Neither change adds a new button in the product. They tighten what the OpenSearch research stack can write into HTML and which tools the executor may call when the brief needs equations or citable papers with a DOI.
LaTeX and math in reports
The executor and assembler are instructed to put raw TeX inside <span class="katex-source">…</span>, optionally with data-display="inline" or data-display="block" for short in-sentence math vs display-style equations. The reader turns those spans into KaTeX in the UI.
Plain $…$ delimiters alone are not treated as math—the stored HTML uses the span convention above. The prompts also warn against leaving formulas only in bold tags, which often read as plain superscript text.
DOI lookup with Unpaywall
The lookup_unpaywall_by_doi tool calls the Unpaywall v2 API for a single DOI and returns the full JSON (for example OA status, best OA location, and oa_locations entries that may include URLs). It does not download the paper itself.
In the agent prompt this tool is described as fallback only: use it after normal search and visit_website when you still cannot read the paper (paywall, errors, or useless content). When Unpaywall shows a legal open PDF URL, the same visit flow can be used on that URL; the prompt notes to watch for bot walls and try another location if needed.
What stays the same
You still start research the same way. The model picks up the new instructions and tools from the server-side prompts; outputs improve when your topic needs math or when a DOI is the cleanest handle on a paper.
Full product history lives on the changelog.