pagemap — a compact page map for a self-trained browsing agent

A DOM walker in JavaScript and a Python renderer produce [ref:role 'name' -> target] lines plus sections, headings, lists and tables, stitch iframes and shadow DOM in, and write the refs back onto the live elements so the agent can click them. Tuned on real pages: a bookshop, Hacker News, Wikipedia, GitHub and Amazon.

PeriodSep 2026StatusResearchAreaTools, agents & research

Highlights

68 tests against a kitchen-sink fixture; Playwright drives the installed Edge, so nothing is downloaded.
The format is the model's input, so it is specified in the README and kept stable; next come a recorder for real browsing sessions and a LoRA fine-tune of a small open model.
Deliberately avoids distilling a commercial model's outputs into the training data.