blackicelisted

Run a human-in-the-loop adversarial review PANEL over mission-critical / records-critical / irreversible code (money movement, tax or regulatory records, data integrity, safety, migrations, auth). It drives one `claude -p` subagent PER reviewer persona PER epoch, iterating until a halting condition (converged / budget / epochs / stall) or a ruin-class "UGLY" circuit-breaker. Use when a change "looks obviously right" but being wrong is expensive, or when the user asks for a deep/rigorous/adversarial/multi-expert review, a "panel", or to "find edge cases, traps, pathologies". NOT for routine, reversible, or well-covered changes — it is deliberately heavy and spends tokens.
aigora-de/blackice · ★ 0 · AI & Automation · score 66

Install: claude install-skill aigora-de/blackice

# blackice — HITL adversarial review panel ## What it is A generalisation of the two-pass adversarial panel (see `two-pass-adversarial-review-pattern.md`) into a **bounded, human-convened iteration loop**. A deterministic Python engine (`kuang/engine/`) owns the control logic — halting, dedup (a coarse signature always-on, plus an opt-in semantic **reduce** into canonical clusters), stall detection, token/time budget, and the **UGLY circuit-breaker** — and binds to Claude Code via `kuang/backends/claude_code/`, which spawns **one `claude -p` subprocess per persona per epoch** (read-only: no edit tools). You (the convening `main` session) supply the scope, run the loop, **synthesise** its output, and gate decisions with the human. You are the *synthesiser, not the judge* — the human decides. ## How to run it 1. **Scope it.** Decide `--why` (the risk being guarded against) and `--what` (the change or code under review), and the review surface. Exactly one mode: - **Diff mode** — `--base`/`--head`: review a *change*. - **Path mode** — `--paths <file|dir> …`: review *existing code* in full (proactive bug-hunting, or a repo with no reviewable diff). Directories expand via `git ls-files` (honouring `.gitignore`); a total-size cap (`--max-surface-bytes`) bounds the surface, naming anything it omits. 2. **(Optional pre-flight)** `--dry-run` to confirm *which* personas were sourced (e.g. from `CLAUDE.md`) and eyeball the assembled prompt. This spawns **no