Agent ergonomics: cheaper act→observe loop, two small tool-surface trims #11

Activity

1. jackwener commented on Oct 22, 2026

   

   jackwener

   on Oct 22, 2026

   Owner

   More actions

   Two more ergonomics gaps from a real task ("brush an X timeline for 100 posts, then save it as Markdown"). Both cost tokens or force hand-rolled loops:

   1. No way to write a file from the REPL.
   The only path to disk is save_file in the MCP tool surface, so content generated inside js has to be round-tripped through the model context:

   nodeRepl.write(mdText);   // 11k chars, needs a maxChars bump to avoid truncation

   then paste the same 11k chars into save_file. That doubles the cost of every "export this page" task. A tools.saveFile(path, content) binding, or a workspace-scoped fs inside the REPL, would remove the round-trip entirely.

   2. No "collect while scrolling" primitive for virtualised lists.
   tab.read() only sees mounted nodes: on X, ~12-18 article elements exist at any moment, and they are recycled as you scroll. Every agent therefore hand-rolls scroll + dedupe:

   for (let i = 0; i < 8; i++) {
     collect();                                       // dedupe by /status/ href
     document.scrollingElement.scrollTop += innerHeight * 1.05;
     await sleep(1400);
   }

   Station adapters do not help here: the twitter.timeline adapter pages through API cursors and never moves the visible page, so it cannot satisfy "scroll the page for me". Suggestions: a tab.collect({ scrolls, key }) helper, and/or a visual: true mode on site commands that also drives the UI.

   (Related: the twitter.timeline cursor call also aborted with MCP request aborted once mid-page, which after the abort left the REPL bindings cleared — see the separate issue about silent state loss.)