Default to full file reads

[mrubens]

Anecdotally Roo seems dumber when it doesn't read the whole file. I think full file reads are a safer default for now.

[user202729]

This PR explicitly makes the tool not do what it tells the AI to do.

With this change, by default, even if the AI provides <start_line> and <end_line>, the whole file will be provided. So that there is no way for the AI to actually read a line range.

There are a few problems with this:

• With very large file, this can lead to overflow of context window after every file read and make the AI stuck in a loop.
• Worse, a single file read may overflow the context window.

Or any other rationale that made #1440 added in the first place.

[mrubens]

This PR explicitly makes the tool not do what it tells the AI to do.

With this change, by default, even if the AI provides <start_line> and <end_line>, the whole file will be provided. So that there is no way for the AI to actually read a line range.

There are a few problems with this:

• With very large file, this can lead to overflow of context window after every file read and make the AI stuck in a loop.
• Worse, a single file read may overflow the context window.

Or any other rationale that made #1440 added in the first place.

Yeah that’s a good point - we should probably make the prompts smarter. And you can always change your settings if you prefer the partial reads, but I and others noticed that the AI seems to make a lot more mistakes when using partial reads so we decided to change the default.

[user202729]

Yes, but the implementation leads to misleading output to the user.

A better way to implement it may be if maxReadFileLine option is -1, then don't tell the AI that <start_line> and <end_line> exists at all (and just act like it doesn't exist). Otherwise it would leads to a confusing situation where the AI says "I will read line 1000--1100 of the file", but actually reads all of it.

This probably involves some refactoring — first getReadFileDescription function needs to be changed, but also we need to pass maxReadFileLine to it somehow. (It isn't currently passed.)

[mrubens]

I agree, that’s a great point. We should fix.

[mrubens]

#3995

[user202729]

in order to test out Roo-Code I try to replicate fixing the issue pointed out above, and immediately see an issue when full file read is not enabled (GPT 4.1 mini):

The cause is because it only reads line 1–40 previously.

That said, the reaction makes sense — it doesn't know the surrounding context yet, so its best guess is the user forget the closing quote there.

It might helps here when, in partial read_file calls, adds e.g. a [10 more lines...] at the end of the returned result. So that the model would know it's not yet the end of the file.

Another idea would be in replace_in_file tool call, disallow the model from changing parts that it has not read (more precisely, if the model want to change line X → Y, it must have read line max(1, X-5) → min(N, Y+5) where N is the total number of lines). This can be checked with 100% certainty without the use of AI, just keep track of what tool calls has been performed.

[also lesson learnt: don't use Roo-Code in development mode to develop itself, otherwise you will get

♻️ File changed: /home/user202729/Roo-Code/src/core/prompts/tools/read-file.ts. Reloading host…

and the extension being killed. I think it's worth adding this as a note to Local Setup & Development
section in the README. ]