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.
RetroPie was the only platform with status: archived, which kept it out
of verify --all, slots, the pipeline and list_platforms unless
--include-archived was typed. release_pack already decides which packs a
release publishes and the installer serves every platform, so the status
set nothing apart. The field, its schema entry and the flag are gone, and
RetroPie is covered by every full run like any other platform.
Step 5 compared every narrowed pack with verify and never the --split parts with the full pack, so a part writing a path the full pack settles passed every check. check_split_packs builds each platform's parts in memory and compares their union with the full pack, path for path and byte for byte.
pipeline.py --offline --with-export, listed as a standard command, passes
no --fetch to step 2e: on a fresh clone the database, verify and truth
were built before the export failed on every platform whose own file it
patches. The run is refused before step 0 when .cache/upstream-native/
holds nothing for such a platform, and tools.md says how to fill it.
generate_emulator_pack resolves with offline=offline and fetches a release asset the cache lacks, while verify_emulator resolved offline and called it missing. verify --emulator now applies pending_release_asset as verify --platform does, accepts --offline, and the pipeline compares emulator packs offline on both sides.
Step 5 builds the --required-only, --no-game-data, --region, --target
and --system packs of every platform in memory and compares their paths
with verify run under the same narrowing. verify_platform gains
required_only, generate_pack a record mode. A run samples three regions,
targets and systems per platform (155 packs, about 3 minutes);
--full-narrowings compares all 1,150 (25 minutes). A test fails when
_narrowings() gains a dimension that is neither compared nor exempted.
A platform pack carried only the trees its own systems declared, so for_platforms reached RetroArch, Lakka and RetroPie alone and every other platform running blueMSX, FBNeo or Dolphin got no tree. The pack, the install list and the pipeline's tree members now add the trees of the libretro cores the platform runs, below the system directory and build folder the core reads, unless the platform installs or never opens them. A platform entry the manifest resolved into a carried tree becomes that tree's member, which the installer can fetch: AmberELEC and EmuELEC no longer hold 192 blueMSX machine configs back. stonesoup reads its tree beside the content and pcee2 compiles its resources in, so neither tree reaches a BIOS pack.
Step 5 built emulator packs one profile at a time, where two profiles never share a path. It now also builds each system pack that gathers several profiles and compares it with verify. The first run found five standalone MSX and ColecoVision packs where verify skipped the data tree of a profile the system filter left without files, and credited msx-emu's copies the pack never wrote. verify now places those trees too: 541 system packs, all equal.
--emulator mednafen_psx refused with 'use --emulator beetle_psx' although the profile names its target. The alias now selects that profile and says so on stderr, in verify, generate_pack and cross_reference alike. Launchers are still refused, and pipeline step 5 skips aliases since they compare under the profile they name.
--skip-packs said --skip-docs would stop after verify while 4d and 4e still run, and the tools.md flag table left 8b out of --skip-docs. The help, the usage and the flag table now name the steps the step table marks.
The dry run, the Batocera walkthrough and the narrowing line of
wiki/installer.md were typed by hand and went stale with each list.
generate_installer_doc.py now runs install.py against the committed
lists in a sandbox, dated by each list, and the pipeline runs it at
step 7.
The render step's npm build wrote Node's compile cache to /tmp, a 4 GB
tmpfs on this host, and a day of pipeline runs filled it to 3.4 GB.
Every child the pipeline starts now carries NODE_DISABLE_COMPILE_CACHE=1.
release_pack: true in the registry is the list a release builds: RetroArch,
EmuDeck, RetroBat, Batocera, RomM and MiSTer, 83% of the downloads of
v2026.10.10. generate_pack --all, --verify-packs, the pipeline consistency
step, release_record and check_release read it. Manifests are still written
for every platform.
generate_site.py writes the Starlight project in website/ and nothing
else: mkdocs.yml, the MkDocs hooks, theme overrides and stylesheet go,
and the wiki is written in Starlight's syntax. Both workflows set up
Node with the official actions/setup-node pinned by commit, run npm ci
from the committed lockfile, build with the strict link validator and
check the result with validate_site.py.
The search weighs a page by what it answers: Start here pages first,
a system page's cards and file names before a guide citing them, deep
guides, launchers, aliases and frozen copies after. Status badges,
hash blocks and the export change lists leave the index. The ranking
sits at the centre of a range where the 13 queries of search_gate.py
all answer, in the search box itself.
839 pages, every URL and 45,611 ids match the MkDocs build. The tests
read the Starlight content and syntax.
verify resolved offline, so a release asset absent from bios/ and the
cache read as missing while an online build fetched and packed it. verify
now reads the database record of the registered path and counts such a
declaration present, listed as fetched at build time. --offline keeps the
old count, and the offline pipeline passes it to its verify step.
Without jsonschema the step printed only the missing module, and the run then read COMPLETE while PR validation runs the same check. The skip now names the install command and the PR check.
Step 2a advanced mame.yml's core_version and upstream_commit to a new
release and left the device ROM sets on the previous one. A plain run
that moves the pin now fetches that release's listxml into .cache/dats
and runs the device merge, which checks version and commit itself.
The refusal tables compared each flag with its default value, so --source full and --group-by system passed beside the modes that refuse every other value. Both now default to None and take their effective value after the checks, in generate_pack and pipeline.
The pipeline locked only --output-dir while it also writes database.json,
install/, README.md and docs/. A run on a scratch output directory could
start beside the run on dist/ and rewrite them. Every run now holds the
repository's .lock first.
Ranking contested core claims sorted the whole gap report, which
reordered every manifest. Only the claims on one contested path move now,
grouped where the first of them stood, and two spellings of a path are
one group where the filesystem folds case (RetroBat's 64DD_IPL_DEV.bin).
An entry without a destination (EmuDeck's system:md5 whitelist) no
longer counts as a pack path in verify, and step 5 prints one status for
counts and paths together. A full offline run compared all 12 archives
and 930 emulator packs path by path, every one equal.
The 3.12 guard rested on generate_readme f-strings that have since been
rewritten. Every script compiles under CPython 3.9.25, 3.10 and 3.11, and
only generate_pack's parenthesized context managers fail under 3.8. Under
3.9 the offline pipeline (--skip-packs) completed with no change to the
committed artifacts, the 15 manifests matched 3.12 byte for byte, and the
RomM pack was identical. CI now compiles every script under a real 3.9.
The help said "Only regenerate DB and verify" while steps 1b, 4d, 7 and
8 still ran and rewrote derivable.json, README.md and docs/. It now names
the steps it skips and points to --skip-docs.
Every step runs from the repository root, but the lock and the purge of
stale packs read --output-dir from the caller's directory. Started from
scripts/, the run locked and purged scripts/dist/ and wrote to dist/,
leaving the stale packs there. The directory is now made absolute before
the lock, dist/ at the root by default.
The lock refusal, "no pack for", "no manifest for" and "no pack to
verify" went to stdout, so a run redirected to a log exited 1 with
nothing on the terminal. A test reads both scripts and refuses an ERROR
line printed to stdout.
Step 5 compared (ok, total) and the core file count. RetroArch, Lakka
and RetroPie read 526/526 on both sides while verify credited four paths
below SGB1.sfc and SGB2.sfc that the pack never writes: a declared file
blocks them as a directory. verify now places the platform files in the
builder's order with the same path rules, and a user_provided or
external declaration keeps its slot on both sides, so no core file takes
it in the report only.
verify --destinations writes the paths each pack holds by the report,
and step 5 compares them with the archive members (notes and data tree
members set aside). Every emulator profile, in each build it has, is
built in memory (PackRecord, nothing read or written) and compared with
verify --emulator path by path. Cost measured here: about 1 s per
platform archive and 11 s for the 930 emulator packs.
--emulator accepts a buildbot core name from a profile's cores:, and a
mistyped target gets the closest names instead of every target. diff_truth
shows destinations and reports unpaired entries as evidence, not removals.
The verify summary counts discrepancies, a shared destination names both
systems, a truth hash conflict names both cores, and pipeline steps carry
their place in the real run instead of N/8. Every GitHub API caller takes
the token from github_token(), which falls back to gh auth token.
Abdessamad Derraz