# Hostile Reviewer

> Gemini-native multi-model adversarial code review — iterative convergence loop for PRs and plans Use when this capability is needed.

- Skill: `tomevault-io/hostile-reviewer` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add tomevault-io/hostile-reviewer`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/hostile-reviewer/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: tomevault-io (https://skillmd.com/u/tomevault-io)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/tomevault-io/hostile-reviewer

---


# Hostile Reviewer Skill (Gemini Edition)

Multi-model adversarial review with iterative convergence, leveraging Gemini's 2M+ token window to perform "whole-project" reviews that identify structural flaws and architectural drift.

## Workflow
1. **Multi-Model Pass**: Gemini orchestrates a review pass across multiple models (Gemini, Codex, DeepSeek-R1).
2. **Finding Aggregation**: Gemini analyzes and deduplicates findings from all models, identifying "weighted-union" risks.
3. **Iterative Convergence**: Automatically applies fixes and re-runs the review until 2 consecutive passes produce no findings above NIT severity.
4. **Disagreement Triage**: Explicitly surfaces and triages disagreements between different models for human consideration.

## Gemini Advantages
- **Deep Architectural Grounding**: Gemini sees the entire PR or plan in the context of the *entire* codebase, identifying breaking changes that local diff-based reviews miss.
- **Superior Aggregation**: More intelligently merges findings from other models by understanding the technical intent behind their critiques.
- **Analytical-Strict Persona**: Maintains a rigorous, PhD-level domain expertise posture across the entire review loop.

## Arguments
- `--pr <N>`: PR number to review.
- `--file <path>`: Plan file to review.
- `--models`: Comma-separated model list (default: codex,deepseek-r1).
- `--passes`: Fixed number of passes (default: iterates to convergence).

---
> Source: [OmniNode-ai/omnigemini](https://github.com/OmniNode-ai/omnigemini) — distributed by [TomeVault](https://tomevault.io).
<!-- tomevault:4.0:skill_md:2026-06-16 -->

