# Manual verification

Use an isolated temporary workspace and fresh Project. Do not use an existing personal conversation as a test fixture.

1. Start with Project selected. After the first turn begins, its identity becomes a static Project label and cannot be downgraded. Ordinary conversations retain their original behavior.
2. Assign a small task with a reusable finding. Verify an existing Worker can be reused or a new one created, the shared index is read first, a context document is written and indexed, and the Coordinator owns notes/archived updates.
3. Inspect the real model request and trajectory: Coordinator has control_workers and lacks replaced native delegation/control tools. Workers retain native tool composition.
4. While a Worker runs, send a Coordinator message through Enter. Verify timely steer admission rather than waiting behind the Worker. Explicit queue remains a later-turn operation. Operate a Worker directly, then verify the next Coordinator interaction observes its updated status/input.
5. Verify the Worker final AI message returns in full. No final AI message must still report a completion with an empty body. Coordinator processes results before normal same-turn closure. Stop remains interruptible without cascading into all Workers.
6. Check capsules at each participating message-segment bottom. Passive replies do not copy or move them; accepted updates and received results create a new seat. Delete a Worker and verify all its capsules and active Project queries disappear.
7. Check native sidebar title alignment, busy animation and ordinary rows. In a Worker composer, Back is transparent and the independent permission picker is absent. Coordinator and ordinary permission pickers remain.
8. Check Workers and notes.md native tabs, close/reopen behavior, a single empty state, newest-first timeline, collapsed input/output headings, Markdown on expansion and native artifact opening.
9. Restart while results are pending; verify no duplicate result delivery. Archive/restore the Project family without restoring independently archived children.

Record the actual runtime, model, task, visible result and limitations. Passing scripted fixtures alone is not real-model/UI acceptance.