ds-spec-loop

English | 简体中文

A reusable Agent Skill that applies Spec programming patterns observed in DeepSeek Harness to software development in other repositories.

ds-spec-loop helps coding agents keep requirements, decisions, implementation, tests, and documentation consistent while building features, fixing bugs, changing architecture or interfaces, improving tests and development processes, or removing code.

Here, Spec programming does not mean writing a long plan before coding. It means using the repository itself as the source of truth: instructions describe constraints, current documentation describes the system as it exists, decision records preserve why choices were made, and code plus tests provide executable evidence.

🧭 Method source

This Skill is an independent community project based on patterns observable in the public DeepSeek Harness repository and its Git history. Those patterns include:

• repository-level and directory-level instructions for agents;
• current-state architecture and package documentation;
• decision records with proposed, implemented, rejected, superseded, and archived states;
• implementation across the actual integration and runtime path;
• acceptance criteria tied to direct evidence;
• updates to code, tests, generated artifacts, and documentation in the same change.

ds-spec-loop turns these observed practices into a portable Agent Skill. It is not an official DeepSeek project, is not endorsed by DeepSeek, and does not redistribute DeepSeek Harness source code.

✨ What the Skill does

Use it for feature development, bug fixes, API or configuration changes, architecture changes, test-strategy changes, development-process changes, code removal, or Spec-to-code review. A spelling correction, formatting-only edit, or other local mechanical change normally does not need this Skill.

🚀 Quick start

Install with the community-maintained cross-agent skills CLI:

npx skills add songyang0603/ds-spec-loop

Then invoke it in Codex:

Use $ds-spec-loop to implement this feature.
Before editing, inspect the repository's instructions, current documentation,
existing decisions, relevant source, tests, and actual integration path.
Keep the decision record, implementation, evidence, and documentation consistent.

Explicit invocation depends on the host:

Common usage

Write a Spec without implementing it:

Use $ds-spec-loop to investigate this change and write the proposed decision record.
Do not implement it. Describe the problem independently of the preferred solution,
compare real alternatives, and connect acceptance criteria to direct evidence.

Continue an existing proposal:

Use $ds-spec-loop to continue the existing proposed decision through implementation,
testing, documentation updates, and the correct lifecycle transition.

Review consistency without changing files:

Use $ds-spec-loop to check whether active decisions, source, public contracts,
tests, generated artifacts, and current documentation agree. Do not modify files.

Simplify or remove code:

Use $ds-spec-loop to investigate the real consumers of this code and plan its removal.
If removal is justified, update the full integration path and verify that obsolete files,
registrations, exports, tests, and documentation no longer remain.

The requested task boundary remains binding: a review-only task stays read-only, and a Spec-only task does not silently become implementation. For Claude Code or GitHub Copilot CLI, replace $ds-spec-loop with /ds-spec-loop.

🧩 How it works

The Skill treats the Spec as a set of connected repository records:

instructions + current documentation + decision records
                           ↕
       source + types + configuration + integration path
                           ↕
       tests + runnable behavior + generated artifacts

• Current documentation explains what the repository does now.
• Decision records explain why a decision was proposed, accepted, rejected, replaced, or retained.
• Source and public contracts implement the decision.
• Tests and runnable paths provide evidence for observable behavior.
• Repository-specific checks enforce facts that can be decided mechanically.

The Agent updates only the records affected by the task. It does not assume that a format check proves semantic correctness, or that changing one module means the feature works through the complete system.

📦 Installation

The Agent Skills specification standardizes the portable SKILL.md package. Discovery directories and invocation syntax remain host-specific.

Cross-agent installer

npx skills add songyang0603/ds-spec-loop

The installer discovers skills/ds-spec-loop/SKILL.md and lets you select the target Agent and installation scope.

GitHub CLI

GitHub CLI also provides a preview Agent Skills installer:

gh skill install songyang0603/ds-spec-loop ds-spec-loop --agent codex --scope user

Replace codex with claude-code or github-copilot for those hosts.

Manual installation

Clone the repository, then copy skills/ds-spec-loop into a supported discovery directory:

Codex and Claude Code also document support for linked Skill directories. For GitHub Copilot CLI, use a copy or its documented directory-registration mechanism. See the official Codex, Claude Code, and GitHub Copilot CLI documentation for current behavior.

Repository contents

skills/ds-spec-loop/
├── SKILL.md
├── agents/openai.yaml
└── references/
    ├── acceptance-and-evidence.md
    ├── adoption.md
    ├── agent-note-lifecycle.md
    ├── decision-classes.md
    ├── documentation-discipline.md
    ├── simplification.md
    ├── system-of-authority.md
    └── templates.md

Agents load the compact SKILL.md entrypoint first and read only the reference files needed for the current task.

💬 Contributing

Ideas, issues, and pull requests are welcome. If something does not work well in your repository, open an Issue and tell us what you were trying to do, which Agent you used, and what happened. That is enough to start the conversation.

See NOTICE for attribution and LICENSE for this repository's license.