/delegate-review

Get a review for the user's request: $ARGUMENTS when the user typed the command, otherwise the request that brought you here.

1. The PR: the link or number given. Otherwise the current branch's PR (gh pr view --json url -q .url). If there is none, ask the user which PR.
2. The reviewers: every CLI named, in any order (vendor words as in the description), as a comma list (claude,copilot), or all for all three. With none named, auto: mesh picks by the size (step 3), what is installed and the quotas, from review.reviewers.<size> (mesh routing shows them), and says on stderr what it picked and why. A PR gets at most review.rounds.<size> rounds, a ceiling, not a quota: a round with nothing you agree with to fix ends its reviews, and after the last one's fix check the fix commits yourself (git diff <reviewed head>..<new head>) rather than asking for a re-review, unless the user asks for one. A re-review after a fix (the user asks you to look again at a PR these reviewers already reviewed, or a fix before a PR's last round) is the same command again: auto asks the same reviewers again by review.re_review (by default without Copilot, whose credits are kept for first looks at the PRs that matter).
3. The size: small for a PR of up to a few hundred lines in a few files (gh pr view <PR> --json additions,deletions,changedFiles), large for anything bigger or riskier (auth, money, data migrations). Every reviewer gets the same size. The size picks the plan phase and the level of a Claude reviewer's /code-review (review.level.<size>); Claude and agy reviewers run at review.effort either way, each in a pane beside you.

4. Run them from the local checkout of the PR's repo (the current one, or ask the user where it is), all in one command, a review round:

   mesh ask --repo <checkout> --review <PR> --size <size> --reviewers <auto | claude,agy,copilot (the named ones) | all>

   • It starts a run if there is none, runs every reviewer at once, waits until each has answered or alerted, prints one line per reviewer and the path of one answers file (answers: …), and closes the reviewers that answered. Read that file, not their inboxes.
   • You are Claude Code: run it with run_in_background; its notification is the whole round. You are agy: run it in the foreground.
   • Copilot runs headless inside the round; while it works, a pane beside you shows what it's doing (mesh peek --follow) and closes when it's done.
   • A reviewer that didn't answer (its line says why: hung, stopped, held at a startup screen, out of quota) stays open: follow its line, and its later answer comes to your inbox. mesh retire <name> it once you have no follow-up. Any other exit code is mesh watch's (see "Questions and reviews" in ~/.config/agent-mesh/ORCHESTRATE.md, the only section you need).
5. Check every finding before you report it: it's a claim, not an order. Trace it in the code or run the failing case, and judge the suggested fix. Keep each finding's blocking or advisory mark (a blocking one you can't reproduce is advisory at most). With several reviewers, merge the same issue found by more than one into one finding and say who found it; agreement makes a finding likelier, not true, so check it all the same. Then give the user each finding with your verdict (agree, agree with a different fix, or disagree and why), and the reviewers' names: mesh ask <name> --body "…" asks one again with everything it read. Fix nothing unless the user asks; a fix they ask for goes on the PR's own branch, not in a new PR.