AI Automation Suggester 1.6.0

Release date: 2026-09-05

Highlights

MiniMax support: Configure a MiniMax API key, global or China endpoint, model, and temperature. The default is MiniMax-M3, with MiniMax-M2.7 also in the catalog. Thanks @octo-patch for PR #178.
LiteLLM model sensor: The sensor now displays the configured model instead of Unknown Model Key, without the repeated warning. Thanks @mjacobs for PR #186, fixing #185.
Ollama native thinking control: The existing Disable Think option now sends think: false, addressing slow or empty responses from models that ignore the /no_think prompt hint (#188). The option remains opt-in.
Current provider defaults: New Google entries use gemini-3.5-flash (#184). New Groq entries use openai/gpt-oss-120b, with migration warnings for retired Llama IDs and a preview label for qwen/qwen3.6-27b (#187).
Long local requests: Setup and options now accept request timeouts above 1800 seconds (#182). The default remains 900 seconds, the minimum remains 10 seconds, and zero does not disable the timeout.
Nested YAML wrappers: Complete Markdown YAML fences inside structured suggestion fields are removed before validation. Newlines and four-space indentation are preserved, including in best-effort recovered JSON. This fixes a reproducible wrapper case investigated for #172.

Update through HACS and restart Home Assistant. No config-entry or stored-history migration is required.
Existing model selections are preserved. If an older Google or Groq model returns a model-not-found error, select a current model in the integration options. Groq's Llama retirement affects free/developer accounts, not every enterprise account.
For Ollama models affected by #188, enable Disable Think. For long-running local models, set a suitable timeout, such as 7200 seconds for two hours.
Existing token budgets are unchanged. For reasoning models or truncated YAML, start with 16000 input tokens and 4096 output tokens, a small entity limit, and no automation/script YAML reading. Higher budgets can increase actual usage and cost. Check the provider's usage dashboard.

initializing is not evidence of an API outage. Startup does not run inference. Run generate_suggestions with all_entities: true and inspect the action result and provider sensor error attributes. Report #177 still needs an explicit generation result.
The existing reasoning_content/reasoning fallback for empty OpenAI-compatible responses is covered by new tests (#127). A completely empty response remains an explicit error.
Reports #127 and #172 remain open for retesting and raw-response evidence. YAML whose newlines are already missing cannot be safely reconstructed by guessing indentation.
Claude Pro/Max subscription authentication is not implemented (#179). The Anthropic provider continues to use supported API key authentication.

89 automated tests pass in clean Python 3.11 and 3.12 environments.
Repository-wide Ruff checks pass.
Release checks include Python compilation, dashboard JavaScript syntax, integration JSON/YAML parsing, HACS, and hassfest.
Provider HTTP behavior is tested with mocks. No live paid-provider inference was performed for this release.

Thanks @octo-patch and @mjacobs for the merged provider contributions, and the issue reporters for their reproduction details.