Bug Description

The bot used via Telegram channel gets stuck in an infinite loop, repeatedly sending the same message. The issue is caused by a corrupted/stuck session state that persists across service restarts. Simply clearing the update-offset-default.json file is not sufficient - the session file must also be deleted.

This appears to happen when an agent task fails (e.g., Python script errors during audio generation or even a simple cron related tasks [screenshots below]) and enters a retry loop. The session state grows large (200KB+) and gets reloaded on every restart, causing the bot to continue sending the same message repeatedly.

Environment

• OpenClaw Version: 2026.1.30
• Installation Method: npm global install
• OS: Linux (Ubuntu/Debian)
• Node Version: v22.22.0
• Telegram: Using polling mode (not webhook)

Steps to Reproduce

1. Ask the agent to perform a task that may fail or timeout (e.g., generate audio/theme song using Python)
2. Let the agent encounter errors and enter a retry loop
3. Wait for the session to accumulate errors
4. Restart the service: systemctl --user restart openclaw
5. Bot loads the stuck session and continues the loop

Expected Behavior

• Service restart should clear any stuck/failed session states
• Or: Failed tasks should be cleaned up automatically after N retries
• Or: Option to "forget" session state on startup

Actual Behavior

• Session files (~/.openclaw/agents/jatin/sessions/<uuid>.jsonl) persist across restarts
• The bot re-loads the stuck session and continues sending duplicate messages
• Manual deletion of both the session file AND update-offset-default.json is required to recover

Workarounds Tried

1. ✗ Service restart alone: systemctl --user restart openclaw - Does not help
2. ✗ Clearing telegram offset only: rm ~/.openclaw/telegram/update-offset-default.json - Does not help
3. ✓ Clearing both session file AND offset: Works

Recovery script that works:

systemctl --user stop openclaw
rm -f ~/.openclaw/telegram/update-offset-default.json
rm -f ~/.openclaw/agents/jatin/sessions/<stuck-session-id>.jsonl
systemctl --user start openclaw

Logs

# From journalctl showing the stuck session being reloaded:
[agent/embedded] Removed orphaned user message to prevent consecutive user turns. 
runId=a5a47ea7-11d3-4f82-9375-39b734ac3fba 
sessionId=8f57b2a1-005e-4edf-bfc6-d20c6469dfa2

Session file sizes observed during stuck state:

• 8f57b2a1-005e-4edf-bfc6-d20c6469dfa2.jsonl: 438KB
• fed57ba4-9950-4118-b499-6757f7dbd400.jsonl: 220KB

Screenshots

Suggestions

1. Auto-cleanup: Automatically delete or archive session files that exceed a certain size threshold
2. Startup flag: Add a --fresh-start flag to ignore existing session states
3. Session timeout: Automatically expire/forget sessions after X minutes of inactivity
4. Max retry limit: Enforce a hard limit on retries for failed tasks