发现最好的 DSH 插件

DeepSeek Harness 插件 — 第 120 页

39768个结果

搜索 DeepSeek Harness 社区插件,比较用途、安装方式与兼容信息,为你的 AI 工作流找到合适的扩展。

@iasiv5/dsh-skins

DSH Web 皮肤插件:多套品牌皮肤侧栏热切换。

视觉工具 · 界面扩展
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:iasiv5/dsh-skins
@infinitepersistence/dsh-serial-console

支持用户与AI共享及审计的嵌入式串口控制台

其他扩展
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:InfinitePersistence/dsh-serial-console
@jiaoqsh/dsh-document

DeepSeek Harness bundle: the read_document tool reads Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, and PDF files as Markdown for the model

效率工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add @jiaoqsh/dsh-document
@jipika/dsh-memory

Progressive long-term memory for DSH: sharded markdown bodies under ~/.dsh/memory/topics/, with only a live-generated map (shard · keywords · count · latest date) injected into the system prompt, entries pulled on demand via grep + read — so the prompt cost stays flat however large the memory grows. Two layers (global + per-project) plus the global rule files, viewable AND editable in DSH Settings, with per-layer switches and an index/full injection toggle. v0.5.0 adds enforcement hooks: a write guard on the native tool pipeline (format/secrets/history rules, optional external command judge, deny/ask with reasons fed back to the model), post-write audit log, a read reminder when memory files are read, a mandatory discipline block, and an optional injection filter command. Zero extra LLM cost, live re-read. v0.5.1 keeps the switches in ~/.dsh/memory/settings.json — read live by the plugin and edited from the settings pane over the plugin's own route, independent of the host settings API. v0.6.0 fixes the index silently degrading to a title-less directory once memory outgrew the hardcoded 9k budget (which is why an agent could see memory injected and still never read it): the index now shrinks per-shard along a newest-N-per-shard ladder under a configurable indexBudget, adds per-shard keywords (routing hint) and latest-entry dates, ships a 'check memory before starting a task' discipline, and makes full mode read the shard bodies (auto-falling back to index when over budget) instead of the retired memory.md navigation file. v0.7.0 makes 'map' the default injection mode (shard-level map only, no entry titles): retrieval = grep the memory dir with task keywords, then read the matched line; layers with <=12 entries stay inlined; the discipline block teaches that two-step path and routes cross-shard lookups (>=3 entries / >=2 shards / >=3k body) to a subagent that returns facts + line numbers instead of raw bodies. v0.9.0 adds the card-memory layer for roleplay sessions and makes every path Windows-safe: home resolution follows os.homedir() (so an undefined or empty HOME no longer roots paths at the drive root), tool and bash paths are normalized before guard matching, and drive-letter cwds produce valid project slugs. v0.9.2 fixes retrieval ranking at scale (IDF-weighted entry scores, df=0 CJK bigram fragments excluded from the coverage gate, BM25-style length normalization, preview windows anchored on the rarest matched token instead of a fixed 240-char head cut; measured locally: shards covered per top-8 2.88->5.75, largest tie block 4.38->1.13, previews containing a matched token 67%->100%): shard-level bonuses no longer inflate every entry in that shard (they were added inside the entry loop, so all hits from one shard shared an identical score and ordering collapsed to date-then-line), a per-shard quota stops one shard from filling the whole result page, and the default scope is the current project layer (other projects' shards — 60% of the pool locally — are skipped but reported in dropped.offScopeLayers/offScopeFiles, with scope:'all' to opt back in). v0.9.3 stops query function words from being searched at all (a CJK stopword set plus Latin stopwords, judged per-token so a real word that merely contains a stopword character stays searchable) and rejects sub-2-character queries, which previously collapsed into an unfiltered whole-library substring match: on the local corpus 'the/of/...'-style Chinese stopword-only queries went from 1270 hits to 0, and a single-character query from 1459 hits to a clean rejection, while the 15-query positive suite kept 14/15 recall with per-query hit counts unchanged (ranking, IDF, length normalization, per-shard quota and preview anchoring are untouched). v0.9.4 adds the mapless injection mode: the global shard map stops living in the system prompt entirely (locally measured: 188 shards = 40364 chars injected every turn) and is replaced by a fixed 370-char pointer, while shard-level routing moves into the memory_search reply as a query-scored top-8 bridge. Measured on 10 real tasks from this machine: entry-level retrieval alone finds the right shard in 6/10, the bridge in 10/10 (the 4 tasks entries miss, the bridge returns at rank #1) because a task phrased in Chinese colloquial prose cannot literally match an English shard name, and bridge scoring only needs topic-level overlap. Injection size becomes independent of memory growth and no longer invalidates the KV prefix when a memory is written. Bridge cost averages 1378 chars and only pays on turns that actually search. Also consolidates the inject-mode whitelist, which was the same ternary chain copy-pasted in three places (host 1 + client 2), into normalizeInjectMode()/INJECT_MODES plus a probe assertion that both tables agree, and keeps the per-shard over-limit warning in the bridge because the map that used to carry it is gone.

记忆与知识库 · Agent 协作 · 消息通知 · 金融数据 · 服务集成
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add @jipika/dsh-memory
@johnnren/dsh-im-humanize

Humanized fork of @xmanrui/dsh-im (base 4.38.0) for DeepSeek Harness: humanized message delivery (send delay + typing two-phase), message_break + streaming coexistence, status reactions, reply quotes, per-bot overrides, and a no_reply reclaim tool. @xmanrui/dsh-im 的拟人化分支:人化发送(延迟+输入中两阶段)、message_break 与 streaming 共存、状态表情、回复引用、per-bot 全量覆盖。

Agent 协作
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:john-walks-slow/dsh-im-humanize
@khorsheed/dsh-ankh-guard

提供基于Git HEAD绿色构建凭证的自我修改重启守护

Agent 协作 · 开发工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:Khorsheed/dsh-ankh-guard
@kirkchinese/dsh-claude-suite

Claude Code, Codex and OpenCode in DSH, in one install: Harness Forge, the capability catalog and Recollect

开发工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add 'github:kirkchinese/harness-forge#2e05143dd2407cfb899f46afc6d5fb824bf8b3d6&path:packages/harness-forge/suite'
@kirkchinese/dsh-harness-compat

Harness Forge: layered Claude Code / Codex / OpenCode compatibility for DeepSeek Harness

开发工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add 'github:kirkchinese/harness-forge#2e05143dd2407cfb899f46afc6d5fb824bf8b3d6&path:packages/harness-forge/plugin'
@kirkchinese/dsh-harness-lab

Test-only DSH plugin used by the Harness Forge conformance runner

其他扩展
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add 'github:kirkchinese/harness-forge#2e05143dd2407cfb899f46afc6d5fb824bf8b3d6&path:packages/harness-forge/testkit/lab-plugin'
@linxin666/dsh-client-ui-activity-heatmap

DSH Web 侧边栏活动热力图:常驻的 GitHub 风格年度网格,可切换显示每日提交数、Token 使用量或预估花费,并展示今日统计与模型成本。

金融数据 · 界面扩展 · 开发工具 · 效率工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:283Gawin/dsh-heatmap
@local/dsh-knowledge-base

个人知识工作台:可控隐私、首次设置、目录地图、快速收集、阅读摘录、引用问答与按需备份。

记忆与知识库 · Agent 协作
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:KingsleyAsher/Personal-Knowledge-Base
@local/dynamic-workflow

Run JSON workflows with /workflow, inspect progress, select subagent models, and resume with valid completed results preserved.

Agent 协作 · 效率工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:orion-nebula-dev/dsh-Dynamic-Workflow
@loserfox/git-identity

DSH Git 提交身份固定插件,通过注入环境变量强制使用环境自身作者身份进行提交。

开发工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:LoserFox/dsh-git-identity
@luziyang2026/dsh-question-nav

会话提问导航点及卡片悬浮预览

效率工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:AbelKeithsun/dsh-question-nav
@max-null/dsh-chat-rail

画卷式消息导航栏与会话锚定过滤插件

界面扩展 · 效率工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:Max-Null/dsh-chat-rail
@maxwell-feng/dsh-windows-ocr

DSH 本地 OCR 插件:利用 Windows 系统内置的 OCR 引擎识别图片文字,仅将文本发送给模型,支持纯文本模型处理图片。

视觉工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:maxwell-feng/dsh-windows-ocr
@michengai/dsh-pua

提供任务推进、角色风味、验收循环和状态卡片

Agent 协作
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:MichengAI/dsh-pua
@my-dsh/dsh-intelligent-ui

在聊天中渲染交互式UI的智能界面插件

界面扩展
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:my-dsh/deepseek-intelligent-ui
@nanmicoder/dsh-plugin-market

经过验证的 DeepSeek Harness 插件市场:可在 Web UI 中发现、检查、安装和移除 DSH 插件。

界面扩展
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:NanmiCoder/dsh-plugin-market
@openguardrails/dsh-tui

为 DeepSeek Harness 智能体提供的交互式终端(TUI)入口。

Agent 协作 · 界面扩展 · 开发工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:openguardrails/dsh-tui
@sakura12/oh-my-deepseekharness

多智能体协同编排层适配插件

技能扩展 · Agent 协作 · 服务集成
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:1225Sakura/oh-my-deepseekharness
@tianbuyu-wwx/dsh-hermes-link

连接 Hermes 代理与 DSH 的双向桥梁

Agent 协作 · 界面扩展 · 服务集成 · 效率工具
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:Tianbuyu-wwx/dsh-hermes-link#1c63009c138d6e46d43c4f154b9cc017cab5ae4a&path:packages/dsh-hermes-link

从发现到安装,找到适合你的插件

  1. 按任务查找

    用功能关键词或分类缩小范围,再按最近更新或 GitHub Star 排序。

  2. 了解插件能力

    查看插件简介、AI 分析和原始 README,确认它能解决什么问题。

  3. 确认后安装

    检查兼容状态、依赖范围与安装说明,复制详情页提供的命令。

关于 DSH Hub 的常见问题

DSH Hub 是什么?

DSH Hub 是 DeepSeek Harness 社区插件目录,汇集插件介绍、源码链接和安装说明,帮助你发现与比较扩展。

收录是否代表已经测试兼容?

不代表。详情页展示目录记录的兼容状态和依赖范围;未知状态不等于兼容,安装前请结合项目文档确认。

如何安装插件?

打开插件详情页,查看安装区域。有命令时可复制到你的 Harness 环境执行;没有命令时,请按项目 README 操作。