explaingit

tetsuo-ai/dad

17ShellAudience · developerComplexity · 4/5ActiveLicenseSetup · hard

TLDR

Grok Build plugin that runs a supervisor pane in tmux to watch a coding agent, detect stalls, and refuse done claims that lack command-output or file-change evidence.

Mindmap

mindmap
  root((dad))
    Inputs
      Objective text
      Mode flag
      tmux window
    Outputs
      Event traces
      Evidence files
      Gate decisions
      Daemon logs
    Use Cases
      Long refactors
      Security audits
      Release plans
      Repo exploration
    Tech Stack
      Bash
      Python3
      tmux
      Git

Things people build with this

USE CASE 1

Run a multi-hour refactor with a supervisor that restarts the agent when it stalls at a prompt

USE CASE 2

Do a read-only repository audit where the agent cannot edit files

USE CASE 3

Drive a high-autonomy yolo build that keeps iterating until tests pass

USE CASE 4

Record JSONL evidence and gate decisions for every claim the worker agent makes

Tech stack

BashPythontmuxGitLinux

Getting it running

Difficulty · hard Time to first run · 1h+

Linux-only, needs tmux, python3, git, and a working Grok Build install with plugin support before /dad will run.

MIT license, free to use, modify, and ship commercially as long as you keep the copyright notice.

In plain English

DAD is a plugin for Grok Build, which is xAI's coding agent command-line tool. The plugin is designed for situations where an AI agent has to keep working on a task for a long stretch without a human watching it, such as a multi-step refactor, a security audit, or a release plan. Its job is to act as a second layer of judgement that decides when the working agent is actually making progress and when it has stalled or declared a task done without proof. Inside one tmux window (tmux is a terminal that can split into multiple panes), DAD creates two roles. A pane called the Son runs the actual work using the Grok agent. A pane called Dad supervises that work, restarts it if it gets stuck at a prompt, repairs the recurring scheduler entries that drive it, and writes a stream of evidence to disk. The supervision logic runs in three loops at different speeds: a fast loop every two minutes for nudges and watchdog checks, a deep loop every twelve minutes for evidence gates, and a strategic loop every thirty minutes for higher-level review. The central rule that the author keeps repeating is that prose is not proof. When the Son writes something like "the change is complete", DAD treats that as a claim that has to be backed by command output, file changes, event traces, or explicit verification gates. Gates cover transcript integrity, action effect, stale git state, and workspace identity. If a check fails, DAD pushes the Son back to work with one bounded next action. Installation goes through the Grok Build /plugins menu, after which a /dad command takes an objective and a mode. Three modes are offered: safe for normal work, review-only for audits that should not change files, and yolo for high-autonomy builds. Other slash commands let the user check status, pause and resume, repair the run, or stop it. The plugin requires Linux, tmux, bash, python3, git, and a small set of standard utilities, and ships a doctor script to verify the environment.

Copy-paste prompts

Prompt 1
Walk me through installing the dad plugin in Grok Build and running my first /dad safe objective on a sample repo
Prompt 2
Show me how dad's evidence gates decide whether a 'task complete' claim from the Son is accepted or rejected
Prompt 3
Explain the three supervision loops in dad (fast 2min, deep 12min, strategic 30min) and what each one checks
Prompt 4
How do I write a dad-doctor.py check for a new platform requirement, and where does it plug into the install flow
Prompt 5
Compare dad's three modes (safe, review-only, yolo) and tell me which one to pick for a production release audit
Open on GitHub → Explain another repo

Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.