OpenClaw Bug: Discord Channel Sessions Bypass Compaction

Severity: High
Discovered: 2026-02-07
Version: OpenClaw 2026.2.6-3 (85ed6c7)

Summary

Discord channel sessions accumulate tokens indefinitely, ignoring contextTokens limit and compaction settings. Eventually causes context overflow errors.

Expected Behavior

With config:

{
  "contextTokens": 50000,
  "compaction": {
    "mode": "default",
    "maxHistoryShare": 0.5
  }
}

All sessions should:

• Send max ~50k tokens of history to the model
• Trigger compaction when history exceeds 25k tokens (50% of 50k)
• Never overflow the model's 200k context limit

Actual Behavior

Discord channel sessions grow unbounded:

Discord #chat eventually hit 211k tokens, causing:

400 {"type":"error","error":{"type":"invalid_request_error","message":"prompt is too long: 211803 tokens > 200000 maximum"}}

Impact

• Silent failure with "typing..." indicator stuck
• User waits 10+ minutes with no response
• Requires manual session reset via /new or transcript deletion
• Recurring issue as all Discord channel sessions trend toward overflow

Steps to Reproduce

1. Configure contextTokens: 50000 and compaction.mode: default
2. Use OpenClaw via Discord channels for several days
3. Run openclaw sessions list --json | jq '.sessions[] | {key, totalTokens}'
4. Observe Discord channel sessions far exceeding contextTokens limit

Workaround

Created session-monitor.py cron that:

• Checks session token counts every 15 min
• Alerts at 80% of 200k limit
• Auto-prunes transcripts at 95%

Environment

• OpenClaw 2026.2.6-3 (85ed6c7)
• macOS Darwin 25.2.0 (arm64)
• Node v25.5.0
• Model: anthropic/claude-opus-4-5