Adversarial Review

Retired from the marketplace on 2026-09-09. Use the adversarial-review skill in Thermos 0.3.0 or later. This source remains only for rollback and historical reference.

Adversarial Review is a momentum-first, skills-only Codex plugin for an explicit high-scrutiny review. Invoke it and Codex infers the active target, then immediately runs exactly three fresh reviewers through native Subagent V2:

Skeptic: correctness, security, edge cases, and broken assumptions
Architect: boundaries, coupling, rollout, and migration risks
Minimalist: scope creep and unnecessary complexity

You do not need to switch permission modes, choose models, describe agent mechanics, or separate unrelated worktree state. Reviewers can use inherited tools and approval handling to run targeted tests, builds, linters, and diagnostics. They investigate; they do not turn the review into an unsolicited repair pass. The parent returns one evidence-backed PASS, CONTESTED, or REJECT verdict.

The standard child wave gets three minutes before interruption is considered. Reviewers still making visible progress can continue for up to five total minutes. Stalled reviewers are interrupted, while a deeper wait happens only when you explicitly ask for it.

You can invoke the skill repeatedly in one Codex task. Each review automatically uses the next free shared run number for its three internal child names, so completed or partial earlier runs cannot collide. The verdict still shows only the useful Skeptic, Architect, and Minimalist labels.

Why this is a plugin

The source format remains a normal SKILL.md with progressive-disclosure references. The plugin is its installable distribution unit for the marketplace and the ChatGPT desktop app; it does not need an app mapping, MCP server, hook, or dedicated agent-profile configuration.

Implicit invocation is intentionally disabled because this workflow consumes three independent V2 child runs. Use $adversarial-review or select the installed plugin; that invocation is enough to start the workflow.

Supported scopes

Name a target when you care about an exact boundary:

a patch, plan, tight file list, or commit range
staged or unstaged changes
a named commit
a branch or pull request against a base
the last assistant turn

Otherwise the skill infers the narrowest relevant target from the active request, tracked changes, branch diff, or current artifact. It excludes unrelated state without asking you to manage the boundary.

Install and use

After this catalog change is released, refresh the configured marketplace and install the package:

codex plugin marketplace upgrade mcook-plugins-local
codex plugin add adversarial-review@mcook-plugins-local
codex plugin list | rg 'adversarial-review@mcook-plugins-local'

Then start a new Codex task. In the ChatGPT desktop app, open Plugins from Work or Codex, choose the marketplace, install Adversarial Review, and start a new task before using the bundled skill.

Acceptance

The marketplace bundle is the canonical discoverable copy. After refresh and installation, use the fast release gate below. Do not run a full three-reviewer task merely to prove installation; it consumes three model runs and is intentionally expensive.

Fast release gate

python3 ~/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py adversarial-review
python3 -m unittest adversarial-review/tests/test_contract.py
python3 -m json.tool .agents/plugins/marketplace.json >/dev/null
if rg --hidden -n '[[:blank:]]$' adversarial-review; then exit 1; fi
git diff --check -- adversarial-review .agents/plugins/marketplace.json
codex plugin list | rg 'adversarial-review@mcook-plugins-local.*0\.3\.3'
diff -qr --exclude=__pycache__ \
  adversarial-review \
  ~/.codex/.tmp/marketplaces/mcook-plugins-local/adversarial-review

The deterministic contract tests and byte-for-byte installed-tree comparison are the normal publish acceptance. See evals/README.md for optional behavioral cases when the user explicitly wants to spend the time and tokens on a live trio.