safari-mcp 2.21.2 → 2.21.3

package/extension/background.js

@@ -2761,7 +2761,16 @@ async function handleCommand(type, payload) {
         if (mainResult && JSON.parse(mainResult).length > 0) return mainResult;
       } catch { return mainResult; }
       const frameResult = await execAcrossFrames(queryFn, [payload.selector, payload.limit || 20], tabId);
-      return frameResult || mainResult;
+      if (!frameResult) return mainResult;
+      // Reaching here means the main frame matched nothing, so every element below came
+      // from a child frame and its x/y are that frame's viewport coordinates. Say so in
+      // the payload: the caller is an agent reading this JSON, and the click tools take
+      // page coordinates — a comment in this file never reaches it.
+      try {
+        return JSON.stringify(JSON.parse(frameResult).map((el) => ({ ...el, frameRelative: true })));
+      } catch {
+        return frameResult;
+      }
     }
 
     default:

package/extension/manifest.json

@@ -2,7 +2,7 @@
   "manifest_version": 3,
   "name": "Safari MCP Bridge",
   "description": "Connects AI agents to Safari — fast JS execution, screenshots, clicks with your real cookies/logins",
-  "version": "2.10.11",
+  "version": "2.10.12",
   "icons": {
     "48": "images/icon-48.png",
     "96": "images/icon-96.png",

package/index.js

@@ -3014,7 +3014,7 @@ server.tool(
 
 server.tool(
   "safari_query_all",
-  "Find all elements matching a CSS selector (returns tag, text, href, value)",
+  "Find all elements matching a CSS selector (returns tag, text, href, value). x/y are relative to each element's own frame viewport — for matches inside an iframe, click by selector or snapshot ref, not by these coordinates",
   {
     selector: z.string().describe("CSS selector"),
     limit: z.coerce.number().optional().describe("Max results (default: 20)"),
@@ -3025,6 +3025,15 @@ server.tool(
       "query_all", { selector: args.selector, limit: args.limit, ..._explicitReceipt(args) },
       () => safari.querySelectorAll(args)
     );
+    // The matches live inside a child frame, so their x/y are that frame's viewport
+    // coordinates while safari_click and friends take page coordinates. Warn in the
+    // result text, which is the only part of this an agent actually reads.
+    if (typeof result === "string" && result.includes('"frameRelative":true')) {
+      return textResult(
+        result +
+          "\n\nNote: these matches are inside a child frame. Their x/y are relative to that frame's viewport (see \"frame\"), not to the page — click them by selector or by a safari_snapshot ref; passing these coordinates to safari_click lands somewhere else."
+      );
+    }
     return textResult(result);
   }
 );

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "safari-mcp",
-  "version": "2.21.2",
+  "version": "2.21.3",
   "mcpName": "io.github.achiya-automation/safari-mcp",
   "description": "Safari browser automation for AI agents — native macOS, zero Chrome overhead. 97 tools via AppleScript + JavaScript.",
   "type": "module",