dsh-plugin-runbook

English · 中文

Turn any DeepSeek Harness (DSH) session into a live, scrubbable data-flow runbook — see everything your agent produced, how it connects, and rerun any step in one click.

Why

You let an agent grind for hours. It touched 300 files, ran 50 commands, spawned subagents, committed to git — and at the end you're left asking: what did it actually make, and how do the pieces fit?

The chat transcript can't answer that. A graph can.

Runbook builds a Jupyter-notebook-style DAG of the whole session: scripts, datasets, figures, docs — every node a real file, every edge real data flow. Scrub the turn slider and watch the pipeline grow like a tree. Hover a node for one-click preview / rerun / LLM-explain. Git history becomes provenance nodes; subagent work gets stitched in; a persistent ledger survives compaction and restarts.

Feature highlights

🔀 Turn scrubber	Drag from turn 1 to now — the graph grows structurally (zoom/pan stay live, no canned animation)
🖥 Mini terminal	Click ▶ on any script: a small corner terminal streams stdout/stderr + exit code — page never hijacked
✨ LLM explain	One sentence on what any file is and where it came from; racing primary × fallback models, 2–6 s typical
📜 Git provenance	Commits are first-class nodes chained in time; files hang off their commit — un-pushed history gets structure too
🤖 Subagent stitching	Child-session runs/edits/reads mined from their own logs; agent→script→artifact edges, purple-coded
💾 Disk scan + static IO	Never-ran, never-committed research code still enters the graph: read_csv/to_csv/savefig + CLI-flag adjacency + mtime inference
📖 PIPELINE.md backbone	Got a curated mermaid pipeline doc? It becomes the view. Missing files render as dashed ghosts — your gap list, visualized. No doc? A zero-API backbone is inferred from static IO
🧹 Shelf	Orphan files collapse into a per-directory tray by default — the pipeline stays the hero
🌐 Global retarget	Type any project path in the header; git/scan/backbone all retarget, persisted across sessions

All local by default. The only optional network call is ✨ explain, through your configured provider.

Install

dsh plugin --profile web add zhan-tz/dsh-plugin-runbook

Open any conversation → Runbook tab. For the backbone view, drop a PIPELINE.md (mermaid flowchart LR + status emoji) in your project root — or just let it infer one.

How it works

session timeline ──┐
git ledger ────────┼──► buildFileGraph ──► layered layout ──► SVG (scrub/zoom/hover)
subagent logs ─────┤         ▲
persistent ledger ─┤         └─ host routes (zstd, regex, static IO, cycle guards)
disk scan ─────────┘

Host (lib/index.js) exposes local routes: /agent-fileview /agent-explain /agent-run /agent-git /agent-subruns /agent-ledger /agent-scan /agent-pipeline. Client (lib/client.js) rebuilds the graph and renders it. See ROADMAP.md for the layered architecture and what's next.

Honest limits

• ✨ fallback model can take ~30 s on very long prompts (a live seconds counter shows it's working)
• The ledger preserves edges from install time; earlier history relies on git commit nodes
• Repos > 60 commits are mined by the most recent 60

Community

Part of the DSH plugin ecosystem. Feedback, issues and PRs welcome — especially pipeline-doc variants from other fields (the mermaid parser is deliberately lenient).

Keywords: dsh plugin, DeepSeek Harness, agent artifacts, provenance, data lineage, pipeline DAG, workflow visualization, reproducibility, Jupyter-style, research notebook, agent observability, multi-agent.

---

中文介绍

DSH 插件：把任何 agent 会话变成可回放、可复跑的数据流运行本。

agent 干了几小时活，碰了 300 个文件、跑了 50 条命令、起了子 agent、提交了 git——聊天记录回答不了"它到底做出了什么、怎么串起来的"。图能。

• 回合滑杆：从第一回合拖到现在，流水线像树一样长出来（缩放平移全程可用）
• 小终端：任意脚本上点 ▶，右下角小终端流出 stdout/stderr + 退出码，不劫持页面
• 悬停动作：👁 预览 / ▶ 重跑 / ✨ 一句话解释（主备模型竞速，通常 2–6 秒）
• git 出处节点：提交成链、文件挂靠——没推送的历史也有结构
• 子 agent 缝合：子会话日志挖掘，agent→脚本→产物紫色边
• 磁盘扫描 + 静态 IO：没跑过没提交的研究代码照样进图（read_csv/旗标邻接/mtime 三层推断）
• PIPELINE.md 主链：有 mermaid 流程文档就以它为骨架，缺失文件渲染为幽灵节点；没有文档就零 API 自动推断
• 孤儿收纳：默认收起按目录分组，主流水线永远是主角

安装：

dsh plugin --profile web add zhan-tz/dsh-plugin-runbook

License

MIT