# Capture environment and reproduction
Both sides: web client in Linux Chromium, 1280x800 CSS pixels, device scale 2, light theme, independent dev servers built from their respective worktrees. Each server started with T3CODE_DEV_SEED=0 and created its own fresh database. Synthetic empty threads in the default project; same Claude provider, selected configuration, permissions, and prompts. Default follow-up behavior is Steer.

Base: f86e21f7cce7324119dadf61064965d83c539455
Head: 55138fa36c48a19015e9b3a5caa48ce35ce4b498

Both recordings show the full interaction from submitting the original prompt. Videos preserve real timing and are cropped identically to the chat panel, excluding the unrelated thread sidebar. Private model fields and filesystem path cells are blacked out by the privacy-only DOM style driver; the composer model selector is hidden. No conversation text or behavior is replaced.

Primary case: startTask, wait for Running python3, sendContext, wait for completion, inspectResult, stop recording.
Promotion case: startTask, wait for Running python3, queueOrdinary, sendContext with queue=true, promoteContext, wait for completion, inspectResult, stop recording. Control+Enter explicitly queues both follow-ups so that the same Send now control is reachable on base and head.

Original prompt:
Use the Bash tool to run `python3 -c 'import time; time.sleep(40); print(1)'` in the foreground, then reply with the word done.

Ordinary queued prompt:
Reply with the word queued.

No tests are used as evidence. AI-assisted; verified with the tests and tools listed above.

Observed: Base primary and promotion cases lose the original done reply; context becomes Steer. Head primary queues context until done. Head promotion moves context before the ordinary queued prompt without steering the active task.