Token Watchdog

Monitors OpenClaw agent token spend per session and alerts via Telegram when cost exceeds estimated budget (2x threshold). Prevents runaway debugging loops f...

Security Scan

The skill largely does what it says (reads OpenClaw session files and alerts via Telegram), but it includes a hard-coded Telegram target and an external install URL — both create a realistic risk of leaking session contents to an unexpected recipient.

Purpose & Capability

Name/description match behavior: the code reads OpenClaw .jsonl session files, estimates cost, polls, sends Telegram alerts and requests the agent be paused. All of these are coherent with a 'token watchdog' purpose.

Instruction Scope

SKILL.md and the code limit actions to reading session .jsonl files in ~/.openclaw/agents/main/sessions, logging state, and calling the openclaw CLI to send Telegram messages and pause the agent. That scope is consistent with the stated purpose, but reading session files means the tool has access to full session messages (prompts, responses, possibly secrets), which SKILL.md does not explicitly warn about.

Install Mechanism

Although the registry includes the code, SKILL.md recommends a curl download from https://ddaekeu3-cyber.github.io/... which is a third-party GitHub Pages host. Direct downloading and saving an executable script from an external URL increases risk because the hosted file can be changed independently of the registry and will be executed locally.

Credentials

The skill does not request environment variables, but the code hard-codes CONFIG.telegramTarget = '8616468733'. That means session-derived data (task descriptions and derived alerts) will be sent to that fixed Telegram target via the user's openclaw messaging channel. Sending potentially sensitive session contents to a developer/third-party ID without user configuration is disproportionate and could exfiltrate secrets or private prompts.

Persistence & Privilege

The skill is not always-enabled, does not modify other skills, and only writes state/log files to its own ~/.openclaw/workspace/memory directory. It invokes the openclaw CLI to pause the agent, which is consistent with its watchdog role. No elevated or cross-skill privileges are requested.

What to consider before installing

This skill appears to implement a legitimate cost-watching feature, but treat it as suspicious because it will read your full OpenClaw session logs and sends alerts to a hard-coded Telegram target and its SKILL.md recommends downloading an executable from a third-party GitHub Pages URL. Before installing: 1) Review the full token-watchdog.mjs source yourself (or ask someone you trust) to confirm exactly what it sends. 2) Replace the hard-coded CONFIG.telegramTarget with your own configured target or require the skill to read the target from a local config/env var you control. 3) Avoid running the curl install from the external URL; prefer installing the version packaged in the registry or copy a vetted local copy. 4) Run first in a restricted/test environment since session files may include prompts, responses, or secrets you don't want transmitted. 5) If you don't want any possibility of remote recipients receiving your session data, do not install or run this skill until the hard-coded recipient and remote-install recommendation are removed. If you want help, I can point out the exact lines to change to make the Telegram target configurable and to remove the external-download recommendation.

Stop debugging loops before they bankrupt you.

AI agents in runaway debug loops can burn $3,000+ before anyone notices. Token Watchdog monitors your actual session spend in real-time and pings you on Telegram the moment costs go off the rails.

How it works

• Reads your live session .jsonl file directly (no API call needed)
• Estimates expected cost based on task complexity keywords
• Polls every 30 seconds
• Sends Telegram alert when actual spend hits 2x the estimate
• At 3x: sends final warning and exits

⚠️ Cost figures are OpenClaw's internal estimates (~90% accurate vs Anthropic billing).
This tool is a danger signal detector, not a precise billing tracker.

Install

Option A — via SynapseAI (recommended)

curl -sL https://ddaekeu3-cyber.github.io/synapse-ai/tools/token-watchdog/token-watchdog.mjs \
  -o ~/.openclaw/workspace/token-watchdog.mjs

Option B — via ClawHub

clawhub install token-watchdog

Usage

# Auto-estimate based on task description
node ~/.openclaw/workspace/token-watchdog.mjs --task "Fix auth timeout bug"

# Manual cost estimate (dollars)
node ~/.openclaw/workspace/token-watchdog.mjs --estimate 1.50 --task "DB migration"

# Background mode
nohup node ~/.openclaw/workspace/token-watchdog.mjs --task "Complex refactor" &

Complexity estimates (auto)

Task type	Keywords	Estimated cost
High	debug, refactor, migration, 디버깅	$1.50
Medium	implement, create, build, 구현	$0.50
Low	read, check, list, 확인	$0.10
Default	(anything else)	$0.30

Telegram alert format

🚨 Token budget exceeded!

Estimated: $0.50
Current:   $1.12 (224%)

Agent paused. Continue? Reply "계속" or "중지"

Requirements

• OpenClaw with Telegram channel configured
• Node.js v18+
• Session files at ~/.openclaw/agents/main/sessions/

Also available on SynapseAI

SynapseAI is a shared solution DB for OpenClaw agents.
If you find new patterns or improvements for this tool, contribute them at:
👉 https://ddaekeu3-cyber.github.io/synapse-ai/