When the render would inject the page title as an h1 (markdown.has_h1 is
False for the body), an article job's text carries the same '# {title}'
line: the title translates in document context and the opening paragraphs
see the heading (works both ways). The heading's pair in the decomposed
result becomes the title fragment — heading text only, pure prose, never
stored as a body chunk; a demoted or merged heading skips the title,
which stays pending for a scoped title job. The job covers the title key
so it is not double-dispatched afterwards.
Prompt: prefer established technical loanwords with English roots over
forced localizations (frontend -> frontti in Finnish, not etupääte).
Verified offline (title extraction, demoted-heading skip, no-injection
path) and live against ollama qwen3.8:27b: the standalone title job gave
'Alkuun pääseminen', the article job then overwrote it with the
in-context 'Aloitus' matching the body's translated heading.
Prompt rules are single-line bullets now (no wrapped continuation lines).
Added the extended-Markdown rules the site syntax needs: all formatting
is syntax and is preserved exactly (only the text is translated), and
single newlines inside paragraphs render as actual line breaks, so line
structure must survive untranslated. The payload is no longer wrapped in
a ```markdown fence — the site's extended syntax (and this document
renderer) makes fences unreliable as delimiters; instead an explicit
sentence marks the instruction/text boundary and the payload rides in
faux <translate> tags (<context> for the hybrid-neighbor reference
blocks). unwrap_output strips echoed markers, still also a whole-output
fence. Re-verified live against ollama qwen3.8:27b: title + article jobs,
URL/code fence/{dates} preserved.
Decomposes a pasted full translation with the same align_article
validation as article-mode LLM results and stores proper Data.trans
fragments, instead of the translation editor's one monolithic patch that
silently goes stale hunk by hunk. Runs against the kanta database
directly, with the server stopped. Verified on a scratch copy of the
localhost db: swe/app fi (qwen3.8 trial output) imports 43/43 blocks and
serves through the hybrid; the degenerate qwen3-next output is rejected
on anchor validation.
OpenAI Chat Completions shape for llama.cpp/hosted APIs, ollama native
/api/chat via api="ollama" — ollama's /v1 endpoint silently ignores
think:false (verified on 0.34.2: reasoning ran despite the flag), which
hybrid models need off. Prompts and sampling from the /tmp/llmtrial
evidence (strict structure rules, temperature 0.2, num_predict capped at
~2.5x estimated source tokens); whole-output fence unwrapping and
single-line enforcement for titles are client-side. Config via JSON +
CLI overrides; pagerite itself carries no LLM specifics.
Hello gains model + modes; Job gains mode. Dispatch routes per connection
capability: article jobs only to article-capable connections and only
while a page is mostly pending; titles/chunks go as markdown (whole
fragment, hybrid-neighbor contexts) or segments jobs. Validation skip
list is now (lang, key, mode). align_article decomposes a whole-article
result into per-chunk stores: non-translatable blocks (fences, HTML,
containers) are verbatim anchors, regions between anchors pair
positionally, mismatched regions and blocks with altered link
destinations/placeholders stay pending.
Verified offline against /tmp/llmtrial outputs: qwen3.8:27b runs pair
105/105 and 43/43 blocks; qwen3:30b-instruct's translated code comments,
qwen3-next's degenerate runs and the structurally broken seedx trial
files (a dropped ::: fence) are all rejected.
Extra paragraphs exist only inside a patch hunk's replace text (anchored
on the preceding original block). Deleting one produced a (X, "") hunk,
and apply_patch removed the text but not a surrounding separator, leaving
an accumulating blank line in the served/editor Markdown. hybrid_markdown
now re-chunks the patched text (join_chunks(chunk_markdown(...))), which
drops the stray blank lines on the read path — repairing gaps left by
already-stored patches too.
Adjacent fix in make_patch: a hunk whose search text occurred more than
once in the page hit the FIRST occurrence at apply time, blanking or
rewriting the wrong instance when identical paragraph text appeared
earlier. Ambiguous hunks now grow block context (preceding block first)
until unique; unique hunks stay context-free so they keep surviving
translation refreshes.
load() closed the editor on every navigation, and the close raced the
navigation: the shell's close-time loadPlain re-render resolved last,
swapping the old page back in and replaceStating its URL. Navigation now
leaves the panel open and the existing pagerite:switch-editor retarget
fires while editing; the exception is /_a, where closeEditor({ navigating:
true }) closes the panel without the re-swap or title restore.
- Banner-corner auth link is now a button opening paskia-js's profile()
dialog (handles login too); auth re-probed when the dialog closes.
- Editor/analytics components call /_api via apiFetch/apiJson: an
expired session opens the login dialog and the request retries.
- pagerite.js: task-checkbox toggle uses apiJson (explicit edit attempt,
reverts on any failure incl. cancelled login); auth probes use
fetchJson (never a dialog); page-cache/navigation stay on plain fetch.
- Add the paskia npm dependency; document the convention.
Preview: the editor render handler now returns the complete article HTML
({cards} tags expanded, children's cards appended when the page has no
tag) and the frontend swaps article.innerHTML wholesale — the old
detach/keep/re-append hack for .cards (which duplicated stacks) is gone.
Semantics: one card per item, no subtree flattening (stacks removed).
Bare {cards} lists immediate children (front page: the other top-level
items); a page-less item is represented by its first leaf page, the
nav-link logic. {cards: path} renders that page alone, path/* its
direct children, path/** all published descendants.
The language-switch trail ping read <html lang> before the view
transition applied the swap, so it reported the previous language (and
read pings never update trail-item languages server-side): multi-
language visits recorded as single-language trails and the analytics
flag combining had nothing to show. The switch ping now passes the
picked tag explicitly.
Language negotiation no longer prefers the article's original language
anywhere in the Accept-Language list — nearly every browser lists
English as a fallback, so translations were almost never served. The
first servable header language now wins, the original counting only in
its natural position.
Area and bar fills at 0.6 opacity (were a near-invisible 0.15); the day
view's 'Last 24 hours' legend entry uses the same full-opacity accent
line as the other entries instead of a translucent bar swatch.
Port the seasonal.py demo algorithm to JS (analytics/seasonal.js): the
smoothed 5-minute history is folded onto a weekly grid with exponential
age decay (7-day half-life for the time-of-day pattern, 42 days for
per-weekday deviations, shrunk by effective weeks of data), capped at
180 days of history where the weights become negligible.
Week view: the fading past-week overlays are replaced by the full
Monday-first 'Typical week' estimate. Day view gains the same estimate
cut to the rolling 24h window as a smooth muted 'Typical <weekday>'
curve behind the bars. Longer ranges are unchanged.
Favicon flush to the badge's left edge, then the referer host and UTM
summary in one link, with a single newline-formatted tooltip (origin,
then each utm_* pair). Fixed --badge-bg/--badge-text/--badge-muted
palette (defined in pagerite.css, deliberately unthemed) so transparent
favicons stay legible in both light and dark themes; own-article trail
links remain plain, lifting referer info visually apart from them.
Replaces the outlined utm-tag pill.
GETs store the resolved page language, activity pings carry <html lang>
(including a new ping on in-place language switches, which also keeps the
switch's GET out of crawler classification), and trail items/crawler hits
surface it. The viewer shows flags discreetly: nothing on single-language
sites or primary-language-only rows, one leading flag for uniform visits,
transition flags on mixed trails, a flag row per crawler.
- UA lines click to copy the raw UA; abuse rows with rotating clients
copy every variation, one per line (uaRaws)
- "Copied!" popup is viewport-fixed at the click point; cell overflow
clipped the old absolute one
- language tags show their Intl.DisplayNames full name on hover
- a 🔗 link after the pretty UA opens the crawler info page uarite
provides
- the emoji-symbol dim-until-hover idiom (opacity 0.7 -> 1) is now one
global .icon-btn class in pagerite.css, shared by the edit pens,
auth links, editor icon buttons and the new UA link
The stored ua_pretty froze each record at the uarite version of its
record time (old records showed disguised Meta crawlers as
"Chrome/145 Windows"). Client now carries a display-time-only
"uarite" field holding the full uarite.UA dataclass (pretty/engine/
os/provider/kind/url), filled when the viewer payload is built, so old
data always follows the current uarite. uarite 0.2.0 fixes the Meta
misdetection itself; _compact_user_agent is gone, tracking.py calls
uaparse directly.
A closing ::: glued to a paragraph (no blank line before it) rode inside
the prose chunk and crossed to the translator as part of the text run;
when the model dropped it, validation passed and the splice lost the
fence — the rest of the page rendered inside the container (seen in the
Spanish translation). Container fence lines (::: openers and closers
alike) are now always their own prose-free chunk, never reaching the
translator. Affected pages re-chunk on next save and re-translate under
the new hashes, repairing themselves.
LangSelect's scoped CSS landed on the shared store chunk, whose stylesheet
the editor never loaded (only the public selector path injected it), so the
editor's language selector rendered unstyled on untranslated pages. Collect
editor stylesheets from the entry's imported chunks too (same traversal as
the langselect assets).
Also: hreflang alternates now skip languages disabled site-wide, and all
selectors (page editor, structure tab, public selector) order languages the
same way — primary first, then the lang tab's geographic grouping.
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.
LeoVasanko