Bug Report

Summary

memory_search tool call causes an uncaught exception that freezes the agent run and makes the gateway unresponsive. /restart (SIGUSR1) cannot recover — requires manual gateway process restart.

Environment

• OpenClaw version: 2026.2.15
• Node.js: v25.4.0
• OS: macOS (Darwin 23.6.0, arm64)
• Memory provider: local (node-llama-cpp + embeddinggemma-300M)
• Channel: Telegram

Error

[openclaw] Uncaught exception: Error: A FileHandle object was closed during garbage collection.
This used to be allowed with a deprecation warning but is now considered an error.
Please close FileHandle objects explicitly.
File descriptor: 58 (~/.node-llama-cpp/models/hf_ggml-org_embeddinggemma-300m-qat-Q8_0.gguf.ipull)

Steps to Reproduce

1. Configure local memory search provider (node-llama-cpp with embeddinggemma-300M)
2. Run on Node.js 25.x
3. Trigger memory_search tool call
4. The embedding model file handle gets garbage collected instead of being explicitly closed
5. Node.js 25 throws a hard error (previously a deprecation warning in earlier Node versions)

Impact

• Agent run freezes mid-tool-call with no response to user
• /restart fails — logs show abort failed: reason=no_active_run suggesting the run is in a zombie state
• Requires manual gateway restart (SSH into host, restart process)
• User experienced ~7 minutes of unresponsiveness before manually intervening

Relevant Log Lines

// memory_search tool starts
{"subsystem":"agent/embedded"} embedded run tool start: tool=memory_search

// Embedding batches running
{"subsystem":"memory"} memory embeddings: batch start (items=1, timeoutMs=600000)
{"subsystem":"memory"} memory embeddings: batch start (items=1, timeoutMs=600000)
{"subsystem":"memory"} memory embeddings: batch start (items=1, timeoutMs=600000)
{"subsystem":"memory"} memory embeddings: batch start (items=4, timeoutMs=600000)

// CRASH — FileHandle GC error
[openclaw] Uncaught exception: Error: A FileHandle object was closed during garbage collection...

// Later: /stop and /restart commands fail to abort
{"subsystem":"diagnostic"} abort failed: sessionId=... reason=no_active_run

Expected Behavior

• node-llama-cpp should explicitly close file handles instead of relying on GC
• If a tool call throws an uncaught exception, the gateway should recover gracefully
• /restart should be able to kill zombie agent runs

Suggested Fixes

1. Immediate: Ensure node-llama-cpp file handles are explicitly closed after embedding operations
2. Defensive: Wrap memory_search tool execution in try/catch so uncaught exceptions don't freeze the gateway
3. Recovery: /restart should force-kill any in-progress agent runs regardless of state