🔧 Part 1: The Upgrade That Nearly Broke EverythingA routine Node.js upgrade (v22.22.3 to v24.18.0) to support the new @openclaw/longcat-provider plugin. After install + Gateway restart, the WhatsApp channel config was silently reset to factory defaults. Allowlist policies gone. Group silence rules wiped. Priyanka DM restrictions removed. The entire security fix was applied through the terminal chat interface — instructing agent Kunia to re-apply all policies, verify the allowlist, and confirm restrictions were restored. No manual config file editing required. 👈 Lesson: Always verify channel configurations after a Gateway restart. A 30-second check can prevent a privacy incident. 🐱 Part 2: Meituan LongCat-2.0 — From Owl Alpha to Open WeightsFor two months, an anonymous model called Owl Alpha sat top 3 on OpenRouter and #1 on Hermes Agent. No company. No press release. This week: unmasked as Meituan LongCat-2.0 from the Chinese food delivery giant. Specs: 1.6T parameters (48B active MoE), 1M context, 35T+ training tokens, 50K domestic ASICs (zero Nvidia), MIT licensed. I grabbed promotional API pricing ($0.30/$1.20 per MTok) immediately. All lightweight non-coding tasks now route through LongCat. ⚔️ Part 3: Arming Kunia — Multi-Platform LLM RosterI armed my agent with a comprehensive LLM roster across providers:
- OpenRouter: LongCat-2.0 (default), DeepSeek V4 Flash, Ling-2.6 Flash
- OpenAI: GPT-5.6 Sol, GPT-5.5, GPT-5.4-mini (OAuth + direct API key backup)
- Google: Gemini-3.5 Pro
Switching is instant via /model gpt-5.6-sol in any session. No lock-in, no restarts. The Bigger PictureThis week was about resilience. Multi-provider = no single outage kills your agent. Proper configs = safety during upgrades. $0.30/MTok = experiment freely. The AI agent stack is maturing fast — run it like infrastructure: backups, monitoring, healthy paranoia. Until next Friday — keep building, keep questioning, keep your configs backed up. — Subhankar Pattanayak
Research Fellow, IMI Bhubaneswar | Founder, TechSambad |
