Angel

Long attachments (>=6000 chars, up to 1MB) no longer dump into context - they're stored as out-of-context documents Angel reads in bounded passes with a read_document tool. Key design turn from Eli mid-build: don't list docs in his standing context every turn (that's clutter, the same mistake as force-feeding memory instead of leaving it behind recall). Instead the doc catalog sits behind a list_documents tool, and the message that attaches a file carries a compact <document id title lines/> pointer so Angel sees it arrive and starts reading right away. Deployed. Stage 1 (background read job, carry-forward, mid-read message injection, quiet progress reporting) is next if the experiment earns it.

shipped a big client polish batch - composer buttons moved to a row below the input with attach as +, rename a conversation by clicking its title in the top bar, archive uses an icon and collapses into an Archived section, date-line separators, PWA install via a network-first service worker, tighter connection tracking (offline/online + faster heartbeat), a Reload menu that clears the SW cache. Also reworked tool-call UX: a tool now announces itself as its name streams, so a long generation shows 'Recording observation...' then 'Recorded observation' (kind only), with truncated attempts cleared back to the last committed boundary so no phantom tool is saved. Conversation naming is exchange-only now and labels the first greeting of the day with the date. One notable reversal: built a 'main conversation + tangents' organizing model, deployed it, then rolled it fully back - Eli doesn't think that way; his real model is one daily conversation plus tangents as flat peers over the single shared context. Now pivoting to the core 'slow reading' feature: Angel reads documents section-by-section held OUT of context via a bounded read_document(id,start,end) tool, recalling and recording integrated observations as he goes. The bet: sequential out-of-context reading produces better-integrated memory than dumping a doc into context. Motivated by two concrete needs - getting Angel current on what Eli's done with Glopus, and setting him up as a domain consultant (Japanese acupuncture for Yael) from thick books not available online (which will need PDF-to-text ingestion after the text loop is proven).