[BUG] Agents continue working after multiple interruptions and /clears

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?

Version: 2.1.42
Login method: Claude Max Account
Model: Default Opus 4.6 · Most capable for complex work
MCP servers: trello ✔
Memory: project (CLAUDE.md)
Setting sources: User settings, Project local settings

Background agents launched via Task tool with run_in_background: true survive multiple user interrupts (Escape), /clear commands, and manual kills via the task
manager UI (k key). The main agent resurrects and delivers output long after the user has explicitly stopped all work.

### Sequence of events:

1. Claude (Opus 4.6) launched 3 background investigation agents via Task with run_in_background: true
2. I interrupted the main agent (Escape) and ran /clear
3. The 3 background agents were still visibly running in the task manager — I navigated to each (arrow down) and killed them with k
4. Status bar confirmed no background processes remaining
5. Despite this, the main agent resurrected and stopped responding to Escape — I had to type "Claude stop" and "stop" as text input to get it to acknowledge
6. I ran /clear a second time
7. ~30 seconds later, a full multi-page response appeared in the terminal as if Claude was actively responding — with zero indication anything was still running

### Key details:

• The k kill in task manager did not actually terminate the agents (or something downstream survived)
• The main agent became unresponsive to Escape after resurrecting
• The status bar showed no background work, but execution was clearly still happening
• Two /clear commands, multiple Escape presses, manual k kills, and typed "stop" commands were all insufficient to halt execution

What Should Happen?

What Should Happen?

1. k in the task manager should hard-kill background agents — confirmed dead, no further output
2. /clear should be a hard reset — kill all in-flight agents (foreground and background) and suppress any pending output
3. Escape should always remain responsive — the main agent should never enter a state where Escape stops working
4. No output should surface after /clear — if agents were killed or conversation was cleared, their results should be discarded, not queued for delivery
5. The status bar should be accurate — if it shows no background processes, nothing should be running

Error Messages/Logs

Error Messages/Logs

  No error messages or stack traces. The issue is behavioral — agents silently continued
  executing and delivered output after being killed and after the conversation was cleared.
  No logs were captured as the output appeared unexpectedly after /clear with no prior
  indication of ongoing work.

Steps to Reproduce

Steps to Reproduce

1. Start a Claude Code session with a codebase available for investigation
2. Give Claude a task that causes it to launch multiple background agents

(e.g., "investigate these 12 findings in the codebase" — Claude will
typically parallelize with 3+ background Task agents using
run_in_background: true)

1. Wait for background agents to appear in the task manager (visible via

arrow-down or status bar)

1. Press Escape to interrupt the main agent
2. Run /clear
3. Navigate to each background agent in the task manager and press k to kill
4. Confirm status bar shows no background processes
5. Wait 30-60 seconds

Expected: Nothing happens. Session is clean.
Actual: The main agent resurrects and delivers a full response. Escape
may stop working. A second /clear may also fail to prevent output delivery.

Note: Reproduction may depend on timing — the agents need to be mid-execution
(not yet complete) when killed. Longer-running investigations (reading many
files) are more likely to trigger this.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Code Version: 2.1.42

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Other

Additional Information

---
Additional Information

• MCP servers active at time of issue: trello
• The background agents were Explore-type subagents performing codebase

searches (Glob, Grep, Read operations across multiple files)

• The final output that surfaced after the second /clear was a complete,

well-formatted multi-page report — suggesting the agent ran to full
completion despite being "killed"

• This suggests the kill signal either didn't propagate to the underlying

process, or the output was already buffered/queued and delivered regardless
of conversation state