feat(mermaid): Markdown 预览安全渲染 Mermaid 图表(strict 清洗 + 懒加载 chunk + 点击放大) (#164)

* feat(mermaid): 新增 mermaid 懒加载 chunk(构建链四处同步 + mermaid 依赖)

tsdown CHUNKS / bundle-route CHUNK_NAMES / chunk-loader ChunkName /
package.json files 同步注册 mermaid chunk;mermaid ^11.16.1 进依赖与
lockfile;chunk 构建注入 uuid 浏览器入口别名,绕开 node:crypto 纯度门。

* feat(mermaid): Markdown 预览安全渲染 mermaid 图表 + 点击放大弹窗

- splitMermaidBlocks 纯函数按 fence 切块,无 mermaid 文件零行为变化;
- mermaid.initialize 走 securityLevel strict + htmlLabels:false,不应用
  bindFunctions(静态图);渲染 SVG 经 mermaid-sanitize 二次清洗:
  删 foreignObject/script/外来 HTML 元素、剥 @*/on* 属性、移除全部
  href/xlink:href,仅接受 SVG 根,解析失败返回空串;
- 点击图表弹出放大弹窗(滚轮缩放/拖拽平移/快捷键,借鉴 PR #75 交互,
  基于已清洗 SVG 克隆重写,不新增攻击面);
- 深浅色跟随重渲、失败回退错误条 + 原码、CodeBlock 同款 chrome。

* test(mermaid): splitter/sanitize 单测 + chunk 清单同步 + 挂载冒烟 diagram.md

- mermaid-blocks.spec 10 例 fence 切块语义;
- mermaid-sanitize.spec 8 例 XSS(foreignObject/on*/@*/href/xlink:href/
  script/外来元素/畸形 XML/非 SVG 根);
- bundle-route / chunk-artifact / manifest-consistency 清单加入 mermaid
  (chunk-artifact 补 addEventListener stub);
- 挂载冒烟新增 diagram.md:client-mermaid.js 往返、SVG 节点文字、
  点击弹窗 Esc 关闭、preview 模式编辑器隐藏。

* docs(mermaid): 安全渲染设计文档 + lazy-chunks 表 + README 双语更新

设计文档记录基线决策(PR #99 架构 + PR #75 点击放大)、安全模型
(strict/htmlLabels/清洗/href 剥离/无 bindFunctions)、验证与取舍。

* fix(mermaid): sanitizer 属性名大小写归一,堵住混合大小写属性绕过(CR P1)

XML 属性名保留大小写,而输出会经 HTML 解析归一化(dangerouslySetInnerHTML),
oNload/HREF/xlink:HREF 这类混合大小写属性此前会漏过清洗。比较前先
toLowerCase(),按原名删除;新增混合大小写事件/链接属性回归用例。

* fix(mermaid): sanitizer 元素黑名单大小写归一,堵住混合大小写元素绕过(CR P1)

XML 元素名保留大小写,querySelectorAll('script') 等选择器匹配不到
<sCrIpT>/<foreignobject>,而输出经 HTML 解析会被归一化为 SVG
script/foreignObject(可执行/可注入 HTML)。改为遍历全部元素按
localName.toLowerCase() 命中小写黑名单后整节点删除;新增混合大小写
script/foreignObject/外来 HTML 元素回归用例(jsdom 实证:旧选择器
命中 0,新逻辑命中 2)。

* fix(mermaid): 单次 Markdown 解析 + 占位交换(CR P1)+ CommonMark fence 规则(CR P2)

P1(跨 fence 语义):弃用逐块 MarkdownText 拆分,改为整篇文档单次
MarkdownText 渲染,布局后把 language-mermaid 代码块交换为 MermaidDiagram
(React 管理的 .md-code-block 宿主保留在树中、display:contents 抑制盒子,
协调安全;块不再 mermaid 时还原原 children)。新增
tests/mermaid-markdown.spec.tsx 与 e2e 跨 fence 引用链接断言。

P2(fence 匹配):支持 3+ 反引号与波浪号开 fence;闭 fence 须同字符且
长度不小于开 fence;反引号 info string 含反引号不视为开 fence;修复
带缩进开 fence 的 info 提取。blocks 单测 +3 例。
This commit is contained in:
Atlas Zeng
2026-08-18 01:22:28 +08:00
committed by GitHub
parent 9bea6c5539
commit 5e416eccef
23 changed files with 2131 additions and 25 deletions
+81 -7
View File
@@ -15,11 +15,12 @@
* fail() strips, no `pageerror`, no plugin-prefixed console errors);
* 4. expands the collapsed panel (openByDefault defaults off), sweeps every
* built-in tab (Files / Source Control / Tasks / Terminal / Browser) —
* including the lazily-fetched terminal chunk — and then opens a seeded
* file through the Files window's tree (in-place mode: the seeded home
* tab itself switches to the file, no new tab), while a response wait
* armed before goto proves the lazily-fetched editor chunk
* (client-editor.js) loaded.
* including the lazily-fetched terminal chunk — and then opens seeded
* files through the Files window's tree (in-place mode: the seeded home
* tab itself switches to the file, no new tab), while response waits
* armed before goto prove the lazily-fetched editor chunk
* (client-editor.js) and the mermaid chunk (client-mermaid.js, rendered
* SVG diagram + zoom modal) loaded.
*
* Deterministic by construction: every wait is on a DOM/network marker, the
* suite is serial (one server instance), and any crash trips the very next
@@ -42,6 +43,11 @@ const WORKSPACE_PATH = process.env.DSH_E2E_WORKSPACE ?? join(tmpdir(), 'dsh-e2e-
* exercise the file-open path (editor chunk = client-editor.js). */
const SEEDED_FILE = 'hello.txt'
/** A markdown file with a mermaid fence, opened through the Files window's
* tree to force the lazily-packed mermaid chunk (client-mermaid.js) to load
* and render a sanitized SVG diagram. */
const SEEDED_MD_FILE = 'diagram.md'
/**
* The plugin's crash markers. The client mounts inside an error boundary that
* renders a strip whose text starts with these prefixes instead of crashing
@@ -54,11 +60,31 @@ const BUILTIN_TABS = ['Files', 'Source Control', 'Tasks', 'Terminal', 'Browser']
let api: APIRequestContext
/** Seed one workspace + one session (plus a file for the editor-chunk probe)
* through the host's unary RPC surface. */
/** Seed one workspace + one session (plus files for the editor/mermaid-chunk
* probes) through the host's unary RPC surface. */
async function seedSession(): Promise<void> {
mkdirSync(WORKSPACE_PATH, { recursive: true })
writeFileSync(join(WORKSPACE_PATH, SEEDED_FILE), 'hello from the mount lane\n')
// The mermaid-chunk probe file: a markdown doc whose preview must fetch
// client-mermaid.js and render the fence into an SVG diagram. The
// reference-style link's definition sits AFTER the fence: it only
// resolves when the preview is one single markdown parse (the mermaid
// path must not split the document into independent MarkdownText blocks).
writeFileSync(join(WORKSPACE_PATH, SEEDED_MD_FILE), [
'# Diagram',
'',
'[before][shared]',
'',
'```mermaid',
'graph TD',
' A[Hello] --> B[World]',
'```',
'',
'[shared]: https://example.com',
'',
'tail text',
'',
].join('\n'))
const workspace = await api.post(`${BASE_URL}/api/workspace.create`, {
data: { type: 'client-request', rpcId: 'e2e-workspace', method: 'workspace.create', payload: { path: WORKSPACE_PATH } },
})
@@ -230,6 +256,54 @@ test('plugin mounts into the DSH shell and survives a built-in tab sweep', async
await page.waitForTimeout(1_500)
await assertNoCrash()
// The mermaid chunk (client-mermaid.js) only loads when a previewed
// markdown file contains a mermaid fence. Open the seeded diagram file
// from the files window's tree (the embedded tree stays pinned while
// hello.txt is open) and require the full round-trip: chunk fetch +
// sanitized SVG diagram in the preview, so a missing/corrupt mermaid
// chunk or a broken render fails the lane.
const mermaidChunk = page.waitForResponse(
(response) => response.url().includes('/sidebar/bundle/mermaid.js'),
{ timeout: 30_000 },
)
const mdRow = sidebar.locator(`[role="button"][title$="${SEEDED_MD_FILE}"]:visible`)
await expect(mdRow, `the seeded "${SEEDED_MD_FILE}" file must appear in the files window's tree`).toHaveCount(1, { timeout: 30_000 })
await mdRow.click({ position: { x: 8, y: 8 } })
await mermaidChunk
await expect(
sidebar.locator('[data-mermaid-diagram] svg'),
'the mermaid fence must render into an SVG diagram in the markdown preview',
).toHaveCount(1, { timeout: 30_000 })
// Labels must survive as real SVG <text> (htmlLabels stays off so the
// sanitizer's foreignObject strip cannot eat the node text).
await expect(
sidebar.locator('[data-mermaid-diagram]').first(),
'the diagram node labels must render inside the SVG',
).toContainText('Hello', { timeout: 30_000 })
// Cross-fence semantics: the reference-style link [before][shared] must
// resolve to the definition that sits AFTER the fence — proof that the
// preview is a single markdown parse and not per-fence fragments.
await expect(
sidebar.locator('a[href="https://example.com"]').first(),
'reference-style links with definitions across a mermaid fence must resolve',
).toContainText('before', { timeout: 30_000 })
// Click-to-enlarge: clicking the diagram opens the zoom modal (portalled
// to document.body), Esc closes it again.
const modal = page.locator('[data-mermaid-modal]')
await sidebar.locator('[data-mermaid-diagram] svg').first().click()
await expect(modal, 'clicking the diagram must open the zoom modal').toHaveCount(1, { timeout: 10_000 })
await page.keyboard.press('Escape')
await expect(modal, 'Esc must close the zoom modal').toHaveCount(0, { timeout: 10_000 })
// The preview/edit toggle is mutually exclusive: in preview mode the
// CodeMirror surface must be hidden (regression guard — a stale css copy
// in the page made the editor stay visible under the preview, breaking
// the toggle semantics).
await expect(
sidebar.locator('.cm-editor').first(),
'preview mode must hide the CodeMirror editor (mutually exclusive toggle)',
).toBeHidden()
await assertNoCrash()
// The plugin's own console prefix must never appear in errors, and no
// unhandled rejection may escape the sweep.
const pluginErrors = consoleErrors.filter((text) => /dsh-better-sidebar|Unhandled/.test(text))