Agent collaboration · DeepSeek Harness plugins — Page 200

让 AI 真正操作浏览器完成任务。支持打开网页、点击、输入、翻页、读取内容和下载文件等操作，可用于网页测试、信息采集和自动化流程，并提供域名限制等安全控制。

让 AI 项目可以在无人监管下持续推进。Orbit 会自动规划任务、分工执行、逐步检查并继续完成后续工作；不同环节可以使用不同模型，让低成本模型承担执行任务、强模型负责规划和审核，从而降低整体 AI 使用成本。Watchdog 还会监控运行异常，在任务卡死或中断时协助恢复。

让 AI 项目可以在无人监管下持续推进。Orbit 会自动规划任务、分工执行、逐步检查并继续完成后续工作；不同环节可以使用不同模型，让低成本模型承担执行任务、强模型负责规划和审核，从而降低整体 AI 使用成本。Watchdog 还会监控运行异常，在任务卡死或中断时协助恢复。

DeepSeek Harness IM bridge plugin: route Telegram & WeChat messages into per-peer agent sessions and reply back

Agent-loop change review for the DeepSeek Harness web GUI: after every agent loop it snapshots the workspace, detects file/code changes, and offers one-click revert plus a right-sidebar diff review panel. 官方双面 dsh 插件（host + client），零依赖。

Family bundle 本地多Agent: the local-agent core plus the four delegation providers (kimi / codex / claude-code / dsh) in one install. Thin meta package: a patch carrying the members' canonical rows, npm dependencies on the members, locale card metadata — no runtime code, no client half.

The session-granted dataset model tools, companion to @khorsheed/dsh-datasets: a tool-only row (provides no service) that agent presets compose to grant the datasets_* tools (`tools`: all / read / authoring / none) per session, plus their tool-guidance prompt section. Deliberately NOT self-mounting — installing it as a dependency only makes the module resolvable; an agent preset's agent.cordis.yml references the row by name.

The session-granted eval tools, companion to @khorsheed/dsh-eval: a tool-only row (provides no service) that agent presets compose to grant eval_conditions / eval_plan_validate / eval_plan_draft / eval_run_status / eval_cells (`tools`: all / none) per session, plus their tool-guidance prompt section. Four reads and one draft — eval_plan_draft writes a plan and its new conditions into the bound repository working copy and validates them, which starts nothing. Deliberately NOT self-mounting — installing it as a dependency only makes the module resolvable; an agent preset's agent.cordis.yml references the row by name.

Render agent-authored ```dsh-card HTML as sandboxed interactive iframes inline in the conversation

Local code-agent harness registry and glue: scoped homes, /<harness> login and sessions commands, delegation provider contracts, and the roster-driven browser settings section

Claude Code harness for dsh-local-agent: scoped CLAUDE_CONFIG_DIR, one-shot claude -p delegation visible in the 子代理 surface, project-file session records

Codex harness for dsh-local-agent: scoped CODEX_HOME, one-shot codex exec delegation visible in the 子代理 surface, rollout-file session records

dsh harness for dsh-local-agent: delegate to dsh itself as a local CLI — scoped DSH_HOME under the shared homes root, one-shot sub-dsh runs visible in the 子代理 surface, caller-supplied session ids, resume across rounds, DeepSeek toggle in settings

Sub-dsh headless bundle for the local-agent-dsh harness: a direct Agent/Session runner over dsh-base — one-shot (caller-supplied --session-id/--resume, prints the final assistant text, exits) or resident (--serve) driving turns over the family stdio wire

Kimi Code harness for dsh-local-agent: scoped KIMI_CODE_HOME, one-shot kimi -p delegation visible in the 子代理 surface, session records and preset wiring

Family-owned subagent delegation tool over the ctx.subagents seam with optional one-shot resume: the model may pass the dsh child session id returned by the first delegation to continue the same CLI conversation in a later round

The session-granted mission model tools, companion to @khorsheed/dsh-mission: a tool-only row (provides no service) that agent presets compose to grant the mission queue tools (`tools`: all / read / none) per session, plus their tool-guidance prompt section. Deliberately NOT self-mounting — installing it as a dependency only makes the module resolvable; an agent preset's agent.cordis.yml references the row by name.

The three community agent presets (dev / dsh-eval / dsh-writing) as declarative `@deepseek-ai/dsh-agent-preset` bundle rows — the 0.1.5 directory presets migrated to the host 0.1.7-rc.1 preset mechanism. Data only: no runtime code, no client half.

Room conversations with native or external coordinators, background delegation, durable goal plans and evidence review, shared member controls, live output and session navigation; the session-granted room model tools (`room_invite` / `room_task` / `room_message`) ship as the package's `./tool` composition entry

侧边对话：全 preset 常驻的随身 Agent——右栏轻量聊天，每个 contextKey 绑定一个持久 agent 会话，引用是不透明文本块（普通会话消息动作「引用到侧边对话」，其他插件经 ctx.sideChat.openWith 供给）

The session-granted typesafe model tool, companion to @khorsheed/dsh-typesafe: a tool-only row (provides no service) that agent presets compose to grant `typesafe_judge` plus its guidance section and the typesafe-decide skill. Deliberately NOT self-mounting — installing it as a dependency only makes the module resolvable; an agent preset's agent.cordis.yml references the row by name.

Pre-configured expert agents with automatic routing for DeepSeek Harness: fixed-persona continuable subagents (requirement analyst / code reviewer / log tracer / SQL analyst), per-expert model routes with failover, dispatch/followup/list/squad/import tools, and a routing prompt section. Dual-face dsh plugin (host + browser half), no DSH source changes.

DSH Agent 上下文浸泡器：在 System Prompt 注入五维认知图鉴，让 CoT 自行调配认知结构；末尾 JSON 快照提取链演化状态（链协议模式，prestep 零干预）

> 一个基于 CQRS 读写分离、由子代理自主进化、由人类锁定根分类骨架的 AI Agent 错误自愈框架。