mthines Routing
Choose the path based on the argument shape:
| Argument shape | Path | Reason |
|---|---|---|
no arg / --report |
Skill("review-loop", "<pr-url> [--critical|--no-ci|--external-review|--interval S]") on the current branch's open PR, or pr-reviewer directly if --report (one-shot, no apply) |
Own PR convergence loop, or one-shot report. |
PR URL or #<n> |
Skill("review-loop", "<pr-url> [--critical|--no-ci|--external-review|--interval S]") — runs the bounded review-apply-simplify loop |
Converges the PR regardless of who authored it (pr-reviewer detects REVIEW_RELATION itself). |
--no-ci, --external-review and --interval S are pass-throughs: this skill never
sets them itself, it only forwards what the user passed (see the note below).
# --report routes straight to pr-reviewer, which --external-review exists to avoid.
# Refuse the pair rather than accepting a flag this path cannot honour, mirroring
# review-loop's refusal of --no-feedback --external-review.
case "$ARGUMENTS" in
*--report*) case "$ARGUMENTS" in
*--external-review*)
echo "--report needs pr-reviewer; drop --external-review or drop --report."
exit 1 ;;
esac ;;
esac
# Resolve the current branch's PR whenever no PR was named. Test for the ABSENCE of a
# PR reference, never for an exact flag string: every flag in the table above is
# combinable, so `= "--report"` left `--external-review`, `--critical`, `--no-ci`, and
# `--interval S` with no PR to act on.
case "$ARGUMENTS" in
*github.com/*/pull/*|*\#[0-9]*) ;; # a PR was named; use it
*) CURRENT_PR=$(gh pr view --json url -q .url 2>/dev/null) ;;
esac
review-changes passes no --no-ci — unlike create-pr and autonomous-workflow
Phase 7, it owns no CI phase of its own, so the loop's CI sub-step is exactly what
keeps a standalone run from converging on a red build instead of a
green-threads-red-build one. Forward --no-ci, --external-review, and --interval
through when the user passes them.
It converges on not red, not on green. review-loop's ci_is_settled() treats
pending as settled — the loop reads check state, never waits for it — so a run can
converge with checks still in flight. Wait for the checks yourself before merging, or
give CI to a caller that owns a watch (create-pr Step 9, phase-7-ci-gate.md) and
pass --no-ci here.
# Default — convergence loop on own or specified PR
Skill("review-loop", "<pr-url-or-number> [--critical|--no-ci|--external-review|--interval S if passed]")
# Report-only (no apply) — one-shot review.
# pr-reviewer is an AGENT — dispatch via the Task tool, NOT Skill():
Task(subagent_type="pr-reviewer", prompt="<pr-url-or-number> [--critical if passed]")
Usage
| Invocation | Effect |
|---|---|
/review-changes |
Convergence loop on the current branch's open PR — pr-reviewer → implement-suggestion --resolve-all → polish simplify → CI, up to 5 iterations, converging until every review thread is resolved (fix or reply) and CI is not red. |
/review-changes --no-ci |
Same, minus the CI sub-step. Pass this when something else already owns CI for the PR. |
/review-changes --external-review |
Waits for an out-of-process reviewer (another agent, a review bot) instead of dispatching pr-reviewer, then applies + resolves + simplifies as usual. Also the path to use where no available tool dispatches a sub-agent (Task, Agent, or another spelling). |
/review-changes --report |
One-shot read-only review via pr-reviewer (no apply). |
/review-changes --report --external-review |
Refused. Print --report needs pr-reviewer; drop --external-review or drop --report. and exit — the same refusal review-loop gives for --no-feedback --external-review, and for the same reason: report-only needs a reviewer to report, and --external-review removes the only one either path owns. Never silently ignore --external-review here. |
/review-changes --critical |
Adds adversarial pre-mortem (Skill("critical", "code")) to each pr-reviewer call. |
/review-changes <PR-URL> |
Convergence loop on the specified PR (self or cross — pr-reviewer detects relation automatically). |
What replaced --comments
The old --comments flag is gone.
Cross-review with line-level inline comments now lives in the pr-reviewer agent and is the default behaviour when a PR is passed.
There is no posting-authorization flag any more. pr-reviewer Step 4 posts a single
visible COMMENT review to POST /repos/{owner}/{repo}/pulls/{n}/reviews
unconditionally, in both relations — the old --publish token and the pending-review
workflow it gated are gone. Step 3 still prints the full proposal to the terminal
before Step 4 posts it, so you always see what was sent.