Reviewer: a stronger model, a fresh context

Reviewer >= author, never weaker

A model reviewing its own output shows measurable self-preference bias; a different model that is at least as strong does not. That is why the reviewer agent is pinned to opus/high regardless of which model wrote the code, and why cross-model review has a direction: Claude may review Gemini’s code, and Gemini 3.1 Pro (high) may do a second pass on Claude’s code, but Claude models inside Antigravity must never review Claude’s own code — same vendor lineage reviewing itself.

Source: home/.claude/agents/reviewer.md (comment above the procedure); home/.claude/MANUAL.md §5 (“Revisor ≥ autor… cadena de revisión cruzada permitida”).

Fresh context, its own cache

reviewer is Read, Grep, Glob, Bash(git *) only — read-only, it cannot edit files. It runs in its own subagent context, with its own prompt cache, so spawning it does not touch the parent session’s cache the way switching the session model with /model would. That is the concrete reason routing rule 6 (docs/routing.md) says never to change the session model to get a review.

Source: home/.claude/agents/reviewer.md frontmatter (tools:, description); home/.claude/CLAUDE.md §Routing rule 6.

Why same-model self-review is not a gate

The evidence the kit cites: same-model self-review is biased toward approval (Panickssery 2024). Using a different, at-least-as-strong reviewer with a forced-choice rubric is the kit’s own practice, not yet backed by a paper — see docs/references.md for what is verified and docs/references-candidates.md for the search trail on the citation this used to carry. This is why both reviewer and /xreview use a rubric with no dimension allowed to be skipped, rather than open-ended commentary.

Source: home/.claude/skills/xreview/SKILL.md (Evidence line).

/xreview: cross-model review, step by step

xreview has two modes, invoked as pack or ingest; it does not auto-trigger from a description (disable-model-invocation: true).

pack:

1. Collect git diff main...HEAD (or the staged diff), SPEC.md, the test command and its last output.
2. Write review/REQUEST-<date>.md: context in <=10 lines, the diff (or a file list if it’s over 800 lines), and the rubric below. Instruct the external reviewer to be read-only and to write its report to review/REPORT-<tool>-<date>.md.
3. Run the external reviewer and save its output. For Antigravity’s agy CLI in headless mode: whole-repo/architectural audits use gemini-3-1-pro at high effort (1M context, independent lineage); a bounded diff uses gemini-3-8-flash-high (fast, separate quota pool from Claude). Antigravity’s Claude models must never review Claude-written code. If agy is missing or errors, say so and stop — never silently fall back to self-review. Codex alternative: codex exec --sandbox read-only.

ingest <path-to-report>:

1. Read the report.
2. Verify each finding against the actual code (open the file:line) and classify it: confirmed / not reproducible / disagree (with reason).
3. Confirmed findings go into HANDOFF.md’s “Next steps” with a verification command each; disagreements get one line each for the owner to arbitrate.
4. Never apply a reviewer’s rewrite wholesale — fix the confirmed gap with the smallest possible diff.

Source: home/.claude/skills/xreview/SKILL.md.

The rubric (both reviewer and /xreview)

Forced choice per dimension — pass / concern / fail — with file:line evidence, no prose without a location:

Dimension
Correctness vs SPEC invariants
Failure modes not handled (empty input, concurrency, auth, partial writes)
Security (injection, secrets, path traversal, permissions)
Duplication added / consolidation or refactor debt skipped
What the tests do NOT cover

reviewer additionally reads SPEC.md and runs the project’s test command, quoting the last 15 lines, and closes with “Would I merge? yes / no, because .” If a dimension is fine, the required phrasing is “pass — checked ” — findings are never invented to look thorough, and passes are never padded with praise.

Source: home/.claude/agents/reviewer.md; home/.claude/skills/xreview/SKILL.md (Rubric).

templates/REVIEW.md

The template a review report is expected to fill: a verdict checkbox (merge as-is / merge with fixes / send back — does not meet SPEC.md), a Findings list (file:line — what’s wrong, why it matters), a Security check line (no secrets, no credentials, no path traversal, no scope creep beyond the task’s allowed files), and, if not merging, a “Sent back for” list of exactly what to fix.

Source: templates/REVIEW.md.

What the reviewer must NOT do

Report only — never edit files, never fix what it finds, never invent findings to appear thorough, never praise a passing dimension. Both home/.claude/agents/reviewer.md and the /xreview skill state this explicitly, and /xreview ingest step 4 extends it: even after a human decides a finding is confirmed, the fix is applied by hand as the smallest diff — never by pasting in the reviewer’s own rewrite.

Source: home/.claude/agents/reviewer.md (“Report only. Do not edit files.”); home/.claude/skills/xreview/SKILL.md.