Bug Report: Delivery-Mirror Session Token Overflow

Date: 2026-04-08
Reporter: Catherine Rouleau (via Stan agent)
OpenClaw Version: 2026.3.13
Session Key: agent:main:telegram:direct:8799946229

Summary

The delivery-mirror session for Telegram DMs accumulated 204,765 tokens, exceeding Anthropic's 200k token limit and causing a request loop. Auto-compaction did not prevent this overflow.

Details

Session info from sessions_list:

{
  "key": "agent:main:telegram:direct:8799946229",
  "kind": "other",
  "channel": "webchat",
  "displayName": "webchat:g-agent-main-telegram-direct-8799946229",
  "model": "delivery-mirror",
  "contextTokens": 1000000,
  "totalTokens": 736865
}

Error message (from last message in session):

"errorMessage": "400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 204765 tokens > 200000 maximum\"},\"request_id\":\"req_011CZqTVoL6iU2FwWBpy3d1Z\"}"

Observable behavior:

• User sent message in solo Telegram chat
• OpenClaw attempted to deliver via delivery-mirror session
• Anthropic API rejected request (204k > 200k limit)
• OpenClaw retried multiple times, appearing as "looping" to user
• User received same diagnostic response 6+ times

Configuration

Auto-compaction is enabled in /Users/ivy/.openclaw/openclaw.json:

"compaction": {
  "reserveTokensFloor": 60000,
  "model": "anthropic/claude-sonnet-4-5",
  "memoryFlush": {
    "enabled": true,
    "softThresholdTokens": 8000
  }
}

Analysis

Why compaction didn't prevent this:

1. Model mismatch: The delivery-mirror session uses model: "delivery-mirror", not the configured compaction model (claude-sonnet-4-5). Compaction may not apply to internal routing sessions.

2. Session type: Delivery-mirror sessions are OpenClaw's internal routing layer. They shouldn't accumulate conversational history like user-facing sessions do.

3. Token accounting: Session shows contextTokens: 1000000 (likely a placeholder) but totalTokens: 736865. Actual API call had 204k tokens, suggesting the delivery-mirror transcript grew unbounded.

Expected Behavior

Delivery-mirror sessions should either:

1. Not accumulate full conversation history (they're routing only)
2. Be subject to aggressive compaction (much lower threshold than user sessions)
3. Auto-reset periodically if they must hold history

Workaround

Immediate fix:

rm -f /Users/ivy/.openclaw/agents/main/sessions/[session-id].jsonl

User-facing fix: In Telegram chat, send: /reset

Impact

• Severity: Medium — blocks user interaction until manual intervention
• Frequency: Unknown (first observed occurrence)
• Workaround: Available (session reset)

Recommendation

1. Delivery-mirror sessions should not accumulate conversational history, OR
2. Apply aggressive auto-compaction to delivery-mirror sessions (e.g., reserveTokensFloor: 150000), OR
3. Add session-type-specific compaction rules to prevent internal routing sessions from overflowing

Environment

• OS: macOS (Darwin 25.2.0 arm64)
• Node: v25.8.1
• Model: anthropic/claude-sonnet-4-5 (default)
• Channel: Telegram (solo DM)
• Session scope: dmScope: "per-channel-peer"

Transcript location (if needed for debugging): /Users/ivy/.openclaw/agents/main/sessions/ddd4e93a-5e9f-4eb0-ba8e-eb05dd2b7fde.jsonl (Note: file was deleted as workaround, may not exist by time this is reviewed)