agent-platform release v4.66.5

Fixed

• The kagent floor is 1.0.3: a ModelConfig can switch an Ollama model’s thinking off (giantswarm/kagent-upstream#81, giantswarm/model-manager#161). A model with Ollama’s thinking capability (Qwen3.5, Qwen3, DeepSeek-R1, Granite 4.2) reasons before every answer unless the chat request says think: false, which kagent’s native Ollama provider never sent: on a CPU an agent turn paid 1.5–2k thinking tokens per model call, and qwen3.5:2b wrote its whole answer into the thinking, which the runtime does not read, so the agent replied with nothing. kagent 1.0.3 adds spec.ollama.think to the ModelConfig and sends it as the request’s top-level think; model-manager writes it (false by default) on the Ollama models it wires that have the capability. components.kagent.versionRange and components.kagent-crds.versionRange are >=1.0.3 <1.1.0, the example customer BOM pins 1.0.3. No installation acts: the kagent controller and UI restart onto 1.0.3 where the range resolved an older release.