feat(MainActivity): cosmetic filter for inline sponsored content (#89)

[jim-daf]

Cosmetic filter for inline sponsored content (refs #89)

Discusses #89

Background

The discussion in #89 (raised by the geminiAssist fork author) points out a structural limit of the current URL allowlist. ChatGPT now injects ad content into the conversation UI from its own domain, so the network layer cannot block it. In a real browser, uBlock Origin would handle this with cosmetic filters. This PR brings the same idea to the WebView with a small evaluateJavascript hook.

Design

One static JS constant, one WebViewClient.onPageFinished override, and a MutationObserver so streamed content is also caught:

@Override
public void onPageFinished(WebView view, String url) {
    super.onPageFinished(view, url);
    if (url == null || url.equals("about:blank")) return;
    if (!url.startsWith("https://chatgpt.com") && !url.startsWith("https://chat.openai.com")) {
        return;
    }
    view.evaluateJavascript(COSMETIC_FILTER_JS, null);
}

The script itself:

(function(){
  var SELECTORS = [
    '[data-testid*="sponsor" i]',
    '[aria-label*="sponsor" i]',
    '[data-message-sponsored="true"]',
    'a[href*="utm_source=chatgpt"][data-sponsored]'
  ];
  function hideAll(){
    for (var i=0;i<SELECTORS.length;i++){
      var nodes = document.querySelectorAll(SELECTORS[i]);
      for (var j=0;j<nodes.length;j++){ nodes[j].style.setProperty('display','none','important'); }
    }
  }
  hideAll();
  if (window.__gptAssistAdObserver) return;
  window.__gptAssistAdObserver = new MutationObserver(hideAll);
  window.__gptAssistAdObserver.observe(document.documentElement, {childList:true, subtree:true});
})();

Why these selectors

ChatGPT does label sponsored content with attributes whose names contain the literal word "sponsored", as the OP noted. The four selectors above match the patterns currently observed without overreaching into normal chat content. The [i] suffix on attribute matches makes the comparison case insensitive, which protects against minor casing changes.

Maintenance

Selector churn is the obvious risk. Keeping the list as a single Java string constant means an update is one constant edit. The MutationObserver is installed once per page (window.__gptAssistAdObserver guard) so repeat calls from onPageFinished are cheap.

Scope

Only chatgpt.com and chat.openai.com URLs are touched. Anything else (Google login redirect for example) gets no script injection at all.

Files changed

• app/src/main/java/org/woheller69/gptassist/MainActivity.java

[Copilot]

Pull request overview

Adds a WebView-side “cosmetic filter” to hide inline sponsored content inside ChatGPT pages, addressing the limitation that network allowlisting can’t block same-origin injected ads.

Changes:

• Introduces a static JavaScript snippet containing DOM selectors and a MutationObserver to hide sponsored UI elements.
• Injects the script via WebViewClient.onPageFinished() for ChatGPT origins (chatgpt.com / chat.openai.com).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This comment says the cosmetic filter is for “chatgpt.com”, but the injection logic also targets chat.openai.com. Update the comment to reflect both domains (or the shared purpose) so future maintenance doesn’t miss one.

Suggested change

	// Issue #89: cosmetic filter for inline sponsored content on
	// chatgpt.com. Selectors are intentionally narrow and live in
	// one place so they can be revised when ChatGPT changes its
	// markup. The MutationObserver covers content streamed into the
	// chat after the initial paint.
	// Issue #89: cosmetic filter for inline sponsored content in the
	// ChatGPT web UI on chatgpt.com and chat.openai.com. Selectors
	// are intentionally narrow and live in one place so they can be
	// revised when ChatGPT changes its markup. The MutationObserver
	// covers content streamed into the chat after the initial paint.

[Copilot]

The url.startsWith("https://chatgpt.com") / startsWith("https://chat.openai.com") check can be bypassed by URLs like https://chatgpt.com.evil.tld/... (still starts with the same prefix). To ensure JS is only injected into the intended origins, parse the URL and compare scheme/host (and optionally allow specific subdomains) rather than using a raw string prefix check.

Suggested change

	if (!url.startsWith("https://chatgpt.com") && !url.startsWith("https://chat.openai.com")) {
	return;
	}
	Uri parsedUrl = Uri.parse(url);
	String scheme = parsedUrl.getScheme();
	String host = parsedUrl.getHost();
	if (!"https".equals(scheme) ||
	host == null ||
	(!host.equals("chatgpt.com") && !host.equals("chat.openai.com"))) {
	return;
	}

[Copilot]

MutationObserver(hideAll) runs hideAll() on every DOM mutation, and hideAll() performs multiple querySelectorAll scans over the entire document each time. On ChatGPT’s streaming UI this can create unnecessary CPU/battery overhead. Consider injecting a single <style> element with the selectors and display:none !important (which will apply to future nodes automatically), or at least debounce/throttle hideAll() inside the observer callback.

Suggested change

	// markup. The MutationObserver covers content streamed into the
	// chat after the initial paint.
	private static final String COSMETIC_FILTER_JS =
	"(function(){" +
	" var SELECTORS = [" +
	" '[data-testid*=\"sponsor\" i]'," +
	" '[aria-label*=\"sponsor\" i]'," +
	" '[data-message-sponsored=\"true\"]'," +
	" 'a[href*=\"utm_source=chatgpt\"][data-sponsored]'" +
	" ];" +
	" function hideAll(){" +
	" for (var i=0;i<SELECTORS.length;i++){" +
	" var nodes = document.querySelectorAll(SELECTORS[i]);" +
	" for (var j=0;j<nodes.length;j++){ nodes[j].style.setProperty('display','none','important'); }" +
	" }" +
	" }" +
	" hideAll();" +
	" if (window.__gptAssistAdObserver) return;" +
	" window.__gptAssistAdObserver = new MutationObserver(hideAll);" +
	" window.__gptAssistAdObserver.observe(document.documentElement, {childList:true, subtree:true});" +
	// markup. Inject a single stylesheet so the rules apply to both
	// current and future matching nodes without rescanning the DOM.
	private static final String COSMETIC_FILTER_JS =
	"(function(){" +
	" var STYLE_ID = 'gptassist-cosmetic-filter';" +
	" if (document.getElementById(STYLE_ID)) return;" +
	" var SELECTORS = [" +
	" '[data-testid*=\"sponsor\" i]'," +
	" '[aria-label*=\"sponsor\" i]'," +
	" '[data-message-sponsored=\"true\"]'," +
	" 'a[href*=\"utm_source=chatgpt\"][data-sponsored]'" +
	" ];" +
	" var style = document.createElement('style');" +
	" style.id = STYLE_ID;" +
	" style.textContent = SELECTORS.join(',') + '{display:none !important;}';" +
	" (document.head || document.documentElement).appendChild(style);" +

[Copilot]

Comment typo: “re runs” should be “re-runs” (or “reruns”).

Suggested change

	// the sponsored markers ChatGPT emits, and re runs on
	// the sponsored markers ChatGPT emits, and re-runs on