What's new

What changed in each version of Readfine, newest first. This server runs version 0.22.0.

Version 0.22.0 · 2026-10-18

Changed

  • After an OPML import, feeds that could not be added (or were added without a check) are listed under Feeds to check with their name, and one click selects the address for copying into Add feed. When such a feed was part of a filter, saved search or catch-up, the warning now names that item.
  • After you save new relevance terms in Settings → Relevance, or import them with an OPML file, the last 7 days are rescored within seconds instead of up to 5 minutes later.
  • Labels and single-label views open faster on a large instance, most of all when nothing in them is unread. Their speed now depends on how many articles you have labeled, not on how many articles the whole instance holds.

Fixed

  • Importing a large OPML file could end on a "504 Gateway Timeout" page, even though the import itself went through. New feeds are now checked several at a time, and the result page always comes back. Feeds that could not be checked in time are added anyway and fetched in the background; the result says how many.
  • An OPML import with several Reddit feeds dropped some of them with "Too many requests (429)". A feed whose site is rate-limiting the server or briefly down is now added anyway and fetched later.
  • On a new instance, relevance scores show up seconds after the first articles arrive instead of up to 20 minutes later, and they also match words that so far appear in just one article. Until then, a note in the reader says they are on the way.
  • If you skipped the topics question at signup and set up terms later in Settings → Relevance, the scores were there but the article list did not show them. Show the score in the article list now starts checked until you save your first terms.
  • Catch me up over a long period could end on a "504 Gateway Timeout" page, while the AI finished in the background and the tokens were spent anyway. The briefing is now generated in the background and shows up on the page when it is ready.
  • After a change of SECRET_KEY, or with a second Readfine on the same host (the demo next to a real install, for example), every action failed with "403" until you cleared your cookies. The page now picks up a fresh security token by itself.
  • After a password change on another device, or with an old login from a second Readfine on the same host, the browser could get stuck bouncing between the reader and the login page ("too many redirects"). The old login is now dropped and you land on the sign-in form.
  • Creating an API token in Settings → API Tokens failed with "403". It works again.
  • In the dark theme, an unstarred star and empty values (such as zeros in the admin tables) were brighter than the text around them. They are now dimmed as in the light theme, and the dormant and inactive badges in Admin → Users stand out better.

Version 0.21.0 · 2026-10-11

Added

  • A bookmarklet for saving the page you are on. Drag Save to Readfine from Settings → Preferences to your bookmarks bar, click it on any page and press Save. The article lands in Saved, like a link pasted there, and Read it now opens it with the Saved list next to it (from the Android share sheet too).

Changed

  • Claude Sonnet 5.5 now runs with thinking off, like the other models, and can be used for scoring. Before, it thought on every request, which cost more and was slower. The AI cost table knows the newest Claude, OpenAI and Gemini models (Claude Haiku, Sonnet and Opus 5.5, GPT-6 Sol and Luna, among others), so their costs are no longer estimates.
  • Briefings go only to your own address, unless the admin allows other addresses in Admin → Settings. A test briefing always goes to your own address, and an account can have at most 5 briefings with other recipients (admins have no limit).
  • The regex tips in the filter editor are now folded under a Regex tips link instead of always taking up space below the conditions.
  • New rate limits: refreshing a feed by hand and adding a scrape feed (10 a minute), testing and applying filters, the term suggestions in Settings → Relevance, and OPML imports (10 an hour). A message in the article chat is limited to 2,000 characters, like the general chat.
  • Search finds Thai, Lao, Khmer and Burmese words inside running text, as it does for Chinese, Japanese and Korean. The basic relevance score doesn't handle these four languages yet.
  • New screenshots, each in a light and a dark version, in landing.example.html and the README. The install dialog of the app (PWA) shows them too.

Fixed

  • A slow feed server could stop all feed fetching until a restart and make logins hang, and a slow custom AI endpoint could hold up scores and summaries for every account. Fetches and model calls now have an overall time limit (for AI: 2 minutes for a score, 5 for a summary or chat, 10 for a digest) and are retried later. Sending mail no longer holds up the app either.
  • Mail over port 465 (implicit TLS) hung on every send until it timed out. It works now. The TLS box in Admin → Settings → SMTP is for STARTTLS on other ports such as 587.
  • An install reached by IP address over plain HTTP could not log in. Set the new SESSION_COOKIE_SECURE=false (setup.sh does it for IP installs), on a home network only.
  • Some sidebar and title bar counts didn't match their list, because they counted articles retention had removed or a grouped story as several rows.
  • A wrong API key or model name now says so everywhere the AI is used, instead of "try again" or the provider's raw response.
  • Labels stayed on articles you had lost access to (after unsubscribing, or unstarring or unsaving an article with no feed). The article sat in the label's list but could not be opened or marked read. Deleting a label also left its "add label" action in your filters, and saving such a filter failed. Both now go with the access or the label, a filter left with no action is turned off, and the upgrade removes the leftovers.
  • Feeds in an encoding such as windows-1250, named only inside the feed, showed garbled accented letters.
  • One failed feed in an OPML import made every feed after it fail too. Each feed now fails on its own, with the reason in plain words, and a rejected import shows a message instead of replacing the page.
  • A scrape feed with an invalid CSS selector failed on every fetch with "Internal error". Such a selector is now refused when you add or edit the feed. Regenerate in the AI selector now uses your note about the previous attempt.
  • A deactivated account kept sending its briefings, spending its AI key and serving its shared article links. All of it now stops until the account is reactivated.
  • Deleting an account now also deletes the articles only that account kept. The per-feed article cap (purge_keep_count) no longer keeps fewer articles than set.
  • Input the app does not accept (a name that is too long, only spaces, or already taken) ended in a server error page. The app and the API now say what was wrong, the API with 400 or 409. Adding a feed through the API whose address cannot be fetched now answers 400 with the reason too.
  • The basic relevance score could not read Hindi, Bengali, Tamil and other Indic text. Their vowel signs were dropped like accents, which left words in pieces. The term table is rebuilt on its own after the upgrade. In search, word* now also works for words that end in a vowel sign, and reading time for these articles is no longer two or three times too long.
  • PATCH /api/v1/articles/{id} now returns the same fields as a GET, and a search with a very long query no longer ends in a server error.
  • Smaller fixes: articles scrolled past just before closing the tab could stay unread, a labelled article could miss its AI score, a feed with an item dated far ahead showed its last article in the future, and renaming or deleting a saved Catch me up configuration could replace the list with an error. Clicking a label in the sidebar no longer makes its count jump left. The Saved count now goes up as soon as you save a link. In dark mode, the "dormant" date and "unverified" badges in Admin → Users are readable again. Removing a feed or folder now also warns about inactive filters scoped only to it. An article opened from a link (in Stats, for example) on a phone or in the 2-panel layout now closes back to the list, where Back used to leave the app.

Removed

  • The API's feed response no longer has favicon_url. It was always null.

Security

  • A feed could hide page code in an article's address, and it ran in the reader's session when they opened the article or its shared link. Addresses are now encoded, and the upgrade cleans articles stored before.
  • A feed server that answered with a bare address made Readfine fetch it, past every check that keeps fetches off private networks. Anyone with an account could reach services inside your network. A feed's response is now only read as content, and every address outside the public internet is refused (including 100.64.0.0/10, used by Tailscale).
  • A private feed's username and password were sent to every site its articles linked to when full-text extraction fetched them. They now go only to the feed's own host. Saving a link in Saved could also give you an article another account had fetched from a private feed.
  • API tokens also worked on every web page, Admin included. They now work on /api/v1 only, so scripts that called web pages with a token have to move to the API.
  • A filter with a badly written regex could freeze the instance for minutes. Such a filter is now switched off at once and marked "regex too slow" in Settings → Filters. Fix the pattern and turn it back on.
  • AI output no longer shows images. An article could make the AI add one, and loading it told the article's author your IP address and text from your other articles.
  • A custom AI endpoint could send a response large enough to run the instance out of memory. Responses are now capped at MAX_FETCH_BYTES.
  • On a shared computer, saved settings pages, the chat and the last open view stayed in the browser after logout, and the next account got them back. Settings pages are no longer stored, and the rest is restored only for its own account.
  • A file full of dead addresses could keep one OPML import running for hours. An import now tries about twice as many feeds as the account has room for (2,000 for admins) and skips feeds you already follow without fetching them.
  • Smaller fixes: the Content Security Policy no longer allows unsafe-eval. Show AI prompt in the scrape setup could run a script from the scraped site's error page. The article list showed the last error, address included, of a feed you don't follow. Any subscriber of a shared private feed could change its fetch interval or scrape selector. One wrong password right after a lockout ended locked the account again.

Version 0.20.0 · 2026-10-04

Added

  • Saved searches. Save in the header of search results keeps the search, with its filters, under a name in the sidebar above Labels, with a count like the other entries. A saved search is not a snapshot: it runs again each time you open it, and it reads like a feed, marked read as you scroll, with Mark all as read covering only what it lists. Without a status of its own, it follows your unread setting. Edit reopens the search to update it, save it as a new one, rename it or delete it. A new account that lists its topics on the welcome screen starts with one, Top picks: the last week's articles scoring 60 or more, best first. In the API, GET /api/v1/saved-searches lists them and GET /api/v1/articles?view_id= returns what one lists.
  • Keyboard shortcuts in the reader: j/k for the next and previous article, o or Enter to open or close one, Space to read on into the next one, Shift+J/Shift+K for the next and previous entry in the sidebar, m read or unread, s star, e archive, t labels, v the original in a new tab, r to reload the list, Shift+A to mark the view read, g a and g s for All articles and Starred. ? lists them, and so does the help page. They follow Inoreader and Miniflux, and the TT-RSS keys n/p, u and s work as well.
  • The OPML export in Settings → OPML now takes along more of your setup: preferences beyond the timezone, relevance terms and your AI texts (interest profile, summary and context prompts), saved searches, catch-ups with their briefing schedules, and each feed's own settings such as full text and retention. Tick what goes in; with only Feeds & folders ticked you get plain OPML. On import, briefings arrive switched off, and a label that a filter, saved search or catch-up uses is created if you do not have it. AI keys and feed passwords are never exported.
  • The welcome screen for a new account has a second step: where your articles should come from. You can start with a few well-known English feeds on topics like technology, science, world news, food and sport, each topic in its own folder, import an OPML file, or add feeds yourself. Topics that match what you typed on the first step are checked for you. Self-hosters can offer their own list or turn it off, see "Starter feeds for new accounts" in the README.
  • What's new at the bottom of the user menu shows which version of Readfine the server runs and opens the release notes of that version, at /changelog. The page is public and linked in the footer of the landing, help and legal pages. Upgrade notes in it are shown to admins only.
  • Admins can pause fetching for accounts nobody uses: Pause feeds after N days without activity in Admin → Settings → Users (off by default, at least 30 days). A feed keeps being fetched while any of its readers is active, so a shared feed keeps going. With mail set up, a warning goes out 7 days before. A reader who comes back gets the feeds going again within about 15 minutes and a one-time note that articles may be missing. Admins, accounts with a briefing and accounts without feeds are never paused, and paused readers get no AI scoring or summaries. Admin → Users and Admin → Feeds mark such accounts and feeds as dormant.
  • With public page counting on, a new account keeps where it came from: the domain of the site that linked to the landing or registration page, a utm_source tag, invite, internal (another page of this site) or direct (no referrer). Admin → Traffic breaks down the accounts created by source under the registration funnel. Nothing else is stored, and switching counting off clears it from every account. The privacy page says so while counting is on.
  • In Settings → Relevance, the table of how your terms did now says "not counted yet" for a term whose words are too rare in the last month's articles to score, instead of "no match". Such a term may be misspelled, or new to your feeds: words from a feed added today are counted overnight.
  • With PUBLIC_URL set, the footer of a briefing email links to your instance and to Catch me up & Briefings, where the briefing is managed. The email sent when a briefing fails links there as well. Without it both stay as they were.

Changed

  • Opening an article that is still waiting for its full text now fetches it right away. Full text used to come only in queue order for the whole instance, about 20 articles a minute, so after a large import or new feeds the article you were reading could show only the feed's excerpt for a long time. An article whose earlier attempt failed keeps its retry schedule and its Retry button.
  • Adding a feed that someone else on the server already reads no longer gives you its whole stored history as unread, which could be hundreds of articles for a busy site. You get the last week unread, and always at least the newest ten, so a blog that posts once a month still shows its latest posts. Older articles are marked read, so you can still find them in the feed and in search, and they do not count as your reading anywhere. The unread ones now get a relevance score right away instead of the next night.
  • A score condition in a filter now reads like the one in the search: at least or below a number, with at least as the default. It compares the whole number the list shows, so at least 75 takes in every article shown as 75 or more. Existing conditions are converted (more than 74 becomes at least 75), so only an article right on the line can come out differently. Equals, which never matched a score, is left as it is with a note in the editor. In the API, score conditions take gte and lt; gt and equals still work.
  • Suggested terms in Settings → Relevance are more often about what you actually read. A word now ranks higher the more often it shows up in the articles you read compared with everything your feeds bring in. Common words like likely or humanity used to make the list just because they are frequent in longer articles.
  • The Mark all as read button in the sidebar only shows up on hover or long press where there is something unread. On a feed with nothing unread, the refresh button moves into its place.
  • A feed whose readers have all been deactivated by an admin is no longer fetched. It used to keep going as long as it had any subscriber.
  • Active on the admin dashboard's Users card now means accounts that read at least one article in the last 7 days (hover for the last 30). It used to count accounts that were not disabled. Admin → Users shows the same number next to the heading. It counts reading, not logins, and leaves out articles marked read by a filter or by duplicate detection.
  • The user table in Admin → Users shows more of how each account uses relevance. Terms counts a user's relevance terms, grayed out while they are still the list from signup or basic scoring is off. Scored 7d counts articles scored by AI in the last week, and Filters adds in brackets how many filters use a score. Hovering AI 7d breaks the number down by operation, and catch-up runs now count towards it. The Role column is gone; the admin gets a small badge next to their name instead.
  • Opening a link into Readfine while signed out, a bookmark or a link in an email, now takes you to that page once you sign in. It used to land on the article list. The warning before feeds are paused uses this to link straight to Settings → Profile, where the account can be deleted.
  • The user picker in Admin → Scoring eval lists only accounts with scored articles in the selected window, most first. Next to each name it shows how many were scored and how many of those engaged.

Fixed

  • Email addresses no longer depend on letter case. Someone who signed up as [email protected] can log in as [email protected], and the same address can no longer register a second account by changing the capitals. Spaces around the address are ignored too.
  • Search finds Chinese, Japanese and Korean words inside running text. Search split text only at spaces and punctuation, so in these languages it took a whole sentence for one word, and a word inside it was never found. A word now matches wherever its characters stand together, in the title, summary or full text. A search for a single character looks at titles only.
  • The full text of some pages lost lines written as plain text blocks beside ordinary paragraphs. Bon Appétit recipes, for one, kept only the quantities of the ingredients (3, ¼, 5); now each ingredient reads as one line with its amount. Full text also comes out right on more sites: on some, Barron's among them, extraction used to return a print notice instead of the article, and on others a related-story promo ended up in the text. Articles fetched before the fix keep the text they have.
  • Article titles from some feeds, The Verge and other WordPress sites among them, showed HTML codes instead of characters, for example Can an ‘eSUV’ e-bike. Such titles are now stored as plain text, and titles already stored are fixed when you upgrade.
  • The sidebar and the article list load faster on a large instance. Opening a label, feed or folder with nothing unread left went through every article on the instance to find that out, and the counts next to Labels, Starred, Archived and Saved did much the same. Both now look only at your own articles. Switching to another view also no longer reloads the whole sidebar each time, only when it is more than a minute old.
  • On a phone, an article opened full screen (and one opened from Same story) has its ··· menu back, so you can generate a summary or context, mark it unread or fetch the full text there too. Star, archive and share are left out of it, since the bar at the top already has them.
  • A search with a score condition (for example, relevance at least 60) is much faster. It used to go through every article on the instance, not just yours, and on a large instance it took over a tenth of a second each time.
  • Search results sorted by relevance or score no longer skip articles further down while you read some of them with the status filter on Unread.
  • In Saved, an article scrolled up under the header at the top of the list is marked read as it goes out of sight. It used to count as read only once it had passed the header too, so one left under it stayed unread.
  • Articles that came in before a word of your relevance terms was counted (for example, from a feed added the same day) are scored again after the nightly recount. They used to keep a score of 0 until you saved your terms.
  • When two of your feeds carry the same article under the same link, the article list no longer counts the copy as another source covering the story. The copy was already marked read as a duplicate, but the row still said "1 other source". It is still listed under Same story at the foot of the article, marked same link.
  • Filters imported from an OPML export keep their order, which matters when one stops the rest. They all used to land at the same position and ran alphabetically. A filter that adds a label named with digits only, like 2024, no longer gets skipped on import, and a filter limited to feeds or folders you do not have comes in switched off instead of applying to every feed.
  • Last active in Admin → Users stays current for someone who reads in an installed app or a tab they leave open. It used to change only when the app was loaded or on login, so such a reader looked inactive, and after a month without a reload their automatic reading profile stopped being updated. Opening an article, reading by scrolling and changing an article over the API now count too.
  • The help, features and legal pages share one header: the Readfine name at the top leads to the home page, with Back to reader beside it when you are signed in. Help and features had nothing to take a signed-out visitor back, and the name on the legal pages led to the sign-in form.
  • When the verification email cannot be sent (the mail provider is down or over its daily limit), at sign-up or with Resend verification email, the page now says so and asks to try again in a few minutes. It used to say the email was sent, so a new account waited for a message that never came.
  • Checking a password no longer holds up everyone else on the instance. It is slow on purpose and used to block all other requests while it ran: thirty sign-ups at once froze the app for several seconds, and so, for a moment, did every password change, email change or account deletion.
  • On the registration page, Firefox and password managers no longer offer to fill your login into Display name. The email field is now marked as the account name.

Security

  • Signing in through the API (POST /api/v1/auth/token) now follows the same rules as the sign-in page. An account whose email is not verified gets no token, and failed attempts count toward the 15-minute lockout after repeated wrong passwords. The API used to skip both, so someone who signed up with another person's address could use the API before it was verified, and passwords could be guessed there past the lockout.
  • A password reset link stops working once an admin disables the account. A link mailed before that still let someone set a new password on the disabled account.
  • An invitation can only be used once even when two people sign up with it at the same moment. Both sign-ups used to go through.

Version 0.19.0 · 2026-09-26

Added

  • Articles are scored for relevance without an API key. Basic relevance matches the title and opening of each new article against a list of terms you keep in Settings → Relevance. A term can be several words, and related forms match too (cycling finds cyclist). Until now only AI scored articles, and only those a filter had labelled. The page suggests terms to add and drop from what you read in the last 30 days, and shows how well each term picked what you read. It matches words rather than meaning, so it is the weaker scorer, but it ranks clearly better than date order. Saving the list scores the last week of unread articles. A new account is asked for a few topics after signing up, and an account with no terms sees a bar above the article list that leads to the setup.

  • The score in the article list, and now also next to the other sources of a story under an article, shows whichever scorer has an opinion, AI first. A basic score is marked with ~, because a word-match 70 is rare where an AI 70 is common. The end of an open article shows both scores side by side. Settings → Stats, Catch me up, briefings and the admin scoring eval handle both. A filter's score condition now picks its source: AI, Basic, or AI, else basic (the number the list shows). Basic filters run as an article arrives, AI ones after AI scoring, and AI, else basic waits for the AI score only when one is on its way. Existing score filters keep reading the AI score. In the API, articles carry lexical_score next to ai_score, and a filter condition can use the fields basic_score and relevance_score (AI, else basic).

  • Search can filter by score (at least or below a number, from any of the three sources) and sort the results by it, and the results then show that source's number. It can also be limited to recent articles (the last 24 hours, 3 days, 7 days, 30 days or 12 months) and to your starred, saved or archived articles. The status can go by what you actually read (30 seconds on it or its original opened, as in Stats) rather than the read mark, and the results header says how many it found. The old "Relevance" sort there is now called "Search relevance".

Changed

  • Search ignores accents, so zpravy finds zprávy and the other way round, and English words match in their other forms (votes finds voting). A word ending in * finds every word it begins (cycl* finds cycling and cyclist). Sorting by search relevance puts articles with the words in the title above those that only mention them in the text, and the results mark the searched words in titles and snippets.
  • On a phone, the shortcut in the top and bottom bar switches between All articles and Starred if you have no labelled articles. Until now it always offered Labels, which for such an account was an empty list.
  • The admin scoring eval counts an article as engaged the way Stats and the Relevance page count it as read: half a minute on it, an opened link, or a star at any point. It used to need a full minute and a star still in place, so its engagement rate came out lower than the reader's own figures.
  • The admin scoring eval greys out an AUC measured on fewer than 10 engaged articles and says there are too few to judge yet. On a new instance two lucky articles could show up as a green 0.93.

Fixed

  • Editing a filter with an AI score condition while AI scoring was off showed the condition as "title or content", and saving turned it into one. The condition now keeps its source, marked as off, and the filter list flags it.
  • In dark mode, the regex tips in the filter editor and the "try web scraping" link after a feed test were dark blue on a dark background and hard to read.
  • A search with no words, only filters (such as status or labels), now finds the same articles as one with words: it includes what you starred, archived or saved by URL, even from feeds you no longer follow. And searching the "no folder" group no longer lists those kept articles as if they were in it.
  • Chinese and Japanese articles were counted as a handful of words, because those languages put no spaces between words. Reading time showed 1 minute for every article, and a feed that delivers full articles in these languages was not recognised as one, so readable extraction ran on every article anyway. Two characters now count as a word.
  • On a phone, the search window now stays above the on-screen keyboard, so you can scroll to the fields at the bottom and edit them without closing the keyboard first.
  • iDNES articles were again stored as the site's consent notice ("iDNES a reklama"). The consent page now names the article it interrupted as its own address, which made it look like the article. Such a page is recognised as a wall again, and the article falls back to the text the feed delivers.
  • Filter text conditions treat width variants as the same text: full-width AI matches ai, half-width katakana matches the normal form, and Korean text stored in decomposed form matches as typed.

Version 0.18.0 · 2026-09-21

Added

  • A feed's address can be changed. Until now it was the one thing about a feed that was fixed for good, so a feed that moved without leaving a redirect, or one whose address needed a small fix, meant unsubscribing and subscribing again, which throws away every article of that feed along with what you had read, starred and saved. The address is now a field in Settings → Feeds → the feed, and changing it keeps everything: only where the articles are fetched from moves. The new address is fetched once before it is saved, so a typo comes back as a message on the field rather than as a feed that quietly stops working, and if it redirects, the address it redirects to is what gets stored. Anything the old address left behind, an error, a counter, a wait imposed after a refusal, is cleared at the same time, and the feed is picked up on the next round rather than at its usual interval. A username and password written into the address are moved into the feed's credential fields, as they are when subscribing.

  • Folders can be kept in an order of your choosing instead of alphabetical. Above the list in Settings → Feeds there is now a Folder order switch: left on Alphabetical everything works as it did, turned to Custom, arrows appear beside each folder and move it up or down one place. The order you set is the order everywhere folders are listed: the sidebar, the folder box on the subscribe form and on a feed's own page, the feed and folder pickers in filters, the API, and an OPML export. Custom starts from the alphabetical order in front of you, so switching to it changes nothing on screen and you rearrange from there. Switching between the two never overwrites what you arranged; Reset to alphabetical, next to the switch, is the only thing that throws it away. A new folder goes to the end.

  • Administrators can see how much traffic the public pages get, without adding an analytics service. Switch Count public page visits on in Admin → Settings → Traffic and a Traffic page appears in the admin menu, with visitors, page views and identified bots over 7, 30, 90 or 365 days against the period before, a chart by day, a breakdown by page and by where the visits came from, and a registration funnel. Only the pages anyone can reach are counted, and a request carrying a session is skipped, so these are visits from outside rather than your own users. Nothing is recorded per visit: no cookie, no IP address, no user agent, only counts per hour and per day, which is why the visitor numbers are an estimate given as a range. It is off until you turn it on, and counting starts from that moment. Telling one visitor from another needs the real client address, so if anything proxies the app, the nginx in the Docker setup included, TRUSTED_PROXY_COUNT or TRUST_CLOUDFLARE has to be set, or the count quietly reads 1; the page says so when it is not.

  • Several feeds covering one event no longer fill the list with the same news. Readfine matches headlines across your feeds inside a 72 hour window, which catches the same story under different addresses and different wording. One article of each story stays in the list, with a line saying how many other sources covered it and how many you have read; that line opens the rest in place (in a list filtered by a label, a folder or a search, only what belongs in that list), and the same list sits at the foot of the article. Reading the row finishes the story: the articles behind it are marked read too, and marking the row unread brings them back. Unread counters over the views that fold count rows, so a folder no longer adds up to the sum of its feeds. Catch me up and briefings fold the same stories. How far this goes is set in Settings → Preferences. Folding is the default; Fold, and don't show a story I've read again also keeps a repeat of a story you read out of your unread list and digests, without deleting it. Off leaves the list alone.

  • Settings → Stats says how much actually came at you. Intake is a funnel over 30 days: everything fetched, then what never reached you (a repeat of a story you had read, an article a filter marked read, a link that had already arrived in another feed), then what story grouping folded under somebody else's row, and at the bottom the rows you really had to go past. Those five add up to the fetched count, so nothing is counted twice, and under them sit how much of it you read and starred. If you use labels, every line gets a second number for the labelled part alone, which is a different funnel rather than a smaller one: a list filtered to a label folds only the articles carrying it. Below that, Score spread cuts everything fetched into four bands (75 and up, 50 to 74, 26 to 49, 25 and below) plus what has not been scored.

  • The address of a shared feed is the administrator's to change, in Admin → Feeds → Edit. On an instance where several people follow one feed, that feed is a single row fetched once for everyone, so its address is not one subscriber's to rewrite; the form says so instead of offering a field that would not work. An administrator gets the field for any feed, with a note of how many people are subscribed, and the change is written to the audit log with both addresses. There is also a checkbox to save an address without fetching it first, for fixing a feed whose host is down at that moment. The current address is never put into the admin form (it can carry an API key), only shown as a redacted hint, so leaving the field empty keeps the feed where it is.

Changed

  • Missed gems in Settings → Stats counts stories rather than articles. Read any article of a story and the rest of its coverage is no longer listed as something you missed, and a story nobody read takes one row instead of one per source, the highest scoring one. Ten rows are now ten pieces of news you can still catch up on.

  • Search results are no longer marked as read when you scroll past them, even with "Mark as read on scroll" on. Searching is looking something up, not reading: you scan the results for the one article you came for, and the rest should still be there, unread, when you go back to your list. This covers everything the search panel returns, including the filter-only views (scope, labels, status) that submit without a search term.

  • The cost figures under AI are priced off this month's price lists. The catalogue now knows Claude Fable 5.1, GPT-6 Astra and the GPT-5.6 models (Sol, Terra, Luna), and Gemini 3.7 and 3.8 Flash. Gemini's current Flash models are running at half price until the end of the year, which is what they are now costed at rather than the standard rate. A model Readfine does not know is still estimated from a mid-range model of that provider and marked with a "~". GPT-6 also gets the same room to reason as the GPT-5 models, so a scoring call to it no longer risks coming back empty.

  • Creating a folder through the API without naming a position now puts it at the end of your folder order rather than at the front. Folder positions are also renumbered by the upgrade to match the alphabetical order they were displayed in, and again by every move, so positions set through the API before this release are not kept. Nothing in Readfine displayed those positions until now, so nothing you could see changes.

  • The sidebar draws faster. Each of its counters ran its own aggregate over the articles table with no index that fit, so every badge read every article on the instance rather than yours, and the cost grew with the size of the database instead of with the size of your list. Since the sidebar is redrawn whenever you read, star or archive something, that was felt on every click. A covering index keeps those queries out of the table: measured on 31 000 articles, the sidebar's time in the database fell from 330 ms to 100 ms, and the bigger the instance, the wider the gap.

Fixed

  • An OPML import that runs into the feed limit says so where you cannot miss it. Importing a file with more feeds than your account can hold stops at the limit, which is how it has always worked, but it said so in one line of grey small print under a green "Import complete" heading, and it never said how many feeds were left behind. Someone moving in from another reader could easily believe their whole list had arrived. The result now turns amber, says the import stopped at the limit, and gives the numbers: how many feeds the account holds and how many of the file's feeds did not make it. Importing the same file again after making room carries on rather than starting over, and it says that too.

  • A feed that comes back as something other than a feed says what came back. A site serving an error page where its feed should be, a WordPress one printing PHP warnings instead of XML, an address that turns out to point at an ordinary web page: all of them reported themselves as "junk after document element", which is the parser describing a document that was never there. The feed's error line now says the server returned a web page rather than a feed, or that the response was empty. The parser's own message is kept for the case it fits, a feed whose XML breaks part way through.

  • Something that fails now says so, instead of leaving you to guess. When a request to the server went wrong the page was left exactly as it had been: the spinner stopped, the panel stayed empty, and that looked no different from a search that found nothing, a star quietly sliding back to where it started being the clearest example. Anything that goes wrong now gets a message wherever nothing more specific already appears, and being over a rate limit and having no connection at all are named, since one means waiting a minute and the other means something else entirely. Panels that refresh themselves on a timer stay quiet, because they try again by themselves.

  • The last article in a list gets marked as read when you scroll past it, in the installed app as well as in a browser tab. A row counts as read once it leaves the top of the list, but the check measured against the top of the screen instead, which installed on a phone sits a title bar and a status bar higher. Every row but the last one is tall enough to cover that difference; the last one, with exactly enough room under it to leave the list and no more, never could.

  • The bar of buttons at the foot of the list covers the same amount of the list in the installed app as it does in a browser tab. On a phone with a home indicator the bar is that much taller, and the list ran on underneath it, so sliding the bar in took a whole article row with it and left one sitting half covered. The list now stops where the indicator starts, which is what the reading shell was already doing on the other three sides.

Version 0.17.0 · 2026-09-07

Added

  • Readfine is in the share sheet on Android. Reading something in another app, share it to Readfine and it goes to Saved, fetched for reading like any link you paste in yourself. Usually there is nothing to do: a window says it was saved, and Back returns you to whatever you were reading. It stops to ask only when asking is the point, which is when more than one link came through and picking the wrong one would be worse than a question, or when none did. Apps are inconsistent about what they hand over, and many send the address buried in a line of text rather than on its own. A save that does not go through says so and hands you the button back, and if you share enough links in a row to hit the limit it says to wait a minute. The headline that came with the share is kept too, so an article whose page cannot be read, a 404, something behind a paywall, a page built entirely in JavaScript, lands in Saved under its own title instead of its web address; a page Readfine can read still names itself. Sharing to an app is something iOS does not offer web apps at all, so this is Android only; installing works on both.

  • Readfine can be installed to a phone's home screen or to a desktop, where it opens in its own window with no address bar eating the top of the screen. Settings → Preferences has an Install button under "This device" that does it in one press, and it only appears when your browser can actually install: otherwise it is in the browser menu as "Install app", and on iPhone and iPad it is Share, then "Add to Home Screen". It is the same Readfine either way, only without the browser around it. iOS keeps an installed app's data separate from Safari, so the first launch there asks you to sign in again. Self-hosted instances need HTTPS for any of this: browsers do not offer installation over plain http, localhost aside.

  • AI can now run on a model of your own. Settings → AI has a fourth provider, Custom (OpenAI-compatible), and an Endpoint field to go with it, so Readfine will talk to anything that speaks the OpenAI API: Ollama or llama.cpp on your own machine, LM Studio, vLLM, a LiteLLM proxy, or a gateway like OpenRouter. A local server needs no API key, so the key field can stay empty, and the check that a scoring model really answers with a number now covers those keyless setups too. One endpoint per account, shared by both slots, but each slot still picks its own model, so a small model can score while a larger one writes summaries. A model that reasons first needs no setup either: scoring answers with one number in ten tokens, which such a model spends on thinking and comes back empty, the state Qwen3 and friends arrive in by default, so Readfine asks the endpoint to skip reasoning for scoring while summaries, chat and briefings keep the room to think. Ollama wants the /v1 suffix on the address and llama.cpp does not, which is the usual first stumble; the help page covers that and the rest of the setup.

  • A model server on a private address has to be allowed by name. Self-hosting on localhost or a Docker network means listing that address in AI_ALLOWED_PRIVATE_HOSTS in the environment, as host and port: AI_ALLOWED_PRIVATE_HOSTS=ollama:11434, comma-separated if you run more than one. Anything not on the list has to be a public address, because on an instance with other people's accounts the endpoint is theirs to type and this is what keeps it from reaching services inside your network. The port is named too, so allowing a model server does not also allow the database sitting beside it. An address that is not on the list says so wherever it comes up, on the article, in chat, in a briefing and on the settings page, instead of reporting itself as a server that will not answer. Feed addresses are unaffected by the setting.

  • How short an article has to be before summarizing it stops being worth the tokens is now yours to decide. Settings → AI has a Minimum article length field under the auto-summarize switch: the fewest characters an article needs before it is summarized on its own. It counts the article the way the model sees it, HTML stripped and whitespace collapsed, title included. It gates the automatic runs and nothing else, so the Summarize button still works on a short article you decide you want one for, and Context, which explains what an article leaves unsaid and so earns its keep on exactly those short pieces, is unaffected either way. Changing the number applies from then on: it does not fill in summaries for articles you starred earlier, and it does not cancel the ones already queued.

  • The cost report explains what it shows. Notes under the table cover the empty cost column, and say that costs follow the model set for each slot now rather than the one that ran at the time. Worth knowing because compatible proxies often serve models under names like gpt-4o, and pricing a local run at OpenAI's rate would have looked exact while being invented.

Changed

  • Article titles are smaller when you read on a phone. They were set at a size meant for a desktop and then scaled up again by the phone's larger base text, which left a headline bigger on a narrow screen than on a wide one. A long one could take four lines before the article itself started.

  • The minimum length for an automatic summary starts at 1 700 characters instead of 1 500, which is roughly 280 words. The old figure let through pieces that were barely longer than the summary they produced. Existing accounts move to the new number too, and the field in Settings → AI puts it back wherever you want it.

Fixed

  • A failed fetch says what the status code means, even when the code is not a standard one. A feed sitting behind Cloudflare that could not reach its own server reported "HTTP 520" and nothing more, because the 52x codes are Cloudflare's own invention and the HTTP library has no name for them. Now those read "HTTP 522 Connection Timed Out" and the like, in the feed row, in the admin fetch log and on the readable view. The same goes for the codes nginx and a few anti-bot front ends make up.

  • The last article in a list no longer disappears behind the title bar on a phone. Scroll to the very bottom and the list would go on a little further than it had to, taking the last row up behind the bar and leaving an empty screen with nothing to hold on to. The room under the last row is there on purpose, so the final articles can be scrolled up far enough to count as read, but it was a whole screen tall while the list itself is shorter than that by the height of the bar.

  • Saving a link no longer takes the reading panel down with it. Paste an address Readfine has not seen before and, a second or two later, the panel articles open into would vanish: clicking anything in the list did nothing at all, and only reloading the page brought it back. While a saved article is being fetched, its row runs a small ticker that keeps the row up to date until the real title arrives, and that ticker was putting its answer where the article should go instead of back into the row it belongs to.

  • Wikipedia articles read properly. They arrived as one unbroken wall of text with a jumble of stray fragments piled on top, and both halves of that turn out to be the same page furniture getting in the way. The [edit] link sitting in every heading was enough for the extractor to throw the heading away, so nothing was left to break the article into sections. The infobox, the summary table of names and dates in the corner, was being taken apart and its contents scattered through the top of the page as loose lines. The [edit] links are now removed before anything else happens, which brings the section headings back, and the infobox is lifted out whole and put back as a proper table. On a wide screen it sits beside the text the way it does on Wikipedia; on a phone it is one line you can tap open, or already open when the box is short enough to read at a glance. In dark mode its label column sits on the box's own background rather than a gray stripe of its own. This covers any MediaWiki site, not only Wikipedia.

  • A picture the feed sent is not lost when the page is read for you. Webcomics were the worst of it: xkcd's feed hands over the strip itself and nothing else, but Readfine then fetched the page, found nothing there but the site's footer joke about Netscape Navigator, and showed you that instead of the comic. Any picture the feed delivered now comes through even when reading the page turns up none of its own, so the comic is there, above whatever the page had to say. Nothing is dropped to make room: an article with its own photos is left exactly as it was, and a news story keeps every word while getting its lead photo back.

  • Documentation, blog posts and project pages keep their section headings. They were arriving as one unbroken wall of text with every heading gone: a post on jvns.ca lost all 21 of them, uv's page on GitHub all 18, and the PostgreSQL SQL reference went further and had a keyword like GROUP BY pulled out of the middle of a sentence and dropped below it as its own code block, leaving the sentence saying "not allowed in a regular clause". Two things were doing it. Readfine reads a page in a mode that is strict about what counts as the article, which is what keeps navigation and "you might also like" blocks out of what you read, and on some sites that strictness throws the headings out along with them. The other is a heading carrying a link with no words in it, such as the small icon GitHub puts beside each one for copying a link to that section, which was enough on its own for the extractor to discard the heading, the way Wikipedia's [edit] links used to. Those icons are now taken out first, and where the headings clearly went missing anyway, the page is read a second time with the strictness relaxed and that result kept only if they really do come back. Man pages, the Rust book, the Kubernetes docs and the SQL reference come out with their headings and their inline code again, and a few blogs whose HTML is malformed, danluu.com among them, get theirs back from the same pass. Ordinary articles are untouched: against a 181 page benchmark of news and blog articles with hand-written reference text, one article of the 181 comes out differently, and it comes out better.

  • Saving a link that redirects in the browser now saves the article. Some sites answer an old address with a small page that bounces you onward, and because the bounce happens in the browser rather than in the response, Readfine stopped at the stub. The saved article's entire text was "Click here to be redirected." It now follows the hop, as long as it stays on the same site.

  • Tables in extracted articles keep their rows and columns. The extractor hands them over in a format of its own that was not being translated, so the table structure was dropped on the way through and the cells came out as a run of disconnected paragraphs in the middle of the article. Any page carrying a table was affected. On Wikipedia the tables get the same treatment as the infobox on top of that, because the extractor also tends to give up partway through a long one and scatter the rest: the climate table on the Prague article was arriving as a column of bare numbers with the temperatures missing altogether. Tables now come through whole, in the place they belong, and a wide one scrolls sideways on its own rather than stretching the article. The navigation footers at the bottom of a Wikipedia page are dropped, since pieces of them were leaking into the article as broken table fragments and they are no use to a reader anyway.

  • The privacy policy and the terms have a link back to the reader. Reaching either from the help page left the browser's Back button as the only way home. Signed-in readers get the link; a visitor who is not signed in does not, because it would only have led to the sign-in form, which is also why help and the feature list stopped showing it to everyone.

  • Pages outside the reading view keep clear of the rounded corners and the home indicator on an iPhone. Settings, help, the legal pages and the sign-in forms are ordinary scrolling pages with none of the reading view's own edges, so nothing had been keeping them out of the way: in landscape a line of text could run into the curve of the screen, and the bottom of a page sat under the bar at the very bottom. This is most obvious once Readfine is installed, where every page is drawn right to the edges, but it was there in the browser too.

  • Saving a link no longer claims you had already saved it when you had not. Pasting the address of an article that reached you through one of your feeds answered "Already saved", which reads as though the save did nothing. It did: the article was in Readfine, from the feed, and saving added it to Saved as asked. The message now says that instead.

  • The article list fills the screen on a tablet, and on a phone held sideways. Those widths use the two-panel layout, where the list is a fixed width with a drag handle beside it for changing it, and that handle only ever worked with a mouse. On a touch screen it was four pixels of nothing: impossible to grab, and leaving the list narrow with a band of empty background next to it. The list now takes the full width there and the handle is gone, since it set a width that could not be set on that device anyway. A laptop with a touchscreen keeps its handle, and the three-panel layout is unaffected.

  • A summary that was never going to arrive no longer leaves you watching a spinner that ends in nothing. Whether an article was long enough got measured against the copy trimmed to your content limit rather than the article itself, so the same piece could pass on the way into the queue and fail on the way out, and the queue said nothing when it gave up. With a content limit below the minimum length, that was every article you starred. Length is now counted on the full text and settled once, when the summary is queued, and a summary that really is skipped says so instead of quietly disappearing. The minimum can therefore sit above the content limit without trouble: those are two separate questions and one no longer cancels the other out.

  • Emptying one of the number fields in Settings → AI no longer hands you back the default. Clearing a field is how you say "no limit", and answering that with 20 000 was a number you never picked. An emptied field now keeps whatever was there, and the field shows it again after saving. A number that is genuinely out of range still falls back to the default and says so.

  • The thousands separator in Settings → AI stopped working after the first save. The fields group digits as you type, but the page is swapped in place when settings are saved, which left the grouping behind on the old page until a full reload.

  • The admin scoring eval no longer reports a number that flatters itself. A filter with an ai_score condition can decide its own score's ground truth: if a low-score rule also marks articles read or archives them, they never reach the unread list, so nobody can engage with them and the score looks prescient for having pushed them down. Where such a rule is active the page now names it, says how much of the sample it covers and how little engagement there is inside that band, and shows a second AUC computed without it. Both figures are labelled as bounds rather than answers, because leaving the band in credits the score for a tail it created, while leaving it out narrows the score's own range and makes any ranking look worse than it is; the real figure sits between them and the way to get it is to drop the mark_read and keep the label. The aggregate over all users only warns, since each threshold is that user's own and there is no shared band to leave out. The note claiming the figure was a lower bound is gone, since with a filter like that running it is the opposite.

  • The scoring eval's window no longer reaches past what retention keeps. It offered up to a year while un-engaged articles are deleted after 60 days, so a longer window held only the starred and archived ones that survive a purge, and those count as engaged by definition. The window now stops short of the purge horizon, the page says so, and the buttons are derived from the retention setting instead of being fixed at 30/90/180/365, which had become three buttons returning the same result.

  • An AI run no longer slows the rest of Readfine down while it sets up. Every call to a provider built its own TLS setup first, which is about 15 milliseconds of work that blocks everything else on the server, and a batch of articles being scored does it once per article. It is now prepared once and reused, so the same step takes a fraction of a millisecond, and reading stays responsive while summaries and scores are being generated in the background. Gemini keeps a little of that cost, because its own setup does more than the TLS part.

  • Readfine no longer leaves a connection to your AI provider open after every call. A connection was opened for each summary, score, chat message and digest, and none of them were closed, so on an instance that had been running a while they added up. They are closed as soon as the answer is in now. Verify in Settings → AI was opening two and using one.

  • Verify in Settings → AI now tells you what went wrong when the endpoint does not answer. A timeout reported itself as "Request timed out.", which says nothing about whether a local model is still loading or the server is not running at all, and an error that carried no message left the line blank with the button still reading "Verifying…", so there was nothing on screen to go on. A timeout now says the model may still be loading and to try again, an unreachable endpoint points at the address and the server, and no failure can leave the line empty.

  • Pictures that the site will not serve to anyone but itself now load. Readfine was telling every image host that the reader was on Readfine, and some of them refuse on exactly that basis: the little locator maps on Wikipedia were the visible case, and they did not just go missing, they left a hole several hundred pixels tall in the middle of the article, because the space had been reserved before the refusal came back. Nothing is said about where you are reading now, which is also one less thing leaking to a publisher's servers while you read. Links out to the original article never carried it in the first place, and video players are unaffected.

  • A page that answers with a browser check says so instead of saving the check as the article. Some sites meet anything that is not a person with a "please enable cookies" or "checking your browser" page, and PubMed is one of them, so an abstract you saved arrived as a ten word article telling you to turn cookies on. Nothing about it looked like a failure, which is the part that mattered: no error, no Open original button, just a very short article. Readfine now recognises those pages and reports them, so you get the address and can open it yourself. It does not retry them, because the next attempt would be answered exactly the same way. The article itself is still out of reach on such a site, since the check wants a browser running its JavaScript.

Version 0.16.0 · 2026-08-19

Changed

  • Settings → Feeds now explains a feed that has stopped fetching instead of just labelling it. A feed switched off after repeated failures showed a gray "disabled" and nothing more, which is the one state that will not clear up on its own, so the row now says fetching was switched off and that saving the feed's edit form turns it back on. A feed the site is refusing gets the same "throttled" badge as the admin tables, with the same explanation behind it. The feed's edit page adds how many times it has failed since it last fetched successfully, and reports a refusing site in the site's own words, in amber rather than red, because nothing is wrong with the feed and it comes back by itself.

  • Catch me up and briefings now always run on the main model, and the choice between the two models is gone from the form. The two slots had been pulling in opposite directions: scoring gives the model ten tokens to answer in, one number per article, so the model there has to be small, and that same limit decided what could write your digests. The slots are now named for what they do, Scoring and Main, instead of Fast and Quality. If you have only the scoring model set up, Catch me up says so and points you at Settings → AI. Briefings that were set to the fast model have been switched off instead of moved across on their own, because the main model usually costs more and a bigger bill arriving on schedule is not something to sign someone up for. They are still in your list, with the reason on them, and switching one back on clears it.

  • The AI cost report has one line for Catch me up & Briefings instead of one per model slot, and each run in it is priced by the model that actually wrote it rather than by whatever is configured today. The Fast total line is gone, since scoring is the only thing on that slot and the line just repeated it; the scoring row now sits above a divider, with one subtotal for the main model and the grand total below. The note under the heading no longer claims output tokens cost four to five times input: each model has its own price for both, and the rule of thumb was wrong for several of them by a wide margin.

  • You stay signed in for 30 days instead of 14. The session was already sliding, so it only ran out if you didn't visit at all, but two weeks is easy to skip in a reader without meaning to leave. Set SESSION_MAX_AGE_DAYS to make it shorter on a shared machine or longer on your own.

Fixed

  • A feed the site kept refusing can now actually be brought back. Fetching stops after ten refusals in a row, and getting out of that state was close to impossible: saving the feed's edit form reset the counters but left the feed deferred for up to a day, and the admin panel's "Clear error" only ever looked at feeds marked with an error, which a refused feed is not. So a feed came back reading as active, still counting "throttled x14", and was fetched by nothing; once the counters were cleared, the leftover deferral started blocking the manual refresh button too, which answered "rate-limited" for the rest of the day. Every way back now clears the whole thing at once: both counters, the message and the wait. In the admin panel the action is called "Reset errors" and is offered on any feed with a failure on it, not just the ones marked error, so a switched-off feed and a refused one can both be revived from there. A paused feed stays paused, and a feed an admin switched off by hand is left alone, since neither of those is a failure.

  • A problem inside Readfine is no longer reported on your feed as the feed's own error, and no longer switches it off. Whatever the fetcher ran into went to the feed row as it came, so after an update whose database migration had not run yet, feeds showed a message opening with a database error and a piece of the query, column names and all. It named nothing you could do about it and it exposed how the inside of the app is put together. Such a case now says only that an internal error happened and that the detail is in the fetch log, where an admin has always had the whole thing. Errors that really are the source's (an HTTP status, a timeout, DNS, a body that will not parse, an address that failed validation) read exactly as they did before. An internal failure also no longer counts against the feed, so it cannot add up to the five consecutive failures that retire one: the gap between an update and its migration hits every feed at once, and what used to come out of it was an instance with every feed switched off, each needing to be turned back on by hand.

  • Error messages are no longer cut off after one line. A feed's last error in Settings → Feeds, and the error columns on the admin dashboard and in the fetch log, all clipped the text to a single line, so on a typical fetch failure you got the opening words and nothing else. They now run to two lines before they trim, with a "Show more" link underneath for the rest, which appears only on messages that are actually too long. The tooltip holding the full message is still there, but it only ever worked with a mouse, so on a phone everything past the first line was unreachable. The readable extraction tables had the opposite problem, wrapping without any limit, so one long message could stretch a row halfway down the page; they follow the same two lines now. A two-line error also makes its row taller, and everything in that row now sits in the middle of it rather than hanging off the top.

  • Generating a digest from further down the Catch me up page now scrolls the result panel fully into view. It scrolled up to the panel, but stopped with the top of it tucked under the header bar pinned across the top of the page, so the first line of the briefing was hidden and it looked like the page had stopped short.

  • In Catch me up, the cost estimate now follows the rest of the form. Changing the period, the read/unread state, the labels, the score threshold or which feeds are included updated the number of articles below the form but left the estimate on the article count from before, so the figure you were reading was for a selection you had already moved on from; only the model or the article limit refreshed it. Turning off "Include article text" did not lower it either, although the digest itself was correctly built without the text, which is most of what you send the model, so the cheaper option looked no cheaper. The count and the estimate now update together on every change.

  • On a phone, opening search from the menu no longer darkens the screen twice over. Search is opened from inside the sliding menu and the menu stays open behind it, so that closing search puts you back where you came from, but both of them dim the screen and the two together came out much darker. Closing search then snapped the screen back to normal brightness, which looked like something had flashed across the whole page. Only the search window dims things now while it is up.

  • The search window no longer opens on last time's filters for a moment. Its contents are fetched, but the window itself appeared instantly, so until the new copy landed you were looking at the previous one, scope and label chips included, even when you had opened it fresh from the menu and it was about to come up empty. It now empties on close and keeps its size while it fills.

  • Running a search, or switching to another feed, no longer flashes the list you were reading. The list you had open was scrolled back to the top while it was still on screen, so for a frame or two you saw your old articles jump before the new ones took their place. It was most obvious when you had scrolled a fair way down, which is exactly when you are most likely to go looking for something. The new list now simply arrives at the top.

  • In dark mode, the bits the browser draws for itself are dark too. Scrollbars, dropdowns, checkboxes and the yellow tint on an autofilled field took their look from the operating system rather than from Readfine, so with a light system theme you got a pale scrollbar running down the side of a dark article list and a white dropdown on a dark settings page. They now follow whichever theme you have picked, including when you switch between them. The landing page and the sign-in screens changed with it.

  • In a narrow window, the article list's scrollbar no longer disappears behind the bar at the top. The bar is pinned over the list, and the list was only padded to clear it, which moves the text but not the scrollbar, so the top of the track ran on underneath the bar (and its bottom under the row of buttons whenever that slid in). The list now starts below the bar, so the scrollbar covers exactly the part of the list you can see. The bottom bar still floats over the list on purpose, since it comes and goes as you scroll and the list would jog every time it appeared.

  • The AI cost report has current prices again. Claude Sonnet 5's launch price was announced as introductory and the increase that was meant to follow it was called off, so the report was charging you half again as much as the model costs. Claude Opus 5, Claude Opus 4.5, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite were missing from the price list altogether, which meant a report on any of them showed an estimate based on a mid-range model from the same provider rather than the real figure. Prices are checked against each provider's own pricing page; everything else in the list was still correct.

  • A model that cannot be told to stop reasoning is now turned down when you pick it for scoring, rather than failing quietly on every article afterwards. Scoring gives the model ten tokens to answer in, which is all a single number needs, and a model that reasons whether or not it was asked to (Claude Fable 5 is the one that does this today) spends them thinking before it writes anything, so every score came back empty, with nothing to say so but the error banner, one article at a time. Saving a Scoring model now asks the model itself first, and a model that answers a scoring-sized request with nothing is refused with the reason instead of being stored; Verify says the same thing. Only the model's own answer counts against it, so a rate limit, a missing key or a provider having a moment leaves the choice to you. Such a model is still fine in the Main slot. On an account where one is already saved, its scoring jobs now give up on the first attempt instead of retrying twice more over the next two hours.

  • On a phone, the bottom of the app is no longer hidden behind the browser's own bar. The layout was sized to the tallest the window ever gets, meaning the height with the address bar retracted, so while the bar was showing (which is most of the time) the row of buttons along the bottom sat underneath it and the end of an article was cut off. It now follows the height that is actually visible and moves with the bar as it slides away. Readfine also keeps clear of the notch and the home indicator now, in both orientations, and the strip of browser chrome around the page takes the colour of the theme you are using instead of staying pale in dark mode.

  • One subscriber of a shared feed can no longer change the login the others fetch it with. A feed's HTTP credentials belong to the feed rather than to you, and go out on every subscriber's request, but the feed edit form offered the fields on any private feed, shared or not, and shared private feeds do exist: the release that moved passwords out of feed addresses marked those feeds private without separating their subscribers. On one of those, anyone subscribed could overwrite the password for everybody, break fetching for everybody by clearing the username, or read the username another subscriber had set. The fields are now offered only to a feed's sole subscriber, and where they used to be, the form says why and what to do instead: add the address a second time with the credentials and you get a private feed of your own, which is how a credentialed feed has always been created. The API refused to put credentials on a feed that wasn't already private, so it could never set them up in the first place; it now allows what the form allows, on a feed you are the only subscriber to, and turns that feed private in the process.

  • Turning readable extraction back on for a blocked feed now sticks. Extraction switches itself off after three page fetches in a row come back refused (403), and those refusals stay recorded on the articles that got them. The check for the next streak read the feed's most recent results without asking when they happened, so once extraction was back on, the old refusals were still the newest ones on record: the first new article that failed made three, and extraction went off again on the spot. The feed got one article to prove itself instead of three, whether you turned extraction back on yourself or the scheduled re-check did. Only what has happened since extraction came back on is counted now, so a feed gets its full three articles, and the same goes for the longer streak that switches off a feed whose pages extract to nothing.

  • A single HTTP 404 no longer switches a feed off. Readfine read a 404 as the host saying the address is gone for good, which is what 410 means but not what 404 reliably means in practice: sites also serve it while their own backend is briefly unwell, in bursts that take out every feed on that site at once. YouTube's feed addresses do it regularly, so two YouTube subscriptions that had been fetching fine all day could both be disabled at four in the morning and stay dead until someone noticed and switched them back on by hand. A 404 now goes through the ordinary retry counter on a shorter threshold of its own, so a feed is retired after five 404s in a row (around six hours) rather than the first one, and a wobble of a few minutes costs nothing. Addresses that really are retired, 410 and 451, still disable the feed immediately, and a feed disabled by an old 404 needs re-enabling once.

  • A feed that has just failed is now re-checked after half an hour, instead of waiting out the full error backoff (two hours at the default fetch interval). Most failures are a site having a moment, over long before the next attempt, so a healthy feed was losing hours to a problem that had already fixed itself. Only the first retry is quick: from the second failure on, the backoff is what it was, so a site that is genuinely down is not polled any harder than before.

  • Signing up with an address that already has an account now shows you the way back in. The message said the email was taken and left you there, so the easiest thing to do was register again under a different address, which is exactly what people did: a second account, a second OPML import, and the reading history left behind in the first one. It now offers logging in or resetting the password, both carrying the address you just typed, so getting back to your account is one field instead of starting over.

Version 0.15.0 · 2026-08-11

Added

  • Save an article by pasting its link, even one that isn't in any of your feeds. A new Saved entry sits under Archived in the sidebar, with a box at the top of the list to paste a URL into. The page is fetched, run through the same readable extraction as everything else, and shows up as an ordinary article you can read, label, summarize and chat about. It appears in the list straight away, first under its web address and then under its real headline once the text has been pulled in, with a spinner while the fetch runs and a red mark (the same one a broken feed gets) on a row that yielded nothing at all. Your filters run on it, so a rule scoped to "All articles" can label or star it as it arrives; scoring deliberately doesn't, since you already decided this one was worth keeping. A saved article is kept for good and is never purged by retention until you remove it from Saved yourself, from the article's ··· menu or the button row under it. Video links save as the video, and a link that redirects to a consent, login or paywall page is recognised as such rather than saved as article text; both still offer "Open original" and a retry. Because Saved is ordered by publication date a pasted link often lands partway down the list, so saving jumps to the article and flashes it.

  • Save a link from outside the app. POST /api/v1/articles/save-url takes {"url": "…"} and does exactly what the Saved box does, so a phone share-sheet shortcut, a bookmarklet or a line of curl can keep an article without opening Readfine. The article comes back straight away, with its address standing in for the title while the page is fetched in the background, so a client that wants the text asks for the article again a moment later. A link you already have is attached to your Saved list rather than duplicated, and says so with a plain 200 instead of the 201 a new save gets. Taking an article back out of Saved is in the API too, as {"is_saved": false}. The endpoint allows the same ten saves a minute as the web form, since each one fetches a page.

  • Videos play in the reader. An article holding a YouTube or Vimeo video, whether you saved the link yourself or it came embedded in the article, shows the thumbnail with a play button and starts the player in place instead of sending you off to the video site. No player is loaded until you press play, so an article you only scrolled past cannot have a video service's cookies set on it; YouTube is loaded through youtube-nocookie.com, and the caption under the thumbnail says which site the player will come from. The thumbnail itself comes from Readfine rather than from the video service (see Fixed below). Videos from a YouTube feed now show the video from the moment the article arrives, with the whole description under it: the feed already carries both, so the 1.4 MB watch page they were being read from is not requested at all any more. Where scripts cannot run, the thumbnail stays a plain link out to the video.

  • Chapter marks in a video's description are clickable. A description that lists "0:00 Intro, 3:40 The build" now turns each of those times into a link that jumps the player to that point, starting it if it isn't playing yet. Nothing else in a description is linked, since it is mostly sponsor and affiliate URLs. Where the page's scripts don't run, a mark still opens the video on its own site at the same moment.

  • The article you are reading is now marked in the list. In the three-panel layout the list gave no clue which of its rows was open in the reading pane, so finding your place after scrolling meant recognising the headline. The open row now carries a blue bar on its left and the same pale fill the sidebar uses for the section you are in, and it stops being dimmed even if you have already read it (which is most of what going through starred articles is made of). It follows the article however you got to it, including a direct link and the Next button on a phone, and stays put when the row is redrawn. The two-panel layout is unchanged, since there the article opens under its own row.

Changed

  • AI summaries of long articles are no longer cut short. The prompt asks the model to match the summary's length to the article, but the room it had was fixed, so a long feature ran out of space and stopped mid-sentence while a short news item was never near the limit. That room now scales with the article, from the previous allowance for short pieces up to roughly three times that for the longest ones. An ordinary news story was cut off too, for the opposite reason: the allowance it got was sized for a summary written as one paragraph, while the models write a few short labelled sections, which takes more room. It is now sized for what they actually produce. The prompt now also puts numbers on the length, since a model given more room will otherwise use it and end up cut off at the higher limit instead. It still scales with the article, from a sentence or two for a brief item to around 150 words for an ordinary one and up to 300 for a long feature, and those are ceilings rather than targets. It also asks for one list of key points instead of several labelled sections. The numbers were needed because the same wording landed differently on different models: what one read as a tight summary, another read as licence for twice the length. Nothing to configure, and the ceiling keeps a custom summary prompt from turning into an essay.

  • Subscribing to a feed now tells you what the feed is called before you commit to it. The box has always had a "Custom title" field, but the feed's own name appeared nowhere on the page, so there was no way to judge whether you wanted a custom one, and the name a feed showed up under was a surprise. Where it mattered most was the list Readfine offers when the address you gave isn't a feed itself: the feeds found on the page were listed by URL alone, with nothing to tell them apart. They now carry their real names, read from each feed rather than from the title attribute on the page's link tag, which is usually missing or just says "RSS". That also names the addresses Readfine guesses at (/feed, /rss.xml), which had nothing before. Picking one from the list fills its name into the custom title's placeholder, and so does testing an address yourself, the same way the feed edit page has always worked: leave the field blank and the feed keeps its own name. None of this costs an extra request. The green "feed added" confirmation also steps aside as soon as you test or add the next one, instead of staying on screen naming a feed you dealt with several feeds ago.

  • When an AI provider returns nothing, the error now says why. "returned no usable content" covered three different situations: the model refused the request, it ran out of room before writing anything, or it genuinely replied with nothing. The notice now carries the provider's own stop reason (and, for Gemini, the reason a prompt was blocked), so you can tell whether to reword a prompt, raise a limit, or just retry.

  • The AI error notice in Settings → AI now says which article the failure happened on, with a link straight to it, so a failure on one odd article no longer looks the same as a broken API key. Errors that belong to no particular article, such as a failed interest profile, show the message alone. If retention has since removed the article, the notice keeps the message and drops the link.

  • Private feeds are now marked in the feed lists. Settings → Feeds and the admin feeds table put a small lock next to a feed that is kept to your account rather than shared with other subscribers on the instance, which until now you could only find out by opening the feed. Public feeds stay unmarked, since that is what almost every feed is. The badge on the feed edit page went gray at the same time: amber means "something needs attention" everywhere else in the app, and a private feed is not that.

  • The help page now covers the reader itself, not just the setup. A new "Getting around the reader" section names the things you would otherwise have to stumble on: the ↻ and ✓ buttons that appear when you hover a feed, folder, label or view in the sidebar (press and hold on a touchscreen), folding folders and the sidebar away, the ··· menu on an article, and the / shortcut for search. New answers explain Saved, sharing an article by link and how to revoke it, what a red marker on a feed means, how long articles are kept and what makes one stay for good, and that there is an API.

  • The admin dashboard is reordered around what actually needs attention, with fetch errors near the top instead of buried under the readable-extraction tables. A row of pills under the counters names every problem the dashboard found and jumps to it, and an instance with nothing wrong says so in one line rather than printing five empty tables. One of those pills carries something the dashboard never reported at all: a feed that failed often enough to switch its own fetching off. Such a feed stops being fetched, so it stops logging errors and drops out of the 30-day list, and the error counter passes over it too because that counts feeds still retrying. It now has a pill of its own, however long ago it gave up, leading to the feeds table where it can be started again or removed. Fetch errors are now one row per feed rather than one per attempt, which is what made the old list mostly the same broken feed repeated: each feed shows how many fetches failed in the last 24 hours, 7 days and 30 days, and the feed's state right now (active, throttled, error, disabled) next to its last error, so you can see at a glance whether it has recovered since. In this one list a feed the site is refusing reads as throttled from the very first refusal, since every row here is on screen because a fetch failed; elsewhere the badge still waits for three in a row, where a lone 403 is noise rather than a problem. The tables also show more than the five rows they were capped at, and say how many more there are. A fourth counter card tracks feeds currently in error, and the readable-extraction numbers now say what they count, which they previously left to guesswork.

  • Fetch failures are no longer logged forever. Nothing ever pruned that table, so a feed that kept failing left one row per attempt for the lifetime of the instance. Those rows are now dropped after 90 days, which is well past anything that reads them: the admin dashboard groups the last 30 days and the fetch log page shows the newest 100 entries, so nothing on screen changes.

Fixed

  • Summaries and relevance scores now work on the newest Claude models. Models of the latest generation (Opus 5, Sonnet 5, Fable 5) reason before they answer even when nothing asks them to, and that reasoning is paid for out of the same allowance as the answer itself. Readfine sizes those allowances for the answer alone, and scoring asks only for a single number, so the whole allowance went on reasoning and the reply came back empty. Every request to Claude now says reasoning should be left off, which puts the allowances back to meaning what they say. A few models refuse to switch it off at all, Fable 5 among them, and those are asked again with room for the reasoning on top of the answer, so a summary comes out whole instead of stopping mid-sentence. Scoring is the exception: it asks for a single number in ten tokens, which such a model cannot answer at any allowance worth paying for, so it says so and names the reason rather than failing blankly. Pick a model that can leave reasoning out for the fast slot. Nothing changes for OpenAI or Gemini, or for the Claude models you were already using.

  • The "Test connection" button in Settings → AI no longer reports success for a model that cannot actually answer. It sent a greeting and checked that the request came back, not that anything was written in reply, so a model failing in the way above reported a working connection while every summary and score errored out. It now reads the answer, and a model that returns nothing fails the check with the provider's own reason.

  • A summary that ran out of room now says so, instead of passing itself off as complete. Readfine stored whatever the model had written by the time it hit the limit and showed it like any other summary, so a summary ending mid-sentence looked like the model's own choice of words. Such a summary is still kept, but the "AI summary" heading now reads "AI summary · truncated". Regenerating one that comes back complete clears the note.

  • A summary that starts on its own now shows its spinner. Starring an article with auto-summary on, or opening a starred article whose extraction finishes and produces one, generated the summary in the background with nothing on screen to show for it, so the only way to find out was to open the article again. The "Generating summary…" spinner now appears in both cases, and a summary that arrives while you are reading is shown straight away. Unstarring while it spins removes it.

  • Consent, login and paywall walls are no longer stored as an article's text. Some sites answer a fetch that carries no consent cookie by redirecting to such a page, and since it comes back as HTTP 200 the notice was stored as the article, filed under the site's own name (iDNES and a Google News link were two cases). A fetch that ends on a page holding the address it interrupted, which is how a wall returns you afterwards, is now recognised from the redirect itself, so it needs no wordlist and works in any language. Such an article says the site returned a consent page and offers "Open original" and Retry, and keeps the link you saved so neither button walks back into the wall. Articles from feeds whose pages answer this way fall back to the text the feed itself delivers. Retrying on a schedule is not attempted, since a wall answers the same way every time; the Retry button still works.

  • Full-text extraction no longer turns itself off on a feed where it works. Readfine watches whether a feed already delivers whole articles and switches extraction off when it does, but that check read the word count stored with each article, and a successful extraction replaces that number with the word count of the page it pulled in. A feed publishing two-sentence teasers then looked like a full-content feed once you had read a few of its articles, and extraction was switched off at the next fetch, with nothing under readable errors because nothing had failed. The check now counts the words the feed itself delivered, which no extraction overwrites. The same measurement decides the setting when you subscribe to a feed someone else on the instance already reads.

  • Video thumbnails no longer reach out to the video service when you open an article. A thumbnail's address pointed straight at img.youtube.com, or at vumbnail.com for Vimeo, so your browser fetched it from there the moment an article opened and handed that host your IP address and which video it was, before any click. Thumbnails are now served by Readfine, which fetches each one once, caches it, and hands your browser only its own address; for Vimeo this also drops vumbnail.com, a third party, in favour of Vimeo's own thumbnail. Articles saved before this update are covered too. The player is unchanged: clicking still loads it from youtube-nocookie.com or player.vimeo.com, which is the point at which those services can set their own cookies. (The exposure had been there since videos in articles first showed a thumbnail.)

  • Adding a feed by filling in the HTTP auth fields no longer fails on the first fetch. The credentials were saved with the feed but not used for the fetch that happens while the feed is being added, so a feed behind HTTP auth answered 401 at the moment of subscribing, and the only way in was to write the credentials into its address instead. Both ways work now.

  • Share → "Original article" copied a link to Readfine instead of the article's own address. The share menu reads the source URL off the article in the reading pane, but a wrapper element added around it later was picked up first, and it carries no URL, so the code fell back to the page you were on. Sharing to Readfine was affected too, in a smaller way: the shared link itself was right, but the title handed to the system share sheet was empty.

  • Article text no longer runs off the side of the screen. Editors like WordPress often glue words together with non-breaking spaces around bold text, so a phrase such as "Wednesday, Anthropic and AMD announced" counted as one 38-character word that a browser cannot wrap, and it hung over the edge of a narrow reading pane. Newly fetched articles now get the longest of these runs broken back into ordinary spaces, keeping the ones that are there on purpose (10 km, 50 %, one-letter prepositions). Articles already stored, and anything else without a break point such as a very long URL, now wrap instead of overflowing.

  • An article no longer jumps around while its full text is being extracted. Opening an article whose extraction had not run yet starts one in the background, and the reader checked on it every two seconds by redrawing the entire article, which re-laid out the text, reloaded the images and flashed the action bar. Most obvious on a phone, where the page visibly hopped every couple of seconds. The check now touches only the small "Extracting full content…" line, and the article itself is rebuilt once, when the text is ready.

  • The counts next to Starred and Archived in the sidebar could stand above the number of articles those views actually open onto. Articles that reach the end of their retention are stripped to a snippet and hidden everywhere they used to be listed, but the two counters were taken straight off your article states without checking, so such an article kept its place in the badge while being absent from the list. Both counters now count what the list shows, and so does the number that appears after you mark a whole view as read.

  • Testing a feed no longer fails in silence when you have tested a lot of them in a row. Ten tests a minute are allowed, and going over that emptied the result box and stopped the spinner with nothing else to show, which read as the test having done nothing at all. It now says the limit was hit and that a minute's wait clears it. Any other failed request in that box reports itself too, instead of disappearing.

  • The article list highlights the row under the mouse again. The row and the list behind it had both ended up white, so in the light theme nothing happened when you moved across the list. Rows now use the same faint gray the rest of the app uses for hovering, and the dark theme keeps the shade it already had. Touchscreens deliberately have no highlight, since a tap would leave one row lit up; that now follows from the device rather than the width of the window, so a narrow window on a desktop keeps its highlight.

  • The amber "throttled" badge is readable in dark mode again. Its fill kept the light theme's pale yellow while the label turned bright amber, so the two sat almost on top of each other and the word was barely there. The fill now darkens with the rest of the theme. The same pairing is used by the 429 tag in the admin rate-limit list and the warning on a feed test in Settings, so those are fixed too.

  • Switching between Settings and Admin on a phone no longer hides the page you landed on. The row of tabs across the top scrolls sideways, and the active tab was being scrolled into view on load, but the browser restored the strip's old position a moment later and undid it. The tab is now brought back into view after that, and left alone as soon as you scroll the strip yourself.

  • The readable-extraction panels in the admin dashboard read straight and no longer scroll sideways. The revival panel's heading counted the feeds still awaiting a probe while its table listed the feeds a probe had already revived, two different sets, so the count looked wrong against the rows; the heading now belongs to the table and the feeds still waiting are named in a sentence below it. A feed could also turn up claiming it was revived by no probe, because saving the feed's form cleared the revival bookkeeping while the date stayed behind; the date now goes with the rest. Both extraction tables had also gained a feed column that pushed them past the width of the page: the feed now sits on a second line under the article title, the columns have fixed proportions, and a screen too narrow for all of it scrolls rather than squeezing the text.

Security

  • A feed's password is no longer kept in the clear. Credentials can be given to a feed two ways, in the HTTP auth fields or written into its address as https://user:password@host/feed, and only the first was encrypted. The second was stored exactly as typed, which put the password in the database and in every backup, in the feed's name when the feed offered none of its own, in the addresses of every article of a scraping feed, and in the OPML file you get when you export your feeds. It also left such a feed in the shared pool, so anyone else on the instance subscribing to the same address fetched it with your credentials. Credentials are now taken out of the address as the feed is added, encrypted like the ones from the fields, and the feed is kept to your account. Feeds you already have are converted on upgrade, their names and their articles' addresses included. Adding a feed does not change: an address with credentials in it still works, and what you see afterwards is the plain address. One consequence worth knowing: an OPML export now carries that plain address, so re-importing a feed that needs credentials asks for them again, which is already how feeds set up through the fields behave.

  • Full-text extraction now connects to the address it checked. Every outside address is verified before Readfine requests it, so it cannot lead to the machine itself or to something on the local network, but the extraction path then passed the host name on to the HTTP client, which looked it up a second time. A name server under someone else's control can answer those two lookups differently, handing a public address to the check and a private one to the connection a moment later. Extraction now uses the same verified-address path the feed fetcher already used, on the first request and on every redirect it follows. This closes an opening that grew more exposed with saving an article by link, which fetches whatever address it is given, on request and as often as asked.

  • A feed's login details are no longer offered to whatever a redirect points at. When a feed authenticates with a username and password, those were attached to the HTTP client rather than to the individual request, so they went out again at every step of a redirect chain, including a step that left the site. A feed host answering with a single redirect elsewhere could collect them that way. They are now sent only while the address being fetched is still on the site they were entered for, and a redirect that leads off it is followed without them. A host moving its own address from http to https keeps them, since that is the same host and the safer of the two.

  • Readfine no longer downloads a page of unlimited size. Nothing put a ceiling on how much a fetch could pull in, which mattered little while the addresses being fetched were feeds you had subscribed to once, but saving an article fetches whatever address you paste, as often as you paste one. A link that turns out to point at a disk image or a video file was read into memory whole. Downloads now stop at 10 MB, counted after decompression, so a small response that unpacks into gigabytes is stopped as well. A page over the limit says so and is not retried on a schedule, since it would be the same size next time; the Retry button still works. The body of a redirect is not downloaded at all any more, since only the address it points to was ever wanted. Self-hosters can change the limit with MAX_FETCH_BYTES.

  • The admin feed list and fetch log no longer show the secrets some feed addresses carry. A feed can authenticate through its own address, either as https://user:password@host/feed or with an API token in the query, and both printed those addresses in full. Both now hide the credentials and the value of any parameter that names a secret, while keeping the rest of the address, which is what tells two feeds on the same site apart. A redacted address is not a working link any more, so it is shown as plain text instead of a dead one. The edit form for a scraping feed also stopped carrying the address in a hidden field; the preview and the AI selector ask the server for it by id, and the server hands it over only for a feed you are subscribed to.

  • The error message on a failed feed no longer quotes the address with its secrets in it. That message is shown in your feed list and in the admin panel, in the latter right next to the address itself, which is printed with any API key in it blanked out. A message that quoted the address in full walked straight around that. HTTP errors were already built from the blanked version; everything else now gets the same treatment, and credentials written into an address are dropped wherever they turn up.

Version 0.14.0 · 2026-07-30

Added

  • Articles with nothing to read now offer the original page. Feeds that carry only a headline and a link (Hacker News is the clearest case) produce articles with no text at all, and opening one showed a blank pane and a line saying so. Such an article now shows an "Open original" button, plus "Retry extraction" when a previous attempt failed, so a source that was temporarily unreachable can be tried again without digging through the menu. Settings → Preferences adds "Open original for empty articles", off by default: with it on, clicking one of these articles opens the source in a new tab straight away, while the article itself still opens in the reader behind it, so it is marked read and stays starrable and labellable. Only articles that will never have text are affected, never one whose extraction is still running or could still be started. Time spent on the source counts towards reading time.

  • The interest profile used by AI scoring can now update itself. Settings → AI has an "Auto-generate" switch next to the Generate button with Off (the default), every 2 weeks and every 4 weeks. Until now the profile only changed when you opened settings and pressed a button, so scoring kept ranking articles by whatever you were interested in months ago. An automatic run saves straight away, keeps the version it replaced and offers a one-click revert, and the settings page always shows when the profile last changed and whether it was you or the schedule. Several things have to line up before it spends anything: the interval has to be up, at least 20 articles you read or starred have to be new since the last change, and there have to be enough reading signals that the profile is not padded with feed names. Output that does not look like a profile is thrown away rather than saved, and any attempt that reaches the model, successful or not, waits a full interval before the next one, so a dead API key cannot cost you a call a day. After three failures the schedule turns itself off and says why, in settings and on the admin dashboard.

  • Bot protection on the registration form, on by default and with nothing to configure. Open registration means the instance will email any address a visitor types, which bots abuse to flood scraped third-party inboxes using your domain as the sender. The form now carries a hidden honeypot field that only a script fills in, and a signed timestamp that rejects submissions arriving faster than a person could type. A caught submission creates no account and sends no email. Only the honeypot, which nothing legitimate can fill in, answers with the normal "check your email" page so the script cannot tell it was caught; the timing check is a heuristic, so it puts the form back in front of you instead, as does a form left open for hours. Correcting a mistyped password and submitting again does not restart the timer either, so a password manager refilling the form in one click cannot trip it. The README explains what else to put in front of a public instance.

  • LOG_OUTBOUND_REQUESTS (off by default): a diagnostic switch that logs one line per outbound HTTP request, covering feed fetches, scraping and readable extraction. Each line records the host, status, HTTP version, elapsed time and any rate-limit headers the server sent (Retry-After, X-RateLimit-*). Per-feed error records only show failures, and only per feed, so they cannot answer how often a host is really being hit when several feeds and the extraction pipeline share it. Turn this on when a site starts returning 403 or 429, read the real request rate and spacing from the log, then turn it off again.

  • LOG_LEVEL (DEBUG/INFO/WARNING/ERROR, default WARNING): how much the app logs. WARNING keeps the log to things that need attention, INFO adds the running commentary from the scheduler and fetcher. Noisy libraries (httpx, APScheduler) stay at WARNING either way, so raising the level surfaces Readfine's own records rather than a wall of third-party chatter.

  • FETCH_SCHEDULE_OFFSET_MIN (minutes, default 0): shifts the four 15-minute feed-fetch ticks off the usual :00/:15/:30/:45. Useful when two instances share a host (for example a staging instance next to production) and you don't want both firing their fetch round at the same wall-clock moment. Set staging to 7 and it polls at :07/:22/:37/:52 instead. The value is folded into 0–14, and the predicted next-fetch times in the UI follow the shifted schedule.

Fixed

  • A filter regex could quietly fail to match on a busy instance. Filter patterns run under a time limit so that a pathological one cannot freeze the app, and a pattern that runs out of time counts as "no match", which means the filter silently does not fire. The limit was 0.1 seconds of wall-clock time, close enough to what an ordinary pattern costs on a long article that a busy fetch round was enough to trip it: a production log showed \bAI\b timing out. The limit is now a full second, which still stops a runaway pattern but leaves normal ones far below it.

  • A feed whose server dropped the connection mid-request is now retried once instead of counting as a failed fetch. Reusing a kept-alive connection races with the server closing it, and HTTP/2 servers close them routinely, so a request could die with nothing sent back through no fault of the request itself. Fetches are plain GETs, so a second attempt on a fresh connection is safe, and it recovers feeds that were losing a poll here and there for this reason.

  • Settings → AI no longer announces a nightly interest-profile generation that cannot happen. The status line checked that a quality model was picked, but not that there was a key to use it with, so an account whose key had been removed (or a restore that lost the encryption key) read "The next generation runs tonight" indefinitely while the job skipped every night without recording anything. It now says which provider is missing a key.

  • Three AI settings actions (saving a provider key, verifying a model, queueing summaries for all starred articles) were served with no rate limit, despite carrying one in the code. The limit was attached to the route in the wrong order, so it wrapped a copy of the handler that nothing ever called. Generating the interest profile by hand had no limit declared at all, which mattered more: it is the most expensive call the app can make, it builds its prompt from the whole reading history, and one stuck button could run it as fast as the provider would answer. It now allows 5 per hour, configurable through RATE_LIMIT_AI_PREFERENCE. A test now checks every rate-limited route is really limited, since the failure left no trace anywhere.

  • The general AI chat no longer scrolls the page behind it once you have sent a message. The message list was set to swallow the scroll at its own end, but only until the first reply arrived, because the server rebuilt the panel without that setting. Most visible on a phone, where reaching the bottom of a conversation started dragging the article list underneath.

  • The label next to the paperclip in an article's chat now updates when you detach or reattach the article, instead of only doing so after your first message. The panel was drawn without the hook the script needed to find it, and a later redraw quietly added it back.

  • Saving Settings → AI with scoring switched off no longer erases the interest profile. The profile field is disabled while scoring is off, and a disabled field sends nothing at all, so the save read it as "cleared" and wiped the text. Fields that were not part of the submit now keep their stored value, which also covers the new auto-generate schedule and the score-in-list toggle.

  • The scoring checkbox in Settings → AI now enables and disables the fields below it as you click it, instead of only after saving. The checkbox carried two id attributes and the script hooked onto the one the browser had thrown away.

  • Feeds on sites that block automated clients no longer get marked as failing and eventually switched off. Some sites (Reddit is the obvious one) refuse a share of requests with HTTP 403 or a bare 429 and let the rest through, in waves that last minutes to hours and hit every feed on the site at once. Readfine counted each refusal as a fetch error, so five of them in a row disabled a feed that was working fine. Refusals are now tracked separately from real errors: the feed keeps its normal state, backs off progressively instead of retrying on its usual schedule, and is only switched off after ten in a row. Settings → Feeds and the admin feed table label such a feed "throttled" once it has been refused three times running, in amber rather than red, and one stray refusal shows nothing at all. Any successful fetch clears the count.

  • Readfine could learn a fetch spacing that no site had actually asked for. A site that reports its rate limit as exhausted also reports how long until the limit resets, and that countdown was read as if it were the sustainable gap between requests. It is not: it depends on when the request happened to land in the site's current window, so the value swung between 0 and 60 seconds on identical traffic, and the "never loosen" rule kept the highest number seen. On Reddit this settled at 78 seconds, throttling every feed on the host for no reason. The countdown is now used for what it is, a deadline before the next request, and the learned spacing only comes from a limit with room left in it or from an explicit Retry-After. The stored values are cleared once on upgrade so nothing carries the old numbers forward; sites that advertise a real limit are re-learned on the next fetch.

  • Admin → Rate limits could list a host at a spacing of 0 seconds, which read as a learned limit but meant the opposite: a single rate-limit response had created the row before anything was learned, and it then stayed forever. Such rows are no longer written, and existing ones are removed.

  • Full-content extraction switched off because a site refused it now gets a second chance instead of staying off for good. Three refusals in a row disable extraction for a feed, for everyone subscribed to it, and only editing the feed by hand ever turned it back on. That made a passing block permanent: the move to HTTP/2 fixed a whole class of these refusals, yet the feeds it fixed stayed disabled, and the damage went past the article body, because labelled articles from such a feed reach AI scoring with the short feed text instead of the full one. A feed in this state is now retried twice, three days after it was disabled and a fortnight after that, and extraction comes back on if the site lets a page through. The retry only looks at whether the page downloads, so a video post or live blog at the top of the feed cannot condemn it. Feeds you switched off yourself are left alone, and so are feeds disabled for the other reason (they already carry full articles). A feed that gets refused again after coming back is not retried a third time. Existing feeds are queued for a retry when you upgrade, spread over a week. Nothing is re-extracted retroactively; articles already stored keep the text they have.

  • Settings → AI spells OpenAI's name the way OpenAI does. The API key list capitalised the stored identifier, so openai came out as "Openai", and the two provider dropdowns showed the identifier untouched, in lowercase. All three now use the providers' own names.

  • Application log records never reached the log at all. Uvicorn configures only its own loggers and leaves the root logger without a handler, so every logger.info() in Readfine was discarded and warnings fell through Python's fallback handler with no timestamp and no source, leaving no way to tell which part of the app wrote a line. Logging is now configured at startup with a proper format. This also means LOG_OUTBOUND_REQUESTS produces visible output; it lifts its own records above LOG_LEVEL, so switching it on is enough.

Changed

  • Settings → AI is now grouped by feature: Scoring, Summaries & context, Chat, and Limits, each under its own heading. The interest profile used to sit in a section of its own at the bottom of the page, far from the switch that decides whether it is used at all; it is now indented under "Enable article scoring" together with "Show score in article list", so it is visible which settings depend on which.
  • A feed that has permanently moved now gets its stored address updated, so each fetch is a single request again. Until now the redirect was followed but never remembered, and the feed walked the same chain on every poll, forever: one feed here cost three requests and 800 ms where one would do, and it counted against the request budget of sites that ration them. The new address is only taken when the site said the move is permanent, the fetch actually produced articles, and nothing changes along the way that should not (a query string, an HTTPS connection, or credentials in the URL, which are neither dropped nor picked up: a site that answers with credentials in its redirect does not get them stored on a feed everyone shares). Existing feeds fix themselves on their next fetch. Adding a feed resolves the address first as well, so a subscription, or an OPML import carrying a years-old URL, lands on the feed you already have instead of creating a second copy of it.
  • Admin dashboard: a "Feed redirect conflicts" section, shown only when there is something to report. It lists feeds that permanently redirect onto a URL another feed already holds, the one case where the address above cannot be updated, so the feed keeps re-walking its redirect. The fix is to merge the pair by hand; the section names both feeds so it can be found.
  • Admin → Users: hovering the "Joined" date now shows the exact date and time of sign-up. The column itself still shows the date alone, which is not enough to see that a burst of accounts arrived within the same minute.

Version 0.13.0 · 2026-07-19

Added

  • Preferences → "Number & date format": a per-user choice of how numbers and dates are written, independent of the interface language. Five profiles cover the common conventions (US, UK/International, Europe, DE/AT, ISO), each differing in the decimal separator, thousands separator and date order, so you can keep the app in English yet see 1 234,56 and 25.06.2026. New accounts are detected from the browser at sign-up; existing accounts start on the Europe profile and can switch anytime. The setting drives numbers across the stats and AI cost views and the numeric date formats throughout the app, including the date shown in briefing emails (in your timezone). Times stay 24-hour for now.
  • A categorized feature list at /features, linked from the landing page and the help guide, plus a matching FEATURES.md at the repo root. Both are generated from one source (backend/app/content/features.yml): the app renders it at runtime, and scripts/gen_features.py projects it into FEATURES.md, so the list is never maintained in two places. CI regenerates the Markdown and fails if the committed copy is stale.
  • Adaptive fetch intervals: a feed left on "Auto" is now polled at a cadence derived from how often it actually publishes, rather than the flat default. Readfine counts each feed's items over the last 7 days and targets an interval a bit shorter than its real publish gap, so busy feeds refresh more often and quiet ones less. New feeds and feeds without enough history keep using the global default, and an explicit per-feed interval still wins. Admin → Settings gains a "Maximum fetch interval" cap for how rarely a quiet feed may be polled; the feed edit screen shows the interval Auto would pick. Cadence is recomputed daily and at startup. The feeds tables (Settings → Feeds and the admin panel) show each feed's effective interval and its predicted next fetch (relative, e.g. "next ~1h") under the last-fetch column, with intervals of an hour or more rendered in hours.
  • Preferences → "Advance after mark all as read" (off by default): after you mark a feed, folder or label read from the sidebar, Readfine selects and opens the next one that still has unread articles, expanding a collapsed folder if needed. Feeds advance across folder boundaries; empty scopes are skipped, and the special views (All articles, Starred, Archived) are left alone.
  • Admin → Feeds: an "Edit" action on the table's ··· menu for shared feed fields (title, status, fetch interval, and the scrape article-links selector with a live preview). Per-subscriber preferences and feed credentials are intentionally left out, since an admin usually is not the subscriber.
  • Admin dashboard: a "Briefing errors" section listing catch-up configs whose scheduled briefing is currently failing (user, config, error, retries, next send). Configs with no scheduled retry, for example when SMTP is unconfigured, sort first and are flagged as needing manual attention; the entry clears once a send succeeds.
  • Admin → Users: per-user columns showing whether an account is currently active, not just its lifetime totals. "Read 7d" counts articles genuinely read in the last 7 days (marked read with at least 30 seconds of dwell, the same signal the reading stats use), "AI 7d" counts AI operations in the last 7 days across summaries, scoring, context, chat and catch-up, and a filter count.

Changed

  • Fetch interval dropdowns (Admin → Settings and the feed edit screens) now show longer intervals in hours, for example 6h or 24h, instead of raw minutes like 360 min. This matches how the feeds tables already render intervals. The admin settings also spell out how the default, minimum and maximum intervals interact with Auto mode (the minimum floors Auto too, and Auto never polls faster than 30 minutes).
  • Readable extraction backfills an article's publication date from the article page when the feed listing carried no date, so undated articles sort and expire correctly instead of all landing at fetch time. It reads the page's structured datePublished (via htmldate) rather than the oldest date on the page, guards against implausible future dates, and never overrides a date the feed already provided. The reader's date updates once extraction finishes.
  • Deleting a feed subscription or folder now cleans up references to it left dangling in filter scopes and catch-up/briefing scopes. As a safeguard against silently widening, a filter or briefing whose scope would empty out is deactivated or disabled instead, and the affected filter and briefing names are surfaced in the feeds settings banner.

Fixed

  • Admin → Feeds: the "Force fetch" button no longer returns a 500 when the fetch fails (for example a rate-limited or erroring feed). A failed fetch rolls back the database session, which expires the loaded feed and admin objects; the handler then read an attribute off one of them and crashed with MissingGreenlet. It now captures what it needs before fetching and reloads the feed's post-fetch state safely, so the usual "Rate-limited, try again in..." toast shows instead of an error.
  • The sidebar now highlights the active category right after opening or reloading the app, not only after a click. On load the highlight was applied to the collapsed rail copy of each nav item (which is hidden) instead of the visible full-sidebar copy, so nothing appeared selected until you clicked a category.
  • Feeds behind some CDNs (most visibly Reddit, via Fastly) that kept failing with 403 Blocked or persistent 429 now fetch normally. The fetcher spoke plain HTTP/1.1, which these CDNs treat as a bot signal and answer with a header-less 403 or a near-zero rate budget, while serving HTTP/2 clients as usual. Server-side fetches (feeds, scraping, readable extraction) now negotiate HTTP/2 when the server offers it and fall back to HTTP/1.1 otherwise. The User-Agent was never the cause.
  • Scrape-type feeds now record a "last published" date in the feeds tables (Settings → Feeds and the admin panel), like RSS feeds already did. The scrape fetcher tracked every other feed field but never set this one, so the column stayed empty even when the scraped listing carried article dates. It now advances to the newest dated link on each fetch and stays empty only when the listing exposes no dates at all.
  • The single-feed API endpoints (GET/POST/PATCH /api/v1/feeds/{id}) now report unread_count computed fresh from the database, like the feed list already did. They previously serialized a cached column that no longer had a live writer on every path, so it could read as stale or zero. The cached column has been dropped (migration 0079); every response counts unread on read.
  • Server-side feed and scrape fetches now pin each connection to the IP address they validated, closing a DNS-rebinding gap: previously the SSRF check resolved the hostname once and the HTTP client resolved it again at connect time, so a hostname whose DNS flipped between the two could pass validation yet connect to a private or cloud-metadata address. Every request and redirect hop now connects to the checked IP, carrying the original Host header and, for HTTPS, the original hostname for TLS SNI and certificate verification.
  • Readable extraction is no longer enabled for Tumblr feeds, which already deliver the full post in the feed. Extracting the page instead pulled in the likes/reblogs "notes" list as the body and duplicated the post text; Tumblr (and other feeds that advertise themselves as full-content via their <generator>) are now treated as full-content at subscribe time, so their feed content is shown as-is. Should extraction still run on a Tumblr page, the notes list and tracking pixels are stripped before extraction as a safeguard.
  • Marking a feed or folder read from the sidebar now refreshes the whole sidebar, so counts that share those articles (labels, other feeds) update right away instead of going stale until the next reload.
  • Readable extraction removes duplicate images: some news sites emit the same photo as several renditions (lead, inline, responsive) and each was extracted, so the body showed the same picture two or three times. Matching on the image filename now collapses them to a single copy.
  • Admin → Feeds: on narrow screens the "Feed" column no longer collapses to a couple of unreadable characters. As the table grew it stopped fitting the panel, and the flexible feed column shrank to nothing while the rest scrolled; it now keeps a readable minimum width and the table scrolls horizontally instead.
  • Admin → Feeds: a host group that contains a failing feed no longer paints its entire header red, which overstated severity and blended into the group separator; only the host name turns red now.
  • Scheduled fetching no longer crashes while arming per-host pacing: after a fetch committed, the scheduler re-read the feed's URL off an expired ORM object, which raised a MissingGreenlet error, failed the fetch, and skipped arming the adaptive per-host spacing. The spacing was therefore never enforced between same-host feeds, so hosts with several feeds (Reddit, YouTube) were fetched in bursts and more likely to answer 403. The URL is now captured before the fetch, so pacing is armed as intended.
  • A feed marked errored in the sidebar now clears its red indicator immediately when a manual refresh succeeds, instead of staying red until the next full page reload. The refresh only swapped the unread count, which sits apart from the error marker; it now updates the marker out-of-band from the feed's fresh status.

Version 0.12.0 · 2026-07-07

Added

  • Copy article action on the article ··· menu (desktop) and the bottom action bar (mobile). It copies the title, source and body to the clipboard as both rich HTML and plain text, so it pastes with formatting and images into rich editors and as clean text everywhere else. Relative image and link URLs are rewritten to absolute so they still resolve after pasting.
  • The Stats backlog cards (labeled and starred) now link straight into the matching reader view, so you can jump from a count to the actual articles.
  • Admin → Feeds: a "Rate limits" view (shown only when some exist) lists the fetch pace Readfine has learned per host, with the host, its spacing, and how and when that was learned, each with a Clear action. Errored feeds in the admin table also show their predicted next fetch, and a feed with no per-feed interval override shows the effective default instead of a blank dash.
  • Admin → Feeds: a "By host / A-Z" toggle that groups the feed list by fetch host, so all of a site's feeds (say, every Reddit feed) sit under one host header with a count instead of a flat alphabetical list. Hosts sort alphabetically, single-feed hosts fall into an "Other" bucket, and any host holding an errored feed floats to the top so problems stay visible. Within a group, and in the flat list, feeds now sort by status first (errored, disabled, paused, active) and then by name. The choice is remembered per browser.

Changed

  • A single HTTP 403 no longer disables a feed. Reddit and YouTube return 403 as a transient anti-bot or rate-adjacent block (datacenter IP, generic user-agent) far more often than as a permanent denial, so 403 now backs off through the error tier like 408/429 and 5xx and only disables after several consecutive failures. Genuinely permanent 4xx (400, 401, 404, 410) still disable immediately.
  • The fetcher reads rate-limit headers (Retry-After, RateLimit-*, X-RateLimit-*) on both successful and 429 responses and applies a per-host cooldown. Once a host reports its budget is spent (for example Reddit's x-ratelimit-remaining: 0), other feeds on that host wait out the reset instead of hammering it with more 429s. The waiting happens inside the fetch round, up to a budget that keeps the round short enough not to miss the next slot; anything over that defers to the next round. Feeds on other hosts still fetch in parallel.
  • Manually refreshing a feed (the sidebar ↻ and the admin "force fetch") now respects a known rate-limit window instead of firing straight into another 429. While the host is cooling down it shows "Rate-limited, try again in …" (seconds or minutes, from the server's reset headers). A bare 403 anti-bot block is treated differently: only the background scheduler paces itself on those, since a manual retry often succeeds.
  • Readfine learns a sustainable fetch pace per host and spaces its requests accordingly, rather than only reacting once a host reports its budget already spent. It reads the pace from a host's rate-limit headers on successful responses and tightens it when the host keeps answering 429. The learned pace only ever tightens, so it never oscillates, and is capped so a feed can't stall forever. This keeps aggressive hosts like Reddit, where a burst of same-host feeds fetched back to back would trip a 403/429, from being throttled. The pace is stored and survives restarts and deploys, so a host isn't re-probed into a rate limit on every restart, and manual refreshes respect it too.
  • Feeds are fetched at whichever of the four 15-minute ticks (:00/:15/:30/:45) first follows their interval, instead of being pinned to the top of the hour. This spreads load across the hour on each feed's own phase rather than piling every hourly feed onto :00, and it improves freshness: an hourly feed first fetched a few minutes past the hour used to wait until the next :00 (up to ~2 h between fetches) and now refreshes about an hour later as intended. Feeds that miss a tick (a host cooldown, a transient error, a restart mid-round) recover at the next tick instead of waiting a full interval.
  • The per-feed refresh button (↻) now reloads the article list when you're viewing that feed, so newly fetched articles appear right away instead of only after re-selecting it. Refreshing a feed you're not viewing still just updates its unread badge.
  • The fetch-interval selector spells out the server default next to the "Default" option (for example "Default (60 min)") on the subscribe, scrape-setup and feed-edit forms, and wraps better on narrow screens.
  • Adding a feed now shows specific messages for rate-limiting (429, including when to retry) and temporary server errors (5xx) instead of a bare "HTTP error {status}".
  • The Feeds, Filters and Labels settings pages and the admin Users page show an item count next to the heading, kept current as items are added or removed without a reload.
  • Labels and filters sort case-insensitively everywhere now: settings lists, label pickers and chips. Previously the database collation put all uppercase names before any lowercase one, so a new lowercase label or filter got stuck at the end of the list.
  • The filter list shows a "priority N" badge on filters whose priority isn't the default, making it clear why a filter sorts and runs ahead of alphabetical order.
  • Briefings sent to extra recipients now put the account owner in To: and the additional recipients in Bcc, so co-subscribers no longer see each other's addresses. The modal also notes that delivery can lag the scheduled time by up to 15 minutes (the scheduler tick).
  • The admin "force fetch" button shows a spinner and blocks double-clicks while the synchronous fetch runs, instead of looking like it did nothing for several seconds.
  • Settings → AI cost estimates now cover the current Anthropic, OpenAI and Google model families. A configured model that isn't in the built-in price list is estimated from a typical model for its provider (shown with a "~" and a note under the table) instead of showing as unknown or zero.
  • The Trend column in the AI cost table tracks estimated cost rather than raw operation count, and the Fast/Quality/Total rows show a trend too (previously blank), so the arrows reflect what actually moves your spend, such as longer articles costing more at the same number of runs.

Fixed

  • Collapsing and expanding the sidebar is instant now. It used to refetch the whole sidebar from the server on every toggle, so the old layout lingered, briefly squished into the new width, until the request returned. Both the collapsed rail and the full sidebar are rendered up front and the toggle just switches between them in the browser, with no round-trip. On the mobile "collapsible" sidebar, opening the overlay no longer reflows the article-list text either, because the rail is a fixed strip now and the list keeps a constant width.
  • Toast notifications (a feed's error when you open it, or a manual refresh result) no longer render at roughly half-width on mobile. They stretch edge to edge with a small gutter on narrow screens and stay centred with a sensible max-width on wider ones.
  • Filters that share a priority run in a fixed order now, exactly the order the Settings → Filters list shows (priority, then name). Their order used to be left to the database, so a "stop on match" filter could behave differently between fetches.
  • The green "Feed added successfully" banner no longer reappears when you refresh the Feeds settings page after subscribing.
  • Articles that carry a label stay visible in their label view even after their feed is deleted or unsubscribed. The view used to inner-join the feed and hide them, leaving the sidebar badge counting an apparently empty category.
  • The article-list loading overlay matches the neutral dark-mode background instead of a blue-tinted grey, and is delayed slightly so quick cached loads don't flash a spinner.
  • The AI cost table's total is no longer quietly understated when a model slot uses a model missing from the price list. That slot used to be added as $0, making the total look complete while dropping part of the cost.

Security

  • Filter regex conditions run under a per-match timeout now, closing a denial-of-service hole. The old create-time heuristic could be bypassed by a catastrophic-backtracking pattern (for example ([a-z]+)*), and because matching ran synchronously on the event loop during fetch, filter tests and retroactive apply, and CPython's re neither times out nor releases the GIL, a single crafted filter could freeze the whole app for every user. Evaluation now uses the regex module with a hard timeout; a timed-out pattern counts as no match, and normal filter behaviour is unchanged.

Version 0.11.0 · 2026-06-30

Added

  • Search is now also a filter view: alongside the text query you can scope to feeds/folders, filter by labels (any / specific) and read status (all / unread / read), and choose the sort (relevance / newest / oldest). Leaving the text empty applies the filters on their own. Search moved from the user menu to an icon in the sidebar.
  • Feeds are now fetched conditionally: Readfine remembers each feed's ETag / Last-Modified and sends them back on the next poll, so an unchanged feed answers 304 Not Modified with no body and the download and parse are skipped entirely. Less bandwidth, and lighter on rate-limited sites.
  • Catch me up & briefings now have a dedicated label filter (any label / specific labels, OR) shown alongside the feed scope, replacing the old "Labeled only" relevance radio. The minimum-score filter is now an independent toggle (shown only when scoring is configured) rather than bundled with labels. The "Since yesterday" period is now labelled "Yesterday+".
  • Adding a feed lets you set its fetch interval from the subscribe form, and owners of a private or solely-subscribed feed can change the interval when editing it. Shared public feeds show the interval read-only (only an admin can change it).
  • Errored feeds now show when they will next be retried, both on the feed list and the feed detail page; a feed auto-disabled after repeated failures says so explicitly instead of leaving the next fetch ambiguous.

Changed

  • Switching between sections (Starred, Labeled, folders, feeds) now shows a brief loading overlay over the article list, so the sidebar highlight no longer appears to change before the list it points at has loaded.
  • Creating, renaming, or deleting a folder immediately updates the folder dropdown in the add-feed form without a page reload.

Fixed

  • Feeds where every item points at one shared link (e.g. a podcast whose episodes all link to the show page) no longer have every new item after the first silently dropped as a duplicate; items are now de-duplicated by links that actually identify a single item, falling back to the unique GUID otherwise.

  • Reddit (and similar) article content built from a header-less layout table no longer overflows the reading panel: such tables now stack the image above the text, images are constrained to the column width, and genuine data tables scroll horizontally instead of overflowing.

  • Text search combined with a read-status filter no longer skips results while scrolling: mark-as-read-on-scroll is disabled for that specific case (where it shifted the offset-paginated result set), leaving plain search and the filter view unaffected.

  • A feed returning HTTP 429 (Too Many Requests) is no longer disabled on the first hit. 429 and 408 are now treated as transient: the feed backs off via the normal error tier and is only disabled after the usual run of consecutive failures. When the server sends a Retry-After header, the scheduler waits at least that long before re-fetching.

  • Adding a feed now costs a single network request instead of up to three. The "Test" step caches the fetched feed briefly and Subscribe reuses it for both the title and the initial article import, so rate-limited sites (e.g. Reddit) no longer return 429 mid-subscribe.

  • When several feeds share a host (e.g. multiple Reddit subreddits), a scheduled fetch no longer requests them all at once. Requests to a given host are now serialized within a fetch round (different hosts still run in parallel), which flattens the burst that made some of those feeds return HTTP 429.

  • Readable extraction that returns no usable content, e.g. a Reddit article page that serves a bot-verification wall (HTTP 200) instead of the article, is no longer saved as a blank "successful" extraction that rendered an empty body. Such articles now show their original feed content, and a feed whose pages keep extracting nothing auto-disables full-content extraction after repeated empties (the same way persistent HTTP 403 blocks already did) instead of re-fetching every page forever.

  • The auto-disabled notice for full-content extraction now states why it was turned off (the feed already delivers full articles, or the site blocked extraction / returned no readable content) instead of always claiming the site blocked it.

  • The article view no longer flickers an endless "Extracting full content…" spinner for an article whose extraction failed and is waiting to retry; it shows the feed content quietly, and the spinner appears only while a first attempt is in flight.

  • "Extract full content" from the article menu no longer momentarily drops the article's star, archive, or label state from the action bar.

Version 0.10.1 · 2026-06-27

Added

  • One-command local demo: docker compose -f docker-compose.demo.yml up brings the app up on http://localhost:8000 with a seeded admin and no setup wizard, for trying it out before a full install. Demo only: plain HTTP, DEBUG=true, and hard-coded throwaway secrets; not for production. See README → Quick demo.

Fixed

  • Infinite scroll in unread/label views could stop early or silently skip articles when rows were marked read while scrolling (the unread set shrank under the numeric page offset). The article list now uses keyset (cursor) pagination, so scrolling reliably loads every remaining article regardless of mark-read-on-scroll.
  • Mobile: the active tab in the horizontal side-nav strip now scrolls into view on load, instead of staying off-screen when the strip was left scrolled elsewhere.
  • Docker: the db healthcheck now probes the actual database (pg_isready -d), so a DB_USER that differs from DB_NAME no longer logs a Postgres FATAL on every check.

Version 0.10.0 · 2026-06-25

Added

  • In-app feedback / bug report: a "Send feedback" item in the user menu opens a form (type, subject, message) that emails all admins via the configured SMTP, with Reply-To set to the sender's account email. Off by default; admins enable it in Admin → Settings (requires SMTP).
  • AI error badge: a red dot on the user menu and the Settings → AI nav item when a background AI call (e.g. scoring) last failed, so credit/quota errors are visible without opening Settings. Self-clears on the next successful AI call, or dismiss it manually via the × on the error panel in Settings → AI.
  • Filter action archive: alongside label / mark-as-read / star, a filter can now archive matching articles (removes them from the inbox and exempts them from retention purge). Available in Settings → Filters and via OPML round-trip.

Changed

  • Stats: the single "Backlog" figure is split into labeled backlog (unread items carrying a label) and starred backlog (your read-later pile); both are now all-time rather than capped at 90 days. Reading streak, per-day reads and the most active hour are computed in your own timezone instead of UTC.
  • OPML import: Tiny Tiny RSS filter scope (feed / category) is now matched by name and mapped to the corresponding Readfine feed/folder scope, instead of being dropped and imported as global. Mixed scoped/global filters still import as global with a warning.
  • Admin → Settings: the SMTP test now shows the underlying error detail on failure, making misconfiguration easier to diagnose.

Fixed

  • Stats: corrected the engagement funnel bars.
  • Mobile: the collapsible sidebar reliably reappears after a refresh instead of occasionally staying hidden.
  • Favicon: app pages now declare the raster apple-touch-icon, so Firefox/Android use it for link previews and home-screen tiles instead of rasterizing the SVG.

Security

  • Migrated JWT handling from the unmaintained python-jose to PyJWT.

Version 0.9.0 · 2026-06-20

First public release. Self-hosted RSS reader with:

  • RSS/Atom feeds and web-scraping feeds (CSS selectors), folders, scheduled fetching
  • Readable extraction (trafilatura → readability-lxml fallback)
  • 3-panel reading UI (HTMX + Tailwind), article states, labels, dark mode
  • Filters (conditions → actions, regex, AND/OR, feed/folder scoping) with retroactive apply
  • AI summaries, relevance scoring, chat over articles, and Catch me up & briefings (Anthropic / OpenAI / Gemini, bring-your-own-key)
  • Per-user settings, admin panel, SMTP, API tokens (JWT), tiered retention/purge
  • OPML import/export, including web-scraping feeds (round-trips via custom outline attributes) and Tiny Tiny RSS compatibility
  • /healthz endpoint (lightweight DB ping, GET + HEAD) for uptime/monitoring probes
  • backup.sh: off-site PostgreSQL backups via pg_dump + restic (encrypted, deduplicated, retention), with a Cloudflare R2 example config. See README → Backups.