AiWatchDog, monitor your agents, increase productivity

Description and use

AiWatchDog offers monitoring and stuck detection for AI coding agents using a weighted ensemble of signals, enabling users to receive alerts only when intervention is truly needed.

Problem

When running AI coding agents, users often return after a break to find the terminal idle without knowing if the agent is working, waiting, or stuck, leading to wasted time and uncertainty.

Outcome

AiWatchDog reduces idle time by detecting stuck agents with a weighted ensemble, providing clear alerts and traceability, and enabling users to intervene only when necessary.

Workflow

• 1. Wrap the agent command with watchdog wrap -- <agent> run "<task>". 2. Watchdog streams execution over WebSocket to the Live Command Center. 3. The detection engine evaluates multiple signals (heartbeat, tool activity, errors, commands, file changes, process state) with assigned weights. 4. If the summed confidence exceeds 60, Watchdog triggers a stuck alert and shows the exact event causing it. 5. Users can review the trace view to understand why the agent was considered stuck and take corrective action.

Evidence

I run a lot of AI coding agents — Kilo, Claude Code, Codex, Aider, my own scripts. And I kept hitting the same wall: I start a task, walk away, come back 20 minutes later, and stare at a terminal with no idea whether the agent is crushing it, waiting for me, or wedged in an infinite retry loop burning tokens.

Source

AI Watchdog — real-time visibility and intervention for AI coding agents. It ingests everything your agent does, decides with a multi-signal confidence model whether it's working / waiting / genuinely stuck, and notifies you only when you actually matter to the outcome.

Source

The naive rule — no output for 5 minutes → STUCK — is wrong, and I refused to ship it. Agents legitimately go silent while compiling, installing dependencies, running Docker, or grinding through a long test suite.

Source

Watchdog runs a weighted ensemble of independent detectors, each producing an explainable signal: ... Confidence is the clamped sum of triggered weights — and every contributing weight is traceable, so the UI can always answer "why does it think you're stuck?"

Source

Live Command Center — every execution streaming over WebSocket

Source

One command connects any agent: watchdog wrap -- kilo run "fix the failing test" Kilo, Claude Code, Codex, Aider, OpenCode, npm test, your own script.

Source