gtapps/ha-build-automation

Build HA Automation

Steps

1. Gather context:

   • Read .claude-code-hermit/raw/snapshot-ha-normalized-latest.json for available entities and services.
   • Use the stored language from OPERATOR.md (## HA hermit section) for alias and description fields.
   • Optionally call GetLiveContext via MCP for current state.

2. Draft the YAML:

   • Always include id: as the first field, using a stable, language-neutral snake_case value (e.g., kitchen_motion_after_sunset_notification). The validate-apply command uses this as the REST config ID — omitting it causes a derived, fragile ID that breaks on alias rename.
   • Use the stored locale for alias and description.
   • Set mode explicitly where concurrency matters.
   • Write to .claude-code-hermit/raw/automation-<automation_id>.yaml.

3. Validate:

   • Run ${CLAUDE_PLUGIN_ROOT}/bin/ha-agent-lab ha simulate <path> to check entity references and policy.
   • Run ${CLAUDE_PLUGIN_ROOT}/bin/ha-agent-lab ha policy-check <path> for a safety assessment.

4. Handle results:

   • severity: "allow" — valid and safe: offer to apply via /claude-code-homeassistant-hermit:ha-apply-change.
   • severity: "ask" (ask mode) — explain which sensitive entities are involved, then offer to apply. The apply step will require explicit operator confirmation.
   • severity: "block" (strict mode) — explain why and create a proposal using /claude-code-hermit:proposal-create.
   • If entities are missing: suggest refreshing context first.

YAML Conventions

• IDs: snake_case, language-neutral, descriptive
• Aliases: stored locale, human-readable
• Descriptions: stored locale, explain the purpose
• Triggers: use platform: explicitly
• Actions: use service: with full domain (e.g., light.turn_on)
• Targets: prefer entity_id over area/device when specific

Safety

Under ha_safety_mode: strict (the default): never draft automations that actuate lock, alarm_control_panel, or security-related cover/button/switch. If the user requests this, explain the safety boundary and create a proposal for manual review. Under ask: draft and run ha policy-check — the severity field in the result drives step 4.