OpenSpec Agent-Own SDD Skill

This repository packages an agent skill for running OpenSpec spec-driven development without a human pressing "next" between every step.

The skill treats the OpenSpec CLI as an agent-readable state machine:

• openspec status --change <name> --json exposes state and dependencies.
• openspec instructions <artifact|apply> --change <name> --json exposes the next artifact contract.
• openspec validate --strict becomes the correction oracle.
• openspec archive <name> -y closes the loop without an interactive prompt.

The result is an agent-owned loop harness: the user states intent, and the agent drives proposal, specs, design, tasks, implementation, validation, and archive until the change is complete or genuinely blocked.

Contents

• SKILL.md — the complete skill instructions.

When To Use

Use this skill when an autonomous coding agent should drive OpenSpec end to end:

• "用 OpenSpec 自主开发"
• "agent-own spec-driven dev"
• "OpenSpec 无人值守 / 非交互"
• "spec-driven 自动跑一遍"
• "drive OpenSpec from an agent"

Requirements

• Node.js >= 20.19.0
• @fission-ai/openspec
• A project that either already has openspec/ initialized or can be bootstrapped with openspec init --tools <tool-id>

Quick Start

Install OpenSpec:

npm install -g @fission-ai/openspec@latest

For a new project, initialize non-interactively:

openspec init --tools codex

For an existing OpenSpec project, avoid re-init and refresh instructions only:

openspec update

Then let the agent follow SKILL.md.

Design Principle

Do not make the human act as the event loop.

If the next step is machine-readable, deterministic enough to choose, and validated by OpenSpec, the agent should take it. Escalate only when there is a real blocker: missing critical information, a hard tool failure, or a design conflict that should not be guessed through.

License

MIT