[FEATURE] Agent view: /new panel to set model, effort, name, permission mode and isolation before dispatching a session

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet (closest ones below were auto-closed as stale, not declined)
• [x] This is a single feature request (not multiple features)

Problem Statement

In agent view (claude agents), the dispatch input can only change the model before a session starts (/model opus). Everything else about a new session has to be decided outside the view or fixed after dispatch:

• Effort: /effort in the dispatch input shows attach to a session to run it. Per-session effort is only possible via claude --bg --effort … in a shell, via claude agents --effort … (applies to every dispatch), or via a subagent file.
• Name: sessions are auto-named from the prompt; the only option is Ctrl+R after the fact.
• Permission mode / isolation (worktree vs in-place): same — flags on the shell command or a subagent file, not from the view.
• Discoverability: nothing in the view tells you that /model works in the dispatch input but /effort doesn't. #67078 was closed with "option A is a gap in our docs".

The result: you dispatch with defaults, attach, change settings, and the first turn already ran with the wrong model/effort — or you leave the view and go back to the shell, which defeats the purpose of having a single dispatch surface.

Proposed Solution

A /new command in the agent view dispatch input that opens a small pre-launch panel — same interaction style as the /diff panel — with fields pre-filled from the current dispatch defaults:

╭─ New session ──────────────────────────────╮
│ Name        fix-quotes-modal               │
│ Model       ▸ sonnet   opus   haiku        │
│ Effort      low  ▸ medium  high  xhigh     │
│ Permission  ▸ default  acceptEdits  plan   │
│ Isolation   ▸ worktree  in-place           │
│ Directory   ~/projects/app                 │
│ Prompt      ______________________________ │
╰─ Enter dispatch · Tab next · Esc cancel ───╯

Key properties:

1. Opt-in, zero added friction. Typing a prompt and pressing Enter keeps working exactly as today. Only /new opens the panel, and Enter on an untouched panel dispatches with the defaults.
2. Inline form for power users, reusing the existing CLI flag names so there's no new syntax to learn:

/new --model opus --effort high --name fix-login -- refactor the login flow

1. Validated choices. The panel only offers models/effort levels/permission modes that are valid for the account and model, instead of failing after dispatch.
2. Optional: save as preset. /new fix loads a saved combination. Today the closest thing is a subagent file per combination (model, effort, permissionMode, isolation in frontmatter), which works but means one file per settings combo and mixes "who the agent is" with "how this run is configured".

Alternative Solutions

• Accept more commands in the dispatch input, the same way /model already works: /effort high, /name fix-login, /permission-mode plan, each setting the next dispatch. Smaller change, but still undiscoverable and multi-step.
• Current workarounds: claude --bg --name … --model … --effort … -w from a shell; claude agents --model … --effort … --permission-mode … (applies to every dispatch in that view); a dedicated subagent file per combination.

Priority

Medium - Would be very helpful

Feature Category

Interactive mode (TUI)

Use Case Example

A typical loop: I review a screen of an app, find 3–4 small defects, and dispatch one background session per fix while I keep reviewing.

1. Defect A is a trivial layout fix → /new, pick sonnet + medium, name fix-modal-width, in-place, Enter.
2. Defect B is a calculation bug that needs a test first → /new, pick opus + high, name fix-total-calc, worktree, Enter.
3. Both rows appear already named and running with the right settings — no attaching to switch model/effort, no throwaway first turn, no shell round trip.

Additional Context

Related requests (all closed by the stale bot, none declined):

• #74735 — Allow opening a new empty chat directly from the Agents view (to pick model / run commands before the first prompt)
• #62211 — Allow per-session model and effort overrides from claude agents dispatch input
• #67078 — Per-session model at dispatch time (closed as completed: /model in the dispatch input; effort/name not covered)
• #62853 — Agent view: slash commands treated as prompts

Checked against the agent view docs (https://code.claude.com/docs/en/agent-view) on Claude Code 2.1.273.