The harness picker — pluggable operators, live
Switch provider, model, and reasoning effort from the run header; every call is stamped with the operator that answered it.
kitsoki treats the LLM as a swappable operator. The run header carries a provider / model / effort picker — claude-native, codex, and synthetic backends — that takes effect on the next turn (substrate semantics, not a mid-call swap). Each agent call is stamped with the exact profile, model, and effort that produced it, shown inline on the trace row. The demo asks each operator to identify itself and watches the per-call provenance change as the picker switches — same recorded input, different operator, auditable output.
Which product behavior does this surface prove?
The first few beats show the capability in the real Kitsoki UI.
The page is generated from the same catalog and deterministic fixtures used by QA.