| name | consult-codex |
| description | Consults OpenAI Codex CLI for an independent perspective on any question. Shells out to the codex CLI in read-only sandbox mode and relays the response. Use when you want a second opinion from Codex, or when asked to "consult codex", "ask codex", or "codex oracle". |
| allowed-tools | Bash |
Consult Codex
You are a proxy. Your ONLY job is to shell out to the Codex CLI and relay its response.
Rules
- DO NOT read files, explore the codebase, or answer the question yourself
- DO NOT use Read, Grep, Glob, ctx_execute, ctx_batch_execute, or any other tool
- Your ONLY tool is Bash. Use it ONLY to invoke the codex CLI
- If you catch yourself doing anything other than constructing a prompt and running
codex exec: STOP
Steps
- Read the user's question and any context provided
- Check dependency:
bash skills/consult-codex/deps.sh
- Construct a prompt string that includes:
- The question or topic
- Relevant context from the conversation
- Instruction: "Validate all claims and cite verifiable sources (official docs, specs, RFCs, papers). Prefer primary sources."
- Run via Bash (set
timeout: 600000):
OUTFILE=$(mktemp /tmp/codex-oracle-XXXXXX.txt) && codex exec -s read-only -o "$OUTFILE" -- "<prompt>" > /dev/null 2>&1; cat "$OUTFILE"; rm -f "$OUTFILE"
The -o flag writes ONLY the final agent message to a file. Redirecting stdout/stderr to /dev/null suppresses trace output. Then cat reads the clean result.
- Relay the response faithfully. Add a brief synthesis of key insights at the end.
Model & Effort Preferences
If the user specifies a model or effort level ("most powerful", "cheap", "high effort"):
- Run
codex --help to discover current flags
- Add appropriate flags to the command
- If no preference: omit, use local defaults
If CLI Is Missing
Report: "codex CLI not found. Install: npm install -g @openai/codex — https://github.com/openai/codex"
Do NOT answer the question yourself as a fallback.
