From Official OpenClaw to NexAgent Lingxiao: Complete Version Timeline & Migration Guide

I. OpenClaw Version Timeline (v2026.3.8 → v2026.4.2)

v2026.3.8 (2026-03-09)

• Backup tools: openclaw backup create / verify with manifest validation
• Talk mode: configurable silenceTimeoutMs auto-send silence threshold
• TUI: infer active agent from current workspace
• Brave web search: new llm-context mode returning LLM-ready grounding snippets

v2026.3.11 (2026-03-12)

• 🔒 Security fix: WebSocket CSWSH (GHSA-5wcw-8jjv-m286), enforce browser origin validation
• sessions_yield: orchestrators can end turn immediately, skip queued tools, carry hidden payload to next turn
• Fast mode: Anthropic (service_tier) + OpenAI both support session-level fast inference toggle
• iOS welcome: live agent overview, auto-refresh on reconnect/foreground

v2026.3.12 (2026-03-13)

• Control UI Dashboard v2: modular overview/chat/config/agent/session views, command palette, mobile bottom tabs, export, search, pinned messages
• Ollama onboarding: first-class Local or Cloud+Local setup, curated model suggestions
• GPT-5.4 fast mode: session-level toggle parallel to Anthropic fast

v2026.3.13 (2026-03-14) — Bug fix

• Fix: session reset loses lastAccountId / lastThreadId
• Fix: Discord gateway metadata fetch failures
• Fix: compaction uses full-session token count for sanity check

v2026.3.22 (2026-03-23) — ⚠️ BREAKING

• Plugin install priority: ClawHub over npm; fallback to npm only when ClawHub misses
• Chrome MCP: remove legacy extension relay path; migrate via openclaw doctor --fix
• image_generate standardized: remove nano-banana-pro bundled skill; use agents.defaults.imageGenerationModel
• New Plugin SDK: openclaw/plugin-sdk/* is the public path; openclaw/extension-api removed, no shim

v2026.3.23 (2026-03-23)

• Qwen/ModelStudio: standard DashScope pay-as-you-go endpoints alongside Coding Plan
• Control UI Knot theme: black-and-red palette, WCAG 2.1 AA contrast, config section icons
• CSP: SHA-256 hashes for inline scripts in script-src directive

v2026.3.24 (2026-03-25)

• OpenAI compat: /v1/models, /v1/embeddings, explicit model overrides on /v1/chat/completions
• /tools dynamic: shows actual available tools for current agent; "Available Right Now" in Control UI
• Teams: official SDK migration, streaming replies, welcome cards, typing indicators, AI labeling
• Skills one-click recipes: coding-agent, gh-issues, whisper-api with dependency auto-install
• Skills Control UI: filter tabs, detail dialog with toggle/install/API key entry

v2026.3.28 (2026-03-29)

• xAI/Grok: Responses API, first-class x_search tool, auto-enable xAI plugin from web-search config
• MiniMax: image-01 generation + image-to-image editing with aspect ratio
• Plugin hooks requireApproval: before_tool_call pauses for user approval via Discord buttons/Telegram/exec overlay
• Unified /approve: handles exec + plugin approvals with fallback
• ACP: Discord/iMessage current-conversation bind for Codex spawning

v2026.3.31 (2026-03-31) — ⚠️ BREAKING

• Remove nodes.run wrapper: all node shell execution through exec host=node
• Plugin SDK compat paths deprecated: migrate to openclaw/plugin-sdk/* + api.ts / runtime-api.ts
• Skill install dangerous-code fails closed: critical findings require --dangerously-force-unsafe-install
• Gateway trusted-proxy: reject mixed shared-token configs; local-direct needs explicit token
• Node commands: locked until pairing approval — device pairing alone no longer enough

v2026.4.1 (2026-04-01)

• /tasks chat-native board: background task list for current session with recent details
• SearXNG provider: bundled web_search with configurable self-hosted instance
• Amazon Bedrock Guardrails support
• Voice Wake: trigger Talk Mode on macOS
• agents.defaults.params: global default provider parameters
• Auth failover: rate-limit retry cap + auth.cooldowns.rateLimitedProfileRotations knob
• openclaw cron --tools: per-job tool allowlists

v2026.4.2 (2026-04-02) — ⚠️ BREAKING

• xAI config path migration: tools.web.x_search.* → plugins.entries.xai.config.xSearch.*; run openclaw doctor --fix
• Firecrawl config path migration: tools.web.fetch.firecrawl.* → plugins.entries.firecrawl.config.webFetch.*
• Task Flow restoration: managed vs mirrored sync, durable state + revision tracking, openclaw flows inspect/recover
• Managed child tasks + sticky cancel: stop scheduling immediately, parent Task Flow settles to cancelled
• api.runtime.taskFlow seam: plugins create/drive managed Task Flows without owner identifiers

II. Official vs NexAgent Lingxiao: Full Comparison

III. How to Organize Custom Skills: A Layered Approach

Official OpenClaw's ClawHub provides a general-purpose skill ecosystem. The key to customization is designing skills in business-aligned layers, not just stacking features.

We organize custom skills into five layers:

🧠 Layer 1 — Memory & Learning (foundation, global)

This layer underpins everything else. All other skills depend on it.

• Memory skill: encapsulates PostgreSQL + pgvector read/write, provides a unified memory query/store interface.
• Self-improvement skill: auto-record errors, correction learning, hot-knowledge cache. The agent gets smarter with use.

Portability: These two skills work for any OpenClaw customization. Use them as-is.

🏢 Layer 2 — Platform Core (your own products/services)

Encapsulates operations on platforms you own and maintain — your website, course platform, APIs.

Principle: one skill per platform, all calls go through Bot Token + REST API, never direct DB access.

📰 Layer 3 — Content Operations (automated content production)

Daily news scraping, AI article writing, auto-publishing. This layer makes the agent a content engine.

• News scrape skill: multi-source RSS + Tavily real-time search, standardized JSON output
• AI writing skill: Claude generates bilingual Markdown, publishes by category
• Search skill: Tavily / DuckDuckGo, multi-key rotation with quota failover

Portability: Make scrape sources and publish targets configurable — works with any blog platform.

⚙️ Layer 4 — Productivity Integration (external SaaS)

One skill per external service you use daily.

• Google Workspace (email / calendar / docs / sheets)
• GitHub (code ops, PRs, issues)
• Reminders / iCal subscriptions
• Cloud deployment (Cloudflare Tunnel, Docker)
• Background task queue (async Claude Code execution)

Principle: these handle your personal/team workflow. Private data stays private.

🔒 Layer 5 — Private Business Skills (never published)

Any skill tightly coupled to your specific business — proprietary CRMs, internal time-clock systems, client-specific workflows — lives only in your workspace and is never shared.

This layer is the real "customization". Everyone's Layer 5 is different. That's the point.

Design Guidance

Layer 1 (memory/learning)    ← universal, reuse directly
Layer 2 (platform core)      ← determined by your products
Layer 3 (content operations) ← generic framework, swap sources/targets
Layer 4 (productivity)       ← pick by your SaaS stack
Layer 5 (private business)   ← fully private, never shared

The AGENTS.md Skills Index routing table manages all five layers under one roof: natural-language intent → skill file path.

IV. Migration Steps: Official OpenClaw → This Setup

Step 1 — Install base OpenClaw

npm install -g openclaw
openclaw onboard  # choose Anthropic, enter sk-ant-...

Step 2 — Configure Discord channel

openclaw configure --section channels
# Select Discord, enter Bot Token and Guild ID

Step 3 — Set Claude Sonnet 4.6 as default model In openclaw.json or agent config: agents.defaults.model: "anthropic/claude-sonnet-4-6"

Step 4 — Deploy Memory Service (strongly recommended)

sudo apt install postgresql-17
pip install pgvector fastapi uvicorn sentence-transformers
sudo -u postgres psql -c "CREATE EXTENSION vector"
cd ~/openclaw-memory-service && uvicorn main:app --host 127.0.0.1 --port 18800

Step 5 — Install custom skills

cp -r skills/ ~/.openclaw/workspace/skills/
# No registration needed — AGENTS.md routing table handles dispatch

Step 6 — Create workspace config files In ~/.openclaw/workspace/:

• AGENTS.md — Skills Index routing + exec policy + memory rules + self-learning rules
• SOUL.md — AI persona baseline (a few lines defining style)
• IDENTITY.md — Name, pronouns, timezone, language preference

Step 7 — Native systemd deployment (optional, replaces Docker) Create .service files for Flask API, Memory Service; enable and start with systemctl.

Step 8 — Configure cron jobs

# Blog auto-fetch + publish — 6AM Vancouver (13:00 UTC)
0 13 * * * cd ~/openclaw-services/blog_fetcher && python3 fetcher.py && python3 auto_publisher.py
# OpenClaw daily update analysis — 6:30AM Vancouver
30 13 * * * python3 openclaw_blog_writer.py
# Self-improvement hot-knowledge promotion — 7AM Vancouver
0 14 * * * curl -s -X POST http://localhost:18800/maintenance/promote-learnings

V. Version Impact on Our Setup

VII. Database Migration: SQLite → PostgreSQL Dual-Database Architecture

Official OpenClaw defaults to SQLite for sessions, memories, and tasks. We completely replace it with PostgreSQL 17, split into two logical databases.

Architecture Overview

Official OpenClaw
└── SQLite (openclaw.db)
    ├── sessions
    ├── memories (session-only, no persistence)
    └── tasks

NexAgent Lingxiao
├── PostgreSQL: openclaw_club (business DB)
│   ├── Course platform (users/lectures/cohorts/assignments/attendance)
│   ├── Blog system (posts/vector embeddings/full-text index)
│   ├── Discord moderation (violations)
│   └── Task tracking (task_runs / task_delivery_state)
│
└── PostgreSQL: memory schema (same PG instance)
    ├── memory.embeddings   ← semantic vector memories (pgvector)
    ├── memory.facts        ← structured facts (domain/key/value)
    ├── memory.files        ← file metadata (MinIO object refs)
    ├── memory.access_log   ← access log (time-decay scoring)
    ├── skill_reminder.*    ← reminders + calendar subscriptions
    └── skill_task.*        ← async task queue

Step 1 — Install PostgreSQL 17 + pgvector

# Ubuntu 24.04
sudo apt install -y postgresql-17 postgresql-17-pgvector
psql --version   # PostgreSQL 17.x

Step 2 — Create database and user

sudo -u postgres psql << 'SQL'
CREATE USER oc_admin WITH PASSWORD 'your_secure_password';
CREATE DATABASE openclaw_club OWNER oc_admin;
GRANT ALL PRIVILEGES ON DATABASE openclaw_club TO oc_admin;
SQL

Step 3 — Enable pgvector extension

PGPASSWORD=your_password psql -h 127.0.0.1 -U oc_admin -d openclaw_club   -c "CREATE EXTENSION IF NOT EXISTS vector;"

Step 4 — Create memory schema (persistent memory system)

PGPASSWORD=your_password psql -h 127.0.0.1 -U oc_admin -d openclaw_club   -f /path/to/agentic-memory/sql/schema.sql

Key tables:

memory.embeddings — semantic vector memories, 1024-dim bge-m3, HNSW index

memory.facts — structured key/value store with JSONB and expiry support

skill_reminder.* — reminders with iCal RRULE repeat rules + calendar feed subscriptions

skill_task.* — async task queue with status tracking and Discord notification flag

Step 5 — Create business tables (openclaw_club)

cd ~/openclaw-services/backend
python3 -c "from app import db; db.create_all()"

Key tables: users, cohorts, lectures, lecture_attendance, announcements, assignments, submissions, blog_posts (with pgvector embedding), task_runs, violations, sessions.

Step 6 — Configure connections

# Memory Service + Flask backend
DATABASE_URL=postgresql://oc_admin:[email protected]:5432/openclaw_club
REDIS_URL=redis://localhost:6379/0

Step 7 — systemd ordering

Ensure oc-api.service and memory service have After=postgresql.service in their [Unit] section.

Step 8 — Backup (recommended)

# Daily pg_dump at 3AM
0 3 * * * pg_dump -h 127.0.0.1 -U oc_admin openclaw_club | gzip > ~/backups/$(date +%Y%m%d).sql.gz

SQLite vs PostgreSQL: Key Differences

Summary

Five core upgrades when moving from official OpenClaw to Lingxiao setup:

1. Persistent memory — pgvector hybrid search; remember everything across sessions
2. Defined persona — SOUL.md gives consistent, opinionated personality
3. Vertical skills — 17 custom skills encapsulating real business logic
4. Granular governance — AGENTS.md unifies routing, exec policy, memory rules in one file
5. Lightweight ops — native systemd, zero Docker, runs on a thin-and-light laptop

This configuration is maintained in the NexAgent workspace and serves as practical material for OpenClaw Club courses.