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 wash stream is small; the entry camera photographs every car in exactly the view the
classifier is trained on. The booth can now queue entry vehicle reads as pure training
material — crop + the camera's class, no order, no operator, no category.
- Core announces every vehicle read (deviceEvents.emitVehicleRead from snapshot.ts); the
Car Wash module listens, samples entry reads in-process (sampleEntry: exactly one in N)
and queues them (enqueueEntry). CARWASH_REVIEW_ENTRY_SAMPLE=N; 1 = every entry (storage
and bandwidth are not the limit — user); 0/unset = off. Forwarded by compose.
- Packages carry kind: "wash" | "entry". Collector: kind column, entry meta validated
without the operator fields, review screen shows an entry sample as such, export has a
kind column, operator agreement computed from wash items only. Setup line shows
"1 in N entries sampled"; status carries entrySample.
- komodo: park-2's four review lines enabled (collector URL by Netbird DNS name, booth-2,
the shared per-booth secret, every entry sampled) — the collector is up on the overlay.
- Tests on both sides. Wiki: vision-review-outbox (entry stream + the internet-feed
assessment), log.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
The far end of the Car Wash review outbox (wiki/concepts/vision-review-outbox.md): a small
Fastify + SQLite service in the monorepo (shares the payload contract and the class
vocabulary via @parking/shared), delivered to art-docker-station by its own stack so
nothing booth-side lands there and nothing of it on a booth.
- POST /ingest: bearer token per booth (constant-time), X-Booth-Id must match, multipart
meta + JPEG (magic checked, 2 MB cap), meta validated against the contract, idempotent on
the item id; crop stored at crops/<booth>/<item>.jpg on the volume + one items row.
- /review + /api/*: the reviewer's screen served by the process (Basic auth, one login):
one pending crop at a time, operator's pick and camera's pick beside it, one button/key
per vocabulary class + unusable + skip; stats per booth and per hashed operator
(agree / disagree / unusable — disagree = the reviewer's class is outside the operator's
category).
- GET /export/labels.csv: reviewed usable rows for training; formula-leading cells are
neutralised (booth-supplied names). Crops stay on the volume for the trainer on the host.
- Booth payload now carries operatorCategory.classes so the comparison needs no site setup.
- Delivery: apps/collector/Dockerfile (monorepo context), docker-compose.collector.yml
(bind to the overlay IP; commented `trainer` profile seam for the GPU), a third build
step in build-images.yml, a `wash-collector` stack in komodo/resources.toml with one
secret per booth referenced from both the collector's token list and the booth's own
stack (park-2 lines templated, commented, DNS name for the URL).
- Tests: app.test.ts (ingest ok/dup/refusals, review + stats + export, config). Image
built and smoke-tested locally (health, ingest, duplicate, auth, verdict, export).
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
Groundwork for the Car Wash pilot (wiki/decisions/venue-modules.md, build-order
steps 1 + 3). No Car Wash code yet; validation is the first module behind the
seam, unchanged in behaviour.
- @parking/shared: MODULE_IDS, ModuleManifest, MODULES (parking required;
validation dependsOn parking), parseEntitledModules / resolveModuleActivation
/ effectiveModules as pure functions.
- DB: site_config.modules_json (migration 0026, hand-written + journal;
additive, nullable = everything entitled).
- Server: modules.ts (entitledModules from MODULES_ENTITLED env, activated
from site_config, effective set, requireModule preHandler → 403
module_disabled); modules/index.ts registers folder-based modules by
iterating the registry (modules/validation); site-config GET exposes
modules/modulesEntitled/modulesActivated, PUT takes the full desired set,
enforces entitlement + dependency rules (400 with reason) and signs one
config_change per module that actually flips; /api/auth/me carries the
effective set; validation routes guarded requireModule → requirePermission.
- Web: lib/modules.ts + modules/{index,validation}; router.tsx spreads
WEB_MODULES into nav + route tree (validate route no longer named there);
Setup → Site "Modules" panel (required shown disabled, dependencies as
hints, server refusal shown verbatim); validation sections + programs fetch
gated on the module; App invalidates the router whenever the session
changes (route-context consumers only re-read on navigation — the nav was
stale after a flip, and after every other setUser too).
- Lavazh validation station retired (STATIONS = ["bar"]; rows untouched).
- Deploy: MODULES_ENTITLED=parking,validation explicit in both booth stacks;
documented in .env.example.
- Tests: modules.test.ts (7); suite 329/329; web build clean; Playwright
round-trip on /setup/site verified live.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
Promotes park-buzi + park-2 to the runtime-configurable desktop backend
address (ConnectScreen) and the desktop CSRF fix. build-images.yml
confirmed green for this sha before bumping.
Promotes park-buzi + park-2 to the just-merged desktop-app fixes (login,
mixed-content routing, WS origin) and the WS_ALLOWED_ORIGINS fix — none of
this was on stage before. Wait for build-images.yml to confirm the image
actually exists before syncing/deploying in Komodo.
Login worked after the mixed-content fix, but the live feed 403'd silently:
tauri-plugin-websocket's connect() runs on Tauri's Rust side, not inside the
webview page, so it never auto-attaches Origin the way a browser WebSocket
would — routes/ws.ts's anti-CSWSH check rejects a missing Origin before
auth. platform-ws.ts now sets Origin: tauri://localhost explicitly.
Also fixes a second, independent gap the above alone wouldn't have caught:
komodo/resources.toml's booth Stacks had WS_ALLOWED_ORIGINS= empty in
production despite .env.example documenting it as required for desktop.
Needs a Komodo sync + redeploy to reach a live booth.
New real booth, same tier as park-buzi: tracks stage, pinned stage-<sha>,
own park_2_* secret refs. Periphery had the known /etc/komodo root_directory
permission bug on --user install (wiki gotcha #9); fixed and confirmed OK
in Core before adding this Stack.
That host is becoming a Traefik edge, and parking's prod override brings its own
Caddy on `network_mode: host` holding :80 — the two cannot share the port. The
lab tier moves to a dedicated bench PC rather than contorting either side.
This also names what has been holding :80 on that box: the edge stack deployed
there on 2026-09-01 failed with "address already in use" and the owner was
recorded as unidentified. It was almost certainly this Caddy.
REMOVING THIS BLOCK DOES NOT STOP ANYTHING. The containers keep running and keep
the port. Destroy park-lab from Komodo Core BEFORE syncing this removal:
DestroyStack names a stack and Core resolves where from its own synced copy of
the definitions, so a sync that drops the block first takes the teardown handle
with it. If that has already happened, remove the containers by hand on the host
— there is no compose project context on a Komodo-managed box.
Three Core secrets are now unreferenced: art_docker_station_jwt_secret,
art_docker_station_event_signing_key, art_docker_station_backup_key. Lab keys
with no real ledger behind them, so they are safe to delete once the stack is
gone.
Claude-Session: https://claude.ai/code/session_01SARfPK19vLBstMWBxubezN
park-lab-old referenced a server removed from Komodo, breaking the resource
sync. Also documents two Periphery incidents from this session: a Core-UI
rename doesn't touch the agent's own connect_as, and upgrading Periphery is
a config-preserving re-run of the installer.
Printer investigation (park-buzi): cover-open on the USB thermal
printer wedges its status offline/faulty, surviving a full reboot,
recoverable only via `docker restart server`. Traced sendRawUsb/
probeUsb end-to-end — no persistent handle in the app layer, so the
leading theory is the container's /dev/usb directory bind-mount
retaining a stale view across the printer's physical re-enumeration.
Not yet confirmed on hardware; documented with repro/confirmation
commands and ranked candidate fixes.
Also registers a new lab bench box, "art-docker-station", as a Komodo
Stack (dev tier, same shape as park-lab, its own isolated secret refs).
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
Adds on top of the un-deployed cd3b534 pin: camera clock sync via ISAPI
(heals the 1970 power-cut reset at the offline→ready edge + daily
backstop). Everything since the deployed f9887c2 rides along: USB
printer close-cancel fix (hardware-verified at the lab), USB device
dropdown (lp1 shows by model name), printers addable without a
controller. No migrations.
Post-deploy validation: pull a camera's power, let it come back, then
docker logs | grep "clock synced" — expect a warn with a huge drift.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
Second [[stack]] block: server park-test (the lab box's Periphery
connect_as), compose files from the dev branch, MOVING TAG=dev (a lab
may float; real booths pin), its own park_lab_* secret refs (per-box
blast radius — never shared with a real booth). park-buzi is untouched
on stage + pinned.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
Promotion of the entry presence-gate bypass (+ signed relay test): merge is on
stage, CI built :stage-6505a4a (pull verified). Bump the park-buzi Stack pin to
the new immutable sha. Deploy is the manual Komodo step: refresh ResourceSync →
Execute → Deploy; watch for [migrate] done (carries 0020 — two site_config
bypass columns, additive, applied at boot against the /data volume).
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
The first :stage image is built and in the registry (stage-39c778f). Pin it in the
IaC so git matches Core's Stack env and a ResourceSync won't revert TAG to the
placeholder. Bump this on each promotion.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
A committed stage-<sha> can never match the commit that introduces it (the pin
commit changes HEAD), so a hardcoded sha here is always stale by one. Make it an
explicit placeholder (stage-REPLACE_WITH_BUILT_SHA); the real immutable sha is set
when you deploy from Komodo Core after CI builds :stage-<sha>. No moving tag on a
booth still holds.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
Model the staging-vs-production split that fleet-deployment-komodo flagged as open.
Three tiers: dev (working, no booth) -> stage (staging booth park-buzi, real-world
test) -> main (production, manual + pinned).
- build-images.yml: trigger on [dev, stage, main]. The tag computation is already
branch-derived, so :stage / :stage-<sha> build with no other change.
- komodo/resources.toml: park-buzi now branch=stage + TAG=stage-<sha> (pinned;
no webhook even on staging). BACKUP_KEY already wired as a per-booth secret.
- komodo/README.md: a Promotion (dev->stage->main) section; per-booth secret list
now includes backup_key; hard-rule #1 generalised to pinned <branch>-<sha>.
- wiki: fleet-deployment-komodo open-item resolved + a Promotion-tiers table;
deploy-trigger choice generalised; container-deployment tag list gains :stage.
Promotion is a merge: when dev is ready, merge dev->stage, CI builds the image,
bump TAG=stage-<sha> in resources.toml, deploy from Core. stage is branched from
dev HEAD so the first real-world test carries the full current app. Per-booth
secrets must pre-exist in Core; migrations run at boot so a promotion auto-migrates
the staging ledger (where a bad migration is caught before production).
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
Retention (keep-last / keep-daily-days) is operational policy the on-site admin
should tune, not a server env var requiring a redeploy -- same reasoning that moved
the target directory to the UI.
- Migration 0017: site_config.backup_keep_last + backup_keep_daily_days (nullable;
null = code default 7 / 30 per field).
- BackupService reads retention fresh each run; status() exposes keepLast +
keepDailyDays. DEFAULT_BACKUP_RETENTION is now a pure code default (env reads gone).
- PUT /api/backup/config accepts keepLast / keepDailyDays (non-negative int, or null
to reset to default; 400 on negative).
- UI: two retention fields on the Backup config card; one Save covers target +
retention. i18n sq + en.
BACKUP_KEY wired into Komodo:
- komodo/resources.toml: BACKUP_KEY=[[park_buzi_backup_key]] (per-booth secret,
alongside JWT / signing keys).
- komodo/.env.komodo.example: documents it as the ONLY backup env var -- escrow it
offsite alongside EVENT_SIGNING_KEY (recovery needs both); target + retention are
admin-chosen in the UI / DB, not env. Server .env.example trimmed to just BACKUP_KEY.
Also carries the small in-progress setup-intro i18n copy trim.
Tests: 218 server tests green, incl. retention persist / reset-to-default / reject-
negative and the updated status shape. Migration applies cleanly (needed a
statement-breakpoint between the two ALTERs). Wiki backup-recovery updated.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
Camera snapshots were stored RAW — the camera's full-res JPEG straight into
the BLOB, no resize/recompress. Measured on the dev DB: 300 snapshots = 81.7 MB
= ~72% of the 114 MB SQLite file (the big ones 2688×1520 / ~600 KB, Hikvision
main stream). They dominated the appliance's single backed-up DB file.
Re-encode on capture (snapshot.ts):
- Downscale each frame to SNAPSHOT_MAX_EDGE (1280px long edge) + recompress at
SNAPSHOT_JPEG_QUALITY (80) via sharp (libvips, Apache-2.0) before storage —
~6-10× smaller (verified 2688×1520 → 1280×724, ~8×), plate still readable,
clean image/jpeg (drops the camera's charset cruft). STORAGE-ONLY: recognition
keeps the ORIGINAL full-res bytes (downscaling hurts OCR). Fail-soft — a
re-encode error stores the original, never drops the snapshot or blocks the
(already-open) path. sharp lives in apps/server (owns the capture path), where
bcrypt already establishes the native-dep pattern.
Disk-pressure retention (snapshot-retention.ts) — a SAFETY VALVE, not the daily
mechanism (the re-encode does that). Daily check reads the DB filesystem used%
(statfs on db.$client.name); no-op unless ≥ SNAPSHOT_DISK_HIGH_PCT (70). Over the
mark: delete the OLDEST until an estimated SNAPSHOT_DISK_FREE_TARGET_PCT (10%) of
disk is freed — never below SNAPSHOT_MIN_KEEP (500) — then VACUUM once to return
space to the OS. A DELETE only frees SQLite pages (disk doesn't drop until VACUUM),
so the loop is driven by estimated freed bytes (SUM(length(bytes))), not a live
disk re-read; the prune owns the DB-locking VACUUM, run daily off-peak. diskUsage
is injectable for tests. None of this touches the signed ledger — snapshots are
unsigned/advisory, referenced only by id.
Tests: encodeForStorage (downscale / clean-type / no-enlarge / fail-soft) +
pruneSnapshots (no-op below mark / delete-oldest-to-target + VACUUM / MIN_KEEP
floor / skip-VACUUM-when-empty). All four snapshot env knobs documented in the
komodo env reference. Full workspace build/lint/test green; the prune smoke-verified
on a scratch DB copy (file shrank after VACUUM).
Existing ~81.7 MB of raw snapshots are unchanged (a one-off re-encode backfill is
a separate optional follow-up). Updated entry-exit-points + technology-stack wiki.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
A loop started by a far/early car would (a) give up before the REAL car settled at
the barrier, and (b) swallow the real car's pushes (the #polling guard dropped them).
So a confident-but-wrong far-car plate could win, or the intended car get debounced
out after the loop ended — wrong car acted on, right car blocked.
Fix: a push that JOINS a running loop now EXTENDS the deadline (lastPush +
ANPR_POLL_WINDOW_MS) instead of being dropped, capped at start + ANPR_POLL_MAX_MS
(30s) so a continuously-busy lane can't slide forever. Each tick still pulls a FRESH
frame, so the loop tracks whoever is at the barrier NOW, not the car that started it.
Per-camera sliding deadline in #pollDeadline (cleared with #polling in finally).
+1 test (push mid-poll keeps the loop alive past the initial deadline); 171 server
tests green. New knob ANPR_POLL_MAX_MS documented in the komodo env reference + the
two concurrency guards written up in lane-presence-and-anpr-entry.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
The Stack example only listed the required set; expand it to the FULL reference so
an operator can see (and tweak) every env without digging through code. Grouped:
- IMAGE SELECTION (REGISTRY, TAG)
- REQUIRED (JWT_SECRET, EVENT_SIGNING_KEY, COOKIE_SECURE — no safe default)
- COMMONLY SET (VISION_ENABLED, WS_ALLOWED_ORIGINS)
- SET BY COMPOSE — don't put in the Stack (VISION_URL, DATABASE_URL, VISION_RECOGNIZER)
- OPTIONAL TUNABLES with code defaults: ports, logging/retention, device+printer poll
intervals, lane/capture TTLs, and the ANPR knobs incl. this session's new
ANPR_POLL_MS=1000 / ANPR_POLL_WINDOW_MS=8000 (raise the window for a slow barrier)
- VISION CONTAINER env (the Python service's own VISION_* vars)
All defaults pulled from the code (process.env.X ?? default). Documentation only.
Adopt Komodo Periphery (over the NetBird mesh) as the booth fleet control plane,
superseding SSH-and-booth.sh. The booth runs the SAME compose files; Komodo Core
drives them remotely. booth.sh is demoted to a break-glass local fallback.
- komodo/resources.toml mirrors the working park-buzi Stack (built by hand in the
Core UI, then exported to TOML — field names match the running v2.2). Stack-only:
servers are created by the agent onboarding OUTBOUND (one-time onboarding key →
Periphery self-registers, auto-rotating keys, booth opens no inbound port), so
there is no [[server]] block. Per-booth secrets via [[...]] refs to Core's store.
- komodo/README.md + .env.komodo.example document the flow and the hard rules
(no webhook; onboarding/outbound/mesh-only; per-booth unique secrets; never
down -v the ledger volume).
- wiki/decisions/fleet-deployment-komodo.md records the decision + threat-model
analysis (Periphery is a root agent → mesh-bound; EVENT_SIGNING_KEY-in-Core is a
fraud-root blast radius until ATECC608 signs; Core is now Tier-0; GPL-3.0 is fine
as external ops tooling). container-deployment reframed (booth.sh = fallback);
index + log updated.
Verified end-to-end against a real booth (park-buzi): onboarded OK, Stack deployed,
all containers green, admin seeded.
Claude-Session: https://claude.ai/code/session_01Xcm6ikLgGoCxxHrxtjkk5V
julian