hydrus/docs/changelog.md

title
Changelog

changelog

!!! note This is the new changelog, only the most recent builds. For all versions, see the old changelog.

Version 515

related tags

• I worked on last week's related tags algorithm test, bringing it up to usable standard. the old buttons now use the new algorithm exclusively. all users now get 'related tags' showing in manage tags by default (if you don't like it, you can turn it off under options->tag suggestions)
• the new algorithm has new cancel tech and does a 'work for 600ms' kind of deal, like the old system, and the last-minute blocks from last week are gone--it will search as much as it has time for, including partial results. it also won't lag you out for thirty seconds (unless you tell it to in the options). it searches tags with low count first, so don't worry if it doesn't get to everything--'1girl' usually doesn't have a huge amount extra to offer once everything else has run
• it also uses 'hydev actually thought about this' statistical sampling tech to work massively faster on larger-count tags at the cost of some variance in rank and the odd false positive (considered sufficiently related when it actually shouldn't meet the threshold) nearer the bottom end of the tags result list
• rather than 'new 1' and 'new 2', there is now an on/off button for searching your local files or all known files on tag repositories. 'all known files' = great results, but very slow, which the tooltip explains
• there's also a new status label that will tell you when it is searching and how well the search went (e.g. '12/51 tags searched fully in 459ms')
• I also added the 'quick' search button back in, since we can now repeat searches for just selections of tags
• I fixed a couple typos in the algorthim that were messing some results
• in the manage tags dialog, if you have the suggested tag panels 'side-to-side', they now go in named boxes
• in the manage tags dialog, if you have suggested tag panels in a notebook, 'related tags' will only refresh its search on a media change event (including dialog initialisation) when it is the selected page. it won't lag you from the background!
• options->tag suggestions now lets you pick which notebook'd tag suggestions page you want to show by default. this defaults to 'related'
• I have more plans here. these related tags results are very cachable, so that's an obvious next step to speed up results, and when I have done some other long-term tag improvements elsewhere in the program, I'll be able to quickly filter out unhelpful sibling and parent suggestions. more immediately, I think we'll want some options for namespace weighting (e.g. 'series:' tags' suggestions could have higher rank than 'smile'), so we can tune things a bit

misc

• the 'open externally' canvas widget, which shows any available thumbnail of the flash or psd or whatever, now sizes itself correctly and draws the thumbnail nicely if you set the new thumbnail supersampling option to >100%. if your thumbnail is the wrong size (and probably in a queue to be regenerated soon), I think it'll still make the window too big/small, but it'll draw the thumbnail to fit
• if a tag content update comes in with an invalid tag (such as could happen with sidecars recently), the client now heals better. the bad tag is corrected live in more places, and this should be propagated to the UI. if you got a warning about 'you have invalid tags in view' recently but running the routine found no problems, please reboot, and I think you'll be fixed. I'm pretty sure the database wasn't being damaged at all here (it has cleaning safeguards, so it shouldn't be possible to actually save bad tags)--it was just a thing to do with the UI not being told of the cleaned tag, and it shouldn't happen again. thank you for the reports! (issue #1324)
• export folders and the file maintenance dialog no longer apply the implicit system:limit (defaults to max 10k files) to their searches!
• old OR predicates that you load with saved searches and similar should now always have alphebetised components, and if you double-click them to remove them, they will now clear correctly (previously, they were doing something similar to the recent filetype problem, where instead of recognising themselves and deleting, they would instead duplicate a normalised (sorted) copy of themselves)
• thanks to a user, updated the recently note-and-ai-updated pixiv parser again to grab the canonical pixiv URL and translated tags, if present
• thanks to a user, updated the sankaku parser to grab some more tags
• the file location context and tag context buttons under tag autocompletes now put menu separators between each type of file/tag service in their menus. for basic users, this'll be a separator for every row, but for advanced users with multiple local domains, it will help categorise the list a bit

Version 514

downloaders

• twitter took down the API we were using, breaking all our nice twitter downloaders! argh!
• a user has figured out a basic new downloader that grabs the tweets amongst the first twenty tweets-and-retweets of an account. yes, only the first twenty max, and usually fewer. because this is a big change, the client will ask about it when you update. if you have some complicated situation where you are working on the old default twitter downloaders and don't want them deleted, you can select 'no' on the dialog it throws up, but everyone else wants to say 'yes'. then check your twitter subs: make sure they moved to the new downloader, and you probably want to make them check more frequently too.
• given the rate of changes at twitter, I think we can expect more changes and blocks in future. I don't know whether nitter will be viable alternative, so if the artists you like end up on a nice simple booru anywhere, I strongly recommend just moving there. twitter appears to be explicitly moving to non-third-party-friendly
• thanks to a user's work, the 'danbooru - get webm ugoira' parser is fixed!
• thanks to a user's work, the deviant art parser is updated to get the highest res image in more situations!
• thanks to a user's work, the pixiv downloader now gets the artist note, in japanese (and translated, if there is one), and a 'medium:ai generated' tag!

sidecars

• I wrote some sidecar help here! https://hydrusnetwork.github.io/hydrus/advanced_sidecars.html
• when the client parses files for import, the 'does this look like a sidecar?' test now also checks that the base component of the base filename (e.g. 'Image123' from 'Image123.jpg.txt') actually appears in the list of non-txt/json/xml ext files. a random yo.txt file out of nowhere will now be inspected in case it is secretly a jpeg again, for good or ill
• when you drop some files on the client, the number of files skipped because they looked like sidecars is now stated in the status label
• fixed a typo bug that meant tags imported from sidecars were not being properly cleaned, despite preview appearance otherwise, for instance ':)', which in hydrus needs to be secretly stored as '::)' was being imported as ')'
• as a special case, tags that in hydrus are secretly '::)' will be converted to ':)' on export to sidecar too, the inverse of the above problem. there may be some other tag cleaning quirks to undo here, so let me know what you run into

related tags overhaul

• the 'related tags' suggestion system, turned on under options->tag suggestions, has several changes, including some prototype tech I'd love feedback on
• first off, there are two new search buttons, 'new 1' and 'new 2' ('2' is available on repositories only).. these use an upgraded statistical search and scoring system that a user worked on and sent in. I have butchered his specific namespace searching system to something more general/flexible and easy for me to maintain, but it works better and more comprehensibly than my old method! give it a go and let me know how each button does--the first one will be fast but less useful on the PTR, the second will be slower but generally give richer results (although it cannot do tags with too-high count)
• the new search routine works on multiple files, so 'related tags' now shows on tag dialogs launched from a selection of thumbnails!
• also, all the related search buttons now search any selection of tags you make!!! so if you can't remember that character's name, just click on the series or another character they are often with and hit the search, and you should get a whole bunch appear
• I am going to keep working on this in the future. the new buttons will become the only buttons, I'll try and mitigate the prototype search limitations, add some cancel tech, move to a time-based search length like the current buttons, and I'll add more settings, including for filtering so we aren't looking up related tags for 'page:x' and so on. I'm interested in knowing how you get on with IRL data. are there too many recommendations (is the tolerance too high?)? is the sorting good (is the stuff at the top relevant or often just noise?)?

misc

• all users can now copy their service keys (which are a technical non-changing hex identifier for your client's services) from the review services window--advanced mode is no longer needed. this may be useful as the client api transitions to service keys
• when a job in the downloader search log generates new jobs (e.g. fetches the next page), the new job(s) are now inserted after the parent. previously, they were appended to the end of the list. this changes how ngugs operate, converting their searches from interleaved to sequential!
• restarting search log jobs now also places the new job after the restarted job
• when you create a new export folder, if you have default metadata export sidecar settings from a previous manual file export, the program now asks if you want those for the new export folder or an empty list. previously, it just assigned the saved default, which could be jarring if it was saved from ages ago
• added a migration guide to the running from source help. also brushed up some language and fixed a bunch of borked title weights in that document
• the max initial and periodic file limits in subscriptions is now 50k when in advanced mode. I can't promise that would be nice though!
• the file history chart no longer says that inbox and delete time tracking are new

misc fixes

• fixed a cursor type detection test that was stopping the cursor from hiding immediately when you do a media viewer drag in Qt6
• fixed an issue where 'clear deletion record' calls were not deleting from the newer 'all my files' domain. the erroneous extra records will be searched for and scrubbed on update
• fixed the issue where if you had the new 'unnamespaced input gives (any namespace) wildcard results' search option on, you couldn't add any novel tags in WRITE autocomplete contexts like 'manage tags'!!! it could only offer the automatically converted wildcard tags as suggested input, which of course aren't appropriate for a WRITE context. the way I ultimately fixed this was horrible; the whole thing needs more work to deal with clever logic like this better, so let me know if you get any more trouble here
• I think I fixed an infinite hang when trying to add certain siblings in manage tag siblings. I believe this was occuring when the dialog was testing if the new pair would create a loop when the sibling structure already contains a loop. now it throws up a message and breaks the test
• fixed an issue where certain system:filetype predicates would spawn apparent duplicates of themselves instead of removing on double-click. images+audio+video+swf+pdf was one example. it was a 'all the image types' vs 'list of (all the) image types' conversion/comparison/sorting issue

client api

• this is later than I expected, but as was planned last year, I am clearing up several obsolete parameters and data structures this week. mostly it is bad service name-identification that seemed simple or flexible to support but just added maintenance debt, induced bad implementation practises, and hindered future expansions. if you have a custom api script, please read on--and if you have not yet moved to the alternatives, do so before updating!
• all ...service_name... parameters are officially obsolete! they will still work via some legacy hacks, so old scripts shouldn't break, but they are no longer documented. please move to the ...service_key... alternates as soon as reasonably possible (check out /get_services if you need to learn about service keys)
• /add_tags/get_tag_services is removed! use /get_services instead!
• hide_service_names_tags, previously made default true, is removed and its data structures service_names_to_statuses_to_... are also gone! move to the new tags structure.
• hide_service_keys_tags is now default true. it will be removed in 4 weeks or so. same deal as with service_names_to_statuses_to_...--move to tags
• system_inbox and system_archive are removed from /get_files/search_files! just use 'system:inbox/archive' in the tags list
• the 'set_file_relationships' command from last week has been reworked to have a nicer Object parameter with a new name. please check the updated help! normally I wouldn't change something so quick, but we are still in early prototype, so I'm ok shifting it (and the old method still works lmao, but I'll clear that code out in a few weeks, so please move over--the Object will be much nicer to expand in future, which I forgot about in v513)
• many Client API commands now support modern file domain objects, meaning you can search a UNION of file services and 'deleted-from' file services. affected commands are
• • /add_files/delete_files
• • /add_files/undelete_files
• • /add_tags/search_tags
• • /get_files/search_files
• • /manage_file_relationships/get_everything
• a new /get_service call now lets you ask about an individual service by service name or service key, basically a parameterised /get_services
• the /manage_pages/get_pages and /manage_pages/get_page_info calls now give the page_state, a new enum that says if the page is ready, initialised, searching, or search-cancelled
• to reduce duplicate argument spam, the client api help now specifies the complicated 'these files' and now 'this file domain' arguments into sub-sections, and the commands that use them just point to the subsections. check it out--it makes sense when you look at it.
• /add_tags/add_tags now raises 400 if you give an invalid content action (e.g. pending to a local tag service). previously it skipped these rows silently
• added and updated unit tests and help for the above changes
• client api version is now 41

boring optimisation

• when you are looking at a search log or file log, if entries are added, removed, or moved around, all the log entries that have changed row # now update (previously it just sent a redraw signal for the new rows, not the second-order affected rows that were shuffled up/down. many access routines for these logs are sped up
• file log status checking is completely rewritten. the ways it searches, caches and optimises the 'which is the next item with x status' queues is faster and requires far less maintenance. large import queues have less overhead, so the in and outs of general download work should scale up much better now
• the main data cache that stores rendered images, image tiles, and thumbnails now maintains itself far more efficiently. there was a hellish O(n) overhead when adding or removing an item which has been reduced to constant time. this gonk was being spammed every few minutes during normal memory maintenance, when hundreds of thumbs can be purged at once. clients with tens of thousands of thumbnails in memory will maintain that list far more smoothly
• physical file delete is now more efficient, requiring far fewer hard drive hits to delete a media file. it is also far less aggressive, with a new setting in options->files and trash that sets how long to wait between individual file deletes, default 250ms. before, it was full LFG mode with minor delays every hundred/thousand jobs, and since it takes a write lock, it was lagging out thumbnail load when hitting a lot of work. the daemon here also shuts down faster if caught working during program shut down

boring code cleanup

• refactored some parsing routines to be more flexible
• added some more dictionary and enum type testing to the client api parameter parsing routines. error messages should be better!
• improved how /add_tags/add_tags parsing works. ensuring both access methods check all types and report nicer errors
• cleaned up the /search_files/file_metadata call's parsing, moving to the new generalised method and smoothing out some old code flow. it now checks hashes against the last search, too
• cleaned up /manage_pages/add_files similarly
• cleaned up how tag services are parsed and their errors reported in the client api
• the client api is better about processing the file identifiers you give it in the same order you gave
• fixed bad 'potentials_search_type'/'search_type' inconsistency in the client api help examples
• obviously a bunch of client api unit test and help cleanup to account for the obsolete stuff and various other changes here
• updated a bunch of the client api unit tests to handle some of the new parsing
• fixed the remaining 'randomly fail due to complex counting logic' potential count unit tests. turns out there were like seven more of them

Version 513

client api

• the Client API now supports the duplicates system! this is early stages, and what I've exposed is ugly and technical, but if you want to try out some external dupe processing, give it a go and let me know what you think! (issue #347)
• a new 'manage file relationships' permission gives your api keys access
• the new GET commands are:
• • /manage_file_relationships/get_file_relationships, which fetches potential dupes, dupes, alternates, false positives, and dupe kings
• • /manage_file_relationships/get_potentials_count, which can take two file searches, a potential dupes search type, a pixel match type, and max hamming distance, and will give the number of potential pairs in that domain
• • /manage_file_relationships/get_potential_pairs, which takes the same params as count and a max_num_pairs and gives you a batch of pairs to process, just like the dupe filter
• • /manage_file_relationships/get_random_potentials, which takes the same params as count and gives you some hashes just like the 'show some random potential pairs' button
• the new POST commands are:
• • /manage_file_relationships/set_file_relationships, which sets potential/dupe/alternate/false positive relationships between file pairs with some optional content merge and file deletes
• • /manage_file_relationships/set_kings, which sets duplicate group kings
• more commands will be written in the future for various remove/dissolve actions
• wrote unit tests for all the commands!
• wrote help for all the commands!
• fixed an issue in the '/manage_pages/get_pages' call where the response data structure was saying 'focused' instead of 'selected' for 'page of pages'
• cilent api version is now 40

boring misc cleanup and refactoring

• cleaned and wrote some more parsing methods for the api to support duplicate search tech and reduce copypasted parsing code
• renamed the client api permission labels a little, just making it all clearer and line up better. also, the 'edit client permissions' dialog now sorts the permissions
• reordered and renamed the dev help headers in the same way
• simple but significant rename-refactoring in file duplicates database module, tearing off the old 'Duplicates' prefixes to every method ha ha
• updated the advanced Windows 'running from source' help to talk more about VC build tools. some old scripts don't seem to work any more in Win 11, but you also don't really need it any more (I moved to a new dev machine this week so had to set everything up again)

Version 512

two searches in duplicates

• the duplicate filter page now lets you search 'one file is in this search, the other is in this search'! the only real limitation is both searches are locked to the same file domain
• the main neat thing is you can now search 'pngs vs jpegs, and must be pixel dupes' super easy. this is the first concrete step towards my plan to introduce an optional duplicate auto resolution system (png/jpeg pixel dupes is easy--the jpeg is 99.9999% always better)
• the database tech to get this working was actually simpler than 'one file matches the search', and in testing it works at ok speed, so we'll see how this goes IRL
• duplicate calculations should be faster in some simple cases, usually when you set a search to system:everything. this extends to the new two-search mode too (e.g. a two-search with one as system:everything is just a one-search, and the system optimises for this), however I also search complicated domains much more precisely now, which may make some duplicate search stuff work real slow. again, let me know!

sidecars

• the txt importer/exporter sidecars now allow custom 'separators', so if you don't want newlines, you can use ', ' or whatever format you need

misc

• when you right-click on a selection of thumbs, the 'x files' can now be 'x videos' or 'x pngs' etc.. as you see on the status bar
• when you select or right-click on a selection of thumbs that all have duration, the status bar and menu now show the total duration of your selection. same deal on the status bar if you have no selection on a page of only durating-having media
• thanks to the user who figured out the correct render flag, the new 'thumbnail ui-scale supersampling %' option now draws non-pixelly thumbs on 100% monitors when it is set higher (e.g. 200% thumbs drawing on 100% monitor), so users with unusual multi-monitor setups etc... should have a nicer experience. as the tooltip now says, this setting should now be set to the largest UI scale you have
• I removed the newgrounds downloader from the defaults (this only affects new users). the downloader has been busted for a while, and last time I looked, it was not trivial to figure out, so I am removing myself from the question
• the 'manage where tag siblings and parents apply' dialog now explicitly points users to the 'review current sync' panel

client api

• a new command, /manage_pages/refresh_page, refreshes the specified page
• the help is updated to talk about this
• client api version is now 39

server management

• in the 'modify accounts' dialog, if the null account is checked when you try to do an action, it will be unchecked. this should stop the annoying 400 Errors when you accidentally try to set it something
• also, if you do 'add to expires', any accounts that currently do not expire will be deselected before the action too, with a brief dialog note about it

other duplicates improvements

• I reworked a ton of code here, fixing a heap of logic and general 'that isn't quite what you'd expect' comparison selection issues. ideally, the system will just make more obvious human sense more often, but this tech gets a little complicated as it tries to select comparison kings from larger groups, and we might have some situations where it says '3 pairs', but when you load it in the filter it says 'no pairs found m8', so let me know how it goes!
• first, most importantly, the 'show some random potential pairs' button is vastly improved. it is now much better about limiting the group of presented files to what you specifically have searched, and the 'pixel dupes' and 'search distance' settings are obeyed properly (previously it was fetching too many potentials, not always limiting to the search you set, and choosing candidates from larger groups too liberally)
• while it shows smaller groups now, since they are all culled better, it should select larger groups more often than before
• when you say 'show some random potential pairs' with 'at least one file matches the search', the first file displayed, which is the 'master' that the other file(s) are paired against, now always matches the search. when you are set to the new two-search 'files match different searches', the master will always match the first search, and the others of the pairs will always match the second search. in the filter itself, some similar logic applies, so the files selected for actual comparison should match the search you inputted better.
• setting duplicates with 'custom options' from the thumbnail menu and selecting 'this is better' now correctly sets the focused media as the best. previously it set the first file as the best
• also, in the duplicate merge options, you can now set notes to 'move' from worse to better
• as a side thing, the 'search distance' number control is now disabled if you select 'must be pixel dupes'. duh!

boring cleanup

• refactored the duplicate comparison statement generation code from ClientMedia to ClientDuplicates
• significantly refactored all the duplicate files calculation pipelines to deal with two file search contexts
• cleaned up a bunch of the 'find potential duplicate pairs in this file domain' master table join code. less hardcoding, more dynamic assembly
• refactored the duplicated 'figure out pixel dupes table join gubbins' code in the file duplicates database module into a single separate method, and rolled in the base initialisation and hamming distance part into it too, clearing out more duplicated code
• split up the 'both files match' search code into separate methods to further clean the logic here
• updated the main object that handles page data to the new serialisable dictionary, combining its hardcoded key/primitive/serialisable storage into one clean dict that looks after itself
• cleaned up the type definitions of the the main database file search and fixed the erroneous empty set returns
• I added a couple unit tests for the new .txt sidecar separator
• fixed a bad sidecar unit test
• 'client_running' and 'server_running' are now in the .gitignore

Version 511

thumbnail UI scaling

• thumbnails can finally look good at high UI scales! a new setting in options->thumbnails, 'Thumbnail UI scale supersampling %', lets you tell hydrus to generate thumbnails at a particular UI scale. match it to your monitor, and your thumbnails should regenerate to look crisp
• some users have complicated multi-monitor setups, or they change their UI scale regularly, so I'm not auto-setting this yet. let me know how it goes
• sadly <100% for super-crunchy-mode doesn't work

unnamespaced search tags

• I am not really happy with this solution, since it doesn't neatly restore the old behaviour, but it does make things easier in the new system and I've fixed a related bug
• a new option in services->manage tag display and search, 'Unnamespaced input gives (any namespace) wildcard results', now lets you quickly search *:sam* by typing sam
• fixed an issue where an autocomplete input with a total wildcard namespace, like *:sam was not matching to unnamespaced tags when preparing the list of tag results
• wildcards with * namespace now have a special (any namespace) suffix, and they show with unnamespaced namespace colour

misc

• fixed the client-server communication problem related to last week's SerialisableDictionary update. I messed up and forgot this object is used in network comms, which meant >=v510 clients couldn't talk to a <=509 server and vice versa version swaps. now the server always kicks out an old SerialisableDictionary serialisation. I plan to remove the patch in 26 weeks, giving us more buffer time for users to update naturally
• the recent option to turn off mouse-scroll-changes-menu-button-value is improved--now the wheel event is correctly passed up to the parent panel, so you'll scroll right through one of these buttons, not halt on it. the file sort control now also obeys this option
• if you try to zoom a media in so that its virtual size would be >32,000px on a side, the canvas now zooms to 32k exactly. this is the max allowed zoom for technical reasons atm (I'll fix it in a future rewrite). this also fixes the 'zoom max' command, which previously would make no action if the max zoom created a virtual canvas bigger than this. also, 'zoom max' is now shown on the media viewer right-click menu
• the 'max zoom' dimension for mpv windows and my native animation window is now 8k. seems like there are smaller technical limits for mpv, and my animation window isn't tiled, so this is to be extra safe for now
• fixed a bug where it was possible to send the 'undelete file' signal to a file that was physically deleted (and therefore viewed in a special 'deleted files' domain). the file would obediently return to its original local file service and then throw 'missing file' warnings when the thumb tried to show. now these files are discarded from undelete consideration
• if you are looking at physically deleted files, the thumbnail view now provides a 'clear deletion record' menu action! this is the same command as the button in services->review services->all local files, but just on the selection
• fixed several taglists across the program that were displaying tags in the wrong display context and/or not sorting correctly. this mostly went wrong by setting sorted storage taglists (which normally show sibling/parent flare) as unsorted display taglists
• file lookup script tag suggestions (as fetched from some external source) are now set to be sorted

file import options pre-import checking

• this stuff is advanced users only. normal users can rest assured that the way the client skips downloads for 'already in db/previously deleted' files now has fewer false negatives and false positives
• the awkwardly named advanced 'do not check url/hash to see if file already in db/previously deleted' checkboxes in file import options have been overhauled. now they are phrased in the positive ("check x to determine aid/pd?") and offer 'do not check', 'check', and the new 'check - and matches are dispositive'. the tooltip has been updated to talk about what they do. 'dispositive' basically means 'if this one hits, trust it over the other', and by default the 'hash' check remains dispositive over the URLs (this was previously hardcoded, now you can choose urls to rule in some cases).
• there is also a new checkbox to optionally disable a component of the url checking that looks at neighbouring urls on the same file to determine url-mapping trustworthiness. this will solve or help explore some weird multi-url-mapping situations
• also, novel SHA256 hashes no longer count as 'matches', just like a novel MD5 hash would not. this helps keep useful dispositive behaviour for known hashes but also automatically defers to urls when a site is being CDN-optimised and transfer hashes are different to api-reported ones. this fixes some watchers that have been using excess bandwidth on repeated downloads
• fixed several problems with the url-lookup logic, particularly with the method that checks for 'file-neighbour' urls (simply, when a file-url match should be distrusted because that file has multiple urls of the same url class). it was also too aggressive on file/unknown url classes, which can legitimately have tokenised neighbours, and getting confused by http/https dupes
• the neighbour test now remembers untrustworthy domains across different url checks for a file, which helps some subsequent direct-file-url checks where neighbours aren't a marker of file-url mapping reliability
• the overall logic behind the hash and url lookup is cleaned up significantly
• if you are an advanced user who has been working with me on this stuff, let me know how it goes. we erected this rats' nest through years of patches, and now I have cleaned it out. I'm confident it works better overall, but I may have missed one of your complicated situations. at the least, these new options should help us figure out quicker fixes in future

boring code cleanup

• removed some old 'subject_identifier' arg parsing from various account-modification calls in the server code. as previously planned, for simplicity and security, the only identifier for these actions is now 'subject_account_key', and subject_identifier is only used for account lookups
• improved the error handling around serialised object loading. the messages explain what happened and state object type and the versions involved
• cleaned up some tag sort code
• cleaned up how advanced file delete content updates work
• fixed yet another duplicate potentials count unit test that was sometimes failing due to complex count perspective

Version 510

notes

• duplicate metadata merge options now supports note merging. you can copy from worse to better or in both directions, with a couple extra conflict-resolution options that are a subset of note import options and have reasonable defaults.
• the default note merge options are to go from worse to better for 'set as better' and both directions for 'they are the same', renaming notes on conflicts. your existing duplicate metadata merge options will receive these settings on update, so if you don't want this, update your settings from the duplicate filter page
• the manage notes dialog gets copy and paste buttons. these will copy all the current notes and paste them to another instance of the panel, using the default (extend if possible, otherwise rename) conflict resolution rules
• if an automatic system like a parser gives a note text that already exists on the file, the Note Import Options now discards it in all cases, no matter the names involved. no more automatic dupes!
• ADVANCED: note import options (and related note add/merge operations that use it) now scan all prefix-matching note names for 'new note is already in file' and 'new note is an extension of a note already in file' tests. this improves a former fix to the 'successive parses of two sites with the same note name but different note text cause one of them to be dupe-added as (2), (3), (4), renames etc...' bug. the initial (1) rename will be scanned and recognised as 'already in file' and ignored or now extended as the settings say, just as if the desired name were hit. thanks to the reports here--I missed the logic the first time around
• it would be nice to have 'manage notes' for multiple files at once--this is still a future goal

notes client api

• the /add_notes/set_notes now takes some new parameters if you want to apply the adapted Note Import Options merge logic rather than figure out renames and extensions yourself
• /add_notes/set_notes now returns the changes it made, which in the new mode may not be exactly what you instructed
• added unit tests and help to reflect the above
• client api version is now 38

misc

• I fixed up how shift/ctrl/drag selection works on taglists. like with the recent thumbnail selection update, you can now 'undo' a shift-select with subsequent clicks or 'drag undo', and the list remembers what was selected beforehand. ctrl-shift-select is also a more reliable 'deselect range'. both mouse drag selection and ctrl-drag selection use this logic, have fewer index bugs, and the ctrl-drag now chooses at the start whether this drag will be selection or deselection based on your initial click that started the drag. have a play with it--overall it just feels better now
• the 'file log' menu now shows a 'reverse' command, which reverses all the imports in the log. if you want to import from oldest to newest with a typical booru, just start your downloader with file imports paused (check the cog icon), and then allow the gallery search to fully populate the list as normaly. once done, hit this new reverse and then unpause the files, and you should be good
• any image files or thumbnails that are completely transparent and have a non-completely-black image now have their alpha channel stripped, just like files that are completely opaque. I believe the instances where this is a mistake outweigh the instances where it is legit, but let me know how we get on--maybe there are some weird mid-gif thumbs or something where this misfires. in the same thing, I reverted the 'psd thumbnails now have no transparency' change from last week. the issue where ffmpeg was sometimes being confused about psd layer masks from earlier should be fixed while letting legit transparency work correctly. the ultimate fix here will be to roll imagemagick into the program, which I am now planning and will start 'running from source' experiments with soon
• the three 'additional fixed time...' settings in options->downloading now have a max value of 3600, for extreme situation testing

boring code cleanup

• updated my serialisabledict/list objects again--they can now handle bytes objects in any position. I will slowly migrate my existing hardcoded bytes serialisation and the old serialisablebytesdict to these freshly flexible classes
• for clarity, across the code, renamed 'duplicate action options' to 'duplicate content merge options'
• refactored duplicate content merge options initialisation, clearing the stuffed init and totuple to nicer get/set
• broke apart how NoteImportOptions does its main note filtering for easier low-level access
• cleaned a ton of note import options code up. the logic here was not great, now it is a bit tidier
• undid whatever nonsense I was doing with taglist ctrl-drag-selection and cleaned up the main click and drag event handling along with its index calculation and 'what was clicked last time' record
• fixed numerous weird logical/position index issues with the taglist and clicking/dragging

Version 509

misc

• added an option 'mouse wheel can "scroll" through menu buttons' to options->gui. this turns off the behaviour where a mouse wheel event over, for instance, the file sort asc/desc button, will change the button's value rather than scrolling the underlying panel. if you found this annoying, you can finally turn it off!
• fixed an annoying 'save service' bug that some users saw last week with the introduction of serverside Tag Filters. some users had an old datatype in their service data storage--a legacy issue--but the system now coerces all datatypes and direct sub-objects to a saveable format on load or update
• the tag washing system now collapses more types of whitespace character to space. mostly this means tab is now converted to space, but some unicode stuff goes too
• the hangul filler character \u3164 is no longer permitted as a namespace or subtag. it can be in longer tags, but isn't allowed on its own (where it appears to be a blank space). (hydev saw one in the wild, probably from some cheeky post title)
• let me know if you run across a newly invalid tag already in your system and the UI goes bananas--ideally hydrus should now catch this and either fix itself or report with a polite note, but let's see. if things go crazy, run database->check and repair->fix invaliid tags
• improved some image transparency detection and slicing logic. it is more accurate and saves more memory now. also, the system that saves thumbnails will more reliably use jpegs when it doesn't need png's transparency
• fixed some PSD thumbs showing a fully transparent transparency layer
• fixed a bug where you could enter capital letters into the namespace colour list in 'tag presentation' options panel
• the default twitter downloaders are all renamed to remove the confusing and technical 'syndication' label
• 'speedcopy' is now an optional supported library. a couple users have suggested this to make network copies on Windows and Linux much faster. I'd like some advanced users who run from source to try adding it to their venvs, and we'll see how it works out IRL in different situations (you can see if it is loaded under help->about)
• if you run from source, the 'advanced' setup route now offers a (t)est Qt install, which sets PySide6 6.4.1 (up from 6.3.21). feel free to try it out--it works well for me, but I want to test it more before trying to roll it to the releases
• in a side thing, thanks to the user who walked me through setting up signed commits to github with my own PGP key. you can see my new key in the contacts help page, id 76249F053212133C, and I am now committing with it. I'm not very familiar with the sheer mechanics of this tech, so bear with me, but I'm pretty sure I can sign or encrypt something if ever needed

macOS build fix

• since v505, many macOS users were unable to boot the built app. it has taken multiple rounds of back and forth with users, but we figured it out. (looks like pyoxidizer updating from 0.22.0 to 0.23.0 simply broke qtpy/Qt bindings, so we force a rollback this week)
• also, the macOS app moves from PySide6 to PyQt6 this week. they are basically the same, but PyQt6 packages into a 258MB dmg, less than half the 548MB PySide6 one!
• let me know if the macOS app gives any more trouble. otherwise, to the people who helped out here, thank you very much for the help!

mostly boring tag filter panel

• removed the 'add' buttons; added 'delete' buttons to the simple whitelist and blacklist panels; added 'block everything' to simple blacklist panel
• the panel now talks about the special sibling and namespace rules when you edit an explicit blacklist-mode-only filter (the tag import options blacklist works this way)
• the 'you didn't need to add that exception' text and 'filter is too complicated for this panel' texts now show/hide rather than waste empty space
• some of the simple-advanced interactions are better, but there's still some logical bork here. mostly stuff like when you hit the 'unnamespace' checkbox in the whitelist panel, it gets needlessly added to the 'except' column in the advanced, rather than just removed from the advanced 'exclude'. I'll fix this up in the near future
• the two namespace checkbox lists are now sized more appropriately
• the white/blacklist panels disable more simply and reliably

boring cleanup

• the confusing 'view this file's duplicates' menu label, which was an artifact of an old submenu label, is removed. if the duplicate menu wants to present the 'view' commands for two locations, it'll title with the respective location, otherwise the commands speak for themselves, no label
• some old 'check(er) timings' nomenclature is renamed to 'checker options' across the board
• the hydrus serialisable dictionary now washes any nested lists or dicts to hydrus serialised equivalents, which should stop situations like the save service bug in future
• the hydrus serialisable list can now handle a mix of hydrus serialisables and python primitives. it also washes its lists or dicts to serialisable equivalents
• improved the data-stability of some image channel slicing
• fixed some PIL fallback thumbnail generation, and improved its 'has transparency' png/jpeg decision-making
• fixed the main thumbnail loader being confused at times about which thumbnail mime to load with. the check I have added is ultra-fast on data we are loading anyway, so we shouldn't notice a difference, but if you get slow thumb loads, let me know
• fixed the media container embed buttons using the file mime rather than the thumb mime when loading thumbnails (again causing transparency issues)
• fixed more generally bad mime handling in the thumbnail generation routine that could have caused more unusual transparency handling for clip, psd, or flash files

Version 508

misc

• added a shortcut action to the 'media' set for 'file relationships: show x', where x is duplicates, potential duplicates, alternates, or false positives, just like the action buried in the thumbnail right-click menu. this actually works in both thumbs and the canvas.
• fixed file deletes not getting processed in the duplicate filter when there were no normal duplicate actions committed in a batch. sorry for the trouble here--duplicate decisions and deletes are now counted and reported in the confirmation dialogs as separate numbers
• as an experiment, the duplicate filter now says (+50%, -33%) percentage differences in the file size comparison statement. while the numbers here are correct, I'm not sure if this is helpful or awkward. maybe it should be phrased differently--let me know
• url classes get two new checkboxes this week: 'do not allow any extra path components/parameters', which will stop a match if the testee URL is 'longer' than the url class's definition. this should help with some difficult 'path-nested URLs aren't matching to the right URL Class' problems
• when you import hard drive files manually or in an import folder, files with .txt, .json, or .xml suffixes are now ignored in the file scanning phase. when hydrus eventually supports text files and arbitrary files, the solution will be nicer here, but this patch makes the new sidecar system nicer to work with in the meantime without, I hope, causing too much other fuss
• the 'tags' button in the advanced-mode 'sort files' control now hides/shows based on the sort type. also, the asc/desc button now hides/shows when it is invalid (filetype, hash, random), rather than disable/enable. there was a bit more signals-cleanup behind the scenes here too
• updated the 'could not set up qtpy/QtCore' error handling yet again to try to figure out this macOS App boot problem some users are getting. the error handling now says what the initial QT_API env variable was and tries to import every possible Qt and prints the whole error for each. hopefully we'll now see why PySide6 is not loading
• cleaned up the 'old changelog' page. all the '.' separators are replaced with proper header tags and I rejiggered some of the ul and li elements to interleave better. its favicon is also fixed. btw if you want to edit 500-odd elements at a time in a 2MB document, PyCharm is mostly great. multi-hundred simultaneous edit hung for about five minutes per character, but multiline regex Find and Replace was instant
• added a link to a user-written guide for running Hydrus on Windows in Anaconda to the 'installing' help
• fixed some old/invalid dialog locations in the 'how to build a downloader' help

client api

• a new /get_files/file_hashes command lets you look up any of the sha256, md5, sha1, sha512 hashes that hydrus knows about using any of the other hashes. if you have a bunch of md5 and want to figure out if you have them, or if you want to get the md5s of your files and run them against an external check, this is now possible
• added help and unit tests for this new command
• added a service enum to the /get_services Client API help
• client api version is now 37
• as a side thing, I rejiggered the 'what non-sha256 hash do these sha256 hashes have?' test here. it now returns a mapping, allowing for more efficient mass lookups, and it no longer creates new sha256 records for novel hashes. feel free to spam this on new sha256 hashes if you like

interesting serverside

• the tag repository now manages a tag filter. admins with 'modify options' permission can alter it under the new menu command services->administrate services->tag repo->edit tag filter.
• any time new tags are pended to the tag repository, they are now washed through the tag filter. any that don't pass are silently discarded
• normal users will regularly fetch the tag filter as long as their client is relatively new. they can review it under a new read-only Tag Filter panel from review services. if their client is super old (or the server), account sync and the UI should fail gracefully
• if you are in advanced mode and your client account-syncs and discovers the tag filter has changed, it will make a popup with a summary of the changes. I am not sure how spammy/annoying this will be, so let me know if you'd rather turn them off or auto-hide after two hours or something
• future updates will have more feedback on manage tags dialog and similar, just to let you know there and then if an entered tag is not wanted. also, admins who change the tag filter will be able to retroactively remove tags that apply to the filter, not just stop new ones. I'd also like some sibling hard-replace to go along with this, so we don't accidentalyl remove tags that are otherwise sibling'd to be good--we'll see
• the hydrus server won't bug out so much at unusual errors now. previously, I ingrained that any error during any request would kick off automatic delays, but I have rejiggered it a bit so this mostly just happens during automatic work like update downloading

boring serverside

• added get/set and similar to the tag repo's until-now-untouched tag filter
• wrote a nice helper method that splays two tag filters into their added/changed/deleted rules and another that can present that in human-readable format. it prints to the server log whenever a human changes the tag filter, and will be used in future retroactive syncing
• cleaned up how the service options are delivered to the client. previously, there would have been a version desync pain if I had ever updated the tag filter internal version. now, the service options delivered to the client are limited to python primitives, atm just update period and nullification period, and tag filter and other complex objects will have their own get calls and fail in quiet isolation
• I fixed some borked nullification period initialisation serverside
• whenever a tag filter describes itself, if either black or whitelist have more than 12 rules, it now summarises rather than listing every single one

Version 507

misc

• fixed an issue where you could set 'all known tags' in the media-tag exporter box in the sidecars system
• if a media-tag exporter in the sidecars system is set to an invalid (missing) tag service, the dialog now protests when you try to OK it. also, when you boot into this dialog, it will now moan about the invalid service. also, new media-tag exporters will always start with a valid local tag service.
• Qt import error states are handled better. when the client boots, the various 'could not find Qt' errors at different qtpy and QtCore import stages are now handled separately. the Qt selected by qtpy, if any, is reported, as is the state of QT_API and whether hydrus thought it was importable. it seems like there have been a couple of users caught by something like system-wide QT_API env variables here, which this should reveal better in boot-crash logs from now on
• all the new setup scripts in the base directory now push their location as the new CWD when they start, and they pop back to your original when they exit. you should be able to call them from anywhere now!
• I've written a 'setup_desktop.sh' install script for Linux users to 'install' a hydrus.desktop file for the current install location to your applications directory. thanks to the user who made the original hydrus.desktop file for the help here
• I fixed the focus when you open a 'edit predicate' panel that only has buttons, like 'has audio'/'no audio'. top button should have focus again, so you can hit enter quick
• added updated link to hydownloader on the client api page

dupes apply better to groups of thumbs

• tl;dr: when the user sets a 'copy both ways' duplicate file status on more than two thumbnails, the duplicate metadata merge options are applied better now
• advanced explanation: previously, all merge updates were calculated before applying the updates, so when applied to a group of interconnected relationships, the nodes that were not directly connected to each other were not syncing data. now, all merge updates are calculated and applied to each pair in turn, and then the whole batch is repeated once more, ensuring two-way transitivity. for instance, if you are set to copy tags in both directions and set 'A is the best' of three files 'ABC', and B has tag 'x' and C has 'y', then previously A would get 'x' and 'y', but B would not get 'y' and C would not get 'x'. now, A gets 'x' before the AC merge is calculated, so A and C get x, and then the whole operation is repeated, so when AB is re-calculated, B now gets 'y' from the updated A. same thing if you set to archive if either file is archived--now that archived status will propagate across the whole group in one action

client api

• the new 'tags' structure in /get_files/file_metadata now has the 'all known tags' service's tags
• the 'file_services' structure in /get_files/file_metadata now states service name, type, and pretty type, like 'tags'
• /get_services now says the service type and type_pretty, like 'tags'. /get_services may be reformatted to a service_key key'd Object at some point, since it uses an old custom human-readable service type as Object key atm and I'd rather we move to the same labels and references for everything, but we'll see
• updated the client api help with more example result data for the above changes (and other stuff like 'all my files')
• updated the client api unit tests to deal with the above changes
• client api version is now 36

server/janitor improvements

• I recommend server admins update their servers this week! everything old still works, but jannies who update have new abilities that won't work until you update
• the petition processing page now has an 'account id' text field. paste an account id in there, and you'll get the petition counts just for that account! the petitions requested will also only be for that account!
• if you get a 404 on a 'get petition' call (either due to another janitor clearing the last, or from a server count cache miscount), it no longer throws an error. instead, a popup appears for five seconds saying 'hey, there wasn't one after all, please hit refresh counts'

boring server improvements

• refactored the account-fetching routine a little. some behind the scenes account identifier code, which determines an account from a mapping or file record, is now cleaner and more cleanly separated from the 'fetch account from account key' calls. account key is the master account identifier henceforth, and any content lookups will look up the account key and then do normal account lookup after. I will clean this further in the near future
• a new server call looks up the account key from a content object explicitly; this will get more use in future
• all the 'get number of x' server calls now support 'get number of x made by y' for account-specific counting. these numbers aren't cached, but should be fairly quick for janitorial purposes
• same deal for petitions, the server can now fetch petitions by a particular user, if any
• added/updated unit tests for these changes
• general server code cleanup

Version 506

misc

• the thumbnail/media viewer's right-click menu now shows all known modified dates for a file (under the top row submenu). any file downloaded in the past few months should have some extra ones, and you can see how the aggregate number is the reasonable minimum of what you have
• added media viewer shortcut actions for 'zoom: 100/canvas fit/default'
• like with the recent system:time update, the system:rating dialog now has nicer labels for the different numerical operators, saying 'more than' instead of '>' and so on
• also on system:rating, the the 'rated' and 'not rated' choices are now folded into the main radio buttons. to say 'is rated in some way', select 'has rating.' to say 'not rated', set 'is' and make the rating blank. to not search that rating, select 'do not search'. I've wired up the click events here a little, too, to flip from 'do not search' to 'is' when you click and so on
• to make it a little easier to get to, the 'view this file's relationships' submenu is bumped up a level, and the parent 'file relationships' menu is moved above the viewing stats row
• thanks to a user, the install_dir/static dir now has an example hydrus.desktop file for Linux users. feel free to play around with it. the user taught me how this stuff works, so I'm going to try to integrate it into my setup scripts in the near future
• I think I fixed a bug where on rare occasion the client would take 30 seconds to close while waiting on a random daemon like 'sleep check'
• I undid last week's Windows auto-darkmode detection in a hotfix. thanks to the users who quickly notified me that this wasn't working well enough IRL. it is now opt-in, using launch parameter --win_qt_darkmode_test, and it applies darkmode 1 rather than 2. if there are no problems with this, then I will make 1 default and 2 opt-in, so let me know how it goes
• the new Windows taskbar grouping identifier now only applies to the source version of the program. if you pinned the built exe to the taskbar, it was not grouping on that pin (issues #1273, #1271)
• added a custom popup message if a subscription query comes up DEAD on the first sync. it was previously firing off the 'didn't find anything on first sync' error by accident
• when you ok the manage options dialog, if you didn't change the thumbnail size, the thumbnail grids across the program no longer purge and regen
• when you ok the manage options dialog, if you changed the media view options, the image tile cache now clears itself
• when you ok the manage options dialog, if the set mpv.conf content hasn't changed, mpv is no longer told to reload it

sidecar paths

• sidecars get more options regarding their file paths. it is all collected in a new 'sidecar filename' box in the normal metadata routing UI, either for sidecar importers or exporters
• first off, a checkbox now allows you to remove the source media file's extension from the sidecar. with 'my_image.jpg', this would change the default sidecar path from 'my_image.jpg.txt' to 'my_image.txt'. I've heard the the new AI/ML artist .txt outputters use this!
• secondly, an ADVANCED String Converter button lets you go bananas and convert the sidecar path to whatever you need using regexes or whatever
• and lastly, it now has live test/result UI so you can put in an example media path and see what the sidecar will be. this thing is populated with sensible defaults and updates the string converter button's internal example text if you change things
• I added some unit tests for these new features

client api

• the /get_files/file_metadata call has several expansions:
• a new tags structure shows all a file's tags in a neater, combined way. it can do everything the 'service_blah_to_blah_tags' structures do while still giving all information efficiently. please migrate to using this structure within the next eight weeks
• hide_service_names_tags is now default True and deprecated. if you are still using it, please move off it; I will remove it in four weeks
• added hide_service_keys_tags to do similar. it is default False for now, but I will make it True in four weeks and then delete it four weeks later just like names
• the time_modified value is now the aggregated modified timestamp, not the local file modified timestamp
• the new time_modified_details value is an Object of domain : timestamp for all known modified timestamps, by domain
• added thumbnail_width and thumbnail_height for files that have proper thumbnails. they are a reliable prediction, but not a promise
• added is_deleted, which refers to whether the file is either in the trash or has been fully deleted from the client
• added has_exif, has_human_readable_embedded_metadata and has_icc_profile to the metadata Object
• the unit tests have been updated to test these changes
• the help has been updated to reflect these changes. also fixed up some little 'you wouldn't actually get that' issues in the mega 'file_metadata' response example
• the client api version is now 35

running from source

• if the venv activation fails in the setup script or launch script, they now stop there with an error message on all platforms
• linux and macOS setup scripts now look to use 'python3' for initial venv setup, falling back to 'python' if that does not exist
• updated the build scripts to always use 'python -m pip' instead of 'pip' or 'pip3' directly. this stops some weirder environments getting confused about which pip to use
• updated the running from source help with several clarifications and little fixes and notes users have contributed

cleanup

• refactored some menu templating functions from the cluttered ClientGUIMedia and ClientGUIResults to the new ClientGUIMediaMenus
• for the new expanded modified dates stuff, cleaned up how the media 'pretty info lines' are sent to a menu
• replaced a crash-prone emergency-error-handling dialog hook in the database migration rebalance routine with a simple popup message
• cleaned up some bad type hints and other linter warnings
• cleaned up some canvas zoom code
• fixed another 'duplicates' unit test that would on rare occasion fail due to a too-specific test
• removed a no-longer needed token declaration from the github build script that was raising a warning