fix(agent): run coord-click loop detection on converted CSS pixels

src/chrome/src/agent/agent.js

@@ -6627,7 +6627,8 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
       }
       let coordCheck = { kind: 'none' };
       if (fnName === 'click' && fnArgs?.x != null && fnArgs?.y != null) {
-        coordCheck = this._checkCoordClickLoop(tabId, fnArgs.x, fnArgs.y);
+        const point = this._resolveClickCoordsForLoop(tabId, fnArgs);
+        coordCheck = this._checkCoordClickLoop(tabId, point.x, point.y);
       }
       const axReadCheck = this._checkAccessibilityReadLoop(tabId, fnName, fnArgs, toolResult);
       const scrollCheck = this._checkNoProgressScroll(tabId, fnName, fnArgs, toolResult);
@@ -8742,6 +8743,21 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
     };
   }
 
+  /**
+   * Resolve the point a click call will actually dispatch at, for loop
+   * detection. Loop detection must measure the same CSS-pixel point the
+   * click dispatches: under downscaled screenshots the model's image-pixel
+   * nudges collapse to the same CSS pixel and the detector must see that,
+   * otherwise "slightly different spot" clicks never register as the loop
+   * they are. Non-numeric args fall back to Number() so the batch guard
+   * (x != null) and the detector keep their existing shape.
+   */
+  _resolveClickCoordsForLoop(tabId, args) {
+    const mapped = this._screenshotClickCoords(tabId, args);
+    if (!mapped) return { x: Number(args.x), y: Number(args.y) };
+    return { x: mapped.x, y: mapped.y };
+  }
+
   _coordinateReconciliationDiagnostic(point, resolution, clickPath, fallbackReason) {
     const target = resolution?.success === true ? resolution.semanticTarget : null;
     const resolved = clickPath === 'semantic';

src/firefox/src/agent/agent.js

@@ -5568,7 +5568,8 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
       }
       let coordCheck = { kind: 'none' };
       if (fnName === 'click' && fnArgs?.x != null && fnArgs?.y != null) {
-        coordCheck = this._checkCoordClickLoop(tabId, fnArgs.x, fnArgs.y);
+        const point = this._resolveClickCoordsForLoop(tabId, fnArgs);
+        coordCheck = this._checkCoordClickLoop(tabId, point.x, point.y);
       }
       const axReadCheck = this._checkAccessibilityReadLoop(tabId, fnName, fnArgs, toolResult);
       const scrollCheck = this._checkNoProgressScroll(tabId, fnName, fnArgs, toolResult);
@@ -6159,6 +6160,21 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
     };
   }
 
+  /**
+   * Resolve the point a click call will actually dispatch at, for loop
+   * detection. Loop detection must measure the same CSS-pixel point the
+   * click dispatches: under downscaled screenshots the model's image-pixel
+   * nudges collapse to the same CSS pixel and the detector must see that,
+   * otherwise "slightly different spot" clicks never register as the loop
+   * they are. Non-numeric args fall back to Number() so the batch guard
+   * (x != null) and the detector keep their existing shape.
+   */
+  _resolveClickCoordsForLoop(tabId, args) {
+    const mapped = this._screenshotClickCoords(tabId, args);
+    if (!mapped) return { x: Number(args.x), y: Number(args.y) };
+    return { x: mapped.x, y: mapped.y };
+  }
+
   _coordinateReconciliationDiagnostic(point, resolution, clickPath, fallbackReason) {
     const target = resolution?.success === true ? resolution.semanticTarget : null;
     const resolved = clickPath === 'semantic';

test/run.js

@@ -11733,6 +11733,87 @@ test('screenshot click scale: from_screenshot converts image px to CSS px', () =
   }
 });
 
+test('coord click loop: helper resolves the CSS-pixel point the click dispatches', () => {
+  for (const AgentClass of [AgentCh, AgentFx]) {
+    const agent = new AgentClass({});
+    const tabId = 9;
+    agent.screenshotClickScale.set(tabId, { scaleX: 2, scaleY: 2 });
+
+    // from_screenshot:true converts image px to the CSS px the click fires at.
+    assert.deepEqual(
+      agent._resolveClickCoordsForLoop(tabId, { x: 100, y: 200, from_screenshot: true }),
+      { x: 200, y: 400 },
+      `${AgentClass.name}: from_screenshot coords not converted`,
+    );
+
+    // Without the flag, coords pass through unchanged.
+    assert.deepEqual(
+      agent._resolveClickCoordsForLoop(tabId, { x: 100, y: 200 }),
+      { x: 100, y: 200 },
+      `${AgentClass.name}: non-screenshot coords not passed through`,
+    );
+
+    // Flag set but no stored scale: no conversion.
+    agent.screenshotClickScale.delete(tabId);
+    assert.deepEqual(
+      agent._resolveClickCoordsForLoop(tabId, { x: 100, y: 200, from_screenshot: true }),
+      { x: 100, y: 200 },
+      `${AgentClass.name}: no stored scale should pass through`,
+    );
+  }
+});
+
+test('coord click loop: batch flow feeds converted CSS px into the loop detector', async () => {
+  for (const [label, AgentClass] of [['chrome', AgentCh], ['firefox', AgentFx]]) {
+    const agent = new AgentClass({ getVisionProvider: async () => null });
+    const tabId = label === 'chrome' ? 52705 : 52706;
+    agent._persist = () => {};
+    agent._skipPermissionGate = true;
+    agent.screenshotClickScale.set(tabId, { scaleX: 2, scaleY: 2 });
+    const seen = [];
+    agent._checkCoordClickLoop = (tid, x, y) => {
+      seen.push([x, y]);
+      return { kind: 'none' };
+    };
+    agent.executeTool = async () => ({ success: true });
+    agent._preflightRichTextToolbarTarget = async () => ({ block: null });
+
+    const messages = [];
+    await agent._executeToolBatch(
+      tabId,
+      [
+        {
+          id: 'loop_click_1',
+          function: { name: 'click', arguments: JSON.stringify({ x: 100, y: 200, from_screenshot: true }) },
+        },
+        {
+          id: 'loop_click_2',
+          function: { name: 'click', arguments: JSON.stringify({ x: 101, y: 201, from_screenshot: true }) },
+        },
+        {
+          id: 'loop_click_3',
+          function: { name: 'click', arguments: JSON.stringify({ x: 50, y: 60 }) },
+        },
+      ],
+      messages,
+      () => {},
+      { supportsVision: false },
+      null,
+      new Set(['click']),
+      1,
+    );
+
+    // The downscaled-screenshot clicks land on CSS pixels, not image pixels:
+    // (100,200) and (101,201) are distinct image pixels but collapse to the
+    // same CSS bucket, so the detector must be fed the converted values.
+    assert.deepEqual(
+      seen,
+      [[200, 400], [202, 402], [50, 60]],
+      `${label}: detector did not see the converted click points`,
+    );
+  }
+});
+
 test('visual target resolution stays private across Chrome and Firefox tools and prompts', () => {
   for (const [label, getTools, prompts] of [
     ['chrome', getToolsForModeCh, [