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。

Context Proxy — 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
C

dsh-context-proxy

Context Proxy

面向 DeepSeek Harness 的轻量按需上下文检索层:通过 sessionQuery 和 subprocess 接口提供 context_query / context_slice / context_grep 工具。

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:EvilIrving/dsh-context-proxy#1e1965f51ace195272697ff55b9dc8d43798e6a6
README兼容性版本

兼容性与来源证明

Context Proxy 以 dsh-context-proxy 发布,当前版本为 0.1.0。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
2026/8/21

版本

0.1.0stable
2026/8/21

相关插件

正在加载相关插件…

最新版
0.1.0
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
★ 0
周下载
0
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

相关插件

继续浏览 memory-context 分类下经过校验的插件。

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。Mnemondsh-mnemon面向 DeepSeek Harness 的可组合三层记忆控制平面:持久化运行时上下文、可搜索的项目文档、可插拔的长期记忆、受保护的策略、WebUI 和无头工具。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。Memory@furongjun1999/dsh-memory灵枢(Lingshu·líng shū)DeepSeek Harness 插件:完整大脑——长期记忆/知识飞轮/自我认知/递归反思接入 DSH,对话自动沉淀进 md_cg 认知图(md 文档)

README

dsh-context-proxy

Thin on-demand context-retrieval layer for the DeepSeek Harness.

Automatic truncation, spill-to-disk, compaction, and token metering are already owned by output-retention, spill-policy, compaction-basic, and session-query. This package adds only the missing piece: three model-facing tools that read already-persisted history back on demand.

Install

dsh plugin --profile <name> add github:EvilIrving/dsh-context-proxy

Or, from a checkout:

dsh plugin --profile <name> add ./dsh-context-proxy

The bundle patch inserts one plugin row (dsh-context-proxy). sessionQuery and subprocess are optional backends (ctx.get): without a sessionQuery backend each tool degrades to an isError result instead of waiting forever.

Tools

ToolBacking seamReturns
context_querysessionQuery.filterEventslightweight { seq, type, text } matches by seq/type/surface/text
context_slicesessionQuery.readEventone event plus a bounded before/after window
context_grepsubprocess (packaged rg) or sessionQuery.filterEventsmatches with { text, citation }

Citations are replay-safe: context_query/context_slice cite session:<seq> (rebuildable from the canonical log), and context_grep cites either a spill path (path:line, fast local path) or session:<seq> (semantic-text fallback when no spill path is supplied or available).

Config

export interface Config {
  readWindowDefault: number  // default before/after for context_slice (default 0)
  grepMaxBytes: number       // rg stdout in-memory cap (default 1 MiB)
  rgGraceMs: number          // rg termination grace (default 5000)
}

Dependencies

  • sessionQuery and subprocess are optional services read via ctx.get(...). A missing sessionQuery backend degrades each tool to an isError result (session query backend unavailable) instead of leaving the plugin waiting forever; the same applies to subprocess for context_grep with an explicit path.
  • tools is a hard dependency (inject).
  • The packaged ripgrep binary comes from @vscode/ripgrep — the same mechanism dsh-tool-fs-search uses — so no system rg install is required and no shell layer exists between the argv vector and ripgrep.

Model Experience

Request context and condition

What the model sees

Three tool schemas (context_query, context_slice, context_grep) are registered into the agent's tool set. Each result is a compact JSON object of matches/items plus citations; no prose is returned for the model to parse ids out of.

Token effect

Zero-direct effect until a tool is invoked; each invocation's cost is bounded by the query/slice window and the grepMaxBytes cap.

KV Cache effect

Append-only: tool results append to the surface; they never rewrite earlier request tokens.

Known Limitations and Deferred Work

  • context_grep line parse is path:line:text — a spill path containing a newline (not expressible in the durable locator) would confuse the parser; absolute spill paths never contain one.
  • Fallback text scan is literal, not regex — when no spill path is supplied, context_grep falls back to sessionQuery's literal, case-insensitive, whitespace-flexible scan, not a regex search.
  • Bundle writing is not sandboxed — nothing here writes files; retrieval only reads through the seams.