tp-pr-iterate

tp-pr-iterate — PR-Iteration Loop Driver

A long-running loop over a single open PR. Each iteration polls for new review comments, classifies them, defers conflicts to a human, calls tp-pr-fix.run_round for the structural subset, and re-enters the wait. Terminates on:

• Two-stable (the one success terminal) — in a single round both review sources are quiet (/code-review returns [] AND no new structural Copilot verdicts) AND every actionable thread has been replied-and-resolved, verified against GitHub: a fresh list_review_threads(pr_url) shows zero unresolved Copilot / code-review threads. Only then is the PR human-ready. A classifier-flip (no structural verdicts this round) is a necessary precondition, never sufficient on its own — flipping on a stale snapshot while threads sit unresolved is the exact failure that makes a PR look "stable" when it isn't.
• Idle timeout (yield to human) — no new comments for 30 minutes AND the prior round was not structural-present. This is a time-based yield, not a verified-stable claim: _poll_step does NOT run the ground-truth zero-unresolved re-fetch (only two-stable does), so the terminal carries the [idle-timeout] transition note and leaves termination_reason unset — consumers must treat only termination_reason="two-stable" as reviewed-stable.
• Mid-loop human push (yield to human) — a non-[tp-pr-fix iter- commit appeared on the branch since last_loop_sha.
• Caps — iteration count exceeded (max_iterations, default 8) or wall-clock budget exhausted (max_wall_clock, default 4h).
• Convergence failure — cumulative diff has grown past 3× the loop-open baseline, OR three consecutive rounds returned structural-present.
• All-conflicting deferral — every comment in a round had an overlapping line_range with another, so the loop couldn't proceed.

Arguments

• {design} — kebab-case design name, validated per skills/_shared/validate-name.md. Resolves the worktree on branch tp/{design} and the design directory.
• --max-iterations N — round cap (default 8). Triggers cap-exhausted.
• --max-wall-clock <duration> — total budget (default 4h). Triggers cap-exhausted.
• --dry-run — fetch + classify + log what the loop would do, but do NOT invoke tp-pr-fix.run_round. Use this to verify classifier output on a real PR before letting the loop write commits.
• --dispose-only — calls thread_dispose.dispose_threads once and exits; no iteration, no fix dispatch, no backoff/poll loop. Use this to reply-and-resolve all open review threads out-of-band (e.g. after a manual head-SHA bump or when threads arrive outside a loop run). This is a flag, NOT a new skill. See the --dispose-only mode section below.