Reading Room 1.3.2

• The model you pick has to exist for your account. Assistant settings now lists the models each CLI on this machine actually offers — Codex's own list for your login (with the reasoning levels each model takes), Claude Code's aliases and effort levels, each CLI's configured default — instead of a list baked into the app. A Codex model or effort that isn't on its list is flagged in the settings and refused before a run starts, with the reason and the models to choose from. Before, a typo'd or unavailable model (gpt-6) booted fine and failed only once the command was in: The 'gpt-6' model is not supported when using Codex with a ChatGPT account, and the run spun.
• An assistant that answers with an error is reported at once. A run or discussion whose assistant replies with an error instead of working — a model it can't use, an expired login, a usage or rate limit — shows the assistant's own message on the card within a second. A fatal one (model / login) stops the run, rolls back and offers Assistant settings; a transient one (rate limit) stays advisory and clears once the assistant carries on.
• CLAUDE.md, AGENTS.md, GEMINI.md are now written for the reader's assistant — the commands, where data lives and the conventions for writing a digest, in ~65 lines — instead of the app's full design notes (Claude Code warned the file was over its size limit).
• Docs: the install pages say which release ZIP exists today; the Gatekeeper command strips quarantine recursively.

Reading Room 1.3.1

• The assistant's first launch works. The app now judges an assistant by what it
  shows, not by how many bytes it printed: Claude's, Codex's and Gemini's trust dialogs
  (and their login screens) are recognised, the terminal comes up on top of the
  Analyze card or the Setup window so you can answer right there, and the command is
  sent only once the assistant's own prompt is on screen. Before, the "answer in the
  terminal below" card hid the terminal behind itself, Show terminal lost the
  Continue button, and the command could be typed straight into the trust dialog — which
  made Claude quit (its default is "No, exit") and left Codex and Gemini idle.
  Continue is offered only when the app cannot tell what the assistant is showing.
• Setup keeps its window. Launching an assistant from the last step no longer closes
  Setup for a small bar at the top: the terminal opens over it, a status line says what the
  assistant is asking, and Apply setup runs from the same window — refusing, with the
  reason, while a trust or login prompt is still up. Closing Setup part-way and reopening
  it resumes where you were; Esc in a field only leaves the field.
• Tutorial: every screenshot is shown whole in a fixed 3:2 frame (the Setup step was
  cropped to its top half), the card keeps one height so Next stays under the pointer,
  images are preloaded, and steps slide instead of re-laying out.
• Starting a flow while a Discussion is running now shows the discussion and offers End
  chat / Stop instead of killing it silently; End chat appears once the discussion has
  actually started. Esc typed in the terminal belongs to the assistant (it no longer closes
  the drawer, a panel or Setup). A missing CLI fails the run with that reason at once.
• A run whose page failed to build is reported as such (the file is kept) instead of
  "Added to your library"; a run whose assistant quit while no window was open is ended
  and rolled back by the server after 30 s, and a page that opens later is told so.
• Assistant choice, model/effort settings and the sort order now travel with the install
  (they survive a port change and a new browser), like the reading state.
• READMEs: the app's is short and points at the website; the website's summarises the app.

Reading Room 1.3.0

• Analyze a PDF from your computer. Choose PDF… in the Analyze panel, or drop
  one (or several) anywhere on it: the file is copied into papers/ and analyzed
  like any other paper. Most fields are not on arXiv, and the app window has no file
  browser of its own — this is how a paywalled or scanned paper gets in.
• A queue, not one paper at a time. Add another stacks papers up; the button
  then reads Analyze N papers and works through them in order, saying which one it
  is on, and finishes with a list of everything it added.
• Per-run options on Analyze: depth, audience and a focus for this run only,
  without typing flags. Your profile's defaults still apply when they are left alone.
• Reading state is saved to disk (user/reading-state.json) as well as in the
  browser. Status, stars, collections and dismissed graph nodes now survive a new
  browser, a different port (the app falls back when 4317 is busy — the library used
  to come back unread), and a restore from backup.
• A run outlives its window. Closing the app while a report is being written no
  longer kills it: the server stays up until the run lands, writes the report, and
  then quits. The run card also shows elapsed time and the assistant's last line, so
  a long job looks supervised rather than hung.
• Remove paper on a report page — it lists exactly what will be deleted and
  warns about comparisons built on that paper, then rebuilds.
• Diagnostics (avatar menu): versions, which assistants were found on PATH, how
  this copy was installed, the library's size and the tail of the server log, with
  one-click Copy.
• Backups before updates, and on a schedule. Every update now writes a PDF-less
  snapshot of the library first; Back up gained a daily/weekly/monthly automatic
  backup (only the five newest are kept).
• Updates follow published releases, not the branch head: a download now updates
  to the newest release (preferring the packaged ZIP for its platform), so it can
  never land mid-work code, and a clone names the newest version tag.
• Linux launcher (ReadingRoom.sh, with a menu entry on first run) and
  scripts/package_release.sh, which builds a release ZIP with the app's
  dependencies already installed — no npm install or native build on first launch.
• Sort the catalogue and the Library (year, recently updated, title, priority,
  reading status); the Library filter now also searches ids, years and summaries.
• Accessibility: the full-screen panels are proper dialogs (focus moves in, Tab
  stays inside, Esc closes, focus returns), run status is announced, graph nodes are
  keyboard-reachable, and the terminal has an optional screen-reader mode.