2.0 KiB
2.0 KiB
2026-02-18 - Chaotic Day with GLM-4.7 🦀
Model Issues
- Session started on GLM-4.7 (free Nvidia model via ZAI)
- GLM-4.7 caused repeated timeouts throughout the day — Anthony kept seeing "Continue where you left off. The previous model attempt failed or timed out."
- Model eventually switched to anthropic/claude-sonnet-4-6 late in the day
Docker Issue Discovered
- Morning Briefing cron failing with "spawn docker ENOENT" error
- Runtime showing "docker/non-main" when it should show "direct/non-main"
- Root cause: OpenClaw trying to use Docker but it's not installed/available
- Anthony's iMac node connected but exec via node not available in this config
Email Check Cron Saga
- Email check cron was spamming notifications every 15 minutes (supposed to be silent)
- Root cause: systemEvent payload always fires a notification, regardless of script output
- Multiple failed attempts to fix it:
- Changed payload text (still fired)
- Changed to isolated session (broke exec access, script failed)
- Changed back to main session with "exec" prefix (still fired)
- Finally DISABLED the cron job entirely
- Lesson learned: systemEvent crons ALWAYS notify. For silent background tasks, use agentTurn isolated with delivery.mode=none, or handle in heartbeat
Exec Tool Availability
- exec tool not available in sandbox mode
- exec host=node requires companion app support
- Anthony's iMac node doesn't support system.run in this config
- This is a significant limitation for automation scripts
Anthony's unfinished question
- Anthony typed "Can you treat my exec" but message kept getting cut off due to model timeouts
- Never got the full question — follow up next session!
Key Lessons
- GLM-4.7 is unreliable for main sessions — too many timeouts
- systemEvent cron payloads are NOT silent — they always fire a notification
- exec tool needs careful host configuration to work
- API keys should NOT be in workspace files (found SendClaw key in TOOLS.md — security issue)