Summary

Background exec sessions launched via the exec tool with background: true are silently terminated (SIGTERM) after approximately 30 minutes. This appears to happen both on context compaction and potentially via a hard timeout.

Reproduction Steps

1. Launch a long-running background exec:

   exec({ command: "while true; do echo tick; sleep 300; done", background: true })
   

2. Continue chatting normally for 30+ minutes
3. Check process status — session shows failed with signal SIGTERM

Observed Behavior

• All background exec sessions die at approximately the 30-minute mark with SIGTERM
• Multiple processes killed simultaneously, suggesting compaction is the trigger
• Exec failed (session-id, signal SIGTERM) notifications arrive bundled with the next user message, not as real-time events
• The agent cannot proactively restart failed processes since it only "wakes" on user messages, cron jobs, or heartbeats

Real-World Impact

I was running:

1. Batch audio transcription (Whisper processing 930 .wav files, ~20 min job) — killed 3 times
2. Tesla location polling loop (every 5 min) — killed repeatedly
3. Both had to be restarted manually by the user noticing and asking

Expected Behavior

• Background exec sessions should survive context compaction
• Or: a configurable maxDurationMs parameter should allow agents to set longer timeouts
• Or: exec failure should trigger an agent turn so the agent can self-heal

Current Workaround

Launch processes as fully detached OS processes:

nohup bash /path/to/script.sh > /tmp/output.log 2>&1 &

This works but loses all integration with the process tool (log, poll, kill).

Related Issues

• Session context lost after agent run timeout — background exec results injected into new session without history #20436 — Session context lost after agent run timeout (related: orphaned exec processes)
• Feature Request: Async exec callback — inject result back to session when process exits #18237 — Feature Request: Async exec callback (related: long-running process lifecycle)

Proposed Solutions

1. Compaction-safe exec: Don't SIGTERM background processes on compaction. Re-attach them to the new context.
2. Configurable timeout: Add maxDurationMs parameter to exec tool for known long-running tasks.
3. Failure callbacks: Exec failures should optionally trigger an agent turn for self-healing.

Environment

• OpenClaw version: 2026.2.17
• OS: macOS (Apple Silicon, M-series)
• Model: Claude Opus 4
• Shell: zsh