Three site commits pushed to main fail checks CI runs: two prose
semicolons, an untyped signature, and GoatCounter answers read as a
dict without a test. GoatCounter's stats now go through one reader
that refuses an answer that is not an object.
The skill and llms.txt said "a quote" where the tool refuses the
typographic ones too, which end a PowerShell string. One constant states
the set for both, and points to device_setup_command.
The skill and llms.txt told an agent to write --dest=<card>/bios bare.
A folder now goes in single quotes, attached to its option, and one that
holds a quote, a backtick or a control character, or starts with -, is
a question for the user.
ARD reserves text/markdown for agent skills, and Lighthouse warned on
the llms.txt entry. skills/setup-bios/SKILL.md holds the steps that set
up a device with the fewest files, and the manifest lists it as a skill.
device_setup_command built a shell command from the agent's values: a
typographic quote ends a PowerShell string, and a path starting with -
reads as an option. Such values are refused, the options are written
attached to their values, and platform must be a setup name.
llms.txt and an ARD manifest name the pages and the JSON API, and lead
an agent setting up a device to --for-games --required-only. Every page
registers four read-only WebMCP tools: the setup command, a hash lookup,
an emulator's files and the site search.
--for-games --required-only on a dArkOS card from a computer kept every
file of the list: --dest skipped the card's detection, so its rom
folders were never found. The games part beside the BIOS folder now
names them. The manifest marks a data tree every core runs without
(optional), and --required-only leaves it out: one GBA and one SNES game
took FBNeo's cheats and samples, 3,472 files no game opened.
validate.yml runs export_native with continue-on-error, and pr_checks
failed the whole run on it. The step is now shown and does not fail
the run, and a test keeps that list equal to the workflow's.
Code moved to sitecollection.py and the triage progress added errors
the baseline did not hold. A reused loop name, two lists without a
type, returns without a value, and a regex match read without a test
for None. Two baseline entries no longer occur and are removed.
The CI runner sets XDG_CONFIG_HOME to its own home, and the installer
now reads it as RetroArch does. Tests that fake HOME then found the
runner's RetroArch folders. The suite clears the three XDG variables
it reads, and a test that exercises one sets it itself.
og:title and twitter:title carry the search title ("mGBA BIOS
files"), an emulator page shares its maker's card, a download page is
dated by its release in the sitemap, and a Switch emulator's files are
called keys.
recipes/archives.json records each of the 2,222 archives database.json lists, .variants/ copies included: members in order with size, crc32 and sha1, the container sha1, md5 and size, exact when the TorrentZip build gives its bytes (1,790) and canonical otherwise (432), the profile and DAT recipes agreeing on its members, and the platforms or profiles pinning its bytes.
romset_recipes.py --capture writes it (pipeline step 1c), --build and --build-all rebuild from held members, loose first, then another archive, then its own copy, read back and checked before the move into place. Under zlib-ng nothing is claimed exact. A test fails when an archive has no recipe or a member is held nowhere.
The importer keeps the archive names the collection holds and every set whose member list equals an archive the collection keeps under another name (aga.zip is isa_aga), under the collection's name with the upstream one in set. All 113 MAME listxml releases (0.177 to 0.289) are imported, not one in five, and FBNeo is read again at its tip: 2,936 MAME recipes and 145 FBNeo ones, compacted as before.
fccpu20.zip, fcisio1.zip and interpod.zip now have recipes and carry slot ROMs beside their own set. Trimmed, each equals the archive under the newer MAME name, so they are kept as they are until the storage decision.
write_torrentzip reads and deflates one member at a time, gives back what it read (size, crc32, sha1) and takes a ZIP64 field only past 32 bits, so a 2 GiB disk image builds without being held whole. TorrentZip layouts are unchanged. zlib_ng_version() names the deflater that cannot reproduce TorrentZip bytes.
Recalbox hashes a ZIP by its members (zip_digest: composite, Bios.cpp:64-68), so its md5 names no container bytes: romset_recipes --missing tried to rebuild it and counted it unreproducible.
validate.yml runs mypy against its baseline and checks that shellcheck
is installed, and pr_checks ran neither. The guard test only knew the
python, vermin and npm commands, so it missed both. A step can now pipe
its output into a second command, as mypy into mypy-baseline filter.
check_platform_folder.py, the script the site publishes, still accepted
a cut md5 by prefix. It now compares whole digests as Batocera does
(batocera-systems:1088), like verify since cut_md5_owner.
Mini vMac opens no MinivMacBootv2.dsk at babdf7b5: disks come from a .cmd
(libretro-core.c:181-195) or disk1.dsk to disk9.dsk
(OSGLUERETRO.c:563-579). The name is the one Recalbox's configgen writes
into its .cmd, already recorded in its launchers: layer. Nine manifests
no longer send the 24 MB image to platforms that never open it.
Abdessamad Derraz