Release Notes
Added
- Maintenance signal on matches. Where a source publishes a last-updated
date, matches now carry one. The detail view reads updated 3 years ago, and
anything untouched for two years or more is flagged ⚑ in the results table
— flagged, never demoted: an unmaintained tool that exists is still prior art,
so it keeps its rank and only gains a marker
- 8 of the 18 sources report a date in the same search response the adapter
already fetches, so nothing costs an extra request: crates.io (updated_at),
npm (package.date), GitHub (pushed_at), Hex (updated_at), Maven
(timestamp), AUR (LastModified), Artifact Hub (ts), and pkg.go.dev
(scraped from its rendered publication date). The remaining sources publish
no date in their search results, and fetching one per match would mean a
request per row that the per-source timeout has no room for; those matches
show no date at all, which is not a claim that they are stale — unless
another source returned the same URL, in which case dedup fills the gap and
the row shows that source's date (see Fixed, below). Hacker News
is left blank deliberately — it exposes when a thread was posted, which says
nothing about whether the thing discussed is still maintained.
src/sources/mod.rs carries the full table, so the gaps read as decided
rather than unfinished
--json output gained a last_updated key on every match: a whole-second
RFC 3339 UTC timestamp, or null where the source reports none.
schema_version stays 1 — the change is purely additive, so existing
consumers keep working untouched- The LLM prompt now shows each match's age, with explicit guidance to treat it
as colour on the prior art and never as a reason to discount it: a tool that
exists but looks unmaintained is still proof the idea has been built, and an
absent date is not evidence of staleness
patent::freshness: normalises every registry's date format — RFC 3339 at
any sub-second precision, epoch seconds, epoch milliseconds, and
pkg.go.dev's rendered Feb 28, 2026 — to one canonical shape, then turns it
into a display label and a staleness flag. Every constructor returns Option
rather than Result, so a registry that changes its date format degrades
that match to "no date known" instead of taking the whole source down. Date
fields are also deserialised leniently: a value whose type changed upstream
would otherwise abort the entire response and put the source dark- The
⚑ marker is documented in the TUI help overlay (?) and in the README
Changed
- The absence-claim scrubber now also catches absence dressed up as
maintenance. Showing the model dates opens a second way to make the one claim
this tool must never make — not "nothing exists" but "no actively maintained
tool exists", which is just as unprovable: only 8 of the 18 sources publish a
date at all, so an undated match is not an unmaintained one. Such headlines
and gaps are now replaced or dropped like any other absence claim, and hedges
("no currently maintained tool") no longer walk one past the filter. Saying
a specific match looks unmaintained is untouched — that is read straight off
the data shown, and is the whole point of the signal
- Ranking is deliberately left similarity-only; recency is not blended into
the score. verdict.rs derives the saturation floor from the similarity
data, so deflating similarity for age would quietly weaken the integrity
guard itself and let a well-trodden space read as "Open". Age belongs in the
output, where a human and the model can weigh it — the rationale is recorded
on rank::score_sort_limit so it is not silently undone later
Fixed
- Dedup no longer discards what a duplicate knew. When two sources return the
same URL — routinely, since 35% of Homebrew formulae have a homepage
byte-identical to a GitHub repo URL — only the first arrival survived, and the
fan-out iterates a HashSet, so which one that is is randomised per process.
The same query really did show a last-updated date and a star count on one run
and neither on the next. The kept match now fills its empty last_updated and
popularity from the duplicates behind it; identity, ordering and every field
it already had are untouched. This predates the maintenance signal — recency
is simply the first field where losing the race became visible. last_updated
and popularity are the whole of it: they are the only optional fields on
Match. Which source is credited for a shared URL still varies per run, so
the name and description shown are still whichever source won the race — that
is untouched here, and no data is lost to it
Breaking (library API)
Match gained a public last_updated: Option<String> field. Match has no
Default impl, so downstream code that constructs one with a struct literal
must add last_updated: None. Reading and matching on existing fields is
unaffected
Install patent 0.10.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/r14dd/patent/releases/download/v0.10.0/patent-installer.sh | sh
Install prebuilt binaries via Homebrew
brew install r14dd/patent/patent
Download patent 0.10.0
