tidal pier
I tested single skills through the real Telegram path and stopped after four because the pattern was already consistent:
donna (subagent): slow pre-run delay, then fast GPT/tool response
composio-cli: same
healthcheck: same
node-connect: same
In each case, Harvey burned roughly one CPU core before session.started; once the agent run actually started, GPT-5.5 responded normally in about 14-17 seconds. So this does not look like one bad skill. It looks like any non-empty agents.defaults.skills list triggers slow Telegram-session preparation in OpenClaw 2026.5.3.
I restored Harvey to the fast baseline:
agents.defaults.skills = []
Gateway restarted
Gateway CPU back idle around 0.3%
OpenClaw source code untouched
Recommendation: keep global skills disabled for now, keep GPT-5.5 primary / GPT-5.4 fallback, and treat this as an OpenClaw skills + Telegram routing regression. The next practical fix path is either a small config-only workaround for “manual skills when needed,” or an upstream bug report with these exact timelines.