A portable correctness review. It reads the diff and reports; it never edits,
stages or commits. A caller that fixes findings (/qa-full Step 3) applies the
fixes itself and re-runs this review to verify them.
The bar for what counts as a blocker is the superskills quality standard,
ENGINEERING_STANDARDS.md in the superskills install (a plugin install exposes
it at ${CLAUDE_PLUGIN_ROOT}/ENGINEERING_STANDARDS.md). It is not expected in
the project under review.
Scope
- With an argument, review exactly that: a range such as
origin/main...HEAD(as/qa-fullpasses it), or paths. - Without one, "pending changes" means the current branch's diff against its
base plus anything uncommitted:
git diff <base>...HEADandgit diff HEAD. Find the base the way/qa-fullStep 1 does (gh repo viewdefault branch, elseorigin/HEAD, elsemain/master). - If there is nothing to review, say so and stop.
Read each changed hunk with enough surrounding code (callers, the function it sits in, the config it reads) to judge it. Verify claims against the actual code rather than inferring from names.
Review
Review the pending changes in this repo. Flag issues a careful maintainer would block on before merge, in roughly this order:
- Correctness: logic errors, broken invariants, missed edge cases, off-by-one, races, lifetime/memory bugs, error paths that swallow failures.
- Security: injection, authn/authz mistakes, secret handling, unsafe deserialization, unsafe defaults, TOCTOU.
- Reliability & performance: blocking the wrong thread, unbounded resource use, accidental O(N²), retries without backoff, silent failure paths.
- Contract risk: behavior changes callers rely on, silently changed defaults, broken backward compatibility, missing migrations.
Then apply checklist.md in this skill's directory (gstack's pre-landing
checklist): run its Pass 1 (CRITICAL) categories, then Pass 2, and honor its
Suppressions list. This skill is read-only, so ignore the checklist's
Fix-First / AUTO-FIXED instructions and its output format (use the Report format
below), and skip the items it assigns to gstack's specialist subagents or to
gstack-only markers.
For each finding cite file:line, explain why it is wrong (not what the code
does), and propose a concrete fix when one is obvious. Be calibrated: if a
finding is not high-confidence, say so or skip it.
What earns a finding
A candidate earns a finding only if every one of these holds:
- It meaningfully affects correctness, security, performance, or maintainability, and it is discrete and actionable — one defect, one fix.
- The diff introduced it. Problems that predate the change get at most a one-line aside labelled "pre-existing" — never a finding, never a blocker.
- The rigor demanded matches the file's neighborhood: a corner of one-off scripts does not need enterprise-grade validation.
- The author of this change would fix it if told; if they would reasonably argue, it is not a finding.
- No speculation: a harm of the form "may break other callers" needs the caller that is provably affected, named — or it is dropped.
- It is not an intentional behavior change dressed up as a bug.
If nothing clears this bar, "No findings worth fixing in " is the correct and complete review. An empty result beats an invented one.
Skip pure style, formatting, naming, and anything a linter or CI already enforces.
Report
One entry per finding, most severe first:
[CRITICAL|HIGH|MEDIUM|LOW] (confidence: high|medium) file:line
Why it is wrong: ...
Fix: ... (or "no obvious fix")
CRITICAL and HIGH are what a caller treats as blockers. End with one line naming the range reviewed and the number of findings by severity.
If you find nothing worth fixing, say so plainly: "No findings worth fixing in ."
Related
/review(gstack): the full review. Prefer it when installed; it goes deeper and applies fixes itself./clean-code: KISS/DRY/SOLID/YAGNI quality cleanups, which this review skips./defense: the deeper security audit of the same diff.