Description

After multiple context compactions (3+) with a large context (~193k tokens), the model started outputting random Hebrew characters (gibberish) instead of coherent responses.

Environment

• Clawdbot version: 2026.1.24-3
• Model: Claude Sonnet 4.5 / Opus 4.5
• Channel: WhatsApp
• OS: Linux

Steps to reproduce

1. Long conversation with many tool calls
2. Context grows to ~190k+ tokens
3. Multiple compactions occur (3+)
4. At some point, responses become gibberish

What happened

• Model started outputting random Hebrew text
• Could not recover within the session
• Required full restart

Workaround applied

• Reduced contextTokens to 100k to trigger earlier/gentler compaction

Questions

1. Is this a known issue with compaction?
2. Is there a way to detect this state before it becomes unrecoverable?
3. Are there recommended compaction settings to prevent this?

Logs

Unfortunately no logs were captured from the incident. Will enable logging for future occurrences.