oracle

joelklabo 的头像作者 joelklabo

Parallelized multi-agent critique system using Gemini, Copilot, Claude, and Codex. Decomposes complex engineering tasks into specialized domains and synthesizes a single authoritative report.

📖 中文导读:Parallelized multi-agent critique system using Gemini, Copilot, Claude, and Codex. Decomposes complex engineering tasks into specialized domains and synthesizes a single authoritative report. 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: oracle description: Parallelized multi-agent critique system using Gemini, Copilot, Claude, and Codex. Decomposes complex engineering tasks into specialized domains and synthesizes a single authoritative report.

Oracle Intelligence Layer

Use this skill to get a deep, multi-model analysis of your code, plans, or architecture.

Workflow

  1. Provide a prompt or a file to analyze.
  2. Oracle automatically determines if it needs code context.
  3. It decomposes the task into 4 specialized expert domains.
  4. It synthesizes the findings into a prioritized Master Report.

Usage

# General usage
~/.codex/skills/oracle/scripts/oracle.sh -p "<prompt>"

# Implementation/Refactor (forces code context)
~/.codex/skills/oracle/scripts/oracle.sh --research -p "Refactor auth.ts"

# Fast path (single model)
~/.codex/skills/oracle/scripts/oracle.sh --simple -p "Explain this regex"

# Apply fixes found in the report
~/.codex/skills/oracle/scripts/oracle.sh --apply -p "Fix bugs in main.py"

Configuration

Overrides can be set in ~/.codex/oracle.json or via environment variables:

  • ORACLE_GEMINI_MODEL
  • ORACLE_COPILOT_MODEL
  • ORACLE_CLAUDE_MODEL
  • ORACLE_CODEX_MODEL

Model Selection

  • Primary: gemini-3-pro-preview (Synthesis), gpt-5.1 (Copilot), sonnet (Claude), gpt-5.2 (Codex).
  • All tools run in non-interactive mode with auto-approval for seamless CLI integration.