Download and execute aqr-20261121-win-install.exe if you have no other reason.
User-facing Changes
--mark WHENchooses what happens to the mark of the web (Windows' Zone.Identifier, macOS' quarantine):- When extracting:
propagate(the default, as before),always(mark every file, as from the Internet if the archive has no mark),keep(only the marks stored in the archive) ornever - When compressing:
keep(the default) ornever - Explorer's and Finder's verbs still extract with the default
- When extracting:
--compactand--moveno longer refuse archives with entry comments, ZIP extra fields, an archive comment or extension keys; they keep them- Windows Explorer:
- All the selected files now reach aqr-win at once, as a drop. Before, selections that included downloaded scripts (
.jsand the like) could be split across several archives, or partly left out when Explorer's security prompt was cancelled - The Aqueous menu no longer disappears when more than 15 files are selected
- All the selected files now reach aqr-win at once, as a drop. Before, selections that included downloaded scripts (
- Developer options:
-Z no-declared-limitdecodes an archive without its declared memory limits, for archives whose limits were declared too low (see Bug Fixes)
Specification Changes
- Extension keys: text string keys of
aqr-opt-metadataandaqr-entry-extrahold a program's own metadata, named by a reversed domain. They must not change anything the specification defines, decoders must not reject them, and programs that write an archive again should keep them, but must drop them from entries whose content changes - Each message queued in a gang's channels counts at least one byte toward the gang's memory limit, so that queues of empty messages are bounded too
Performance Optimizations
- From level mid up, four batches' worth of data is let in at once instead of two, so a batch holding one long file no longer leaves the other threads idle
- Base64 and hexadecimal blocks in large texts are now searched for in one pass on all threads
- NTFS images with a fragmented MFT no longer take time quadratic in their number of records
- AC-3 and DivX 3 streams split into many small chunks decode in a few seconds, not tens of seconds
Bug Fixes
- Files with several outputs sharing one module run (such as some installers' executables) declared too little memory. They compressed, but
-dand-Kthen failed with "memory limit exceeded". Archives made by earlier versions can be decoded with-Z no-declared-limit - Appending (
-a):- An append cancelled through
--ipcnow restores the archive exactly as it was. Before, it was left without its footer, and could still be written to after being cut back - An archive that is among its own inputs is skipped. Before, it was read as an input, which failed on Windows
-a,-j,--delete,--tagand--compactare refused on archives made with-A. Before, they left blocks that were not mirrored
- An append cancelled through
- Extraction:
- A file that cannot be written in full (a full disk, for example) is removed, not left behind cut short
--ipc: cancelling, or closing standard input, during-dkeeps the files already extracted and removes the one being written. Before, files were left cut short--ipc: a frontend closing its end right after the lastendno longer removes the archive just made- Lines that are not UTF-8 from the frontend are ignored, not taken as a cancel
- Windows:
- Files named with a trailing dot or space, or with a device name (
CON,NUL.txt, …), can now be compressed - Folders under
\\wsl$no longer warn about alternate data streams for every file - An archive that could not be replaced named its path twice in the error
- Files named with a trailing dot or space, or with a device name (
- Self-extracting programs:
- They are written to a temporary file next to the destination and renamed when complete, so a failed write no longer leaves a partial program in place
- They find their archive even with data appended after it, such as an Authenticode signature
- They name extracted files the way
aqr -ddoes
- Archives that could not be extracted:
- Uniform GPU textures (a large BC7 or ETC2 texture of one color, for example) compressed fine, but the decoder rejected them
- A JPEG with a damaged scan among separate sequential scans was compressed into a formula that failed to decode
- DivX 3 video at 1080p or more exceeded its own memory plan when run as WebAssembly
- A large DivX 3 AVI made by 20261112 could crash the sandbox worker that decodes it
- A checksum on an input cut into pieces was not checked
- ZIP output: names of 65536 bytes or more made a broken central directory. They are now an error
- MS-DOS times (ZIP, ALZ, EGG) past 2107 became 1980; they now stop at the end of 2107
- ELF files with more than 65279 sections lost their section names
- A panic in the steps that take apart zstd frames, DEFLATE streams inside files or git packs, or in a module run during compression, no longer aborts the whole archive. The files are kept as they are, and the panic is reported
- Corrupt or crafted input no longer causes panics, aborts, endless loops or huge allocations across most readers and decoders. Among them:
- Containers: RAR, 7z, ZIP, tar, EGG/ALZ, WIM, OLE, ASAR, CLIP, PSD, DEX, ARSC, DWARF
- Disk images and file systems: WOF and LZNT1 files, GPT, ISO, HFS, ext, FAT, APFS
- Streams: Ogg, IMA and MS ADPCM, AC-3, DivX 3, LZX, XPRESS, LZNT1, AZO, Snappy, zstd, BCJ2
- Archive metadata and formula blocks
- Output lengths declared in the gigabytes now fail cleanly in WebAssembly too
- Module runs: spawning tasks with memory beyond a module's declared maximum, or queueing empty messages without end, is now an error. Archives using shorthand module blocks now decode in the sandbox