Handler responses may be redundant

[JRegimbal]

With LLMs going into preprocessors, handlers are starting to use their output near-directly (e.g., display with context, put through TTS). This may pose problems to users in the future. The conceptual model for IMAGE assumes that handlers will represent preprocessor data in different ways, which is not the case with LLM text. Users then run the risk of receiving multiple representations from handlers that include the same information from the LLM over and over.

We should find some workable solution to this, either in terms of server management or making handlers/the orchestrator "smarter" so redundant info is less likely to be displayed.

[jeffbl]

Note discussion in #939, although the proposal there is a hack that breaks the existing IMAGE model.