agent-platform release v4.66.2

The model-manager floor is 1.0.8: the Ollama ModelConfigs it wires carry num_ctx, so agent prompts are no longer cut on hosts with less than 24 GiB of VRAM (giantswarm/model-manager#157). Ollama picks its default context window from the host’s VRAM: 4,096 tokens below 24 GiB, which is every CPU-only host. It fills a longer prompt by dropping everything between the first tokens and the tail, and an agent turn loses its system prompt and tool schemas without an error. model-manager 1.0.8 writes spec.ollama.options.num_ctx on every Ollama ModelConfig it creates or re-wires: its chart value ollama.contextLength, default 32768, capped at the model’s own context length. Loads pre-warm at the same window. components.model-manager.versionRange is >=1.0.8 <2.0.0 and the BOM pins 1.0.8. The previous range already admitted 1.0.8; the floor keeps an installation from resolving an older release. Installations without an Ollama backend see no change.