576df58776dc0d106b1535cf13b7d9acf1fa0a90
26
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 tajniak81andClaude Opus 5
|
576df58776 |
Go the way the owner's phone already goes
Control had two transports and neither fitted the ordinary customer. OCPP waits for the charger to dial in, which needs a public endpoint it can reach, a certificate, and a firmware willing to talk to our CSMS. Modbus TCP dials the charger, which needs the server on the charger's own network. Between them they cover a charger we host and a charger we stand next to; the common case is a charger behind someone else's router, and that had nothing. It was never unreachable, though. The charger holds a connection open to Anker's own broker — it is how the mobile app drives it from anywhere, and it is the mqttStatus register the Modbus snapshot has been reporting all along. So a third control mode joins that broker as the account: get_user_mqtt_info issues a client certificate, mTLS to aiot-mqtt-eu.anker.com:8883, and commands go out on the same topics the app publishes on. Nothing on the customer's side has to be forwarded, addressed or certificated. What travels is not an API call. The payload is a JSON envelope around a base64 binary frame the device itself speaks — marker, little-endian length, message type, name/length/type/value fields, XOR checksum — so mqttframe.go is a codec rather than a client, written from the message maps in anker-solix-api and anchored on the one frame that project documents byte for byte. A frame whose fields do not tile exactly up to the checksum is refused rather than half-read: these arrive over a link we do not control, and a truncated frame must not read as a charger reporting zeros. Two of the charger's habits shape the rest. It publishes nothing unless asked, so a status read arms a telemetry trigger and waits for the next frame, and a poll inside that window answers from what has since arrived. And a broker connection costs a fetched certificate and a TLS handshake while the plugin manager builds a throwaway instance per request — so the connection lives on the account's shared session beside the auth token, for exactly the reason the token lives there, and closes itself after five idle minutes. The transport also sees two signals no other one does: the boost flag, and the plug and start countdowns. The package doc has said since the first commit that they are never set and the derived mode must do without them. Here they are set, so a charger that has been told to start and is counting down a delay says so rather than sitting in "preparing", and "skip the delay" is offered only while there is a delay to skip. The clients generalise instead of growing a second layout. Both snapshots name the same quantities the same way, so what was Modbus-only in the readouts is now whichever transport read the charger — ModbusStatus becomes ChargerStatus on the phone, mb becomes dev on the web. What each transport can be *told* still differs, and the buttons branch on that: reset and clear-limit stay with OCPP, the timeout and phase registers with Modbus, skip-delay with the cloud. A command a transport has no equivalent for is refused by name, saying which one has it. The cost is worth saying plainly. This leans on Anker's cloud being up and on an unofficial protocol the app may change under us, where Modbus leans on nothing but the LAN. And it is checked against the reference implementation's own worked example rather than against hardware — there is no charger on this end to point it at. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
62cb691f98 |
Everything the register map carries, sorted the way it gets asked about
The Modbus snapshot reported about half of what one poll already brings back. The rest was read into the block and thrown away: line-to-line voltages, reactive and apparent power per phase, the PWM flag, the control-pilot voltage, and the identity block's product number, rated power and current range. All of it now decodes — no extra requests, the registers were in hand already. Added alongside it: the control block, read back over FC03. It answers a question the live registers cannot, which is what the charger is *set* to as opposed to what it is doing — a boost that was asked for reads there while the live block still reports none running. Best effort, so a charger that refuses it still reports its state. Two registers the spec leaves blank are decoded on the hardware's evidence. The control-pilot voltage reads 11873 while the CP signal register reports state A, which that enum names as 12 V, so the register is millivolts. The identity block's current range is in amps, whatever its unit column says about watts and kVA. The charging card lays this out in sections rather than a wall of forty numbers: per-phase measurements as the matrix they are, then live state, then settings, then the device itself, with alarms surfacing only when a word is non-zero. Strings in en/da/pl. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
aaa89dfe10 |
Relays that run at 33 degrees, not 331
The two relay temperatures came back as 331 and 319 from a charger sitting idle with nothing plugged in. The spec's gain column says 1 for both, so we reported them as 331 °C and 319 °C — a reading that would have meant a fire rather than a wallbox at room temperature. The gain is 10. The same table hands the maximum current setting a unit of watts and the timeout a unit of amps, so its unit and gain columns are not load-bearing here; what settles the alignment is the LED brightness two registers earlier, which reads exactly 100 at gain 1, and the fact that the neighbouring registers all decode as tabulated. Read back from the charger afterwards: 33.1 °C and 31.9 °C. The field becomes a float, as the voltages and currents beside it already are. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
cf4fd14b56 |
The table the charger actually keeps its measurements in
Modbus mode never returned a reading: every status poll came back as "the charger did not answer", though the charger was answering all along. It was refusing the question. The A5191 splits its map across two tables where the spec's single 2xxxx column suggests one — 20000-20100 are input registers and reject FC03 with an illegal-address exception at every address in the range, while 21000-21005 really are holding registers and read back over FC03. We inferred one space from the spec's layout and asked for all of it with FC03. The client learns FC04, sharing a body with FC03 since the two differ only in which table the server consults, and the plugin's two measurement reads move to it. Writes stay on FC06, where the controls already live. Confirmed against an A5191 on firmware 1.0.6.1: identity, live block and the control registers all decode as the spec tabulates them. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
f7472bada3 |
Reach the charger where it is, instead of waiting for it to call
OCPP asks the charger to dial us: a public endpoint, a TLS certificate, and a route in through the customer's router. Our own handler then demanded two more things the V1 does not offer — TLS on a charger that connects over ws://, and Basic auth credentials the Anker app has no field for — so every connection was turned away before the upgrade. Anker publishes a Modbus TCP register map for this charger, and it inverts the problem: we dial the charger, on its own network, with no inbound reachability to arrange. That works for a charger behind a router that OCPP cannot reach at all. internal/modbus is the protocol, hand-rolled against the spec like the MQTT and WebSocket clients beside it. The plugin's modbus.go is the V1's map: the same 0-8 status enum the cloud already reports, per-phase measurements, and the writable registers behind start, stop, current limit, boost and phase mode. A new "modbus" control mode routes the existing control endpoints down it, so the REST surface, the rate limit, the confirmation step and the audit trail are the ones already there. The commands the register map has no equivalent for say so by name rather than failing as unknown, and a current below the charger's 6 A floor is refused because it pauses the charge rather than slowing it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
190ae923a6 |
The Anker token outlives the request that fetched it
The manager builds a throwaway plugin instance for every per-user call —
HealthCheckWith, InvokeWith, InvokeBatchWith each construct, Init, probe and
Shutdown. The auth token lived on that instance, so it died with the HTTP request
that fetched it: opening the Anker panel signed in once for the health probe and
again for the charger list, and a page that also asked for OCPP info signed in a
third time. Every refresh, a fresh login.
Anker throttles passport/login per IP per minute and answers code 26161 ("Failed
to request.") once tripped, so this is the shape of the failure the panel has been
reporting; the cloud has also historically kept one token per account, so each of
those logins could evict the one the mobile app was holding.
Tokens and the login backoff now live in a package-level session keyed by the
account signing in, so every instance configured for that account shares one
login. Re-configuring the same credentials keeps the token; a different account,
or the same account on the other regional server, gets its own session. Sessions
unused for a fortnight are pruned, so an edited password does not leave its entry
behind for the life of the process.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
tajniak81andClaude Opus 5
|
c1aee0fac1 |
One refused sign-in, not five: the chargers poll no longer locks the account
A chargers poll asks four cloud views. Each called apiRequest, each found no token, and each ran its own login — so a login Anker refuses was offered four times in one poll, and the next poll spent the fifth. Five is what disables the account for ten minutes, which is how "code 26161: Failed to request." turned into "your account has been disabled" on the very next attempt. The plugin now remembers a refused login instead of repeating it: the failure is cached and replayed to every caller until a backoff window passes — a minute at first, doubling to fifteen, or the full ten minutes when Anker says it has already locked the account (code 10019). New credentials clear it, so a fixed password is tried at once. chargerInventory signs in once up front. A login the cloud refuses is not four views failing, so it is reported as itself rather than as three warnings with the lockout notice buried in the last one. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
a809980d8b |
Anker health: count the chargers the panel lists, not the ones one endpoint admits to
The probe still asked get_user_bind_and_not_in_station_evchargers and read its userBindEvChargersCount, so it reported "0 EV charger(s) bound to account" for an account whose two chargers the panel was listing directly underneath — the same blind spot the capability was just moved off, left behind in the health check. It now takes the same inventory the chargers capability returns and counts that. Authenticated with nothing on the account is degraded rather than ok, following Greencell's rule: the half we address answers, and the empty half is the account or the country that picks the regional server, so the message says so instead of reporting a healthy connection to nothing. A count reached with some view missing says how many views stayed silent, because the number is then a floor rather than a total. The web panel colours degraded amber, as it already did for Greencell. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
e138fad3f4 |
Anker: every charger on the account, not just the ones outside a station
get_user_bind_and_not_in_station_evchargers is the only list the connector ever asked for, and its name says exactly what it withholds. A charger that belongs to a system is not in it. Its userBindEvChargersCount, though, counts every charger bound to the account — so an owner with two chargers in a system got "authenticated; 2 EV charger(s) bound to account" from the health probe and an empty list from the capability that is supposed to show them. A working login that finds nothing. So the capability now asks every view the cloud has and merges them by serial. The standalone list still answers for chargers standing on their own; get_site_list walks the systems and reads each one through get_scen_info, falling back to get_system_running_info where that is silent — the power-service / HES split charger-state already knows; and get_relate_and_bind_devices contributes model, firmware and the Wi-Fi flag, and discovers anything in the A519 family that the first two missed. Whichever way a charger was registered, one of the three has it. The merge is first-writer-wins per field rather than last view overwriting: the standalone record knows the name, the site record knows the live state, and neither should blank what the other established. A view that fails is a warning on the document instead of an error on the call, because one dead endpoint should not cost the chargers the other two found. Only losing all three is a failure. When nothing comes back at all the response says so in its own words and names the remaining suspect — country picks the regional server, and the wrong one authenticates happily and shows an empty account. The other half of "not showing any chargers" was that neither client ever showed a list. The serial was a text box, and the number is printed on a charger hanging on a wall. Both apps now list what the account holds — name, serial, model, site, state, an offline badge — and hand the serial to the OCPP control card instead of asking anyone to go and read it. Where control is off the list still stands on its own, as the answer to the first question an owner has after entering credentials. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
a7cab50e06 |
Apprise: a gateway to hand a message to, not a hundred protocols to carry
Apprise is a Python library that speaks 100+ notification services behind one URL grammar — mailto://, tgram://, ntfy://, discord://. None of that is portable to a server that takes no dependencies, and none of it needs to be: caronc/apprise-api wraps the library in HTTP and is meant to run as a container beside us. So the connector carries no notification protocols of its own. It posts a body to an endpoint the operator runs and lets Apprise fan it out, which is also why adding a service later costs nothing here. Targets are addressed one of two ways and configKey is the switch. Stateful means the URLs live on the Apprise server under a key, narrowed by a tag expression, and recipients are then edited there — no credential for any downstream service is ever held in DriverVault. Stateless means the URLs travel with the request, from a secret config field, which is simpler for one destination and worse for ten. A call that names its own key or urls takes that destination alone rather than merging with the configured one: honouring a caller's URLs while still falling back to the configured key would deliver the message somewhere nobody asked for. baseUrl is Required, which no other connector's address is. Toyota, Anker and Greencell leave everything blank at the global layer because the superadmin → org → user cascade exists to fill it in, and a blank there means "let the user choose". There is no cascade behind this one — a notification gateway is infrastructure the operator runs, not an account a driver owns — so nothing further down can supply the address, and a blank is simply a plugin that cannot work. Better to fail at enable than at the first notification nobody sees. Three limits are choices rather than gaps. /add and /del are not implemented: the Apprise config belongs to the operator, we post to it, and a connector that can delete a notification config has a wider blast radius than one that can only send through it. privacy=1 is forced on /json/urls rather than offered as a parameter, so a target listing reads mailto://user:****@host and downstream tokens stay on the Apprise side of the wire. Attachments are remote URLs the Apprise server fetches; multipart upload is the API's own path for files and not ours. Health follows the rule Greencell set. A reachable server whose config holds nothing to notify is degraded, not down: the half we address works and the missing half is the operator's config. Two cases earn their own line — a config key set against a server running with stateful mode disabled can never resolve, and /status answers 417 rather than 500 when Apprise finds a problem with itself, so that is a parsed answer and not a transport failure. A proxy that strips our Accept header gets the same codes back as plain text, which is read rather than called unreadable; an HTML error page from something that is not Apprise is not, and a test pins the difference. Notifications needed a category of their own, and that is the one change outside the plugin: the constant, the tab order in PluginsCard.vue, and the label in all three panel languages. The cost is now written down in the plugins README beside the Descriptor example, since the previous five categories predate anyone having to add a sixth. The plugin's tests run against an apprise-api stand-in built from that project's views.py — both notify paths, the override rules, 204-as-empty against 424-as-failure, and every health branch. builtin_test.go is the other half: the blank-import list in builtin.go is a silent failure mode, since a connector left out of it compiles, passes its own tests, and never appears in the panel. What is not covered is a live instance; there is no Docker on this machine, so the wire contract comes from reading upstream's source rather than from running it, and a smoke test against a real deployment is still worth doing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
e648634ce1 |
The plugin list, grouped by what a plugin actually is
Category has been in the plugin contract since it was written — apis-external, drives-external, drives-local — and every builtin declared the same one, so it grouped nothing. Two of the three talk to a wallbox and one talks to a car manufacturer, and those are different questions an operator arrives with: the Toyota card is where a driver's account gets linked, the Anker and Greencell cards are where a charger's broker and credentials live. So vehicles and chargers join the constants and the three builtins say which they are. The panel groups on that field rather than on a list of names, which is what keeps an external plugin from needing panel code. Tab order mirrors the constants; a category with nothing in it gets no tab, and a single group hides the bar entirely, so an install with one connector looks exactly as it did. A category the panel does not recognise — or an empty one — falls to the external-APIs tab rather than vanishing, because a plugin nobody can see is a plugin nobody can disable. The selected tab falls back to the first group when its own goes away, which is what removing the last external plugin does. Registration still asks only for name, base URL and provider, so a plugin registered at runtime lands under Other APIs until its manifest names a category. That path already works and is the honest default: the panel is guessing about a service it has never spoken to, and the service can say. The header lockup is the other half. It was a copy of the Web App's mark rather than the same mark, and copies drift — a 32px icon against 28, a 24px wordmark against 21.6, "Driver" at text-strong instead of white, "Vault" a step lighter than brand-400. The Web App's Logo.vue moves in verbatim, props included. The one thing it cannot inherit is which variant to render: the Web App's rail is always dark, while this panel flips with its own theme toggle, so on-dark is bound to the theme and the hand-rolled bar fills that existed to survive that flip are gone. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
340a81b0d6 |
Greencell: the charger on your own broker, not a cloud it never had
The HabuDen has no cloud API to connect to. It is commissioned over Bluetooth in
the Greencell GC app, pointed at an MQTT broker the owner runs, and from then on
publishes there — so the connector is an MQTT client rather than an HTTP one,
and nothing in it reaches Greencell. The wire contract is Home Assistant's own
greencell component and the greencell_client 1.0.3 library beneath it, which is
the only published description of the topics: a BROADCAST on /greencell/broadcast
draws device announcements, and /greencell/evse/{sn}/ carries current in
milliamps, voltage, power under "momentary", the EVSE state, and the access level
chosen in the app.
That meant an MQTT client, and the server takes no dependencies, so internal/mqtt
is hand-rolled the way internal/ocpp's RFC 6455 layer is. It is scoped to what
this connector needs and says so: QoS 0 for everything we send, clean session,
no reconnect — a connection lives for one plugin call, which is exactly how the
manager builds and tears down an instance. Inbound PUBLISH is accepted at QoS 0,
1 and 2 with the acknowledgements each requires, because the QoS of a delivery is
the broker's choice and not ours; an unacknowledged QoS 1 is redelivered forever.
Read-only, and the reason is worth writing down rather than rediscovering. A
device in EXECUTE mode accepts START, STOP, SET_CURRENT and QUERY — but the topic
those go to appears in no source: not Greencell's integration page, not
greencell_client, and Home Assistant ships sensor-only for that same reason.
Publishing to a guessed topic would be a control feature whose failure mode is a
driver believing they stopped a charge. So the access level is reported, and
commandTopic is the seam: an operator who has watched their own broker and found
theirs sets it, and a state read then sends QUERY — the one command a READ-mode
device also honours — instead of waiting out the charger's publish cadence. The
day the topic is public, control is a payload away from the same field.
What the cascade resolves here is a broker, not an account, so host, port, TLS and
credentials resolve together from the highest layer that names a host: an
organization's address paired with a user's password would address a broker with
credentials never meant for it. The serial, the QUERY topic and the listen window
each describe the charger rather than the endpoint, so each resolves on its own.
Two reading rules the tests pin. A phase the device did not report stays nil
rather than zero, because zero amps on a charger is a real measurement — a JSON
null decoding to 0.0 was a live bug until a test caught it — and a partial read
returns with received/complete flags instead of failing, since a device that
publishes some topics on a slower cadence is still worth reading. And a reachable
broker with no charger on it is degraded, not down: the half we configure works
and the missing half is the device. The plugin's end-to-end tests run against an
in-process broker written to the raw wire format, so a bug in the client cannot
hide behind a matching bug in the fixture.
The apps get the third connector card. The panel needed nothing — it renders a
plugin's ConfigFields itself — but the per-user panes are still hand-written per
integration, which is now three near-copies and the argument for the generic
version already noted in the plugins README. The web form splits the broker from
the charger because the server resolves them differently. The phone card is a
declarative config against the shared widget, which gained a number field type, a
degraded state that reads amber rather than red, and a fix for a locked field
that was covering its own displayed value with dots. Twenty keys in three
languages across both apps; Greencell, HabuDen and the literal QUERY join the
proper nouns that stay in English.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
tajniak81andClaude Opus 5
|
5e6b8b4b1c |
Anker Solix: the charger's mode, and the modes it can be moved into
The connector was written against anker-solix-api v3.7.0 and upstream is at 3.8.1 now. The reassuring half of the check first: nothing we depend on moved. The passport/login ECDH exchange, the headers, and every endpoint path this plugin calls are identical across v3.7.0...v3.8.1 — the only apitypes movement touching an EV charger was get_device_rfid_cards being reordered within its own dict. The 400 new lines in charger.py are the A2345 USB charger, which shares a filename with our device and nothing else. What did land for the V1 is two entries in the release notes, and both are MQTT: 3.8.0 gave standalone chargers the usage-mode entity they were missing, 3.8.1 added a switch that reads those modes as a plain on/off so EVCC and its like have a binary to hold. We control chargers over OCPP, not MQTT, so the command path is not ours to port. The reading of state underneath it is, and that half does come over the cloud. So charger-state. The status code arrives under two different names depending on which system family a site belongs to — operating_state inside a scene's charging_pile_list, evChargerStatus inside HES system running info — and upstream's poller quietly renames both to ev_charger_status on ingest, which is the tell that they are the same number. We ask both and merge, because a site answering only one of them is the normal case rather than a fault; the call fails only when neither view is there. chargerMode and chargerModeOptions then follow ev_charger_mode_state and ev_charger_mode_options as written, including the rule that a stopped charger is startable only from standby, and the binary is the same one 3.8.1 chose: everything that is not stop_charge counts as on. The gap worth naming is that the boost flag and the plug and start countdowns reach upstream over MQTT and never over the cloud, so three of the six modes cannot occur here. That is not a bug to be found later — chargerMode takes them as parameters and the callers pass their zero values, so the day an MQTT source exists the derivation is already correct and only its inputs change. The package doc says so in the scope list beside the other limits. Five endpoints upstream has had all along and we never exposed come with it, all EV-charger-scoped: the site scene, energy_analysis under device_type ev_charger, a charger's RFID cards, Anker's own OCPP endpoint list, and one vehicle's details. charger-status takes the featuretype it was hardcoding at 1, since upstream's exporter asks for both 1 and 2 and there was never a reason for us to see only half. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
22a22ec43a |
Toyota: the status route the car answers, not the one it retired
Toyota put the /v1/global/remote read routes behind AWS SigV4 in mid-2026. A bearer token is no longer a credential there, so the doors-and-windows card has been asking a gateway that answers 403 — the one section of the provider tab that could only ever have been in error. The MyToyota app reads that state from /v1/vehicle/status now and pytoyoda followed it in 5.2.0; so does the connector. The electric route did not move, and the comment above the endpoint block says which of the two namespaces each one lives in, because the obvious tidy — sweep the rest onto /v1/vehicle/* — would break the ones that still work. The same migration gave the climate reads a home worth porting: /v1/vehicle/ climate-status is what the cabin is doing, climate-settings the preset it was told to do it at. Both are GETs with a vin, both are new cards on the tab, and their headings are in all three languages on both apps. Nothing about the tab's plumbing changed to hold them — a section is an id, an action, and whatever JSON comes back, which is the point of that shape. Left where they are: the POST wake calls. Upstream refreshes a stale reading by waking the modem, and this connector is documented as read-only, so climate and status show what the car last reported rather than what it would say if asked twice. The cost is a reading that can be hours old, and it is the honest one to pay for a connector that promises not to touch the vehicle. Two tests keep the migration from being undone by hand: one fails if any advertised capability points back at a retired route, the other if a capability is advertised without being wired into Invoke, which is the way the next endpoint would go missing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
ee4ac441be |
Plugins: drop the plugins.json migration, and the volume it needed
The project has no public installs, so there is nothing to migrate from. MigrateLegacyFile, the file-backed Store it read through, PLUGINS_FILE and the legacy path threaded through the Server all go. What is left is one store, PocketBase, and a plugins package that touches no filesystem at all. That was the last thing keeping api_data alive, so the volume goes too. All four compose files now declare exactly one volume, pb_data, and the standalone API Server compose declares none - it talks to an external PocketBase and has nothing of its own to keep. Backing up the stack is backing up one path again. Both images get simpler for it. The API Server image loses VOLUME /data and the su-exec entrypoint that existed only to fix a mounted volume's ownership, so it goes back to a plain USER app; its working directory is now /app and holds nothing. The AIO image loses its second volume and chowns only /pb/pb_data. One consequence worth stating plainly, because it is a small regression rather than a no-op. The panel's Settings -> PocketBase and Settings -> Web App screens write .env in the working directory, which is now ephemeral. In the multi-container stack that changes nothing: compose sets all five of those keys as container environment, and loadDotEnv only applies a key that is not already set, so the file could never win a restart there anyway. In the AIO image it did win for POCKETBASE_ADMIN_EMAIL/_PASSWORD, which are not in that container's environment - so a service account fixed from the panel now lasts only until the container is recreated. Both READMEs say so. Moving those two screens into the app_settings singleton would close it properly; the PocketBase URL and credentials cannot follow, since they are how the database is reached in the first place. go build, go vet and go test ./... pass; the compose files parse and each resolves to a single pb_data volume. Not verified: no Docker CLI here, so neither image was built. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
660af5736a |
Plugins: create the settings collection instead of waiting for it forever
|
||
|
tajniak81andClaude Opus 5
|
9bd5c523c4 |
Plugins: the global layer moves into the database, beside the other two
The integration cascade stored its top layer differently from the two below it: org (L2) and user (L3) plugin config lived in PocketBase, in a pluginSettings field, while the global (L1) layer sat in a plugins.json next to the binary. That split was accretion rather than design - the file was the whole store in the v1 MVP, and the per-tenant layers were later built on PocketBase and layered on top of it instead of replacing it. It also cost something real. plugins.json was a second state store with different durability from pb_data: its own volume, its own ownership, its own backup. Losing pb_data is unmissable; losing api_data was silent, which is how "every plugin comes back disabled after a redeploy" happened. L1 now lives in the app_settings collection - one record keyed "global", holding its settings in a pluginSettings field, the same mechanism and the same field name the layers below use. The documents still differ in shape, because only L1 carries enable state and the registration of external plugins, but the storage is no longer a special case. The Manager grows a Store seam (PocketBase in production, file for the import, memory for tests) and, more importantly, a loaded gate. Settings in a database mean the store can be unreachable at boot - a cold stack, or a service account still to be set from the panel. That must not read as "no plugins configured", or the first save would write emptiness over real settings. So until a read succeeds the Manager stays unloaded, every mutation is refused, /api/admin/plugins* answers 503, and a background retry backs off to two minutes. The same gate covers a document that will not parse: it is never replaced by one built from an empty map, which is a stronger guarantee than the .corrupt backup it replaces. Writing to a store also revealed a hole in the previous fix. Classifying a save failure as errPersist was left to each Store, and a store that returned a plain error would fall through to the "saved, but the plugin failed to start" branch and be reported as a 200 - the same silent-success bug through a different door. The Manager now classifies, whatever the Store returns; a test pins it. Upgrades are automatic: on the first boot that finds no settings in the database, an existing plugins.json is imported and renamed to plugins.json.migrated. The import is refused if the store is merely unreachable, or if the file does not parse, so a stale or broken file can never overwrite live settings. /data is still needed - the panel rewrites .env there when it retargets PocketBase - but plugin settings no longer depend on it. 21 tests in internal/plugins cover both stores, including the production path against a fake PocketBase: create-then-update of the singleton, round-trip across a restart, an outage that leaves settings intact, a missing collection reading as not-ready rather than empty, and the import running exactly once. go build, go vet and go test ./... pass. Schema changes are mirrored into scripts/setup-pocketbase.mjs as that file requires. Not verified: no Docker CLI here, so no image was built and the bootstrap of app_settings against a real PocketBase is untested outside the fake. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
c173ca3653 |
Plugins: a save that fails should say so, not vanish on redeploy
Reported symptom: every plugin comes back disabled after redeploying the image, having been enabled before it. The persistence design was already right - each compose file mounts api_data:/data and points PLUGINS_FILE at /data/plugins.json - so the fault was that a failed write to that file was invisible. Three defects, each confirmed with a test before being fixed: A failed write was reported as success. Upsert set rec.Enabled before it persisted, and the handler folded the resulting error into the same 200-with-warning used for "saved, but the connector failed to start". The panel reloaded, read the in-memory record and showed the plugin enabled; only a restart revealed that nothing had reached the disk. A save that fails now rolls back in memory and returns 500, so the panel row shows the error instead of "Saved". A corrupt state file silently wiped the rest. Load returned an error, main.go logged it and carried on with an empty record set, so the next toggle overwrote plugins.json and took every other plugin's config with it. An unreadable file is now moved aside to plugins.json.corrupt, and persistLocked writes through a temp file + rename so an interrupted write cannot produce that corrupt file in the first place. A state file holding "null" panicked the server with "assignment to entry in nil map" on the next save, and a null entry nil-dereferenced in Load. Both now decode to "nothing configured". Two changes make the next such failure loud rather than silent. StartPlugins probes writability at boot and warns that plugin changes will not survive a restart. And the API Server image gains the root entrypoint the AIO image already had - chown /data, then drop to app via su-exec - because a host bind mount (API_DATA=/srv/...) or a volume created before /data existed arrives root-owned, and the unprivileged process cannot write to it. Not addressed here: a deployment that never reuses the named volume (docker compose down -v, a renamed compose project, an anonymous volume from a bare docker run) loses the file whatever the code does. The new boot warning tells the two apart - writable but empty means the volume is the problem, not permissions. go build, go vet and go test ./... all pass. The Dockerfile change is reviewed but not built: there is no Docker CLI on this machine, so the su-exec privilege drop follows standard Alpine practice rather than an observed run. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 5
|
358ee68f94 |
Cars: create a car from a manufacturer service, with a per-car data tab
A car can now be imported straight from the account its owner already has
with the manufacturer, and every reading that service exposes shows up on
the car's own tab. MyToyota is the first provider.
API Server — internal/api/vehicleproviders.go adds a generic layer over a
plugin that can enumerate vehicles and read data about them. Adding the
next manufacturer is one vehicleSource adapter plus a line in
vehicleSources(): no new endpoints, no Web App changes.
GET /api/vehicle-providers providers + connect state
GET /api/vehicle-providers/{p}/vehicles the caller's vehicles
POST /api/vehicle-providers/{p}/import create a car from one
GET /api/cars/{id}/provider live snapshot for the tab
POST /api/cars/{id}/provider link / unlink a car
POST /api/cars/{id}/provider/sync re-apply provider data
Two properties shape it. Credentials are always the caller's own, resolved
through the same global -> org -> user cascade as the integration settings,
so a shared car shows provider data only when that vehicle is on the
viewer's account — the owner's credentials are never borrowed. And upstream
shapes are not modelled: these are unofficial APIs, so the layer searches
payloads by key name for the readings worth promoting (odometer, fuel,
battery, range) and flattens the rest to dotted key/value pairs alongside
the raw JSON. A renamed field costs one blank value, not a broken page.
The Toyota gate and its wording now live in toyotaSource, so the older
/api/integrations/toyota/vehicles endpoint and the new ones cannot drift.
Manager.InvokeBatchWith shares one transient plugin instance across a batch
of actions. The tab pulls seven capabilities, and InvokeWith builds a fresh
instance per call — which for a connector that authenticates lazily means a
fresh OAuth login per call. Batching logs in once.
cars gains provider + provider_vehicle_id (schema.go and
setup-pocketbase.mjs both). carPayload deliberately omits them, so an
ordinary car edit can neither reassign the car nor break its link;
carProviderPayload writes the link on its own.
Web App — Dashboard grows an "import from service" button beside "add car",
shown only once an account is connected, opening CarImportModal: pick the
vehicle, choose what to pull (identity / fuel type / dates / odometer, all
on by default), import. ProviderPanel becomes the car's first tab, ahead of
Information, labelled with the service: headline readings, the vehicle
record, one card per capability with its raw response, and an offer to take
the provider's odometer when it is ahead of the stored one. On an unlinked
car the tab instead offers to link it, VIN-matched. Info stays the default
selection — landing on the provider tab would fire a login on every car
page view. Full en/pl/da translations.
Tests cover the payload walking, Toyota normalization, import-selection
defaults, and — through the real handler chain against a stand-in
PocketBase — that every route is registered and that a closed gate is soft
on a listing (200 + a reason the UI can show) but hard on a write (4xx, so
a caller cannot read the reply as a created car).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
tajniak81andClaude Opus 4.8
|
a0eb5e4e9d |
Docs: refresh every README against the current code
Verified each documented command, path, port and env var against what the code actually does, and corrected the drift. Phone App. Was still titled Car Control. The navigation description was also stale: the app moved to a RootShell bottom nav (Garage, Charging, Settings, and Users for admins), so the Settings gear and admin action the dashboard bullet described no longer exist. Adds the Charging screen, noting its public tab is placeholder data and only the Home tab's OCPP control is real, and rebuilds the lib/ tree, which had lost i18n.dart, theme.dart, widgets/ and three screens. Web App. Node 18+ was wrong. The installed Vite is 8.1.2, whose engines field is ^20.19.0 || >=22.12.0 - Node 18 is EOL and cannot build this. API Server. The config table gained OCPP_REQUIRE_TLS, OCPP_PUBLIC_URL, PB_BOOTSTRAP and DRIVERVAULT_SUPERADMIN_*, plus a note that PLUGINS_FILE and the panel-written .env resolve against the working directory (a volume, in Docker). Plugins. Per-tenant credentials sat under "not yet implemented", but /api/integrations/* has done exactly that for both built-ins for a while. Narrowed the roadmap item to the genuinely missing generic version. New Docker/README.md and Docker AIO/README.md: the root README's component table linked those directories as documentation but neither had any. The root README now points at them. All 8 markdown files pass a relative-link check. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 4.8
|
1e76c2b7f9 |
Refresh docs and fix Docker builds for current layout
READMEs: correct the auth model (PocketBase token relay, not JWT/sessions), document the full feature set (technical checks, fuel, maintenance, documents, reminders, attachments, integrations, OCPP charging control), the shipping built-in connectors (toyota, anker-solix), and the current endpoint surface. Docker: build against the current repo layout — Go 1.26, cmd/server entry point, Web App source under web/. Add the missing Web App Dockerfile (Go BFF) and .dockerignore, drop the obsolete AUTH_SECRET, modernise CORS var naming, and standardise on drivervault-* naming. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 4.8
|
a1519f6e89 |
Add OCPP control for the Anker Solix EV charger (Own/Proxy CSMS)
The Anker Solix connector was read-only (cloud monitoring only). Add an
OCPP 1.6J control path with a per-user, cascading control mode:
- off monitoring only (default, unchanged behavior)
- own DriverVault is the charger's Central System (full control)
- proxy DriverVault relays to Anker's cloud and injects commands
New internal/ocpp subsystem (stdlib-only, hand-rolled RFC 6455): a CSMS
with session management, inbound dispatch, and typed control commands
(RemoteStart/Stop, SetChargingProfile current limit, ChangeAvailability,
Reset, UnlockConnector, TriggerMessage, Get/ChangeConfiguration). Own- and
proxy-mode paths are verified end-to-end against a simulated charge point.
The charger connects to /ocpp/{serial}, authenticated with OCPP Basic auth
(serial + a per-charger control token) resolved to the owning user via an
in-memory token index. Control REST endpoints mirror the monitoring ones and
reuse the same cascade gate plus a live-session check. controlMode is a new
cascade field (global -> org -> user) advertised as a select on the plugin.
Frontend: control-mode select + provisioning card in Settings, and a real
Start/Stop/limit/reset control panel in Charging, gated on the active mode.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
||
|
tajniak81andClaude Opus 4.8
|
0793b5ec8e |
Add a built-in Anker Solix V1 Smart EV Charger plugin
A Go re-implementation of the auth and read-only data flow from thomluther/anker-solix-api, scoped to the V1 Smart EV Charger (A5191) and adapted to DriverVault's plugin contract. Login is a custom ECDH (P-256) + AES-256-CBC password exchange against passport/login, yielding a ~7-day auth token plus gtoken = md5(user_id) for subsequent requests; a fresh login covers expiry and 401/403. The country code routes to the EU or global Anker server. Exposes read-only capabilities (chargers, charger-status, charge-stats, charge-orders, ocpp-info, devices, sites, vehicles) with a health check that reports the bound-charger count. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 4.8
|
5e435c5f77 |
Add per-user Toyota integration with a settings cascade
Let each user run the Toyota Connected plugin under their own MyToyota
credentials and enable/disable it for themselves in the Web App, while a
superadmin (and, in an organization, an org admin) can impose settings
from above. Resolution is a cascade — top wins, and a lower level only
fills fields the levels above left blank:
- org user: API Server (superadmin) -> org admin -> user
- org-less user: API Server (superadmin) -> user
The MyToyota email + password resolve together as a pair from the highest
layer that supplies an email; brand resolves on its own; enablement is
strictly per-user, gated by the global master switch and the org gate.
API Server:
- plugins.Manager gains RawConfig / HealthCheckWith / InvokeWith so the
cascade can read global config and probe/invoke under a per-caller
resolved config.
- internal/api/integrations.go resolves the cascade and serves
GET/PUT /api/integrations/toyota, POST .../health, GET .../vehicles.
Secrets and inherited usernames are masked before leaving the server.
- The toyota builtin's credentials are no longer required at the global
layer, so the master switch can be enabled without global credentials.
- setup-pocketbase.mjs adds a pluginSettings JSON field to the users and
organizations collections (the user and org layers of the cascade).
Web App:
- api.js gains getToyota/saveToyota/testToyota.
- Settings grows an Integrations section: an enable toggle, credential
fields with locked / "inherited from" states, a brand select, an
org-scope switch for admins, and a live test-connection button.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
||
|
tajniak81andClaude Opus 4.8
|
5a729abd71 |
Add a built-in Toyota Connected Europe plugin
Port pytoyoda's authentication and read-only data flow to a Go builtin plugin behind the existing plugin contract. Implements the three-legged ForgeRock/OAuth2 login (authenticate callback loop, authorize, token exchange), silent refresh with full re-auth fallback, and the full Toyota gateway header set with backoff on 429/5xx. Exposes read-only capabilities: vehicles, telemetry, location, health, status, electric, notifications, and service-history. Config takes a MyToyota email/password and a Toyota/Lexus brand select. Europe-only and read-only, matching pytoyoda's limitations. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
tajniak81andClaude Opus 4.8
|
ae6ed4ac1e |
Rebuild API Server on the PilotVault structure
Mirror PilotVault's API Server layout and add the superadmin console,
plugin system, runtime PocketBase settings, and user/organization
management. The car domain (cars, service records, parts, sharing) is
carried over unchanged apart from the auth switch.
Layout: main.go -> cmd/server/main.go; module carcontrol/api ->
drivervault/apiserver. internal/api is split by concern (auth, users,
orgs, settings, plugins, status, health, respond).
Auth: replace the server-minted HS256 JWT and the sessions collection
with a PocketBase token proxy. /api/auth/login relays PocketBase's
{token, record}, and every protected request re-resolves that token
against PocketBase, so a role change or deletion takes effect at once
instead of waiting out a token. AUTH_SECRET is obsolete and internal/auth
is gone. Per-device session listing/revocation goes with it: PocketBase
tokens are stateless. Changing a password rotates the user's token key,
which invalidates every token already issued.
Roles: add superadmin alongside user/admin, plus an organizations
collection and users.organization. Admins are scoped to their own
organization; superadmins span all of them. Guards prevent changing your
own role, deleting your own account, an admin touching a superadmin, and
deleting an organization that still has members.
Plugins: new internal/plugins package with one contract over two kinds --
builtin (compiled in) and external (any HTTP service, registered at
runtime with no rebuild). State persists to plugins.json; secrets are
masked on read and preserved when saved back at the mask.
PocketBase settings: /api/admin/pb-config applies a new connection at
runtime and persists it to .env. It deliberately does not require a
working service account, so a wrong or unreachable connection can still
be fixed from the panel.
Panel: rebuilt as the superadmin console -- login gate, status, users,
organizations, PocketBase, plugins, and the endpoint reference.
Clients: update the Web App and Phone App for the PocketBase token shape,
the move of user management to /api/users ({users}/{user} envelopes, with
password resets folded into PATCH), and the removal of sessions. Both now
mirror the server's real guards rather than the old last-admin rule, and
parse PocketBase's field-level error shape.
Config: modern POCKETBASE_*/API_ADDR names with legacy PB_*/PORT
fallbacks, so existing .env files keep working. Also fixes /api/status
probing the Web App on 8090 instead of DriverVault's 5173.
Run scripts/setup-pocketbase.mjs to add the organizations collection and
grow users.role; every client must log in once more.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|