Oracle Intelligence Layer

Oracle is a multi-agent critique and research engine. It runs Gemini, Copilot, Claude, and Codex in parallel to analyze plans, code, or architecture from diverse perspectives (Security, Implementation, Strategy, Engineering).

Features

• Unified Synthesis Engine: standardizes on a high-quality "Decomposition → Specialist Research → Synthesis" loop.
• Smart Context Injection: Automatically detects if a task requires workspace context (files, tree, snippets) or treats it as an abstract query to save time/tokens.
• Research Mode (--research): Forces deep workspace analysis even for abstract queries.
• Fast Path (--simple): Skips decomposition and sends the request directly to a single generalist (Codex/GPT-5.2) for speed on simple tasks.
• Auto-Apply (--apply): Automatically detects and applies code patches generated by the agents.
• Cognitive Scaffolding: Uses mandatory Chain-of-Thought (CoT) reasoning for high-stakes planning and synthesis.
• Resource Governance: Built-in token budgeting and context capping.

Usage

# Auto-mode (Decomposes complex tasks, simple for short ones)
./scripts/oracle.sh -p "How do I optimize this React component?"

# Workspace Research (Grounded in your files)
./scripts/oracle.sh --research -p "Explain how the routing logic works."

# Fast Path (Single Agent)
./scripts/oracle.sh --simple -p "Fix typo in README"

# Apply Changes
./scripts/oracle.sh --apply -p "Refactor the error handling in scripts/."

# List Models
./scripts/oracle.sh --models

Configuration

Settings are loaded from ~/.codex/oracle.json.

{
  "gemini_model": "gemini-3-pro-preview",
  "copilot_model": "gpt-5.1",
  "claude_model": "sonnet",
  "codex_model": "gpt-5.2"
}