mirror of
https://github.com/smkrv/ha-text-ai.git
synced 2026-07-21 22:54:00 +08:00
SMKRV
5af733b1b0
Features: - disable_thinking toggle (issue #11) with per-provider semantics: OpenAI classic gets /no_think soft-switch, OpenAI reasoning gets reasoning_effort=low, DeepSeek gets both, Anthropic no-op (thinking opt-in), Gemini 2.5+ gets thinking_budget=0 - OpenAI reasoning model support (o1/o3/o4-mini/gpt-5 family): max_completion_tokens, developer role, reasoning_effort - Per-request disable_thinking override in ask_question service - Provider-specific temperature clip (Anthropic 0-1, others 0-2) Security: - Require API key re-entry on provider change (OptionsFlow) - Validate Anthropic json_schema before system-prompt concatenation - Symlink protection in history file operations - Hardened secret redaction regexes (sk-*, AIza*, x-api-key) - get_history service: hard cap on limit (default 10, max 100) Reliability: - Retry on 502/503/504 in addition to 429/timeout - Honor Retry-After header on 429 - Exception chaining (raise ... from err) - _strip_think_blocks handles nested/dangling tags - normalize_name sha256 fallback for empty-collapse inputs UI/housekeeping: - api_key uses TextSelector(type=PASSWORD) - DeviceInfo entry_type=SERVICE - Services unregister on last entry unload - Dependabot for GitHub Actions - persist-credentials=false in checkout steps - manifest requirements pinned with upper bounds - Ignore docs/plans/ (working artifacts) License: PolyForm Noncommercial 1.0.0 -> MIT No breaking changes: service response shape, sensor attributes, entity IDs and on-disk formats unchanged.