[BUG] When a background agent spawns its own `run_in_background: true` bash process, that process becomes orphaned when the agent exits.

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When a background agent spawns its own run_in_background: true bash process, that process becomes orphaned when the agent exits.

What happened:

1. Spawned background haiku agent with prompt "poll pd-check.sh every 60s"
2. Agent created a bash while true loop calling pd-check.sh
3. Agent ran that loop with run_in_background: true (nested backgrounding)
4. Agent said "polling started" and exited
5. Main thread received <task-notification> that agent completed
6. Bash loop (b504a70) continued running independently - no one watching

Orphaned bash process writing to a file nobody reads. Main thread loses track of it. Agent exited thinking its job was done.

What Should Happen?

• Subagent shouldn't be able to background things.
• Or: background processes spawned by background agents should be tracked/surfaced to main thread

Error Messages/Logs

⏺ Exactly. Here's what happened step by step:

  1. I spawned haiku agent with "poll pd-check.sh every 60s"
  2. Agent read pd-check.sh to understand it (one-shot script)
  3. Agent wrote its own bash while true loop wrapping pd-check.sh
  4. Agent ran that loop with run_in_background: true
  5. Agent said "polling started" and exited
  6. Background bash process (b504a70) continues independently

  So haiku essentially recreated what pd-monitor.sh does - it wrote its own polling wrapper around the one-shot script.

  The irony: we have pd-monitor.sh sitting there unused because haiku built its own equivalent on the fly.

  Two options:
  1. Keep pd-monitor.sh - for manual terminal use, or if you want to run monitoring without spawning an agent
  2. Delete pd-monitor.sh - trust that haiku will always create a sensible loop

  The agent approach is actually more flexible - haiku can adapt the loop logic (different intervals, smarter detection) based on the prompt. The static script is more predictable.

❯ I think the agent is better because right now from the perspective of this Claude code window, I don't have any access to bash script and there's nobody checking it. You would have to check it. So I think the whole point of this was that an agent checks it and looks
  at the result and then reports asynchronously back to this thread, the main thread, if it finds something. Now I think we could do that. Yeah, we could actually do this without even a sub-agent, right? We could just have it running here in the background, but you
  would have to check it then every now and then.

Steps to Reproduce

Pretty easy, just ask for a subagent to background a process.

Claude Model

Not sure / Multiple models

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.5

Platform

Google Vertex AI

Operating System

macOS

Terminal/Shell

Other

Additional Information

_No response_