openclaw - 💡(How to fix) Fix [Bug]: Agent prep stages take ~11s per turn, 5x slower than 2026.4.15 [1 comments, 2 participants]

Bug type

Crash (process/app exits or hangs)

Beta release blocker

No

Summary

After upgrading from 2026.4.15 to 2026.4.29, every agent turn spends ~11s in local prep stages (core-plugin-tools 4.5s, system-prompt 2.7s, stream-setup 2.7s) before any API call. The same setup on 2026.4.15 completes end-to-end in 2-3s.

Steps to reproduce

1. Start OpenClaw 2026.4.29 on WSL2 with the attached config.
2. Send any message to the agent.
3. Observe ~11s prep delay before the first API byte is sent.

Expected behavior

In 2026.4.15, prep stages complete quickly and the agent responds within 2-3s end-to-end under the same WSL2 environment, same model, and same config.

Actual behavior

Gateway logs show prep stages consistently at ~11,000ms total:

• core-plugin-tools: ~4,500ms
• system-prompt: ~2,750ms
• stream-setup: ~2,750ms

The actual DeepSeek API call (measured via curl) completes in ~0.6s. End-to-end user-facing latency is 12-20 seconds per turn.

OpenClaw version

2026.4.29

Operating system

Linux 6.6.87.2-microsoft-standard-WSL2 (x64) on Windows 11

Install method

npm global

Model

deepseek/deepseek-v4-flash (thinking off)

Provider / routing chain

openclaw → api.deepseek.com

Additional provider/model setup details

openclaw → api.deepseek.com

Logs, screenshots, and evidence

Gateway prep stage trace (5 consecutive runs, consistent):

  core-plugin-tools:    4,443ms / 4,448ms / 4,496ms / 4,517ms / 4,476ms
  system-prompt:        2,754ms / 2,731ms / 2,826ms / 2,770ms / 2,724ms
  stream-setup:         2,749ms / 2,674ms / 2,734ms / 2,692ms / 2,785ms
  ──────────────────────────────── ─────────────────────────────
  totalMs:            ~11,000ms every turn

  Diagnostic warnings present:
  eventLoopDelayMaxMs=4479, eventLoopUtilization=0.6

  Direct API test (curl to api.deepseek.com): 0.6s total
  Config: minimal (deepseek provider + feishu channel only)

Impact and severity

Affected: Single-user WSL2 deployment on 2026.4.29 Severity: High (5x latency regression vs 2026.4.15) Frequency: Every agent turn, 100% reproducible Consequence: Agent feels unresponsive; 12-20s wait per message vs previously 2-3s

Additional information

Last known good: 2026.4.15 First known bad: 2026.4.29 No intermediate versions tested. Workaround: none available on 4.29; the only workaround is downgrading to 4.15. Config is minimal — only DeepSeek provider and Feishu channel enabled.