codex-oracle

name	codex-oracle
description	Query OpenAI Codex (GPT-5.2-codex) for second opinions, alternative implementations, or specialized knowledge. Use when you want to consult another AI model, need a different perspective on a coding problem, or want to verify an approach.
argument-hint	[prompt]

Codex Oracle

Invoke OpenAI's Codex CLI to get responses from GPT-5.2-codex.

Prerequisites

• codex CLI installed (npm i -g @openai/codex)
• Authenticated via codex auth login (uses account auth, not API key)

Usage

Run the script with a prompt. Codex runs in read-only mode and sees the same working directory.

# Determine script location (project-local or global install)
CODEX_SCRIPT=".claude/skills/codex-oracle/scripts/ask_codex.sh"
[[ ! -f "$CODEX_SCRIPT" ]] && CODEX_SCRIPT="$HOME/.claude/skills/codex-oracle/scripts/ask_codex.sh"

# Default: gpt-5.2-codex with medium reasoning
bash "$CODEX_SCRIPT" "What's the most efficient way to implement a bloom filter in Rust?"

# With model override
bash "$CODEX_SCRIPT" "Review error handling in this file" codex-mini-latest

# Full output (all thinking/exec traces)
bash "$CODEX_SCRIPT" "Debug this complex race condition" gpt-5.2-codex xhigh 600 full

Arguments

Arg	Default	Description
$1	(required)	The prompt
$2	gpt-5.2-codex	Model (gpt-5.2-codex, codex-mini-latest)
$3	xhigh	Reasoning effort (low, medium, high, xhigh)
$4	600	Timeout in seconds (10 min)
$5	brief	Output verbosity (brief = final answer only, full = all traces)

When to use

• Need an alternative implementation approach
• Want to verify Claude's solution with another model
• Specialized domain where Codex may have different training
• Debugging: "Why might this code fail under X conditions?"
• Security review or vulnerability analysis (Codex has strong cybersec capabilities)

Prompting tips

• Be specific: include language, constraints, context
• For code review: reference file paths (Codex sees the repo)
• For debugging: include error messages verbatim
• Use xhigh effort for complex multi-step reasoning

Limitations

• Read-only: Codex cannot modify files (by design)
• Timeout default 10 minutes for xhigh reasoning
• No streaming—waits for full response