openclaw - 💡(How to fix) Fix [Performance]: 3-7 minute agent response latency on Windows — 66 plugins loaded, 49 disabled, 48 skills missing requirements [1 comments, 2 participants]

Bug type

Crash (process/app exits or hangs)

Beta release blocker

No

Summary

[Performance]: 3-7 minute agent response latency on Windows — 66 plugins loaded, 49 disabled, 48 skills missing requirements

Steps to reproduce

1. Fresh OpenClaw v2026.4.26 install on Windows 11 via npm install -g openclaw
2. Configure single Anthropic provider (claude-sonnet-4-6)
3. Configure single Slack channel (Socket Mode)
4. tools.profile: messaging, tools.deny: ["exec"]
5. 6 workspace files totaling ~15k chars
6. Send any message via CLI: openclaw agent --agent main --message "Say only: hello" --thinking off
7. Observe 3-7 minute response time for trivial messages

Expected behavior

Agent response in under 30 seconds for simple messages. The Anthropic API call itself completes in 1-2 seconds (network latency to api.anthropic.com confirmed at 0.19s). The overhead appears to be in plugin/skill resolution, not in model inference.

Actual behavior

Every agent invocation takes 3-7 minutes regardless of message complexity. Both first and subsequent calls in the same gateway session show the same latency — no warm-up benefit.

openclaw doctor reports:

• Plugins: Loaded 66, Imported 0, Disabled 49, Errors 0
• Skills: Eligible 7, Missing requirements 48

The gateway startup log shows dynamic npm package installation for bundled runtime deps (pdfjs-dist, node-edge-tts, @mozilla/readability, etc.) on first boot with Slack enabled.

99% of Anthropic API cost is cacheWrite ($1.53 of $1.54 total), suggesting the full system prompt + skill/tool context is being re-uploaded on every call rather than cache-hit.

Session context is small: 11k/1000k (1%). Network round-trip to api.anthropic.com is 0.19s. The bottleneck is not context size or network.

OpenClaw version

2026.4.26 (be8c246)

Operating system

Windows 11, Node.js v22.18.0

Install method

npm install -g openclaw

Model

anthropic/claude-sonnet-4-6

Provider / routing chain

Anthropic API direct, single provider, no routing/proxy

Additional provider/model setup details

tools.profile: messaging tools.deny: ["exec"] Plugins disabled via config: bonjour, browser 6 active plugins reported by gateway: acpx, device-pair, memory-core, phone-control, slack, talk-voice Heartbeats disabled (HEARTBEAT.md empty)

Logs, screenshots, and evidence

CLI test 1: "Say only: hello" --thinking off → 3-4 minutes
CLI test 2: "What day is it today?" --thinking off → 6 minutes
Slack DM test: "Good morning. Who are you?" → 17 minutes (likely
includes Slack session queue overhead)
Slack DM test 2: "What is 2+2?" → no response after 20 minutes

Gateway startup (clean): 24-28 seconds
Slack provider start: 5-6 minutes after gateway ready

Doctor output:
  Plugins: Loaded 66, Imported 0, Disabled 49, Errors 0
  Skills: Eligible 7, Missing requirements 48

Hypothesis: OpenClaw resolves all 66 plugins and checks all 48
missing-requirement skills on every agent invocation, not just on
gateway startup. Disabling this scan or caching the resolution
would likely reduce per-call latency to seconds.

Impact and severity

High for interactive use. Agent responses are correct but take 3-7 minutes, making conversational interaction impractical. Async use (send message, check later) is a functional workaround but significantly degrades the user experience.

Additional information

This was a fresh v2026.4.26 install (upgraded from v2026.4.24 to fix the Slack ESM Windows path issue in #74011). The latency was present on v2026.4.24 as well but was less noticeable because we weren't testing interactively.

Would appreciate guidance on:

1. Whether there's a config option to skip loading disabled plugins
2. Whether skills with missing requirements can be excluded from per-call resolution
3. Whether the cacheWrite-dominant cost pattern indicates the prompt cache key is not being reused across calls