fullstack-expert
by adithya-hmt
Cordis-native,基于证据的全栈工程学科,适用于DeepSeek Harness代理
Cordis-native, evidence-driven full-stack engineering discipline for DeepSeek Harness agents
安装
dsh plugin --profile web add github:adithya-hmt/fullstack-expertGitHub 源码安装:首次需按提示配置 allowBuilds 构建授权后重试
安装与环境配置指引、插件开发教程见 DSH 中文社区文档 ↗
安装即在你的机器上以你的权限运行第三方代码——它可读写文件、使用凭据、访问网络,DSH 的工具审批不会为插件代码加沙箱。「检测到 manifest」仅代表发现 dsh.bundle / dsh.plugin 清单,不构成兼容性或安全审查;安装前请审阅源码,不熟悉的插件先在不含密钥的环境试用。
README
目录
Make coding agents show their work. Fullstack Expert is a Cordis-native workflow and safety layer for DeepSeek Harness coding agents. It makes repository inspection, vertical-slice planning, sensitive-operation boundaries, and fresh verification evidence explicit—without introducing a second agent framework, shell, filesystem, browser, database, or model provider.
This is not a smarter model. It is a more disciplined engineering loop around the Harness capabilities you already use.
- Install:
QUICKSTART.md - See it in action:
docs/demo.md - Latest release: GitHub Releases
- Questions and bugs: GitHub Issues
What V0.1 provides
- A scoped methodology prompt: understand → inspect → design → plan → implement → test → review → verify.
- Two deep model-facing operations:
fullstack_plan: bounded repository inspection plus a vertical-slice plan.fullstack_check: read-only project checks and explicit evidence recording.
- A deterministic
FullstackExpertRuntimestate seam for phase, tasks, evidence, and compact event facts. - Embedded narrow skills registered through native
ctx.skills. - Complexity-governor decisions based on requirement, reuse, native platform, dependencies, and safety controls.
- Permission classification and an approval-aware pre-execute gate that delegates sensitive operations to native Harness approval and fails closed when approval is unavailable.
- Optional, read-only Supabase project evidence detection.
- Session/tool outcome observation using native
session/eventandtools/result; the Harness session log remains the durable source of truth. - Provider-neutral operation: the plugin never selects or hard-codes an LLM adapter. GPT-5.6 Max is configured in the Harness model route.
Honest scope
V0.1 does not provide a browser driver, shell executor, filesystem backend, dev-server registry, Supabase SDK, database client, deployment adapter, or model provider. It reuses those capabilities when the selected Harness composition exposes them. Browser verification is therefore guidance plus evidence recording until a stable browser tool is composed. Supabase operations are detection/advice only; migrations, linking, login, pushes, resets, pulls, function deploys, and external writes remain explicit sensitive actions.
Install as a profile bundle
Fastest path
With the DeepSeek Harness CLI installed, add the tagged release directly from GitHub:
dsh plugin --profile web add git+https://github.com/adithya-hmt/fullstack-expert.git#v0.1.1
dsh --profile web --dump-config
dsh --profile web
Replace web with your profile name. The first command installs the package and activates its declared patch layer. The second confirms the composed row before booting. Once npm publication is enabled, the shorter @deepseek-ai/fullstack-expert package spec will work too. For a complete setup, troubleshooting, and a local-checkout path, see QUICKSTART.md.
Manual profile overlay
Build/publish this package into a profile's dependency closure, then add the supplied cordis.patch.yml overlay after the profile's base bundle:
- insert:
- id: fullstack-expert
name: '@deepseek-ai/fullstack-expert'
config:
registerSkills: true
The overlay is intentionally small. The profile must already compose native Harness rows such as dsh-fs, dsh-fs-local, dsh-shell, a shell executor, dsh-tools, dsh-skill, and the relevant model adapter.
For a user-owned preset, copy the shipped standard preset rather than editing the install, then add the row to the copied agent.cordis.yml. Give the copied preset a preset.yml with a human-readable name and description, or it may appear only as its directory id in pickers. A plugin row that contributes a Service must be placed in the host composition or in an entry-local isolated realm when its consumers are session-owned.
GPT-5.6 Max / OpenAI-compatible CLI proxy
Keep provider configuration outside this package. Configure the existing OpenAI-compatible CLI proxy using the Harness provider/model settings surface, selecting:
provider: <your proxy adapter>
model: GPT-5.6 Max
The exact provider id, endpoint, credential binding, and CLI command are deployment-specific. Do not put credentials in this repository or prompt context. The plugin consumes the provider-neutral agent/session interfaces only.
Usage
- Start with a composition containing the plugin and native Harness filesystem/tool/skill rows.
- Ask for a feature, bug fix, review, refactor, or application build.
- The agent should call
fullstack_planfor substantial work, inspect actual files, state assumptions/unknowns, and implement a vertical slice with native tools. - Use
fullstack_checkfor bounded structure/command/Supabase checks and to record non-verification review facts. Verification evidence must come from a native Harness check; V0.1 rejects self-attested verification records. - Completion is only justified after a native
verificationevidence record passes and relevant failures are absent.
The native shell tool reports its [exit code: N] marker; investigate every non-zero result. For browser work, use the composed browser capability and record actual console/network/UI evidence rather than treating a build as proof.
Development
Requirements: Node.js >=22 and the Harness peer composition installed by npm ci.
npm ci
npm run verify
npm run test:consumer
npm run test:profile
test:consumer installs the packed tarball into a temporary consumer and mounts fullstack_plan and fullstack_check against the real peer packages. test:profile installs that tarball into a disposable Harness profile and checks the composed patch layer; it requires dsh and pnpm (or uses the documented pnpm fallback).
The repository is intentionally dependency-light. Source is plain ESM JavaScript and the package publishes the source entrypoints plus the Cordis patch, matching the plugin's plain-runtime contract.
Documentation
docs/architecture.mddocs/methodology.mddocs/skills.mddocs/workflows.mddocs/complexity-policy.mddocs/verification.mddocs/permissions.mddocs/supabase.mddocs/model-provider.mddocs/extending.mddocs/demo.mddocs/releasing.mddocs/reference-methodologies.mddocs/harness-seams.md
Community and promotion
If Fullstack Expert helps with a real project, share the task and verification evidence (without secrets) in an issue or discussion. Add the dsh-plugin topic to help DeepSeek Harness users discover it. A short demo is more useful than a claim: start with docs/demo.md.
License and provenance
The plugin implementation is MIT-licensed. The methodology report in docs/reference-methodologies.md records the pinned MIT sources consulted for Superpowers, mattpocock/skills, and Ponytail, with citations and a clear separation between inspiration and copied code. No reference implementation text is bundled.
原始 README: https://github.com/adithya-hmt/fullstack-expert/blob/main/README.md ↗
同类插件
查看全部 →
deepseek-harness
从仓库或系统描述生成经过校验的自包含交互式架构图、流程图、时序图、数据流图和生命周期图。

dsh-plugin
通过 DSH MCP 客户端挂载 Ouroboros 的纯配置包,在 DSH 中提供 36 个涵盖需求访谈、Seed、执行、评估与演化流程的工具。

dsh-tongflow
基于 TongFlow 的“片场”插件,用于图片、配音、音乐与视频制作:agent 为每个资产生成 TongFlow 工作流文件(.tongflow.json)并通过 TongFlow 插件执行,内嵌工作流画布,按镜头/角色/take 组织项目,附漫剧模板;以 @tongflow 开头的会话进入 Studio 界面。

helloagents
AI 编码 CLI 的工作流层:技能、项目知识、交付检查、更安全的配置写入与可恢复执行

dsh-ai-novel-writer
安装专用 AI 小说创作预设与工作台:提供带修订号的本地项目资产、紧凑侧边工作台,以及需要原生审批的逐文件变更。

rea
用 agent 逆向任何东西:从应用行为到原生二进制