Introduction

Coding agents have a common flaw: when a single function could solve a problem, they build an abstraction layer, add a configuration set, and append irrelevant error handling. To constrain this behavior, the common approach is to stuff a “Please follow YAGNI” into the system prompt. The effect varies by model, and you have to rewrite it every time you switch profiles.

dsh-ponytail takes a different path. The DSH (DeepSeek Harness) philosophy is “Everything is a plugin”. This plugin turns ponytail’s “laziest senior engineer” coding style into 6 skills distributed with the package. After registering with the skill registry, they can be used without the user writing any configuration. Below is an introduction to its features, installation, and how it works.

What is this

One-sentence positioning: dsh-ponytail (npm package name dsh-ponytail-skills) is a skill plugin that ported DietrichGebert/ponytail to the DeepSeek Harness Cordis plugin architecture. Current version 0.1.3, MIT license, maintained by gongyijie85.

It should be noted that this is an unofficial port: the skill content is adapted from upstream ponytail (MIT, © DietrichGebert), and the DSH port part is © gongyijie85. The skill files are identical to upstream (verbatim), and metadata like argument-hint and license are passed through unchanged.

Six Skills

The plugin registers skill providers to the host layer of the ctx.skills registry. Six skills are distributed with the package, located at skills/<name>/SKILL.md:

Skill Purpose
ponytail Core mode: Enforces “laziest but functional solution”, YAGNI → stdlib → native → one-liner → minimal code, supports lite / full (default) / ultra intensity levels
ponytail-review Code review specifically for over-engineering, one finding per line (position, what to delete, what to replace with)
ponytail-audit Repository-wide over-engineering audit (repository version of ponytail-review), sorted output “maximum cut first”
ponytail-debt Collects all ponytail: comments in the code into a debt ledger, preventing delayed rot
ponytail-gain Shows ponytail’s practical gain scoreboard (median baseline)
ponytail-help Cheat sheet for all modes/skills/commands

All skills are callable by both the model and the user (upstream does not set disable-model-invocation).

Installation and Activation

Supported DSH version is >=0.1.1-rc.2 (verified in production, compatible with higher versions). The npm package name dsh-ponytail is already taken by a project with the same name, so this package is published as dsh-ponytail-skills. Three installation methods:

# npm
dsh plugin --profile web add dsh-ponytail-skills

# GitHub
dsh plugin --profile web add github:gongyijie85/dsh-ponytail

# Local development
dsh plugin --profile web add D:\plugins\dsh-ponytail

After installation, restart the profile (dsh web), and the skills can be loaded using the skill tool. The entire plugin has zero runtime dependencies and only uses Node built-in modules.

How it Works

The plugin has two layers:

  1. Bundle Layer: cordis.patch.yml inserts the plugin line into the dsh-base layer, allowing subsequent layers to locate it by ID.
  2. Provider: lib/index.js calls ctx.skills.registerProvider(...), scans the skills/ directory, parses name / description from YAML frontmatter, and returns the complete skill definition, with resourceBase pointing to the skill directory.

A noteworthy detail is that the upstream skill files use description: > to fold multi-line frontmatter; the provider parses the folded scalar by connecting it with a single space according to YAML semantics. Therefore, the skill content can be used verbatim without any rewriting.

Typical Usage

After installation, tell the agent “Write this code in ponytail mode”, or let the model automatically adopt it during coding. The priority order of the mode is:

YAGNI → Reuse existing code → Standard library → Native platform capabilities → One-liner → Minimal code

Intensity levels are divided into three tiers: lite / full (default) / ultra. Just say “stop ponytail” to exit.

Applicable Scenarios and Precautions

Suitable for two types of scenarios: first, developers doing coding tasks on DSH who want concise implementations by default; second, those wanting to clean up over-engineering in existing code, who can use ponytail-review and ponytail-audit to scan line-by-line and repository-by-repository respectively.

Please note before installation:

  • The plugin runs with the permissions of the current dsh process; it is recommended to check the source code and license before installing.
  • This is an unofficial port, license is MIT: Skill content © DietrichGebert, DSH port © gongyijie85, see the LICENSE file in the repository for details.

Summary

What dsh-ponytail does isn’t complicated: it takes a mature “laziest senior engineer” style and installs it into the host layer of ctx.skills via DSH’s plugin mechanism. Six skills cover scenarios like coding, reviewing, auditing, and debt recording. Zero runtime dependencies; install and restart the profile to use.

  • Community directory page (independent site, no official affiliation with DeepSeek / Hyperscience): https://www.skillhub.cn/plugins/gongyijie85/dsh-ponytail
  • GitHub repository: https://github.com/gongyijie85/dsh-ponytail