Skip to main content
A watcher is a read-only check that Kepler runs in the background on a schedule. It costs nothing while nothing changes. When it fires, Kepler wakes the session that set it, looks into what happened, and comes back with a suggestion. Nothing is changed without you.

​
Arm a watcher

Ask in plain words, in any session:
Kepler turns the request into a check it can run without writing anything, confirms what it will watch and how often, and arms it. The watcher is listed under Watching in the session’s This thread panel, opened from the indicator at the top right of the conversation, with a stop control beside it.

​
When it fires

The session wakes. A card appears in the thread saying what the watcher saw, and Kepler follows it with a short investigation and a suggestion: what changed, what it likely means, what you might do next. If you have a channel connected and you are away from your desk, the alert can reach your phone too. See Chat from your phone. Because the watcher woke the session on its own, Kepler stays on the safe side. It reads and reports. Anything it proposes follows the session’s posture, so in Observe and Assist nothing changes without your approval.

​
Watchers are read-only, always

A watcher can only run commands that read. That holds in every posture, including Yolo. If the check Kepler drafts needs a command it does not recognise, it stops and asks you to trust that command as read-only, the same card you see in a normal session. The Run once option is never offered for a watcher, because a watcher would run it again and again.

​
Watchers that judge in plain words

Most watchers compare plain output: a pod is not Ready, a number crosses a line, a job exits. Some conditions need judgement, such as “wake me if the logs show anything like a connection storm”. For those, each new output is read by a model that answers one question: does this match?
  • With a decision model picked, the watcher asks it. Each check takes about a second and spends no tokens, so the watcher can check as often as it runs. It wakes you only on a yes the model is sure of.
  • Without one, the watcher uses your worker model, with a token budget for each watcher and a minimum time between checks.
Either way, output that has not changed since the last check is not judged again, and long output is read in full, page by page. A check that fails is skipped and tried again on the next tick. Kepler never switches to another model.

​
See and stop watchers

  • In the session: open This thread from the top right of the conversation. Watching lists the watchers armed there. Press the stop control to end one.
  • Everywhere: Settings > Watchers lists every watcher across every session, with counts for how many are watching, how many checks have run, how many alerts were sent, and how many have ended. Open one for its checks and signals.
You can also just ask:

​
Example watchers by situation

​
Limits worth knowing

  • Kepler has to be running for watchers to check. Under Settings > General > While you are away, switch off Kepler stops when you close the window so watchers keep going with the window closed. On macOS, allow Kepler under System Settings > General > Login Items > Allow in the Background.
  • On Windows in this release, Kepler does not run in the background, so watchers pause when the window is closed.
  • A watcher belongs to the session that set it and wakes that session. Open that session to see the card.

​
Where to go next

Chat from your phone

Let a fired watcher reach you when you are away.

Postures

Why a woken session still asks before it acts.

Everyday workflows

The investigation that usually comes before a watcher.

Settings reference

The Watchers section in Settings.