The SDK broadened its AI capabilities while hardening feature flag evaluation and improving operational reliability. Prompt management shipped in v7.8.0, then expanded to support versioning and multimodal capture across OpenAI, Gemini, and Anthropic integrations. Local flag evaluation gained semver targeting, device_id bucketing, and ETag caching to reduce polling overhead. The library also tightened retry logic for timeouts and 401 errors, exposed the flag definition cache provider for custom implementations, and surfaced flag evaluation failures through a new $feature_flag_error property.
Feature flag evaluation expanded to support device_id as a bucketing identifier for local evaluation, while AI token tracking gained the ability to detect value overrides through a new $ai_tokens_source property. The month also hardened reliability across the board—fixing IP address property naming, preventing HTTP request storms on 401 errors during flag refreshes, and adding Retry-After header respect for 408 responses.