openclaw - 💡(How to fix) Fix [Bug]: QMD boot probe triggers llama.cpp build on ARM Linux — searchMode: "search" should skip vector probe [1 comments, 2 participants]

Summary

When memory.backend: "qmd" is enabled with searchMode: "search" (BM25-only), OpenClaw still calls qmd status at boot to probe vector capability. On ARM Linux without Vulkan (e.g. Oracle Cloud aarch64), this causes QMD to attempt a full CMake + llama.cpp compile, fail, hit OpenClaw's 5s containment timeout, and leave all agents with Vector: unavailable and 0 chunks indexed.

Environment

• OpenClaw version: 2026.4.1 (da64a97)
• QMD version: 2.0.1
• OS: Linux 6.17.0-1009-oracle (arm64, Oracle Cloud)
• Node: v22.22.2
• Config:

{
  "memory": {
    "backend": "qmd",
    "qmd": {
      "searchMode": "search",
      "sessions": { "enabled": true },
      "scope": { "default": "allow" },
      "limits": { "timeoutMs": 8000 }
    }
  }
}

Observed Behavior

After gateway restart, openclaw memory status shows for all 4 agents:

Provider: qmd (requested: qmd)
Model: qmd
Sources: memory, sessions
Indexed: 0/N files · 0 chunks
Dirty: no
Vector: unavailable
Vector error: QMD status probe failed: qmd status timed out after 5000ms
Batch: disabled (failures 0/0)

Gateway log confirms qmd status triggers CMake + llama.cpp build (Vulkan required, not available on this ARM server):

CMake Error: Could NOT find Vulkan (missing: Vulkan_LIBRARY Vulkan_INCLUDE_DIR glslc)
ERROR OMG Process terminated: 1

Expected Behavior

When searchMode: "search" is configured, vector capability is not needed at query time. OpenClaw should either:

1. Skip the vector probe entirely when searchMode: "search" — no need to check if llama.cpp works if we're never going to use it
2. Pass a flag/env var to qmd status instructing it not to initialize the embedding runtime (if QMD exposes such an option)
3. Degrade gracefully — if the vector probe fails/times out but searchMode: "search" is set, still index and serve BM25 results instead of leaving chunks at 0

Currently the timeout is contained (gateway stays responsive), but the QMD backend is left fully non-functional even though BM25 search would work fine.

Root Cause

The root cause is in QMD itself — qmd status should not trigger a build. Filed separately: tobi/qmd#491

However, OpenClaw can defensively fix this on its side by not probing vector capability when it isn't needed per the configured search mode.

Additional Context

• Bug #58643 (collection naming collision) was confirmed fixed in v2026.4.1 ✅
• This is the remaining blocker for QMD on ARM Linux without GPU
• The 5s timeoutMs containment added in v2026.4.1 prevents the prior gateway hang, but QMD still ends up non-functional