Agent 連携 · DeepSeek Harness プラグイン — 211 ページ

Skills, MCP and instruction-file manager for the dsh web GUI: browse/enable/import/delete skills across project and user roots, manage MCP servers with a REAL connection through @deepseek-ai/dsh-mcp-client (enabled servers connect and register their tools on ctx.tools; enable/disable actually connects/disconnects), and read/edit/create/delete the AGENTS.md-family files the harness reads every turn. Hot-pluggable — mounted via a profile bundle patch + node_modules symlink, no dsh source changes.

DeepSeek Harness (dsh) skill plugin: extract structured data from JavaScript single-page apps when a page fetch times out or returns only obfuscated/bundled JS. Registers the spa-api-extract skill on ctx.skills. Pipeline: curl raw HTML -> grep for fetch()/api endpoints and script src -> request the JSON API (carrying the token query param) -> parse with Python to a UTF-8 file. Read-only by default; no telemetry.

Model-level proxy routing + retry for DSH LLM requests, configurable live from the DSH 设置 page (模型代理). Selected models' baseURL hosts route through the configured proxy (proxyHost:proxyPort) while everything else stays direct, via a custom undici global dispatcher (RetryAgent over RoutingDispatcher); loopback is always direct.

DSH command-line tool quarantine zone: intercept direct shell execution, route every command through submit/review/execute-by-id with full audit

GitHub CLI 管理插件：gh 安装检测/自愈（含 Windows 网络代理自愈）、多账户认证管理、Git 仓库高频操作，全部在 DSH 设置页（GitHub CLI）呈现。

Persistent Playwright browser tools and interactive browser panel for DeepSeek Harness

DSH 客户端插件：把 Agent 输出当答卷批改 —— 选中文字打批注、打分（100 分制扣分/加分），本地高亮留痕，一键把结构化批改反馈发给 Agent

DSH LLM provider hub with a native model-configuration panel, gateway routes, model catalogs and live model discovery. Supports three wire protocols, custom User-Agent, model overrides and live routing without hand-editing settings.yaml.

Replay real DeepSeek Harness turns against Standard, Minimal, Anchored, or plugin candidates with frozen request-surface evidence

Claude Code-style `paths:` rule injection for DeepSeek Harness (DSH): inject rules from ~/.dsh/rules, <project>/.dsh/rules and <project>/.claude/rules into the model context.

GitHub CI co-pilot for DeepSeek Harness: PR review, CI failure fixing, issue triage and release notes. Everything is a plugin.

Feishu/Lark document design & layout skill for DeepSeek Harness (dsh) — restructure long-form docs into management-readable, brand-consistent, evidence-traceable formats with charts, illustrations & covers. 飞书文档设计排版技能插件：将战略方案/经营报告/GTM/PRD/复盘等长文编排为管理层可扫描、品牌一致、证据可追溯的正式文档。

Feishu/Lark bid-proposal document generator skill for DeepSeek Harness (dsh) — structured pitch documents with cover, background, analysis, solution, pricing, team & cases. 飞书讲标方案文档生成器技能插件：自动生成封面/项目背景/需求分析/解决方案/服务内容/报价明细/团队介绍/案例展示等章节。

Generate structured Feishu/Lark spreadsheet quotes with auto cost & profit calculation. 飞书电子表格报价单生成器：自动计算成本、人力、利润率、付款节点，含项目明细、资源分配、ROI 预估等工作表。A DeepSeek Harness (dsh) plugin bundle that installs the lark-quote-sheet-generator skill.

Evidence-driven Feishu/Lark document writing skill for DeepSeek Harness (dsh) — turn Feishu wiki/docs/minutes/Base/chats & local files into structured docs with conclusions, evidence & actions. 飞书文档证据化编写技能插件：从业务资料到可验证、结构化、可执行的飞书成稿（战略方案/经营分析/GTM/复盘/PRD/管理汇报等）。

Thaliris plugin for DeepSeek Harness using native agents, subagents, Settings and shared Web/Desktop client extensions.

DSH plugin that injects a solo orchestrator preset with explore/editor/verify/general subagents and subagent-only minimal anchoring.

Generic workspace base for the dsh harness (local + remote workspace providers with per-workspace routing), with SSH as one concrete provider: enter SSH mode from the session header, browse/edit remote files in a new left panel, let the local harness operate the remote server through remote_* agent tools, and manage hosts through the ssh_* tools and the SSH panel (legacy dsh-ssh merged in).

DSH Profile Bundle composing token optimization with Lean formal-proof plugins

Token-parameter, evaluation, semantic-update, and optimizer-provider contracts

Git-backed parameter, insight, and optimization state transitions

Domain-neutral DSH trajectory projection and token accounting

Composite Lean kernel and agentic white-box Loss providers

Agentic group-relative semantic optimizer for token parameters