Operations · Intermediate

Verified

Automated verification

  • Passed every automated gate: schema, uniqueness, HTTPS and live source link, prompt-injection and safety scan, approval-step rule.
  • Drafted by deepseek-flash via DeepSeek API.
  • Second-model review: kimi-code/k3.1-pro (no critical or high findings).
  • Verified .
  • Automated checks and AI review only, not human review.
How verification works

Review coding agent command logs for risky entries

Flag and summarize shell commands an agent ran, ready for human review

  1. On a scheduleDaily command log review
  2. ThenClassify commands by risk
  3. ResultDigest of flagged commands

Each day the bot reads the command history left behind by an automated coding agent, groups entries by session and marks anything that looks destructive, network-facing or outside the project directory. It writes a short digest of what ran and what the agent appeared to be doing, lists flagged entries with their exact command text, and leaves any cleanup decision to the person reviewing.

Quick answers

What does it do?

Flag and summarize shell commands an agent ran, ready for human review. Result: Digest of flagged commands.

What does it need?

Tools: GitHub, Slack, Terminal, Notes. Intermediate: some setup.

How do I run it?

On a schedule: Daily command log review. Copy the prompt below into a bot in BotBento.

Prompt to start with

BotBentoRecommended
You audit the shell activity of an automated coding agent. Read the command log for the past day, group entries by session, and note the apparent goal of each session. Classify every command as routine, needs a second look, or risky, judging by whether it deletes or overwrites files, changes permissions, opens network connections, installs software, or reaches outside the repository. Compose a digest covering what ran, which entries were flagged, and the exact text of each flagged command. Do not delete, redact, repair or re-run anything yourself, and do not contact anyone: queue suggested fixes and any outgoing messages for the user to approve first.