Keep browser evidence inside useful output bounds

Output size is part of the interface

A browser page can contain far more information than an agent needs for one decision. Lupin therefore returns bounded text, links and controls rather than an unlimited DOM dump. The current snapshot includes up to 16,000 characters of body text, one hundred visible links and sixty selected visible elements. It sets truncated when body text exceeds its text limit. Treat these limits as part of the retrieval contract. A successful request can still provide an incomplete observation, and completeness must be assessed separately from HTTP success or a readable summary.

Ask a smaller evidence question

Before reading a long page, define the specific claim or action that the observation should support. An agent investigating one configuration option does not need every navigation link and product announcement on the site. Prefer the exact primary documentation page rather than a broad homepage when possible. If the site offers a supported paginated view or authorized export, use it deliberately. Repeatedly reading the same unchanged document does not increase the snapshot's cap. Nor does asking the model to infer the missing section make the unobserved material part of the evidence.

Understand the control window

The element list includes a limited set of visible interactive elements and may not provide a selector for every entry. A control missing from this list could be beyond the cap, inside another browsing context or represented by a custom component. Do not turn that absence into a claim that the application lacks the feature. Inspect a screenshot when it helps establish layout, navigate to a more focused page or hand the ambiguity to an authorized operator. Keep the target tab explicit so reducing the evidence scope does not accidentally switch the page under examination.

{"action":"read","tab_id":"TARGET_ID_FROM_TABS"}

Handle image bounds separately

The screenshot action captures a JPEG viewport and rejects image data beyond its current size bound. It is not a full-page archive, and its encoded payload can be awkward for clients that treat every result as text. Use native MCP image content or decode the API payload in the display layer that needs it. Avoid logging the full image string or sending it repeatedly to downstream steps. If the visual question can be answered by a fresh page read, use that smaller observation. If it cannot, record the image's limited scope and any capture failure accurately.

Preserve limits in the final answer

When summarizing a bounded observation, retain the source URL, retrieval time and any truncation caveat relevant to the conclusion. Do not describe an entire document as reviewed when only its initial rendered text was available. For workflows that require complete record sets, add an explicit pagination or export stage with a stopping rule and deduplication. Store only the useful evidence outside the session and close the browser when the task finishes. Output budgeting makes the workflow more reviewable when it narrows the question; it becomes misleading when a concise answer silently hides missing source material.

Sources & verification

Reviewed against the current Lupin route, provider adapter and shared input contracts, plus the linked official references. Examples were not run against live services.

• dom.spec.whatwg.org ↗Checked 2026-11-05
• chromedevtools.github.io · devtools-protocol · tot · Page ↗Checked 2026-11-05

Prepared with AI assistance by Lupin editorial. How we research and maintain these guides.