Agent 협업 · DeepSeek Harness 플러그인 — 284 페이지

SQL 优化模式：EXPLAIN 分析、索引策略、N+1 解决、查询改写。受 wshobson/agents（38k★ MIT）启发。

Packaged multi-agent adversarial code review for DeepSeek Harness — one finder per lens in parallel, then a verifier per finding whose task is to refute it, so only confirmed defects are reported.

SQL 优化模式：EXPLAIN 分析、索引策略、N+1 解决、查询改写。受 wshobson/agents（38k★ MIT）启发。

并行多角色代码评审小队：/review 一条命令派出安全/正确性/测试/风格四名只读评审员子代理（每人可单独指定模型与思考强度），汇总为按严重度分组的结构化报告。Parallel multi-role code review squad: /review dispatches security/correctness/tests/style reviewer subagents (each on its own model) and aggregates one severity-grouped report.

Claude Code-style rewind for DeepSeek Harness: edit a past user message and re-run on a forked session (out-of-tree plugin)

Agent instruction manager for DeepSeek Harness: edit, refine, and template the AGENTS.md files the harness actually reads — the user-global ~/.dsh/AGENTS.md and per-workspace AGENTS.md — from a sidebar panel under the Skill Center.

Agent instruction manager for DeepSeek Harness: edit, refine, and template the AGENTS.md files the harness actually reads — the user-global ~/.dsh/AGENTS.md and per-workspace AGENTS.md — from a sidebar panel.

Rich branching survey system for DeepSeek Harness (DSH) Web GUI: the ask_survey tool extends the flat ask_user_question flow — up to 150 questions, per-option branching (select C -> a different range of questions follows), delayed hover insights with sources, expandable Mermaid diagrams, multi-select, free-text, quick mode (one-click whole-survey answer templates), and reroll/push/discuss pre-flight actions. Host-authoritative pending surveys, rendered as a wizard in the same conversation composer seat.

Percent-progress scoreboard for DeepSeek Harness: the tracking_write/tracking_checkpoint tools plus a docked board below the todo pill — rows with live percents derived from artifact truth, checkpoints with host-captured git state, and pursue/align/dismiss operator actions that reach the agent as instructions.

Percent-progress scoreboard for DeepSeek Harness: the tracking_write/tracking_checkpoint tools plus a docked board below the todo pill — rows with live percents derived from artifact truth, per-row record dialogs (long-form detail + clickable sources), checkpoints with host-captured git state, and pursue/delegate/scout/align operator actions that reach the agent as instructions.

DSH plugin: free web search tool (DuckDuckGo via ddg-kit with system-proxy support, automatic Bing fallback) + front-loaded vision understanding tool (vision_understand: understand images via a vision-capable model configured in DSH when the conversation model has no image modality) + composer-level paste-image capture (paste/drop an image in the conversation input → vision model → description, bypassing DSH's image gate on text-only models) + drag-drop file stash (drag/paste/attach any file → stashed into the session workspace, attachment list injected into the next message via agent/pre-step — clean draft, replayable history) + token auth toggle (turn off token verification for open access, effective immediately, persists across restarts). Official bundle plugin, install: dsh plugin --profile web add github:LingyeSoul/dsh-rider#main。DSH 插件：免费网络搜索（duckduckgo_search 工具）+ 前置视觉理解（vision_understand 工具）+ 对话输入框粘贴图片捕获（粘贴/拖入图片 → 视觉模型 → 描述，绕开 DSH 对纯文本模型的图片拦截）+ 对话窗口拖拽上传任意文件（拖入/粘贴/选择文件 → 暂存进会话工作区，附件清单随下一条消息经 agent/pre-step 自动注入，草稿零污染）+ Token 验证开关（关闭后免 token 直接访问，立即生效，重启保持）。

Discipline for one thing: the delivery has to meet what the person meant, not only what they typed. A reading with unstated needs, a plan that serves them, children that attack and watch it, and an independent review whose approval says what the passing scripts do not prove.

RingCentral Team Messaging IM channel plugin for deepseek-harness (dsh)

Zero-interruption audit + fuse blocking for DeepSeek Harness. / 零打扰 Agent 审计与保险丝拦截插件。

Role and permission layer for DeepSeek Harness: one Markdown file per agent role (persona + model + tool whitelist), least-privilege delegation through team_delegate, and a strict write/read-only split.

Host-level TOOL / MCP / SKILL permission guard for role-directed team delegation in DSH. Enforces role frontmatter fields (tools / disallowedTools / mcp_servers / skills) at the tools/pre-execute waterfall for spawned role children, and filters the model-facing skill catalog down to the role's whitelist so out-of-scope skills are invisible, not just call-denied.

Native RoleHub discovery, verified role Sessions, and Agent Team Room integration for DeepSeek Harness.

Event-sourced roleplay agent runtime for DeepSeek Harness

Schema-owned event-sourced state for roleplay agents

Global fixed Writer and isolated task-subagent catalog for Roleplay

Role-based subagent delegation for DeepSeek Harness: one delegate tool over a folder of role files, with a roster that ships and stays overridable

Roomcomm for DeepSeek Harness: shared rooms where your dsh agent talks to other AI agents (Claude Code, Codex, OpenClaw, Hermes, other dsh instances) — MCP tools + a bundled skill, zero setup

Complexity-routed subagent delegation for DeepSeek Harness: the model picks a configured runtime tier (provider, model, max tokens, reasoning effort) per delegation, with per-tier spawn chains.

A lightweight, localized task-routing plugin for DeepSeek Harness.