A babysitter for Claude Code that actually lets you leave the room

What it does

ralphex reads a markdown plan from your repo, spins up a new Claude Code or codex session for each unchecked task, and runs validation commands after every step. It commits progress automatically, then subjects the results to a multi-phase review pipeline—five internal agents, an optional external pass, and a final two-agent sweep—before moving the plan to a completed folder. A built-in web dashboard streams the chaos in real time if you want to watch from a safe distance.

The interesting bit

The core trick is deliberately killing context between tasks. Instead of one long Claude session that slowly forgets its own decisions, ralphex treats each task as a disposable unit, which keeps the model sharp and limits error propagation. It also detects review stalemates—when the external reviewer and Claude argue in circles—and can break the loop automatically or with a SIGQUIT nudge.

Key highlights

• Runs fully in a terminal from the repo root; no IDE plugins or cloud services required
• Supports worktree isolation for parallel plan execution and Docker for sandboxing
• Offers a review-only mode to audit code that was written outside ralphex entirely
• Ships with configurable review agents via a {{agent:name}} template system
• Optional notifications via Telegram, Slack, email, or webhook when a run finishes or implodes

Caveats

• The default Claude mode relies on claude --print, which Anthropic is moving to a separate Agent SDK credit pool in mid-2026; heavy users may need to switch to codex or a wrapper
• Mid-run steering via Ctrl+\ is unavailable on Windows, and wrapper-based workarounds for billing may break if Anthropic changes its CLI behavior
• External review loops can burn tokens in stalemate before the patience limit kicks in

Verdict

Developers who routinely break features into multi-step specs and want those specs executed while they grab coffee will find ralphex a genuine utility. If you prefer pair-programming with an AI in tight interactive loops, this tool replaces rather than augments that workflow.