Coding Agent Loops

The proven pattern for running AI coding agents in production. Stop losing work to crashed sessions.

The Problem

Long agent sessions stall. Claude Code times out. Codex crashes. You lose hours of work with nothing to show. No notification when it finishes or fails.

The Solution: Ralph Loop Pattern

Run coding agents in short focused loops inside persistent tmux sessions. Each loop starts fresh with no accumulated context. The agent picks up where it left off via files and git history. Completion fires a wake event so you know immediately when work is done.

tmux -S ~/.tmux/sock new -d -s my-feature \
  "cd /repo && ralphy --codex --prd PRD.md; \
   openclaw system event --text 'Ralph loop finished' --mode now; \
   sleep 999999"

What's Included

• SKILL.md with full OpenClaw, Claude Code, and Codex integration
• Ralph loop pattern: battle-tested, production-proven session management
• PRD-based workflow: markdown checklists that survive agent restarts
• Parallel agent setup: run multiple coding agents on separate tasks simultaneously
• Completion hooks: instant Telegram/OpenClaw notifications when jobs finish
• tmux socket hardening: avoid macOS socket-reaping that kills sessions silently
• Session monitoring commands: check progress without interrupting the agent

Compared to Running Agents Directly

| Direct | With Coding Agent Loops | |---|---| | Session dies, work lost | tmux survives disconnects | | No retry on failure | Ralph retries automatically | | No notification on finish | Instant wake event | | Context rot accumulates | Fresh context every loop | | One task only | Parallel agents on separate tasks |

Verified with Claude Code (claude-sonnet-4-6), OpenAI Codex CLI, and OpenClaw on macOS and Linux.

Built by ClawGear. Production-tested.