resolve-review
Work through a code-review report's findings and fix them in priority order, with confirmation on anything non-trivial. Trigger on:- resolve review, address findings, fix the review, clear blocking it
SKILL.mdskills/resolve-review/SKILL.md
GitHubResolve Review
Takes a code-review report and works through its findings, fixing them one at a time
in priority order. This skill remediates — it does not re-judge whether something is a
real issue. If a finding looks wrong on inspection, say so and skip it rather than
silently ignoring it or arguing it away.
Assumptions (flag if wrong)
- The report is an
open-statuscode-reviewoutput atspecs/NNN-<slug>/review.md, with findings grouped under Blocking / Should-fix / Nit headers, each with file:line, what's wrong, why it matters, and a suggested fix. - Fixing scope is limited to what's named in the finding — not an invitation to refactor around it.
Read .specify/memory/constitution.md first for the pipeline map and the review status
vocabulary.
Arguments
report— path to the review report to work from. If omitted, look for the most recentspecs/*/review.mdwithstatus: openand confirm it's the right one before proceeding.which— which severity tier(s) to address:all(default) — Blocking, then Should-fix, then Nit, in that order.blocking/should-fix/nit— just that tier.
Workflow
Read the report. Parse out findings per severity tier, each with its file:line and suggested fix.
Work the findings via the constitution's Item-execution loop, in tier order: Blocking, then Should-fix, then Nit (Nit only if
whichincludes it). Two review-specific points on top of the loop:- A finding that looks wrong on inspection — not a real bug, or the fix would break something else — is a skip with reason, not a silent pass. This skill remediates; it doesn't re-review the other findings.
- For Nit, ask once up front — apply all, or list for the user to pick — rather than prompting per item.
Set the report's
status:toworked-onand bumpupdated:. Neverclosed— onlycode-reviewcloses a thread, after re-checking the fixes.Offer to re-run
code-reviewon the same scope to confirm the findings are actually resolved, rather than trusting this pass's own fixes. Don't run it automatically — ask first, since it may not be needed for anit-only pass.
Boundaries
- Don't fix things the report didn't flag, even if you notice them along the way — note
them instead (a one-liner in the summary, or point at
/noticeif it looks like a recurring pattern). - Don't reduce a Blocking or Should-fix finding's severity to make it easier to skip.
- If more than a couple of findings turn out to need real design decisions rather than
straightforward fixes, stop and flag that the report may need
/designitrather than direct patching — don't improvise architecture mid-fix.