dsh-mindmap

dsh-mindmap

简体中文

A DeepSeek Harness plugin that turns a plain Markdown file into a live mindmap. The working directory is the document: the chat is the editor, the AI edits the .md step by step, and the right-side floating panel re-renders the mindmap in real time.

Project status: pre-1.0. The current feature set (see CHANGELOG) is implemented and covered by unit tests, but cross-version compatibility beyond the development environment is not yet certified.

The core idea

• Open a Markdown file — it is a mindmap.
• The chat is not the main character; it is the assistant that edits the mindmap next to you.
• You say one sentence, the AI edits the .md one step, the panel follows instantly.
• Mindmap = Markdown: diffable, shareable, and git-friendly by nature.

Features

• Four tools (mindmap_create / mindmap_open / mindmap_get / mindmap_update) — plain Markdown files in the session working directory; the root node title is the filename and stays in sync both ways (renameRoot renames the file, collisions are rejected).
• Live panel with zero extra channels — the panel consumes the session snapshot (mindmap_* tool results), so every AI edit re-renders immediately.
• Reliable open, recoverable loading state — AI create/open results always expand the panel (a structural-fingerprint selector drives snapshot recomputation even when the host reuses the nodes array reference); clicking a .md first reads it through the local read-only route, while AI fallback loading still recovers from case-only path mismatches, inline tool errors, and a ~30s watchdog timeout with one-click retry.
• Refresh from disk — 刷新脑图 reloads the current file while preserving drafts and attachments, without submitting a model request. Fresh disk content stays visible until a later tool result arrives; overlapping reads, closing, and session changes cannot let an old request replace the new view. Failed refreshes retain the canvas and show a retryable error.
• Draft protection — uses the host's published input snapshot for text, references, and attachments. Pending or unreadable input is preserved; files still open through the read-only route and editing commands remain available for manual submission.
• Graceful parse failure — if a document's Markdown ever breaks the parser, the panel never crashes: the mindmap area shows an explicit parse-failed state with the error, the original file is left untouched, and you can switch back to the directory tab or let the AI fix the content and reopen. Normal documents never reach this path.
• Floating right panel or native sidebar tab — when dsh-better-sidebar is installed, the mindmap registers as a native single-instance tab (dsh-mindmap:mindmap) inside Better Sidebar, with a compact one-row toolbar (mindmap list, current mindmap, and export on the same line); the header 思维脑图 button opens or focuses that tab. When Better Sidebar is absent, the panel falls back to a standalone floating right panel toggled by the 思维脑图 button — drag-resizable (280px ~ 80% viewport), persisted, and pushing the chat left (layout-push) so the two never overlap. AI create/open/view intents always open or focus the panel/tab and switch to the target document, including when it is currently closed or the same document is opened again. The mode switch is fully reversible: if Better Sidebar is unloaded mid-session, the standalone panel and layout-push CSS are restored automatically.
• Directory tree tab — a persistent tree of the session working directory (served by plugin-owned read-only routes), lazy-loaded per directory; right-click to create a mindmap in the inbox or inside a directory; left-click a .md renders it through the read-only route first, then hands it to the AI for editing when the draft is empty. Labeled 目录 in standalone mode and 脑图列表 in sidebar mode.
• Single-mindmap mode — two tabs only: the tree/list tab and 脑图 (the current mindmap); opening another .md replaces the previous one.
• "What you see is what the AI edits" — when the visible mindmap differs from the AI's working document, the panel automatically asks the AI to open it, keeping the chat focus in sync.
• MarkGrove-style mapping — heading hierarchy, nested lists (empty items become placeholder nodes), code blocks as leaf nodes, paragraphs promoted to their own nodes (019 block concept), stable structural IDs, and orthogonal connector lines between nodes.
• Centered canvas with zoom and pan — the mindmap opens centered in the canvas; the top-right zoom bar (out / percent / in / 适配 / 全图) auto-fits on open without shrinking 13px node text below an effective 12px. Larger maps remain scrollable. 全图 instead tries to frame the entire tree, even below the readable floor, but stops at 25%: exceptionally large maps can still overflow. Manual zoom spans 25%–300%, and automatic re-fitting after AI edits pauses when you choose a zoom or 全图; 适配 restores it. Clicking a node focuses it at a readable scale near the left-quarter anchor; large subtrees may extend beyond the viewport. The ≤250 ms transition has no first-frame jump, normally changes zoom by at most ×2/÷2 per click (except when restoring readability after manual zoom below 12px), and writes frames directly to the canvas DOM without re-rendering the tree. New interactions interrupt it; when an AI edit pushes a selected node completely out of view, the canvas minimally scrolls it back without changing your zoom. Narrow panels fit by height instead of width and reserve a stable scrollbar gutter. Pan with the middle button anywhere, the left button on blank canvas (including Mac trackpad click-and-drag), or Space + left button on a card. Dragging follows the pointer even at scroll edges; a 4px threshold keeps clicks and the selection ring distinct from pans.

GitHub で README 全文を読む →