home-assistant-manager — Claude Code skill: how to install

---
name: home-assistant-manager
description: Manage Home Assistant configuration safely and fast — edit and deploy YAML (automations, blueprints, scripts, scenes, templates, MQTT), validate with ha core check, deploy via git or rapid scp, reload-vs-restart correctly, verify changes from logs, traces and entity state, and build Lovelace dashboards. Use for any Home Assistant config, automation, template, or dashboard work over SSH/hass-cli/MCP.
---

# Home Assistant Manager

Operate a remote Home Assistant instance precisely: make a change, get it live, prove it
worked. Optimize for the fewest safe round-trips.

## Assumptions
- The repo you're editing **is** the HA `/config` dir, git-connected to the instance.
  Edits aren't live until pulled on the instance.
- `[email protected]` in examples is a placeholder. Resolve the real user/host once
  (project CLAUDE.md, `~/.ssh/config`, or ask) and if it isn't recorded in the project
  CLAUDE.md yet, add it so future sessions skip this step.
- Access via one or more of: `hass-cli` (REST), SSH `ha`, or an MCP server (see below).
- Only edit `.yaml`/`.yml`/`.md`. Never read/write `.env` or `secrets.yaml`; use `!secret`.

## Remote access — pick the right tool
- **SSH `ha`** — always works, needs no local env. Use for `ha core check|restart|logs|info`.
- **`hass-cli`** (REST) — state/service calls, but needs `HASS_SERVER`/`HASS_TOKEN` in the
  shell *before* the session starts. If they're unset, hass-cli falls back to the wrong
  host (localhost) and errors — don't retry, check `[ -n "$HASS_TOKEN" ]` once, then use
  SSH or MCP instead.
- **MCP** (preferred when available) — first-class tools for live state/control, no env
  juggling. Official `mcp_server` integration (HA core ≥2025.2) or community `ha-mcp`
  (richer, 80+ tools). Use it instead of shelling out when present.

## The deploy pipeline (the one canonical flow)
Changes are not live until step 4.
1. Edit YAML locally.
2. Validate: `ssh [email protected] "ha core check"` (slow, ~30-60s — see "when to
   skip" below).
3. Commit + push: `git add … && git commit -m "…" && git push`.
4. **Make it live:** `ssh [email protected] "cd /config && git pull"`.
5. Apply: **reload** if possible, else **restart** (table below).
6. Verify (next section).

**Rapid iteration:** skip git and `scp` straight to the instance, then reload — good for
dashboards and tight test loops. Commit to git only once stable.
`scp automations.yaml [email protected]:/config/` → reload.

**When to skip `ha core check`:** it parses the whole config and is slow. For an isolated
YAML edit you're confident in, a domain reload surfaces errors faster and the logs tell
you immediately. Always run it before a *restart* or for `configuration.yaml` changes.

## Reload vs restart
| Change | Action |
|--------|--------|
| automations, scripts, scenes, groups, template entities, themes | **reload** the domain (`hass-cli service call automation.reload`, etc.) |
| `configuration.yaml` core, new integrations, platform sensors (min/max), MQTT sensor/binary_sensor platforms, dashboard registry (`lovelace_dashboards`) | **restart** (`ssh … "ha core restart"`, ~30s) |

Prefer reload. Never restart without a passing `ha core check`. Before risky changes
(core `configuration.yaml` surgery, removing an integration), snapshot first — it's cheap:
`ssh [email protected] "ha backups new --name pre-<change>"`.

## Verify — don't assume it worked
1. Reload/restart the right domain.
2. For automations, **trigger manually** for instant feedback:
   `hass-cli service call automation.trigger --arguments entity_id=automation.<id>`
   (or call the service via MCP). This **bypasses `conditions` by default** — it proves the
   actions, not the gate. To test conditions too, pass `skip_condition: false` or exercise
   the real trigger, then read the automation's trace in the UI.
3. Read the logs filtered to your change:
   `ssh [email protected] "ha core logs | grep -iE '<name>|error' | tail -20"`.
   Good: `Running automation actions`, `Executing step …`. Bad: `Invalid data for
   call_service`, `TypeError`, `Template variable warning`, `Error executing script`.
4. Confirm the real outcome: device/sensor state (`hass-cli state get <entity>`), or ask
   the user for notification-type actions.
5. On error: fix → re-pull/scp → reload → re-check. Loop until clean.

## Automations — write modern syntax
HA 2024.10 renamed the keys; legacy syntax still works but don't emit it in new code:
top-level `triggers:/conditions:/actions:` (plural), `trigger:` not `platform:` inside a
trigger, `action:` not `service:` for calls. Every automation gets a stable `id:` (traces
and UI editing need it) plus an `alias`.

**Full automation reference** (syntax table, `mode:` behavior, blueprints, trace debugging,
pitfalls) → read [`reference/automations.md`](reference/automations.md) when writing or
debugging automations.

## Templates — the precision rules
- Always coerce types before comparing: `states('sensor.x') | int(0) < 7`. Bare states are
  strings; `'5' < 7` raises `TypeError`. Provide a default (`int(0)`) so startup `None`
  doesn't error.
- Test in **Developer Tools → Template** before committing.
- `state_attr(...)` returns `None` if the entity/attr is missing — guard it.

## Conventions
- Surgical edits; preserve comments; 2-space indent.
- Validate before restart; prefer reload; verify from logs.
- Use context7 MCP for current HA docs before non-trivial or unfamiliar config.

## Dashboards
Lovelace dashboards live in `.storage/lovelace.*` (JSON). UI edits show on a browser refresh;
**direct file edits (scp/git) may not appear until a `ha core restart`** — HA caches the
lovelace store in memory. Adding a *new* dashboard to `.storage/lovelace_dashboards` also needs
a restart. `scp` + refresh is the fast loop; validate JSON first:
`python3 -m json.tool .storage/lovelace.x > /dev/null`. After deploying, **validate the UI
visually in the browser** (see the reference) — logs/state won't catch a broken card or a
mis-sorted popup.

**Full dashboard reference** (view types, card catalog, template cards, tablet layout,
pitfalls, debugging) → read [`reference/dashboards.md`](reference/dashboards.md) when doing
UI work. Modern HA: native **sections** view (drag-drop grid, badges, `heading` cards) and
feature-rich **tile** cards now cover most needs without custom cards; reach for Mushroom
only when you want its specific look.

## Quick reference
```bash
# Validate / apply
ssh [email protected] "ha core check"
ssh [email protected] "ha core restart"
ssh [email protected] "cd /config && git pull"     # make pushed changes live

# Logs
ssh [email protected] "ha core logs | grep -iE 'error|<name>' | tail -20"

# State / services (needs env loaded, or use MCP)
hass-cli state get <entity>
hass-cli service call <domain>.reload
hass-cli service call automation.trigger --arguments entity_id=automation.<id>

# Rapid deploy
scp <file>.yaml [email protected]:/config/ && hass-cli service call automation.reload
```