daily_hacker_news_digest

Requirements

• Node.js ≥ 20
• An OpenAI-compatible LLM API (OpenAI, Groq, OpenRouter, Ollama, etc.)

Install

Install from a clone of this repository (the package is not published to npm yet):

git clone https://github.com/just-claw-it/n8n-to-claw.git
cd n8n-to-claw
npm install
npm run build

Run the CLI:

node dist/cli/index.js convert workflow.json

Optional — install the n8n-to-claw command globally from the repo:

npm install -g .
n8n-to-claw convert workflow.json

Usage

From a local workflow JSON file:

n8n-to-claw convert workflow.json

From the n8n REST API:

n8n-to-claw convert \
  --n8n-url https://my-n8n.example.com \
  --api-key <n8n-api-key> \
  --workflow-id <workflow-id>

Useful flags:

# Parse only — summarise nodes/warnings, skip LLM
n8n-to-claw convert workflow.json --dry-run

# Print full IR + prompt + known node list, then exit (no LLM call)
n8n-to-claw convert workflow.json --inspect

# Write a reproducible debug bundle (IR + prompts + LLM + validation traces)
n8n-to-claw convert workflow.json --debug-bundle

# Print LLM prompt + raw response + tsc output to stderr
n8n-to-claw convert workflow.json --verbose

# Override output directory
n8n-to-claw convert workflow.json --output-dir ~/my-skills

# Overwrite an existing skill output without prompting
n8n-to-claw convert workflow.json --force

# Print CLI version
n8n-to-claw --version

# Verify LLM_* and that the OpenAI-compatible endpoint is reachable (use before convert, especially local Ollama)
n8n-to-claw check-llm

Debug logging:

DEBUG=n8n-to-claw n8n-to-claw convert workflow.json

Debug bundle artifacts (--debug-bundle):

• debug-bundle/meta.json — source, status, model/timeouts, retry history
• debug-bundle/ir.json — normalized IR snapshot (raw payload stripped)
• debug-bundle/warnings.parse.json + warnings.transpile.json
• debug-bundle/prompt-attempt-<n>.txt + llm-response-attempt-<n>.txt
• debug-bundle/validation-attempt-<n>.json

Environment variables

Required:

Optional:

Copy .env.example to .env, fill in values, then source .env (or export the variables in your shell).

After setting LLM_*, run n8n-to-claw check-llm to confirm connectivity. That uses the same env as convert and fails fast with hints if the URL is wrong, Ollama is not running, or a remote agent cannot reach your machine’s localhost.

Example (OpenAI):

export LLM_BASE_URL=https://api.openai.com/v1
export LLM_API_KEY=sk-...
export LLM_MODEL=gpt-4o

Example (Groq / GroqCloud — not xAI Grok):

export LLM_BASE_URL=https://api.groq.com/openai/v1
export LLM_API_KEY=gsk_...
export LLM_MODEL=llama-3.3-70b-versatile

Example (local Ollama — transpile sends a large system prompt + workflow; an 8B CPU model can take many minutes per attempt. ollama run "ping" is not comparable workload.)

export LLM_BASE_URL=http://127.0.0.1:11434/v1
export LLM_API_KEY=ollama
export LLM_MODEL=llama3.1:8b
export LLM_TIMEOUT_MS=900000
# Optional: cap output length to reduce generation time (may truncate; retry may still help)
# export LLM_MAX_TOKENS=2048

Output

Skills are written to ~/.openclaw/workspace/skills/<workflow-name>/:

~/.openclaw/workspace/skills/my-workflow/
├── SKILL.md                  # OpenClaw skill descriptor
├── skill.ts                  # Node.js implementation
├── credentials.example.env   # Generated if the workflow uses credentials
├── warnings.json             # List of every degraded/stubbed node
└── skill-meta.json           # Provenance: tool version, workflow fingerprint, transpile status

If the generated skill.ts fails TypeScript validation after two LLM attempts, the output goes to draft/ instead:

~/.openclaw/workspace/skills/my-workflow/
├── draft/
│   ├── SKILL.md
│   └── skill.ts              # Fix TypeScript errors, then move up one level
├── warnings.json
└── skill-meta.json           # Still at skill root (includes status: draft when applicable)

Model recommendation

Minimum: GPT-4o or Claude Sonnet tier.

Small Ollama models (7B–13B) will produce syntactically broken or logically incorrect skill.ts output. The tool will catch compilation errors and retry once with the error injected into the prompt, but logical correctness is not guaranteed even with large models — always review generated skills before using them.

Pipeline

┌─────────┐    ┌─────────────────┐    ┌───────────────────────┐    ┌─────────┐
│  Input  │───▶│  Parse (IR)     │───▶│  Transpile (LLM)      │───▶│ Package │
│ file/   │    │  normalize n8n  │    │  prompt → SKILL.md +  │    │ write   │
│ n8n API │    │  JSON to IR     │    │  skill.ts → tsc check │    │ to disk │
└─────────┘    └─────────────────┘    └───────────────────────┘    └─────────┘

Graceful degradation

All degraded nodes are listed in warnings.json with name, type, and reason.

Node coverage

A generated dashboard lists every node type that appears in test-fixtures/, its resolved category, and whether it was mapped via EXACT_MAP, a prefix/suffix fallback, or unknown. Regenerate it with npm run coverage:nodes (requires a successful npm run build). After npm run build, you can refresh only the markdown with npm run coverage:nodes:write. CI fails if the committed file does not match a fresh regeneration.

**479 node t