scholar-labs-search

name	scholar-labs-search
description	Search Google Scholar Labs through BrowserOS MCP to find the top 3 best-fit papers for a topic, with optional seminal or survey mode and manual handoff for sign-in or anti-bot interruptions. Use when Codex needs an on-demand literature search against Scholar Labs rather than a full literature review.

Scholar Labs Search

Use BrowserOS MCP to inspect Google Scholar Labs and return the top 3 best-fit papers for the user's topic. Treat Scholar Labs as the primary search surface. Use Exa only after selection to verify or improve outbound links.

Setup

Require BrowserOS running locally with MCP enabled before using this skill.

If the browseros MCP server is not configured in Codex, stop and tell the user to add it from BrowserOS Settings -> BrowserOS as MCP. Do not guess the MCP URL.

Use this config shape:

[mcp_servers.browseros]
url = "http://127.0.0.1:9001/mcp"

Workflow

1. Prepare the search query.
   • If the user gives keywords, use them as-is.
   • If the user gives a short problem statement, rewrite it once into a compact Scholar query using only the main topic and method terms.
   • If the first query returns zero usable results, retry once after dropping the weakest qualifier.
   • If the retry also returns zero usable results, return 0 papers found and stop.
2. Open Scholar Labs in BrowserOS.
   • Navigate to https://scholar.google.com/scholar_labs/search?hl=en.
   • Prefer new_page over new_hidden_page for active searching. Hidden pages are acceptable for inspection, but query submission may not fire reliably there.
   • After opening the page, use take_snapshot or get_page_content to confirm that the Ask Scholar textbox is visible.
   • Submit the query using the working BrowserOS flow described in workflow.md. Do not assume fill plus Enter will submit.
3. Handle blocked states exactly as described in workflow.md.
4. Extract up to the first visible results page, N <= 10.
   • Capture visible title, authors, year, venue or source, Scholar link, short annotation, citation signal if present, and result position.
   • Leave missing fields blank instead of inferring them.
   • Deduplicate obvious duplicate versions by normalized title plus year. Keep the higher-positioned or richer card.
5. Rank the candidates exactly as described in ranking.md.
6. Use Exa only after the top 3 are selected.
   • Replace the Scholar link only if identity is verified as described in ranking.md.
7. Return the result.
   • Default output: 3 papers.
   • For each paper, include title, authors, year, venue or source, best verified link, and a one-line reason it was selected.
   • If the user explicitly asks for all candidates, return the visible first-page candidates with one-line notes instead of only 3.
   • If fewer than 3 usable results remain, return the smaller set and say why.

Scope

Use only the first visible Scholar Labs page in v1. Do not paginate. Do not turn this into a full literature review. Do not silently fall back to OpenAlex or Semantic Scholar if Scholar Labs remains blocked.

BrowserOS Notes

• The reliable tool sequence is usually:
  1. new_page
  2. take_snapshot or get_page_content
  3. evaluate_script to set #gs_as_i_t, dispatch input, and click #gs_as_i_s
  4. exec_command with sleep 6 to wait for Scholar Labs to render results
  5. take_snapshot and get_page_content to poll for visible results
• fill and press_key Enter may leave the query in the textbox without submitting it.
• The visible circular button #gs_as_i_nsb is not the submit action for a new query. The submit control that worked in practice is #gs_as_i_s.
• When running multiple searches in one session, click New session before issuing the next query so result sets do not mix.
• Once the result cards are visible, capture only the first visible page and stop. Ignore deeper loading states such as Looking for more results....

Failure Rules

Stop instead of guessing when any of the following happens:

• BrowserOS MCP is unavailable.
• Scholar Labs remains blocked after one manual resume.

In those cases, say what failed and keep the response narrow.