hi! could you try invoking the model for each TODO one at a time? it may be a bit more difficult to get thematic consistency between items, but you'll definitely get that one-to-one correspondence between inputs and outputs.

point of comparison: i ran my iOS26 app all night running foundation models in a Task { }, no rate limits. Phone got warm but no persistent errors. (haven't tried macOS, visionOS though)

Specifically see docs/schema.md in the zip file, you want the section titled Schema with guided generation

hii! looks like this is explained in the schema.md document included in the adaptor toolkit, see adapter_training_toolkit_v0_1_0.zip in docs/schema.md, specifically the section called Schema with instructions. They talk about: considerations for the proper training.jsonl format for guided generation turning your data structure into JSON schema safety tuning (required, but formulaic) unfortunately it's not quite a tool calling example, but could you generate a well-formed data structure using guided generation that represents your tool call? You'd have to dispatch the actual call yourself in your business logic, but it's better than nothing... alternatively if you're brave, you could consider looking at the patterns generated for tool calling by messing with base_model.pt and mimic that format, the framework is conceptually fairly simple hope that helps? (another user)