codex-opinion

codex-opinion

A Claude Code plugin that gets a second opinion from OpenAI's Codex CLI on your work.

Prerequisites

• Claude Code — authenticated (claude in terminal)
• OpenAI Codex CLI — authenticated (codex in terminal)

Both must be logged in and working in your terminal before using this plugin.

Install

claude plugins marketplace add ehzawad/codex-opinion
claude plugins install codex-opinion@codex-opinion

Persists across sessions — no flags needed.

For development

git clone https://github.com/ehzawad/codex-opinion.git
claude --plugin-dir ./codex-opinion/plugins/codex-opinion

Usage

/codex-opinion:codex-opinion

With a custom instruction:

/codex-opinion:codex-opinion focus on security vulnerabilities

Claude Code also triggers the skill automatically when you ask for a second opinion in natural language — no slash command needed:

ask codex what it thinks about this diff
get a second opinion on my changes

How it works

When invoked, Claude gathers your diff, plan, or context and pipes it to codex exec. Codex uses your configured model and settings from ~/.codex/config.toml, reads the codebase, runs commands, and does deep analysis. Claude reads the response and reports back.

sequenceDiagram
    participant U as User
    participant C as Claude Code
    participant S as ask_codex.py
    participant X as Codex CLI

    U->>C: /codex-opinion:codex-opinion
    C->>C: Gather diff / plan / context
    C->>S: Pipe context via stdin
    S->>X: codex exec --json
    X-->>S: JSONL events
    S->>S: Extract final message
    S-->>C: Codex's analysis via stdout
    C-->>U: Reports findings

Session management

One Codex session per project, stored at ~/.local/state/codex-opinion/{project-hash}.json. Follow-up calls resume the prior Codex thread so it builds on its accumulated codebase knowledge — across Claude Code sessions, not just within one.

If the session has expired server-side, the script logs a notice and starts fresh.

flowchart TD
    A[Invoke /codex-opinion:codex-opinion] --> B{Session file exists<br/>for this project?}
    B -- Yes --> C[codex exec resume session_id]
    C --> D{Resume succeeded?}
    D -- Yes --> E[Extract response]
    D -- No --> F["Log: session could not be resumed"]
    F --> G[Start fresh session]
    B -- No --> G
    G --> H[Save session metadata]
    H --> E
    E --> I[Return to Claude]

JSONL protocol

The script communicates with codex exec --json via JSONL events on stdout:

sequenceDiagram
    participant S as ask_codex.py
    participant X as Codex CLI

    X->>S: {"type": "thread.started", "thread_id": "UUID"}
    Note over S: Captures session ID
    X->>S: {"type": "turn.started"}
    X->>S: {"type": "item.completed", "item": {"type": "agent_message", "text": "..."}}
    Note over S: Captures last agent message
    X->>S: {"type": "turn.completed", "usage": {...}}
    Note over S: Returns final message to Claude

Security

Codex runs with --dangerously-bypass-approvals-and-sandbox — no approval prompts, no filesystem sandbox. This gives Codex full read/write access to your machine so it can thoroughly inspect and analyze the codebase. Do not use this plugin on untrusted repositories or with untrusted input.

Configuration

The script uses your Codex CLI defaults — model, reasoning effort, and other settings come from ~/.codex/config.toml. No model is hardcoded. Sandbox and approval settings are overridden by the plugin (see Security above).

License

MIT