Second Opinion

Second Opinion lets you consult a second AI model mid-session without interrupting your primary agent. Use it to pressure-test a design, get an outside read on a tricky bug, or sanity-check a plan before committing to it.

How It Works

When you trigger a consult, muxd:

1. The agent (or you) writes a focused summary of the question or context to send.
2. The summary is forwarded to the configured consult model as a single-turn request — no tool calls, no conversation history.
3. The response is displayed in a separate view marked with a crystal ball (🔮) so it is clearly distinct from the main conversation.
4. The response is never fed back to the primary agent. It is purely for your reference.

Because no conversation history is sent, the consult model starts from a clean slate every time.

Usage

/consult Slash Command

Trigger a consult from the TUI at any time:

/consult Should we store refresh tokens in the database or in a signed cookie?

The text after /consult is sent directly as the prompt to the consult model. You do not need the agent to be idle — you can consult while a task is running.

Via the Agent

You can also ask the primary agent to initiate a consult on your behalf:

"Get a second opinion on this database schema design before we proceed."

The agent calls the consult tool, writes a focused summary of the question, and the response appears in the separate consult view.

Configuration

Set the model used for consults:

/config set model.consult claude-haiku-3-5

Any model available in your provider configuration can be used. Defaults to the cheapest available model. The consult model is intentionally kept separate from your primary model so you can use a fast, cheap model for quick checks or a more capable model for deep review.

Consult View

The response appears below the main conversation in a distinct panel:

🔮 Second Opinion (claude-haiku-3-5)
─────────────────────────────────────
Signed cookies are generally preferable here because they avoid
a database round-trip on every authenticated request. The tradeoff
is that you cannot invalidate individual tokens server-side without
a blocklist — consider whether that matters for your threat model.

The panel disappears when you send the next message to the primary agent.

Constraints