Follow-ups, stopping, and cancelling
How follow-up prompts queue behind a running turn, what Stop does, which session states can be prompted again, and when to start over.
This page explains what happens when you send another prompt into a session that is already working, how to stop a turn, and which session states can still be prompted.
These actions require access to the specific session as well as the relevant workspace permission. On team-owned sessions, current owning-team membership is required even for Administrators and Owners. Private prompting also requires being the owning user or an explicit collaborator; Owner break-glass read access alone is not enough. See Multiplayer and sharing.
Sending a follow-up while a turn runs
You can send a prompt at any time. If the agent is still working on an earlier prompt, the new one is queued and runs after the current turn ends. Prompts run in the order you sent them:
Prompt 1 (processing) -> Prompt 2 (queued) -> Prompt 3 (queued)Every harness behaves the same way here: OpenCode, Claude Agent, and Codex hold follow-ups until the running turn finishes. A follow-up never interrupts the agent mid-turn.
A session may hold at most 50 unfinished (pending or processing) prompts; see Limits.
While a turn runs the composer placeholder changes to "Add a follow-up..." and the send button reads "Queue". Queued prompts stack above the composer in run order, each row marked with a clock icon. Each row has a "Remove queued prompt" button if you have the sessions.lifecycle permission. Removing a prompt puts its text back in the composer so you can edit it and send it again.
Unsent text, including a removed queued prompt restored into the composer, is saved in tab-scoped sessionStorage for your account and this session. It survives reloads in that tab, stays after a failed send, and clears after a successful send. Retrying an unchanged restored draft reuses an unconfirmed send's request identity; editing the text starts a new request identity. See The composer for storage and sign-out behavior.
A prompt sent while the sandbox is still booting waits until the sandbox is ready, then dispatches. The header shows the boot phase while you wait.
What a good follow-up contains
The agent already has the conversation and the workspace. A useful follow-up adds something it does not have: new information, a concrete correction, or the next step.
The failing test is tests/billing/test_proration.py::test_partial_month, not the
one you changed. Revert the edit to invoice_totals.py, then fix the rounding in
proration.py so the partial-month amount is rounded half-up to two decimals.
Run the billing tests before you finish.Avoid restating the original task or sending "any update?" style prompts. They queue behind the running turn and cost a full turn to answer. See writing prompts.
Stopping the current prompt
While a prompt is running, a red stop button appears in the composer for people with sessions.lifecycle. Its tooltip reads "Stop current prompt; queued prompts will continue", and that is exactly what it does:
- The running turn is interrupted.
- Anything already committed or written in the sandbox stays as it is.
- Prompts in the queue are not removed. The next queued prompt starts.
If you want the session to go quiet, remove the queued prompts first, then stop the running turn. Stopping is the right move when the agent is heading in the wrong direction and you want to redirect it with a corrected prompt.
Cancelled sessions
cancelled is a terminal session status. A cancelled session cannot be prompted again, and its sandbox is stopped.
The web app has no cancel control; in practice a session is cancelled only when a parent agent cancels it with cancel-child, and nested children are cancelled with it by default. For a session you started yourself, stop the turn and archive the session instead. See child sessions.
Resuming completed or failed sessions
A session whose last turn ended (completed) or whose last turn or boot failed (failed) still accepts prompts. Send a new prompt and the session picks up where it left off:
- After inactivity, the provider may restore a saved snapshot into a new sandbox, or resume a retained one in place. The header shows "Starting..." while the sandbox becomes ready.
- If the boot itself failed before a snapshot was saved, the next prompt can start fresh. If a saved recovery point is incompatible or its outcome is unknown, recovery is held rather than discarding it for a fresh checkout.
A prompt you type before the sandbox is ready waits and then runs. If the session failed because of the sandbox rather than the task, see recover from a failed run.
Archived sessions
Archiving files a session away; it does not delete anything. The composer's send button is disabled on an archived session. To prompt it again, restore it first, either with the "Unarchive" button on the session page or from Settings › Data Controls, which lists archived sessions and restores them. Restoring needs the sessions.lifecycle permission.
When to start a new session instead
A follow-up keeps the agent's context, the branch, and the sandbox state. That is usually what you want, but not always.
| Situation | Action |
|---|---|
| The agent is still on task and you have a correction or new detail. | Send a follow-up. |
| The agent is going the wrong way right now. | Stop the current prompt, then send a corrected follow-up. |
| You want the session to go idle. | Remove queued prompts, then stop the current prompt. |
| The session is completed or failed and the work should continue. | Send a follow-up. The sandbox restores saved state or resumes in place; after a boot failure with no saved snapshot it can start fresh. Incompatible or unknown recovery is held. |
| The session is archived. | Unarchive it (session page or Settings › Data Controls), then prompt. |
| The session is cancelled. | Start a new session. Cancelled is terminal. |
| The target changed (different feature, different repository, different base branch). | Start a new session. Follow-ups cannot change the repository set or the base branch. |
| The sandbox environment is corrupted (broken dependencies, half-applied migration, wrong global state). | Start a new session so the agent gets a clean workspace. |
| The conversation contains misleading context (a wrong assumption the agent keeps building on). | Start a new session with a corrected prompt rather than arguing it out in follow-ups. |
| You want the newest managed skills. | Start a new session; skills are pinned when a session is created. |
Troubleshooting
Next steps
The session page
A tour of the session page, from the header and timeline to the composer, the right sidebar, and the details overlay on small screens.
Reviewing changes and evidence
Where to find the session's diff, the screenshots and command output the agent produced, and how to review a session's pull request before approving it.