Stop Protocol

Skill Transparency & Observability Protocol (STOP)

Stop. Look inside. Understand what your skills actually do.

停下来，看清楚——你的 Skill 到底做了什么。

[English](#english) | [中文](#中文)

---

What is this?

STOP is an open specification for Agent Skill observability. It defines how skills declare their capabilities, emit execution traces, and verify outcomes — so developers and users can understand what a skill did, why it did it, and whether it succeeded.

Why?

Agent Skills today are black boxes. You invoke one, something happens (or doesn't), and you have no structured way to understand the execution. As the skill ecosystem scales (40,000+ skills on platforms like SundialHub), this opacity becomes a real problem:

• Debugging is guesswork — when a skill fails, you read logs and pray
• Trust is binary — you either trust a skill completely or not at all
• Composition is fragile — chaining skills without observability is like piping commands with no stderr
• Security review is manual — no standard way to audit what a skill actually does

STOP fixes this by bringing SRE-grade observability (Logs, Metrics, Traces) to the skill layer.

Spec Documents

Document	Description
spec/manifest.md	Skill Manifest (skill.yaml) — declarative capability description
spec/trace.md	Execution Trace format — structured runtime telemetry
spec/assertions.md	Post-condition assertions — verifiable success criteria
spec/levels.md	Observability levels — progressive adoption from L0 to L3

Quick Example

# skill.yaml (Manifest)
sop: "0.1"
name: github-issue-creator
version: 1.0.0
description: Create GitHub issues from bug reports or feature requests

inputs:
  - name: repo
    type: string
    required: true
    description: GitHub repository (owner/repo)
  - name: title
    type: string
    required: true