restart-claude-agents

name	restart-claude-agents
user-invocable	true
description	Explicit /restart-claude-agents: restart Claude Code background agents so they load updated plugins.
allowed-tools	["Bash(*restart-claude-agents*)"]
effort	low
disable-model-invocation	true

Restart Claude agents

Restart running Claude Code background agents so a fresh process loads the plugin versions on disk. A running agent holds its plugins in memory from the moment it started; there is no in-process reload (see the /reload-plugins section of the how-plugins-work reference), so picking up an updated plugin means starting a new process. claude --bg --resume does exactly that: it stops nothing on its own, but a fresh process replays the agent's conversation and loads the current plugins, so the agent keeps its context and continues where it left off.

The helper at ${CLAUDE_PLUGIN_ROOT}/bin/restart-claude-agents does the work. It only ever touches kind: background agents; interactive sessions (your live terminals) are never restarted.

Workflow

1. List first. Run the helper's list command and show the running background agents to the user, so it is clear what is about to be restarted:

   node "${CLAUDE_PLUGIN_ROOT}/bin/restart-claude-agents" list
   

2. Restart. Without arguments, restart restarts every idle background agent and reports any busy ones it skipped:

   node "${CLAUDE_PLUGIN_ROOT}/bin/restart-claude-agents" restart
   

   • Pass one or more agent ids to restart exactly those, busy or not: restart <id> <id>
   • --force includes busy (non-idle) agents in a bulk restart
   • --cwd <path> limits to agents whose working directory is under <path>
   • --dry-run reports what would be restarted without touching anything

3. Report the old to new job id mapping the helper prints. Each restarted agent gets a new job id and session id (resume forks the session), so the ids in claude agents change.

What is preserved

The restart re-applies each agent's original launch flags, read from its own job state at ~/.claude/jobs/<id>/state.json (which Claude Code writes for every background agent): permission mode (so a bypassPermissions / YOLO agent comes back YOLO), the disallowed-tools deny list, custom settings, and the agent's goal (the stored intent is re-passed, re-arming the goal-redrive loop). Combined with --resume, the agent keeps its conversation, its permissions, its safety net, and its mission across the restart.

The only fallback: if an agent has no readable job state, it gets a plain context-preserving resume under default permissions and the helper flags it in the output. That is the exception, not the normal path.

Requirements

Claude Code, with Node.js on $PATH (the helper is a Node script). The helper drives the claude CLI it finds on $PATH; override with CLAUDE_BINARY if claude is not the right name.