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。

Project Brain — DeepSeek Harness 插件（DSH Plugin）
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
← Plugins
P

dsh-project-brain

Project Brain

DSH 的持久化项目智能与记忆插件：架构分析、跨会话上下文、TODO，以及可选的混合检索

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:yj-liuzepeng/dsh-project-brain#19acb88f20c3451bc9f9630def478363eebc2de4
README兼容性版本

兼容性与来源证明

Project Brain 以 dsh-project-brain 发布，当前版本为 1.3.1。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
web
发布来源
github
Registry 更新时间
2026/9/17

版本

1.3.1stable
2026/9/17
1.3.0stable
2026/9/15
1.2.0stable
2026/9/14
查看其余 4 个版本收起版本
1.1.1stable
2026/9/12
1.0.0stable
2026/9/9
0.7.0-beta.2beta
2026/8/30
0.7.0-beta.1beta
2026/8/30

相关插件

正在加载相关插件…

最新版
1.3.1
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
web
许可证
MIT
发布源
github
GitHub
★ 13
周下载
0
最近提交
2026/9/17
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

README

dsh-project-brain

English · 简体中文

A persistent project intelligence and memory plugin for DSH. It analyzes the active workspace, explains the architecture, preserves decisions and development history, and restores the right context in future sessions.

Instead of making every new conversation rediscover the repository, dsh-project-brain stores structured knowledge inside the project itself and keeps getting more useful as the project evolves.

Current release: 1.3.1 — adds import / export / local backup restore (cross-machine sync). End-to-end 20/20 smoke suites + 46/46 host-acceptance pass; data format fully backward-compatible with v1.3.0.

v1.3.1 新特性：📦 导入导出 / 备份恢复

  • Dashboard 顶部新增 💾 备份 / 📥 恢复 / ↶ 回滚 三个按钮（与项目名 + techStack chip 同一行）
  • 一键把整个项目脑打成 zip（含 timeline + cache + manifest + sha256 校验）→ 跨机器传文件 → 在另一台机器一键还原
  • 导入前自动预览影响清单（即将覆盖的脑 / 旧脑备份路径 / rootPath 改写）；主按钮文案「覆盖并恢复」明确提醒
  • 每次导入/回滚前自动备份当前脑到 .project-brain.backup-<ts>/；可随时回滚到任一历史备份
  • 4 个新 Tool：project_export / project_import / project_rollback_backup / project_cleanup_backups

GitHub release · DSH community showcase · MyDSH listing

Preview

Project Brain activation entry (before first scan)

Initial activation screen showing the three core capabilities (read project / remember across sessions / continue from where you left off) and the Start Project Brain button on the current workspace path

Project card with detected stack and live status

Project card showing project name, detected stack chips (GitHub Actions / Monorepo), description, last-updated timestamp, and a three-stat summary (pending TODOs / project memories / completed) with phase progress bar

Dashboard overview with SuggestionCard and Quick Actions

Dashboard header (Dashboard · project overview), SuggestionCard with AI recommendation and confidence score, four Quick Actions (rescan / organize TODOs / organize memories / project overview), and the Overview tab showing detected tech stack / language usage / dev entry points

Architecture tab — lane-based layered overview

Architecture tab with project positioning, architectural style tags, plugin-style layered architecture, and DSH LLM enhancement badge; layers shown as left-label + right-component lanes (interface / tools / storage / memory / build)

Architecture tab — expanded runtime path for a single layer

Selecting the session-extraction-and-hybrid-retrieval layer reveals its runtime path: a horizontal numbered flow (memory subsystem → embedding → hybrid retrieval → config-gated vector lookup → Brain storage) with a step-by-step description, files, and risks

TODOs and development timeline

Activity tab: completed TODOs on the left (real-time memory, session summary, cross-session continuation), and the timeline on the right showing memory-organize events, project_rescan events, and recent memory-add entries with timestamps

Project memories with type badges and importance stars

Memories tab: Current Core (per-round injection) section with type chips (decision / bug / architecture / lesson / note), importance stars, summary line, date, and View Details button per card; dormant memories collapsible at the bottom

Settings — vector retrieval and Embedding config

Settings tab · Retrieval and Embedding section: hybrid retrieval mode, enable vector switch, Embedding URL / model / API key input fields, vector dimension; top green banner reminds users that unsaved changes do not take effect

Settings — retrieval weights and probe button

Settings tab · Retrieval weights section: keyword / vector / importance / confidence / recency weight inputs (0.15 / 0.25 / 0.3 / 0.1 / 0.2), each with a one-line hint describing what it does; small light-blue Test Vector Connection button probes the live embedding endpoint

Git history tab (VSCode-style)

Git history tab: branch switcher (main, 32 commits, 1 other branch), HEAD hash, auto-refresh 30s toggle; commit timeline with graph dots, commit subject, short hash, author, relative time, and changed-file summary; branch chips on the right

Why use it?

AI coding sessions are usually short-lived, while real software projects are not. Important context gets scattered across conversations: why an architectural choice was made, which workaround fixed a production bug, what is currently in progress, and which files form the critical path.

dsh-project-brain turns that context into a local, workspace-scoped knowledge layer:

  • Understand a repository quickly — detects languages, frameworks, tooling, entry points, manifests, README metadata, symbols, and imports.
  • Explain architecture, not just folders — produces a report covering project purpose, architectural style, conceptual layers, component responsibilities, relationships, runtime flows, key files, reading order, and risks.
  • Reuse the active DSH model — architecture analysis uses the provider/model already selected for the current DSH Session; no separate chat-model key is required.
  • Remember across conversations — decisions, requirements, architecture notes, bugs, lessons, changes, TODOs, and timeline events persist in the workspace.
  • Restore context automatically — a new Session receives high-value memories, active TODOs, and recent activity through the DSH system prompt.
  • Keep workspaces isolated — the Host resolves the live Session workspace; one project's memory is never guessed from another project.
  • Work without embeddings — local BM25 retrieval is the zero-configuration default. OpenAI-compatible embeddings are an optional semantic upgrade.
  • Stay usable when AI is unavailable — architecture generation and memory retrieval have local fallback paths.

How it works

Active DSH Session
       │ trusted live Session cwd
       ▼
Workspace scanner ───────► project facts and source evidence
       │
       ├──► current DSH LLM ───► semantic architecture report
       │          │ unavailable / timeout / invalid output
       │          └────────────► local architecture fallback
       │
       ├──► .project-brain/ ───► memory, TODOs, timeline, architecture
       │
       ├──► context injector ──► next Session system prompt
       │
       └──► Connection RPC ────► Dashboard and TodoStrip

The browser-side UI cannot choose arbitrary filesystem paths. Runtime actions are resolved by the Host against the active DSH Session.

Local project data

Each workspace owns its data:

.project-brain/
├── project.json          # project metadata, stack, languages, and entry points
├── architecture.json     # purpose, layers, components, flows, and evidence
├── memory.jsonl          # long-term structured project memories
├── todo.jsonl            # project tasks and status
├── timeline.jsonl        # scans, memory events, TODO events, and sessions
├── codegraph.json        # optional legacy AST graph
└── cache/
    └── embeddings.jsonl  # optional, derived, and safe to rebuild

Project data survives plugin upgrades and conversation changes. Uninstalling the plugin does not delete .project-brain/.

Memory model

Memories can represent decision, requirement, architecture, change, bug, lesson, issue, or context.

Every new V2 memory includes importance, confidence, lifecycle status, source, timestamps, optional tags, and related files. Archived, superseded, and deleted memories are excluded from normal retrieval, automatic prompt injection, status counts, and the Dashboard.

The plugin writes memories through three paths:

  1. The Agent records durable decisions and lessons with project_memory_add.
  2. Session shutdown records a deduplicated Git change summary and uses the active DSH model to extract a small set of durable, evidence-based project facts. Credentials and non-text/tool output are excluded; unavailable or invalid model output degrades safely to Git-only memory.
  3. Architecture-diff analysis can generate architecture/change memories.

The Organize Memory action first previews merge and archive candidates. It only writes changes after confirmation.

Session extraction is configurable with sessionSemanticMemoryEnabled, sessionSemanticMaxChars, sessionSemanticMaxItems, and sessionSemanticTimeoutMs. Disable it to keep Session summarization Git-only.

Retrieval

The default local retriever uses BM25 over titles, content, tags, related files, and types. Ranking also considers importance, confidence, recency, stable memory types, and result diversity.

Optional hybrid retrieval combines keyword and vector scores. Enable it in the plugin settings:

retrievalMode: hybrid
vectorEnabled: true
embeddingBaseURL: https://your-provider.example/v1
embeddingModel: your-embedding-model
embeddingApiKeyEnv: PROJECT_BRAIN_EMBEDDING_API_KEY

Embedding indexes are updated lazily by content hash. Missing credentials, network errors, invalid vectors, or dimension mismatches automatically fall back to local keyword retrieval.

Enabling a remote embedding service sends memory titles, content, tags, and related file names to that service. Use a local model or a trusted provider for confidential repositories.

Architecture analysis

Initialization and rescans first collect deterministic local facts. When the current DSH Session has an available model route, the plugin sends bounded project evidence—README content, manifests, relative paths, symbols/imports, and selected source excerpts—to generate a conceptual architecture report.

Safety and reliability controls include:

  • no absolute paths in the architecture prompt;
  • configurable source inclusion and file/node limits;
  • strict JSON extraction, repair, and schema normalization;
  • validation of evidence paths and component relationships;
  • no model-generated HTML rendering;
  • automatic local fallback on missing routes, timeout, or invalid output;
  • architecture fingerprints to avoid unnecessary repeated analysis.

Set architectureLlmIncludeSource: false to send structure only, or architectureLlmEnabled: false for fully local analysis.

Quick start

Install the tagged public beta into the DSH profile you use:

dsh plugin --profile web add github:yj-liuzepeng/dsh-project-brain#v0.7.0-beta.2

For example, use --profile web with dsh web, or --profile desktop when your Desktop distribution runs that profile. Restart the active DSH process after installation or upgrade; for DSH Desktop, quit and reopen the application completely. The package contains prebuilt Host and Client bundles, so users do not need Node.js or a local build.

Runtime compatibility

  • The Host plugin provides workspace analysis, 13 project_* tools, persistent memory, context injection, TODOs, and architecture analysis to a compatible DSH profile.
  • The Dashboard and TodoStrip are a DSH Web client plugin (platform: web). They work in DSH Web and in Desktop distributions that host the DSH Web client.
  • A pure CLI/TUI client does not render the Dashboard. Core Host capabilities may still be used when that profile provides the plugin's required DSH services; non-Web profile coverage is still being validated during beta.

For source development, Node.js 22.19+ or 24+ is required:

git clone https://github.com/yj-liuzepeng/dsh-project-brain.git
cd dsh-project-brain
npm install
npm test
npm run build
npm run verify:release
npm run verify:install

Source contributors can link the repository into a target DSH profile workspace and enable dsh-project-brain in its bundle configuration.

After installation:

  1. Open a repository as the current DSH workspace.
  2. Open the Project tab.
  3. Click Start Project Brain.
  4. Review the Overview, Architecture, Work Activity, and Project Memory tabs.
  5. Continue in a new conversation and verify that project context is restored.

See INSTALL.md for detailed setup and troubleshooting.

Tools

CapabilityTool
Initialize or rescanproject_init, project_rescan
Status and continuationproject_status, project_continue
Long-term memoryproject_memory_add, project_memory_list
TODO managementproject_todo_add, project_todo_list, project_todo_update, project_todo_done
Query and maintenanceproject_ask, project_dream
Git + LLM architecture diffproject_diff

Dashboard quick actions run these workflows in the background and provide loading, confirmation, success, error, and retry states.

Privacy and security

  • Project knowledge stays under <workspace>/.project-brain/ unless an explicitly enabled model capability sends bounded content to its configured provider.
  • Session semantic memory sends bounded user/assistant text to the provider already selected by DSH; tool output, system messages, and recognizable credentials are removed first. Set sessionSemanticMemoryEnabled: false to disable it.
  • Client RPC cannot provide a filesystem path; the Host uses the live Session workspace.
  • Tool execution prioritizes the trusted live Session path over model-provided arguments.
  • Uninitialized projects are not silently populated by the Session summarizer.
  • Vector retrieval is disabled by default.
  • Embedding credentials are resolved through DSH Credentials or an environment-variable reference and are not written into memory files or Client RPC responses.
  • Release builds do not embed the developer's workspace, Session IDs, project memories, credentials, or local filesystem paths.
  • .project-brain/ may contain internal decisions and bug history. Review it before committing it to Git.

Development and release verification

npm test
npm run build
npm run verify:release
npm run verify:install
npm audit

The automated suite covers runtime workspace resolution, cross-workspace isolation, cross-Session memory, architecture local/LLM paths, Memory V2, BM25/hybrid retrieval, Session deduplication, scanner behavior, Dashboard/TodoStrip contracts, theme tokens, Git pack/delta handling, and LLM protocols.

verify:release validates package metadata, entry points, the npm file allowlist, and scans release files for local paths, Session IDs, credentials, private keys, logs, archives, and project-brain data. verify:install then packs the exact release artifact and installs it in a clean temporary project to catch peer-dependency or missing-build failures before users do.

Current limitations

  • Session semantic extraction is intentionally bounded and only records a few durable facts. Explicit memory/TODO tools remain the most reliable way to preserve critical intent.
  • The current DSH model route may not exist before the Session makes its first normal model request. If so, chat once and rescan.
  • project_diff uses its own OpenAI/Anthropic-compatible configuration; architecture initialization uses the current DSH Session model.
  • Git multi-pack-index (MIDX) support is not yet complete.
  • Host bundle upgrades require restarting the active DSH process; Desktop users should quit and reopen the application.

Documentation

  • 中文说明
  • Installation
  • Release checklist
  • Changelog
  • Design
  • Specification
  • Store listing

License

MIT

相关插件

继续浏览 memory-context 分类下经过校验的插件。

Memory Plugin@openviking/dsh-memory-plugin适用于 DeepSeek Harness 的 OpenViking 记忆与上下文套件Contextdsh-context用于上下文洞察和管理的 DeepSeek Harness 插件，提供上下文仪表板和上下文命令，帮助了解上下文的构成及其演变过程。Weknora@wxg-prc-cpg/dsh-weknora适用于 DeepSeek Harness (dsh) 的 WeKnora 知识检索工具：通过自有知识库进行语义搜索、文档阅读以及 RAG/代理回答。Memsearch Dsh@zilliz/memsearch-dsh适用于 DeepSeek Harness 的 MemSearch 插件：在多个代理之间共享 Markdown 记忆，支持捕获、步骤前上下文注入、记忆召回技能和技能候选审核面板。