Problem

Gateway process RSS grows continuously when Discord EventQueue has sustained backlog, eventually hitting the Node.js heap limit (~4GB) and crashing with FATAL ERROR: Ineffective mark-compacts near heap limit Allocation failed - JavaScript heap out of memory.

Environment

• OpenClaw: 2026.2.26
• macOS 26.3 (arm64)
• Node.js: v22.22.0
• 7 Discord bot accounts, 30+ channels/threads

Evidence

OOM Crash (Feb 25)

[2662:0x7bd40c000] 4828960 ms: Mark-Compact 4022.6 (4123.9) -> 3988.4 (4112.5) MB
FATAL ERROR: Ineffective mark-compacts near heap limit Allocation failed - JavaScript heap out of memory

Memory Growth Pattern (Mar 1)

• 11:01 AM: Gateway restart, RSS ~200MB
• 11:56 AM: RSS 537MB
• 12:04 PM: RSS 730MB
• 12:10 PM: RSS 1058MB (proactive restart at this point)
• Growth rate: ~100MB/min during heavy EventQueue backlog

EventQueue Backlog

• 17,480+ slow listener events in ~1 hour
• DiscordReactionListener: 1-12 seconds per event
• DiscordMessageListener: 30-87 seconds per event
• Rate: ~70 slow events/minute sustained

Preceding Symptoms (from error log)

[discord] Request queue is full (1000 / 1000)
[discord] Slow listener detected: DiscordMessageListener took 1497.2 seconds
[diagnostic] lane wait exceeded: lane=session:agent:circe:discord:channel:... waitedMs=1826140 queueAhead=62

Impact

• Gateway crashed and was down for 5h17m (LaunchAgent KeepAlive did not recover quickly)
• 3 weekly cron jobs (soul-evolution) were skipped
• All Discord-based agent communication was interrupted

Root Cause Analysis

The EventQueue processes Discord events (reactions, messages) synchronously. When LLM responses take >30s, the listener times out but the event remains in the queue. With 7 bot accounts receiving events across 30+ channels, the backlog grows faster than it drains. Each queued event holds references that prevent GC, causing RSS to grow linearly until OOM.

Related Issues

• Expose Discord EventQueue listenerTimeout as a configurable option #24458 — EventQueue listenerTimeout not configurable
• [Bug]:Discord EventQueue timeout (30s) too aggressive for AI responses #27851 — 30s timeout too aggressive for AI responses
• Agents periodically stop responding — EventQueue listener timeouts during heavy session I/O #29107 — Agents stop responding during heavy session I/O
• bug: DiscordMessageListener SIGUSR1 storm on long LLM responses — 30s EventQueue timeout not configurable #28566 — SIGUSR1 storm on long LLM responses

Suggested Fix

1. Short-term: Merge PR fix(discord): expose EventQueue listenerTimeout as configurable option #28945 (configurable EventQueue timeout)
2. Medium-term: Add EventQueue backpressure — drop/skip events when queue exceeds threshold
3. Long-term: Process Discord events in a worker thread to prevent event loop blocking

Workaround

Periodic openclaw gateway restart (every 4-6 hours) to clear the EventQueue and reset RSS.