Skip to content

Zonit.Extensions.Ai 10.6.0

Latest
Latest

Choose a tag to compare

View all tags
@Alerinos Alerinos released this 21 Jul 22:43
v10.6.0
49b86df

First release since 10.2.2 — this consolidates the entire 10.3.x → 10.6.0 line into one drop.

New models

  • Claude Sonnet 5 (Anthropic).
  • Grok 4.5 (grok-4.5, xAI).
  • OpenAI GPT-5.6 tier — new Sol56, Terra56, Luna56 models (the Sol/Terra/Luna naming replaces pro/mini/nano).
  • Grok Imagine Video 1.5 (grok-imagine-video-1.5) — xAI's current standard model for text-to-video and image-to-video, at 480p/720p/1080p.

New modalities & providers

  • Speech synthesis (text-to-speech) via a new ElevenLabs provider (ISpeechLlm) — voice catalog, selectable audio formats, DI registration, and per-character billing.
  • Video generation, made robust: dedicated IImagePrompt / IVideoPrompt prompt types with ready-made ImagePrompt / VideoPrompt (and ImagePromptBase / VideoPromptBase for templated prompts). Image/video generation can no longer be miswired with a plain text prompt. Input is now validated centrally against each model's declared channels (ILlm.Input): passing a video into a text/image-only model, or a prompt with neither text nor a source file, is rejected before any API call — the Asset value object detects the real media type from the bytes, so a wrong extension can't slip through.

Features & improvements

  • Global HTTP/SOCKS proxy for AI providers (config section Ai:Proxy) with per-provider opt-out — solves regional locks (e.g. Grok in the EU).
  • Strong ReasonEffort typing for OpenAI reasoning models — compile-time-safe effort levels (OpenAiReasonEffort / OpenAiReasonEffortExtended); the effort→wire mapping is centralized.
  • OpenAI agent session now resubmits the full configuration (tools, instructions, format, reasoning) on every previous_response_id continuation — fixes agents that stopped calling tools after the first turn. Includes a regression test.
  • ConversationInfo in the agent context (IRunContext) — exposes MessageCount and IsEmpty so tools and sub-agents can react to conversation state.
  • Resilience: recovery from double-encoded JSON in AI responses, with a dedicated test.

Breaking changes

  • Image/video generation overloads now take IImagePrompt / IVideoPrompt instead of IPrompt<Asset>. SimpleImagePrompt and ImagePromptBase already implement the new interface; migrate any custom IPrompt<Asset> image/video prompts to ImagePrompt / VideoPrompt (or the *PromptBase classes). Text/string overloads are unchanged.
  • ReasonType removed for OpenAI reasoning models — use OpenAiReasonEffort / OpenAiReasonEffortExtended.

Full changelog: v10.2.2...v10.6.0

Assets 2
Loading