DeepSeek Harness 插件

sdlc-harness

A DeepSeek Harness bundle that runs an engineering practice against a locally served model(英文原文)

跳到安装方式

来源信息

GitHub 仓库
brunodantas/sdlc-harness
最近更新
2026年8月18日
分类
工具与能力
GitHub stars
0
载体类型
plugin
目录证据
上游声明已找到 dsh.bundle
证据路径
package.json#dsh.bundle
核对版本
0.1.0-rc.8
上游核对日期
2026-08-21

该证据由上游目录提供。本站没有安装、运行或安全审核这个插件。

安装

默认先复制一段 Prompt,让 Agent 读 GitHub 仓库和源码;需要自己装时再切到命令。

复制这段 Prompt,发给 DSH、Codex 或其他 Agent,让它先读 GitHub 仓库和源码。

请先不要安装或执行任何命令。阅读这个插件的 GitHub 仓库、README 和关键源码,然后用清楚、直接的方式回答以下问题,帮助我判断它是否适合我的需求:

1. 这个插件是什么,解决什么问题;
2. 适合哪些用户和典型使用场景;
3. 安装后如何使用,并给出一个最小使用示例;
4. 有哪些已知限制,以及隐私、安全、兼容性或维护风险;
5. 给出“推荐 / 有条件推荐 / 不推荐”的明确建议和理由。

请区分仓库明确说明、根据源码推断和未知信息。证据不足时请明确说明,不要猜测或照抄 README。

GitHub:https://github.com/brunodantas/sdlc-harness
插件名:sdlc-harness
作者:brunodantas

检查来源文件

安装前先看这个插件目录里的 README 和其他文件。

文件资源管理器3 个文件
README.md来源说明 · 只读预览

sdlc-harness

A DeepSeek Harness bundle that runs one engineering practice, its skills and its rules, against a model you serve yourself. It is a bundle, not a fork: the substrate is pinned and never modified, and this repo adds one patch layer that supplies a persona plus one command that installs the practice's skills from their upstreams at pinned releases.

The practice is a fixed build flow. Every step is a skill, invoked by its bare name:

flowchart TB
  subgraph spec ["Spec"]
    direction LR
    A["/grill-with-docs"] --> B["/to-spec"] --> C["/to-tickets"]
  end
  subgraph ticket ["Each ticket, in its own context"]
    direction LR
    D["/to-cases"] --> E["/to-tests"] --> F["/implement"] --> G["/code-review"]
  end
  subgraph ship ["Ship"]
    I["/build-pr"]
  end
  spec --> ticket --> ship

Quickstart

Six steps, in order. No prior Claude Code experience is assumed.

1. Serve a model. Any OpenAI-compatible endpoint whose chat template advertises tool support. The substrate always sends the native tools field and has no text-parsing fallback, so a model without declared tools cannot run here.

2. Install the substrate and pnpm. Node 22 or newer; pnpm is required because dsh plugin shells out to it.

npm i -g @deepseek-ai/dsh@0.1.0-rc.7 pnpm

The version is pinned on purpose: this bundle is verified against rc.7, and the installer warns when it finds any other release.

3. Clone this repo and add it to the browser profile.

git clone https://github.com/brunodantas/sdlc-harness.git ~/sdlc-harness
dsh plugin --profile web add file:$HOME/sdlc-harness

Use the web profile (created on first use) rather than building your own: it already composes the browser surface, which is not installable on its own.

4. Install the practice.

node ~/sdlc-harness/bin/install.js

This fetches every skill at its pinned release into ~/.dsh/skills, writes the practice's rules to ~/.dsh/AGENTS.md with their on-demand reference beside them as ~/.dsh/practice-reference.md, records what it resolved in ~/.dsh/skills.lock.json, and generates an agent preset named SDLC Harness.

5. Boot. A key is required even for a local server, and any non-empty string satisfies it.

DEEPSEEK_API_KEY=local-none DEEPSEEK_BASE_URL=http://localhost:8080/v1 dsh --profile web --port 7777

6. Select the preset. Open the printed URL and, in the session's mode menu, choose SDLC Harness in place of Standard mode. This does not persist: a session started without it runs the substrate's own persona and rules, and looks identical until the agent's behaviour tells you otherwise.

Then invoke a skill by its bare name, for example /to-cases path/to/spec.md, and send it with the send button (Enter does not send while the skill autocomplete is open).

What gets installed

Forty skills, flattened into one unnamespaced root:

  • brunodantas/skills at v0.2.0 (cc6da61): build-pr, harden-tests,

ilities, to-cases, to-tests

  • mattpocock/skills at v1.2.3 (6acc160): 35 skills; the full roster is

in [skills.manifest.json](skills.manifest.json)

Both sources are MIT licensed and installed unmodified. Their licences travel with the skills and apply to your copies; this repo's own MIT licence covers only the bundle.

Known gaps

The biggest gap is judgment, not structure. The structure is all here: skills load, rules are read, the flow runs in the order the practice prescribes. But in the acceptance run, a locally served 24B ran the whole chain to completion and froze a test file that contained no assertions. Read [docs/worked-example.md](docs/worked-example.md) before trusting anything this harness produces unattended.

The rest, including everything else the acceptance run found, lives in [docs/dsh-parity-delta.md](docs/dsh-parity-delta.md). The headlines:

  • Skill tool restrictions do nothing. The frontmatter parser silently drops

allowed-tools, so every skill runs with the full tool set.

  • The hook bridge is not composed. No PostToolUse or similar automation

runs here.

  • Permission behaviour is the profile's defaults. No equivalent of

acceptEdits, and plan mode is advisory rather than enforced.

Where the rest lives

Vocabulary[CONTEXT.md](CONTEXT.md)
Ranked architectural characteristics[docs/architecture.md](docs/architecture.md)
Decisions that bind[docs/adr/](docs/adr/)
Feature-by-feature map[docs/dsh-parity-map.md](docs/dsh-parity-map.md)
The measured gap[docs/dsh-parity-delta.md](docs/dsh-parity-delta.md)
One ticket, start to finish[docs/worked-example.md](docs/worked-example.md)

Licence

MIT. See [LICENSE](LICENSE).