Context

Running OpenClaw 2026.3.13 natively (no Docker) on a Hetzner CPX32 VPS:

• CPU: 4× AMD EPYC Genoa vCPU
• RAM: 8 GB
• Disk: 160 GB SSD (56% used)
• OS: Ubuntu, Linux 6.8.0-101-generic (x64)
• Node.js: v22.22.0
• Gateway: systemd-managed, loopback-only

Current config highlights

Request

Perform a systematic analysis of potential performance bottlenecks that could cause reasoning slowness in this native setup. Specifically:

1. Context assembly & prompt size

• With contextTokens: 800000 and models that have 200k native windows, how does OpenClaw handle the mismatch? Is there unnecessary token counting, truncation overhead, or wasted prompt assembly work?
• How does workspace file injection (AGENTS.md, SOUL.md, USER.md, TOOLS.md, etc.) scale? Are these re-read and re-tokenized on every turn?

2. Memory search overhead

• memorySearch with session recall enabled — does the embedding search (OpenAI text-embedding-3-small) happen synchronously before each model call?
• What is the latency profile of memory search on a VPS with no GPU and external API calls?
• Is there any caching of embedding results or does it re-embed the query every time?

3. Compaction pipeline

• Compaction uses a separate model call (claude-sonnet-4-6). When does this fire vs. the actual reasoning call? Is it sequential or parallel?
• On a 4-vCPU box, is the Node.js event loop blocked during compaction token counting?

4. Subagent orchestration

• With maxConcurrent: 3 subagents + maxConcurrent: 4 agents, what is the actual concurrency behavior?
• Are subagent spawns serialized behind a single event loop? Do they share the same Node.js process or fork?
• What is the overhead of the announce-back mechanism?

5. Telegram channel latency

• streaming: off + blockStreaming: true means the full response must complete before delivery. For long reasoning chains this adds perceived latency.
• queue mode — does message queueing add artificial delays between user messages and processing start?
• timeoutSeconds: 90 on Telegram — does this cause premature aborts on complex multi-tool reasoning turns?

6. Hook overhead per turn

• 4 internal hooks enabled (boot-md, bootstrap-extra-files, command-logger, session-memory). What is their per-turn cost?
• Does session-memory hook do synchronous I/O on every turn?

7. Auth fallback chain

• 2 Anthropic accounts, 2 OpenAI-Codex accounts with 2h billing backoff. Does the fallback probing add latency on every request, or only on failure?

8. Node.js / OS level

• 4 vCPUs with LimitNOFILE=524288 — is there any event loop contention when running gateway + multiple subagent sessions?
• Is there measurable GC pressure at 8 GB RAM with large context windows?

Expected outcome

A documented analysis (or instrumentation plan) that identifies:

1. Measured or estimated latency contribution of each layer
2. Quick wins — config changes that reduce latency without code changes
3. Code-level improvements — where the gateway pipeline could be optimized
4. Instrumentation gaps — what timing data is missing today to diagnose slowness

Environment

OpenClaw 2026.3.13 (61d171a)
Node.js v22.22.0
Linux 6.8.0-101-generic x64
AMD EPYC Genoa 4 vCPU / 8 GB RAM
Native systemd deployment (no Docker)