Hack: remove line numbers from reads and diffs #1889
Conversation
|
|
Here is another problem with line numbers pointed out by @hannesrudolph: Actual content in the file:- GitHub Users page showing:
- List of GitHub usernames
- Link status
- Option to create new user or link to existing user
- Pagination for large datasetsread_file content result35 | - GitHub Users page showing:
36 | - List of GitHub usernames
37 | - Link status
38 | - Option to create new user or link to existing user
39 | - Pagination for large datasetsContent match that AI tries to replace in
|
63ddbee to
a660b24
Compare
* timeouts for individual servers * changeset * remove logger * use const and descriptive function for time settings
This reverts commit 87670a3.
a660b24 to
d6fe15b
Compare
this tool does not work without line numbers
This reverts commit 41c871d.
The insert_content tool was intended for appending content only, but the code allowed arbitrary line numbers which could corrupt files. This change enforces that line_number must be 0 (append-only mode) to prevent file corruption. Signed-off-by: Eric Wheeler <roo-code@z.ewheeler.org>
Add documentation showing how to use insert_content in conjunction with write_to_file to handle files that exceed output limits by breaking them into multiple append operations. Signed-off-by: Eric Wheeler <roo-code@z.ewheeler.org>
Add line count check to writeToFileTool to only require apply_diff for files that are more than 25 lines long. This allows write_to_file to work on small existing files while still protecting larger files that need more careful modification. Signed-off-by: Eric Wheeler <roo-code@z.ewheeler.org>
Roo read_file/apply_diff Hackfest
Today @hannesrudolph and I we are troubleshooting strange Claude hallucinations related to line numbers and other issues. While troubleshooting that, I hacked together these few commits.
Do not merge this! --- at least not yet. I am sharing this with the community for feedback and for 8000 others who may wish to test the behavior.
what does it do?
This branch:
apply_diffif a file exists (write_to_file is denied)Background
While there are places where line numbers are useful, line numbers in file reads have a few issues:
|symbol because it interferes with the line number separator and confuses the model (just try to edit a markdown table, I dare you...)123 |There are places where line numbers are useful, here are a few:
read_filestart/end rangesSo, I have built this into my daily driver and will report my findings and continue to refine this.
Long term goals