# AGENTS.md — Project state, review findings, open items, and working rules **Last updated: 2026-08-11** (full-repo review). This file is the entry point for every new agent session: what this repo is, what is finished, what is missing, and how to work here. Detail lives in the linked documents — this file points, it does not duplicate. > **Maintenance rule (binding):** whenever you change this repo in a way that affects anything > recorded here — completing an open item, making or reversing a decision, adding/removing a > component, discovering a new gap — **update this file in the same session** (tick the checkbox, > adjust the status table, bump the "Last updated" date). This file must always be written in > **English**, even though the rest of the project is German. **How this file is loaded:** Claude Code does not read `AGENTS.md` natively; the root `CLAUDE.md` imports it via `@AGENTS.md` (official recommended pattern). Other agents (Codex, Cursor, Copilot) read `AGENTS.md` directly. Edit content here, not in `CLAUDE.md`. --- ## Working rules ### Karpathy guidelines (adopted as project standard) From [`multica-ai/andrej-karpathy-skills`](https://github.com/multica-ai/andrej-karpathy-skills), behavioral rules derived from Andrej Karpathy's observations on LLM coding failures. They bias toward caution over speed — for trivial tasks, use judgment. 1. **Think before coding.** Don't assume. Don't hide confusion. Surface tradeoffs. State assumptions explicitly; if multiple interpretations exist, present them — don't pick silently. If a simpler approach exists, say so. Push back when warranted. If something is unclear, stop and ask. 2. **Simplicity first.** Minimum code that solves the problem. Nothing speculative: no features beyond what was asked, no abstractions for single-use code, no unrequested configurability, no error handling for impossible scenarios. If 200 lines could be 50, rewrite. 3. **Surgical changes.** Touch only what you must. Don't "improve" adjacent code, comments, or formatting; don't refactor what isn't broken; match existing style. Remove only orphans **your** change created; mention (don't delete) pre-existing dead code. Every changed line should trace directly to the request. 4. **Goal-driven execution.** Define success criteria, loop until verified. "Fix the bug" → write a test that reproduces it, then make it pass. For multi-step tasks, state a brief plan with a verify step per item. ### Claude Code practices (current guidance, 2026) - Keep this file **concise and specific** — instructions concrete enough to verify ("run X before Y"), markdown headers/bullets, no dense prose. Context files reduce adherence as they grow; target roughly ≤200 lines of actual instruction. - Don't record here what the code or git history already shows (directory dumps, dependency lists) — record pitfalls, rationale, decisions, and conventions that differ from defaults. - Conflicting instructions get picked arbitrarily — when updating, **remove** superseded rules rather than stacking corrections. - Verify loading with `/context` (file must appear under **Memory files**); edit via `/memory`. --- ## What this project is Private replacement for the Audi *connect plug & play* app (discontinued end of 2026), for exactly one vehicle (**Audi RS 4 Avant competition**), single user, self-hosted on Home Assistant, access via Tailscale only. Read-only toward the vehicle — no remote control. Covers: vehicle status, trip log, fuel log with Shell receipt parsing, service forecasting, insurance/tax, tire management. **License rule (hard):** Audi Type fonts, four-rings SVG, and model-badge SVGs are cleared **only for this one private, unpublished installation**. Therefore `design-system/` is brand-free (it gets uploaded to Claude Design), while the HA panel and the future sideload-only app may use the real assets. Never mix these the other way around. **Reading order for a new session:** 1. This file (overview + open items) 2. `SPECIFICATION.md` — authoritative for the HA panel, incl. §7 "Known Gaps" 3. `COMPANION_APP_ARCHITECTURE.md` — authoritative for DataMetric360 (decided, barely built) 4. `AUDIT_2026-08-10.md` — accessibility/platform audit of the panel (partly done, rest below) 5. `bauauftrag.md` — **historical only**; code has diverged (see SPECIFICATION.md §7) ## The three projects in this repo | Area | What | Status | |---|---|---| | `homeassistant/` | HA panel (`panel_custom`): pyscript backend + vanilla-JS frontend | ✅ **finished, in use** | | `design-system/` | React component library `@audi-dash/ui`, brand-free, feeds Claude Design | ✅ done as a kit (20 components, 1,690 lines) | | `companion-app/` | **DataMetric360** — successor app (Capacitor iOS/Android + HA iframe); will **replace** the panel | 🚧 data layer only (~930 lines TS), **no UI** | | `design/` | Export of the Claude Design draft for DataMetric360 | 🚧 interim (wrong model, main screens only) | Root files: `dashboard-muster*.html` = original static prototype (superseded, reference only), `bauauftrag.md`/`.html` = original build brief (historical), `DESIGN_BRIEF_DATAMETRIC360.md` = the prompt for the Claude Design project. **`homeassistant/` in one paragraph:** 10 pyscript scripts + 4 modules (`pyscript/modules/`); trip detection via the iPhone WLAN sensor with pause tolerance, two-stage trip completion via HA history screening (odometer often updates only on the next trip), fill-up detection on fuel-level rise, Shell PDF parser (`data/shell_beleg_parser.py`, subprocess, the only tested part of the repo), tire km counter, backup, self-update, image management. Frontend: one file `www/audi-dashboard-app.js` (3,130 lines, custom element, no framework/bundler), 5 tabs + ~19 detail routes, cache-busting via `audi-dashboard-version.json` + loader stub. Entity IDs configured centrally in `pyscript/modules/einstellungen.py` (the one file edited before install). Deploy: `update.ps1` (robocopy to Samba share) or — still inactive — self-update from git. **`companion-app/` — what exists:** dependency-free, strictly typed TS data layer (`src/api/`): REST client (`rest.ts`), WebSocket client with auth flow + reconnect backoff (`live.ts`), persistent offline write queue (`warteschlange.ts`, strict FIFO), environment detection capacitor/iframe/browser (`umgebung.ts`), types + entity table (`types.ts`), facade `DataMetricApi` (`index.ts`). Smoke test ran 7/7 green against Docker HA `audi_ha_test` (localhost:18123) — unauthenticated only; token-authenticated reads/writes untested. **DataMetric360 architecture (short — details in `COMPANION_APP_ARCHITECTURE.md`):** - Future data source: **Teltonika FMM003** on the CAN bus, fully replacing the iPhone WLAN sensor and the VAG integration. - **Current data path (since 2026-08-11): Codec JSON → MQTT/TLS → Mosquitto add-on → HA.** Traccar is **dropped** (the decision log deliberately remains in the architecture doc, marked ÜBERHOLT). - Frontend access: HA REST + WebSocket with a long-lived access token (no `hass` object). - External access: Cloudflare Tunnel + reverse proxy with a path allowlist; HA itself stays unreachable. Domain **`datametric360.app`** registered (all-inkl, 2026-08-11). - Distribution: **sideload only** (license reason) — hence real Audi assets are allowed there. - Trip detection moves to FMM003 ignition; the pause-tolerance feature (`fahrten_pausenzeit_min`) is a deliberate product decision and must be preserved. --- ## Review findings (2026-08-11) ### HA panel — known gaps (most also in SPECIFICATION.md §7) - **GPS is dead schema:** `start_lat/lon`, addresses, `route`, `avg_speed_kmh` never populated. The trip-detail map draws a **fabricated** line via `fakeTrack()` (`audi-dashboard-app.js:414`) — not a real track. - **No GPS fallback in trip completion:** trips without an odometer match stay `status="offen"` forever (`modules/fahrtabschluss_logik.py:16-20`). - **RAM-only state:** running trip (`_fahrt_start_ts`) and fuel low-water-mark (`_tiefststand_pct`) do not survive an HA restart. Deliberately deferred hardening. - **Inactive features:** `BATTERIE_SENSOR = ""` (entire battery-history feature is a no-op) and `UPDATE_REPO_URL = ""` (self-update inactive), both in `pyscript/modules/einstellungen.py`. `www/bilder/` has no vehicle photos (all slots show placeholders); `steuer.faellig` unset. - **Robustness:** `profil_lesen()` in `modules/profil.py` does not handle a missing/corrupt `fahrzeugprofil.json` — all callers throw. - **Tests:** only the receipt parser is tested (`data/tests/test_shell_beleg_parser.py`, 10 real receipts) — and its test PDFs are gitignored, so a fresh clone can't run it. Backend and frontend: no tests, no CI. - **Leaflet via CDN:** trip map needs public internet in addition to the Tailscale tunnel. ### Documentation drift (small fixes; align docs to code) - `homeassistant/README.md:99`, `INSTALL.md:204`, and the header comment `audi-dashboard-app.js:13-15` claim the statistics view shows sample numbers — **false**; `vStat()` computes real values from `TRIPS`/`FILLS`. - `INSTALL.md` step 4 names variables that no longer exist (`DOORS_SENSOR`, `WINDOWS_SENSOR`, `LOCK_ENTITY`, `BATTERY_VOLTAGE_SENSOR`) — actual names: `TUER_SENSOREN`/`FENSTER_SENSOREN`/ `TUERSCHLOSS_SENSOREN`/`BATTERIE_SENSOR`. - `homeassistant/README.md:110` mentions the removed 97% full-tank rule (removal documented in `belegverarbeitung.py:18-20`); the README file list omits 5 pyscript files. - Obsolete comment `belegverarbeitung.py:41` ("TODO: Datei ablegen" — file has long existed). ### Audit leftovers (`AUDIT_2026-08-10.md` §4, deliberately left open) - 🟠 Swipe-to-delete has no gesture-free fallback — screen-reader/switch-control users cannot delete trips/fill-ups. Needs a design decision (long-press vs. "…" button vs. action sheet). - 🟡 Popup close-by-tap-outside is not keyboard-reachable (needs a quick manual check). - 🟡 Self-host Leaflet JS/CSS (tiles necessarily stay remote). - Audit's own note: these three may be better done in DataMetric360 than retrofitted — the panel gets replaced anyway. ### companion-app / design-system - **Not wired together:** `companion-app` does not reference `@audi-dash/ui` anywhere (no dependency, no import, no workspace root). The link exists only in prose. - Both packages: no `node_modules`, no `dist` — smoke tests need `npm install` first (design-system additionally `npm run build`; `scripts/smoke.mjs` imports from `../dist/`). - No unit tests, no Storybook (substitute: SSR smoke over 21 cases in design-system). - `design/` export (2026-08-11) has two known, already-commissioned fixes not yet in the export: shows **RS 6 instead of RS 4**, and **16 sub-pages** the HA panel already has are missing (list in `design/README.md`). --- ## Open items ### A) Build DataMetric360 (the big block) - [ ] Fix the Claude Design draft (RS 4, not RS 6) and extend it by the 16 missing sub-pages; then re-export to `design/` - [ ] `companion-app`: set up Vite + React + Capacitor scaffold; wire `@audi-dash/ui` as a real dependency (possibly add a workspace/monorepo root) - [ ] Implement the screens from the design draft on top of the existing `DataMetricApi` layer - [ ] Add Audi assets (fonts/rings/badges) at implementation time from `homeassistant/www/` — **never** into `design-system/` - [ ] Secure storage for the LLAT (iOS Keychain / Android Keystore via Capacitor plugin; the `ablageSetzen()` hook already exists) - [ ] Onboarding: manual token paste (required); QR scan only if it stays simple (QR generated locally under HA `/local/`, architecture §3) - [ ] Authenticated smoke tests of the data layer (reads, service calls, queue round-trip) against `audi_ha_test` with a real token - [ ] Offline UX per design brief (offline marker, visible pending queue) ### B) Infrastructure / commissioning (partly waits for FMM003 hardware) - [ ] Switch `datametric360.app` nameservers at all-inkl to Cloudflare ("full setup") — prerequisite for the tunnel; domain carries nothing else, so this is consequence-free - [ ] Decide hostname split (app on apex + API on `api.` subdomain, or vice versa) - [ ] Choose reverse proxy (Nginx Proxy Manager vs. Traefik) — **can be done before hardware** against the existing HA API - [ ] Define the reverse-proxy path allowlist (depends on final entity/service names) - [ ] Install/wire the FMM003; record firmware version (Codec JSON is firmware-dependent) - [ ] Generate TLS certificates for Mosquitto + device (small private CA); FMM003 refuses MQTT without them - [ ] Decide broker reachability for the vehicle: port-forward 8883 vs. VPS broker with Mosquitto bridge over Tailscale - [ ] Capture the first real Codec JSON message (`mosquitto_sub`/MQTT Explorer) and build the field mapping from it — **do not guess beforehand** (explicit decision) - [ ] Move trip detection to FMM003 ignition (reuse the `fahrterkennung.py` pattern, keep pause tolerance); then remove `sensor.iphone_wifi_connection` from `einstellungen.py` ### C) Maintain the existing HA panel (low priority — being replaced) - [ ] Fix documentation drift (statistics claim, INSTALL variable names, README gaps, obsolete TODO comment) — text-only changes - [ ] Harden `profil_lesen()` against missing/corrupt `fahrzeugprofil.json` - [ ] Decide whether the 3 audit leftovers get fixed here or only in DataMetric360 - [ ] Optional: persist the RAM-only states (trip start, fuel low-water-mark) — deliberately deferred; may become moot with the FMM003 switch - [ ] Upload vehicle photos to `www/bilder/`, set `steuer.faellig` (operational data, not code) ### D) Once the panel is superseded - [ ] **Archive** `audi-dashboard-app.js`, don't delete (settled decision, architecture §1) --- ## Working conventions (observed — keep them) - German is the project language: identifiers, comments, commits, UI texts. Exceptions: `design-system/` uses English props/JSDoc (Claude Design audience) — and **this file**, which is English by rule. - Decisions are logged, including rejected ones (see the Traccar section of the architecture doc); superseded sections stay in place marked "ÜBERHOLT" rather than being deleted. - Numbers in `de-DE` format; font weights 300/400 only, never ≥600 (only those cuts exist). - Real vehicle/movement data stays local (`data/` contents and receipt PDFs are gitignored). - Security principle: HA is never publicly exposed; only narrowly scoped surfaces (MQTT broker, proxy allowlist) may be exposed, each by explicit decision.