Redteam

---

name: redteam description: Use when a decision, design, diff, or PR needs adversarial third-person scrutiny before it is accepted or merged. Spawns independent opponents that argue the opposite and demand evidence, then returns a deterministic per-decision verdict (cleared / revise / blocked) and an overall gate. Use inside the worker gauntlet, before any autonomous merge, or any time a choice should be stress-tested.

Red Team — Adversarial Decision Auditor

Announce: "Using the redteam skill to adversarially audit ."

Question every material decision in opposition, in the THIRD PERSON, and force an evidence-backed defense or a revision. Default stance: a decision is NOT justified until its defense survives opposition.

When to use

• Inside the worker gauntlet (self-redteam after implementation; intent-check before merge).
• Before any autonomous merge on a real-money surface.
• Any time you want to stress-test a design, diff, or PR.

Inputs

• Target: a decision list, design doc, diff, or PR number.
• Intent: the issue's acceptance criteria + the mission spec's success criteria + constraints.

Procedure (follow exactly)

1. Frame. State the target and the intent in 2-3 lines. If the intent is missing, get it before proceeding — you cannot judge a decision without knowing what it is for.

2. Extract decisions. Enumerate the material decisions (see references/severity-rubric.md, "What counts as a material decision"). For each record: id, a one-line summary, the surface it touches (standard vs elevated, per the rubric), and a severity estimate.

3. Oppose. For each decision, dispatch independent opponent subagents using references/opponent-prompt.md verbatim (fill the {{slots}}):

   • standard surface: 3 opponents
   • elevated surface: 5 opponents Dispatch them concurrently — one message, multiple Agent calls.

4. Defend / revise. Collect each opponent's JSON verdict. For every refutation, answer with concrete evidence (code, tests, spec citation). An unanswered or hand-waved refutation STANDS (it counts as refuted: true). Do not perform agreement; cf. superpowers:receiving-code-review.

5. Aggregate (deterministic). Write the opponent results as JSON in the references/verdict-format.md input schema, then run: node bin/aggregate.mjs <opponents.json> It applies the majority/unanimous + severity-floor rules and prints the per-decision verdict and the overall gate. Do not compute the verdict by hand.

6. Report. Emit the human-facing report (references/verdict-format.md): per decision — challenged alternative, the agent's defense, verdict, severity, required action — then the final GATE: pass|fail line.

Rules

• Elevated-surface decisions require a UNANIMOUS clear; any standing refutation fails them.
• The gate passes only if EVERY decision is cleared. Both revise and blocked fail the gate.
• Never soften a verdict to be agreeable. The opponents' job is to be right, not kind.
• If aggregate.mjs exits non-zero or prints an error instead of a verdict, treat the gate as FAILED (fail closed). Never read a missing verdict as a pass.