apps/trainer (parking-trainer): inspect / train / evaluate / publish. Reads the wash
collector's SQLite + crops read-only off its volume; time split (validation = newest
slice); thin classes dropped; damped class weights; `features` mode (frozen ImageNet
backbone, on-disk feature cache, seconds to retrain) and `finetune` mode (light
augmentation). CPU-only torch from PyTorch's wheel index. ONNX export checked against
the torch model; NO model file below the validation floor (exit 3, report still written);
exit 2 = not enough labels. `evaluate` scores a shipped model on labels reviewed after
training + the unlabelled pile; `publish` PUTs a version folder to a Gitea generic package.
Light core deps; the `train` extra is heavy — CI syncs without it, torch tests skip.
apps/vision: BodyTypeClassifier (bodytype.onnx + sidecar = the preprocessing contract:
crop margin, input size, RGB 0-255, normalisation inside the graph) and
RefinedVehicleDetector over YOLOX — refines only `car` or a class the classifier trained
on, min-confidence, `detector_class` on the result; path set but no file = phase B off
without an error; a broken file is a health detail. models/bodytype.version (tracked,
empty) pins the published version the Dockerfile fetches at build (BuildKit secret;
a pin that cannot be fetched fails the build). Verified: a trainer model gives identical
probabilities inside the vision service; both images built and smoke-tested.
Delivery: parking-trainer image in build-images.yml, the `trainer` compose profile on the
collector stack (CPU, read-only data, TRAINER_OUT), commented TRAINER_OUT/PUBLISH_TOKEN in
the wash-collector stack, .dockerignore for both Python contexts, trainer deps synced in CI.
Wiki: bodytype-classifier-training rewritten as built (+ one fleet model not per site,
secrets/access, where the crops live), opencv-anpr-service §Phase B, vision-review-outbox,
vision-service-packaging, fleet-deployment-komodo, index, log.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
The operator's category choice is a hypothesis, not truth (user, 2026-09-06): each wash
order with a vehicle read queues a package for a trusted reviewer over the private overlay
(Netbird); the verdict becomes the phase-B training label and the per-operator error rate.
wiki/concepts/vision-review-outbox.md.
- Boxes: the vision service returns the vehicle bbox; snapshot.ts stores the vehicle and
plate boxes on the read as FRACTIONS of the analysed frame (the stored snapshot is a
downscaled copy); vehicleForIdentity() returns them.
- carwash_review_outbox (migration 0031) + review-outbox.ts: crop = detector box + 8 %
margin, ≤ 640 px, plate blurred in place from the plate box; payload carries a
pseudonymous booth id and a keyed operator hash — no site name, no plate, no OSD, no
bystanders; multipart POST with a per-booth bearer; 2xx → sent (image dropped);
400/404/413/415/422 → abandoned; anything else → backoff 1 min·2^n capped 6 h; voided
orders and items older than 14 days abandoned unsent. Nothing queued while unconfigured.
- Enqueue is fire-and-forget off the intake path in createOrder; the loop runs every
CARWASH_REVIEW_INTERVAL_SEC (60) and stops on close.
- GET /api/carwash/review/status (site:read) + a "Remote review" line in Setup → Car wash.
- Env CARWASH_REVIEW_URL / _TOKEN / _BOOTH_ID (all three or off) documented in
.env.example and forwarded by compose.
- Tests: review-outbox.test.ts (crop + blur on a synthetic frame, config/pseudonyms,
queue/drain/backoff/abandon, through the app). Wiki: new concept page, index,
venue-modules As built, log. The collector is not built.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
The app plumbing for venue-modules.md §"Vehicle category from vision"; the model is the
open half (no bundled recognizer emits body_type yet, so the desk shows nothing until
phase A lands in the vision service).
- Shared: VEHICLE_CLASSES vocabulary, VehicleRead, CARWASH_VISION_THRESHOLD_DEFAULT,
reason code carwash.categoryDowngrade; settings/order/lookup views carry the read.
- Vision contract: /analyze vehicle.body_type + confidence (service schema); the Node
client normalises to the vocabulary and drops the rest.
- Record: snapshot.ts stores the read in the plate's device_events row (or its own when
the plate was unreadable); vehicleForIdentity() resolves it like the plate.
- Car wash: carwash_categories.vision_classes (site mapping "car, sedan → Vetura"),
carwash_config.vision_threshold (signed config_change when it moves), four vision
columns on orders — migration 0030. Lookup returns vision + suggestedCategoryId.
- Desk pre-selects the mapped category and shows the read + snapshot thumbnail; Setup
offers class chips per category and the threshold. Operator decides.
- Flag: a read at/above the threshold whose mapped category prices HIGHER than the chosen
one signs one `anomaly` (both categories/prices, operator, snapshot) and stores its id on
the order. Equal/upgrade/unsure/unmapped → nothing. Recorded only, never blocks, no
reason prompt (user, 2026-09-06).
Tests in carwash.test.ts; wiki venue-modules (As built), opencv-anpr-service, log.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
Add a dev CLI (uv run python -m vision_service.cli <image>) that runs a recognizer on
an image file and prints the parsed plate(s) + confidence + region — fast feedback with
no HTTP. Also a package.json `recognize` script and a vision-recognize entry point.
Verified fast-alpr for real: installed the `alpr` extra, downloaded the YOLOv9 + CCT
ONNX weights (~11MB, cached offline under ~/.cache), and ran recognition on the
project's test image → "5AU5341" at 1.000 confidence, region "Czech Republic", ~40ms
on CPU, via both the CLI and POST /analyze.
Fixes result parsing against the actual fast-alpr API: ocr.confidence is a LIST of
per-character confidences (not a scalar) — reduced to one plate confidence via the MIN
(a plate is only as trustworthy as its weakest character); also surface ocr.region.
Extracted the per-result mapping into a pure plate_from_alpr_result + _reduce_confidence
and unit-tested them (no model weights needed). 7 tests pass; ruff + mypy strict clean;
full turbo build/lint/test green.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
Skeleton of the host-side vision service per the packaging decision: a Python/FastAPI
app at apps/vision/, uv-managed, wired into the Turbo graph via a thin package.json
shim (dev/lint/test/build → uv/uvicorn/ruff/pytest). A per-package turbo.json sets
build outputs [] so the no-op build is warning-free.
Endpoints: GET /health (readiness + model version) and POST /analyze (raw
octet-stream body, so Node POSTs Snapshot.bytes directly; empty→400, oversize→413,
recognizer-not-ready→503). The recognizer is a Protocol with a StubRecognizer (no
models, boots/tests offline — the dev/CI default) and a FastAlprRecognizer (the real
MIT YOLOv9+CCT/ONNX stack, lazily imported; missing models ⇒ ready=False, not a crash)
— the device-adapter pattern applied to the model. fast-alpr + onnxruntime are an
optional `alpr` extra, so `uv sync` needs no model download.
Verified: turbo run lint|test|build includes @parking/vision and stays green; uv run
mypy strict-clean; uvicorn boots and serves /health + /analyze live; pnpm workspace
6→7. Not built yet: the Node VisionClient adapter, a Dockerfile + model fetch, and
Job 2 (vehicle verification). Updates the packaging decision (As-scaffolded) + log.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
julian