概览
@njuptlzf/dsh-ponytail
README / ZH
插件文档
dsh-ponytail
English · 点击下方「中文」切换中文
A DeepSeek Harness plugin that installs the Ponytail "lazy senior developer" ruleset. It mirrors upstream's two-layer split:
- an always-on, mode-switchable system-prompt section built from the compact ruleset (~2.5 KB), and
- the full ruleset + 5 companion skills as on-demand
/skills.
This repo vendors no skill content. On first boot it fetches once from upstream DietrichGebert/ponytail, caches to disk, and never re-fetches or hand-syncs afterwards.
What it does
- Always-on section (lean) — the compact ruleset (
.agents/rules/ponytail.md, ~2.5 KB) as a system-prompt section, cached at~/.dsh/ponytail-cache/compact.md. Switch the level per session with/ponytail lite|full|ultra|off;offempties the section so the tokens are actually saved. Defaultfull. - On-demand skills (user-only) — the full ruleset (
ponytail) and the 5 companions (ponytail-help/ponytail-audit/ponytail-debt/ponytail-gain/ponytail-review) written once to~/.dsh/skills/<name>/SKILL.md, invocable from the/palette. Each file's frontmatter carriesdisable-model-invocation: true, so these skills stay user-invocable (shown in/with a "user only" tag) but are not advertised to the model — noavailable_skillscatalog cost, no auto-load on "ANY coding task". The always-on compact section is the only model-facing copy.
Install
dsh plugin --profile web add @njuptlzf/dsh-ponytail
# without a global dsh:
npx @deepseek-ai/dsh plugin --profile web add @njuptlzf/dsh-ponytail
# from git instead: dsh plugin --profile web add github:njuptlzf/dsh-ponytail
Restart dsh web after install. The first boot fetches once and caches, so later boots — online or offline — never touch the network. To update the skills, delete the corresponding cache/skill files and restart.
Usage
/ponytail— report the current level;/ponytail lite|full|ultra|off— switch it. Mode is per-session (agent.id), so one conversation'soffdoes not affect another.- The 6 skills stay in the
/palette, tagged "user only"; type/and pick one (e.g./ponytail-review) to trigger it manually — the model never pulls them in unprompted.
dsh-ponytail vs codex-ponytail
| codex-ponytail (upstream Codex plugin) | dsh-ponytail (this plugin) | |
|---|---|---|
| Mechanism | Codex plugin: SessionStart + SubagentStart + UserPromptSubmit lifecycle hooks |
DSH bundle: systemPrompt section + commands service |
| Always-on content | full ruleset, mode-filtered, at SessionStart |
compact ruleset (~2.5 KB) every step |
| Per-turn cost | full rules only at session start / compact / subagent; one mode line per turn | ~2.5 KB re-rendered per step (DSH systemPrompt per-step semantics) |
off |
UserPromptSubmit stops injecting, tokens saved |
section renders empty, tokens saved (parity) |
| Levels | lite/full/ultra filter the full ruleset's per-mode rows |
level carried as header + behavior; per-mode detail lives in the on-demand full skill |
| Session isolation | single-session process → process-level flag file | multi-session process → per agent.id (== session.id) |
| Compaction | SessionStart re-fires on the compact matcher |
compaction only shatters history, never the system prompt (inherent) |
| Subagents | SubagentStart + PONYTAIL_SUBAGENT_MATCHER |
global section inherits into subagents via composition (no patch expected) |
| On-demand detail | ponytail skill |
ponytail skill (parity) |
| Fallback | inline getFallbackInstructions |
TODO (see OPTIMIZATION_PLAN.md) |
Structure
dsh-ponytail/
├── index.js # host plugin: fetch once → compact always-on + mirror full/companions
├── package.json # dsh.bundle.patch declaration
├── cordis.patch.yml # mount row
├── OPTIMIZATION_PLAN.md # full design rationale + comparison
├── README.md
└── LICENSE # MIT
Attribution
Skill content comes from DietrichGebert/ponytail (MIT), fetched at runtime. The plugin wrapper is MIT © 2026 Kasakaze (njuptlzf).
中文 · 点击上方「English」切换英文
一个 DeepSeek Harness 插件:把 Ponytail「懒高级工程师」规范接入进来,对标上游的两层结构:
- 常驻、可按档切换的 system-prompt section,用 精简规则(~2.5 KB)构建;
- 全文规则 + 5 个同伴技能作为按需
/skill。
本仓库不存技能内容——首次启动从上游 DietrichGebert/ponytail 抓取一次并落盘缓存,之后不再联网、永不手工同步。
它做什么
- 常驻 section(精简) — 精简规则(
.agents/rules/ponytail.md,~2.5 KB)作为 system-prompt section,缓存于~/.dsh/ponytail-cache/compact.md。用/ponytail lite|full|ultra|off按会话切档;off会让 section 渲染为空、真的省回 token。默认full。 - 按需 skills(仅用户可用) — 全文规则(
ponytail)和 5 个同伴(ponytail-help/ponytail-audit/ponytail-debt/ponytail-gain/ponytail-review)写入~/.dsh/skills/<name>/SKILL.md,/面板可调。每个 skill 的 frontmatter 都带disable-model-invocation: true:技能保持仅用户可调用(在/里显示、带「仅用户可用」标签),但不对模型公开——不进available_skillscatalog、不会因描述写着 "ANY coding task" 被自动加载。常驻精简 section 是唯一面向模型的那一份。
安装
dsh plugin --profile web add @njuptlzf/dsh-ponytail
# 无全局 dsh:
npx @deepseek-ai/dsh plugin --profile web add @njuptlzf/dsh-ponytail
# 或者从 git 安装:dsh plugin --profile web add github:njuptlzf/dsh-ponytail
安装后重启 dsh web:首次启动联网抓取一次并缓存,之后无论在线离线都不再联网。要更新技能,删对应缓存/技能文件再重启即可触发重抓。
使用
/ponytail— 报告当前档;/ponytail lite|full|ultra|off— 切档。档位按会话(agent.id)隔离,一个会话的off不影响其它会话。- 6 个技能常驻
/面板、标「仅用户可用」;打/看候选、手工选一个(如/ponytail-review)触发——模型不会未经请求就自动加载它们。
dsh-ponytail vs codex-ponytail
| codex-ponytail(上游官方 Codex 插件) | dsh-ponytail(本插件) | |
|---|---|---|
| 接入机制 | Codex 插件:SessionStart + SubagentStart + UserPromptSubmit 三个 lifecycle hook |
DSH bundle:systemPrompt section + commands 服务 |
| 常驻内容 | SessionStart 时注入全文(按 mode 过滤) |
每步注入精简规则(~2.5 KB) |
| 每轮成本 | 全文只在会话开始/压缩/子代理时;每轮只一行 mode 标记 | 每步重渲 ~2.5 KB(DSH systemPrompt 的 per-step 语义) |
| off | UserPromptSubmit 停止注入、省回 token |
section 渲染为空、省回 token(对齐) |
| 强度档 | lite/full/ultra 过滤全文对应档的内容 |
档位以 header + 行为承载;分档细节落在按需全文 skill |
| 会话隔离 | 单会话进程 → 进程级 flag 文件 | 多会话单进程 → 按 agent.id(== session.id)隔离 |
| 压缩/遗忘 | SessionStart 的 compact matcher 在压缩时重注入 |
压缩只 shadow 历史、不动 system prompt(天然免疫) |
| 子代理 | SubagentStart + PONYTAIL_SUBAGENT_MATCHER |
全局 section 随 composition 继承子代理(预期无需补丁) |
| 按需详情 | ponytail skill |
ponytail skill(对齐) |
| 兜底 | 内置 getFallbackInstructions |
待补(见 OPTIMIZATION_PLAN.md) |
结构
dsh-ponytail/
├── index.js # host 插件:只抓一次 → 精简常驻 + 镜像全文/同伴
├── package.json # dsh.bundle.patch 声明
├── cordis.patch.yml # 挂载行
├── OPTIMIZATION_PLAN.md # 完整设计说明与对比
├── README.md
└── LICENSE # MIT
归属
技能内容来自 DietrichGebert/ponytail(MIT),运行时抓取。插件包装层 MIT © 2026 Kasakaze (njuptlzf)。
