# v1.5.0 — Run agents from the desktop app

The Agent Room stops being a viewer: it can now actually launch agents —
**confirmation-first**, exactly as the AGENT_ROOM design doc always
specified (preview → confirm → run). Nothing ever runs without your click.

## `/run <task>` and Run here

- New **`/run <task>`** composer command: builds
  `relay-baton run "<task>" --primary <selected agent> --pretty` — the
  Codex/Claude header toggle now picks who actually goes first.
- **`/plan`, `/execute`, `/handoff`** build real launching commands too
  (handoff dropped `--no-run`).
- Every one of these opens the existing confirm modal showing the **exact
  command**; a new **Run here** button executes it via the bundled sidecar
  CLI (your local agent logins and quota — relay-baton still never calls
  provider APIs). Copy-to-terminal stays available.

## Live run card

Agent output streams into a live timeline card (`--pretty` formatting):
tool calls, assistant text, and the closing token/cost receipt appear as
they happen. A **Cancel** button kills the process; on exit the card gets a
green `done` or red `exit N` chip and the dashboard panels refresh. One run
at a time; output capped at the last 400 lines; the card survives timeline
reloads like any v1.4.1 result card.

## Under the hood / compatibility

- Tauri capability additions: scoped `shell:allow-spawn` for the sidecar and
  `shell:allow-kill` for cancel (execute-only capabilities would have denied
  streaming runs).
- The GUI remains business-logic-free: it spawns the CLI and displays lines.
  All gates (handoff quality, token diet, redaction, guardrails) run in the
  CLI exactly as from a terminal.
- Guide/tooltip strings updated in all 4 GUI locales — the app is now a
  "confirmation-first dashboard", not a read-only one.
- CLI behavior unchanged; versions bumped in lockstep.