Restore explicit per-subagent Fast control: Standard Astra coordinator with Fast Luna workers

Problem

I use Codex Desktop and want a straightforward mixed-model workflow:

• GPT-6 Astra coordinator: Standard speed for planning, delegation, and final review.
• GPT-5.6 Luna subagents: Fast speed for bounded research and execution tasks.

My available Astra usage is limited enough that I avoid enabling Fast for it. However, I still want the responsiveness of Fast Luna workers. Model choice, reasoning effort, and service tier are separate decisions, and I need to control them independently.

Requiring the entire agent tree to share the root service tier prevents this workflow. Keeping Astra on Standard also prevents explicitly selecting Fast for Luna; enabling Fast at the root defeats my intention to keep the coordinator on Standard. This is frustrating because using a capable coordinator with smaller, faster workers is a normal multi-agent workflow.

Subagent service_tier configuration appears to no longer take effect

The issue is not only the absence of a service_tier parameter in spawn_agent: setting service_tier = "fast" in a subagent's custom role configuration appears to no longer take effect. With the Astra root conversation on Standard, an explicitly Fast-configured Luna subagent appears to follow the root's Standard tier instead of honoring its own setting.

For example, the intended role-level setting is:

service_tier = "fast"

Please restore the effectiveness of this explicit subagent configuration, or clearly surface when it is overridden. A configuration that can be written but has no effect makes it difficult to know whether the requested speed is actually being used. This is an apparent configuration-effectiveness regression, not a claim that I have independently verified the server's applied tier or billing.

Relevant behavior change

PR #41308 explicitly removed per-spawn and role-level service-tier overrides and made subagents follow the root thread's service tier. In the currently exposed Desktop tool schemas, spawn_agent also has no service-tier parameter.

This report concerns the missing control and the resulting workflow limitation; I am not claiming a measured billing discrepancy or providing a controlled runtime reproduction.

Requested fix

Please restore explicit, user-authorized per-subagent service-tier overrides, while keeping root-tier inheritance as the safe default:

1. A child with no explicit override inherits the root service tier.
2. A user can configure a role such as Luna to use Fast while Astra remains Standard, or authorize an equivalent per-spawn selection.
3. Root-tier changes continue to update inheriting children, while explicitly overridden children have clearly documented behavior.
4. The UI and runtime status clearly show each agent's effective model and service tier. Unsupported tiers should be surfaced rather than presented as successfully enabled.

The important requirement is user control. If allowing the model to choose Fast freely creates unexpected usage, a human-authored role setting or an explicit user opt-in would address that concern without banning mixed service tiers altogether.

Expected outcome

A Standard Astra coordinator can launch an explicitly configured Fast Luna worker without switching Astra to Fast. Users who have not opted into an override retain the current inheritance behavior.

Related issues

• #42665 — Restore custom-agent service tier overrides for Standard coordinators and Fast workers.
• #35187 — Allow spawn_agent to select Normal or Fast mode per subagent.
• #42612 — Custom-agent Fast configuration ignored when the parent runs Standard.

This adds the concrete Desktop use case of Standard Astra + Fast Luna. If you prefer to track it under one of those existing issues, please consolidate it there; the key request is to restore this supported workflow.

Start with the current Desktop spawn_agent tool schema and custom role configuration, then review PR #41308 and related issues #42665, #35187, and #42612. Confirm how root-tier inheritance and explicit service_tier = "fast" settings are handled; done means an opted-in Fast Luna worker can run under a Standard Astra root, with effective tiers clearly surfaced and inheritance preserved by default.