dsh-builtin-browser 0.2.0 → 0.3.0

package/src/browser-electron/virtual-cursor.ts

@@ -22,15 +22,46 @@ export type CursorAction = 'move' | 'click'
 /** Element id of the overlay root, stable so repeat calls update one node. */
 const CURSOR_ID = '__dsh_agent_cursor__'
 
+/**
+ * Where the overlay was last placed, so an unchanged position is not repainted.
+ *
+ * A single tool call issues several CDP commands (locate, click, cursor), and the
+ * cursor is repainted for each one that moves it. Repainting to the same pixel is a
+ * wasted round-trip *and* a visible flicker: the transition restarts from the same
+ * place, which reads as a stutter rather than as movement.
+ */
+const lastPlaced = new WeakMap<object, { x: number; y: number; action: CursorAction }>()
+
+/**
+ * Whether a paint at this position would change anything on screen.
+ * @param handle - the view the overlay belongs to.
+ * @param x - viewport x in CSS pixels.
+ * @param y - viewport y in CSS pixels.
+ * @param action - the action being reported.
+ * @param force - paint even when unchanged (a click ripple must always fire).
+ * @returns true when the overlay should be painted.
+ */
+export function cursorNeedsPaint<TView>(handle: TView, x: number, y: number, action: CursorAction, force = false): boolean {
+  if (force) return true
+  const key = handle as unknown as object
+  if (key === null || typeof key !== 'object') return true
+  const previous = lastPlaced.get(key)
+  return previous === undefined || previous.x !== x || previous.y !== y || previous.action !== action
+}
+
 /**
  * The in-page expression that materializes or moves the overlay. Idempotent: the
  * first call creates the nodes, later calls only move them (and pulse).
  * @param x - viewport x in CSS pixels.
  * @param y - viewport y in CSS pixels.
  * @param action - whether to pulse a click ripple at the point.
+ * @param label - optional short description of the operation, shown beside the
+ *   pointer. The overlay is the only "the agent has taken over this tab" signal, so
+ *   saying *what* it is doing is what turns it from a decoration into an account of
+ *   the agent's actions that a human can follow without reading the tool log.
  * @returns a self-contained expression that evaluates to `true`.
  */
-export function cursorScript(x: number, y: number, action: CursorAction): string {
+export function cursorScript(x: number, y: number, action: CursorAction, label?: string): string {
   return `(() => {
   const id = ${JSON.stringify(CURSOR_ID)}
   let root = document.getElementById(id)
@@ -38,25 +69,56 @@ export function cursorScript(x: number, y: number, action: CursorAction): string
     root = document.createElement('div')
     root.id = id
     root.setAttribute('aria-hidden', 'true')
-    root.style.cssText = 'position:fixed;left:0;top:0;width:0;height:0;z-index:2147483647;pointer-events:none;transition:transform 120ms ease-out;will-change:transform'
+    root.style.cssText = 'position:fixed;left:0;top:0;width:0;height:0;z-index:2147483647;pointer-events:none;transition:transform 190ms cubic-bezier(.22,.85,.24,1);will-change:transform'
     const svg = '<svg width="22" height="22" viewBox="0 0 22 22" style="position:absolute;left:-2px;top:-2px;filter:drop-shadow(0 1px 2px rgba(0,0,0,.45))">'
       + '<path d="M4 2 L4 18 L8.4 13.8 L11.2 19.6 L13.8 18.4 L11 12.8 L17 12.6 Z" fill="#ffffff" stroke="#111111" stroke-width="1.4" stroke-linejoin="round"/></svg>'
-    const ripple = '<span style="position:absolute;left:0;top:0;width:34px;height:34px;margin:-17px 0 0 -17px;border-radius:50%;border:2px solid rgba(64,160,255,.9);opacity:0"></span>'
-    root.innerHTML = svg + ripple
+    const ripple = '<span data-dsh-role="ripple" style="position:absolute;left:0;top:0;width:34px;height:34px;margin:-17px 0 0 -17px;border-radius:50%;border:2px solid rgba(64,160,255,.9);opacity:0"></span>'
+    // The bubble sits below-right of the pointer tip and is capped in width so a long
+    // label cannot cover the page it is describing.
+    const bubble = '<span data-dsh-role="bubble" style="position:absolute;left:18px;top:16px;max-width:280px;padding:3px 8px;border-radius:5px;background:rgba(17,17,17,.92);color:#fff;font:12px/1.45 -apple-system,Segoe UI,Roboto,sans-serif;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;opacity:0;box-shadow:0 1px 3px rgba(0,0,0,.4)"></span>'
+    root.innerHTML = svg + ripple + bubble
     ;(document.body || document.documentElement).appendChild(root)
   }
   root.style.transform = 'translate(${x}px, ${y}px)'
-  ${action === 'click' ? `const ripple = root.lastElementChild
+  ${action === 'click' ? `const ripple = root.querySelector('[data-dsh-role="ripple"]')
   if (ripple && typeof ripple.animate === 'function') {
     ripple.animate(
       [{ transform: 'scale(.25)', opacity: 0.95 }, { transform: 'scale(1.6)', opacity: 0 }],
       { duration: 550, easing: 'ease-out' },
     )
   }` : ''}
+  ${label !== undefined && label !== '' ? `const bubble = root.querySelector('[data-dsh-role="bubble"]')
+  if (bubble) {
+    bubble.textContent = ${JSON.stringify(label)}
+    if (typeof bubble.animate === 'function') {
+      // Fade in, hold, fade out: the label is a narration of the action, not a
+      // permanent caption, so a stale one never lingers over later work.
+      bubble.animate(
+        [{ opacity: 0, transform: 'translateY(-3px)' }, { opacity: 1, transform: 'translateY(0)' }, { opacity: 1, offset: 0.6 }, { opacity: 0 }],
+        { duration: 2200, easing: 'ease-out', fill: 'forwards' },
+      )
+    } else {
+      bubble.style.opacity = '1'
+    }
+  }` : ''}
   return true
 })()`
 }
 
+/**
+ * Forget the remembered position for a view.
+ *
+ * Called when the document is replaced: the overlay lived in the old document and
+ * went with it, so the cache would otherwise claim the pointer is already where it
+ * needs to be and skip every later paint — the pointer would simply never return
+ * after a navigation.
+ * @param handle - the view whose cache should be dropped.
+ */
+export function forgetCursor<TView>(handle: TView): void {
+  const key = handle as unknown as object
+  if (key !== null && typeof key === 'object') lastPlaced.delete(key)
+}
+
 /**
  * Draw (or move) the pointer overlay on one view's page.
  *
@@ -69,6 +131,7 @@ export function cursorScript(x: number, y: number, action: CursorAction): string
  * @param y - viewport y in CSS pixels.
  * @param action - whether to pulse a click ripple.
  * @param evaluate - the provider's evaluate bridge (keeps the CDP coupling in one place).
+ * @param label - optional short description of the operation, shown beside the pointer.
  */
 export function paintCursor<TView>(
   handle: TView,
@@ -76,7 +139,14 @@ export function paintCursor<TView>(
   y: number,
   action: CursorAction,
   evaluate: (handle: TView, expression: string) => Promise<unknown>,
+  label?: string,
+  force = false,
 ): void {
   if (!Number.isFinite(x) || !Number.isFinite(y)) return
-  void evaluate(handle, cursorScript(Math.round(x), Math.round(y), action)).catch(() => {})
+  const roundedX = Math.round(x)
+  const roundedY = Math.round(y)
+  if (!cursorNeedsPaint(handle, roundedX, roundedY, action, force)) return
+  const key = handle as unknown as object
+  if (key !== null && typeof key === 'object') lastPlaced.set(key, { x: roundedX, y: roundedY, action })
+  void evaluate(handle, cursorScript(roundedX, roundedY, action, label)).catch(() => {})
 }

package/src/tool-browser/index.ts

@@ -161,15 +161,26 @@ function parseFillValue(v: string | undefined): string | number | boolean {
   return v ?? ''
 }
 
-/** Format a snapshot element list for the model. */
+/**
+ * Format a snapshot element list for the model.
+ *
+ * Coordinates are opt-in. They are only meaningful to a caller that intends to
+ * click a pixel position, which means a vision pass; for the far more common case
+ * of locating an element by role/label and passing a semantic target, they are the
+ * bulkiest and least useful part of every line. Omitting them keeps the listing
+ * readable — which is exactly what a model without image input has to work from.
+ * @param snapshot - the snapshot payload.
+ * @param options - `coords: true` to include each element's viewport position.
+ */
 function formatSnapshot(snapshot: {
   url: string
   title?: string
   elements: readonly { ref: number; kind: string; label: string; x: number; y: number; frame?: boolean }[]
   truncated?: boolean
   challenge?: { blocked: boolean; kind?: string; reason?: string }
-}): string {
-  const lines = snapshot.elements.map(el => `[${el.ref}] ${el.kind}: ${el.label}${el.frame === true ? ' (iframe)' : ''} (${el.x},${el.y})`)
+}, options: { coords?: boolean } = {}): string {
+  const showCoords = options.coords === true
+  const lines = snapshot.elements.map(el => `[${el.ref}] ${el.kind}: ${el.label}${el.frame === true ? ' (iframe)' : ''}${showCoords ? ` (${el.x},${el.y})` : ''}`)
   const header = `URL: ${snapshot.url}${snapshot.title !== undefined ? `\nTitle: ${snapshot.title}` : ''}`
   const body = lines.length > 0 ? lines.join('\n') : '(no interactive elements found)'
   const tail = snapshot.truncated === true ? '\n(snapshot truncated)' : ''
@@ -241,7 +252,10 @@ export function apply(ctx: Context, config: Config = {}): void {
           },
         },
       },
-      render: (_args, value) => [{ type: 'text', text: formatSnapshot(value) }],
+      // browser_open is an overview of what just loaded, not a locator: it has no
+      // `coords` parameter, so coordinates stay off here (browser_snapshot owns
+      // that switch for callers that actually target pixels).
+      render: (args, value) => [{ type: 'text', text: formatSnapshot(value, { coords: (args as { coords?: boolean }).coords === true }) }],
     },
     timeoutMs,
     isConcurrencySafe: () => false, // opens tabs / navigates; exclusive within a task
@@ -267,8 +281,10 @@ export function apply(ctx: Context, config: Config = {}): void {
 
   ctx.tools.register(defineTool({
     name: 'browser_snapshot',
-    description: 'Return an AI-friendly snapshot of the current shared-browser page: numbered interactive elements (inputs, buttons, links) the model can cite. Use this to understand an interactive page before driving it.',
-    parameters: {},
+    description: 'Return an AI-friendly snapshot of the current shared-browser page: numbered interactive elements (inputs, buttons, links) the model can cite. This is the primary way to understand a page, and it needs no image input — every listed element can afterwards be addressed by a semantic target (browser_click, browser_type, browser_check, …) rather than by coordinates.',
+    parameters: {
+      coords: { type: 'boolean', description: 'Also print each element\'s viewport coordinates. Off by default; only useful when a vision pass is going to click a pixel position.' },
+    },
     output: {
       schema: {
         type: 'object',
@@ -304,7 +320,7 @@ export function apply(ctx: Context, config: Config = {}): void {
           },
         },
       },
-      render: (_args, value) => [{ type: 'text', text: formatSnapshot(value) }],
+      render: (args, value) => [{ type: 'text', text: formatSnapshot(value, { coords: args.coords === true }) }],
     },
     timeoutMs,
     isConcurrencySafe: () => true,
@@ -329,6 +345,7 @@ export function apply(ctx: Context, config: Config = {}): void {
     parameters: {
       includeHidden: { type: 'boolean', description: 'Include hidden elements (default false).' },
       maxNodes: { type: 'number', description: 'Maximum nodes (default 500, range 10-5000).' },
+      coords: { type: 'boolean', description: 'Include each element\'s viewport coordinates. Off by default: a semantic target needs no coordinates, and they are pure noise unless a vision pass is going to click a pixel position.' },
     },
     output: {
       schema: {
@@ -361,14 +378,24 @@ export function apply(ctx: Context, config: Config = {}): void {
           },
         },
       },
-      render: (_args, value) => {
+      render: (args, value) => {
         const nodes = value.nodes as Array<{ ref: number; role: string; name: string; value?: string | null; states: string[]; depth: number; tag: string; x: number; y: number; frame?: boolean }>
+        // Built for reading, not for imaging: containment is shown as indentation
+        // (the depth was already collected and simply thrown away), an empty state
+        // list adds nothing, and coordinates appear only when asked for — a model
+        // working from the DOM never needs them, and they are the bulkiest part of
+        // every line.
+        const showCoords = args.coords === true
         const lines = nodes.map(n => {
           const valuePart = n.value !== undefined && n.value !== null ? ` value="${String(n.value).slice(0, 60)}"` : ''
-          return `[${n.ref}] ${n.role} "${n.name}"${valuePart} (${n.x},${n.y}) states=[${n.states.join(',')}]${n.frame === true ? ' (iframe)' : ''}`
+          const indent = '  '.repeat(Math.max(0, Math.min(8, Number(n.depth) || 0)))
+          const coordsPart = showCoords ? ` (${n.x},${n.y})` : ''
+          const statePart = Array.isArray(n.states) && n.states.length > 0 ? ` states=[${n.states.join(',')}]` : ''
+          return `${indent}[${n.ref}] ${n.role} "${n.name}"${valuePart}${coordsPart}${statePart}${n.frame === true ? ' (iframe)' : ''}`
         })
         const header = `URL: ${value.url}${value.title !== undefined ? `\nTitle: ${value.title}` : ''}`
-        return [{ type: 'text', text: `${header}\n\n${lines.length > 0 ? lines.join('\n') : '(no accessible interactive nodes)'}${value.truncated === true ? '\n(truncated)' : ''}` }]
+        const hint = showCoords ? '' : '\n(coordinates omitted — pass coords: true to include them, or click by semantic target)'
+        return [{ type: 'text', text: `${header}\n\n${lines.length > 0 ? lines.join('\n') : '(no accessible interactive nodes)'}${value.truncated === true ? '\n(truncated)' : ''}${hint}` }]
       },
     },
     timeoutMs,
@@ -601,7 +628,7 @@ export function apply(ctx: Context, config: Config = {}): void {
 
   ctx.tools.register(defineTool({
     name: 'browser_click',
-    description: 'Click in the shared browser. Two ways: (1) semantic target — pass target {by: css|text|xpath, value, index?} and the element is located, scrolled into view and clicked at its center; (2) viewport coordinates (x/y) — use with browser_screenshot when a vision model locates an element on the screenshot (covers icons, image buttons, and canvas that DOM locators cannot target). Provide exactly one of target or x/y.',
+    description: 'Click in the shared browser. Prefer the semantic target: pass target {by: css|text|xpath, value, index?} and the element is located from the DOM, scrolled into view and clicked at its centre — this needs no image input and is the only form that works without vision. The alternative is viewport coordinates (x/y), which are meaningful only after a vision model has located the element on a browser_screenshot; coordinate clicks are refused outright under the non-visual strategy. Provide exactly one of target or x/y.',
     parameters: {
       target: {
         type: 'object',
@@ -613,8 +640,8 @@ export function apply(ctx: Context, config: Config = {}): void {
         },
         description: 'Locate the element semantically and click it (css/text/xpath).',
       },
-      x: { type: 'number', description: 'Viewport x coordinate (CSS px), e.g. from a vision model reading the screenshot.' },
-      y: { type: 'number', description: 'Viewport y coordinate (CSS px).' },
+      x: { type: 'number', description: 'Viewport x coordinate (CSS px), taken from a browser_screenshot read by a vision model. Prefer a semantic target — it works without image input.' },
+      y: { type: 'number', description: 'Viewport y coordinate (CSS px), taken from the same screenshot as x.' },
     },
     output: {
       schema: { type: 'object', additionalProperties: false, properties: { clicked: { type: 'boolean', required: true } } },
@@ -1089,7 +1116,7 @@ export function apply(ctx: Context, config: Config = {}): void {
 
   ctx.tools.register(defineTool({
     name: 'browser_screenshot',
-    description: 'Capture the current shared-browser page as a screenshot (PNG default, JPEG optional). Use for visual confirmation of layout, charts, designs, or CAPTCHAs, or to feed a vision tool (read_image) that locates elements visually. Supports full-page capture, save-to-file, JPEG encoding, and downscaling (maxWidth/maxHeight) to cut vision-tool token cost. JPEG is only available on the self-hosted native path; the desktop-shell path returns PNG.',
+    description: 'Capture the current shared-browser page as a screenshot (PNG default, JPEG optional). This is for models that can read images: layout checks, charts, designs, CAPTCHAs, or locating an element by eye before clicking its coordinates. A model without image input gains nothing from it — browser_snapshot, browser_a11y and browser_content carry the same page as text, and browser_scrape extracts structured data. Supports full-page capture, save-to-file, JPEG encoding, and downscaling (maxWidth/maxHeight) to cut vision-tool token cost. JPEG is only available on the self-hosted native path; the desktop-shell path returns PNG.',
     parameters: {
       fullPage: { type: 'boolean', description: 'Capture the full scrollable page instead of the viewport (default false).' },
       savePath: { type: 'string', description: 'Absolute file path to also save the image to (e.g. for read_image vision location). Must resolve inside the configured downloadDir (default: the system Downloads folder, localized names such as ~/下载 included); an existing file is never overwritten.' },
@@ -1338,6 +1365,8 @@ export function apply(ctx: Context, config: Config = {}): void {
     parameters: {
       limit: { type: 'number', description: 'Maximum entries to return (default 30, capped at 200).' },
       domain: { type: 'string', description: 'Only visits whose hostname contains this text (case-insensitive).' },
+      query: { type: 'string', description: 'Only visits whose URL or title contains this text (case-insensitive).' },
+      session: { type: 'string', description: 'Only visits recorded by this task/session label — use it to separate your own pages from another session\'s or the human\'s.' },
     },
     output: {
       schema: {
@@ -1384,7 +1413,14 @@ export function apply(ctx: Context, config: Config = {}): void {
       const rawLimit = typeof args.limit === 'number' && Number.isFinite(args.limit) ? args.limit : 30
       const limit = Math.max(1, Math.min(200, Math.floor(rawLimit)))
       const domain = typeof args.domain === 'string' && args.domain.trim() !== '' ? args.domain.trim() : undefined
-      const entries = browser.visited({ limit, ...domain !== undefined ? { domain } : {} })
+      const query = typeof args.query === 'string' && args.query.trim() !== '' ? args.query.trim() : undefined
+      const session = typeof args.session === 'string' && args.session.trim() !== '' ? args.session.trim() : undefined
+      const entries = browser.visited({
+        limit,
+        ...domain !== undefined ? { domain } : {},
+        ...query !== undefined ? { query } : {},
+        ...session !== undefined ? { session } : {},
+      })
       return {
         count: entries.length,
         entries: entries.map(entry => ({