OpenClaw not responding

An OpenClaw agent that sits there is the most common complaint after cost, and the diagnosis is almost always one of eight things. Upstream now ships a triage flow that gets you to a diagnosis in about two minutes. Here it is in plain language, followed by the causes that fill the forums, and what changes when someone else runs the Gateway for you.

The first sixty seconds

Run the ladder in order: triage, status, gateway probe, gateway status, doctor, channels probe, then follow the logs. Healthy reads as reachable, runtime running, connectivity probe ok, no blocking doctor errors, and every channel reporting a live transport. The triage command writes a sanitised diagnosis you can hand to an agent, a forum or support.

The usual suspects

• Pairing pending: an unknown sender messaged it and nobody approved the request, so it ignores them by design. The production version is a Gateway stuck in a pairing-required reconnect loop.
• Mention gating in a group: it drops group messages until it is mentioned, and says so in the logs.
• An allowlist mismatch: the sender, the channel or the model is filtered by policy. A model that is not on the allowlist makes cron jobs and subagents fail silently.
• The tool profile: a minimal profile allows almost nothing, so it answers but cannot act.
• The provider: a rate limit, an expired credential such as a Google OAuth testing token, or a refusal, visible in the logs as a 429 or a 403.
• A delivery failure that looks like a job failure: a cron job marked failed with a 401 on the Telegram or gateway side while the work itself succeeded.
• A plugin or channel disabled after an upgrade, so the assistant is running and the door it used to answer on is closed.
• The Gateway itself: down, restarting, or killed for memory on a small host.

It just sat there for 2 hours doing nothing.

Slow is not the same as stuck

A reply that arrives minutes late is usually a busy model, a long context being re-read, or a heartbeat turn occupying the session. The OpenClaw heartbeat explains that last one, and why it is also the reason an idle agent is never quite idle.

What changes on a managed host

The ladder above assumes a shell on the box. On Diali there is no box: the activity feed shows provisioning, updates and channel events, the dashboard has pause, restart and reinstall, and the Gateway is watched for you. Pairing, allowlists and the tool profile are still yours to set, because they are policy, not plumbing. Hosted OpenClaw on Diali describes the rest.

Before you post in a forum

• Run triage and attach its output; it is the first thing every helper will ask for.
• Check pairing and mention gating before you reconnect anything.
• Read the logs for the three signatures: mention required, pairing request, blocked.
• If it is slow rather than silent, look at the provider and the heartbeat first.

And if the word Gateway in all of this is the part you never understood, The OpenClaw Gateway, explained is the explanation.