v1.4.1

Highlights

• Reasoning models via the OpenAI Responses API (GPT‑5, etc.). Inputs are input_text, system prompt goes into instructions, and we add reasoning.effort when appropriate.
• Assist works cleanly: return a ConversationResult compatible with HA’s agent manager; no more as_dict crashes.
• Async‑safe OpenAI client: inject Home Assistant’s shared HTTPX client to eliminate the load_verify_locations warning in logs.

Why it matters

• The Responses API is the path forward for reasoning‑class models; this release aligns with the latest schema and keeps you future‑ready.
• Options are now editable from the gear (no re‑add needed). Uses OptionsFlowWithReload.

Upgrade guide

1. HACS → Update to v1.4.1
2. Restart Home Assistant
3. Settings → Voice Assistants → set Conversation agent = Extended OpenAI Conversation
4. Open the integration card → Configure (gear) and set:
   • Model: gpt-5
   • Strategy: auto
   • Reasoning effort: medium
   • Max tokens: 800–1200
   • Leave temperature/top‑p at defaults (not sent to reasoning models)

Azure OpenAI

• Set Base URL and API version (e.g., 2024-10-21); the integration automatically uses the Azure client.

What we fixed from real logs

• 400 invalid input type: using input_text + instructions resolves
  Invalid value: 'text' .... :contentReference[oaicite:5]{index=5}
• as_dict crash: implemented a compatible shim so Assist can serialize the result. :contentReference[oaicite:6]{index=6}
• Blocking SSL warning: pass HA’s shared HTTPX client to OpenAI. :contentReference[oaicite:7]{index=7}