DeepSeek Harness Plugin Hub

发布与管理完整 Harness Profiles,发现适合你的插件。

探索

插件目录环境预设文档中心动态

社区

发布插件联系我们报告问题

相关链接

Plugin Hub GitHubDeepSeek Harness 官方项目系统状态隐私说明
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

独立、非官方社区项目,与 DeepSeek 官方无隶属、授权或背书关系。

Lazydeepseek — DeepSeek Harness 插件(DSH Plugin)
← Plugins
L

lazydeepseek

Lazydeepseek

LazyDeepSeek — LazySeries 系列中的 DeepSeek Harness (dsh) 成员:从 LazyZCode 移植而来的技能、钩子、MCP 服务器、命令、代理以及 Ralph/ULW 循环。

插件会安装到这里;不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add github:elvinzhao10/LazyDeepSeek#b351568500bc9d2cd44c571916cd70578101200e
README兼容性版本

兼容性与来源证明

Lazydeepseek 以 lazydeepseek 发布,当前版本为 1.3.5。Plugin Hub 会校验它的 manifest,并保存精确安装来源,便于复现安装结果。

DSH 兼容范围
*
运行环境
any
发布来源
github
Registry 更新时间
2026/10/3

版本

1.3.5stable
2026/10/3
1.3.4stable
2026/10/1

相关插件

正在加载相关插件…

最新版
1.3.5
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
未提供
文件数
未提供
Surface
any
许可证
MIT
发布源
github
GitHub
★ 1
周下载
0
最近提交
2026/10/3
查看源码 ↗
README Badge

点击下方 Badge 复制 Markdown,粘贴到 README 即可。

这是你的 Plugin?认领权益 · 优先安全扫描

验证 package.json 声明的 GitHub 仓库,即可管理这个公开页面。认领后,Hub 会优先安排当前版本的安全扫描,并在通过后公开展示结果。

认领这个 Plugin →
报告问题
DeepSeek Harness Plugin Hub
ProfilesPlugins分类动态文档登录管理 Profiles
ProfilesPlugins分类动态文档登录

相关插件

继续浏览 developer-tools 分类下经过校验的插件。

Web App@deepseek-ai/dsh-web-appdsh 浏览器界面捆绑包:位于 dsh-base 之上的 Web 补丁层,加上运行时粘合插件(提供前端 dist、Web 界面提示符、bash 运行时变量和 URL 行)Sdk Minimal@deepseek-ai/dsh-sdk-minimal独立的最小 SDK 配置包:JSON-RPC、一个 DeepSeek 适配器、持久化 Shell 和 JSONL 会话Sdk App@deepseek-ai/dsh-sdk-appdsh SDK 配置包:基于 dsh-base 提供 stdio JSON-RPC 服务和进程生命周期管理Experimental Inspector@deepseek-ai/dsh-experimental-inspector用于 Host 调试和 Client Runtime 检查的实验性跨 realm CDP hub

README

LazyDeepSeek

LazyDeepSeek

Describe the work. Keep the plan. Prove the result.

LazyDeepSeek helps you use structured, evidence-based workflows in DeepSeek Harness. It prepares local package assets and guidance; a host is only considered ready after it is observed in a fresh session.

Platform status · Get started · Host routes · 1.3.5 notes · Family · Docs

Current package version: v1.3.5. HOST READINESS: PENDING. Local checks and release archives prove package behavior; a fresh host session must prove loading, command/skill execution and MCP connections.

What's in 1.3.5

  • Runtime checks execute real package and lifecycle paths; unknown exercises and failed checks cannot report success.
  • Core and optional TypeScript LSP requirements are checked separately, including unsupported runtimes.
  • Hook input is bounded before parsing and stays out of process arguments.
  • Platform guides distinguish current native capabilities, legacy routes and integrations awaiting live acceptance.

This is a maintenance release. It includes the workflow foundation introduced in the family since v1.3.0 and subsequent reliability work. For Kimi and DeepSeek, that describes inherited family behavior, not prior public releases of these ports. The details below describe the cumulative v1.3.5 experience; the release notes distinguish this patch's fixes from inherited features. No new speed, token-saving or cost claim is made.

Family milestoneWhat you get in the current package
v1.3.0 foundationNatural-language entry, editable plans, durable decisions and verification tiers.
v1.3.1 reliabilityClearer execution intent, safer isolation and evidence comparisons.
v1.3.2 handoffRevision-bound verification-report contracts; generic completion APIs have separate limits.
v1.3.3 hardeningHost-specific hook, MCP and publication repairs.
v1.3.4 maintenanceTransactional run integrity, safer lifecycle and native adapter repairs.
v1.3.5 repairsReal runtime exercises, bounded hooks, dependency updates and current platform guidance.

Just ask, or use a command — both work

Two entry routes converge on the same execution authority and gates:

  • Natural language: describe the work plainly — "Fix the typo in the welcome label" — and the smallest sufficient workflow is selected and run.
  • Explicit commands: /lazy-ulw-plan <idea> builds a new plan, and /lazy-start-work <plan> executes a known plan. Same authority, same gates.

No command is required for a clear implementation request. Conversely, asking to explain, quoting a command, or saying "plan only" never touches your files: the persisted execution_intent stays plan_only until you actually ask for execution, and a vague "ok" with several open questions never grants execution by itself.

Plans you can edit while work runs

Plans are Markdown you own. Edit them mid-run; the harness reconciles your changes at execution boundaries instead of overwriting them:

  • Cosmetic wording and ordering edits preserve existing evidence.
  • Semantic edits (acceptance, dependencies, verification commands) invalidate only the affected task and its dependents — unrelated work is untouched.
  • Your checkbox is an assertion, not a verdict: a checked box alone never counts as verified completion, and unchecking reopens the task.

Decisions the harness remembers

Cross-plan decisions live in a durable ledger (.lazydeepseek/decisions/ledger.jsonl). When plan two hits a question plan one already answered — with evidence — it recalls the decision instead of re-asking you. Contradictions are surfaced as supersessions, defects become scoped corrections that block only the affected work, and nothing in memory can override your current instructions.

Verification sized to the change

The workflow calls for verification sized to the change: a documentation inspection (V0), a focused check (V1), an integration scenario (V2), or a comprehensive security/release gate (V3, normally in CI). Valid evidence may be reused while its inputs match; affected, missing or stale checks must rerun. Native execution still needs acceptance in the selected host.

Milestones, decision gates, and full state/version semantics are shared byte-identically with LazyBuddy, LazyTrae, and LazyQoder (see plugins/lazydeepseek/contracts/lazyseries-shared-semantics.v1.json). The DeepSeek Harness route boundary is unchanged: package selection never proves host activation.

Recommended: install with AI help

Open an AI coding assistant in your project and paste this:

Help me install LazyDeepSeek from https://github.com/elvinzhao10/LazyDeepSeek for this project. Read the repository's AGENTS.md and install guide. Verify the root package manifest and run safe package checks first. Guide me through the dsh git-spec install for that GitHub repository pinned to a commit sha, then verify the installed plugin in a fresh session. Ask me before installing or enabling the plugin.

The assistant can run local checks and guide the host steps. Installing or enabling the plugin grants it code-execution trust, so approve those actions in DeepSeek Harness after reviewing the source.

Manual setup

Direct profile setup

  1. Confirm the pinned host: dsh --version reports exact 0.2.0-rc.2.
  2. After approval, run the git-spec install with the repository URL pinned to a full commit sha (or the GitHub release tarball, or ./lazydeepseek for a local development checkout). The npm registry is not used for this package. Installed plugins are enabled by default; no build scripts run because the prebuilt lib/ is committed. To upgrade an existing installation, install the new verified release SHA into the same profile and restart the session so the generated hooks configuration is parsed fresh.
  3. You need Node.js LTS 24 (recommended) or 22 (supported alternative) — the lifecycle also accepts Node.js LTS 20 for compatibility — and Git on PATH for the local launchers. Try it in a new task: skills appear via the Skill tool when the model invokes a loaded skill; registered commands may appear as interactive entries such as /lazy-ulw-plan where supported; observe them rather than inferring visibility from registration.

Native onboarding

bash scripts/install.sh verifies Node.js LTS 20+ and Git, validates the root dsh bundle manifest and route contract, runs the package load-check and plugin doctor, and prints the profile-scoped dsh plugin install steps with the GitHub URL (copied to the clipboard on macOS). With --project <absolute-project-root> it additionally runs the durable lifecycle onboard. Inside DeepSeek Harness, the /lazy-onboard, /lazy-update, and /lazy-offboard commands walk onboarding, update, and receipt-safe removal step by step. The git-spec profile steps above stay canonical, and package checks never prove host readiness — the script ends with HOST READINESS: PENDING until a fresh session shows one real skill/command and all six MCP connections.

Durable lifecycle

Manual setup is available when you prefer complete control. You need Node.js LTS 24 (recommended) or 22 (supported alternative) — the lifecycle also accepts Node.js LTS 20 for compatibility — and Git, plus Python 3.10+ available as python3. Start from the verified origin https://github.com/elvinzhao10/LazyDeepSeek and follow the installation guide.

The git-spec install above is the primary route. For a durable install that survives source checkouts, run onboard once to create a durable installation; after that, use the stable launcher for update, status, and safe offboard:

node "<install-root>/LazyDeepSeek/launcher.js" status --project "<absolute-project-root>"

What “ready” means

  • Package readiness means the copied package and local checks are valid.
  • Host readiness needs a fresh host session, one real Skill or command, and every expected MCP connection.

Until that is observed, the honest result is HOST READINESS: PENDING. Local files and load checks never prove that a host loaded the plugin.

The default package doctor does not discover or execute dsh from PATH. Optional host-manifest validation is an explicit action using bash scripts/lazydeepseek-plugin-doctor.sh --host-validator /absolute/path run from plugins/lazydeepseek/. The release verifier runs classified shell regressions serially, all package tests/*.test.js with conservative Node concurrency, and Python tests under tests/ and tooling/; its JSON names those outcomes separately.

Choose one route

Pick one route per project:

  • DeepSeek Harness git-spec install (dsh-plugin-git-sha) is the default full-plugin route: skills, commands, native role tools, seven bridged hook events, and six MCP rows are composed from one bundle; fresh-session loading is observed separately.

Skills plus manual MCP connectors are a recovery-only (manual-skills-mcp-fallback) option. Do not run that fallback beside a full-plugin route for the same project. Stop the session, remove only LazyDeepSeek's previous entries through the host UI, choose one route, and start a new session to verify it.

Design mindset

Start with the result you want and how you will know it worked. Then use the smallest amount of structure that fits the task. You can simply describe the work in plain language; the modes are guidance, not commands you need to memorize. The dual-entry routing introduced in v1.3.0 picks one of these for you.

ModeUse it whenExample request
DirectThe change is small and clear.“Fix this error and run the relevant test.”
AssistedYou need help understanding an unfamiliar area or failure.“Help me find why this command fails, then verify the fix.”
PlannedThe work has several parts or important choices.“Make a plan for this feature before changing files.”
OrchestratedThe work affects a release, security, or a risky change.“Review this release and prepare it for publication.”
Long-horizonThe goal needs to continue across sessions.“Keep working on this migration with checkpoints.”

Keep host changes deliberate

LazyDeepSeek does not automate credentials, OAuth values, private registries, or trust settings. It asks for approval before any host-managed action and keeps safe package checks separate from native plugin installation and connector changes.

Package inventory

SurfaceCountRole
Skills19Host-facing workflow policies for planning, execution, review, and verification.
Commands20Named host entry points (slash menu entries) for those workflow policies.
Agents13Specialist role definitions for planning, implementation, QA, security, and context.
Hook events9 declared / 7 bridgedSessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStart, SubagentStop bridge natively. PermissionRequest and PostToolUseFailure are degraded synthesis; child events are advisory.
MCP declarations6Local services for ledger, verification, status, context, code intelligence, and docs.

LazySeries family

One workflow philosophy. Six host integrations. Choose the sibling for the host you use; each keeps its own native adapters, installation route and acceptance evidence. These packages run independently.

SiblingTarget host
LazyBuddyCodeBuddy CLI / IDE · WorkBuddy
LazyTraeTraeCode / TraeWork / TraeCode CLI
LazyQoderQoder CLI / IDE / app
LazyZCodeZCode
LazyKimiKimi Code CLI · Kimi Work (experimental)
LazyDeepSeek ← you are hereDeepSeek Harness 0.2.0-rc.2

The family shares planning, evidence, decision-memory and completion contracts. The first shared verification core is vendored in every package; product adapters keep host setup and permissions explicit. Matching contracts do not make host capabilities interchangeable. In particular, Kimi Work remains experimental for LazyKimi, and DeepSeek's synthesized events are not native hooks. Use each sibling's host guide before installation.

Technical reference and evaluation

The source-level explanation lives in docs/README.md. It maps the package structure, request flow, state model, security boundaries, MCP lifecycle, and release checks with diagrams tied to the implementation.

For a capability-by-capability comparison with the original LazyCodex design, including what LazyDeepSeek implements and where it intentionally differs, see lazydeepseek-evaluation.md.

LazyDeepSeek is primarily inspired by LazyCodex. Attribution and its relationship to OmO are recorded in NOTICE. It is an independent implementation and does not require LazyCodex or OmO at runtime.

Learn more

  • Install and verify a host
  • Workflow playbooks — how the modes pick work
  • Evidence and completion — what "done" proves
  • Host routes and recovery
  • Release notes
  • Documentation index

License

MIT. See NOTICE for attribution and provenance.

Contributing

Issues and pull requests are welcome. Read CONTRIBUTING.md for development checks, release expectations, and guidance for reporting sanitized reproduction details. Report vulnerabilities privately according to SECURITY.md.