Skip to content
Back to skills

Review

ASecurity

Request a read-only adversarial Codex review with explicit model, effort, budget, and compact findings.

  • 28 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added May 27, 2026
data-aipythonbashsecurity

Security analysis

A100/100

Scanned September 24, 2026

npx -y skills add Borda/AI-Rig --skill review --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Review?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Review
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/borda-review/badge)](https://www.skillsdirectory.com/skills/borda-review)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: review
description: Request a read-only adversarial Codex review with explicit model, effort, budget, and compact findings.
argument-hint: '[--model MODEL] [--effort LEVEL] [--timeout-seconds N] REVIEW_INSTRUCTIONS'
allowed-tools: Bash
---

# Ask Codex to Review

Parse `$ARGUMENTS`: required review instructions; optional `model`, `effort`, `timeout-seconds`, `depth`, `run-id`, `workspace`. Reject empty review instructions. Not a general question; for that use `/bridge:advise`. Effort omitted: classify complete scope, pass level explicitly. Preserve caller-supplied level. Tiers: `minimal` = trivial lookup or one-line mechanical confirmation; `low` = narrow mechanical change or settled fact; `medium` = bounded implementation, diagnosis, or review; `high` = cross-file, adversarial, architectural, or security judgment; `xhigh` = unusually broad and consequential; `max` = explicit caller request only.

Run `python "${CLAUDE_PLUGIN_ROOT}/bin/bridge_call.py" review --task "<instructions>"`; pass each option separately. For quoted text you did not author, use scratch file + `--task-file <path>` instead of `--task`; mutually exclusive. Bridge runs read-only ephemeral Codex execution with adversarial-review prompt; default soft budget 300 seconds. Never resume review.

Return compact JSON envelope. Keep bounded transcript at bridge-reported workspace-relative artifact path. Never inline peer `details`.

If `status=blocked`, open the JSON file referenced by `incident` and inspect its `fault`; `incident` may be `null` — then report blocked with no incident detail. For `output-limit`, report no review verdict. Split by file, module, or symbol; use fresh read-only calls with bounded tool output. If `status=timeout`, report the returned `effort` and the unreviewed scope; narrow the scope or raise `--timeout-seconds` — never resend at a higher effort against the same budget. If `status=refused`, report its `fault`: a request at inherited depth one or greater is refused (`recursion-depth`); never retry — the request already exceeded permitted cross-host chain depth. Reconcile every scope; disclose unanswered work. Never treat transcript fragments as a completed review.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.