DSH HUB
HomePlugin StorePlugin PacksCommunityRankingsResourcesPublish Guide
Plugin source
Back to catalog

pfmiles /

pfmiles/dsh-gsd

Verified

GSD, the spec-driven AI coding workflow (72 skills + engine + guard hooks), installed into DeepSeek Harness (DSH) as a native plugin bundle with a single command.

★ 0 Stars0 Forks0 IssuesN/A Community rating0 Confirmed installs
View on GitHubProject homepage
READMESource: main@a0529acf

license: MIT DSH

dsh-gsd

给 DeepSeek Harness 一条命令装上 GSD。

dsh-gsd 把 GSD——规格驱动的 AI 编码工作流(72 个 skills + 引擎 + 守卫 hooks)——打包成 DeepSeek Harness(DSH)的原生插件。装好之后,你的会话直接拥有完整的 GSD 工作流能力。

已经用脚本装过?迁移到插件见 docs/MIGRATION.md(含回退路径与排障)。

为什么用 dsh-gsd

  • 一条命令安装。无需脚本、无需手工步骤、不写 ~/.dsh 任何文件。
  • DSH 原生插件。可在插件管理器里 enable / disable / remove,升级与回滚 = 换一个版本号重跑同一条命令。
  • 全量 GSD 能力自包含。72 个 gsd-* skills、GSD 引擎(workflows / templates / agents / commands)、守卫 hooks 全部随包携带。

快速开始

# 1. 安装
dsh plugin --profile web add dsh-gsd@1.20.1-1.dsh022

# 2. 重启 dsh web(或对应 profile 的服务)即生效

# 市场(dshmarket)与宿主 Discover 的卡片给的是不带版本号的等价命令:
dsh plugin --profile web add dsh-gsd

装好后会话内的 skill 目录会出现 72 个 gsd-*,执行 /gsd-health 可直接自检。接下来像平常一样对话即可:/gsd-new-project、/gsd-quick 等命令随手可用,直接开始推进开发。

也可以从 GitHub 仓库安装:dsh plugin --profile web add github:pfmiles/dsh-gsd

卸载 / 升级

# 卸载:下个进程启动后一切消失
dsh plugin --profile web remove dsh-gsd

# 升级 / 回滚:同一条命令换一个精确版本号
dsh plugin --profile web add dsh-gsd@1.20.2-1.dsh022   # 升级:GSD 引擎 1.20.2
dsh plugin --profile web add dsh-gsd@1.20.1-1.dsh022   # 回滚:回到 GSD 1.20.1

版本号方案

版本号 = <gsd-ver>-<rev>.dsh<line>;升级与回滚永远是同一条命令换一个精确版本号。安装只认精确版本(prerelease 版本不参与 range 匹配,@^/@~/@*/@latest 形态匹配数为 0)。

字段 含义 示例(1.20.1-1.dsh022) EN
gsd-ver 打包的 GSD Core 引擎版本 1.20.1 bundled GSD Core engine version
rev 修订号:同一 (gsd-ver, DSH 行) 对内从 1 起、逐次递增 1 revision id — starts at 1, +1 per release of the same pair
dsh<line> 适配的 DSH 发布行(去点标签):dsh022 = DSH 0.2.2 行、dsh021 = DSH 0.2.1 行 dsh022 dotless DSH line label

环境要求

  • DeepSeek Harness >=0.2.2-rc.1 <0.3.0-0(engines.dsh 自动校验,不满足会拒绝安装)
  • Node ≥ 24
  • macOS / Linux / WSL

版本不匹配被 engines.dsh 拒绝时,宿主给出的 allow-version 是显式风险豁免(精确 name@version × 精确 DSH 版本 + acceptRisk)——本插件不推荐使用,也绝不会自动触发。

当前状态

首个正式版本 1.20.1-1.dsh022 已就绪:8 项 UAT 已在干净机器(一次性 /tmp 新 DSH_HOME,零 GSD 残留)全部通过;npm 首发、GitHub Release 与 dshmarket 收录随后落地。本版本的变更见 CHANGELOG.md 的对应条目。

启动开销

插件给每次会话启动加了多少开销?三点测量,口径固定为事件时间戳之差:会话记录里 session.createdAt 到首个 turn/start.time 的差(两个值都由 DSH 以毫秒写进 session.v4.jsonl.zstd)。

点 含义 实测
① 基线启动 无插件 profile 9 ms
② 带插件启动 同一 home、同一模板 + 插件 279 ms
③ 插件段 ① → ② 的差 ≈270 ms

测量条件:同一 DSH_HOME、同一 --from-default-profile headless 模板成对造 profile(只差一次插件安装)、两 profile 交替各 3 轮(n=3)、单机、暖页缓存。差值是中位数之差,不是每轮配对差的中位数。

墙钟不可用作开销。 同一轮从进程启动到答案返回的墙钟在 2 560–3 834 ms 之间漂移,而它由 LLM 往返主导 —— 用它算「插件成本」等于在量模型延迟。 技能目录的 watch(长驻进程热重载)是 not measured。 一次性会话每次都是新进程、启动时重读技能目录,watcher 对它结构性不适用;而数 fsevents 需要 root。本段只给启动开销,不给 watch 数字。

自举构建(顺带一提)

这个项目由它自己开发:插件装入验证实例后,由它自带的 skills / 引擎 / hooks 驱动本仓库的全部开发流程(dogfooding)。感兴趣可看 .planning/。

构建者与复验者见 BUILD.md——GSD_VERSION 用法、在线/离线四档、exit codes 0/1/2、gates 契约、上游完整性交叉核对与验证实例复验序;上游包元数据(SBOM)见 THIRD-PARTY-NOTICES.md。

License

MIT,见 LICENSE。含上游内容:GSD Core (MIT, © 2026 Open GSD)、DeepSeek Harness (MIT)。


dsh-gsd

GSD for DeepSeek Harness, one command away.

dsh-gsd packages GSD — a spec-driven AI coding workflow (72 skills + engine + guard hooks) — as a native DeepSeek Harness (DSH) plugin. Once installed, full GSD workflow capabilities are available in your sessions.

Installed with the shell script before? See docs/MIGRATION.md to migrate (rollback and troubleshooting included).

Why dsh-gsd

  • One-command install. No scripts, no manual steps, nothing written to ~/.dsh.
  • A native DSH plugin. Enable / disable / remove in the plugin manager; upgrade or rollback = re-run the same command with a different version.
  • Fully self-contained. All 72 gsd-* skills, the GSD engine (workflows / templates / agents / commands) and guard hooks ship inside the package.

Quick Start

# 1. Install
dsh plugin --profile web add dsh-gsd@1.20.1-1.dsh022

# 2. Restart dsh web (or the target profile's service) — done.

# The dshmarket / host Discover card generates the version-less equivalent:
dsh plugin --profile web add dsh-gsd

After restart, 72 gsd-* skills appear in the session skill catalog, and /gsd-health runs a self-check. From there just talk to your agent as usual — /gsd-new-project, /gsd-quick, and 70 more commands are ready to drive your development.

GitHub install: dsh plugin --profile web add github:pfmiles/dsh-gsd

Uninstall / Upgrade

# Uninstall — everything disappears on the next process start
dsh plugin --profile web remove dsh-gsd

# Upgrade / rollback — re-run the same command with a different exact version
dsh plugin --profile web add dsh-gsd@1.20.2-1.dsh022   # upgrade: GSD engine 1.20.2
dsh plugin --profile web add dsh-gsd@1.20.1-1.dsh022   # rollback: back to GSD 1.20.1

Version scheme

Version = <gsd-ver>-<rev>.dsh<line>; upgrade and rollback are always the same command with a different exact version. Install specs are exact only — prerelease versions do not participate in range matching, and @^/@~/@*/@latest forms match zero candidates.

Field Meaning Example (1.20.1-1.dsh022) EN
gsd-ver the bundled GSD Core engine version 1.20.1 bundled engine
rev revision id — starts at 1, increments within the same (gsd-ver, DSH line) pair 1 revision
dsh<line> dotless DSH line label — dsh022 = DSH 0.2.2 line, dsh021 = DSH 0.2.1 line dsh022 line label

Requirements

  • DeepSeek Harness >=0.2.2-rc.1 <0.3.0-0 (engines.dsh is enforced automatically; incompatible hosts are rejected)
  • Node ≥ 24
  • macOS / Linux / WSL

When a mismatch is rejected via engines.dsh, the host-offered allow-version is an explicit risk exemption (exact name@version × exact DSH version + acceptRisk) — this plugin does not recommend it and never triggers it automatically.

Status

The first public release, 1.20.1-1.dsh022, is ready: all 8 UAT items pass on a clean machine (a throwaway /tmp DSH_HOME with zero GSD residue); the npm first publish, the GitHub Release and the dshmarket listing follow. What changed in this release is in the matching CHANGELOG.md entry.

Startup overhead

What does the plugin add to session startup? Three points, measured on a fixed clock: the difference between session.createdAt and the first turn/start.time in the session record (both written by DSH in milliseconds into session.v4.jsonl.zstd).

Point Meaning Measured
① baseline profile without the plugin 9 ms
② with plugin same home, same template + plugin 279 ms
③ plugin segment ① → ② delta ≈270 ms

Setup: one DSH_HOME, paired profiles built from the same --from-default-profile headless template (differing only by one plugin install), alternated over 3 rounds each (n=3), single machine, warm page cache. The delta is a difference of medians, not the median of per-round paired differences.

Wall clock is not a cost measure. Time from process start to answer on the same setup drifts between 2 560 and 3 834 ms, dominated by LLM round-trips — using it as "plugin cost" measures model latency. Skill-directory watch (hot reload in long-lived processes) is not measured. One-shot sessions are a fresh process every time and re-read the skill directory at startup, so a watcher structurally does not apply; counting fsevents needs root. This section reports startup cost only — no watch numbers.

Bootstrapped (a fun aside)

This project is built by itself: once installed into a verification instance, its own skills / engine / hooks drive this repo's entire development process (dogfooding). Curious? See .planning/.

Builders and re-verifiers: BUILD.md covers GSD_VERSION usage, the four offline/online tiers, exit codes 0/1/2, the gates contract, upstream integrity cross-checks, and the verification-instance re-verify sequence; upstream package metadata (SBOM) lives in THIRD-PARTY-NOTICES.md.

License

MIT — see LICENSE. Bundles upstream content: GSD Core (MIT, © 2026 Open GSD), DeepSeek Harness (MIT).

DSH HUB

A community index for DSH plugins. Not an official GitHub or DeepSeek AI product.

CommunityResourcesAPIAbout
—/ 5

No ratings yet

Verified DSH bundle

Commit a0529acf57e1

Community comments

No comments yet. Be the first to write one.