review-impl

Review an implementation you did not write.

Target: $ARGUMENTS

Use the unbiased-review skill. Read references/implementation-review.md, references/isolation-protocol.md (required before executing anything), references/output-contract.md, both rubrics, and references/stack-notes.md.

This is the only review where execution is available, so use it. Reading a diff tells you what the author meant; running it tells you what they did.

Establish the change completely first:

git -C <target> status --short
git -C <target> diff --staged --stat
git -C <target> diff --stat
git -C <target> ls-files --others --exclude-standard

Untracked files are part of the change. Reviewing staged-only content on a tree with untracked new test files reviews the wrong change and misses the tests entirely, which inverts every TDD finding.

Establish the approved bar before judging scope. Locate the spec and plan in the target — including uncommitted/untracked docs in the working tree — and any saved spec/plan review in the reviews location. Those set expectations, not the author's summary of them. If none exists, that absence is a finding and blocks APPROVE. No session decides scope, defers work, or records a follow-up without the human's approval; an unapproved deferral is a finding, never something you bless (Hard rules 3, 8).

Then: three-way check against spec and plan (and any prior report's finding IDs), the suite run in a scratch clone with real numbers, the red-green proof for the change's core assertions, and architecture graded by import graph rather than folder names.

The skill's Hard rules bind hardest here. Never run a test or build inside the target — clone to the scratch root defined in references/isolation-protocol.md (a temp path outside any target tree) and work there. Reproduce before reporting: no empirical failure goes in the report until you have proved it exists in target state and not just in your copy. Remove the scratch clone at the end and confirm the target's git status is unchanged.