TigerAI A2A Code2n8n Skill Pack — User Manual

TigerAI A2A Code2n8n Skill Pack — v1.0 Production-Grade Methodology AI-driven build of enterprise-grade n8n workflows from natural-language sticky notes or existing codebases. Per A2A directive · 4-Tier dependency security CI-enforced · Path B real-vendor-sandbox runtime evidence.

🌐 English | 繁體中文 📖 Why Code2n8n? Read the Code2n8n manifesto — why enterprises need n8n more in the AI-coding era, not less.

The Code2n8n positioning: AI Coding (Claude Code / Codex / Antigravity) is great at writing code. n8n is great at making code manageable by an enterprise. This pack is the bridge — describe a requirement, or point at an existing system (Apps Script / Express / Lambda / Docker stack), and get a runnable n8n workflow that IT, operations, and managers can all read, audit, hand off, and govern.

🎯 One-sentence positioning: This Pack is a porting / review / validation / governance method. n8n editions + your IT carry the enterprise platform capabilities.

🤖 AI agent consuming this Pack? Read the A2A (AI-to-AI) directive before running any Code2n8n flow. It is available in 11 languages — every consuming LLM gets the gate in its primary language so the deterministic triggers and forbidden-phrase tables actually fire: English · 中文 · 日本語 · 한국어 · Français · Deutsch · Español · Tiếng Việt · ภาษาไทย · Bahasa Melayu · Bahasa Indonesia.

The directive gates when you may use the words “validated” / “tested” / “production-ready”, what tools you must invoke, what evidence schema you must emit, and what phrases are forbidden until the gate passes. Skipping it produced the v0.27.0 incident captured in examples/einvoice-n8n/REFLECTION.md. Human reviewers use docs/code2n8n-vv-checklist.md instead.

🆕 What’s new — v1.0 absorbed paid-tw/einvoice into n8n + thanks

We recently ported the upstream SDK paid-tw/einvoice (Taiwan unified e-invoice SDK, MOF MIG 4.0 spec, 5 providers: Amego / ECPay / ezPay / ezPay cross-border / ezReceipt) into n8n workflows and completed Path B’s three legs:

• ✅ Port: SDK → 80-line Hono svc wrapper + 14 n8n workflows
• ✅ Security review: 22 SEC entries (20 ✅ FIXED + 1 OPEN mitigated + 1 documented) + 4-Tier external-dependency security CI auto-enforced
• ✅ Real-vendor-sandbox runtime PASS: Amego 10/10 SDK capability against real Amego public sandbox — 11 real invoice traces queryable in Amego back-office (AA26515011 ~ AA26515020)

V&V evidence (per the A2A directive): see examples/einvoice-n8n/tests/v0.40-amego-full-coverage-report.md and the closing report tests/v0.41-final-validation-report.md.

🙏 Thanks to paid-tw/einvoice maintainers

The MIT license made this v1.0 milestone possible — it let the Pack wrap the SDK into a svc + n8n governance layer for public release. The Pack consumes the SDK via npm install with exact pins and does not vendor source (per SEC-017).

We’re also contributing an upstream-issue draft back: docs/upstream-issues/paid-tw-einvoice-sec-021-scheduled-issue.md — a real-sandbox finding that the SDK’s capabilities[] declaration vs issue() runtime behavior is inconsistent (SCHEDULED_ISSUE is accepted on Amego despite not being listed), with a reproducer and proposed fix. The Pack already mitigates this with the einvoice-capability-aware-gate workflow, but an SDK-level guard would be more reliable for all consumers.

What this Pack is — and what it is not

Receipts for the ✅ column: docs/responsibility-matrix.md (status per claim) and tests/REPORT-v0.24.1-evidence.md (fresh dated evidence).

📊 The whole pack in one picture: Natural-language intent or an existing program system → Code2n8n Skill Pack (Cookbook + 2,061 reference workflows as a design-lookup corpus, not validated templates + DSL v1.2 + 15 skills + 4 enterprise patterns) → decides what logic stays as code vs lifts into an n8n node → emits a reviewable, hand-off-able, cross-system n8n workflow. by n8n Taipei Ambassador Morris Lu

---

🔄 Two Code2n8n paths

This pack does more than turn sticky notes into workflows. It supports two directions:

Path A: from zero
natural language / yellow sticky note
  → sticky-note-to-workflow
  → n8n workflow

Path B: port an existing system
Apps Script / Express / Lambda / Netlify Functions / Docker stack
  → code-to-workflow (inventory, dedicated security gate, partition, port, validate, version/rollback evidence)
  → code modules + n8n workflow + migration docs

Code2n8n does not transliterate every line of Python or JavaScript into nodes. It re-partitions the system: complex algorithms stay in code, while triggers, cross-system wiring, retries, human approvals, notifications, and execution history lift into a visible, manageable workflow.

AI Coding solves “how is the function built”; Code2n8n solves “how is the capability modularized and governance-traced” (per A2A directive + 4-Tier external-dependency security CI gates — see v1-claims-and-evidence.md); n8n solves “how the modules cooperate across the whole enterprise.”

🧪 Case study evidence — samples of the methodology applied, not the deliverable

The deliverable is the methodology + 15 SKILLs + templates + V&V gate (§10 two-layer + A2A 11-language + §1.6 lexical) + main/critic architecture above. Case studies below are evidence of methodology applied to real codebases, kept in examples/ as reference. More will be added as new GitHub repos go through the code2n8n-pipeline SKILL. Each case maintains its own SECURITY-REVIEW.md and (where the case surfaced new methodology gaps) REFLECTION.md.

Each row is one sample. When a case surfaces a new failure mode (Code v2 contract drift, ghost cron, webhook-register asymmetry, etc.) the methodology upgrades — the case becomes a permanent record of what the upgrade was for.

🛠️ Responsibility boundary: The third block of the hero diagram (“n8n Enterprise Orchestration”) lists SSO / IAM / audit log / HA — n8n Enterprise ships these out of the box, the Pack does not reimplement them. The Pack’s job is to make sure Code2n8n-produced workflows land cleanly on top (IAM-friendly, queue-safe, rollback-traceable). The split between Pack / n8n Enterprise / your IT, and the workflow-design rules that follow, are in docs/enterprise-setup.md.

📊 Per-claim honest status (what’s done / partial / out-of-scope, row by row): docs/responsibility-matrix.md.

---

🤖 This is an Agentic Engineering Example

This entire project was authored using AI Agentic IDEs (Antigravity / Claude Code) — from spec to n8n workflows, every artifact was produced through human-AI agent collaboration.

This Skill Pack is itself a working demo of Agentic Engineering:

Agentic footprints in this repo

• skills/ — plugin.json registers 15 Claude Code / Antigravity skills (incl. v0.30.0 code2n8n-pipeline auto-pilot); each SKILL.md is co-authored by humans and AI
• .agent/workflows/ — Antigravity-native agentic workflows (e.g. /install-n8n-pack one-shot installer)
• cookbook/ — 8 natural-language → workflow examples showing how to “talk to” the AI
• spec/sticky-note-three-layer.md — Three-layer structure spec that forces reviewable AI output
• research/patterns.md — 7 canonical skeletons + anti-patterns mined by AI from 2,061 real workflows
• reference-workflows/ — AI training corpus (Zie619/n8n-workflows, MIT, secrets scrubbed)

Who should study this project

• Developers / PMs learning how to use an AI agent as an engineering teammate
• Teams who already have Apps Script, Express, Lambda, Netlify Functions, or a Docker stack and want to evaluate what stays in code vs what moves into n8n
• Teams evaluating whether Antigravity / Claude Code can replace hand-written skills / workflows
• Anyone curious what real human-AI co-authored engineering output looks like

💡 In other words: this isn’t just “a Skill Pack for n8n” — it’s also an open case study of how AI agents build a real product.

👥 You (the user) can build n8n workflows the same way

Once you install this Skill Pack, you can author your own n8n workflows with the same agentic approach — no node syntax to learn, no code to write:

Typical interaction (30-second mental model):

You ──> AI: "Every weekday 9am, pull Shopify orders, build a daily
             report, email it to the boss; on failure post to Slack #ops"

AI ──> You: ✅ workflow.json generated (Schedule → Shopify → Code → Email + Error → Slack)
             ✅ Yellow sticky: your original requirement, preserved
             ✅ Blue sticky: which credentials, constraints, test method
             ✅ Deployed to your n8n via API, webhook URL: https://...

🎯 The core idea: Users don’t need to memorise n8n node syntax — clear requirements are enough to get a structured, reviewable, maintainable workflow. Calling it deployment-grade still requires credential setup, live validation against real sandbox, and the V&V A2A directive two-layer gate must PASS — see v1-claims-and-evidence.md for how we verify each claim.

If you already have code, don’t rewrite it into a sticky note. Just say:

“Use code-to-workflow to inventory this project, decide what stays in code vs moves to n8n; do the security audit first, then emit SDD, workflow, and validation results.”

See 02-USAGE-MODES.en.md (three intent-driven modes) and 03-FIRST-WORKFLOW.en.md (15-minute hands-on); for porting existing code, go straight to code-to-workflow.

---

📖 Reading order (strongly recommended)

---

⚡ Understand it in 90 seconds

What it does

You drop a yellow sticky note on the n8n canvas and write (in any language):

Every day at 9 AM, fetch sales data and email the daily report to my boss.
On failure, notify Slack #ops.

You ask AI to build it. The canvas now shows a complete workflow:

┌─ Yellow sticky: your requirement (preserved as-is)
├─ Middle: AI-generated nodes (Schedule → HTTP → Code → Email)
└─ Blue sticky: AI's notes (credentials needed, assumptions, limitations, how to test)

No code. No syntax to learn. No need to memorize n8n node names.

Four usage modes

The first three start from intent. The fourth starts from existing code. Full Code2n8n methodology: skills/tigerai/code-to-workflow/SKILL.md.

---

📂 Pack contents

TigerAI-Code2n8n-Skill-Pack/
├── README.md / README.zh.md ← You are here
├── CODE2N8N.md              ← Code2n8n manifesto (positioning + thesis)
├── 01-INSTALL.md/.en.md       ← Install
├── 02-USAGE-MODES.md/.en.md   ← Three intent-driven usage modes
├── 03-FIRST-WORKFLOW.md/.en.md ← Hands-on tutorial
├── 04-FAQ.md/.en.md           ← Common questions
│
├── cookbook/                  ← 8 copy-paste recipes (each has plain-language + DSL fold)
│   └── 00-INDEX.md/.en.md
│
├── skills/                    ← 15 skills (plugin.json manifest matches on-disk)
│   ├── _vendor/                  6 vendor n8n-skills (MIT)
│   └── tigerai/                  8 TigerAI execution skills
│       ├── code-to-workflow/        ← Marquee: existing code / system → n8n
│       ├── n8n-security-governance/ ← Security + version control + CI/CD + rollback gate
│       └── n8n-code-to-native/      ← Code node → native n8n nodes
│
├── spec/                      ← Technical specs (for engineers)
│   ├── sticky-note-three-layer.md
│   └── sticky-note-dsl.md
│
├── examples/google-workspace-admin-workflow/    ← 1,373-line Apps Script → n8n
├── examples/line-ai-customer-service/           ← Cloud LINE CS → n8n + admin UI
├── examples/line-ai-customer-service-onprem/    ← On-prem Docker + Qdrant RAG (NOT deployable as-is)
├── examples/tigerai-flagship/ ← 3 enterprise-grade examples (with SDD)
├── reference-workflows/       ← 2,061 public workflows (AI corpus)
├── research/                  ← Research artifacts
├── tests/                     ← Three rounds of acceptance reports
│
├── CHANGELOG.md / VERSION
├── LICENSE                    ← Pack-wide MIT license
├── install.sh / install.ps1   ← Install scripts (supports Claude Code & Antigravity)
├── .agent/workflows/          ← Antigravity-exclusive workflows (e.g., /install-n8n-pack)
└── plugin.json                ← Skill manifest

📝 The /install-n8n-pack slash command lives in .agent/workflows/install-pack.md as an Antigravity-native workflow, not a Skill — so there is intentionally no skills/tigerai/install-tigerai-n8n-pack/ folder. The manifest and disk both have 14 entries.

---

🎯 Suggested reading paths by role

I’m new to n8n (never built a workflow)

1. This file → 01-INSTALL.en.md → 03-FIRST-WORKFLOW.en.md
2. After your first workflow runs, browse cookbook/00-INDEX.en.md for your scenario
3. Stuck? → 04-FAQ.en.md

I’m experienced with n8n, evaluating this Pack

1. This file → 02-USAGE-MODES.en.md
2. Read tests/REPORT-3.en.md: historical acceptance baseline (v0.9.0 R3)
3. Browse any of the three Code2n8n case studies under examples/ for current evidence
4. Browse examples/tigerai-flagship/: enterprise-grade SDD examples

I’m an engineer / integrator

1. This file → spec/sticky-note-three-layer.md + spec/sticky-note-dsl.md
2. Porting existing code: skills/tigerai/code-to-workflow/SKILL.md
3. Security, version control, CI/CD, and rollback gate: skills/tigerai/n8n-security-governance/SKILL.md
4. Building from scratch intent: skills/tigerai/sticky-note-to-workflow/SKILL.md
5. skills/tigerai/n8n-api-bridge/SKILL.md: n8n REST API SOP
6. research/patterns.md: 7 standard skeletons + anti-patterns

I have existing code I want to move into n8n

1. Read CODE2N8N.md first to understand the “keep in code / lift to flow” split
2. Use code-to-workflow for source inventory, partitioning, and Step 1.5 security audit
3. Compare against the three empirical case studies: Google Workspace admin, LINE cloud, LINE on-prem
4. Pass static lint + n8n REST import, then end-to-end with real credentials
5. If security findings remain unfixed, per marquee skill hard rule §3, publish a SECURITY-CAVEATS.md — see the on-prem LINE CS example for the template

I’m distributing this to my team

1. This file → run 01-INSTALL.en.md end-to-end
2. Read 04-FAQ.en.md to prepare for team questions
3. Hand the entire folder to teammates and ask them to start at this README

---

✨ The three-layer structure (one diagram)

┌─────────────────────────────────────────────────────┐
│ 🟡 Layer 1 (yellow sticky): User intent              │
│    "Every day at 9 AM..."                            │
│    ← AI never modifies this. Always the source of    │
│      truth.                                          │
├─────────────────────────────────────────────────────┤
│    Layer 2: AI-generated nodes & connections        │
│    Schedule → HTTP → Code → Email                   │
├─────────────────────────────────────────────────────┤
│ 🔵 Layer 3 (blue sticky): AI's commentary            │
│    • Why each node was chosen                        │
│    • Required credentials                            │
│    • Assumptions and known limits                    │
│    • How to test                                     │
└─────────────────────────────────────────────────────┘

---

🛠️ Pain points this Pack solves

---

🧪 Code2n8n empirical case studies

The third case deliberately preserves the upstream POC’s security defects and documents them in SECURITY-CAVEATS.md. This isn’t “failed acceptance swept under the rug” — it’s Code2n8n’s core principle: AI-written software that runs is not automatically software an enterprise can deploy.

The fourth case (einvoice) is the first to ship as CLEARED with real-vendor-sandbox runtime evidence — see its final validation report.

---

🛡️ V&V + Security capabilities (what the Pack actually enforces)

Pack 對外宣稱「能補足 AI Coding 的稽核、安全、透明可控」（full story）— 此 section 給工程主管 / 稽核 / 法遵窗口看真的 ship 了什麼。

V&V two-layer gate（依 A2A directive 11 國語言 machine-readable）

§1.6 Lexical schema-before-claim rule：訊息含「validated / 驗證通過 / production-ready / X/X ok / 全綠」等受限字眼前，同訊息更早位置必須先 emit 完整 evidence schema。Critic gate lexical regex enforce — AI Coder 無法用「我覺得 OK」繞過。

4-Tier external-dependency security（v0.36-39 ship 完）

「我們對外部 GitHub 進來後有做 security check 嗎? 有對惡意程式處理嗎?」 — v0.35.0 自查答：有但是 advisory 等級。v0.36-39 升到自動 gate。詳見 external-package-security-posture.md。

惡性程式偵測（concrete patterns scanned）

security-scan.mjs MALICIOUS_JS_PATTERNS 對 workflow JSON Code 節點 jsCode 抓 9 條：

自製惡意 fixture 驗 7 error + 4 warning 全捉；Pack 30 個正當 workflow 0 false positive。

外部來 workflow 進 Pack 走 scripts/ingest-external-workflow.mjs 三道 gate（scanner 0 error / 雙人 review marker _pack_ingest 含 submitter ≠ reviewer / node digest spot-check + JSONL audit log）。

Case study 驗證信心對照

einvoice 案例是 Pack 內第一個 case study ship 真實 vendor sandbox runtime evidence。其他 case study 屬「結構層通過 + caller 需自己對接 production」。

---

📊 Historical baseline acceptance (v0.9.0 R3)

The numbers below were the real-environment baseline for three-layer workflow generation as of v0.9.0 R3; the current pack version is in VERSION, and the three Code2n8n case studies above plus tests/REPORT-v0.24.1-evidence.md are the current-version evidence layer that supersedes pure-generation acceptance.

Details: tests/REPORT-3.en.md.

---

🔢 Version & changelog

Current version: see VERSION. All changes: CHANGELOG.md.

---

📜 License

The whole pack is now MIT-licensed. See the root LICENSE file.

• skills/_vendor/: MIT — from czlonkowski/n8n-skills, see skills/_vendor/LICENSE
• reference-workflows/: MIT — from Zie619/n8n-workflows. API tokens, bearer tokens, and other secrets present in the original files have been replaced with placeholders (e.g. YOUR_API_TOKEN_HERE) before redistribution.
• examples/line-ai-customer-service-onprem/: derived from MIT-licensed scorpioliu0953/ai_customer_service, attribution chain in the example’s CREDITS.md.
• examples/einvoice-n8n/: derived from MIT-licensed paid-tw/einvoice TypeScript SDK (5 providers, MIG 4.0). The Pack’s svc/ Hono wrapper + 14 n8n workflows + SECURITY-REVIEW + Amego sandbox runtime test runner are MIT-authored by TigerAI on top of the upstream SDK. SDK adapter source code is consumed via npm install @paid-tw/einvoice* with exact-pin (per SEC-017); no SDK source is vendored into this repo.
• The rest (TigerAI-authored skills, cookbook, specs, docs, install scripts, manifesto, marquee code-to-workflow skill): MIT (Copyright © 2026 Morris Lu / TigerAI).

Full third-party notices: THIRD_PARTY_NOTICES.md.

---

🆘 Stuck?

Tell Claude / ChatGPT:

“I’m new to this. Following the TigerAI Skill Pack README, currently on [filename], hit [problem].”

The AI will diagnose. Or check 04-FAQ.en.md first.