Changelog

they reach refto.one the way models and downloads do — the proxy in Settings → Network, or the system's (System Settings) when it is set to automatic — and, when an automatic proxy cannot be reached, they go direct instead of failing. Where the direct way is slow, activating no longer said “Cannot reach refto.one” the first time and the daily check no longer failed with a timeout.

after changing Refto's language, Settings → About → Show Changelog no longer showed the notes in the language before until opened again. While it reads the notes again it says Refreshing… over those shown, and if that fails, says so.

EPS files without a preview inside whose thumbnail failed because no Ghostscript would draw them — none installed, too old, its use turned off, or it could not start in its sandbox — are drawn again, in the background, once one would: after brew install ghostscript or brew upgrade ghostscript (noticed within a minute, or when the library opens), after turning it on in Tools → EPS, or after an update of Refto changes Ghostscript's sandbox. Files that failed for another reason stay as they are. When Ghostscript draws nothing, the file's page now says why — not installed, turned off, too old, no sandbox it could start in, or it failed (an error, too long, a picture too big) — instead of always “Install Ghostscript”, and the log notes it in a line (the file's path and how gs ended, never what it printed).

when every item of the AI's batch is deleted or moved to the Trash before it is analysed, its card goes without a “done” card; one that went through some counts only those.

items moved to the trash or deleted while their AI analysis waited (behind a backlog, or held while the Mac is short of memory) now leave it at once — the jobs of nothing but such items are cancelled and the sidebar's AI analysis card finishes — instead of staying until those jobs had run.

what you start that writes much — an import from Eagle, a .refpack or .eaglepack, files dropped or pasted, files an agent or the API adds, a Finder mirror, a video saved with yt-dlp, a preview copy of a video — now checks first that it leaves the disk Refto keeps free (2 GB or 2 % of it, whichever is more); before, a copy from another disk could write the system disk down to nothing. When it does not fit, an import stops whole and says how much it needs and how much is free — what came in stays, and run again once there is room it goes on where it stopped — instead of listing thousands of items as failed; a disk found full partway (another app filled it) stops it the same way. A video may take at most what the temporary folder's disk and the library's leave over that floor (and 4 GB, as before), and an .eaglepack is unpacked only when all of it fits.

yt-dlp runs in a process group of its own, so a cancelled download (or one over its size) no longer leaves the ffmpeg it started merging or fetching the stream, writing into a folder already removed.

each time the AI had stored an item's vectors, the next search by meaning (the search field, the notch shelf's search, agents, plugins) read every item's vectors again — about 0.3 s and 250 MB of memory churn in a library of 50,000 items, for every search while an import was analysed. It now reads only the items that changed: about 14 ms instead of 0.3 s and 11 MB instead of 250 MB; what it finds is the same.

a grid cell whose larger thumbnail failed to load (not made yet, say) showed its small one for the rest of the run, blurry when zoomed in. It is now tried again after 10 seconds, then less and less often (at most every 5 minutes).

as the AI read the text of a long scanned PDF (a few thousand pages of Chinese), each time it saved what it had read held up a rating, a tag or a rename made meanwhile by up to half a second; the text is now prepared for search before.

watching a folder with its subfolders, Refto went through every file it knew once for each subfolder as it started — about 1.3 s for 1,000 subfolders of 20 files, at every launch; now about 0.06 s.

the activity card under the sidebar copied and sorted every waiting job each time one moved on — about 34 MB of memory churn for each update with 10,000 files queued, up to ten times a second. It now only counts them; the Tasks page sorts them when open.

the vectors of tag names the AI suggests from were all written again (about 85 MB for 10,000 tags) every time a new tag appeared, and those of renamed or deleted tags were kept for good. New ones are now added to the end of the file, renamed or deleted tags' go, and at most 20,000 are kept.

the AI cut the sound of a video Apple's frameworks cannot read (WebM, MKV, AVI…: through ffmpeg) by reading all of it into memory at once — about 450 MB more for each hour, while a newly added video was analysed in the background. It now reads a window at a time: a few MB, however long the video; what the AI finds is the same.

the daily clean-up now also removes what a web page's capture left in a library when Refto quit or crashed in the middle (they no longer go into full backups either), the larger thumbnails of items deleted from the trash or given another version as their file, unfinished video copies, frames and waveforms in the caches, the thumbnails of format extensions and plugins being made, and — once unused for a month — the caches of libraries Refto no longer lists. Hover strips and waveforms are kept under 512 MB a library, the oldest going first (they are made again when needed).

a word found in recordings' transcripts or PDFs' text no longer has each hit's line looked up in every text, every search: only the hits on show (the grid's first screen, an agent's page) are placed at their second or page, the others as you open them; a long PDF's text is no longer cut into pages to find a line in it. In a library of 100,000 items (300 transcripts, 300 long PDFs), a search for a common word in the grid takes about 20 ms instead of 47 ms and makes 4.5 MB of garbage instead of 39 MB; the notch shelf's search reads its hits once (to 300) and shows the next pages from them — scrolled to its end, 5 ms instead of 0.5 s and 370 MB; an agent paging through a search 20 at a time, 0.2 s instead of 1 s.

Refto looked for each missing linked file in its folder and the one around it again on every check (when a library opens, then every 10 minutes): 500 files deleted from a folder of 3,000 were 3–6 s of work each time. A folder is now listed once for all its missing files, and a file missing already is looked for again after half an hour, then twice as long each time (6 hours at most), or as soon as its folder changes: 1–10 ms. Those checks, the scheduled backups' list of linked files and Settings → Collect's posts without their videos also read the library on their own connection, not in front of the grid's.

clicking Card, Screenshot or Live page above an item's page left the keyboard on the switch, so ← → switched the views instead of turning the pictures (or the items). A click now gives the keyboard back to the page; reached with Tab, the switch still takes ← → as before.

a picture shown for the first time in the Screenshot view slid in as the page's colour while it was being decoded. The pictures beside the one shown are now decoded ahead, and a small version stands in until the full one is ready.

on an item's page, the posts you turned away from no longer keep their Screenshot view's pictures decoded (up to 64 MB each, two of them beside the one shown); going back to one shows the picture you left it on, small for a moment as it is decoded again. A long animated GIF on its page now plays whole, a little smaller, within 64 MB of frames (160 MB before), instead of stopping after its first frames.

closing Refto's window (it keeps running in the menu bar and the notch) or minimizing it no longer keeps its pictures decoded for hours — the grid's thumbnails (64 MB or more), an item page's picture, the pages beside it and a post's pictures (up to a few hundred MB). Once the window has been away 30 seconds they go; shown again, it looks as you left it (the item, its zoom, the slide) and decodes what shows, as after switching libraries. Another library's window that stays open keeps its own. The notch shelf and the captures' shelf let go of their cards' pictures 30 seconds after they are closed or put away the same way.

messages such as "It is on the board already" covered the board's tools at the bottom; on a board they now sit above the toolbar. Other pages are unchanged.

the outline of a picture clicked or right-clicked on the card (Card view) was square at the grid's round corners — its inside edge showed a sharp corner under the rounded one. It now runs inside the picture with the grid's corners, in every layout (one picture, two, three, four, a note's album), in Light and Dark Mode; a video's focus ring too.

the open notch shelf reaches into the menu bar beside the notch, and where Refto's icon sat under it (a menu bar with few icons), a click on the icon went to the shelf instead — the shelf stayed open and kept the keyboard. With the pointer on the icon the shelf now lets the click through: the icon closes the shelf as it should. An icon hidden behind the notch is left alone.

with another app in front (Finder, say), after using the notch shelf opened from the menu bar icon, the shelf closed by itself as the pointer left it on the way to the icon, and the click then opened it again with the keyboard — what you typed next went to the shelf instead of the app you were in. A click on the icon right after the shelf closed that way now counts as that close. And a shelf closing now gives the keyboard back whenever its window has it.

closing the notch shelf's search with ×, or the shelf itself (after a drag out, from the menu bar icon, with Esc), left the keyboard with the shelf's hidden window — what you typed next was lost until you clicked Refto's window. The keyboard now goes back to Refto's main window when Refto is in front, else to the app you are in, without bringing Refto forward.

adding an item again to a board that has it said "They are on the board already".

with Refto in front, a card dragged out of the shelf opened from the menu bar icon closed the shelf as the pointer left it, and the drag (its picture) went with it. The shelf now stays open, as it was, until the drag ends — dropped or cancelled — and then closes if the pointer is away.

with words in the shelf's search, clicking Refto's main window (its search field, say) left the keyboard with the shelf, and what you typed went into the shelf's search. The click now moves the keyboard to the window clicked; the shelf stays open with its words and results, and clicking its field again goes on typing there.

opening it made the row of tabs 2 points taller, so the tabs and items moved a little. The field is now as tall as the tabs.

deleting one item said "Moved to the trash%!(EXTRA int=1)" (also "Already in the library", "Link copied", "Organized" and the shelf's locked item note for one).

the open shelf's window is larger than what it shows, and the transparent strip under it (and beside it) took the clicks meant for the windows below — a right-click on a board card there did nothing while the shelf was open. Now the shelf takes the mouse only with the pointer on it (or content dragged onto it); elsewhere clicks reach what is under it. With Refto's own Reduce motion on (Settings → General), not only the system's, the shelf opens and closes at once and its search field appears without growing out of the button.

the first time a newly installed Refto reads your Desktop, Documents or Downloads folder (or iCloud Drive, a volume), macOS asks whether it may; until you answer, an agent's add_items only saw its job running and some added the file again. Refto now looks first: while macOS is asking, the agent is told to ask you to answer the prompt; when macOS does not let Refto read the folder, the file is skipped (macos_permission_denied) with where to allow it (System Settings → Privacy & Security → Files and Folders). Anything still being added when the wait ends says how to wait longer, and the same request again returns the same job instead of adding the file twice.

with a video's page open, a link to a PDF (refto://item/…, also with ?page=) showed the video's last frame, still, in place of the PDF until a page in the column was clicked. Leaving a video or a sound for any other item (a PDF, a picture, a link, a document, another video) or closing its page now lets go of the view it played in, and the next item opens in a fresh one.

about one opening in six of a video the preview plays as it is (an MP4 of H.264 or HEVC) was taken as not playing when its first frame took more than 3 seconds, and the video was converted for the preview — and so every time after (a large video converts for a while, Preview copy at the sidebar's foot). Refto now converts a video for good only when it reports it cannot decode it; one slow to start gets 10 seconds, counted only while its player shows in a window in front, and then is converted that once. Videos marked so before play as they are again.

while Converting for the preview… showed, the player still showed the PDF opened before. The video's thumbnail shows until it plays.

going from a video that played to one the preview converts (an MKV, an AVI…) sometimes left the video before playing on under Converting for the preview…, and some 15 seconds later said This file cannot be played here. though the copy was made in under 3 seconds. The video before now stops as you go, and once the copy is made it plays; a video that has data but shows no picture (stuck at 0:00 with the one before's picture, or black while its time runs on, as a damaged MP4's does) is converted that once after 10 seconds, or says Can't preview: when it cannot be.

when the preview's copy of a damaged video fails, the page says Can't preview: and why, with Open with the default app and Show in Finder, instead of Converting for the preview… forever; the log has it, and opening it again does not start the same failing conversion over and over.

the temporary stream files of converted videos (*.live.mp4 in Refto's cache) are removed once the video is closed or its conversion fails, and leftovers after a while.

when an agent asks to change the library or to add files from outside your home folder while you are in another app, Refto's Dock icon now bounces until you come to it, the menu bar icon gets its dot, and the notification is posted — without taking the front from the app you are using. Coming to Refto shows the main window with the question (restored if minimized); once the last question is answered or times out, the bouncing stops (also when the question came just after you switched away: macOS can still count Refto as in front for a moment, and the bounce is asked again). Before, the question could wait unseen until the agent gave up.

its question did not show until another key was pressed (any dialog whose open state was worked out each time could stay invisible). An agent connected from more than one client window says how many (3 client windows) in its row — counting each client process once (ChatGPT keeps about ten Refto mcp running for one window: one, not “10 sessions”), and only those that sent something in the last 90 seconds: Refto mcp now ends its session when its client stops it (SIGTERM, SIGHUP) or quits without closing it (an orphan), which ChatGPT does each turn, so the row no longer counts 19 sessions for 4 conversations.

the pictures agents get from render_board and get_thumbnail (a contact sheet too) are JPEGs of at most 600 KB of base64, made smaller when they must be (imageScale then says how to map them back); an 820 KB PNG of a board reached ChatGPT's Codex as a megabyte of text and doubled a moodboard task's tokens (2.47 M). The copy written for you (file, markdown) stays a sharp PNG.

run_tool, upscale_image and erase_image on a tool not installed or turned off answer at once with why (tool_not_installed…) and what to do (how), instead of asking you to allow a change first.

tag:"dark mode", folder:"Fintech & Banking" and the like in an agent's query (MCP, the refto command, the local API) found nothing; a quoted value is now one value, as in Refto's search field. folder: also takes a folder's path (folder:"Mobile/Onboarding"), in the search field too. A search's total is how many it found (it was the page's), so more is right. An Excel workbook's text is read from its cells — inline text, numbers and every sheet (only its shared strings were, so many workbooks had none); those already in a library are read again.

the knob and the played part of the bar under a video or a sound (on its page, in the quick look, in a window of its own, on a board) went on in jumps — a 10-second video's about 13 points every 150 ms. They now move a little each frame of the display, the page itself built only once a second for the time shown (and the readings of the video's time, a little off, are caught up gently, never back); a long video's bar is painted only when it moves half a pixel, a paused one, one out of sight or a window in the Dock not at all. A sound's waveform fills as smoothly.

a video page saved to have its video downloaded (the local API’s video: true) became a new video item that kept only the link’s name, note, first folders and tags — its rating, the date added, its pin, its place in a stack and on boards, and whatever you did to the link while the video downloaded were lost. The video now takes all of it (a rejected tag stays rejected), and the link is gone instead of a second card of it waiting in the trash (one with annotations or highlights still goes to the trash). A video the library had keeps its own name and date, and gets the link’s folders, tags, the higher rating and its note.

making a library in a folder not named as one (~/Pictures/Refto itself, by a script) is now refused instead of turning that folder into a library; a new library’s name typed with .reflib (or .actlib) no longer gets a second extension (Mood.reflib.reflib). Libraries made before as .actlib open as ever.

under its subfolders' cards it said Nothing here yet; it now says Everything is in the subfolders and how many items they hold (each once), and that files dropped there go into the folder itself. With empty subfolders too it says so.

a WebM (and an MKV, Ogg, AVI… — what macOS does not read) is now cut, listened to for speech, transcribed and timed word by word through the built-in ffmpeg at once, instead of asking the engine first, which logged unsupported_media: Cannot read for each. A WebM without a sound track is recorded as having no sound, not as a failed speech detection and transcript, and a transcript of a WebM made from its panel now times its words too (they were left untimed). Videos marked failed before: choose Analyse again in their menu.

when an agent arranges a board by tag, colour or kind (MCP arrange_board), the sections' labels — Other, Untagged, the colours and the kinds — follow Refto's language; they were always Chinese. A tag named like one of them stays a section of its own. Asking to highlight, upscale or erase, an agent's first change now says which.

tags made only to suggest them on other items (a YouTube or Dribbble page's keywords: design, uidesign, ui design tutorial) are no longer suggested to every item that looks alike — a tag joins the AI's vocabulary once you take it. Suggestions also leave out the words your library's name and the item's folders already say (unless you tag things with them), one of each way of writing a name (ui design, uidesign), and a page's keyword that is a chosen tag plus words like “tutorial” or “for beginners”.

a tag the AI (or the file info: a website, a camera) made only to suggest it — a page's keyword, a site's name — is gone from the tag list once its suggestions are all rejected or dropped, and it is still not suggested to those items again. Tags you made, coloured, starred, grouped or took stay, even empty. Libraries from before are tidied once (only the tags that can be told made by a suggestion).

a post or page that writes colours as #1D2026, #fff or #A1B2C3 no longer gets them as hashtag tag suggestions; hashtags such as #design, #UI2026, #c4d or #cafe still count.

after zooming a PDF out (or in) with its controls, the next item no longer shows zoomed too — another PDF opens at its fit again, and an X post's card, a saved web page or a video at their own size, their text no longer small. Within the same item (another of its pages, its quick look and its page) the zoom stays.

in the inspector’s Add to folder, the search field had collapsed to a magnifier on top of the first folder; it now sits above the list, which scrolls under it. ↑↓ mark a folder, Return adds to it (a search marks its first match), very deep folders stay readable. The boards drop-down’s search had the same fault.

a pasted picture (⌘V, Paste, a thumbnail from the clipboard) no longer leaves a copy in the temporary folder, and what interrupted work leaves there goes after a day. The AI models compiled for this Mac (tens to hundreds of MB each, for AI Upscale and the AI eraser) go with their model, when it is updated, and after a month unused; a Live Photo's sounds keep to 64 MB; of the indexes saved before restoring a backup, the latest 3 stay. A video ffmpeg cannot read no longer holds its thumbnail, a frame made its cover or a Live Photo's sound for ever: a minute at most.

tagging or untagging 1,000 items — as an X list's full check does — takes under a second instead of several (their files are written several at a time).

(browser extension 0.6.4): each list's progress has its own place in the extension's storage, written every few pages instead of every page, all of them together kept under 6 MB (the list synced longest ago is forgotten first: its next sync reads it again, nothing saved twice); should the storage be full, the sync stops and says so.

a little after starting, as it drew the covers of the posts saved before — or a post's cover as it was saved, or a board's picture (export, Copy as picture, an agent's look at a board) — Refto could quit at once, the window drawing at the same time. Those pictures are now drawn between the window's frames, one at a time, and the window keeps moving smoothly meanwhile. Should Refto ever quit unexpectedly, its report is now kept in ~/Library/Logs/Refto/crash.log (with the version and the time), and the log notes it at the next start.

during an import, an X sync or the posts’ covers being drawn, the window is rebuilt at most 10 times a second instead of once per event (an import of 100 pictures: 133 rebuilds, now about 20); an X sync’s card at the sidebar’s foot shows how many it has seen without a moving bar, so the window is no longer redrawn every frame for the whole sync, and any bar that cannot say how far it is stops moving after 5 seconds. Checking a new file against the library for duplicates looks it up directly instead of reading every item (in 30 000 items about 4 ms each time, now well under 0.1 ms).

a library holding, in place of one of its small files (its settings, quick access, a page's details, a sidecar), a link to a device or a special file no longer makes Refto read without end as it opens — every start, since Refto opens the last library again — and such a file counts as missing; a file far larger than Refto writes is refused. A very long name — of an item, a tag, a folder, a smart folder, a board — no longer stops the window: names are cut to 1,000 characters (an item's) or 255 (the others'), line breaks become spaces, as they are saved and as they are read. The library menu no longer waits for a library on a disconnected network volume.

opening a post with a picture from the grid (double-click or Return) or pressing Space on it no longer blows its thumbnail up into one big picture that then turns into the post's card at once; the picture now flies to its own place in the card, and the card — avatar, text, counts — fades in around it before the picture hands over seamlessly (a card without the picture, or with it in another shape, cross-fades with it once it lands). Closing the quick look fades the card out with the panel instead of dropping it. In the quick look a picture grows from its cell keeping its shape and place all the way, rather than sitting small in the panel's corner and jumping to size as it lands; pictures, video posters and web cards in it have round corners concentric with the panel's (its radius less its padding), going from the thumbnail's corners on the way. With Reduce motion everything just fades.

a video, a sound or a page's media now stops once nothing on screen shows it. Before, it played on with no way to tell from where: in the column beside the grid once no item was chosen (and under the quick look, which played the item too), on a board as a presentation started or another board opened, in an SWF movie or a plugin's preview after its item's page was left, in a post's card behind its live page loading, and anywhere in the main window once it was closed into the Dock. Minimizing a window or closing the main window pauses what plays in it, as Photos does; hiding Refto (⌘H) or switching to another app does not. Back, it shows paused where it was.

(extension 0.6.1): its settings, popup, menus and notes, and what it shows on pages — the drop zone, the site buttons, X's Sync to Refto, its question, progress and notes — are now all in the language of Refto's interface (Refto's local API tells it), where the settings could be in Chinese and X's sync in English. The extension's settings can choose Chinese or English instead (Connection → Language). Until it has reached Refto once, the browser's interface language.

no longer stutters as it opens and closes, in a large window too: as the panel flew, the clipping of what is in it was drawn again on the CPU every frame (9 ms a frame in a 1280×800 window on a MacBook Air, up to 30 ms in a large one, longer than a refresh); it now takes about 2 ms. Space after a pause no longer uploads the grid's thumbnails again before the first frame, and a large picture decoded while the panel flies shows once it lands. In Low Power Mode the window behind darkens without blurring. Space held down opens it once, as Finder's, instead of opening and closing it with the key's repeat.

closing it (Space, Esc, a click beside it, Return to the item's page) or an arrow to another item stops a video's or a sound's audio at once; it could play on after the quick look closed.

content dragged onto what was just saved (the bar under the notch, or the island) shows Release to collect at once, as on the notch; it stayed the last save's bar before. Dragged away with nothing let go of, the bar comes back; dropped, the new save has a bar of its own. The bar has a Back button at its right end: the open shelf as it is — Recently Collected with what came in first and ringed, the tabs, pin — to drag it out or look at it right away.

the item turned to shows from its top. The gesture that turned the page — scrolling an X post's card to its end and on — glided on after the fingers lifted and scrolled the next post's card (or its screenshot, snapshot, reading view, live page, a PDF or a document) half way down as it showed; a text turned to took that glide too, and the end of the drag as it slid under the fingers. The glide now stops with the turn; the next gesture scrolls the new item as ever.

pinching in on a PDF's page goes back to the grid again, as on a card (its pinch was lost the same way). A PDF fitting its view wide — a slide, say — no longer shrinks instead each time, and one zoomed in (by pinching, a two-finger double tap or its zoom buttons) pinches back to its fit and stops there; pinch once more to go back.

on an X or GitHub post's card (and its screenshot, snapshot and reading views): the page no longer shrinks under the fingers and springs back as they lift unless pinched very small — the rest of the pinch was lost once the page's picture stood in for it. Everywhere (pictures too) the page now closes under 85 % of its size, or under 94 % when the fingers are still closing fast as they lift; opening them again as they lift keeps it open.

no longer looks as if accepting did nothing: "Accept N tags" says how many suggestions are left and why (less confident than the threshold, or not tags); an item answered no longer comes back for a moment while the AI keeps adding suggestions; the sidebar's count only counts what the page lists (a suggested name the item has already, or a locked item, kept it from reaching zero).

finds one picture saved again: bigger or smaller, sharper or blurrier, recompressed or in another format, on a wider canvas or with a margin, or cut a little (a site's thumbnail beside the full picture, a design shot with and without its background). Such a group says Same picture; pages of one app that only share their frame are not taken for one. Pictures are compared by themselves in the background as they come in (and those already in a library once, after the update), with their progress at the sidebar's foot and on the page — no need to click Check existing items first.

suggests which one to keep and why — the most pixels, then the largest file, then the first collected — and shows where the items of a group differ (format, dimensions, size), the best of each in the accent colour.

stays quick with a picture saved thousands of times: comparing 4,000 copies of one screenshot took 17 minutes and wrote 8 million rows into the library; it now takes about a second and one row a copy. Such copies are one group (its card shows 24 of them and says how many more), and the page with thousands of groups scrolls as smoothly as with a few. Comparing pictures in the background takes a pause after half a minute of work and goes on by itself.

no longer slow down when thousands of files arrive at once (an archive unpacked in Downloads): the notch lists the first 500 and says the others were left where they are; files moved away one by one no longer rewrite the list each time.

the grid, search and the sidebar no longer wait behind the AI reading every item in the background (30,000 items: the grid's first page came 0.2 s late each time a library opened or a batch of analysis ended); the item page and the Tools page no longer make garbage on every frame (1–2 MB a frame while a video plays or the page scrolls); the sidebar and Settings → Tags with thousands of tags, To review while the AI keeps adding suggestions, and the grid during a large import do much less work each time.

comparing a library's pictures for Duplicates the first time takes half the memory (30,000 items: 2.1 GB of work, now 0.9 GB). It, the posts' covers drawn in the background, the background capture of older links and X's Full check now read on the background's own connection, so the grid and search no longer wait behind them. The background capture reads its list of links once a round instead of before each link (3,000 links: about 50 ms and 10 MB each time, now well under a millisecond); the posts' covers read each saved page's details once instead of twice; trying again what the AI gave up no longer holds up saving while it looks for it.

the web log and yt-dlp's no longer write the part of a link after its path (a 小红书 link's xsec_token, a signed media link). At most 50 browsers stay paired with the local API (the one unseen longest goes first), and the sidebar's foot shows at most 8 list syncs.

a page's HTML is parsed up to 8 MB and 250,000 nodes (the browser extension's "save the whole page" and the rendered page could send any size: 8 MB of tags took 1.5 GB of memory, now about 100 MB); a capture stops after 5 minutes and its snapshot's resources after 2 minutes, 300 at most, 4 at a time from one site (a page of pictures that never answer held a network slot for hours); a browser's screenshot is read up to 25 megapixels, a library's web.json up to 4 MB.

older links get their covers in the background, and posts their drawn covers, only in libraries made or trusted on this Mac (nothing is written into another's library, nor its avatars decoded, until you trust it); Settings → Collect says when a library is paused and can trust it.

on the Mac, a picture whose header says more than 64 megapixels (a PNG of 1 MB can say 16000×16000: a gigabyte decoded) is no longer handed to ImageIO for its thumbnail, a thumbnail shown in the grid, an avatar or a post's picture: its thumbnail comes from Quick Look, outside Refto, or none shows. A picture of more than 268 megapixels is not decoded whole, as it was not by Refto's other decoders.

requests that send something are served 16 at a time, large bodies 2 at a time (others wait, then get 429), each token 6,000 requests a minute, 64 connections; a request's body must arrive within 5 minutes. A note is at most 64 KB everywhere (a 60 MB note took 200 MB of disk); a request has at most 200 tags and folders.

the shelf's cards (whatever library it shows), a saved post's pictures in its Screenshot view and on its card while the page loads, a board's thumbnails, and MCP's get_thumbnail and contact sheet (a version's file too) refuse a picture whose header says more than 64 megapixels, as the grid does, instead of decoding up to 268; agents' thumbnails are decoded small on the Mac rather than whole.

a plugin's page can no longer save a file into Downloads (a download link, a blob) or go to another web site in its panel, window or preview; it moves between its own pages only, opens no window, and a web link clicked in it opens in your browser when the plugin may open links (its open-url permission). The format extensions' pages (ePub, SWF, EPS) and a board's PDF follow the item page's rules too: links out open in the browser, nothing is downloaded.

a format extension's component, the yt-dlp release and an .actplugin are all unpacked the same careful way — no file outside the folder, no link, nothing overwritten, at most 20,000 entries and a size counted as it is written (a zip that claims to be small and unpacks to gigabytes stops at the limit). Paths are checked the same way everywhere: moving Refto's data folder into a folder of it whose name starts with ".." is refused like any other folder of it; a built-in tool's export folder is checked once .. is resolved; removing a plugin named "." no longer touches the other plugins. Names Refto makes for files (a board's export, a dropped text, a library, a backup) no longer start with a dot (a hidden file) nor hold control characters.

the released app no longer takes ACT_ENGINE, ACT_FFMPEG, ACT_RECORDER, ACT_CJXL, ACT_OXIPNG, ACT_LIBWEBP or REFTO_API_TOKEN from its environment (a program of yours could set one with launchctl setenv, and what Refto started in place of its engine, ffmpeg or recorder would have had Refto's screen recording, microphone and folder permissions, or the local API's token would have been known beforehand). It says in its log that it ignored one; development builds and the scripted checks still take them, and refto mcp still takes REFTO_API_TOKEN.

an EPS file with a preview inside (a TIFF, Photoshop's image or thumbnail, an EPSI preview) is shown by it, and no PostScript is run; only one without is drawn by the Ghostscript you installed — before, Ghostscript drew every EPS file when it was there, its -dSAFER the only guard, and that has been got round more than once (CVE-2024-29510 was used with EPS files). Ghostscript now runs in a sandbox that lets it read only the file, itself and the system's fonts and libraries, write only a new temporary folder, and reach no network and start no other program — no sandbox, no drawing; never older than 10.03.1 (Settings → Tools → EPS says when yours is, and to run brew upgrade ghostscript, as does the file's page); on a page of a size Refto sets (its long side at most 4096 pixels, whatever the file asks for), one page, within 30 seconds.