I believe this is because extension_custom_bindings.js (http://src.chromium.org/viewvc/chrome/trunk/src/chrome/renderer/resources/extensions/extension_custom_bindings.js?revision=HEAD&view=markup) references chrome.windows.WINDOW_ID_NONE. However,  with http://crrev.com/115252, the chrome.windows namespace is no longer available if the extension doesn't request access to the windows API.

The WINDOW_ID_NONE constant should probably be moved to a place where all bindings can see it, regardless of which APIs the extensions uses.

How about we always insert constants into the bindings? Or at least have an option for doing so in the constant definition, like export=true?

Then we'd add this even without access to the windows API:

chrome.windows = {
  WINDOW_ID_NONE: -1,
  WINDOW_ID_CURRENT = -2
}

I'll take this bug since I introduced it with the refactor.

It's more complicated to fix than just adding an "export=true" property to the API json. For a start, the only json that gets passed to schema_generated_bindings in the first place is those which the extension has access to; which "windows" isn't necessarily one of.

So for this approach either:
 - we export all the extension APIs, and schema_generated_bindings.js decides which to inject and which to ignore. It can inspect properties with "export=true" etc etc. This is a gross amount of complexity though, because SGB doesn't have knowledge of this yet, it's complicated enough, there are no tests... etc.
 - just always inject extension APIs and go back to just having the permissions checks on the browser side. This is obviously not an option, since it's the reason for all this in the first place.

Always inserting constants would have a similar problem, but we'd need the additional complexity of having a heuristic to decide which properties are constants. I guess that's why you mention the export=true thing.

But taking a step back, having export=true, I think, defeats the purpose of having these stricter permissions checks and conditional inclusion of APIs. Here are a couple of other ways we could fix this:

 - perhaps, arguably, we should  always be injecting "windows" anyway. We always inject "tabs" because there are functions in tabs that require no permissions, so we have no choice... and since they share the same permission, maybe this means they should both be injected? It would add irritating complexity to the permissions set, but, oh well.
 - just write -1.
 - copy the constant somewhere else and use it there.

The following revision refers to this bug:
    http://src.chromium.org/viewvc/chrome?view=rev&revision=118573

------------------------------------------------------------------------
r118573 | kalman@chromium.org | Fri Jan 20 16:53:25 PST 2012

Changed paths:
 M http://src.chromium.org/viewvc/chrome/trunk/src/chrome/renderer/resources/extensions/extension_custom_bindings.js?r1=118573&r2=118572&pathrev=118573

Fix crash in extension_custom_bindings.js by defining WINDOW_ID_NONE locally
rather than looking up chrome.windows.

BUG=110694
TEST=as in bug

Review URL: https://chromiumcodereview.appspot.com/9138009
------------------------------------------------------------------------

(No comment was entered for this change.)