appendix-e-sources
Appendix E: Sources & References
Primary sources used across the handbook, grouped by topic and verification status.
All major claims in this handbook are sourced. This appendix lists primary sources by category.
OpenClaw — Primary Sources
| Source | Type | Verified |
|---|---|---|
| openclaw/openclaw | Source code | ✅ Local fork at /Users/mafr/Code/github/openclaw |
docs/concepts/system-prompt.md | Architecture reference | ✅ Verified |
docs/gateway/heartbeat.md | Heartbeat interval (30m default) | ✅ Verified |
src/agents/skills/workspace.ts:390 | SKILL.md discovery | ✅ Verified |
src/agents/tools/sessions-send-tool.ts:99 | A2A sessions | ✅ Verified |
src/agents/skills/refresh.ts:85-93 | File auto-discovery | ✅ Verified |
| openclaw v2026.3.28 release | Async approval, xAI/Grok, MiniMax, plugin gating | ✅ GitHub Release |
| openclaw v2026.4.1 | Follow-up release: /tasks, SearXNG, cron allowlists, exec/plugin hardening | ✅ GitHub Release |
NanoClaw — Primary Sources
| Source | Type | Verified |
|---|---|---|
| qwibitai/nanoclaw | Source code | ✅ GitHub |
NanoClaw Security Model — docs/SECURITY.md | Trust model, container isolation, OneCLI Vault, mount allowlist | ✅ Fetched |
| VirtusLab: NanoClaw deep-dive (Feb 25, 2026) | Architecture analysis, AI-native philosophy, ~500 LOC claim | ✅ Fetched |
| VentureBeat: NanoClaw CVE story (Feb 11, 2026) | Security background | ✅ |
| Blink.new: OpenClaw vs NanoClaw comparison (Mar 24, 2026) | Feature comparison | ✅ |
SecureClaw & Security Ecosystem
| Source | Type | Verified |
|---|---|---|
| adversa-ai/secureclaw | Plugin/skill | ✅ GitHub |
| Adversa AI launch post | OWASP v2.0 coverage | ✅ |
| SecureClaw: 5 AI Security Frameworks | MITRE, NIST, EU AI Act mapping | ✅ |
| prompt-security/clawsec | Cross-platform security suite | ✅ GitHub |
NemoClaw / NVIDIA
| Source | Type | Verified |
|---|---|---|
| NVIDIA Newsroom announcement (Mar 16, 2026) | Official press release | ✅ |
| NVIDIA/NemoClaw GitHub | Source | ✅ GitHub |
| CNBC: Jensen Huang “next ChatGPT” | Jensen Huang quotes | ✅ |
| TechMarketer: GTC 2026 full breakdown | Full architecture description, all Jensen Huang quotes verified | ✅ Fetched full text |
| ZDNET: NemoClaw security stack | OpenShell description | ✅ |
Peter Steinberger / OpenClaw Background
| Source | Type | Verified |
|---|---|---|
| steipete.me: OpenClaw, OpenAI and the future (Feb 14, 2026) | Steinberger’s own post | ✅ Fetched full text |
| Lex Fridman Podcast #491 (Feb 11, 2026) | Peter Steinberger interview | ✅ |
| Fortune: Who is Peter Steinberger? | Background profile | ✅ |
| OfficeChai: More than a decade in the making | Career history | ✅ |
Agentic Control Plane — Claude Code, Cursor, Windsurf et al.
| Source | Type | Verified |
|---|---|---|
| Claude Code architecture deep-dive (Mar 31, 2026) | Full architecture: agentic loop, ~26 tools, 5-layer permissions, Teams, CLAUDE.md | ✅ Fetched full text |
| Claude Code internals Part 3: Message Structure | Messages API usage | ✅ |
| claude-code/issues/20625 | Structured output schemas | ✅ |
| cline/cline/issues/9848 | XML tool format leak bug | ✅ GitHub |
| OpenAI acquires Windsurf $3B | Acquisition | ✅ |
| Google acquires Windsurf executives $2.4B | Talent deal | ✅ |
| Lovable $6.6B valuation, acquisition push (Mar 24, 2026) | Market signal | ✅ |
| HatchWorks: AI wrapper moat (Mar 31, 2026) | Thin wrapper analysis | ✅ |
API Layer
| Source | Type | Verified |
|---|---|---|
| Portkey: Three API formats compared (Feb 23, 2026) | Chat Completions vs Responses vs Messages | ✅ Fetched full text |
| LiteLLM docs: v1/messages → /responses mapping | Parameter translation | ✅ |
| LiteLLM supply chain attack (Mar 26, 2026) | Security incident | ✅ |
| Anthropic: Advanced tool use (Nov 24, 2025) | tool_use blocks, Agent SDK | ✅ |
Workforce Disruption & Governance
| Source | Type | Verified |
|---|---|---|
| Oracle restructuring coverage (BBC, Mar 31, 2026) | Large-scale AI-linked restructuring reports | ✅ |
| Larry Ellison: “build more software with fewer people” | Direct quote | ✅ |
| HBR: To Thrive in the AI Era, Companies Need Agent Managers (Feb 12, 2026) | Agent Manager role, Zach Stauber quote | ✅ Fetched |
| McKinsey: Trust in the Age of Agents (Mar 5, 2026) | Four-layer accountability model | ✅ |
| McKinsey: State of Organizations 2026 (PDF) | Nine shifts, human-agent collaboration | ✅ |
| McKinsey: The Agentic Organization (Sep 26, 2025) | Agentic operating model | ✅ |
| IJRSI: When AI Agents Act | Governance, accountability, risk matrix | ✅ |
| Singapore AIGL: Model AI Governance Framework for Agentic AI v1.0 (Feb 2026) | Four principles | ✅ |
Model Evaluation — Sources
| Source | Type | Verified |
|---|---|---|
| Gemma 4 vs Qwen 35B: Agentic Tool Calling (Apr 3, 2026) | Real-world benchmark: 4% diff vs claimed 50–80% | ✅ Fetched |
| Small LLMs for Agentic Tool Calling — arXiv 2512.15943 (Dec 2025 / rev. Mar 2026) | SLM fine-tuning outperforms large models on tool-bench; 350M model 77.55% pass rate | ✅ arXiv |
| EFFGEN: Agentic framework for SLMs — arXiv 2602.00887 (2026) | Prompt compression 70–80%, 11.2% gain for SLMs vs 2.4% for large models | ✅ arXiv |
| TinyLLM: SLMs on edge — arXiv 2511.22138 (Nov 2025) | Accuracy differences across model scales for function calling; medium-sized 1–3B outperform ultra-compact | ✅ arXiv |
| TheAgentCompany: Real-world agent benchmarking — arXiv 2412.14161 (Dec 2024 / rev. Sep 2025) | Best agent completes 30% of workplace tasks; performance varies dramatically by scaffold | ✅ arXiv |
| Why LLM Benchmarks Fail Your AI Agent — bestaifor.com | 0.95^10 problem; GPT-4 SWE-bench: 2.7% vs 28.3% across scaffolds; Llama 4 benchmark gaming documented | ✅ Fetched |
| DABstep: Data Agent Benchmark — Adyen + HuggingFace | Best reasoning agent 16% accuracy on real financial agent tasks | ✅ HuggingFace |
| Qwen3.5-27B reasoning suppressed with tools — GitHub Issue #89 (Mar 2026) | Tool presence triggers “fast tool-call mode,” bypasses extended reasoning — trained behavior, not prompt-configurable | ✅ GitHub |
| TinyLLMs for Tool-Calling Agents — Medium, Mar 2026 | Majority of agent decisions are constrained/structured; small models handle these reliably | ✅ Fetched |
n8n vs Zapier
| Source | Type | Verified |
|---|---|---|
| n8n 70+ AI nodes | Feature count | ✅ |
| Zapier vs n8n comparison (Apr 1, 2026) | Platform philosophy | ✅ |
Flowwink / Internal
| Source | Type |
|---|---|
liteit.se | Platform — Flowwink CMS/CRM/FlowPilot |
autoversio.ai / autoversio.com | Partner — private AI on-premise (Uppsala, Sweden) |
/Users/mafr/Code/github/openclaw | OpenClaw local fork — all source code claims verified here |
| Flowwink production deployment | Edge functions, skills, memory architecture |
Notable OpenClaw Forks — GitHub API verified April 2, 2026
| Repo | Stars | Verified |
|---|---|---|
| jiulingyun/openclaw-cn | 4,604 | ✅ API |
| DenchHQ/DenchClaw | 1,459 | ✅ API |
| OpenBMB/EdgeClaw | 905 | ✅ API |
| sunkencity999/localclaw | 78 | ✅ API |
| jomafilms/openclaw-multitenant | 41 | ✅ API |
| remoteclaw/remoteclaw | 16 | ✅ API |
| romiluz13/ClawMongo | 15 | ✅ API |
| DenchHQ: OpenClaw Is Early React | Blog (Mar 26, 2026) | ✅ |
| EdgeClaw 2.0 blog | EdgeClaw architecture | ✅ |
OSS Insight — Ecosystem Analysis
| Source | Type | Verified |
|---|---|---|
| OpenClaw forks wave analysis (Mar 26, 2026) | Rewrite wave analysis | ✅ |
Canonical source registry is maintained at src/content/SOURCES.md.
Last updated: April 11, 2026. All URLs verified active at time of research.