tyres: use HA's unit metadata instead of assuming kPa (#39)

[ThinkOffApp]

Fixes #39 — @eclass reported "tyres holding steady at thirty bar" on 11 Sep. Thirty bar is ten times a car tyre, and it had been saying so for a week.

The trace, raw value to spoken value

@codexmb asked for this before any conversion was chosen, and it changes what the fix is:

1. GET /api/states returns each entity with state and attributes.unit_of_measurement.
2. mercedesme.py reads attrs (it already did, line 327) but passed only the state string to _norm_value.
3. The value was filed under the hardcoded key tires_kpa whatever unit HA actually reported.
4. agent.py emitted "tyre pressures kPa <n>" — a bare number plus a unit word.
5. The model converted it itself, and a decade went missing.

So the bug is not a missing conversion. The unit was fetched and thrown away.

What changed

• _pressure_kpa() converts from bar / psi / kPa / mbar / hPa explicitly, and returns None on an unknown or missing unit rather than assuming kPa. An unlabelled pressure is not a kPa reading.
• The fetch passes attributes.unit_of_measurement through for tyre entities.
• fmt_tyres_bar() renders bar to one decimal, as the issue asks, so the model never converts at all.
• A physically implausible reading (outside 1.0–4.5 bar) is flagged as a sensor fault instead of stated as the pressure.
• webchat.py was interpolating the raw dict into the prompt (tyres {'front_left': 250, ...} kPa); it now uses the same formatter.

Tests

tests/test_tyre_units.py, 7 cases: every unit HA might report, the refusal on an unknown unit, 30 psi no longer becoming thirty-anything, bar-with-one-decimal output, wheel ordering, the implausibility flag, and non-pressure entities untouched.

Run in isolation and with the suite. tests.test_dash_loop fails on main as well — pre-existing, unrelated to this change.

Not verified

Which unit this particular car's HA reports is still unconfirmed — the Mini's snapshot is 59 h stale and the local one is empty. "30 meant psi" remains a hypothesis. The fix removes the assumption either way, and the implausibility flag catches the symptom regardless of cause. Worth confirming against the live car when it is next reachable.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
carwatch-dev	Ready	Preview	Sep 18, 2026 4:56pm UTC

[chatgpt-codex-connector]

To use Codex here, create an environment for this repo.