Skip monitor auto-start inside agent/teammate Claude Code processes

[sophieow]

Problem

monitors.json when:"always" auto-starts client.py in every Claude Code process — including agent-team teammates / subagents spawned as child claude processes (claude --agent-id worker@team --parent-session-id <sid> …, argv[0] sometimes the versioned install path). With a 30-worker fleet we observed:

• one bus listener per worker, auto-named from cwd into collision chains (myproject, myproject-2, myproject-2-2, …) with no human session behind them;
• broadcast reaching mid-task workers — the reaction policy treats incoming text as an instruction, a real derail hazard (mirrors the concern behind the skill's own SUBAGENT-STOP header).

Fix

_in_agent_context(): walk up to 15 ancestors; return True when an ancestor is a claude binary (argv[0] claude* or the versioned install path …/claude/versions/<v>) carrying --agent-id / --parent-session-id as exact argv tokens — a shell ancestor merely quoting that text must not match. main() skips only the auto-start path (no --name); an explicit --name still connects an agent deliberately. Fail-open on any error (psutil missing, AccessDenied mid-walk) so the normal connect path is never blocked.

Testing

• 9 unit tests in tests/test_agent_guard.py; full suite 208 passed.
• Live-verified on macOS, Claude Code 2.1.204/2.1.205: real teammate processes carry the flags as separate argv tokens (versioned-path argv[0]); interactive sessions run as bare claude and are not matched; a bash -c ancestor quoting the flag text is not matched.

🤖 Generated with Claude Code