coding-agent-loops（编码 Agent 循环）

Coding Agent Loops

Run AI coding agents in persistent, self-healing sessions with automatic retry and completion notification.

Core Concept

Instead of one long agent session that stalls or dies, run many short sessions in a loop. Each iteration starts fresh — no accumulated context. The agent picks up where it left off via files and git history. This is the "Ralph loop" pattern.

Prerequisites

• tmux installed
• ralphy-cli: npm install -g ralphy-cli
• A coding agent: codex (Codex CLI) or claude (Claude Code)
• Stable tmux socket: always use ~/.tmux/sock (default /tmp socket gets reaped by macOS)

Quick Start

Single Task

tmux -S ~/.tmux/sock new -d -s my-task \
  "cd /path/to/repo && ralphy --codex 'Fix the authentication bug'; \
   EXIT_CODE=\$?; echo EXITED: \$EXIT_CODE; \
   openclaw system event --text 'Ralph loop my-task finished (exit \$EXIT_CODE) in \$(pwd)' --mode now; \
   sleep 999999"

PRD-Based Workflow (Preferred for Multi-Step Work)

tmux -S ~/.tmux/sock new -d -s feature-build \
  "cd /path/to/repo && ralphy --codex --prd PRD.md; \
   EXIT_CODE=\$?; echo EXITED: \$EXIT_CODE; \
   openclaw system event --text 'Ralph loop feature-build finished (exit \$EXIT_CODE) in \$(pwd)' --mode now; \
   sleep 999999"

Parallel Agents on Separate Tasks

ralphy --codex --parallel --prd PRD.md

Session Management

Check Progress

tmux -S ~/.tmux/sock capture-pane -t my-task -p | tail -20

List Active Sessions

tmux -S ~/.tmux/sock list-sessions

Kill a Session

tmux -S ~/.tmux/sock kill-session -t my-task

The Completion Hook (Mandatory)

Always append this to tmux commands:

; EXIT_CODE=$?; echo "EXITED: $EXIT_CODE"; \
openclaw system event --text "Ralph loop <name> finished (exit $EXIT_CODE) in $(pwd)" --mode now; \
sleep 999999

Why each part matters:

• EXIT_CODE=$? — captures the agent's exit code
• echo "EXITED: $EXIT_CODE" — visible in tmux pane output
• openclaw system event — fires a wake event so OpenClaw notifies you immediately
• sleep 999999 — keeps the shell alive so output remains readable

PRD Format

Ralph tracks completion via markdown checklists:

## Tasks
- [ ] Create the API endpoint
- [ ] Add input validation
- [ ] Write tests
- [x] Already done (skipped)

Ralph validates that all items are checked before accepting a completion signal from the agent.

When to Use What

Key Principles

1. Always use tmux — background exec processes die on gateway/host restart. tmux sessions persist.
2. Always use the stable socket (~/.tmux/sock) — the default /tmp socket gets cleaned up.
3. Always add the completion hook — without it you won't know when the agent finishes.
4. Log active sessions — record running sessions in daily notes or a tracking file so you don't lose awareness.
5. Verify before declaring failure — after a process ends, check git log, git diff, and process output before concluding it failed.
6. PATH in tmux — tmux may not inherit your full PATH. Prepend /opt/homebrew/bin: if tools aren't found.

Troubleshooting

• Agent exits immediately: Check ~/.codex/log/codex-tui.log for auth errors. May need codex auth login.
• Ralph marks tasks done but nothing committed: Ralph can mark PRD tasks complete even when the agent fails silently. Always verify via git log --oneline -3 and git diff --stat.
• API rate limits (429s): Common when running multiple parallel agents. Ralph's retry handles this, but reduce parallelism if persistent.
• Session disappeared: tmux sessions can die from OOM or system restarts. Check with tmux -S ~/.tmux/sock has-session -t <name> and restart if needed.