Data model: nodes are real content entries (writings, portfolio,
bookshelf), edges are shared-tag connections — the same relationship
"related posts" already uses, drawn instead of listed. No decorative
or fake data.
- src/lib/constellation-data.ts: builds the node/edge graph from the
actual collections at build time.
- src/lib/constellation-scene.ts: vanilla Three.js scene (no
framework) — fibonacci-sphere layout, auto-rotation, pointer
raycasting with a title tooltip, click-through to the entry's page.
Honors prefers-reduced-motion (skips rotation and the spring-eased
entrance, which reuses src/lib/spring.ts) and checks WebGL support,
falling back to a plain count for browsers/settings that lack it.
- src/components/Constellation.astro: wires it up, serializing the
graph as inline JSON for the client script to consume.
- Homepage rebuilt around it: hero constellation above a 3-column
dashboard (latest writings/portfolio/bookshelf), using a new `wide`
layout variant on Base.astro (dashboard needs more than the 42rem
article width).
Also: removes "life and beliefs" from all site copy (site.config.ts,
the llms.txt generator + its output) per request — description is
now "Computer science and mathematics."
Verified in-browser: real WebGL scene renders with correctly
positioned/colored nodes and edges, hover shows the right title via
raycasting, layout holds at the new wide container width.
scripts/generate-llms.astro-site.mjs covers writings, portfolio,
bookshelf, and blogs (unblocked now that site.config.ts exists from
the feeds/meta work), wired into apps/site's build script.
Also fixes .gitignore: the bare `public` pattern (meant to ignore the
Zola build's root-level output directory) was matching at any depth,
so apps/site/public/ and apps/cms/public/ were being silently
excluded from every commit despite `git add` reporting them as
staged for directory adds (the ignore-file warning only surfaces for
an explicitly named path, not a recursive directory add — easy to
miss). Scoped to `/public` so only the root output dir is ignored.
Picks up robots.txt, the Astro scaffold's favicon, llms.txt, and
Strapi's own public/robots.txt, none of which were actually tracked
until now.
apps/cms/src/utils/bookshelf-covers.ts fetches book covers from Open
Library by ISBN (real image, cached into apps/site/public/covers/)
and backfills missing title/authors for papers from Semantic Scholar
by DOI/arXiv id. Wired into the sync middleware in content-sync.ts:
runs once per entry (cover_image already set means "done"), and only
on live create/update through Strapi, not during the boot-time
import (so a restart doesn't hammer external APIs).
Also renames the bookshelf `status` field to `read_status`
everywhere (schema, content/README.md, seed content, Astro zod
schema, bookshelf page, sync mapping). Strapi's content-manager API
treats a top-level `status` key as its own reserved draft/publish
parameter even with draftAndPublish disabled, so an attribute
literally named `status` silently rejected every write — caught this
while testing the enrichment hook.
Verified end-to-end: created a book by ISBN with no cover_image,
confirmed a real cover image was downloaded and cached, cover_image
was set on the entry and written to the file, and the cover/file were
both removed on delete.
apps/cms/Dockerfile.dev + root docker-compose.yml give a one-command
way to run the CMS in a container for testing (docker compose up
cms), with the shared content/ directory mounted so it matches the
host dev setup. apps/cms/Dockerfile is a production multi-stage build
for later Dokploy deployment. Uses node:24-bookworm-slim to match the
local Node version and get glibc prebuilt binaries for better-sqlite3
without needing a native build toolchain in the final image.
Verified: both images build, the dev container boots Strapi
successfully, admin panel and API routes are reachable, and the
mounted content/ volume resolves at the same relative path the future
sync provider will expect.
andwati