Critical Bug: /stop and Abort Commands Fail to Interrupt Active Tool Execution in OpenClaw

A critical issue has been identified in OpenClaw where the /stop command, abort functionality, and even gateway restarts fail to interrupt actively executing tools. This means once a tool starts running, you have no reliable way to halt it — a significant problem for long-running or runaway operations.

The Problem

When you issue a /stop command or attempt to abort an operation in OpenClaw, you expect the current task to halt immediately. However, issue #29034 reveals that active tool calls continue executing regardless of these interruption attempts.

This creates several problematic scenarios:

• Runaway shell commands — If your agent launches a long-running exec command (compiling a large project, running an infinite loop, or processing massive files), you cannot stop it
• Expensive API calls — Tools making external API requests will complete even after you try to cancel
• Browser automation gone wrong — The browser tool navigating through dozens of pages won't respond to stop commands
• Resource exhaustion — Background processes continue consuming CPU, memory, and network bandwidth

Why This Matters

The inability to interrupt tools is more than an inconvenience — it's a control issue. When you give an AI agent access to your system, the implicit contract is that you maintain ultimate authority. The /stop command is your emergency brake. When it doesn't work, that contract is broken.

Consider debugging scenarios where you realize mid-execution that your agent misunderstood your intent. Maybe it's about to send emails to the wrong recipients or delete files in the wrong directory. The few seconds between recognition and completion matter enormously.

Workarounds (Limited)

Until this is fixed, your options are constrained:

1. Kill the gateway process entirely — Use pkill -f openclaw or similar, but this terminates all sessions and requires a full restart
2. Set aggressive timeouts — Configure shorter exec timeouts in your agent configuration, though this limits legitimate long-running tasks
3. Avoid dangerous tools in critical contexts — Temporarily disable exec or browser tools when you can't afford runaway operations

What Needs to Change

The fix likely requires implementing proper signal handling at the tool execution layer. Each tool wrapper needs to:

• Register for cancellation signals
• Implement graceful shutdown logic
• Kill child processes when the parent operation is cancelled

This is tracked in openclaw/openclaw#29034. If this issue affects your workflow, consider adding your voice to the discussion — the more data points on real-world impact, the faster it gets prioritized.

In the meantime, exercise caution with tools that could run longer than expected, and keep your finger near that process kill switch.