Per-run model choice on the subagent authorization card

Overview

When a subagent tool call needs approval (the tool authorization gate is in ask mode), the Authorization required card now carries a model picker. Before tapping Allow once / Allow for session / Always allow, the user can choose which model executes this delegated run. The choice is a one-off: nothing is persisted on the chat, the agent, or the project — the next subagent call uses the chat's model again (or the agent's pinned model).

Usage

The picker appears only on subagent approval cards. It lists the same union the main chat model picker shows:

• project-defined models (settings.models)

• the per-provider live catalog (GET /api/ai/models/live)
  deduplicated by (provider, modelId), sorted by provider then id. The first row is a "(chat default)" entry that keeps the run on the chat's current model (or the agent's model pin, when the called agent has one). The trigger shows (chat default) <modelId> until the user picks an override. Below it, a thinking dropdown offers the picked model's presets plus an "Inherit chat thinking" first row (the same ThinkingSelectField the Agents editor uses).

  The card's picker also shows the Pinned and Recent bookmark sections exactly like the chat top bar, using the same source of truth: pins live in localStorage (per project), and the recent list is the server-backed model_recent SQLite cache, refreshed when the sheet opens. Picking a model on the card records it as recently used; pinning/unpinning works right on the card.

  Pick a model (and optionally a thinking level), then choose any non-deny action. The decision payload carries

{
  "modelOverride": { "providerId": "openrouter", "modelId": "anthropic/claude-sonnet-4" },
  "thinkingLevel": "medium"
}

and the nested call runs on that model — with that reasoning effort — for this invocation only. Deny still ignores the pick.

Precedence

Model resolution for the delegated run, most specific first:

1. Authorization-card pick — the user explicitly approved this run on a model, so it wins over everything else.

2. Agent model pin — a named agent's modelId (Settings → Project → Agents) applies when no card pick was made.

3. Chat model — the fallback for a generic subagent with no pin.

   Thinking level follows the same precedence: the card's pick wins, then the agent's thinkingLevel, then the chat's inherited level.

Regression check

With debug Chrome available at http://127.0.0.1:9222 (or CDP_URL), run:

node scripts/test-auth-model-picker.cjs

The test mounts the real AuthModelPicker on an isolated about:blank tab with an in-memory localStorage shim, seeds a pinned model and a server-backed recent entry, opens the sheet, and asserts both bookmark sections render. It does not modify app data.

Related

• Tool authorization — the gate the card belongs to.

• Agents — named subagent personas and their model pins.

• Model picker — the main picker whose data feeds this one.