Handling Browser 5MB Limit Error (image exceeds 5 MB maximum)

📌 Context of the Problem

The underlying AI engine (Vertex AI) has a hard limit of 5 Megabytes per image payload.

Actions that trigger this crash immediately:

1. Instructing the browser_subagent to perform a long scroll taking screenshots of the entire page ("Take a full page screenshot by scrolling down").
2. Requesting the subagent to capture web components with extremely high-definition load, gradients, and simultaneous overlays.
3. Accumulating more than 2 screenshots from a modern display in a single interaction (Response).

The reported error looks exactly like this: messages.62.content.1.tool_result.content.1.image.source.base64: image exceeds 5 MB maximum: 6255808 bytes > 5242880 bytes and usually returns an HTTP 400 Bad Request or HTTP 401 Unauthorized.

🛑 Preventive Actions (BEFORE using the Browser Subagent)

When you decide to visually audit an application using the temporary browser, ALWAYS explicitly add these constraints inside your TaskPrompt:

1. NO INFINITE SCROLLING (Full-Page Screenshot): Never ask it to review the structure by taking a screenshot of every piece.
2. PRIORITIZE THE DOM: For text and HTML class audits, specifically instruct the subagent to use browser_get_dom.
3. MAXIMUM 1 OR 2 SCREENSHOTS: Add as a strict guideline inside the prompt sent to your subagent sibling: "CRITICAL: Do NOT return more than 1 or 2 screenshots total to avoid crashing the AI engine. Use DOM reading for the rest of your context."
4. SEPARATE MOBILE/DESKTOP VERSIONS: Never order it to repeatedly Resize -> Screenshot -> Resize -> Screenshot. Ask for it in different iterations or mitigated parallel tasks.

🛠️ Reactive Actions (WHAT TO DO IF IT ALREADY HAPPENED)

If during your planning a user sends you a 400/401 error with this footprint (TraceID), it means you have broken the memory with your visual requests:

1. Immediately Acknowledge the Cause: Do not lie to the user, explain that the returned images weighed more than 5MB.
2. Use Other Tools: Momentarily stop the "full visual audit". Let the context rest.
3. Re-Audit Exclusively Analytically (DOM): Launch the browser_subagent again but with the severe order: "DO NOT take any screenshots at all. Try to describe what you see exclusively by reading the DOM (classes, ids, texts).".
4. Trust User Feedback: If the subagent failed due to visual memory, ask the user what they saw on their screen to patch it with code (view_file, multi_replace).

Remember: We are efficient at writing code, not consuming Google's memory trying to view 10000-pixel-high web pages in a single view.