Get Started with the Skills Pipeline: Auto and Manual Skills

One agent.yaml. Multiple SKILL.md files. Skills run in sequence — automatically or after you approve. No orchestration code, no framework. Just a folder and YAML.

What Is the Skills Pipeline?

A skills pipeline lets you wire multiple AI skills together in a single agent.yaml. Each skill is a folder containing a SKILL.md file. Skills execute in order — and you control what runs automatically vs what waits for your approval.

OE Runtime supports two skill trigger types:

• Auto — the skill fires immediately after the previous one completes. No human needed.
• Manual — the runtime pauses and waits for your approval before running the next skill.

Both trigger types work the same way across all three run interfaces: CLI (terminal), HTTP server, and MCP (Claude Code, Cursor, Windsurf). Everything is configured in one agent.yaml — no separate chain files, no project registry.

What You Will Build

A pipeline with three skills in one folder:

1. main-task — auto: greets you and gives a history fact
2. followup — auto: summarizes the previous skill's output immediately
3. review — manual: runs only after you approve

What You Need

• Node.js 18+ for npx, or the OE Runtime binary from openenthrium.com/runtime
• An API key from OpenAI, Anthropic, or any supported LLM provider
• A terminal (for CLI and HTTP server modes) or Claude Code / Cursor / Windsurf (for MCP mode)

Create the Project Folder

Create a folder called skill-pipeline/ with this structure:

Create agent.yaml

This is the only orchestration file you need. It lists all three skills in order with their trigger types:

name: My Skills Pipeline
description: Runs three skills in sequence — two auto, one manual.

skills:
  - path: ./main-task
    trigger_type: auto

  - path: ./followup
    trigger_type: auto

  - path: ./review
    trigger_type: manual

Create the SKILL.md Files

---
name: main-task
description: Greets you and shares a fun fact about today in history.
license: Apache-2.0
metadata:
  author: Open Enthrium
  version: "1.0"
---

You are a helpful assistant.

## Step 1: Greet

Say hello, mention today's date, and give one fun fact about today in history.

---
name: followup
description: Runs automatically after main-task. Summarises its output.
license: Apache-2.0
metadata:
  author: Open Enthrium
  version: "1.0"
---

You are a summarizer assistant.

## Step 1: Summarize

The previous skill just ran and sent you its output as context.
Summarize it in one sentence and say "Auto skill complete ✅".

---
name: review
description: Runs only after you approve. Confirms the task with context from previous skills.
license: Apache-2.0
metadata:
  author: Open Enthrium
  version: "1.0"
---

You are an assistant that confirms tasks.

## Step 1: Confirm

The human approved running this skill.
Acknowledge it, briefly reference what the previous skills output,
and say "Manual approval skill complete ✅".

Create oe-config.json

{
  "llm": {
    "provider": "openai",
    "model": "gpt-4o",
    "apiKey": "sk-..."
  }
}

Replace sk-... with your actual API key. Anthropic users: set "provider": "anthropic" and "model": "claude-sonnet-5".

Run It — Three Modes

The same folder works in all three delivery modes. Pick the one that fits your workflow.

npx -y @openenthrium/oe-runtime@latest skill-pipeline/agent.yaml

{
  "llm": {
    "provider": "openai",
    "model": "gpt-4o",
    "apiKey": "sk-..."
  },
  "server": {
    "enabled": true,
    "port": 3333,
    "apiKey": "your-secret"
  }
}

npx -y @openenthrium/oe-runtime@latest --serve --config oe-config.json

curl -X POST http://localhost:3333/run-agent \
  -H "x-api-key: your-secret" \
  -H "Content-Type: application/json" \
  -d '{"file":"/path/to/skill-pipeline/agent.yaml"}'

{
  "success": true,
  "output": "The previous skill greeted you... Auto skill complete ✅",
  "pending_skill_chain": {
    "chain_id": "lc9f2kx4m",
    "skill_name": "review"
  },
  "duration_ms": 3241
}

curl -X POST http://localhost:3333/approve-chain \
  -H "x-api-key: your-secret" \
  -H "Content-Type: application/json" \
  -d '{"chain_id":"lc9f2kx4m","approved":true}'

{
  "success": true,
  "approved": true,
  "output": "The human approved! ... Manual approval skill complete ✅",
  "pending_skill_chain": null,
  "duration_ms": 1876
}

How the Skills Pipeline Works

When the runtime processes an agent.yaml, it walks the skills: list in order:

1. Auto skills run immediately. The previous skill's full output is passed as context: "Context from previous skill: [output]. Now execute your task."
2. Manual skills pause. In CLI, the runtime prints a prompt. In HTTP, a pending_skill_chain object with a chain_id is returned. In MCP, the result describes the pending skill.
3. Context flows forward — each skill receives the accumulated output of all skills before it. Nothing is lost between steps.

agent.yaml Skills Reference

skills:
  - path: ./skill-folder     # directory containing SKILL.md
    trigger_type: auto        # fires immediately after previous skill

  - path: ./another-folder
    trigger_type: manual      # pauses for human approval

• path — directory containing SKILL.md, relative to agent.yaml
• trigger_type — auto or manual (defaults to auto if omitted)
• Skills run in declaration order — top to bottom
• Each skill can use {{param}} placeholders if params are declared in agent.yaml

When to Use Each Trigger Type

• Auto — fully automated pipelines: fetch data → transform → send report → log result
• Manual — human-in-the-loop workflows: draft a document → pause for review → send to customer
• Mix them — auto-process first, then pause before taking an irreversible action