), a two-phase gate (agent self-check then independent hook check), failure memory appended to a `.claude/pua-loop-history.jsonl` file so past failed approaches aren't repeated, stall detection that escalates behavior after repeated promise rejections, and unbounded iteration by default. A setup script (`setup-pua-loop.sh`) starts the loop, and the skill will auto-append a `--verify` command (for example `npm test` or a curl health check) when it can infer one from the task. The mode deliberately disables AskUserQuestion and forbids the agent from saying it cannot solve the problem, driving toward the completion signal.

Target users are developers who want a hands-off "run until verified" loop for tasks with a checkable success condition (fix all tests, build an API to a passing health check). The main limitation and caveat is exactly that autonomy: it runs shell verify commands and setup scripts without pausing for the user and, by default, never stops on its own until the Oracle passes or an abort/pause marker or Ctrl+C intervenes.