Cross-Model Review

Use a fresh reviewer to challenge an artifact against the original request. A different model family can add useful diversity, but agreement is not proof. Preserve the user's configured model and effort unless the user explicitly requests a change. Skill guidance does not override the task's scope, existing authorization, or host permissions.

Choose the Review Surface

Run the chosen CLI's --version and relevant --help before relying on flags. Tool names, wait limits, permissions, and model availability belong to the current host. An external CLI is still delegation; it does not bypass a host prohibition.

For CLI details and a copyable launcher, read references/cli-flags.md. Use references/prompts.md for the requested review mode. Load references/failure-recovery.md only when a launch fails or stalls.

Freeze a Reviewable Artifact

Capture the original request, repository path, review base, head SHA, and owned file list. For working-tree review, include staged, unstaged, and relevant untracked files. A SHA alone does not describe uncommitted work; retain the diff and file contents or a content digest. Keep unrelated work outside the packet.

Freeze the reviewed files while the reviewer reads. Continue independent work elsewhere. If the artifact changes, identify the delta and re-review affected behavior. A metadata-only commit does not erase execution evidence for an identical tree, but an old review must never be described as examining a new SHA without a content comparison.

Write a Brief That Can Disagree

Give the reviewer the user's actual request and the artifact before your interpretation. Include necessary constraints and relevant repository instructions. Label your proposed explanation as a hypothesis. For a first independent pass, avoid supplying the conclusion you want confirmed.

A useful brief names:

• The question and exact review scope.
• The important behavior, interfaces, or failure modes to check.
• Available evidence and the environment needed to reproduce it.
• Permitted actions and required output.

Existing test results are receipts, not an instruction to trust the author. Let the reviewer repeat a decisive check or choose an independent probe. Avoid numerical confidence cutoffs, mandatory finding counts, and prompts that assume a defect exists. Read-only review does not mean tests are harmless: untrusted code needs an isolated environment without credentials.

For code, request each finding's trigger, impact, location, and evidence. For a specification, request an unmet requirement or a concrete counterexample. For a fact-check, request a verdict per claim with a primary source. For a design consult, request a recommendation and the evidence that would change it; a consultation needs no PASS.

Launch Once, Retain the Handle

For long Claude CLI reviews, use scripts/run_claude_review.py and inspect its saved status with scripts/review_status.py. Read references/observable-reviews.md for launch, authentication, scope, and recovery details. The helper retains events and stderr without requiring another reviewer process. Native host progress tools or a direct CLI capture remain suitable when they provide enough visibility.

Capture the full output to a unique file, and print that path before launch. Keep prompt content in a file or single-quoted heredoc. Do not interpolate untrusted text into shell code. A quoted "$(cat "$review_prompt")" reads a trusted prompt file as one argument; an unquoted heredoc can execute embedded shell syntax.

A running process or cell ID means the job is active. Poll that same handle with the host's supported wait operation. Wait duration controls when the host yields, not whether the reviewer succeeds. Use legal wait values and leave room for user updates; do not hardcode a five-minute initial wait into every harness.

Do not pipe a live review through head or tail as the sole capture. head can close the pipe early; tail commonly withholds output until EOF and discards earlier findings. Read the saved result after completion. Silence, transcript size, and elapsed time do not establish a verdict.

Authentication must follow the user's intended provider and billing route. For a subscription-backed Claude review, remove an inherited ANTHROPIC_API_KEY from that child process when it would select unintended API billing. Preserve it for an explicitly API-backed workflow. Never print credentials or change persistent auth configuration to make a review run.

Adjudicate Findings

A finding is a claim until its cited path and trigger survive a disproof attempt. Inspect the exact artifact reviewed before changing code. A reviewer can identify the wrong similarly named file, invent a CLI flag, or flag behavior already fixed at HEAD.

Repeat passes of one model and distinct review lenses add coverage of their own: in one exploratory vulnerability study, four passes of a single model found about as many issues as one pass from each of four families. Treat them as diversity alongside a second family, not as substitutes for it. Agreement among any of them increases investigation priority; it does not promote a candidate to confirmed.

A missing grep match does not establish absence. Compare base and head when claiming the change introduced a defect. Static tracing can confirm a deterministic error if the complete path is established; report the actual evidence tier without pretending execution occurred.

Fix verified issues and inspect the fix for new regressions. Preserve reviewer independence: switch explicitly from review to implementation when the user authorized both. Do not let a read-only reviewer silently edit the artifact it is judging.

Converge on Evidence

One pass is the default. Run a second only to close a blocking finding whose fix is itself non-trivial; for a small fix, the coordinator verifies closure by reading the diff. Keep a compact findings ledger when a second round is needed. Each entry has an identity, evidence, disposition, and verification status. Subsequent briefs carry the prior finding verbatim and the changed artifact so the reviewer checks closure rather than rediscovering the conversation.

A second pass asks only for closure of the carried findings. If the same objection returns without new evidence, settle the underlying contract or report the remaining disagreement. Avoid open-ended loops driven by a nonzero suggestion count. Nonblocking preferences do not expand an authorized fix into a redesign.

An interrupted review cannot grant PASS over unread material. Process exit zero means the CLI completed; the substantive report determines the review verdict. Distinguish model review, static analysis, and executed checks in the receipt. None replaces the others.

Evidence and Maintenance

Reviewed 2026-09-04 and 2026-10-27; sources and their limits are in references/evidence.md.

Anti-Patterns

What This Skill is NOT

• Not permission to send source to an unapproved provider, publish a review, or modify someone else's branch.
• Not protection against shared stale knowledge or a misleading brief.
• Not a substitute for runtime verification, product judgment, or understanding the finding.