@mehmoodqureshi/chrome-mcp 0.9.2 → 0.9.3

package/dist/src/mcp/locate.js

@@ -10,10 +10,13 @@
  * redeploy.
  *
  * A locator closes that: `{ role: 'button', name: 'Sign in' }` resolves through
- * one snapshot, server-side, and the caller never sees the tree. Matching runs
- * strongest-first (exact, then case-insensitive, then contains) so an
- * unambiguous name wins outright, and an ambiguous one fails loudly with the
- * candidates rather than silently clicking the first row.
+ * one snapshot and the caller never sees the tree. The page does the matching
+ * itself (`collectSnapshot` with a locator returns only the strongest-tier
+ * hits, and stamps refs on those alone), so what crosses the bridge is a
+ * handful of nodes rather than 400; this module re-scores them — same tiers:
+ * exact, then case-insensitive, then prefix, then contains — so an unambiguous
+ * name wins outright and an ambiguous one fails loudly with the candidates
+ * rather than silently clicking the first row.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.hasLocator = hasLocator;
@@ -60,6 +63,7 @@ async function resolveLocator(ex, loc, opts = {}) {
         tabId: opts.tabId,
         interactiveOnly: false,
         max: 400,
+        locator: want,
         frameId: opts.frameId,
         allFrames: opts.allFrames,
     });
@@ -79,10 +83,13 @@ async function resolveLocator(ex, loc, opts = {}) {
     }
     const describe = (n) => `${n.role} "${n.name}"`;
     if (winners.length === 0) {
-        const sample = snap.nodes
-            .filter((n) => !want.role || norm(n.role) === norm(want.role))
-            .slice(0, 8)
-            .map(describe);
+        // A page that scored in place reports what it had of that role as `nearby`;
+        // a backend that returned the full tree leaves it to us.
+        const sample = snap.nearby ??
+            snap.nodes
+                .filter((n) => !want.role || norm(n.role) === norm(want.role))
+                .slice(0, 8)
+                .map(describe);
         throw new validators_1.McpToolError(`no element matches ${JSON.stringify(want)}. ` +
             (sample.length
                 ? `Closest by role: ${sample.join(', ')}. `

package/dist/src/mcp/tools.js

@@ -91,7 +91,7 @@ exports.TOOL_DEFINITIONS = [
     { name: 'press', description: 'Press a key (with optional modifiers).', inputSchema: { key: zod_1.z.string(), modifiers: zod_1.z.array(zod_1.z.string()).optional(), tabId: tabIdField, failOnAuthWall: authWallField } },
     { name: 'hover', description: 'Hover over an element.', inputSchema: { ...TARGET_PROPS, ...LOCATOR_PROPS, ...FRAME_PROPS, tabId: tabIdField, snapshotAfter: snapshotAfterField } },
     { name: 'scroll', description: 'Scroll the page or to an element.', inputSchema: { ...TARGET_PROPS, ...FRAME_PROPS, x: zod_1.z.number().optional(), y: zod_1.z.number().optional(), deltaX: zod_1.z.number().optional(), deltaY: zod_1.z.number().optional(), tabId: tabIdField } },
-    { name: 'screenshot', description: 'Capture a PNG screenshot (page or element).', inputSchema: { ...TARGET_PROPS, ...FRAME_PROPS, fullPage: zod_1.z.boolean().optional(), tabId: tabIdField } },
+    { name: 'screenshot', description: 'Capture a screenshot (page or element). Default is JPEG (quality 70) at CSS-pixel size, which is several times smaller than PNG and reads fine. Pass format:"png" for lossless, quality 1-100 for JPEG, scale 2 for device pixels on a Retina display or 0.5 to shrink.', inputSchema: { ...TARGET_PROPS, ...FRAME_PROPS, fullPage: zod_1.z.boolean().optional(), format: zod_1.z.enum(['jpeg', 'png']).optional(), quality: zod_1.z.number().optional(), scale: zod_1.z.number().optional(), tabId: tabIdField } },
     { name: 'get_text', description: 'Get visible text of the page or an element.', inputSchema: { ...TARGET_PROPS, ...FRAME_PROPS, tabId: tabIdField, maxBytes: maxBytesField } },
     { name: 'get_html', description: 'Get HTML of the page or an element. Output is capped (see maxBytes) and cut at a tag boundary; narrow it with `selector` rather than raising the cap when you can. Password field values are always blanked.', inputSchema: { ...TARGET_PROPS, ...FRAME_PROPS, outer: zod_1.z.boolean().optional(), tabId: tabIdField, maxBytes: maxBytesField } },
     { name: 'snapshot', description: 'Accessibility snapshot: interactive elements with refs to target by `ref` (more reliable than guessing CSS selectors). Pass diff:true to get only what changed since the last snapshot of this tab - far cheaper in a click/read loop. Password fields appear as secret:true with no value.', inputSchema: { interactiveOnly: zod_1.z.boolean().optional(), max: zod_1.z.number().optional(), diff: zod_1.z.boolean().describe('Return added/removed/changed elements since the previous snapshot of this tab instead of the whole tree').optional(), failOnAuthWall: authWallField, ...FRAME_PROPS, tabId: tabIdField } },
@@ -201,15 +201,15 @@ const GATE_CONTEXT = 'cannot resolve the target tab URL for the policy gate';
  *
  * Prefers a URL the backend already reported over asking again: the extension
  * rides the tab's landing URL home on every result frame, which is what keeps a
- * gated call to ONE round-trip instead of two. That cache only ever describes
- * the active tab, so it is bypassed whenever an explicit `tabId` is in play.
+ * gated call to ONE round-trip instead of two. The active-tab cache serves calls
+ * without a `tabId`; the per-tab cache (fed by results for that tab and by any
+ * `tabs_list`) serves explicitly-targeted ones, so a parallel batch over N tabs
+ * gates on the one listing that opened it rather than N more.
  */
 async function gatedUrl(ex, tabId) {
-    if (!tabId) {
-        const known = ex.cachedActiveUrl?.();
-        if (known)
-            return known;
-    }
+    const known = tabId ? ex.cachedTabUrl?.(tabId) : ex.cachedActiveUrl?.();
+    if (known)
+        return known;
     let tabs;
     try {
         tabs = await ex.tabsList();
@@ -547,13 +547,20 @@ exports.TOOL_HANDLERS = {
     },
     screenshot: async (a, ctx) => {
         await gate(ctx, 'screenshot', { tabId: tabId(a) });
+        const format = (0, validators_1.optionalString)(a, 'format');
+        if (format !== undefined && format !== 'jpeg' && format !== 'png') {
+            throw new validators_1.McpToolError('"format" must be "jpeg" or "png"');
+        }
         const shot = await ctx.ex.screenshot({
             tabId: tabId(a),
             ...frameOpts(a),
             fullPage: (0, validators_1.optionalBoolean)(a, 'fullPage'),
             target: (0, validators_1.optionalTarget)(a),
+            format,
+            quality: (0, validators_1.optionalNumber)(a, 'quality', { min: 1, max: 100 }),
+            scale: (0, validators_1.optionalNumber)(a, 'scale', { min: 0.1, max: 4 }),
         });
-        (0, workspace_1.saveScreenshot)(shot.dataBase64);
+        (0, workspace_1.saveScreenshot)(shot.dataBase64, shot.mimeType === 'image/jpeg' ? 'jpg' : 'png');
         const caption = shot.truncated ? `(truncated; full height ${shot.fullHeight}px)` : undefined;
         return (0, envelopes_1.imageResult)(shot.dataBase64, shot.mimeType, caption);
     },

package/docs/BLUEPRINT.md

@@ -173,7 +173,7 @@ export interface EvalResult { ok: boolean; value?: unknown; type?: string; error
 export interface WaitResult { matched: boolean; ref?: string; waitedMs: number; }
 export interface ActionOk   { ok: true; }
 export interface ScreenshotResult {
-  dataBase64: string; mimeType: 'image/png';
+  dataBase64: string; mimeType: 'image/png' | 'image/jpeg';  // jpeg q70 @ CSS px by default; format/quality/scale opt-in
   width: number; height: number; truncated: boolean; fullHeight?: number;  // fullPage cap metadata
 }
 export interface DownloadResult { path: string; backend: BackendKind; bytes: number; mimeType?: string; suggestedName?: string; }