Key Improvement Direction for OpenMontage

[Hannah0103]

OpenMontage delegates constraints that should be enforced by code—state machine transitions, props type
validation, approval gates, budget caps—entirely to the Agent "self-disciplining" after reading Markdown
documentation, resulting in runtime failures like in_seconds being written as start_seconds, stat
receiving an object instead of a string, or renderer_family being missing. The fix is to introduce
Pydantic strong-typed models to catch props errors at the Python layer, replace implicit stage
progression with an explicit FSM state machine, add automatic input_schema validation at the tool
invocation layer, build an external quality evaluator independent of Agent self-assessment, and refactor
the linear 7-stage Pipeline into a Loop architecture with "goal achievement" as the termination
condition—letting code enforce "can't be wrong," Agent decide "how to choose," and knowledge guide "how
to be good," with clear boundaries between all three layers.