Watchdogs

A watchdog is a per-session supervisor that monitors a coding session and intervenes when the agent gets stuck — answering questions, approving permissions, or escalating to you when it can't handle the situation.

Creating a watchdog

Right-click a session card and select Add Watchdog, or use the watchdog toggle in the session header. Provide natural-language instructions that tell the watchdog how to handle decisions — for example, "approve all file edits but escalate any git push."

Toggle Watching on/off to activate or pause the watchdog.

What watchdogs detect

Watchdogs monitor two types of stuck states:

Pending questions

When the agent asks an AskUserQuestion (e.g. "Which approach should I take?"), the watchdog detects it via hook sidecar files and evaluates the question against your instructions.

Permission gates

When the agent hits a tool permission prompt (e.g. "Allow bash command?"), the watchdog can auto-approve or deny based on your instructions.

The watchdog also does a periodic conversation review — it checks the session's recent activity against your instructions and can intervene with a nudge if the agent appears to be going off-track. This runs on a cooldown and only when new messages have appeared since the last review.

How watchdogs respond

When a stuck state is detected, the watchdog makes an LLM call to evaluate the situation against your instructions and decides to:

• Select an option — answer a multiple-choice question
• Approve — grant a permission request
• Escalate — surface the situation to you

The watchdog interacts with the session by sending terminal keystrokes, mimicking how a human would respond.

Escalation

After 10 retries or 2 convergence failures, the watchdog escalates to you with a "Watchdog stuck" banner and a Take action button in the session card.

Argus exclusion

Sessions that are children of an Argus manager skip watchdog monitoring entirely — their Argus parent handles supervision directly.