Move the visible frontend section above At a glance, remove four requested subsections, and restore all 33 renderer entries. README warnings preserved. Validation: format/evidence passed with 117 captures and local preview reviewed. Release-note removal excluded from this commit after approval review rejection.
Replace the Overview hero with the Games capture, document current library/media/save and QR workflows, group 33 cores with system icons, and retain every permanent warning. Keep source-build status and known limits explicit.
Verification: bash tools/verify.sh format evidence — PASS, 116 captures. Source audit: all local links resolve and 33 cores retained. Local browser: 69 images loaded, no page overflow. Evidence: evidence/readme-beta-refresh. No publication.
Keep all 18 core entries, ten screenshots and permanent warnings. Add illustrated highlights, linked contents and grouped system artwork; remove development internals and shorten public descriptions.
Validation: bash tools/verify.sh format evidence — PASS, 115 captures. Browser: 57 images loaded, no horizontal overflow. Evidence: evidence/release-beta-illustrated-notes/{capture,expectation}.json. Local preview only; no publication.
Keep opt-out available and restore the enabled default on each new draft. Browser coverage verifies source visibility, labels, manual opt-out and the default download flow.
Validation: tools/verify.sh format unit build integration evidence; 185 unit and 185 integration tests pass. Build fd8f648f deployed; defaults and config preservation verified. Evidence: evidence/webui-auto-media-default/capture.json. Replay: tools/verify.sh evidence.
Recapture Download Media with Nintendo - Game Boy Advance collapsed, hiding the game name and match. Validate updated screenshot in the local release preview. tools/verify.sh format evidence passes (113 captures). Evidence: evidence/webui-release-collapsed-recap; replay: tools/verify.sh evidence.
Replace four captures and their captions/catalog metadata. Validation: tools/verify.sh format evidence; 112 captures pass, all 15 preview images load. Evidence: evidence/webui-release-bust-a-move/capture.json. Replay: tools/verify.sh evidence.
Capture the requested Neo Geo B library, Add Game, game inspector and Download Media views from the console WebUI, with explanations and a screenshot digest catalog. Preserve the permanent warning and local QR preview.
Validation: tools/verify.sh format evidence; 111 evidence captures pass. All 15 local preview images load. Evidence: evidence/webui-release-gallery/capture.json; replay: tools/verify.sh evidence.
Fix the blocked 3D box and disc icons without changing the CSP. Check all ten media selectors decode/render in the production-server browser regression.
Validation: tools/verify.sh format unit build integration evidence; 185 unit and 185 integration tests pass. Build 8135742c deployed at Picker; all ten icons visually verified; config hashes unchanged. Evidence: evidence/webui-media-icons/capture.json. Replay: python3 tools/evidence.py compare evidence (110 captures pass).
Add a Media Inspector-based workflow for new backups, editable metadata,
LaunchBox lookup and manual/automatic media. Keep users in a live job view
until completion, then open freshly loaded artwork. Preserve prior metadata
edits and rank literal database titles above enhanced editions and variants.
Use the approved color icons, real region flags and hardware artwork for all
50 recognized systems. Restore RetroArch branding and bundle attribution.
Preview native savestate screenshots and invalidate them on replacement.
Validation: PS5_FRONTENDS='es-de picker' PS5_VULKAN_TEMPLATE="$PWD/build/picker-template"
bash tools/verify.sh format unit build integration evidence
185 unit and 185 integration tests; all shipped targets and evidence pass.
Base PS5: seven disposable-file/API checks, asset byte comparisons and
configuration preservation pass; test files removed, Picker restored.
Evidence: evidence/webui-add-game/capture.json
Replay: python3 tools/evidence.py compare evidence
Verified actual Picker, XMB, RGUI, and ES-DE captures with zbarimg and live WebUI HTTP 200. Original Picker and configurations restored. Raw images stay ignored; sanitized hashes and procedure in evidence/webui-qr-console-captures. tools/verify.sh format evidence: 107 captures pass.
Verified with tools/verify.sh format evidence (106 captures); evidence/release-note-beta1 records the updated Markdown digest. Local browser preview confirms yellow warning and linked 1.7beta4 release. Permanent warning diff audit passes.
Present the 18 added cores, Games library, media downloads, shared artwork,
storage support, stylus controls and fixes since v0.6.7-alpha.6. Preserve the
README warning verbatim and keep paragraphs on single lines for release rendering.
Apply the owner's preview edits: omit the Beta notes section, overview image
and its caption. No release, tag or package is published.
Verification: tools/verify.sh format evidence (106 captures, zero failures).
Rendered Markdown inspected in the local browser: four tables, warning image
loaded, no horizontal overflow, requested removals confirmed.
Evidence: evidence/release-note-beta1/{capture,expectation}.json.
Replace Azahar's inverted cursor in Vulkan and software rendering with the
same supersampled white pen, dark outline and lower-left nib as DeSmuME.
Keep clipping anchored to the touch point and preserve the input behavior.
Verify: tools/verify.sh format
PS5_FRONTENDS='es-de picker' PS5_VULKAN_TEMPLATE="$PWD/build/picker-template" tools/verify.sh unit build integration evidence
182 unit and 182 integration tests pass; native core ABI/provenance pass.
The parity regression fails on the old cursor and passes at 1x, 5x and 18x.
Evidence: evidence/azahar-stylus-parity/{capture,expectation}.json
Fresh base-PS5 launch draws the stylus in both captured frames and exits
cleanly; stored eboot and DS/3DS core hashes match the staged package.
105 evidence records replay successfully. Core-switch crash is unresolved.
Rasterize a slim outlined, antialiased pen at the DS output resolution.
Cast Azahar's unsigned screen dimensions before multiplying signed deltas;
the production-type regression fails before this change and passes after it.
Verification: tools/verify.sh format; PS5_FRONTENDS='es-de picker'
PS5_VULKAN_TEMPLATE="$PWD/build/picker-template" tools/verify.sh unit build
integration evidence; tools/verify.sh format evidence. 182 unit and 182
integration tests passed, 104 evidence records replayed.
Evidence: evidence/stylus-polish/{capture,expectation}.json records native
core provenance, stored console SHA-256 readback, and clean production
fresh Azahar (600 frames) and DS Close Content -> Load Content -> Azahar
runs on build 9917ddb4. Reproduce host checks with tools/verify.sh and
recorded console commands with the ignored diagnostic fixtures.
The original owner load crash is unresolved, not claimed fixed. The
active-content scripted RELOAD fault was an invalid reproduction because
it unloaded the core from inside its input callback. Diagnostic handler
changes are excluded from this commit and the deployed production build.
Adapt Proton contact tracking for relative hover and physical press/drag/release.
Keep touchpad Select in menus and other cores. Add resolution-scaled mouse
motion and clipped stylus cursors to DeSmuME and Azahar (Vulkan/software),
with reproducible core patches, build provenance and licence attribution.
Verified: format; 182 host/integration tests; all shipped native targets;
core ABI, manifest and licence checks; 103 evidence captures replayed.
Deployed build 6da657c4 to the base PS5; stored SHA-256 of eboot.bin and both
cores matches the stage. Physical controller/TV acceptance remains pending.
Reproduce:
PS5_FRONTENDS='es-de picker' PS5_VULKAN_TEMPLATE="$PWD/build/picker-template" tools/verify.sh
python3 -m unittest tests.test_input_ps5 tests.test_touchpad_stylus
Evidence: evidence/touchpad-stylus/capture.json and expectation.json.
Raw logs: klog/touchpad-stylus/ (ignored).
Bind WebUI in the common action initializer instead of the custom keyboard
branch. Preserve patch markers so existing configured trees update too.
Production callback/dispatch regression fails before the fix and passes
after it; picker and ES-DE binaries remain byte-identical.
Resolve the integration fixture race exposed by this run: browser covers
now use valid PNG data and assertions require decoded images.
Validation: PS5_FRONTENDS='es-de picker' PS5_VULKAN_TEMPLATE="$PWD/build/picker-template" tools/verify.sh
Format, unit (177 tests), and build passed. After correcting the fixture,
tools/verify.sh format integration evidence passed (177 tests).
Evidence: evidence/webui-qr-actions/capture.json and expectation.json;
reproduce with python3 tools/evidence.py compare evidence/
Deployed build 0bdc8dd8: 6 changed files verified; exact stored-byte eboot
digest matches. Owner confirmed picker/ES-DE; XMB/RGUI live recheck pending.
Picker Triangle toggles and Circle dismisses; XMB gets a WebUI tab left of
Settings; RGUI shows WebUI first; ES-DE replaces Scraper with WebUI.
Share runtime IPv4 discovery and an unchanged MIT Nayuki QR encoder.
Verify: PS5_FRONTENDS='es-de picker' PS5_VULKAN_TEMPLATE="$PWD/build/picker-template" tools/verify.sh
All five gates pass; 176 tests in each pass. Independent QR decoding,
address changes/offline recovery, picker input and menu routing tested.
Evidence: evidence/webui-qr/capture.json and expectation.json; replay with
python3 tools/evidence.py compare evidence/
Deployed build 700259e0: 9 changed files verified, 13614 retained. Exact
stored-byte digests match all three frontend binaries. On-console scans
and controller checks remain pending; no title launch performed.
python3 tools/deploy-title.py verified 9 changed files and skipped 11689
unchanged files using the prior console deployment record. A separate
SELF-disabled FTP readback verifies the exact stored eboot SHA-256; live
HTTP page assets match the staged files. The old daemon is still running,
so metadata sorting on the restarted server is not yet claimed.
Artifact: evidence/webui-games-sorting/capture.json and expectation.json.
Verification: bash tools/check-manifest.sh; tools/verify.sh evidence
(100 captures, zero failures). Raw captures remain in ignored
klog/webui-games-sorting/deploy/.
Add All/#/A-Z buttons, remembered sort field and direction, and global
sorting before pagination. Expose stored release date, genre, developer,
publisher and rating in the library response; missing details stay last.
Verified: tools/verify.sh format; tools/verify.sh unit build integration
evidence. All five gates pass: 169 tests per pass, PS5 title and daemon
built, staged manifest checked, 100 evidence captures replayed.
Final format/evidence rerun also passes.
Reproduce feature checks:
python3 -m unittest -v tests.test_scraper.Scraper.test_library_sort_details tests.test_scraper.Scraper.test_games_page_in_a_browser
Artifacts: evidence/webui-games-sorting/capture.json and expectation.json
(host results and staged artifact hashes). Console deployment/live check
remain pending; no console verification is claimed.
The owner saw a job "only download after all three passes". The job record says
otherwise: the NeoGeo games already had nearly every picture and their details; what
was missing were manuals and videos, which LaunchBox and libretro do not have. Their
passes checked all 204 games in ~10 s each and found nothing to give; ScreenScraper (one
game at a time on this account, level 1) has been downloading since, in real time (78
files, 70 MB in 3 min). But the card hid it: a game handed to the next source counts as
pending, so nothing moved during a pass.
- Job.passes (kept in the job file): per pass, the games its source checked, those it
gave something, and how many kinds; /api/scraper/job gives them with the games that
reached each source.
- The card: one step a source (done, running with "checked X of Y" and a bar, or
waiting), and what it found.
Evidence: tests/test_scraper_screenscraper.py (the chain's passes: libretro 5 checked,
3 games, 6 kinds; ScreenScraper 5, 3, 6); a screenshot of the steps. Console: the job
record and the monitor's log above. tools/verify.sh: PASS (format unit build integration
evidence).
On the console the daemon stopped minutes after every start, the page then refused to
connect, and the status said RetroArch was closed while the picker ran. Its log
(webui-daemon.txt): "serving the WebUI" 6 s (9 s the time before) after its start,
then "linked: picker", "unlinked: title", "unlinked: picker", and 20 s later "no program
of the title linked: the title is closed". The WebUI's start (reading the library, once
for each of 16 saved scraping jobs) ran on the link loop: the picker's hello waited past
its 5 s, the picker gave up for good, and the daemon, linked to nothing, stopped.
- The daemon starts the WebUI on a thread of its own (4 MiB stack, joined before a
stop): hellos are answered meanwhile.
- The saved jobs are read against one scan of the library, not one each.
- src/webui_link.cpp: a hello not answered is asked again every second, never given up
while the daemon is there.
Evidence: tests/test_webui_daemon.py test_hellos_are_answered_while_the_webui_starts (a
start made 7 s slow: both hellos answered in under 1.5 s, the WebUI up later with the
right frontend). tools/verify.sh: PASS (format unit build integration evidence).
The daemon died at start on the console (SIGSEGV, fault address 0) and never came up:
the backtrace (klog, symbolized against build/webui-daemon/ps5-retroarch-webui.elf)
runs from a scraper worker's download through curl_easy_perform into
mbedtls_ssl_write_client_hello -> mbedtls_ctr_drbg_random_with_add ->
mbedtls_entropy_func -> mbedtls_sha512_update. curl 8.22 shares one CTR-DRBG (prediction
resistance on) among all its connections and relies on mbedTLS's own locks for it; our
mbedTLS 3.6.7 was built without them. With 32 workers (a034b6e) starting handshakes
together the shared generator was torn; the job resumed by itself at every start, so
every start crashed.
tools/build-curl.sh now sets MBEDTLS_THREADING_C and MBEDTLS_THREADING_PTHREAD
(scripts/config.py) before building mbedTLS. Why a flag in a shared script: curl's
mbedTLS backend requires it whenever handles run on several threads; the ES-DE build
that links this curl gains the same locks.
Evidence: libmbedcrypto.a's ctr_drbg.o now calls mbedtls_mutex_lock (pthread mutexes in
threading.o); the daemon's mbedtls_ctr_drbg_random calls the lock before and after its
work (objdump). tools/verify.sh: PASS (format unit build integration evidence). The
console run follows the deploy.
Two races of the funnel (a034b6e), found by repeated test runs:
- Between a source's result and the funnel moving the game to the next pass, the game
looked settled; another worker could end the job then, and the game stayed pending.
- A worker clearing its hold after the funnel queued the game could clear the hold of
the worker that had already taken it on.
Each worker marks the games it holds with its own number (Item.holder, not saved); a pass
or the job ends only when no game is held; a worker frees only its own hold; a held game
is never taken by another worker nor reset by Resume.
Evidence: the three scraper suites (tests/test_scraper*.py) passed five runs in a row
(failed 1 in 3 before). tools/verify.sh: PASS (format unit build integration evidence).
No checkerboard and no dark stage behind pictures: a logo, a 3D box or a disc shows the
page through its transparent parts, in the game overlay, the full-size view and the
Games tab's cards (all four views). Only a video keeps a black backdrop.
Evidence: screenshots with real transparent PNGs (3D box in the overlay, logo on a card)
checked by eye; tests/webui_scraper_browser.cjs unchanged and passing. tools/verify.sh:
PASS (format unit build integration evidence).
- The owner's design: the first source goes through every game at full speed, as if no
other came after it; when its pass is over, the next source gets only the games still
missing something, for what they miss; and so on. Before, each game went through every
source in turn, and ScreenScraper's one-game account limit held the whole job to one
game at a time.
- Job.pass (the running source) and Item.pass (the source a game waits for), kept in the
job file; a pass ends when none of its games is pending or in flight. A choice in
"Games that need you" runs its source again, then those below; Resume sends a failed
game back to the top (what it has is kept).
- ScreenScraper's thread limit is a slot around its own work only; a game that gets its
slot after a pause asks nothing.
- 32 workers on the PS5 (was 4); the PC helper downloads 16 at once (was 4). libretro's
game list of a system is fetched once, not once a worker.
- Fixed a race the extra workers exposed: Resume turned games still held by a worker
back to pending, so a second worker took them too (a video downloaded twice).
- The page: the job says which pass runs and how many games wait for the next source;
the source hint describes the funnel.
Evidence: tests/test_scraper_screenscraper.py (libretro's pass ends before ScreenScraper
is asked about any game; with a one-game account libretro runs several at once and
ScreenScraper one; a quota pause mid-funnel resumes with nothing fetched twice);
tests/test_scraper.py (cancel, resume and restart with 32 workers). The three scraper
suites passed three runs in a row. tools/verify.sh: PASS (format unit build integration
evidence).
- LaunchBox has no API: its database is one download (Metadata.zip, ~110 MB). The
console fetches it once a month, streams Metadata.xml and Mame.xml out of the zip
(minizip, already in the daemon) and keeps a small index a platform in
library/.launchbox (games, alternate names, images of the types used); the zip is
deleted. Images come from its public image server: no credential in any address, so
PC mode works too.
- Matching: a game chosen by the user, else its title or an alternate name (words as
libretro's matching), else the closest names ("Name [id]"), the same title taken
unasked; arcade romsets ("mslug3") by MAME's name. 80 library systems mapped to
LaunchBox platforms (arcade systems to Arcade, Neo Geo MVS/AES...).
- Media: front, back and 3D box, gameplay and title screenshots, clear logo, disc or
cartridge, fan art, in the region asked for; details: description, developer,
publisher, release date, genres, players, rating (0-1).
- It joins the source chain; a failed database download pauses the job (one message,
the other workers wait for the resume).
Evidence: tests/test_scraper_launchbox.py against a fake LaunchBox (titles, an alternate
name, a romset, regions, details, an ambiguous game and a search, the database fetched
once and no image twice, PC mode through the helper, a libretro -> LaunchBox chain, a
failed download paused then resumed). Against the real database on the host: index in
6 s (140 MB, 79 platforms); 143 of 150 FBNeo romsets of the base console matched.
tools/verify.sh: PASS (format unit build integration evidence).
- The check covered only the 15 cores with option catalogs (webui/core-metadata/
index.cfg); the 18 added since (3DO, Neo Geo CD, Atari, PC Engine, PC-FX, Amiga,
Dreamcast...) were never looked at. It now lists cores/*_libretro.so and reads each
core's .info (its display name, its library name for the per-core system folder).
- Rules where the .info files are wrong: PS1 and Saturn need one BIOS of any region,
not all of them; 3DO and Neo Geo CD need one of their BIOS files (and Neo Geo CD its
LO ROM) though the .info calls all optional; PC-FX needs its BIOS; PC Engine CD games,
Amiga Kickstarts and the Dreamcast BIOS are needed only for some games.
- The page: "Needed" (counted in the badge) and "Only for some games" apart; each item
says why, lists the files any one of which will do, and the folder they go in.
Evidence: tests/test_webui.py test_alerts_follow_installed_cores_and_custom_system_folder
(Saturn one-of, a custom system folder, a core without a catalog checked by its rule,
PUAE as some games, an unknown core by its .info); a preview of all 66 built cores with
no BIOS listed what each needs. tools/verify.sh: PASS (format unit build integration
evidence).
The daemon outlives a deploy or an update: the console served new page files from an
older server (Game details greyed out, no source chain: the server reported neither).
The server keeps the page build it started with (webui/version.json) and reports it
in /api/status; the page compares it with the version.json now on disk and, when they
differ, asks to close and reopen PS5 RetroArch. A source an older server describes
without "details" counts as having them unless it is libretro.
Evidence: tests/test_scraper.py test_the_server_reports_the_page_build_it_started_with;
tests/webui_scraper_browser.cjs (the notice with newer page files). tools/verify.sh:
PASS (format unit build integration evidence).
- Options.sources (sources=libretro,screenscraper): per game, the first source is
asked for every kind and the details; each next one only for what is still missing,
so a free source spends nothing of a limited one's quota. Each source keeps its own
name for the game (Item.matches, kept in the job file). A game no source could tell
offers the first ambiguous source's choices; once chosen, a later source can offer
its own for what is left ("Some found; for the rest, ScreenScraper offers these").
A quota pause mid-chain resumes without asking again for what was found.
- A chain downloads on the PS5 (one source at a time through a PC); the order is
remembered with the settings; ScreenScraper's thread limit applies whenever it is in.
- The page: "Sources, in order": each source a row with its number, a switch (or Sign
in), arrows and drag to reorder; the hint says who is asked first; media tiles say
"Asked of libretro, then ScreenScraper"; the job and recap name the chain.
Evidence: tests/test_scraper_screenscraper.py: a chain where ScreenScraper is asked
only for videos (no picture), resolved through both sources' choices; a quota pause
with no cover fetched twice after the resume; PC refused for a chain; the order
remembered. tests/webui_screenscraper_browser.cjs: sign in joins the chain, reorder,
turn off, sign out. tools/verify.sh: PASS (format unit build integration evidence).
- The game overlay: each kind is a wide row with a large stage (the whole picture,
never cropped; logos on a checkerboard), a click opens it full size (Escape or a
click closes); the video plays in place with a big play button and its controls;
the manual opens as a PDF; missing kinds are short dashed rows to drop a file on.
- The Games tab: "Show" box art (front), 3D box, logo, or cartridge or disc on the
cards, remembered in this browser; a game without that picture shows its drawing.
Evidence: tests/webui_scraper_browser.cjs (the rows, the view choice kept across a
reload, games without a 3D box drawn as such); screenshots with real PNGs and an MP4
checked by eye. tools/verify.sh: PASS (format unit build integration evidence).
- Options.details (details=1): a job fetches the games' details (description,
developer, publisher, release date, genre, players, rating) only when asked; media
alone, details alone, or both. Details follow "Media you have": filled when missing,
replaced only when replacing everything, and a hand-edited field never. Remembered
with the settings; kept in the job file.
- libretro has pictures only: its tile says so, a details-only libretro job is refused
with the reason; each source says whether it has details.
- The recap counts details like a kind (new / kept / missing).
- The page: a "Game details" tile under the media tiles; the summary names it.
Evidence: tests/test_scraper_screenscraper.py test_details_alone_or_media_alone (media
without details, details without media and no media fetched, kept on a second run, an
edited description kept when replacing, neither refused) and the recap totals;
tests/test_scraper.py test_libretro_has_no_details; the browser test's tile.
tools/verify.sh: PASS (format unit build integration evidence).
- GET /api/library/game: a game's details and, for each of the ten kinds of media,
whether it is stored (type, size, uploaded) and the types it takes.
- PUT /api/library/media: the user's file, streamed to a hidden file beside its final
name and renamed in place when whole (token required; PNG/JPEG for pictures, MP4/WebM
for video, PDF for manuals: what every frontend shows). It becomes the kind's only
file and is marked in the .meta ("uploaded"): no scrape replaces it, even when
replacing everything.
- The page: clicking a card opens an overlay with the game's details, then one row per
kind (a preview, a description, its file); missing kinds greyed; Replace / Add, or a
file dropped on the row, with progress. The card's tags name every kind properly.
Evidence: tests/test_scraper.py test_a_game_shows_its_media_and_takes_the_users_own
(refused types, unknown game and kind, a path outside the library, no token; the
user's cover replaces the scraped one and survives a replace-everything scrape);
tests/webui_scraper_browser.cjs (overlay rows, greyed missing kinds, a refused GIF,
an uploaded PNG shown at once, the checkbox still selecting). tools/verify.sh: PASS
(format unit build integration evidence).
- Each game records the kinds this job stored (kept in the job file). GET
/api/scraper/recap lists, per system and game, the kinds downloaded, already there
and missing, with totals per kind.
- The finished job card shows it: per-kind tiles (new / kept / missing, a bar; a click
narrows the list), filters (got new media, missing something, not found, unchanged),
a search, and a collapsible list per system with each game's kinds as pills and why
a game was not found.
- "Replace what's already there" becomes "Media you have": "Download only what's
missing" (default) or "Replace everything", each explained, and named in the summary.
Evidence: tests/test_scraper_screenscraper.py (recap totals and per-game kinds, a second
job telling kept from new), tests/webui_scraper_browser.cjs (recap tiles, pills,
filters, search; the media-you-have hints). tools/verify.sh: PASS (format unit build
integration evidence).
- Sign-in: clicking ScreenScraper opens a login form. The name and password go to the
console in a request body, are checked with ScreenScraper (ssuserInfos), then kept in
config/private/screenscraper.cfg (0600, folder 0700, out of the 0777 repair, never
served). No answer carries the password. A green check on the chip when signed in;
Manage shows the quota, switches account or signs out.
- Engine: jeuInfos by CRC, file name and size, then jeuRecherche ("Name [id]"
candidates; the same title taken unasked, others resolved from the page). Details in
the language asked for (description, genre, developer, publisher, players, rating
0-1, release date); media per kind and region; the account's maxthreads kept; 429
waits; 401/403/423/426/430/431 pause the job with a message, resumable.
- The developer account: generated scrambled into build/screenscraper_dev.h from .env;
releases carry it scrambled (owner's decision), PS5_SCREENSCRAPER_EXCLUDE=1 leaves it
out. https only (media links upgraded, no https->http redirects). PC mode is refused
for ScreenScraper: its media links carry the developer account, which never leaves
the console. lint-shell fails on a plain copy in any tracked or shipped file.
- ES-DE: release dates become YYYYMMDDT000000.
- Systems table generated from ES-DE's ScreenScraper ids (MIT notice kept).
Evidence: evidence/scraper-screenscraper (6 host tests against a fake ScreenScraper API,
fake accounts, plus the leak check); reproduce with
python3 -m unittest -v tests.test_scraper_screenscraper; bash tools/lint-shell.sh
tools/verify.sh: PASS (format unit build integration evidence).
Not verified: a live sign-in and job against api.screenscraper.fr (owner's run).
Flags for USA, Europe, Japan and Korea and a globe for World, on the region choice
(now flag chips, like the other choices) and in place of a game card's region tag; a
console icon for PS5 and a PC icon for This PC → PS5; a badge for each source:
RetroArch's own invader for libretro thumbnails, and original drawings for
ScreenScraper, LaunchBox and EmuMovies (not their owners' logos, which are theirs to
license). Inline SVG, no image requests; the console icon is a generic outline.
Checked against the console at 1440 px (the Download Media tab, a scraped SNES card).
Command: tools/verify.sh: all five PASS (149 tests).
The scrape panel left the Games page for a Download Media tab: "Download from" PS5
or This PC → PS5 (the PC choice says it downloads on the PC, then transfers to the
PS5), source chips, media tiles with descriptions on hover or focus, a Games scope,
region and a replace switch, a live summary and Start bar, animated job stats and the
games that need the user folded under one summary. The tab's state is declared before
the page runs (opened straight on #media it failed: library read before its let).
- Jobs keep their download counters across restarts; hidden files a dead writer left
in the library are swept at start (older than ten minutes).
- run-title --menu-capture=N pictures the whole frame, menu included, at frame N.
Evidence: evidence/retroarch-shared-media (base PS5 13.40): the XMB History tab
captured at frame 900 with Donkey Kong Country 2's scraped box art drawn, and no
/app0/thumbnails folder on the console: the art is the shared library's. The tab
checked at 1440 px and 390 px against the console.
Command: tools/run-title.sh --pad-script=... --menu-capture=900; tools/verify.sh: all
five PASS (149 tests).
The WebUI's Games page scrapes box art, screenshots, title screens and logos from
libretro's thumbnail server into /app0/library/<system>/<kind>/<game>, one store for
every frontend: RetroArch's thumbnail lookup asks it first (patch 0114, which also
gives thumbnails_directory "default" its folder: it was empty and stopped every
lookup), EmulationStation has it as its media folder and fills its game lists from
its .meta files, the Games page shows it, a future frontend calls ps5_library_media.
- src/scraper.{h,cpp}: jobs saved and resumed without repeats, exact names, exact
titles tags aside, ranked candidates for the user, fill-missing by default; PS5 mode
on the daemon (runs with the page and the title closed) and PC mode through
webui/ps5-media-helper.py (the console identifies; the bytes pass through the PC).
- The page: the method asked first and remembered on the console, sources and every
media type described, version tags on each card. ScreenScraper, LaunchBox and
EmuMovies are listed as coming (ScreenScraper's developer credentials requested).
- The daemon's HTTP is libcurl over mbedTLS 3.6 with the shipped CA bundle (Sony's SSL
failed every handshake from the payload, 0x8095f00c); its updater uses it too.
- The shared library: one path a game read from outside the title (14 games were
listed twice), a disc's listed .bin dropped beside its .cue.
- The daemon: workers on 1 MiB stacks (a 64 KiB frame overflowed one), stale hand-over
on transfers in flight rather than idle keep-alive connections, pid in its log.
Evidence: evidence/scraper-libretro (base PS5 13.40): every system in 148 s (520
complete, 28 partly, 181 kept, 61 ambiguous, 234 unmatched); cached runs fetch
nothing; HTTPS through curl; PC mode with the helper killed: 90 images each sent once;
a job finished with the title closed; 1,024 games, 0 duplicates. Not seen live:
RetroArch and ES-DE drawing the media (owner to look), updates through the daemon.
Command: build/tmp/scraper/live.py; tools/verify.sh: all five PASS.
Gates run inside an if, where bash ignores set -e, so a gate reported only its last
command: a failing test in make test-integration passed the integration gate because
check-notices.py, after it, passed (seen 2026-10-07 on a flaky scraper test). Each
command but the last now returns on failure, in the format and integration gates.
Command: tools/verify.sh: all five PASS, 148 tests OK in unit and in integration.
Measured first: one libmicrohttpd thread with 64 KiB a connection held uploads to
13 MB/s, and four connections to 15.3, on a link where FTP reached 25 / 30.5 MB/s.
PS5-Suite's design was reviewed (persistent connection pool, small files per
connection, large files as pwrite ranges, big socket buffers) and an engine built
from scratch on the WebUI's HTTP (src/webui_transfer.{h,cpp}, docs/WEBUI.md):
- a pool of four server threads, 512 KiB a connection, 4 MiB socket buffers, writes
in 1 MiB blocks; non-transfer routes under one lock as on the single thread;
- PUT /api/upload/batch: many small files in one request (a framed stream), each
renamed into place when whole, existing names skipped and named unless replace;
- upload sessions (/api/upload/session, /part, /commit): a large file's ranges in
any order on any number of connections, committed once covered;
- Range downloads (206); every upload still atomic, never over an existing name;
- the page plans batches, sessions and single uploads onto six lanes.
Evidence: evidence/webui-transfer-engine (base PS5 13.40, PC on Wi-Fi 6). Uploads
26.0 MB/s on one connection and 29.9 MB/s for one file in six parts (FTP: 25 /
30.5); 400 x 32 KiB at 215 files/s (FTP 105). The page against the console: 605
files in 20.1 s, all equal by FTP readback, duplicates refused, a cancelled 200 MiB
file left nothing. ThreadSanitizer under a mixed parallel load: no report.
Command: build/tmp/transfer/bench.py; node tests/webui_transfer_browser.cjs;
tools/verify.sh: all five PASS.
Remaps never worked: the config's input_remapping_directory "default" was taken
as a folder named "default" (upstream resets it for its other folders, not this
one) and the port set no default remap folder, so Save Game/Core Remap File
failed with default/<core>/<game>.rmp and no remap ever loaded. The PS5 frontend
now names /app0/config/remaps (created 0777), and patch 0113 resolves "default"
to it. Game overrides and game core options already saved and loaded.
Evidence: evidence/remaps-per-game (base PS5 13.40, Snes9x, DKC2, Quick Menu by
pad script). Before: "Error saving remap file.: default/Snes9x/Snes9x.rmp". After:
game remap saved and "Game-specific remap found" on the next launch; core remap
saved (input_player1_btn_b = "8") and "Core-specific remap found"; game overrides
and game core options found on every launch. Test files deleted afterwards.
Command: tools/run-title.sh --game=snes9x:<DKC2> --pad-script=build/tmp/remap/*.txt;
tools/verify.sh: all five PASS.
The WebUI only ran inside RetroArch: a frontend change (LoadExec) ended it with
every upload in flight. It is now a payload of its own (daemon/webui_daemon.cpp,
webui/ps5-retroarch-webui.elf) serving the title's folder on 6769. eboot.bin, the
picker and ES-DE each keep a link to it (src/webui_link.h, 127.0.0.1:6770) and the
first to find none starts it through the ELF loader (9021) or websrv (8080 /hbldr).
It stops 20 s after the title closes, never mid-transfer; an update's Install
closes the title and the daemon installs. With no loader, eboot.bin serves the
WebUI while RetroArch runs, as before. The page names the frontend shown.
Downloads were handed to MHD as a descriptor (sendfile), which stopped at 33,300
bytes on the console; they are read in 256 KiB blocks.
Evidence: evidence/webui-daemon (base PS5 13.40). 200 MiB upload across picker ->
RetroArch -> picker and 60 MiB across picker -> ES-DE: one token throughout, 201
Created, every 256 KiB block equal on the console; daemon stopped 20 s after the
title closed; a 1,917,648,896-byte download's sha256 equals the FTP copy.
Command: tools/run-title.sh --picker-test=600:retroarch --retroarch-frames=1800
--watch 110 with build/tmp/webui-handover.py; tools/verify.sh: all five PASS.
Host tests: tests/test_webui_link.py, tests/test_webui_daemon.py.
Users reported FTP permission refusals again. Since v0.6.7 the title usually
hands itself over at once (to the picker or ES-DE, through LoadExec, which ends
the process), so the start-up repair that opens /app0 to FTP was cut short; and
the picker and ES-DE, other executables, wrote with their own modes.
- Files are 0777 as folders are (were 0666): the link-time wrappers and the
repair (src/permissions_ps5.cpp).
- ps5_permissions_settle waits for the repair (at most a minute) before every
handover: to a frontend, closing the title, the relaunch test.
- ES-DE's library files and folders, the remembered frontend
(src/ps5_frontend_choice.h) and the picker's test record are 0777; the picker
opens RADV's shader cache to FTP after its first 30 frames and again when it
hands over.
Verified on the base PS5: the trace reads "permissions: /app0 checked=17992
repaired=174 refused=0" then "handover waited 596.2 ms"; after a home-screen
launch held at the picker and closed there, an FTP walk of
/data/homebrew/PPSA99169 finds 17992 entries, 0 not 0777 (before: 175 files
at 0666) (evidence/permissions-0777). tools/verify.sh format, unit and
evidence PASS.
ES-DE's VideoFFmpegComponent ends its audio filter graph with asetnsamples,
which tools/build-ffmpeg-esde.sh did not enable. The graph failed
("setupAudioFilters(): Couldn't add graph filter: Filter not found") and
every game video stayed a black frame, with no sound, while images showed.
Verified on the base PS5 with a synthetic H.264/AAC clip as a game's video
(and a cover, since Alekfull NX shows its basic variant, with no video, for a
game with neither cover nor miximage): before, a black frame and the error;
after, the clip plays in real time, its burned-in clock advancing with the
captures (5.3, 9.3, 15.3 s at 10, 14, 20 s), sound playing, no filter error.
es-de.bin carries the filter. The owner's ES-DE settings were restored and the
test media removed. evidence/issues-25-27 (both issues); tools/verify.sh
format, unit and evidence PASS (92 captures).
ES-DE's library came from RetroArch's playlists alone, so a game copied into
content/ appeared only after a Manual Scan in RetroArch. ps5_library_load_content
now also scans content roots after the playlists (a playlist's entry still
wins): /app0/content and the USB and extended drives ShadowMountPlus mounts.
- Each folder at a root's top that names a platform the title has a core for
is that system's; the platform table gains the systems the title now runs
(3DO, DOS, ScummVM, Neo Geo CD, NGPC, WonderSwan Color, SuperGrafx, PC-FX,
Virtual Boy, Pokemon Mini, Atari 5200 and Jaguar, NAOMI, Atomiswave), most
specific first ("NeoGeo CD" is not the arcade Neo Geo).
- In it, every file a core of the platform takes; .zip and .7z too for a
cartridge system (RetroArch opens them), not for a disc system (an archived
disc's tracks are not extracted). Left out: firmware (named in a core's
info, a variant of one, or [BIOS]/Kickstart/boot ROM), and the tracks of a
folder that has a .cue, .gdi or .m3u.
- A DOS game is a folder started by the program named like it (doom/DOOM.EXE);
a ScummVM game is a folder started from a data file in it.
- A path too long for the buffers is skipped, never truncated.
es-de.bin passes the roots (frontends/es-de/ps5/library_ps5.cpp).
Verified: tests/ps5_library_test.c (folder names, firmware, tracks, DOS,
ScummVM, a drive, archives by kind, no duplicates) and esde_library_test PASS.
On the base PS5: ES-DE lists 36 systems and 1025 games; E.T. (Fixed), found
only by the scan, starts in RetroArch (Stella) and returns to ES-DE
(evidence/issues-25-27, next commit). tools/verify.sh format and unit PASS.
Each from my local fork with a PS5 platform commit, through
tools/core-fork.sh:
- Opera (PS5_Opera, 3DO): FreeDO's modified LGPL forbids commercial use,
so it joins the non-commercial parts (text from its sources' headers,
tooling/notices/texts/opera-freedo.txt).
- PUAE (PS5_PUAE, Amiga): GPL-2.0-or-later; it boots Kickstart 1.3 from
system/kick34005.A500, AROS without one.
- NeoCD (PS5_NeoCD, Neo Geo CD): LGPL-3.0-or-later; BIOS in system/neocd/.
- ScummVM (PS5_ScummVM: scummvm/scummvm v2026.3.0, every engine, no
OpenGL): GPL-3.0-or-later. Its Makefile empties LDFLAGS, so the title's
link flags go in as PS5_LDFLAGS (without them the wrapper linked
libSceNet, which the core loader refuses); AR carries its flags (ar cru);
its pinned libretro deps are kept across clean checkouts.
build-title.sh stages the four; notices list them; README: systems, forks,
credits, licences (thirty-three cores).
Verified on the base PS5 (13.40), runs 224859-230600: BattleSport (from
its .iso: the archive's .cue names a file it does not hold), 3D Pool,
Metal Slug and Day of the Tentacle load, run and exit with 0
(evidence/four-cores-2026-10-06). Disc images inside archives do not load:
RetroArch extracts only the .cue; unpacked into folders on the console
they do. tools/verify.sh format, unit and evidence PASS (91 captures).
Before a release: publish the four forks.
FreeBSD's <ctype.h> inlines isalpha() and the rest as a look-up in
__getCurrentRuneLocale()'s table below _CACHED_RUNES and calls ___runetype()
above it. The console's libc exports the classification functions but
neither of these, so ScummVM's core left the title's link with both
undefined. The table is built once from the exported isalpha, isdigit,
tolower and the rest. It sits outside the PS5_RETROARCH_RADV guard: the
platform layer that replaces the other shims in a RADV build has neither.
Verified: the file compiles with and without PS5_RETROARCH_RADV and defines
both; the title links with ScummVM, whose Day of the Tentacle runs on the
base PS5 (evidence/four-cores-2026-10-06, next commit). tools/verify.sh
format and unit PASS.
Each from my fork beside this repository (all local, not yet on GitHub),
upstream plus a PS5 platform commit (no --no-undefined, no LTO, zstd without
trace hooks), built through tools/core-fork.sh:
- PicoDrive (PS5_PicoDrive): the 32X SH-2s on the interpreter (the
recompiler's static read-write-execute buffer faults on the PS5); frames
go out as XRGB8888. Non-commercial licence, like Snes9x and FBNeo.
- Beetle PCE (mednafen_pce, with SuperGrafx and CD), Beetle PC-FX, Beetle
VB, Beetle NeoPop, Beetle Cygne: GPL-2.0-or-later.
- PokeMini (GPL-3.0-or-later), Handy (zlib; its text is in the source
headers, tooling/notices/texts/handy-zlib.txt), Virtual Jaguar
(GPL-3.0-or-later), Stella 7.0c, a5200 (built-in Altirra BIOS; its
Makefile empties LDFLAGS, so the link flags go in as PS5_LDFLAGS),
ProSystem: GPL-2.0-or-later unless noted.
- DOSBox Pure 1.0-preview7 (now on Codeberg): the dynamic x86 core's code
cache is ps5_exec_allocate memory, and stdio replaces the VFS wrapper,
whose funopen the console lacks.
build-title.sh stages all thirteen; the notices list each; README: the
systems, forks, credits and licences (twenty-nine cores). On the console the
owner's BIOS files were moved where the cores look: system/pcfx.rom,
system/lynxboot.img, system/bios.min.
Verified on the base PS5 (13.40), one game per core with
tests/pad/boot-and-start.txt (runs 212433-220552): every core loads, runs at
its system's rate (Virtual Boy 50.3, Pokemon Mini 72, the rest 60 frames a
second) and exits with 0; pictures inspected
(evidence/new-cores-2026-10-06). tools/verify.sh format, unit and evidence
PASS (90 captures, 0 failed); build and integration not rerun this step.
Before a release: publish the thirteen forks the notices and README name.
PicoDrive's RGB565 frames (320 pixels, rows of exactly 640 bytes, so the
upload's one-memcpy path) reached the screen half-height, doubled and
miscoloured; Beetle PCE's padded RGB565 rows and every 32-bit frame were
right. In gfx/drivers/vulkan.c the core frame texture is now always
B8G8R8A8, and a 16-bit frame is widened row by row as it is copied in. A
16-bit core is not lent the frontend's framebuffer (runloop.c's
GET_CURRENT_SOFTWARE_FRAMEBUFFER), since that texture is no longer its
format; the guard is in the runloop because withdrawn 0077 replaces
vulkan_get_current_sw_framebuffer while the edits apply.
Verified on the base PS5: PokeMini, Handy, Beetle NeoPop, Beetle Cygne,
a5200 and ProSystem, all RGB565, show correct pictures
(evidence/new-cores-2026-10-06, next commit). tests/test_frontend.py pins
the block count at 249 (three added). tools/verify.sh format and unit PASS.
tools/build-flycast.sh builds Flycast's libretro core from my fork
../PS5_Flycast (b95c532f: upstream flyinghead/flycast 59ed35a7e plus five
commits) through tools/core-fork.sh, like the other forked cores:
- the PS5 libretro build: no --no-undefined, 16 KiB pages;
- the SH-4, ARM7 and AICA DSP recompilers' code in the console's executable
memory (ps5_exec_allocate, near the core's code);
- anisotropic filtering takes the level chosen, not its index (upstream
bug: "16" gave 4x, the default "4" gave 2x);
- PS5 defaults: 2880x2160, per-pixel transparency, 16x filtering; the
widescreen hack stays off;
- no locale switching and no serial pseudo-terminal: the console exports
neither uselocale/newlocale nor grantpt/unlockpt/ptsname, which the
title's link refused.
The port began as rpf16rj's (pull request 3), credited in the fork's
commits; its networking (DCNet DNS and socket changes) and the
ps5_dc_rumble hook, which the title does not export, are left out.
tooling/flycast/ps5-toolchain.cmake is the cores' link contract with
libzip's Annex K and Windows probes answered no. The notices list Flycast
as GPL-2.0-or-later: its LICENSE is the GPLv2 text, and its file headers
say "version 2 or any later version", none "only". README: the system, its
resolution, the fork and the credit.
Verified on the base PS5 (13.40): runs 210509, 210719, 211230. The core
loads (core loader, Vulkan OIT renderer on RADV), Sonic Adventure (USA, Rev
A) boots through the HLE BIOS, plays its film, title screen and attract
demo at 2880x2160 with sound, and START reaches "Select a memory card";
RetroArch exits with 0 (evidence/flycast-sonic-adventure, pictures
klog/pad-210719 and klog/pad-211230). About 60 frames a second in the demo,
30 in the film (its own rate), early stalls of 0.5-1.4 s.
Gates: tools/verify.sh format, unit and evidence PASS (89 captures, 0
failed); build and integration wait for MAME's rebuild. Before a release:
publish github.com/mihawk-99/PS5_Flycast, which the notices and README name.
Mihawk