Feature request: Native read-only second-opinion Advisor for high-impact decisions

Summary

Codex should support an optional, bounded second-opinion workflow for high-impact or high-uncertainty decisions.

At a qualifying checkpoint, the main agent would create a concise decision brief, consult a read-only Advisor, and then integrate the result while explaining which recommendations it accepted or rejected.

For example, a user might use gpt-5.6-luna at Max as the main agent and consult gpt-5.6-sol at Extra High or Max only for selected Advisor reviews. The specific model names should remain user-selectable examples rather than hard-coded behavior.

Requested behavior

• Support an explicit “Ask Advisor” action and an optional automatic mode.
• Trigger only at meaningful checkpoints, such as architecture choices, security boundaries, data migrations, destructive actions, hard-to-reverse design decisions, or repeated failed attempts.
• Prepare a decision brief containing the goal, constraints, available options, current recommendation, risks, and unresolved uncertainty.
• Provide the Advisor only the context needed for the review.
• Keep the Advisor read-only: no file changes, tool side effects, external actions, or recursive Advisor calls.
• Let the main agent retain responsibility for the final decision and report how the advice affected it.
• Provide controls for the Advisor model or reasoning mode, invocation count, token or cost budget, timeout behavior, and privacy/context scope.
• Make it visible when and why an Advisor was consulted, what context was shared, and which model or mode produced the review.

Why this matters

Using the strongest reasoning mode continuously can be unnecessarily expensive and slow. A bounded second opinion at a few consequential decision points could improve reliability while keeping normal work on the preferred main model.

This is not intended to create a model debate on every task. The value is a selective, bounded, read-only, and auditable quality gate before important decisions.

Related issues, but different scope

• #14039 concerns model selection and routing capabilities.
• #31965 concerns exposing built-in Chat/Quick Chat as a callable consultation primitive.

Those capabilities could help implement this feature, but they would not by themselves provide checkpoint detection, decision-brief generation, bounded read-only review, or synthesis by the main agent.

Success criteria

A user can enable the workflow, define its limits, see when and why an Advisor was consulted, inspect the review and the main agent’s final synthesis, and verify that the Advisor could not perform side effects.