review-board

name	review-board
description	Multi-lens, multi-model review conductor. Audits a diff/PR/branch by spawning several scoped Claude review agents in parallel (security, concurrency, performance, API-contract) WHILE a separate model (Codex) reviews independently, then reconciles every finding into one cross-model table: convergence (high-confidence), net-new-from-Codex (verify), and sharpened disagreements. Use when user says "review board", "full review", "multi-lens review", "audit this with all the models", or invokes /review-board. This is the heavyweight lane of the review ladder — for a routine review that posts to GitHub use /pr-review; to spend this only where it pays, /risk-gate first.
user-invocable	true
argument-hint	[PR number | branch | (default: working diff)]
allowed-tools	Bash, Read, Write, Glob, Grep, Agent

Review Board

Conducts a panel review of a change: N Claude lens agents run in parallel, each scoped to one failure-class, while a separate model (Codex) runs an independent review. Their findings are reconciled into a single ranked verdict.

The value is two kinds of diversity at once: lens diversity (each Claude agent only looks for one class of bug, so nothing is "averaged away") and model diversity (a different model has different blind spots). Running Codex concurrently with the lenses hides its latency behind work you were doing anyway.

Pairs with the codex-second-opinion plugin. If that skill / the Codex CLI is not present, this skill still runs Claude-only and says so.

---

Step 1: Resolve the target & the diff

• PR number / URL → check it out / fetch the diff.
• Branch name → diff against the merge base with the default branch.
• No argument → review the working diff (git diff + staged).

Collect the changed-file list and a unified diff:

git diff --merge-base "$(git remote show origin 2>/dev/null | sed -n 's/.*HEAD branch: //p')" --name-only

Fall back to git diff --name-only HEAD for the working-tree case. If there are no changes, stop and say so. Create the run dir: ./.review-board-out/findings/.

Step 2: Choose the lenses

Default lenses (one parallel agent each):

Lens	Looks only for
security	injection, authz/authn gaps, secret handling, SSRF, unsafe deserialization
concurrency	races, async-timing, process lifecycle, lock/ordering, idempotency
performance	N+1s, unbounded work, allocation hotspots, blocking calls on hot paths
contract	API/schema/type drift across layers, breaking changes, nullability

Read the full checklist for each lens from ${CLAUDE_PLUGIN_ROOT}/skills/review-board/references/lenses.md.

TEMPLATE: Add/remove lenses for your stack (e.g. accessibility, i18n, migration-safety). Each new lens is just another section in lenses.md — no new agent file needed; the agent is parameterized.

Skip a lens if no changed file is relevant to it (e.g. skip contract when no schema/type/API files changed). State which lenses you skipped and why.

Step 3: Launch Codex (separate model) in the BACKGROUND first

So it runs concurrently with the Claude lenses:

1. Check availability: which codex. If absent → skip this branch, note it, continue to Step 4.

2. If the codex-second-opinion skill is installed, follow its workflow to build the neutral brief and run it. Otherwise, build the brief inline: write a self-contained /tmp/review-board-codex-brief.md that (a) describes the change, (b) gives a file map with line hints, (c) lists invariants to verify, (d) asks for independent findings first, then confirm/refute, and (e) says read-only, do not edit.

3. Run it backgrounded so it doesn't block the lenses:

   cd <repo-root>
   codex exec --sandbox read-only "$(cat /tmp/review-board-codex-brief.md)" < /dev/null > ./.review-board-out/codex.out 2>&1 &
   

   The < /dev/null is mandatory: codex exec probes stdin for additional prompt input, and a backgrounded run with stdin open blocks forever waiting for EOF (observed 2h40m hang).

   Run the review on Codex's flagship tier (the -sol suffix in current generations) — if the configured default is a mid/speed tier, add -m <current>-sol. A cheap-tier reviewer weakens the cross-model signal this whole step exists for.

Do not wait on it yet.

Step 4: Fan out the Claude lens agents — in parallel

Spawn one review-lens subagent per chosen lens, all in a single message so they run concurrently (and concurrently with Codex). Give each:

• lens — the lens name (e.g. security),
• checklist — that lens's section from lenses.md,
• changed_files + the diff,
• output_dir = ./.review-board-out/findings/.

Each writes findings/<lens>.md (findings table) and returns a one-line summary. Do not perform the lens analysis yourself — that defeats lens isolation.

Step 5: Collect Codex, then reconcile

• Wait for the backgrounded Codex run; read only its final synthesis:

  awk '/^codex$/{last=NR} END{print last}' ./.review-board-out/codex.out
  

  then read from that line. (Skip if Codex was unavailable.)
• Read every findings/<lens>.md.
• Validation pass (kills false positives): before surfacing any finding, verify it yourself against the diff — distrust the report; a lens (and Codex) can over-flag. Drop anything below confidence 0.7 or that you can't confirm from the code.
• Tag each surviving finding by provenance: [CONSENSUS] (≥2 sources flagged it independently — highest confidence, lead with these) / [CROSS-VALIDATED] (a lens and Codex agree) / [SINGLE-SOURCE] (one source — verify before trusting).
• Reconcile into ./.review-board-out/REVIEW-BOARD.md per the contract below.

Step 6: Report

Print the path to REVIEW-BOARD.md and the headline: convergence count, net-new count, and the single highest-severity item. Done. (If invoked on a PR and the user wants it posted, hand off to pr-review rather than posting here.)

---

Output contract — REVIEW-BOARD.md

In this order:

1. Verdict — one paragraph: ship / fix-first / needs-discussion + why.
2. Cross-model table — | Finding | Severity | Lenses (Claude) | Codex | Your verification |.
3. Convergence — items flagged independently by ≥2 sources. Lead here; highest confidence.
4. Net-new from Codex — what only the separate model caught. Mark each "verify in code" — a different model is a strong signal, not gospel.
5. Disagreements / partly-confirmed — where sources differ; include Codex's evidence, which often sharpens a claim rather than negating it.
6. Per-lens detail — each findings/<lens>.md verbatim under an ## <lens> heading.

Error handling

Scenario	Action
No changes in target	Stop; tell the user the diff was empty.
Codex CLI absent / codex-second-opinion not installed	Run Claude-only; mark the Codex column "n/a" and note model diversity was not achieved.
Codex run errors / times out	Note it in the verdict; reconcile the Claude lenses; don't fail the whole review.
A lens agent writes no file	Note the missing lens in the report; reconcile the rest.
A lens has no relevant changed files	Skip it; list it under "lenses skipped".