[FEATURE] Per-session model selection in `claude agents` dispatch input

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

The claude agents dispatch input has no way to select a model for an individual session before dispatching it. The only available model controls are:

• Global default — claude agents --model <model> or the model setting, applied to every session dispatched in that agent view instance.
• Post-dispatch override — attach to a running session with Enter/→, run /model, select a model, then detach with ←.

This means there is no way to dispatch multiple sessions with different models without either re-launching agent view with a different --model flag or attaching to each session after the fact to change the model. When running heterogeneous workloads — e.g. one Opus session for a complex refactor and one Sonnet session for a lighter documentation task — this adds significant friction.

Related: issue #43956 touched on per-session model selection for dispatch but was closed without resolution and was broadly scoped (Cowork, mobile, Dispatch). This request is narrowly scoped to the TUI dispatch input in claude agents.

Proposed Solution

Add a keyboard shortcut in the dispatch input (e.g. Ctrl+M or Alt+M) that opens a model picker — similar to the existing /model picker — before the session is dispatched. The selected model would apply only to that single dispatched session, leaving the global dispatch default unchanged.

The chosen model could be shown as a small indicator next to the dispatch input before pressing Enter:

[opus] Refactor the auth module and add tests

After dispatching, the indicator clears and the next prompt reverts to the global default.

Alternative Solutions

• Inline syntax — a prefix like /model:opus <task> typed directly in the dispatch input. Simpler to implement and consistent with how /model already works in session, but less discoverable.
• --model flag in the dispatch input text — parse --model <name> from the prompt before stripping it. Consistent with the CLI UX but fragile if the task description happens to contain --model literally.
• claude --bg --model <model> "<task>" — already works from the shell, but requires leaving agent view entirely, which defeats the purpose of the dispatch input.

Priority

Medium - Would be very helpful

Feature Category

Interactive mode (TUI)

Use Case Example

1. I open claude agents to dispatch several parallel tasks.
2. One task is a complex multi-file refactor — I want Opus for it.
3. Another task is updating inline doc comments — Sonnet is sufficient and cheaper.
4. Currently I have to: dispatch both with the global default → attach to each → /model → detach. That's 6–8 extra keystrokes per session just to set the model.
5. With a pre-dispatch picker (Ctrl+M → select model → Enter to dispatch), I set the model in one step before typing the task, with no context switching.

Additional Context

The v2.1.172 changelog references an "agent dispatch model picker" and bugfixes around it (org-disabled models appearing, availableModels not being enforced). That picker appears to be /model autocomplete suggestions within the dispatch input, which affects the global default — not a per-session override. This request is for the complementary per-session UX that is still missing.