Build Out Parameters for Char Limit Based on Model

[jevers9788]

Character limit generation count is hard coded for gemini 2.0 flash at 40k, which is beyond the context limit
of many other models. Lets allow users to configure this and set to the limit on the model being used if user doesn't configure

[nickcdryan]

Noted. The output token limit became the real issue for generating the memory in learnings.txt where eventually it would just get truncated, so there's this soft warning that if you're approaching the limit and should add new memory but consolidate old memory such that nothing is lost. It'll probably be needed elsewhere at some point, maybe the input limit

Maybe this gets added to the LLMClient config per model so it's handled automatically for each model we support? (unless the user wants to set it lower to control cost)

The rough math (for gemini tokenizer, but also roughly the same with others) is:

num_tokens * 5 = num_characters

so

max_output_tokens * 5 = max_output_characters

Gemini: 8192 * 5 = ~40,000

So if we add the output token limit to the config for each model we can just set the max_output_characters, plus some headroom.