Codex Second Opinion

Delegate a task to OpenAI's Codex CLI for an independent review or second opinion.

Input

<task> $ARGUMENTS </task>

If no task is provided, ask the user what they want Codex to review.

How to Use

Run Codex in non-interactive mode with read-only sandbox:

codex exec --full-auto --skip-git-repo-check -s read-only -C "<current working directory>" "<prompt>" 2>/dev/null

Prompt Construction

Build a self-contained prompt for Codex that includes:

1. Context: The repo path, what project this is, and what folder/files are relevant
2. The task: What you want Codex to do (review, validate, suggest, etc.)
3. Constraints: Tell it to only read files, not modify anything

Keep it unbiased. Do not include your own conclusions or preferred answer. Present the problem neutrally so Codex forms its own opinion.

Be specific about files. Tell Codex exactly which files to read — it has access to the repo but needs direction.

Example

codex exec --full-auto --skip-git-repo-check -s read-only -C "/path/to/project" \
  "Review the file at src/config.py. Is the retry logic correct? Read the file and give your assessment. Do NOT modify any files." 2>/dev/null

Troubleshooting

• Sandbox restrictions: If Codex output is empty, retry with dangerouslyDisableSandbox: true on the Bash call. The codex CLI needs network access and may spawn subprocesses that hit sandbox limits.
• Output truncation: Codex may run many tool calls (file reads, experiments), producing very large output. Keep prompts focused and ask specific questions rather than open-ended reviews. If output is truncated, the empirical results are still valuable — look for the data in the tool call results even if the final summary is missing.
• --skip-git-repo-check: Always include this flag to avoid errors in repos with uncommitted changes.

After Codex Responds

1. Report the Codex response to the user
2. Note where you agree or disagree with Codex's assessment
3. Let the user decide how to proceed