Seemingly ineffective tool calls cause high token usage

[maliayas]

In my experience so far, Crush can make multiple "view" requests for the same file like once before edit, once after edit (to see if it's edit worked :/). Also it tries to read partial content via "offset" parameter however that may cause not-so-smart edits due to limited context and later trying to read more parts of the same file to fix what it has caused.

Also when it comes to "edit" tool, I've noticed that AI offers edits one by one, each one causing another request that just wastes tokens. I tried to force it to offer all edits at once by editing CRUSH.md but no luck. I've checked the system prompt of Crush and saw this:

• IMPORTANT: All tools are executed in parallel when multiple tool calls are sent in a single message. Only send multiple tool calls when they are safe to run in parallel (no dependencies between them).

Maybe this is the reason? Then however, description of the "edit" tool tries to overwrite that but with no luck:

Remember: when making multiple file edits in a row to the same file, you should prefer to send all edits in a single message with multiple calls to this tool, rather than multiple messages with a single call each.

I know from my Aider experience that AI is pretty much capable to do many edits at once successfully in a single request. That saves a lot of tokens.

To give a general clue: in one of my tests, I tried to do the same job with Aider and Crush. For Aider it costed $0.47, for Crush it costed $1.88 - results were almost similar while Aider being slightly better because I've asked for a revision as well (included in the price) while with Crush I've left it alone and didn't ask any revision.

What do you think and how is your experience? How can Crush be improved in terms of token effectivity?