feat(dispatch): remove a chosen lane from Tiled Dispatch

[Juliusolsson05]

Adds two commands so a Tiled Dispatch slot can be reclaimed at a position of your choosing.

The problem

Tiled Dispatch's size is a single count, and shrinking by count always drops the tail:

const lanes = next < tiled.lanes.length
  ? tiled.lanes.slice(0, next)   // <- always the last lanes
  : buildAutoLanes(prev, next, tiled.lanes)

With seven lanes open and the finished agent in lane 3, going 7 → 6 removes lane 7. You then re-select several lanes by hand to get back to the arrangement you wanted.

The obvious workaround doesn't work either. Closing the agent in lane 3 doesn't shrink the grid — clearTiledLaneSessions empties that lane and auto-fill re-homes another agent into it. Count stays 7.

So there was no way to shrink the tiled grid at a chosen position.

Two commands, not one flag

Command	Agent	Lane
Close Agent and Remove Lane	closed	removed, count −1
Remove Lane	keeps running, stays in the index	removed, count −1

The default is destructive, and a command that sometimes ends a session is the kind of thing that surprises you at speed. The titles carry the difference: Close is this catalog's established verb for ending a session (Close Focused Session, Close Tab, Close Old Agents), so the destructive one leads with it. Putting "Remove Lane" first would bury the irreversible half, and "Remove Agent" was rejected outright — it reads non-destructive and isn't.

Close first, splice second

closeSession runs its own confirmation for irreversible closes. Splicing the lane first would shrink the grid while you were still deciding, and a declined confirm would leave the layout changed with the agent alive — the worst of both.

That required closeSession to report whether it actually closed, so it now returns Promise<boolean>. The two orchestration call sites adapt locally to their Promise<void> contract rather than widening a cross-process type for one caller.

Details worth a look

• Focus is not a plain clamp. A lane removed before the focused one shifts indices down, so holding focusedLane constant would silently move focus to the next agent. Adjusted explicitly, then clamped.
• Lane 0 is a real agent lane — it just has no mini-list, because it's selected from the full index. Removing it promotes lane 1 into that role. No special-casing, but the descriptions say so rather than pretending the grid is homogeneous.
• ratios are dropped on removal, matching setTiledLaneCount — they're positional and a removal invalidates them.

Tests

tiledLaneRemoval.test.ts — 8 cases over the pure removeLaneFromTiled: middle removal keeps neighbours, focus follows its agent when an earlier lane goes, clamps when the last lane was focused, refuses at the floor and on bad indices, always drops ratios, never mutates its input.

Verification

• npx tsc -b → exit 0
• NODE_ENV=test npx vitest run → 255 files, 1758 tests passing
• npm run check:keybindings → OK
• npm run test:contract → satisfied

Not verified: appearance and the live close-confirm flow — I haven't run the app.

🤖 Generated with Claude Code