holoscrape

HoloScrape drives the person's OWN already-signed-in Chrome. There is no cloud browser and no separate profile: every page you read is the page they would see, with their logins and their IP. Treat their browser as someone's desk you are working at.

This guide is organised by the SITUATION you are in, not by tool. Every fact in it was paid for by a failed session; the measurements are kept because a rule without its number gets argued with. The tool descriptions are short on purpose and the conditional warnings arrive as "hint" lines on the replies that prove them, so this is where the procedure and the history live.

Before the first call: which tool, and what it costs

EVERY TOOL NAMED IN THIS DOCUMENT EXISTS ON THIS SERVER. Your client may have shown you only a few of the TEN — schemas are fetched on demand, and a keyword search returns the top matches, not the set. So the tools you were handed are NOT the tools there are. Before concluding that something is impossible, look for it here by name and ask for that exact name.

The ten: current_page, tabs_list, tab_here, page_study, page_grow, page_state, list_extract, page_harvest, results, resolve_x_video. Two of them carry most of the surface: page_state takes the pseudo-paths (@dom, @html, @map, @collect, @await, @fetch, @net) and results takes actions (status, stop, list, get, export, download, merge, guide). Counting those as separate tools is how this line used to say "twenty-one", which sent a reader hunting for eleven tool names that do not exist.

Measured, twice: a session reported "there is no direct navigate-this-tab-to-a-URL primitive" and re-clicked through a list page for every record. tab_here is precisely that primitive and had been shipped for weeks. Another reasoned its way to visiting 250 pages by hand with page_harvest sitting unread. Both were correct about what they had been SHOWN.

• tab_here — POINT AN EXISTING TAB AT A URL. The navigation primitive. Use it instead of tab_here newTab:true (which leaves a tab per destination for someone to close by hand) and instead of pressing links (page_grow mode:"walk" runs INSIDE the page, so an ordinary link destroys the frame it is running in and the answer is lost — the tab arrives, you get nothing).
• page_harvest — MANY PAGES, ONE CALL, NO ROWS THROUGH YOU. A list page and the records behind its links. Measured on the same 250 pages: 52 minutes and 733 calls by hand, 3m13s and one call with this. If you are about to visit a set of addresses you already hold, this is the call.
• current_page — start here for "I have a page open, scrape it". It takes no arguments and returns a tabId to pass to page_study or list_extract.
• tabs_list — only when current_page is not the page they meant and you need to ask which; for the ordinary case prefer current_page.
• page_study — the cheap ranked look: what repeats, how it paginates.
• page_state path:"@dom()" — read what is rendered, no ranking. The fix when a run came back short, and the only way to read an icon rail whose names live in aria-label.
• page_state path:"@html( :: )" — the markup itself. Reach for it the MOMENT another tool refuses something you can plainly see. Reading the markup answers in a minute what an hour of inference does not.
• list_extract — many UNIFORM rows you want followed through pages and saved as an exportable table. It must first agree something is a list; a 23-icon rail and a chat log are not, and it will return one row and fragments rather than refuse. That is not a bug in the page.
• page_state (a store path) — for records whose fields the DOM never renders.

WHAT A CALL COSTS. Every description ends with a tier, because the server knows it and used to keep it to itself. "Cost: minutes" (list_extract, page_harvest) presses, scrolls or walks pages one after another, so budget for it and prefer a cheaper tool when one will answer. "Cost: seconds" (tab_here, page_grow) loads or moves a page. "Cost: instant" (current_page, tabs_list, page_study, page_state, results) reads what is already there and is cheap enough to call before guessing. An agent that does not know a walk runs for minutes calls it to answer a question one read would have answered; one that does not know page_study is instant avoids it and guesses instead. Both are avoidable. Two tools are USUALLY instant and OCCASIONALLY long: page_state answers a property read in milliseconds but also runs a gesture-paced @collect that pages a chat backlog, and page_grow takes up to 50 hops.