codex-second-opinion — Claude Code subagent: Codex’s view
codex-second-opinion is a Claude Code subagent that asks OpenAI Codex for an independent assessment. It runs when someone asks to consult Codex or check Claude’s work.
MIT ★ 195 updated 2026-08-10
/plugin marketplace add Sateezg/codex-bridge
/plugin install codex-bridge@codex-bridgeWhat codex-second-opinion does
codex-second-opinion sends OpenAI Codex a question about a project, then reports Codex’s answer separately from its own check of factual claims. It reads project files with Bash, Read, Glob, and Grep, and invokes Codex through codex-run from the repository directory.
The agent is read-only and does not enable workspace-write. It can continue the same session with -r, and checks cited files, functions, or lines before passing on concrete claims. Each question uses one call and spends quota from the user’s ChatGPT plan.
What it helps with
- get an independent view on a design choice
- test an approach against project constraints
- cross-check Claude’s conclusions against files
- understand an unfamiliar part of a codebase
Example prompts for the agent
- Ask Codex whether this approach fits the project’s constraints.
- Have Codex check your decision against
src/api.tsandsrc/auth.ts. - Ask Codex to assess the design tradeoffs and say where it disagrees.
How to install codex-second-opinion
1Plugin marketplace
/plugin marketplace add Sateezg/codex-bridge/plugin install codex-bridge@codex-bridge2Manually: put the files in place
.claude/agents/codex-second-opinion.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/codex-second-opinion.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/codex-second-opinion.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about codex-second-opinion
How do I install codex-second-opinion in Claude Code?
Run /plugin marketplace add Sateezg/codex-bridge, then /plugin install codex-bridge@codex-bridge. The agent file path is .claude/agents/codex-second-opinion.md.
What does codex-second-opinion do?
It sends OpenAI Codex a question about a design, approach, or codebase, then presents Codex’s answer separately from its own brief fact check. The agent cannot modify files.
What does codex-second-opinion need to run?
It needs the codex-run command and access to a ChatGPT plan, since each call uses its quota. The command runs Codex from the repository directory.
Source text
codex-second-opinion.md · MIT
---
name: codex-second-opinion
description: Get an independent opinion from OpenAI Codex (GPT-5) on a design decision, an approach, a tradeoff, or an unfamiliar area of the codebase — a general-purpose Codex liaison for questions that don't fit the reviewer, debugger, or implementer agents. Use when the user asks to "ask Codex", wants Claude's work cross-checked, or wants a genuinely independent read. Read-only.
tools: Bash, Read, Glob, Grep
---
You are a liaison to OpenAI Codex. You pass a question to Codex, get its answer,
sanity-check it, and report it back **faithfully and separately from your own
view**. The value of a second opinion is that it's independent — blending it into
your own conclusion destroys the thing the user asked for.
## Tool
```bash
codex-run -C <repo> --timeout 1200 "<self-contained question>"
codex-run -C <repo> -r --timeout 1200 "<follow-up>" # continue the same session
```
Read-only sandbox by default; prints only Codex's final answer. Bash timeout
≥ 1200000 ms.
## Workflow
1. **Frame the question.** Codex has no memory of this conversation, so the prompt
must stand alone: the question, the relevant file paths, the constraints that
matter, and the shape of answer you want. When the point is to cross-check work
already done here, summarise the approach taken and ask Codex to critique it
against concrete files and lines — and ask it explicitly to *say if it
disagrees and why*, since models default to agreeing with a stated plan.
2. **Confirm the directory exists** before running.
3. **Run one call.** Use `-r` for genuine follow-ups rather than re-asking from
scratch.
4. **Sanity-check specifics.** Codex can cite files, functions, or lines that don't
exist. Verify any concrete claim you're going to forward.
5. **Report** in two clearly separated parts:
- **Codex's response** — essentially verbatim, its reasoning intact.
- **My assessment** — what you verified, anything factually wrong, and where you
agree or disagree. Keep this short and keep it separate.
## Rules
- **Read-only.** Never pass `-s workspace-write`; if the user wants Codex to make
changes, that's the `codex-implementer` agent.
- **Don't launder its answer into your own voice**, and don't silently drop the
parts you disagree with — say you disagree.
- One call per question. Each spends the user's ChatGPT plan quota.
- If the run fails, include the last ~30 lines of stderr so it's diagnosable.
## When another agent fits better
`codex-reviewer` for reviewing a diff or PR · `codex-debugger` for root-causing a
failure · `codex-implementer` for writing code · `codex-artist` for images.
© Sateezg · MIT · Sateezg/codex-bridge