Prevent roo from reading large files

[QuinRiva]

Auto-approve read works well in most instances, but in some instance it will encounter a very large file. As soon as roo attempts to read the file, the whole task needs to be discarded as it will flood the context window irrevocably.

Instead, roo should check the file size of non-code files (e.g. csv, json, etc.) before attempting to read them. Particularly in the case of csv and json - it can get all the info it needs by reading a sample.

[bertvanbrakel]

Could it delegate to a new subtask to read the thing, extract relevant info, then pass the summary/extracted data back?

With the new parent/child feature, maybe this is straightforward?

If over X then read using model-y, and extract z.

Maybe an MCP tool?

[QuinRiva]

Given the appropriate prompting it has been pretty successful at taking the appropriate actions. But yes I think a more robust approach is probably in order.

[patcon]

First off, I just want to say thanks for everything you do here as maintainers <3

With that said, I will admit that I'm writing this from a headspace very unbecoming of an open source product user, and I can catch the unhealthy "indignant" feeling that I despise seeing others arrive with. But I share that only to convey the emotional valence of this issue (as a quite broke human who works on many unpaid community projects, and just threw a dollar at absolutely nothing, yet again)

As context: I very regularly lose whole dollars to an action that should cost pennies. Just being able to set a read restriction of "do not read json or csv files", I can literally imagine this (at your scale of users) this would save your users (and the environment) huge sums of money and resources on compute. I keep looking for the setting and being baffled that the "allow read" doesn't offer any more granular control than a cost ceiling.

Much love to you all. Consider this simple feedback, with some emotional undercurrent :)