Open a document on an infinite canvas so the human can read it and dive down rabbit holes. Start a NEW hole with { title, content } (or { title, file_path }), or RESUME a saved one with { hole_id } (use list_rabbitholes to find it). When opening content fetched from a URL or repo, pass the document's own URL as base_url so relative images and links resolve. For local images that are not on the web, pass assets and reference them as . For a PDF already processed with ingest_pdf, pass ingest_id when starting the new hole and reference the returned asset names as . The canvas opens in the browser and this call BLOCKS until the human acts. It returns status='branch_request' when the human selects text and asks a question — answer it with answer_branch. A branch_request with EMPTY selected_text is a follow-up question about the parent document as a whole (a chat reply beneath it) — answer conversationally in that document's context. A branch_request may carry a 'lens' (explain | eli5 | example | deeper) — the question text spells out the style the human tapped; honor it. One marked saved=true was asked while no agent was listening — answer it like any other. On a resumed hole the first branch_request carries a 'rehydration' field with the whole tree (and any saved_asks); read it to reload your context. Long waits periodically return status='keep_listening' with hole_id; immediately call open_rabbithole { hole_id } to keep listening, and do not re-send content. If the host reports a tool timeout (e.g. timed out awaiting tools/call), also re-call open_rabbithole { hole_id }; nothing is lost and asks are saved. It returns status='session_closed' when the human clicks Done or closes the tab.

Extract a local PDF into Rabbithole image assets and per-page text. Produces 2x page render PNGs named page-001.png, page-002.png, etc. plus opportunistic embedded rasters named embed-p001-01.png when the PDF contains extractable images. The agent should compose markdown itself, using page renders as the dependable figure source and embedded rasters when they are cleaner, then call open_rabbithole with the returned ingest_id (or pass hole_id here to attach assets directly to an existing hole). For arXiv links, prefer fetching the HTML version and opening that markdown with base_url instead of ingesting the PDF.

Answer one pending branch request from an open Rabbithole. Called after open_rabbithole or answer_branch returns status='branch_request'. Write a focused, well-formatted markdown answer to the human's question about their selection - use selected_text, parent_node_title, and lineage for context (you already hold the documents you authored). If selected_text is empty, answer conversationally about the parent document as a whole. If the request has a 'lens', match that style.

Authoring vocabulary:

• Base notation: GFM markdown, $...$/$$...$$ and (...)/[...] math, and highlighted language-tagged code fences.

• If the answer is content fetched from a URL or repo, pass its document URL as base_url so relative images and links resolve.

• If the answer uses a local image, pass assets: [{ name, file_path }] and reference it as ; use this for screenshots, generated diagrams, and other non-web images.

• Use ```show when a concept is spatial or structural: architecture, memory layout, relationships.

• show dialect: HTML/CSS/inline-SVG only; no scripts. Scripts and unsafe attributes are stripped.

• show craft: prefer HTML/CSS layout with flexbox/grid over absolute SVG coordinates.

• Design visuals for about 380px card width; make them fluid and keep labels short.

• Use theme tokens, never hardcoded colors, so visuals match light and dark themes: --fg, --fg-bold, --fg-dim, --fg-faint, --node-bg, --bar-bg, --border, --border-focus, --accent, --accent-contrast, --code-bg, --hl, --hl-strong, --warn, --font-ui, --font-doc, --font-mono.

• Example show:

<style>.flow{display:grid;gap:8px}.box{border:1px solid var(--border);padding:8px;border-radius:6px}</style>
<div class='flow'><div class='box'>Parse</div><div class='box' style='background:var(--hl)'>Render</div></div>

• Streaming choreography: send prose in 1-3 sentence chunks as usual.

• Emit each visual fence contiguously, ideally in one chunk; readers see a placeholder until the fence closes.

• Interleave prose -> visual -> prose when useful. Use a visual only when it genuinely carries the explanation.

Finish streaming by sending the remaining final chunk in a normal call with a short 'title'. Partial chunks concatenate verbatim: include your own spacing/newlines and never repeat text already sent. The final call blocks and returns the next event. If it returns status='keep_listening', immediately call open_rabbithole { hole_id }; if the host reports a tool timeout (e.g. timed out awaiting tools/call), do the same. Do not re-send content; asks are saved.

List saved Rabbitholes (most recently updated first) so you can resume one by hole_id via open_rabbithole. Returns id, title, last-updated time, and node count for each.