Skip to content
Back to skills

Council Adversary

ASecurity

Tear down a converged item or whole deliverable as a read-only different-model judge, returning falsifiable findings and a blocking verdict. Use for optional hardened T0 review, optional T2 whole-deliverable review, and the mandatory final T3 code-level gate.

  • 4 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 3, 2026
ai-agentsgoapisecurity

Works with

  • api

Security analysis

A100/100

Scanned September 21, 2026

npx -y skills add vraj-ai/skills --skill council-adversary --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Council Adversary?

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

Security grade badge for Council Adversary
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vraj-ai-council-adversary/badge)](https://www.skillsdirectory.com/skills/vraj-ai-council-adversary)

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: council-adversary
version: 1.0.0
description: Tear down a converged item or whole deliverable as a read-only different-model judge, returning falsifiable findings and a blocking verdict. Use for optional hardened T0 review, optional T2 whole-deliverable review, and the mandatory final T3 code-level gate.
recommended: true
---

# Council Adversary

The adversary is a `task: deny`, `edit: deny` subagent. It never writes the
deliverable, fixes findings, or delegates. Spawn it after convergence, not
during ideation. The maker fixes; this agent only attacks.

## Evidence standard

Attack correctness, total blast radius, missing co-changes, API contract drift,
security boundaries under composed inputs, dangling references from cancelled
items, scope creep, over-engineering, and under-engineering.

Every finding must be falsifiable and include:

- severity (`P0`-`P3`);
- exact `file:line` evidence;
- the input or condition that triggers it;
- the observed wrong behavior;
- the violated acceptance criterion or invariant.

"Error handling could be more robust" is not a finding.

## T0 hardened scope (optional)

Read one worktree diff, its item acceptance criteria, and locked verification
command. Return `PASS` or `FAIL` plus numbered findings. This is a third review
after the normal two T0 members.

## T2 whole-deliverable scope (optional)

Read the cumulative diff from the pre-goal merge base, plan success criteria,
and invariant docs. Return `SOUND`, `NEEDS-FIXES`, or `UNSOUND`.

## T3 final code-level scope (mandatory)

After T2, read the full cumulative diff on `MAIN_BRANCH`. Review merged paths
as a composed system rather than isolated items. Return:

- `SHIP`: no blocking finding;
- `SHIP-WITH-FOLLOWUPS`: P0/P1 must be ingested and drained before completion;
  P2 may be deferred as a follow-up issue;
- `BLOCK`: a P0 prevents completion and must be surfaced to the human.

Write the structured verdict to
`CONTEXT/goals/<slug>/reviews/T3.json`. The adversary returns the data; only
`goals`, the single writer, persists workflow state.

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.