Agent collaboration · DeepSeek Harness plugins — Page 276

Live data-flow runbook for DeepSeek Harness (DSH): scrubbable turn-by-turn DAG of every artifact an agent produced — mini-terminal rerun, LLM explain, git provenance nodes, subagent stitching, disk-scan static IO inference, PIPELINE.md backbone. Local-first.

Cut token cost without cutting model intelligence — a DeepSeek Harness (dsh) bundle plugin: structure-aware compression + TOON-style lossless encoding first + reversible spill-to-disk + real-time dashboard.

Minimal runnable DeepSeek Harness plugin scaffold: one model-facing tool registered through ctx.tools.

DeepSeek Harness plugin: cron-driven scheduled items. Each item carries a title, a prompt, and a croner expression; executions spawn a fresh agent session in a bound workspace (or the default cwd) and run on schedule or on demand. Includes a full-page management UI.

run_script tool for DeepSeek Harness: write multi-line scripts to a temp file and execute, avoiding Windows inline-quoting pitfalls.

run_script tool for DeepSeek Harness: write multi-line scripts to a temp file and execute, avoiding Windows inline-quoting pitfalls.

Plugin discoverability for DeepSeek Harness: search dsh plugins across npm and the awesome-dsh-plugin list, look up package metadata, and browse the curated list — all from inside dsh agents.

Blocks model-facing tools from reading secret and PII files (.env, keys, credentials) in DeepSeek Harness. Deny with reason via the tools/pre-execute waterfall.

Blocks model-facing tools from reading secret and PII files (.env, keys, credentials) in DeepSeek Harness. Deny with reason via the tools/pre-execute waterfall.

DeepSeek Harness 会话黑匣子插件：全量旁路采集会话事件（轮次/工具/报错），一键导出美化 Markdown / HTML 复盘报告

Render local image files inline in the DSH conversation via a global show_image tool.

DeepSeek Harness host bundle that implements slashx.request.v1 and slashx.response.v1

科研批量/多子问题任务自动拆分 + 子代理并行求解 + 结果合并（DSH 插件）

DeepSeek Harness 表情包插件：AI 在聊天时根据情境挑选并发送表情包图片（蓝发鱼娘/蓝色大肥鱼系列 109 张），聊天内直接内嵌显示。

In-harness plugin store for the dsh web GUI: a curated catalog (awesome-dsh-plugin) plus GitHub `topic:dsh-plugin` discovery, rendered as a compact settings tab. Only bundles whose root manifest declares dsh.bundle are installable.

Model-first subagent management for the DeepSeek Harness: spawn subagents defaulting to a cheap/fast worker model and encode the analyze/implement-vs-design/accept routing convention. Profile-mounted static plugin.

Task Runner 任务拆解模式：主代理把大任务拆成独立子任务、派给子代理（默认本地模型）并行执行再综合结果；并发数与 worker/fallback/orchestrator 模型可在设置面板图形化配置。安装时自动把 agent preset 装进 ~/.dsh/.agent-presets/。

{{DESCRIPTION}}

在 DSH 会话中提供终端智能体面板，作为对话与轨迹之外的第三种会话形态。

Coordinated session/workspace checkpoints, safe rewind, DAG branching, and reflection for DeepSeek Harness

Cross-session and cross-project memory manager for dsh (silent turn-end sedimentation + native web client embedding)

把 DSH 会话事件（中断 / 结束 / 报错 / 等待操作）聚合到一个 IM 通道，推到 QQ / 飞书，带可点按钮与定向回复

Low-token web search for dsh: compact, sanitized {title, url, snippet} results with a pluggable, auto-failing-over backend chain (Tavily / Exa / SearXNG / Brave / Google CSE) and a token budget you control

Remembers which DeepSeek Harness tools have been failing, across sessions, and tells the next session — so an agent stops rediscovering a dead endpoint one failed call at a time.