👻 Revenant
Which of your dependencies are already ghosts?
Revenant predicts which open-source projects are about to die, holds a funeral for the ones that already did, and helps you resurrect your code with a living successor.
Two readings of "ghost in the machine" at once: zombie dependencies haunting your
dependency tree — and, in the spirit of Día de los Muertos, an ofrenda that honors
the projects (and people) that gave us so much before they went quiet.
Signals, not verdicts. A ghost score is a probability estimate, never an accusation. Maintainers owe nobody anything.
Live demo · Screenshots · What it does · How it works · Use cases · Quick start · Verified numbers
Live demo
A brand-new checkout — or a free Render instance with zero env vars — boots with a fully
populated product: the seed index and showcase tombstones ship in-repo, and every missing
credential degrades to a named, honest fallback instead of a broken page.
Screenshots
| Scan & land | Haunting report |
|---|
| |
| Tombstone & Ofrenda | Cementerio index |
|---|
| |
Full desktop + mobile set (16 shots, verified zero horizontal scroll at 390px) in docs/screenshots/.
What it does
🔍 Scan
Paste a package.json, requirements.txt, or pyproject.toml — or scan a public repo
URL — or exhume a single package by name. The scanner streams progress per phase
(parse → resolve → signals → score → autopsy) over SSE, so a 40-dependency manifest
never looks hung. Unknown registries and missing packages degrade to an explicit
Unknown band instead of a fake green.
📊 Haunting Report
Every dependency gets a ghost score (0–100) — the estimated probability of
abandonment within 12 months — and a band: Alive / Fading / Haunted / Departed /
Unknown. The report page carries a stacked band-distribution bar, URL-synced
filters and sorting, humanized times ("3.6 years ago"), a copy-as-Markdown table
for pasting into PRs, a downloadable PNG card, and an inline Resurrect action
on every dead or dying row.
⚰️ Tombstones
Each departed project gets a full memorial page: vital signs (commit-pulse
sparkline with a flatline marker, release gap, bus factor, issue-response latency),
a cause of death with numbered, kind-labeled evidence items, a score
contribution chart showing exactly which signals pushed the score up or down
(heuristic-v1 or tabpfn, always labeled), a dignified epitaph, the project's
last words, and living successors. Nonexistent packages return a real HTTP 404 —
a tombstone is never faked.
🕯️ Ofrenda
Under each tombstone, a three-tier altar renders every recorded contributor as a
candle scaled by commit share — papel picado across the top, marigold petals on the
steps, name cards beneath. You can light a candle (persisted locally, no
backend). Where contributor records are unavailable, a single candle burns with an
honest "records unavailable" note.
🕯️ Seance
Talk to the spirit of a project. Seance answers are retrieval-grounded — the
model can only speak from the project's own commit history, releases, registry
records, and docs. When the corpus doesn't contain an answer, the spirit says
"I don't remember that" rather than hallucinating, and every reply carries a
grounded/unverified badge plus evidence citations.
🌱 Resurrection
For anything dead or dying, Revenant ranks maintained successors using registry
pointers, dependency-graph data, and successor metadata — deterministically, so it
works with zero API keys — then enriches the migration plan with an LLM when one is
configured. The report stores dependency specs so the UI shows a real manifest
diff (remove old, add successor). The "Open PR" button stays disabled with an
honest tooltip until GitHub sign-in is genuinely configured — nothing is ever
committed without explicit human approval.
🪦 Cementerio
The Graveyard Index: 180 real npm + PyPI packages ranked by ghost risk, with a
"Recently departed" strip, band filter chips with live counts, sorting, debounced
search (still functional as a plain GET form without JavaScript), pagination, and
"mobile card" layout below sm.
⚔️ Arena
The same autopsy, run across open-weight models side by side — with latency bars,
precomputed real bouts labeled with their snapshot date, and accessible voting.
Judges see actual Gemma 4 output even on a zero-credential deploy.
🔬 Science
Dataset card, methodology, signal-attribution weights for heuristic-v1, live index
stats, and honest metrics — ROC-AUC/PR-AUC charts render when
data/eval-results.json exists, otherwise an honest date-labeled empty state.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
