The OpenClaw 2026.11.5 release, the current stable version, changes what happens to a message that arrives while the agent is still working on an earlier one. The release notes say such messages now “steer compatible work or queue quietly instead of requiring a manual resend.” The same line of the notes also describes personal and corporate security presets. The project’s own documentation states that those presets were removed, and a config file that still contains the setting has to be cleaned up before the Gateway will start. Both points matter to a solo operator or a small business that talks to one OpenClaw Gateway through Telegram, WhatsApp, Discord or the web interface.
The problem: a second message during a running task
Pull request 185477, merged on October 27, 2026, states the problem in one sentence: messages sent to a busy session could be refused or could require a manual /followup resend. Under the new behavior, a message that the running task can accept is steered into that task. A message that cannot be steered is queued for a later turn that belongs to the person who sent it. The pull request says this happens without automatic queue-status messages and without a resend.
For a business owner who sends a correction shortly after a request, the practical effect is that the correction is not lost and does not produce an error reply. It either reaches the task that is running or runs when that task ends.
How steering works by default
The command queue documentation lists the defaults that apply on every inbound channel when nothing is configured: queue mode steer, collaborative steering on, a 500 millisecond debounce, a cap of 20 queued messages per session, and a drop policy of summarize, which keeps compact summaries of the oldest entries when the cap is reached.
The steering queue documentation describes the limits of steering. A tool call that is already running is not interrupted. It finishes, and the new message becomes visible to the model after the tool results and before the next model decision. If the agent had planned several tool calls in sequence, the ones that had not started are skipped and each receives the result Skipped to process an incoming message. so the model knows they did not run. Tool calls launched in parallel are never skipped for steering.
The web interface, called the Control UI, shows the state of a waiting message. According to the same page, a message queued behind the current task is labeled “Queued behind the current turn.” A waiting steer can show a Send now link that interrupts supported running tools without ending the task. When a tool cannot be stopped safely, the interface shows “Steering…” instead.
Four queue modes
The queue documentation lists four modes, set with the /queue command for one session or with messages.queue.mode in the config:
steer: send the message into the running task. This is the default.followup: do not steer. Run each message as its own turn after the current one ends.collect: do not steer. Combine compatible waiting messages into a single later turn.interrupt: stop the running task and run the newest message.
Two channel details apply. On Discord and Telegram, collect behaves like followup, so each message keeps its own turn. In Discord server channels and threads, an ordinary message from a different person always waits as a separate turn, even when collaborative steering is on.
Shared chats: whose permissions a steered message uses
The setting that deserves attention in a shared chat is messages.queue.allowCollaborativeSteering, which the messages configuration reference lists as true by default. When one person’s message steers a task that another person started, the task keeps the identity and permissions of the person who started it. The steering documentation states the consequence directly: the second person can influence actions using permissions the first person has and the second does not.
A solo operator who is the only person in the chat is not affected. A small team, or a business that lets a client or contractor into a group chat with the agent, has a decision to make. Setting allowCollaborativeSteering to false keeps each participant’s work separate. Another person’s message then waits quietly and runs as a new turn under that person’s own identity and current policy. A stricter option, collaborativeSteeringFallback: "reject", refuses the message with a notice and requires an explicit /followup.
The steering documentation adds one caution about stopping work. A plain-language request to stop, sent as a normal message, is queued or steered like any other message and does not confirm that the task stopped. The Stop control and authorized stop commands use a separate cancellation path that reports the actual outcome.
The security presets in the notes are not in the release
The release notes for 2026.11.5 say that “personal and corporate security presets provide coherent defaults.” Pull request 185477 did add a security.preset setting with those two values. One day later, on October 28, pull request 186381 removed it. That pull request says the removal lets execution, filesystem, sandbox, steering and conversation-access policy be configured independently, and that the quiet queue behavior was kept.
The Doctor config migrations page describes the current state: the security.preset field has been removed, a configuration that contains personal or corporate must drop that field before startup, and openclaw doctor --fix removes it through its existing backed-up repair. Doctor does not expand a removed preset into individual settings. An operator who wanted the restrictions a preset implied has to set the sandbox, approval and steering controls directly.
An operator who reads the release notes and adds security.preset to a config file would therefore add a field the Gateway does not accept. Earlier coverage of credential scoping in OpenClaw describes controls that do exist in current releases.
Checklist for a small operation
- After updating to 2026.11.5, send a second message during a long task and confirm it is steered or queued without an error reply.
- Decide who else can message the agent. If anyone with fewer permissions shares a chat with it, consider setting
messages.queue.allowCollaborativeSteeringtofalse. - Use
/followupfor a request that is unrelated to the running task, so it runs as its own turn. - Use the Stop control to stop a task. Do not rely on a chat message that asks the agent to stop.
- Search the config for
security.preset. If it is present, runopenclaw doctor --fixand set any needed restrictions directly.
The 2026.11.5 release also keeps queued messages across Gateway restarts, which this site covered in its report on restart-safe replies, and it changed how failed updates are handled, covered in the report on update rollback. Operators who run the agent in group chats can find setup details in the guides to OpenClaw chat apps and Telegram groups and topics. The AI agent harness guide compares how other agent systems handle input during a running task.
Sources
- OpenClaw v2026.11.5 release on GitHub
- OpenClaw pull request 185477: make busy-session messages automatic
- OpenClaw pull request 186381: remove personal and corporate security presets
- OpenClaw documentation: Command queue
- OpenClaw documentation: Steering queue
- OpenClaw documentation: Configuration, messages and talk
- OpenClaw documentation: Doctor config migrations

