keepgoing

Gives Claude Code or Codex one last chance to finish required follow-through before ending a turn. By default it runs once, then honors the stop_hook_active loop breaker so the agent can stop normally.

For experiments that truly need relentless continuation, infinite mode keeps rejecting every stop until you interrupt the agent. It is deliberately opt-in.

Combine with hype for the ultimate scientific buddy: one hook that reminds Claude to believe in itself on every prompt, another that refuses to ever let it give up.

See the main installation guide for Claude Code and Codex setup. Codex requires you to review and trust the plugin hook before it will run.

Customization

Infinite mode

Set KEEPGOING_INFINITE=1 in the environment where you launch Claude Code or Codex. Infinite mode ignores the loop breaker and can keep consuming tokens until you interrupt the agent, so it should be enabled only intentionally.

For Claude Code, it can also be set per project:

{
  "env": {
    "KEEPGOING_INFINITE": "1"
  }
}

Completion messages

Set KEEPGOING_MESSAGES_FILE to a text file of your own encouragement — one message per line; blank lines and # comments are ignored. Useful for tailoring the completion check to the task at hand. Export it in the shell where you launch your agent, or set it per project in .claude/settings.json:

{
  "env": {
    "KEEPGOING_MESSAGES_FILE": "./keepgoing-messages.txt"
  }
}

How it works

Whenever the agent first tries to finish its turn, the Stop hook (hooks/keepgoing.py) returns {"decision": "block", "reason": ...}. The agent receives that reason as a continuation prompt. When it stops again, the hook sees stop_hook_active: true and allows the turn to end unless infinite mode is on.

Codex supplies the same CLAUDE_PLUGIN_ROOT compatibility variable used by Claude Code, so both clients run the same hook definition and Python script.