pr-looplisted

# PR Loop Drive one or more same-repository Issues into a reviewed pull request, or drive an existing pull request through review and fix rounds until no actionable feedback remains. The top-level agent owns every repository and GitHub mutation. Delegate only planning, review, and feedback analysis to fresh, independent, read-only subagents through the active runtime's native subagent mechanism. ## Core Invariants - Use real native subagents with fresh context. Do not emulate them in the parent context, launch nested coding-agent CLIs, or require fixed agent names, models, providers, or configuration files. - Treat each accepted subagent as a terminal leaf: it performs its assigned role directly and does not re-enter `pr-loop` or delegate again. - If a required independent subagent cannot be launched, report `unsupported` and stop. Retry only when the runtime proves rejection occurred before execution was accepted; never duplicate ambiguously accepted work. - Every accepted subagent dispatch must reach a terminal result within a finite runtime- or caller-enforced deadline. A still-running poll is not failure; on terminal failure or deadline expiry, stop or cancel and reap the accepted dispatch without launching replacement work. - Bind every review and feedback decision to an exact PR head SHA. If the head changes before an action based on that decision, discard the stale result and restart from the new head. - Treat subagent output as advisory. The top-level agent valida