Replace the pre-classified store (visits/crawlers/abuse lists written at
collection time, plus abuse_ips and in-memory pending/session tables) with
a raw append-only log: one Get record per document GET (full path, true
HTTP status, referer origin, preload flag) and one Msg per /_ws activity
message. Visitor/crawler/abuse classification, visit grouping (30-minute
inactivity gap), status/referer/UTM attribution and all aggregates are
derived in Store.display(), so future rule changes never invalidate stored
data. The viewer payload keeps its exact shape.
Fixes structurally:
- Abuser 404s on slug-format paths showed up as "articles read": the
not-found branch recorded the request twice through separate status
plumbing. Each request is now recorded once with its true status.
- 404 trail links never rendered red: cache-served navigations issue no
GET and the only real GET (the idle preload) was discarded before status
recording. Preloads are now recorded with pre=True, never counted, and
used for status attribution.
- formatAbuseRows merged a path's 404 probes and 200 reads into one entry;
the collapse is now keyed by (path, status class).
Rule improvements enabled by the redesign:
- The plain-404 abuse threshold counts within a 1-hour sliding window, so
long-time readers accumulating misses never classify (scanners spray).
- Hidden (admin) clients never trigger abuse classification: editing means
visiting not-found pages.
- /.well-known/ probes (RFC 8615, e.g. Chrome devtools) are never abuse
evidence; //foo-style empty path segments are an instant telltale.
- Visits can no longer open on an external exit URL; favicon fetches skip
hidden clients' referers/exits.
Legacy analytics.json files are set aside as .bak-legacy on startup.
Past-week overlays can run far above the autoscaled y range, and the svg
is overflow: visible for the axis labels — wrap the data paths (bars,
skyline, area/line curves) in a clipped group so they cannot paint outside
the plot rect.
Visits with under 5 s of total reported reading time are JS-running bots:
display() converts them to crawler hits (one per trail page, with referer
and UTM query) and excludes them from every aggregate. Crawler referers
join the favicon fetch origins and render with their icon in the crawler
table. Abuse hits now record the real response status, so the abuse table
splits 404 probes from the articles the abuser actually read (200 GETs),
shown as trail links like the visitor/crawler tables.
All pending titles are now offered before any article chunks (stable
sort, menu order kept within each kind) — a page's name in the menu is
its most visible string.
Every request-driven kanta transaction now passes user= from the
Remote-User header the SSO/forward-auth proxy sets (the editor socket
reads it from the WebSocket headers; the translator worker keeps its
client key). Action labels are short identifiers naming the object, not
sentences: page / page:{lang} / page:title / page:{lang}:title /
page:language / page:slug / page:delete / structure:reorder / settings /
translate:reset for admin actions, translate:{lang}[ :title ] for worker
submissions (title results identified via the job kind, now tracked in
the connection state).
Weight-ratio mapping alone was routinely off by a word and could glue a
mark to its neighbor (losing the space between). Now each mark's source
words are aligned to the translation's words by form similarity
(sequence ratio + shared prefix, case-folded, capitalization bonus) with
cheap skip penalties, so inflection, dropped articles/prepositions and
reordering don't break the match; slices are cut exactly at word
boundaries. Alignments without an anchor pair fall back to the weight
ratio (still the CJK path).
Implement comprehensive content localization, admin panels for editing each language, AI translation interface with automatic updates when base language version is changed.
- SEO tags for all language URLs
- Uses accept-language by default, ?lang=en overrides temporarily
- User edits patched on top of translations
- RTL language supportReviewed-on: #1
word-break: keep-all does not suppress breaks at hard hyphens (an
explicit UAX #14 break opportunity), so `--arg` could split after the
dashes. Inline code is short — nowrap is the safe fix.
Space-separated brace attrs on a lone image's own line (e.g.
{style="max-width: 40em"}) are consumed onto the paragraph but leave
an empty text token in the inline children, which defeated the
lone-image check — the image stayed a bare inline <img> without the
figure wrapper (and without lightbox/zoom treatment). Strip empty text
tokens before the check.
Clicking a figure image opens a full-viewport overlay: the image as
large as fits (100vw, flex-shrunk to leave exactly the caption's
height) with the caption below. Click or any key closes it; wheel and
touch scrolling stop at the overlay (overscroll-behavior: contain).
Styling lives in the base theme: blurred dark backdrop (theme-tunable
via --lightbox-bg / --lightbox-text), soft shadow, reduced-motion-aware
open animation.
::: aside / .margin blocks no longer split the column segments: they
stay inside the .colseg at their anchor point, so the surrounding text
is measured and laid out as a single columned layout (split-off sides
previously lost .cols when too short on their own).
The zone placements (multicol side zone, sidebar track, wide gutter)
now position the boxes absolutely off the article's left border —
unaffected by any column layout inside — with the vertical spot coming
from the unset top (where the box occurs in the text). Narrow widths
keep the in-column float fallback. Trade-off: out-of-flow boxes no
longer stack via clear, so boxes anchored close together may overlap.
Inline code gets padding-inline, keep-all word-break, and a color nudged
30% toward --muted from the inherited color, so code inside
accent-colored text keeps its hue.
All color-mix calls now rely on the default interpolation space (oklab)
instead of declaring srgb/oklab explicitly.
migrate_v1 now also rebuilds the legacy flat pages store as the menu
tree (moved from the app lifespan, raw-dict level); migrate_v2 now also
backfills missing AVIF/WebP/JPEG derivatives on disk (moved from the
lifespan) and drops the obsolete version field. The version render
counter was cache-invalidation state, not database state: replaced by an
in-memory render generation that clears the page-body LRU and feeds page
ETags. The legacy Page struct and Data.pages/version fields are removed;
old databases lose the stale keys on re-serialization.
Uploaded images (SVGs rasterized, GIFs excepted) are stored as the
original (hash.orig.ext, internal only, never served) plus AVIF primary
and WebP/JPEG fallback derivatives re-encoded from it. Pages link the
bare hash; the server serves a format only when Accept lists it
explicitly (image/avif -> AVIF, image/webp -> WebP, else JPEG) with
vary: accept, while an explicit extension pins the format. Favicons go
through the same pipeline at 192px. migrate_v2 rewrites old
/_f/{hash}.avif article links, a startup backfill creates missing
derivatives, and twitter:image pins the .webp variant for X's scraper.
uploadImage and insertTable no longer inject at the cursor: on a
non-empty line (e.g. inside an existing image tag) the block goes on a
fresh blank-separated line after it, never into it.
PUT /_api/files now runs raster uploads through mediapreview.dispatch
(temp file for format routing: pyvips, ffmpeg for HEIC/HEIF/AVIF),
storing the untouched original as <hash>.orig<ext> and serving the
AVIF derivative <hash>.avif in links. SVG/GIF and failed conversions
fall back to plain <hash><ext> storage. FileStore.delete removes the
whole hash pair. Adds mediapreview[standard] dependency.
Cursor-driven editor→page scroll sync pinned the cursor's page position
at a fixed window height, so every cursor move dragged the page along.
Now the page scrolls only when the cursor's mapped position crosses a
viewport edge margin, and just enough to bring it back inside.
- Image insert always adds an empty "" caption with the cursor inside.
- Fenced blocks (``` code, ::: aside) share one toggle: clicked inside
one it is removed and the content selected; otherwise the selection
(expanded to whole lines) is wrapped, cursor left on the opener line.
- Code button: inline wrap toggles (selection preserved, backtick runs),
line-spanning selections make fenced blocks.
- Link button toggles: clicked inside [label](url) it unwraps.
- Block classes via two pickers (placement, AA size) with current-class
indication and a "normal" reset; placement replaces ::: container
names, fences get their own attribute line.
- New .small/.large/.huge text-size classes (0.7/1.5/3em); base body is
exactly 1rem so the scale is uniform.
- .left/.right floats generalized from figures to any block.
- Toolbar polish: non-emoji glyphs, scaled-up symbols, borderless
hover/active states, reordered (B/i and size last).
- Editor panel no longer closes on Escape.
- CodeMirror editors capture Tab/Shift-Tab for indent/dedent.
_block_attrs now applies a trailing {...} to the block when it ends the
last text line after whitespace (some text {.small}), not only on a line
of its own; a space is what keeps the braces off an image/link ending the
line. Glued-to-text braces stay literal.
Nitro's orange accent selection fill clashed with accent-colored text.
Introduce --selection-bg (base: accent 30% mix, as before) used by both
::selection and the editors' selection layer; nitro overrides it with a
neutral grey. CodeMirror's selection needed a baseTheme with its exact
&light/&dark selectors to win, and is now hidden when unfocused like a
normal input.
Live-filtering through slugify ate hyphens and spaces (trailing hyphens
are stripped) and jumped the cursor to the end on mid-text edits. Now
only spaces→hyphens and lowercasing happen oninput (length-preserving,
cursor stays put); the full slugify runs at commit before the server
call, which keeps its own validation.
LeoVasanko