feat: #99 workflow knowledge

[pkarw]

Well, I implemented it - #99 - I guess it MIGHT do a good job limiting dispatching offspring (by taking the agent too many context planning tasks) - however, it's pretty damn hard to find it out at the moment. To be perfectly honest, I don't really like it, so it's up to you, @grabbou, if you decide to merge it or not.

I mean - I believe in a better quality change in dispatching only after refactoring and A/B kind of testing different dispatchers on all example probes. Right now, I guess.

I'm not sure about the naming: knowledge - as in the future - knowledge maybe should be something more related to the gathered shared context between the agents. I mean, in the end, it's the same meaning, a different way of how it will work.

[pkarw]

Note that the backward compatibility mode makes it not compliant with the kind of Liskov Substitution Principle (LSP) and, therefore, more difficult to debug - especially for newcomers.

So if merging in, maybe we should disable backward compatibility and always add only the background to the agents even when it's empty. @grabbou

The other thing is that it's pretty essential to pass this information - till now embedded into description like, for example, file names - otherwise, agents started hallucinating during the tests - inventing file names that don't exist, etc.

[grabbou]

I proposed an alternative approach in #108, in case we wanted to go with knowledge field. I will let you decide which approach should we go with.

[grabbou]

Going to close this one in favor of #108 as per discussion.