Babysit step 6 on Claude Code: say how to run watch-pr (background in the main session, bounded in a subagent) #232

Problem

On Claude Code, the poteto-mode Babysit playbook does not say how to run the PR watcher. Agents either hold it in the foreground or wrap it in a babysit subagent that polls CI.

skills/poteto-mode/playbooks/babysit.md step 6 (identical in installed 1.4.1 and main at 9368993, 1.15.0):

On GitHub, run skills/poteto-mode/scripts/watch-pr/watch-pr --owner "$base_owner" --repo "$base_name" --pr "$pr" under the installed plugin. … The bare command polls until a terminal verdict, which is drive behavior. … Run drive and background under /loop in dynamic mode. Rearm the watcher after every push wave and every verdict you act on. Watcher output drives wakeups. Never add a second sleep loop.

watch-pr --timeout defaults to 0 ("deadline; 0 disables it") in scripts/watch-pr/cli.ts. On Claude Code that means:

• Foreground in any session: the bare watcher has no deadline. It hits the 10-minute Bash cap and blocks the conversation until then.
• Background in the main session (run_in_background: true): this works. The exit notification re-invokes the session, which matches "Watcher output drives wakeups".
• Inside a subagent: a subagent cannot idle. Its own background notifications arrive only between its tool calls (see In-process subagents are never re-invoked when their own run_in_background Bash / Monitor task completes anthropics/claude-code#88423). A subagent that backgrounds the watcher ends up polling its output file.

Evidence

From one month of Claude Code transcripts on one machine (2026-09-19 to 10-19), where babysitting ran as one background subagent per PR:

• 227 polling loops, 9.5 h. 9.3 h of that was inside subagents, almost all of them babysit subagents.
• By what the loop waited for: the subagent's own test or build log marker (3.7 h), its own background task output file (2.9 h), and CI or GitHub checks (2.3 h).
• The longest single babysit subagent spent 1.44 h polling.

Suggested text

Add to step 6, for Claude Code:

On Claude Code, run the watcher from the main session with Bash run_in_background: true and act on its exit notification. Inside a subagent, run it in the foreground with --timeout 570 and re-arm on timeout. Do not background it and then poll its output file.

Optionally, mention the same thing in references/codex-tools.md. Codex exec_command already returns with a session id, so it is unaffected.

Not verified

• Whether /loop dynamic mode on Claude Code already re-arms a backgrounded watcher correctly. I did not test that path.