Steer a running Claude run

When Claude Code is halfway through a task and you think of something else — “also check the logs”, “use the staging database instead” — you have two choices for the message you send from Telegram:

• queue (the default) — the message waits until Claude’s current turn finishes, then runs as the next turn.
• steer — the message goes straight into the run that is already working. Claude reads it at its next step (the next time a tool finishes) and folds it into the answer it is producing.

Steer is Claude Code only and needs a live session, which means a permission mode is set (/planmode or /config → Permission mode) and [watchdog] live_sessions is on (the default). Other engines always queue.

Steer one message#

Send /steer followed by your text while a run is working:

/steer also tell me the hostname

You’ll get ↪️ Steered into the current run. as a reply. Once Claude has read it, the run’s progress message shows a ↪️ steer received: also tell me the hostname row, and the final answer covers both the original request and the steer.

If Claude has already finished its turn when the message arrives (the session is idle), there is nothing to steer: the message simply runs as the next turn and its answer replies to it — no Steered acknowledgement.

If the steer arrives after Claude’s last tool call (it is already writing its answer), Claude finishes the first answer and then answers the steer as its next turn — you get a second reply, threaded under your /steer message. Nothing is lost either way.

Queue one message#

/queue <text> sends a message the normal way even when steer is the default:

/queue when you're done, write a summary

Make steer the default#

• Chat — send /steer with no text, or open /config → ↪️ Follow-up and pick Steer. Every plain message you send while Claude is working is steered.
• Forum topic — send /steer inside the topic to set it for that topic only.
• Everywhere — set the default in untether.toml (applies without a restart when watch_config = true):

[transports.telegram]
followup_mode = "steer"

Send /queue (no text) or pick Queue in /config to switch back. Clear override in /config returns the chat to the untether.toml default. In a group, only admins can change the default with a bare /steer or /queue; anyone allowed to use the bot can still send /steer <text> or /queue <text>.

Voice notes follow the default: a transcribed voice note sent to a steer chat is steered.

What always queues#

• Files, photos and albums (they need saving before the prompt exists)
• Forwarded messages
• Commands
• A message sent while Claude is asking you a question — a plain-text reply answers the question instead

When a steer can’t be delivered#

The message is queued as normal and Untether tells you why:

With steer as the chat default you only see these while something is actually running, once per run — a steer chat with nothing running just starts a normal run. An explicit /steer <text> tells you too, with two exceptions: while Claude is waiting on a question (AskUserQuestion) the text is queued without a note, and if your model or mode changed since the session started, the queued run posts its own notice when it restarts the session with the new settings.

When a follow-up didn’t run#

In either mode, one notice can appear on a follow-up after the run ends:

Cancelling a run doesn’t drop a follow-up already queued behind it: it runs next, so you don’t need to send it again.

Related#

• Inline settings menu — the /config Follow-up page
• Interactive approval — a steer sent while an approval is pending is read after the tool is approved or denied
• Commands reference
• Claude runner: live sessions