review-loop
Iterative multi-model review-and-fix loop. Repeatedly runs the review-multi-model panel over the working tree, applies the verified fixes, and re-reviews - review → fix → review → fix - until a full pass finds zero verified issues (or a safety cap / no-progress guard stops it). Use when asked to "review-loop", "loop the review", "keep reviewing and fixing until clean", "review until no issues", "iterate review-multi-model until it's clean".
uiverify · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add uiverify/uiverify --skill review-loop技能说明
name: review-loop description: Iterative multi-model review-and-fix loop. Repeatedly runs the review-multi-model panel over the working tree, applies the verified fixes, and re-reviews - review → fix → review → fix - until a full pass finds zero verified issues (or a safety cap / no-progress guard stops it). Use when asked to "review-loop", "loop the review", "keep reviewing and fixing until clean", "review until no issues", "iterate review-multi-model until it's clean". argument-hint: '[pr-number-or-url] [--claude-only] [--codex-only] [--max-iterations N]'
review-loop - iterate review-multi-model until the diff is clean
Run the review-multi-model panel in a loop: review → apply
verified fixes → re-review → fix → … until a full multi-model pass returns zero
verified findings. This exists because a fix can introduce a new defect - the
review that clears one issue often surfaces another the previous pass couldn't see,
so a single review+fix is rarely a fixpoint. The loop drives the diff to an actual
fixpoint instead of stopping after one round.
CRITICAL: Do NOT enter plan mode. Execute directly, iterate, and present the per-iteration progress and a final convergence summary inline. Do not write results anywhere except review-multi-model's own scratch dir.
It loops over the WORKING TREE only
The loop's whole premise is that round N re-reviews exactly the code round N-1
just edited. review-multi-model's fix step (--fix) writes to the working tree, so
the loop only round-trips fixes when the thing being reviewed is also the working
tree - i.e. review-multi-model's branch mode (working tree vs. merge-base).
- Branch mode (default): the loop works as designed - each round diffs the working tree, so it sees the prior round's fixes.
- PR argument: the PR must already be checked out (ideally in a worktree) so
the loop can review-and-fix its working tree. Loop in branch mode against the
PR's own base branch - its merge target (e.g.
release/1.x), which may not be the repo default. Prefix review-multi-model's branch-mode block (in the same Bash call - shell vars don't persist between calls) with its base override set to that branch:REVIEW_BASE=$(gh pr view <n> --json baseRefName -q .baseRefName) || exit 1- abort if that lookup fails rather than letting the base silently fall back to the
repo default. Fetch that base fresh once before looping (
git fetch origin "$REVIEW_BASE") so a stale localorigin/<base>doesn't fold already-merged changes into the diff. Do not loop ongh pr diff: that reflects the pushed PR, never the local uncommitted fixes (which this loop does not commit), so no round could observe its own fixes.
- abort if that lookup fails rather than letting the base silently fall back to the
repo default. Fetch that base fresh once before looping (
--stagedis not supported and is not a pass-through flag. review-multi-model fixes the working tree without re-staging, sogit diff --stagedwould show the unchanged index every round and the loop would never converge. For a one-shot staged review, run/review-multi-model --staged(its--fixwrites to the working tree, sogit addthe result afterward if you want it back in the index).
If the user passes a PR that is not checked out, say so and stop before iteration 1
(gh pr checkout <n> first) - do not switch their branch/worktree yourself, and a
fix loop that cannot see its own fixes is pointless anyway. --claude-only /
--codex-only pass through unchanged; the fix step is always on (it's a fix loop).
What one iteration is
One iteration = one complete review-multi-model run with its fixes applied: its review
phase (compute the working-tree diff → dual reviewers → adjudicate → present)
followed by its fix step (--fix - apply the verified, agreed fixes). Follow
the review-multi-model skill verbatim each round - do not reimplement the
review; this skill only wraps it in a loop. Every round starts fresh (new diff,
new scratch dir, two fresh reviewers), so a regression a fix introduced shows up
next round.
The loop
Each iteration is a review followed - only if needed - by a fix:
- Review (review-multi-model's review phase). Record the verified-finding count, whether any consequential Unverified items were flagged, whether the full requested panel actually ran, and a fingerprint of each finding (see below).
- Decide from that review, before touching code:
- Clean (zero verified findings, no consequential Unverified, full panel ran) → converged, stop and report success.
- A no-progress or only-un-appliable stop condition fires → stop, report.
- Otherwise → go to step 3.
- Fix (review-multi-model's fix step): apply the verified fixes, then begin the next iteration - a fresh review that confirms them. A clean confirming review is the only thing that lets you call the diff clean; a fix on its own never does.
--max-iterations bounds the run when you pass it; it is unset by default. Convergence is
always decided by a clean review, never assumed after a fix. If you reach the cap
and the last review still had findings, you will have applied that round's fixes
without a further review to confirm them - report those as applied but
unconfirmed, list the findings open as of the last review, and suggest re-running
with a higher --max-iterations to confirm.
Finding fingerprint (for the no-progress guard)
Fingerprint a finding by the defect and where it lives - the file plus the specific construct/region it points at (or the thing a fix would change) - matched semantically. Two findings are "the same" if they describe the same defect in the same place, even if the two rounds word it differently ("missing null guard" vs. "nullable value dereferenced") or the line number drifted because edits above it shifted the file. Do not match on exact wording or line number - that lets one persistent defect masquerade as a new one and churn to the cap.