The local gate list only covered `scripts/export_openapi.py` + a clean
`openapi/` diff, but `frontend/src/api/schema.d.ts` is generated from that JSON
and CI re-runs `npm run codegen` with `git diff --exit-code`. d07a083 changed a
route docstring, refreshed openapi.json, and skipped codegen — local green,
remote red on a one-line comment diff. Spell out both steps and note that route
docstrings feed the OpenAPI description too.
Also add AGENTS.md as a symlink to CLAUDE.md so other agent tooling picks up
the same contract.
d07a083 changed the /api/energy/prices docstring (sell now deducts sell_fee)
and refreshed openapi/openapi.json, but frontend/src/api/schema.d.ts was not
regenerated, so CI's "check codegen is in sync" step failed. Comment-only diff.
The price chart keyed its X axis on formatLocalTime() "HH:mm" labels, which
repeat across a today+tomorrow range. Recharts resolves axis tooltips by value
(combineTooltipPayload -> findEntryInArray), so hovering a slot after midnight
matched today's identically labelled point: the tooltip showed today's prices
and the active dot jumped back to today's position instead of following the
cursor.
Key the axis on the ISO instant instead (buildChartRows, sorted by slot start)
and format down to HH:mm in the tick formatter; the tooltip label now carries
the date so today and tomorrow are distinguishable.
Also mark the price slot currently in effect by default: findActiveSlotIndex()
locates the slot containing now, rendered as a ReferenceDot on the buy and sell
lines plus a caption, re-evaluated every 30s.
Regression test drives a real mousemove over a sized chart in jsdom and asserts
the resolved slot and active-dot position.
Tibber's API `total` already includes the buy-side inkoopvergoeding
(verified from production data: total = spot×1.21 + energy_tax 0.11085 +
inkoopvergoeding 0.0248). Under net metering Tibber pays back
`total − verkoopvergoeding` per returned kWh (NL: EUR 0.28 -> 0.2552), so the
two EUR 0.0248 fees do NOT cancel — the feed-in price sits 0.0248 below buy.
Model the verkoopvergoeding as a first-class, always-subtracted contract
field `energy.sell_fee` (default 0.0248) instead of folding it into
`sell_adjust`. New sell formula:
sell = total − energy_tax − sell_fee − sell_adjust
`sell_adjust` now carries only the net-metering energy-tax refund
(= −energy_tax). Applied in both the billing strategy and the /prices
endpoint; recorded in the pricing snapshot. Frontend renders the field
automatically (dynamic profile form). Docs (references, m6) corrected to
drop the wrong "fees cancel" premise.
priceInfoRange is a historical cursor connection whose range ends at "now": it
never returns upcoming slots. With it, the DB only ever held prices up to the
last hourly refresh, which caused two problems:
1. The price chart could only show history up to now, never a forward curve.
2. Worse, per-slot billing was subtly wrong. _tibber_strategy looks up the
price via `starts_at <= t0` (nearest slot at or before the period). Because
a period's exact 15-min slot was not fetched until the next hourly refresh
(~1h later), intra-hour periods were billed with the PREVIOUS quarter's
price and then locked in by the immutability guard — never corrected.
Switch to priceInfo(resolution: QUARTER_HOURLY) { today tomorrow }, which is
forward-looking AND quarter-hourly: today is always the full local day (96
slots) and tomorrow fills in once Tibber publishes day-ahead prices (picked up
by the next hourly refresh). Every 15-min slot's exact price is now in the DB
before the slot closes, so each period finds its own slot (accurate billing)
and the live current-price entity stays fresh.
Verified against the live Tibber API: fetch_price_range returns 192 points
(96 today + 96 tomorrow), spanning local today 00:00 → tomorrow 23:45, with
future slots present (previously 0).
test_future_window_counts_0_days relied on the real wall-clock date and assumed
7/1→8/1 2026 was entirely in the future; once that range started elapsing the
window counted fixed-fee days and the assertion failed. Pin local_now to
June 25 2026 (as the sibling window tests already do) so the case stays
deterministic.
priceInfoRange is a Relay-style cursor connection over the subscription's
entire price history. With no cursor, first:96 returned the OLDEST 96 slots,
anchored at the subscription start date — a fixed window that never advanced.
For a contract added mid-period this meant refresh_prices kept re-upserting the
same day-one slots forever, so GET /api/energy/prices found nothing in the
today+tomorrow window and the UI showed "No Tibber price points available".
Use last:192 to return the newest 192 quarter-hourly slots (2 days), which ends
at the latest published slot and advances daily, fully covering the prices
endpoint's today+tomorrow window.
- Add CHINT DDSU666 profile (ddsu666.yaml): FC03 holding registers, voltage/
current/active power (kW)/reactive power (kvar)/PF/frequency, import+export
active energy. Word/byte order left at big-endian as a documented best guess
(manual has no float example) — to be confirmed on-device.
- Add DDSU666-Modbus-Protocol.md reference extracted from the official manual,
plus the source PDF (parity with the SDM120 reference).
- Generalize driver.read_blocks to dispatch FC03 (holding) or FC04 (input)
based on a function_code argument (default 4, SDM120 behaviour unchanged);
the code is validated before any connection is attempted.
- Wire profile.function_code through the CLI read command, the background
poller, and the device /test endpoint — previously the profile field was
declared but never honoured (read path was hardcoded to FC04).
- Tests: default -> FC04, function_code=3 -> FC03 holding, invalid FC rejected
before connecting.
Design for explicit meter lifecycle (epochs) so the register-difference
billing engine never computes a delta across a meter swap (2G->4G or house
move). Meter-centric (no Home entity); per-meter cumulative reset; retroactive
declaration with recompute; delta sanity guard as backstop. 7 atomic task cards.
- MQTT import_cost_total / export_revenue_total now anchor fixed-fee / tax-credit
accrual at max(contract start, first recorded period), so standing charges for
untracked pre-recording days are no longer folded into the running total.
- Add energy.import_cost_today / energy.export_revenue_today (monetary,
total_increasing) that reset at server-local midnight, for per-day cost
aggregation in Home Assistant alongside the running totals.
- ContractForm add-version mode prefills the contract's open version values, so
only the fields that actually change need editing.
- Store UTC, compute day boundaries in server local timezone (new app/services/timezone.py).
- summarize: fixed fee / tax credit accrue only for elapsed whole local days,
integrated across contract versions — no future-day inflation, no version-switch collapse.
- MQTT import_cost_total / export_revenue_total getters delegate to summarize (no inline
apportionment), anchored at the active contract's earliest version effective_from so the
total sensors stay monotonic and never jump backward.
- effective_from: naive datetimes interpreted as server-local wall-clock, then stored UTC.
- frontend ContractForm sends local-midnight naive datetime (was UTC midnight).
- get_costs_summary default window uses server-local "today".
A disabled device with historical readings still cannot be deleted by default
(409 guard stays, to prevent accidental data loss). A new explicit opt-in lets
the user remove it for real:
- DELETE /api/modbus/devices/{uuid}?cascade=true removes the device's readings
and exposed_entity_toggle rows, then the device, in one transaction (FK order:
readings before device). Best-effort clears the HA discovery configs first
(MQTT not connected -> no-op, never blocks the delete). Default (no cascade)
unchanged. Returns 200 with deletion counts.
- Frontend: a 'Force Delete (all data)' button appears only after the 409, with
a red irreversible warning; cascade is sent only on that explicit action.
- Also corrected a stale comment: FK RESTRICT IS enforced at runtime now.
Follow the standard MQTT Discovery split: discovery config stays under the HA
discovery prefix (homeassistant/.../config, required), but the actual state and
availability are published under our own prefix (default home_automation/). The
config payload's state_topic/availability fields point at the new prefix so HA
reads state from there.
- new config ha_state_topic_prefix (default home_automation).
- build_discovery_payload + all publish_* split discovery_prefix (config topic)
from state_prefix (state/availability). energy-cost still omits availability;
modbus availability structure unchanged. Tests updated.
Subscribe the separate DSMR tariff topic (dsmr/meter-stats/electricity_tariff,
1=dal/off-peak, 2=normal/peak), keep the latest slot in memory, and make the
manual-contract buy/sell_price_now sensors show the matching tariff's price
(tariff 1 -> dal, 2/unknown -> normal). Tibber (single 15-min price) unchanged.
Billing is untouched (register-split already handles tariffs correctly).
- new config dsmr_tariff_topic; subscribe managed by apply_dsmr_subscription
(restart-free on config save). In-memory tariff is lock-guarded.
- import_cost_total / export_revenue_total cumulative getters unchanged.
So HA can show the full payable: net = import - export = Total Payable, with
both entities staying monotonic-positive (no negative-value/reset pitfalls).
- import_cost_total = SUM(import_cost) + days*(network_fee+management_fee)/30
- export_revenue_total = SUM(export_revenue) + days*heffingskorting/365
(days = calendar days since the active version's effective_from, incl. today)
- state_class total_increasing -> total (monetary-compatible; values still only
increase). device_class/key/names: key unchanged, names clarified.
- buy/sell_price_now and DeviceInfo untouched. Decimal money math. Tests added.
Backend serializes SQLite-read (naive) UTC datetimes without a Z, so the browser
was parsing them as local time -> the UI effectively showed UTC values.
- New src/utils/datetime.ts: treats a tz-less timestamp as UTC, then formats in
the browser's local timezone with hour12:false (no AM/PM).
- Applied to Energy tabs (contracts/prices/costs/DSMR), per-device readings &
trends, and the meter readings table on the Records page.
- CostView Today/This month now use local day boundaries (queries stay UTC ISO).
- Untouched: location/poo (raw Zulu strings), the map, and query-window params.
- Tests are timezone-independent (verified under TZ=America/Los_Angeles).
Root cause (found in the live DB): register_at() returned the latest reading
for any boundary beyond the DSMR data range (a future instant or a long gap),
so future periods got start==end -> delta 0 and were written as degraded=False
'ok' rows with cost 0. recompute_range() over a future end then filled the whole
month with such fake rows, which then occupied the slots so the per-minute tick
(immutability guard skips non-degraded rows) never recomputed the real data.
- register_at: ignore a reading older than one period (15min) before the
boundary -> returns None -> period is marked degraded instead of a fake 0.
- recompute_range: only process closed periods (t1 <= now); never future ones.
- Tests cover staleness window, out-of-range -> degraded, recompute skips
future, and a regression that normal closed periods are unaffected.
The energy-cost device has only sensors (no online/offline heartbeat), but the
discovery config still declared an availability topic that nothing ever publishes
'online' to — so Home Assistant marked every energy-cost entity 'unavailable'
even though the state topic was being published (visible in MQTT Explorer).
- DeviceInfo gains provides_availability (default True, keeps Modbus behavior).
- _energy_cost_provider sets provides_availability=False.
- build_discovery_payload only emits availability/availability_mode for devices
that publish a heartbeat; HA treats the rest as always-available.
- Test asserts energy-cost configs omit availability.
- New 'DSMR' tab showing the latest parsed telegram (GET /api/energy/dsmr/latest)
as a key/value table — the source-of-truth view so you can see ingest is
landing data without reading the DB.
- Auto-refresh toggle (~10s); loading/error/empty states (empty explains the
likely cause); null phase values shown as a dash.
- useDsmrLatest extended with optional auto-refresh. Tests added.
Config hot-reload (no restart):
- MqttManager.unsubscribe(); apply_dsmr_subscription(settings) re-applies the
DSMR subscription (subscribe/unsubscribe/topic-change, fresh snapshot so the
sample interval also takes effect).
- Called from lifespan AND PUT /api/config, so toggling DSMR ingest in the UI
takes effect immediately without restarting the app.
Decouple ingest from the DSMR telegram id (overflows / resets to zero):
- Migration 20260624_12: drop UNIQUE(source_id), make recorded_at UNIQUE.
- Idempotency now keys on recorded_at (telegram timestamp); the table's own
autoincrement PK is the stable identity. source_id kept only as a reference.
- Regression test: colliding telegram ids no longer drop new data.
- routes/api/energy.py + schemas/energy.py: GET prices (tibber points or manual
tariff, buy/sell consistent with the pricing strategies), GET costs (time
window + limit, ascending), GET costs/summary (summarize passthrough), GET
dsmr/latest, POST costs/recompute (idempotent, <=366d guard, CSRF), POST
tibber/test (three-state, token never echoed, CSRF).
- main.py registers router; OpenAPI re-exported; tests for all endpoints.
Rework the M5 design doc data model and add manual-testing surface:
- Two-layer model: rename energy_meters/energy_readings -> generic
modbus_device + modbus_reading; readings become a JSON payload blob
instead of a wide fixed-column table (aggregation via SQLite
json_extract, generated-column index as a future lever).
- Protocol knowledge moves to read-only YAML profiles (data+functions,
no OOP inheritance); deployment/config (friendly_name, unit_id, host)
stays on the device row; multiple devices share one profile.
- UUID = stable internal identity, also anchors HA discovery unique_id
(Z2M rename model); device metadata for HA derived from the profile.
- Naming scheme C: modbus_* / /api/modbus/devices at the data+API layer,
'Energy' as the first domain view in the frontend.
- New task M5-T01B: ConfigPage Accordion sectioning (avoids a second
in-page side nav conflicting with the T01 app sidebar).
- Manual testing: read-only modbus_cli 'probe' subcommand (FC03/04 only,
no register writes); mqtt/test now publishes a test message viewable
in MQTT Explorer; no MQTT CLI (manual verification via UI + HA).
Promote four now-decided directions from Future Ideas into a new
"下一阶段：已确定要做" section (roadmap altitude, not yet broken into task cards):
TOTP second factor for the public dashboard, frontend optimization (scope TBD),
MQTT / IoT integration, and a settings-page long-lived API token (PAT-style,
related to but distinct from M3's mobile OAuth). Future Ideas is now an empty,
purpose-stated bucket for not-yet-decided ideas.
Multi-arch image builds (linux/amd64,linux/arm64) emulate the foreign arch via
QEMU. The unpinned `FROM node:22-slim` frontend-build stage was built per
target arch, so Node ran under emulation and V8's baseline JIT crashed (SIGTRAP
/ exit 133 on `npm ci`) on the CI runner's QEMU.
Pin the stage to --platform=$BUILDPLATFORM so the static, arch-independent SPA
dist/ is built once on the native build host and COPYd into each arch's runtime
stage. Node never runs emulated; both arch images are still produced.
Record TOTP (RFC 6238) as a deferred hardening idea for the now public-facing
Web dashboard: second factor on the single-admin login, with CLI-only password
reset and a CLI TOTP reset/recovery path that works even if the recovery codes
are lost (no lock-out dead end). Not M2.5, not scheduled — parked under a new
Future Ideas section.
Post-M2 self-walkthrough polish, batched into one commit.
Map / heat:
- fix heat-layer white-screen crash after login (add layer to map before
setLatLngs; an off-map leaflet.heat layer has a null _map and throws)
- normalize each heat layer to the densest pixel cell visible in the CURRENT
viewport (maxZoom:0 so intensity factor f=1) and recompute on moveend/zoomend,
so sparse poo data reaches red and stays normalized at any zoom level
- dark CARTO basemap tiles when the color scheme is dark
UI:
- dark-mode toggle in the top-right, beside the settings gear
- switch top-right nav (records / theme / settings / logout) to Feather icons
with hover tooltips
- home: Grafana-style quick time-range presets + back/forward shift buttons,
placed between the From/To pickers and Apply; fix Select/tooltip z-index
(Leaflet stacking) and the shift-button height alignment
API client:
- stop flooding GET /api/session with 401s: the session probe and the login
endpoint own their 401s (no global redirect), which fixes the logout hang and
the spinning login page
Compose:
- rename docker-compose.override.yml -> docker-compose.dev.yml as an explicit,
non-auto-layered dev stack (8001, -dev container names, prod-copy ./data DB);
update tests/test_deployment.py (read dev.yml, tolerate the !override tag) and
the README "Docker Compose" section
Tests:
- pixel-grid peak counter, time-range presets, heat-layer ordering regression,
and 401-redirect regression
- README: add 前端 v2 (React SPA) section (dev/build/codegen/hosting/gates),
update directory listing, drop stale Jinja descriptions
- architecture-overview: retire '不引入前后端分离' constraint; reflect SPA + JSON API
- roadmap: mark M2 done
- remove orphaned jinja2 dependency (recompile requirements*.txt; no other churn)
- delete empty tests/test_auth.py stub; drop dead _extract_csrf_token in test_api_data
- verified image still builds and app imports with the slimmer deps
- Dockerfile: node:22-slim stage runs npm ci + npm run build; python runtime
stage COPY --from copies dist to /app/frontend/dist (matches SPA_DIST_DIR);
runtime image has no node
- .dockerignore: exclude frontend/node_modules and frontend/dist from context
- .github/workflows/frontend.yml: npm ci + codegen-sync + lint/typecheck/test/build
- tests/test_deployment.py: skip COPY --from sources in the context-existence
check; assert the multi-stage frontend build wiring
- verified with a real docker build (image serves SPA, no node at runtime)
tliu93