Get a second opinion from Codex without leaving your Claude Code chat

The problem

Before a risky change (a migration, a retry loop, a change to auth) it helps to ask a model from the other family. Doing that by hand means copying the question and the relevant code into the other app, explaining the context again, and pasting the answer back.

• The other app doesn't see your repo unless you open a chat there in the same folder.
• Letting one agent start the other by itself (from its shell, say) gives you no say over what it sends, which model it uses, or what the second agent may do.
• A subagent in the same app is a second pass, but from the same model family.

How Pairbar solves it

1. Install Pairbar's tools for Claude and Codex
   Settings → Integrations → Pairbar tools for Claude and Codex → Install… on the Claude row (and the Codex row, for the other direction). Then start a new chat: the apps load plugins when a chat starts.
2. Ask the chat for a second opinion
   Say "get a second opinion from Codex on this migration before we run it", or type /pairbar:advisor in Claude ($pairbar:advisor in Codex). The chat calls second_opinion with its question and the context it thinks Codex needs.
3. Read the request and approve it
   "Second opinions and reviews" is Ask me to start with, so the request waits for you at the top of Pairbar's chat list: who asks, that it will start a Codex chat, "Read-only (no commands, no edits)", the budget line, and under What it will be sent the whole brief exactly as Codex will get it. Approve or Decline, there, from the notification's buttons, or in ⌘K.
4. Codex answers, read-only
   Pairbar starts a Codex chat in the Claude chat's folder that can only read files: no commands, no edits. It shows under the asking chat's row with its state, Open in Codex and Stop, and is an ordinary chat in Codex too.
5. The reply comes back to Claude
   The Claude chat collects the answer with wait or get_result. If it ended its turn instead, Pairbar delivers the result into it once it's idle (with Deliver results into the chat that asked on, as it is to start with).

What the Codex chat can do

• Read files in the asking chat's folder, and nothing else: no commands, no edits, no other tools (a Claude helper gets only Read, Glob and Grep; a Codex helper runs in the read-only sandbox).
• It gets the question and context as framed data, with a line saying who asked and that the text is its task, not instructions to do anything else.
• It answers and ends. A helper that runs longer than 30 minutes is stopped.

Which model answers

A second opinion uses the Second opinion profile unless the chat names another: the other app, its strongest model, high effort, reads files only. A chat can also pass its own model and effort, or reviewers: 2 to ask both apps (two tickets). When the other app has less plan left than Keep this much of each plan for you (25% to start with) and the asking chat's own app has enough, the asking chat's app answers with its strongest model, and the ticket says so.

Not asked every time

Tick Allow requests like this from this chat for 1 hour before Approve, and that chat's second opinions and reviews run without asking for the next hour, within the budgets. Or set Second opinions and reviews to Allow in Settings → Integrations → Agent actions. Pause agent actions refuses every request at once. See approve agent actions.

Without Pairbar

Ways to get a second model's view without Pairbar:

• Copy the question into the other app in a new chat in the same folder. Reliable; you carry the context and the answer across yourself.
• Run the other CLI from the chat's shell, for example codex exec in a read-only sandbox or claude -p, and let the chat read its output. It works, but nothing asks you first, and you set the model, access and limits by hand.
• Use a subagent in the same app for an independent pass. It's the same model family, and it runs inside the chat's own turn.
• Hand off the chat when you want the other app to take over the work, not just comment on it.