Implements the last Phase 2 piece: a working face detector/embedder behind
the new `faces-onnx` cargo feature (mirrors how `plugins` gates wasmtime).
Inert by default — the default build is unchanged and ships the no-op
analyzer.
Pipeline (InsightFace/immich pattern): SCRFD detection with 5-point
landmarks → least-squares similarity alignment to the canonical 112×112
template → ArcFace embedding → L2-normalized 512-d vector.
- face_geometry.rs (always compiled, unit-tested): SCRFD anchor/distance
decode, NMS, the closed-form (complex-number) similarity transform,
bilinear affine warp, NCHW normalization, L2-norm, Laplacian sharpness.
11 unit tests cover the error-prone math with no model needed.
- onnx_face_analyzer.rs (feature `faces-onnx`): wires the geometry to ONNX
Runtime via `ort` (load-dynamic, so libonnxruntime is dlopen'd at runtime
and the crate builds without it). Inference runs on spawn_blocking; each
session is serialized behind a Mutex. Loads via `ort::init_from` (fallible)
not ORT's lazy loader, which would panic under `panic = "abort"`.
- config: FacesConfig + OXICLOUD_FACES_{ORT_DYLIB,DETECTOR_MODEL,
EMBEDDER_MODEL,DET_SIZE,DET_THRESHOLD,NMS_THRESHOLD,INTRA_THREADS}.
- di: build_face_analyzer() loads the real analyzer when the feature is
compiled in and runtime+models are configured; any missing piece or load
failure degrades to the no-op analyzer (logged) so startup never fails.
- ort/ndarray added as optional deps; example.env documents the setup.
Models and the ONNX Runtime dylib are operator-provided at runtime and are
never committed. Cannot be exercised in CI (no models/dylib); the geometry
is unit-tested and the ONNX seam is isolated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Updates the implementation status from "not started" to reflect the
landed face-recognition stack and notes the deviations from the original
plan (BYTEA embeddings + in-Rust cosine instead of pgvector; a single
FaceAnalyzerPort instead of split detector/embedder ports; union-find
connected-components clustering). Flags the remaining piece: the real
ONNX analyzer (ort + operator-supplied models), which adds a crate and
can't be exercised in this environment.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Adds the client side of Phase 2 (People). A new "People" tab in the
Photos sub-navigation lists identity clusters from GET /api/people and
drills into a person's photos using the existing photos lightbox.
- people.js: peopleView with list/drill-in/rename, reusing .photos-grid
tiles and photosLightbox; rename via Modal.prompt + PATCH /api/people/{id}
- people.css: person grid, circular avatars, single-person header,
loading/empty states — all design tokens, no raw colors
- places.js: People tab wired into the Moments|Places sub-nav, revealed
only when GET /api/people is reachable (capability probe); _switchTab
now toggles three views
- index.html: load people.css + people.js
- en.json: photos.tab_people + people.* labels (other locales fall back
to English via i18n)
The tab stays hidden unless OXICLOUD_ENABLE_FACES is on (the API 404s
otherwise), so this is inert by default.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Phase 2 increment 6 — the People HTTP API (mounted only when
OXICLOUD_ENABLE_FACES is on; every handler is caller-scoped):
- GET /api/people list identity clusters
- GET /api/people/{id}/photos a person's photo file ids
- PATCH /api/people/{id} name / rename a person
- POST /api/people/{id}/hide hide / unhide
- POST /api/people/merge merge two clusters
- POST /api/people/recluster re-run clustering
- DELETE /api/people/data erase all face data (opt-out)
- GET /api/people/faces/{file_id} face boxes for lightbox tagging
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Phase 2 increment 5:
- FaceIndexingService: a FileLifecycleHook that, on image upload, detects +
embeds faces in a background task and stores them. Dedup-aware (clones an
identical blob's faces instead of re-running inference), reindexes on
overwrite, and relies on the DB cascade for deletes. Completely inert when
no model is ready.
- DI: registers the hook in the FileLifecycleService chain and exposes
PeopleService in AppState — both gated on OXICLOUD_ENABLE_FACES, both
using the default no-op analyzer until the operator wires a real ONNX
model.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Phase 2 increment 4:
- recluster(user): union-find connected-components over the user's face
embeddings (cosine threshold); groups of >= min_faces become persons, and
an existing person's name is preserved across reclusters. O(n^2) — fine
for moderate libraries; an ANN index is the documented scale-up.
- caller_id-scoped use cases for the HTTP layer: list_people (non-empty
clusters, cover thumbnail, most-photographed first), person_photos,
faces_for_file, rename, hide, merge, and delete_all (right to erasure).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Phase 2 increment 3: FacePgRepository implements FaceRepository.
Embeddings stored/read as BYTEA (512 little-endian f32), bbox as REAL[].
Every query is user-scoped. Covers face CRUD, person CRUD (create / rename
/ cover / hide), files-for-person, and a transactional delete-all-for-user
(right to erasure).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Phase 2 increment 2:
- domain: Face, Person, BoundingBox, DetectedFace (512-d embeddings).
- ports: FaceAnalyzerPort (detect + embed from raw bytes; decodes internally
so the application layer stays image/ML-crate agnostic) and FaceRepository
(user-scoped face/person persistence).
- DTOs: PersonDto, FaceBoxDto.
- NoopFaceAnalyzer — reports is_ready()==false and returns no faces, so the
whole People pipeline compiles and runs inert until the operator wires a
real ONNX-backed analyzer.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
First slice of Phase 2 (People / faces):
- migration: `faces` schema with `faces.persons` and `faces.faces`.
Embeddings are stored as BYTEA (512 x f32) — no pgvector extension
dependency; similarity is computed in-app (pgvector/VectorChord is the
documented scale-up). Cascade deletes (by user and by source file)
satisfy the right to erasure.
- OXICLOUD_ENABLE_FACES feature flag, OFF by default (biometric data,
opt-in per deployment).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Records the as-built choices: static-file basemap via ServeDir Range
(not the pmtiles crate), HTML thumbnail markers (not deck.gl), default-on
flag, and the pending browser smoke-test + operator basemap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Adds a Places tab to the Photos section: a self-hosted MapLibre GL map of
the user's geotagged photos.
- places.js: lazy-loads vendored MapLibre + pmtiles.js on first open; draws
the server-aggregated clusters (GET /api/photos/geo) as HTML thumbnail
markers (no glyphs/sprites, no client-side clustering); pan/zoom refetches
for the new viewport; click a cluster to zoom in, or a single photo to
open it in the lightbox.
- Optional vector basemap read directly from /basemaps/basemap.pmtiles over
HTTP Range (label-light Protomaps style, light/dark aware); falls back to
a themed background when no basemap is present. ODbL attribution shown.
- "Moments | Places" sub-nav wired into the Photos section.
- enable_places now defaults on, since the route + UI are ready.
No new backend serving code: tower-http ServeDir already serves static/
with Range, so the .pmtiles basemap is just a static file.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Vendored for the self-hosted Places map (Approach A: client-side PMTiles
decoding, basemap served as a static file via the existing Range-capable
ServeDir):
- maplibre-gl 5.24.0 (BSD-3) + its CSS
- pmtiles 4.4.1
Also adds static/basemaps/ with a .gitignore (*.pmtiles is operator-provided,
never committed) and a README explaining how to drop in a basemap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Phase 1 server side, gated on OXICLOUD_ENABLE_PLACES (off by default):
- migration: partial index on storage.file_metadata(longitude, latitude).
- FileBlobReadRepository::list_geo_clusters — plain-SQL grid aggregation
(no PostGIS) scoped to the caller's own non-trashed photos, returning a
centroid, count and a representative file id per non-empty cell.
- PlacesService (caller_id-scoped; user-scoped data needs no authz check,
mirroring RecentService) with a zoom→cell-size mapping.
- GET /api/photos/geo?bbox=w,s,e,n&zoom=N returning GeoCluster[]. The route
is mounted only when the Places service is present, and is registered in
the OpenAPI path list.
The map frontend (PMTiles serving + MapLibre module) is deferred pending
the basemap-sourcing and vendoring decision.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Adds an implementation-status section: Phase 0 done (with commit refs),
0.2 implemented via a flattened PhotoDto rather than widening FileDto, the
two Phase 0 loose ends (embedded map pin, drag-marquee, HEIC, sub-nav),
and Phase 1 (Places) marked backend-in-progress.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
- Photos zoom via wheel, double-click and two-finger pinch (1–5x) with
drag-to-pan; a touch swipe navigates prev/next when not zoomed.
- A new info button toggles a panel showing date, size, dimensions, camera
and GPS coordinates (resolving the old geoloc TODO), pulled from
/api/files/{id}/metadata.
- Zoom/pan state resets on every item change and on close.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Adds a Grid/Justified toggle to the photos toolbar. Justified mode packs
tiles into Flickr-style rows scaled to the container width using each
photo's real aspect ratio (from the new /api/photos width/height, falling
back to 1:1 when missing). It composes with the virtualized renderer:
per-group materialization and the off-screen spacer height estimates are
both layout-aware. The choice persists in localStorage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
list_media_files now LEFT JOINs storage.file_metadata and returns each
photo's pixel dimensions next to the sort date. The endpoint wraps FileDto
in a flattened PhotoDto carrying width/height, so the gallery can lay tiles
out at their true aspect ratio (justified layout) without a second per-file
metadata round-trip and without layout shift. FileItem gains optional
width/height. No change to FileDto or its other construction sites.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
- Tiles are focusable (tabindex / role=button / aria-label) with a
:focus-visible ring; Enter opens the lightbox (or toggles in selection
mode), Space toggles selection.
- Shift-click extends the selection from the last anchor across the
timeline; the range is tracked in the selection Set so it spans
dematerialized (off-screen) groups, with visible tiles updated at once.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
- Add Modal.confirmDialog() (Promise<boolean>, built on openPanel so it
inherits the overlay, animation, focus-trap and Escape handling) and use
it to replace native confirm() in the photos batch-delete and lightbox
single-delete flows.
- The lightbox now reflects the real favorite state when an item opens
(previously the star always started empty) and toggles favorites through
the favorites module so its cache stays in sync.
- Add photos.delete_* i18n keys (English; other locales fall back to en).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Consolidates the research into a concrete, file-by-file plan following the
repo's hexagonal conventions (AuthZ in services with caller_id, audit logs,
feature flags, sqlx migrations, vanilla JS/CSS):
- Phase 0: gallery polish (virtualization done; dimensions in /api/photos,
justified layout, lightbox zoom/pan + map pin, a11y, sub-nav).
- Phase 1: Places — MapLibre + self-hosted Protomaps PMTiles served from
Axum (pmtiles crate) + SQL grid aggregation (no PostGIS).
- Phase 2: People — ort (ONNX) detect+embed, pgvector storage, incremental
threshold clustering; runtime-downloaded models; opt-in/GDPR.
Includes vendoring/dependency table and the open decisions to resolve.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
The photos timeline rendered every tile into the DOM and grew it
unbounded on infinite scroll, degrading on large libraries. Each
date-group is now a <section> whose grid is materialized (tiles
inserted) only while near the viewport and dematerialized (emptied,
height frozen as a spacer) once it scrolls away, driven by an
IntersectionObserver rooted on the scroll container. DOM nodes stay
bounded by a few screens regardless of library size.
- Grouping (day/month/year), infinite scroll, multi-select, video
thumbnails and fade-in are all preserved.
- Selection state and the video-thumbnail cache survive the
materialize/dematerialize cycle.
- Falls back to full rendering when IntersectionObserver is unavailable.
- Spacer heights are estimated from grid geometry and re-estimated on
resize.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JW6ghFMDtnRYuYNzZhb47M
Standard CalDAV/CardDAV clients (Thunderbird, DAVx5, Apple
Calendar/Contacts) failed to connect, mounted collections read-only, or
could not discover address books, even though curl worked. Three
protocol-compliance gaps caused this:
1. Missing Basic-auth challenge on /caldav and /carddav.
The 401 returned for these surfaces carried no `WWW-Authenticate`
header (only /webdav did). Spec-compliant clients never send
credentials preemptively the way `curl -u` does — they wait for the
challenge — so Thunderbird never authenticated and failed with
"discovery failed" / 401. Extend the challenge to all DAV surfaces via
shared `is_dav_path` / `dav_basic_auth_challenge` helpers.
2. Calendars always advertised read-only.
The `current-user-privilege-set` write gate compared `owner_id`
against the literal string "current_user_id", which never matched a
real UUID, so `<D:write/>` was never emitted and clients mounted every
calendar read-only. Thread the caller's id through the CalDAV adapter
and grant write when the caller owns the calendar.
3. CardDAV discovery was incomplete.
There was no `/.well-known/carddav` route and the root PROPFIND
exposed neither `current-user-principal` nor `addressbook-home-set`,
so clients could not locate address books. Add the well-known redirect
and root/principal discovery responses mirroring the CalDAV adapter.
Adds unit tests for the auth challenge predicate, the calendar
owner/non-owner privilege split, and the CardDAV root/principal discovery
responses.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016cVV9nRQjP6G6a8zbNUWMw
D-Prep (migration 20260801000002_drop_access_grants) drops
storage.access_grants and replaces it with storage.role_grants — one
row per role assignment instead of N rows per permission bundle. The
load seeder still spoke the old per-permission shape and failed every
nightly with `relation "storage.access_grants" does not exist`.
Both seeder call sites already grant the read bundle (single
permission), which maps cleanly to the `viewer` role; switching them
to the new schema is a one-row INSERT with the role name. The
conflict key drops `permission` since uniqueness is now per
(subject, resource).
Adds a `ref` input to workflow_dispatch so the nightly can be triggered
from main and run the scenarios against a feature branch that does not
yet carry the workflow file (e.g. feat/drive). Empty input falls back
to `github.ref`, so cron and bare dispatch are unchanged.
The regression-issue title and body now show the tested ref instead of
`github.sha`, which always resolves to the workflow's ref (main) under
workflow_dispatch and would otherwise mislead.
prepare migration of permission to roles
this simplify drastically database (permission are now simply defined in code)
and will permit reuse of the same ReBAC engine to define owners of drives
mapping:
```
Role::Viewer => &[Permission::Read],
Role::Commenter => &[Permission::Read, Permission::Comment],
Role::Contributor => &[Permission::Read, Permission::Create],
Role::Editor => &[
Permission::Read,
Permission::Comment,
Permission::Create,
Permission::Update,
],
Role::Owner => &[
Permission::Read,
Permission::Comment,
Permission::Create,
Permission::Update,
Permission::Share,
Permission::Delete,
Permission::Manage,
],
```
The KNOWN BUG pin for `GET /api/s/{folder-token}/file/{file_id}` was
stale — the route now returns 200 + body for files inside the share's
subtree and 404 for anything outside it. Replaces the sidestep
comment with two positive assertions:
- in-share: 200 + Content-Disposition references the file name
- out-of-share (caller-owned file in a different folder): 404,
matching "no such file" so the response can't be used to
enumerate foreign file ids
Coverage now exercises the actual recipient-side download path that
NC desktop and web clients use; the file-share variant (item_type=file)
moves down to test 8b.
Adds a teardown DELETE for the outsider hello.txt so the next test in
the runner (permissions.hurl) can re-upload its own hello.txt into
admin's home folder without hitting the live-name unique index.
handle_head was declaring `Content-Length: file.size` while writing
`Body::empty()` — on a keep-alive connection the client waits forever
for N bytes that never come. Hyper now derives Content-Length: 0 from
the actual body, which is honest about what's on the wire.
RFC 7231 §4.3.2 suggests HEAD return the same headers as GET, but
lying about Content-Length is worse than omitting it: NC and Sabre
clients use PROPFIND for size anyway, and curl -I (and any client
applying HEAD semantics) gets the same ETag/MIME/Last-Modified it
needs. Caught by the F6b test which uses `curl -X HEAD` to read the
current ETag before a conditional PUT.
Also adds `nc_status_propfind_depth0` to lib/dav_helpers.sh so the
F11/F11b assertions ("did the intermediate parent get auto-created?")
can compile.
When the client sends `MOVE /trashbin/{id}` with a `Destination` header,
handle_restore now resolves the destination path and returns 412
Precondition Failed if a live file or folder already sits there —
matching Sabre/DAV and the NC desktop client's expectation. There is
no `Overwrite: T` workflow for trash restore in either reference
implementation (silently replacing a live file with an undeleted one
is a footgun), so the refusal is unconditional.
The destination header is extracted at the dispatch site as an owned
String so the future stays Send-compatible (`&Request<Body>` is not
Sync because the body trait object is Send-only).
`extract_nc_subpath_from_dest` is promoted to `pub` so trashbin_handler
can share the same URL parser as handle_move.
Claude