[App] Add user-confirmed model and reasoning routing for persistent task creation

Start at the codex_app.create_thread entry point and compare its current model and thinking handling with the managed defaults in requirements.toml. Review the official config schema linked in the issue and the related issues for existing subagent routing. Done means persistent task creation supports explicit user approval or edits, batch mappings, auditability, and no silent fallback when a pair is unavailable.

What variant of Codex are you using?

Codex App (Desktop), with persistent user-visible tasks created through codex_app.create_thread. The local Codex host used for verification is codex-cli 0.146.0.

What feature would you like to see?

Add a first-class, human-in-the-loop model and reasoning routing flow for persistent task creation.

The desired flow is:

1. The coordinating agent evaluates the proposed task.
2. It proposes an exact model and reasoning-effort pair for that task.
3. Codex shows the proposal to the user in a structured confirmation/edit UI.
4. Only after the user approves or edits the pair does Codex create the persistent task.
5. The created task records the proposed, approved, and effective model/reasoning values.

For multiple tasks, the app should support confirming a task-to-model/reasoning mapping as one batch before creation.

Why this is needed

The current create_thread tool accepts model and thinking, but its agent guidance says not to specify a model unless the user explicitly requests a specific model. Otherwise, the fields are omitted and each new task inherits the configured global default.

That is safe, but it prevents an orchestrator from applying task-specific routing unless the user already knows which exact model to request. In practice, a user may want the agent to make the technical recommendation while retaining final control over cost/capability.

In one observed workflow, a coordinator classified several materially different tasks but omitted model and thinking because no exact model had been explicitly requested. All persistent tasks therefore inherited the same global default (gpt-5.6-sol, medium) instead of using the intended per-task routing.

Existing configuration does not cover this surface

The current config schema provides:

• agents.default_subagent_model
• agents.default_subagent_reasoning_effort
• features.multi_agent_v2.expose_spawn_agent_model_overrides

Those settings apply to spawned subagents, not persistent user-visible App tasks created through create_thread.

The managed [models.new_thread] defaults in requirements.toml are static defaults, not a per-task recommendation and approval flow. I could not find a config.toml option for user-confirmed persistent-task routing.

An optional configuration shape could be something like:

[tasks]
model_routing = "confirm" # default | confirm

The exact name is not important; the important behavior is that confirm permits the agent to propose a model/reasoning pair but requires explicit user approval before task creation.

Acceptance criteria

• The agent may propose an exact supported model and reasoning effort for a new persistent task.
• The user can approve or edit both values before creation.
• No persistent task is created until confirmation is received.
• Batch creation presents a clear task-to-model/reasoning mapping.
• The approved values are passed explicitly to create_thread.
• The resulting task exposes the effective model and reasoning effort for auditability.
• If the approved pair is unavailable on the destination host, Codex asks for a new choice instead of silently falling back to the global default.
• The feature remains opt-in and preserves the user's final authority.

Additional information

This is distinct from subagent model routing:

• #31814 covered model overrides for Multi-Agent V2 subagent spawns.
• #32749 exposed model/reasoning overrides for that spawn_agent surface.
• #26767 added global subagent runtime defaults.

It is also narrower than #14923, which requested first-class cross-thread orchestration. Codex App now exposes create_thread; this request concerns the missing user-confirmed model/reasoning selection step when creating those persistent tasks.

Official config schema: https://developers.openai.com/codex/config-schema.json