deep-review
One reviewer, reading deeply, instead of a squad of narrow ones. On this team's own commits it found every High the squad found plus two more, and 93% of what it reported was real (the squad: 46%). Benchmark and sources: references/sources.md.
Step 1 - Pick what to review
- Uncommitted work (the usual case): snapshot it in a private index, so your index, files and branch are untouched and untracked files (respecting .gitignore) are included:
The snapshot is an unreferenced commit nobody pushes. If the tree holds edits unrelated to this session, say so in the report.IDX=$(mktemp) GIT_INDEX_FILE=$IDX git read-tree HEAD GIT_INDEX_FILE=$IDX git add -A HEAD_SHA=$(git commit-tree "$(GIT_INDEX_FILE=$IDX git write-tree)" -p HEAD -m "deep-review: working tree snapshot") rm -f "$IDX" - A clean branch:
HEAD_SHA=$(git rev-parse HEAD). - Base in both cases:
BASE_SHA=$(git merge-base origin/<default branch> "$HEAD_SHA")(fetch first). - A PR handed over by review-pr, or a PR the user names: exactly its base and head.
- No diff between base and head: say so and stop.
Step 2 - Gather the intent, verbatim
Collect what the change is supposed to do: the Monday ticket text, the PR description, or the plan the session followed. Quote it; never write your own summary of the requirements, because a summary written by someone who already knows the problems steers the reviewer toward them. If there is none, pass "none given".
Step 3 - Run one reviewer
Fill in references/reviewer-prompt.md (placeholders {REPO}, {BASE_SHA}, {HEAD_SHA}, {DESCRIPTION}, {REQUIREMENTS}).
- Claude Code: exactly one
Agentcall -subagent_type: general-purpose,model: opus,effort: high- with the filled prompt. Do not split the diff across agents and do not add a second reviewer. Large diffs take longer (334 files took about 34 minutes); tell the user it is running. - Codex: do the review yourself in this session, following the filled prompt, with no sub-reviewers. Recommend high reasoning effort if the session is lower.
The reviewer is read-only: no edits, commits, pushes, posted comments or network calls.
Step 4 - Present the result
Before reporting no findings, revisit removed safeguards, deleted instructions, and tests that assert permissive behavior using the replacement-flow check in references/reviewer-prompt.md. A passing test can encode a regression. An acknowledged limitation in the PR description remains reviewable; acknowledgment alone is not acceptance.
Validate each finding against the exact reviewed head and omit findings already fixed there. Score confidence separately from severity, as defined in the reviewer prompt.
- Summary first: ready to merge (yes / no / with fixes) and the counts by severity.
- Then every issue with confidence >= 50, sorted by severity and then confidence, each with file:line, what is wrong, why it matters and the fix. Mark those >= 80. Do not drop the 50-79 band: on the benchmark, a reproduced crash scored 55.
- Name anything that touches permissions, access control or visibility rules as needing human review (Article 9). Never "fix" those silently.
- Include the reviewer's "Declined to judge" list in one line each.
Step 5 - When asked to investigate or fix
- Verify each finding against the code before changing anything; say per finding: confirmed / partly / wrong, with the reason.
- Fix only the confirmed ones, then run this skill again on the fix diff (
BASE_SHA= the commit before the fixes). A bug that escaped to the PR reviewer came from an unreviewed fix.
Never
- Never run
squad-revieworagent-skills:code-reviewfrom here, and never start them as a fallback. If the review cannot finish, report it as incomplete. - Never write a code-review attestation: no
<!-- code-review: ... -->line and no.git/code-review-attestationnote. If a repo'scode-review-checkthen fails a PR, tell the user the ways out:/review-pr, pasting a line from a review that did run, or theno-review-neededlabel.