-
Notifications
You must be signed in to change notification settings - Fork 3.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Bug]: CodeAct repeatedly edits the wrong file #1890
Comments
I also had the same wonder when I saw SWE agent commands. Apparently LLM often forgets that it needs to open a file before edits it. |
Yep! I think we should plan for one major iteration of the existing file operations now that we have a good way of comparing the performance. |
I vote for including the file name as the first parameter supplied by the AI to the edit command, as suggested above by rbren. (This mix-up happened a bunch of times on the very first day that I used OpenDevin.) |
I'd take this up since I'm already on the agentskills with append_file and get to the edit_file changes for this PR afterwards. cc @rbren @xingyaoww |
I encountered something funny with this symptom recently. The agent was trying to edit a javascript file but it forgot to open it. Since the last file it opened was an HTML file, the agent simply wrote javascript code to HTML file. When it browsed the webpage, it saw those javascript code printed out - and it thought that meant the code had been working! |
Proposal: right now |
+1 |
@tobitege i've been thinking about the same thing too -- feel free to starts a PR and add "filepath" to the edit command! |
Is there an existing issue for the same bug?
Describe the bug
I'm asking CodeAct to repeatedly switch between editing two related files. Often File A is open, then CodeAct sends an
edit
command intended for File B. This usually happens when I issue a new instruction.It'd be neat to show/emphasize the currently open file in CodeAct's prompt.
Side note: maybe we should just stop making
open
a thing, and put filenames directly into the commands? The agent has to look back over previous commands to determine current state, instead of just having a deterministic instruction it can write.Current Version
Main, as of `0ecba83e53c6f2abfb0fdbaf134842a0298234f6`
Installation and Configuration
Model and Agent
gpt-4-turbo, CodeAct
Reproduction Steps
repeatedly prompt codeact to edit different files
Logs, Errors, Screenshots, and Additional Context
The text was updated successfully, but these errors were encountered: