First half of replacing the stub text domain with a media collector. Nothing
references this yet - the store is standalone and fully tested before anything
depends on it.
Blobs are addressed by SHA-256 and sharded two levels deep, so the same picture
re-uploaded at a dozen addresses costs one file. Downloads stage in a sibling
temp directory on the same volume and are promoted by rename, which is what
keeps blobs/ free of truncated files: a crash leaves a stray .part that the next
startup sweeps, never a half-image indistinguishable from a real one.
The SQLite index holds provenance separately from content, so purging one source
leaves blobs another source still references - that is what ref_count buys, and
it is recomputed rather than incremented because the item upsert can replace a
row pointing at a different blob. The seen_url journal deliberately outlives a
purge: without that, the next run downloads again exactly what the user just
deleted. Terminal outcomes are split from retryable ones so a flaky network does
not permanently lose content.
The showcase gives every item a dated, named path via hard links - a second name
for one file, not a second file. Hard links are a filesystem privilege rather
than a guarantee, so it degrades to copying and records which it achieved; the
UI has to be able to admit that. Names suggested by the origin are treated as
hostile: only the last path segment survives, Windows device names are pushed
aside, and the extension comes from the sniffed kind, never from the remote.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The parser will need to move between proxies, so this adds the module it will
sit on: pluggable sources, a pool that hands proxies out and learns from the
outcome, three rotation strategies, and a page to drive it.
Sources are IProxySource implementations. The public proxifly/free-proxy-list
feed is fetched as the combined all/data.json through jsDelivr and filtered
locally — one conditional request beats four per-protocol ones that can disagree
mid-publish — and cached for the five minutes upstream takes to regenerate. A
feed that is down keeps serving its last payload rather than emptying the pool.
The user's own list lives in proxies.custom.json beside the settings, takes a
pasted blob, and names the lines it could not parse instead of quietly dropping
them.
Both knobs the pool exposes are settings, as asked: rotation is Sticky (default,
the only one that keeps site sessions coherent), RoundRobin or WeightedRandom;
liveness is either a parallel sweep of the whole pool or a probe at hand-out
time. Free lists are a few percent alive, so skipping verification entirely
means mostly waiting on timeouts.
Two invariants worth keeping, both of which cost a bug to find:
Availability is decided by the quarantine, never by Health. Excluding everything
that has ever failed made the quarantine window dead code and discarded proxies
permanently on their first hiccup, which is exactly wrong for addresses that
flap constantly. Health only orders the candidates now.
A probe verdict does not touch the success/failure counters. Those are about
real requests, and letting a sweep over a few thousand proxies rewrite them
would drown the evidence weighted selection reads.
SOCKS needs no extra package — .NET resolves socks4/socks4a/socks5 in WebProxy —
but a proxifly record with "protocol": "https" is still an HTTP proxy reached
over http:// with CONNECT, not an https:// scheme.
115 new tests. Also fixes a pre-existing flake: a command gated on another
command's IsExecuting cannot be driven straight after its Execute() completes,
because IsExecuting is published on the output scheduler.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Greenfield skeleton for a parser desktop app. The domain is deliberately a
placeholder — IParser<TIn,TOut> plus two sample parsers — so the shell is
runnable and verifiable end to end before real logic lands.
Layers run one way: Core (no Avalonia, no IO) <- Infrastructure <- UI <- Desktop.
UI is a class library rather than the exe so headless tests build real views
without dragging in Program.cs, Serilog or the container.
Adaptive layout is built from what Avalonia actually offers, since it has no
AdaptiveTrigger or media queries: ResponsiveLayout observes Visual.Bounds and
projects a breakpoint onto both an attached property and :compact/:medium/
:expanded pseudoclasses, with 24px hysteresis so dragging a window edge cannot
make the layout flap. Pane state lives in the view model because a style setter
loses to a local value permanently; styles own only the visual variance.
Stack notes worth remembering: Avalonia.ReactiveUI is deprecated in favour of
ReactiveUI.Avalonia, and ReactiveUI 24 runs on the Primitives engine (RxVoid,
ISequencer, Signal<T>) and no longer self-initialises. Avalonia.Headless.XUnit
12.x requires xUnit v3. InvariantGlobalization must stay false or Semi.Avalonia
throws in its static constructor.
102 tests across three projects, including headless guards for the two failures
that are otherwise completely silent: a stylesheet whose selectors match nothing,
and a light palette too low-contrast for cards to read.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Leonid PershinandClaude Opus 5