Allow task handoffs to carry model and reasoning-effort recommendations as executable settings

Problem

When a task prompt includes a recommendation such as:

Model recommendation: GPT-5.6 Terra / Medium / Standard

Codex treats it only as text. It does not preselect the recommended model or reasoning effort. A stricter prompt guard such as If the active task is not using these settings, STOP is even less useful in practice: Codex stops and waits for the user to manually change the UI settings.

This creates unnecessary friction in workflows where another assistant prepares Codex tasks and can already determine an appropriate model/effort for each task.

Current workflow

1. ChatGPT prepares a Codex task and recommends a model/reasoning level.
2. User copies/sends the prompt to Codex.
3. User must separately change the model and reasoning effort in the Codex UI.
4. If the prompt contains a strict mismatch guard, Codex stops instead of helping apply the recommendation.

For long engineering workflows with many tasks, this becomes repetitive and error-prone.

Suggested improvement

Support model/reasoning recommendations as structured task metadata rather than relying on prompt text alone.

Possible UX options:

• ChatGPT/Codex handoff carries model and reasoning_effort as task metadata.
• Codex shows an Apply recommended settings action before execution.
• Optionally allow automatic preselection of the recommended settings, with the user still able to override them.
• If the requested model is unavailable, show a clear fallback/warning rather than requiring a prompt-level STOP instruction.

Why this would help

It would let task-planning assistants recommend a cheaper/faster model for routine work and a stronger model for security/architecture work without forcing users to manually synchronize prompt text with separate UI controls for every task.

The important distinction is that the recommendation should become an actionable execution setting, not merely text inside the prompt.

Expected behavior

A task created with a recommendation such as GPT-5.6 Terra / Medium should either:

1. open with those settings preselected, or
2. present a one-click confirmation to apply them before execution.

The user should retain final control and be able to override the recommendation.

The issue names no files, tests, or entry points. Start by locating the task handoff flow and the model and reasoning-effort selection paths, then determine how recommendations can be represented and surfaced without removing user control. Done means a handoff can preselect or offer one-click application of settings, with clear fallback behavior when a model is unavailable.