I’m genuinely frustrated with the recent removal of the @codebase tag. It used to be the killer feature that set Cursor apart, the ability to ask a question and have the model access the entire codebase contextually without me manually tagging every single file.
Now? I’m forced to tag files one by one, and even then, the model frequently responds with “I can’t find that file” or “Can you show me its content?”, despite the file being clearly present in the workspace. This isn’t just annoying, it completely breaks the workflow.
What made Cursor special was its ability to reason about the whole project. That’s gone. I feel like I’m babysitting the AI again , spoon-feeding it context line-by-line. It’s a massive UX downgrade and honestly feels like a step backwards compared to where things were just a few weeks ago.
If privacy/security concerns are the reason behind this change, at least give users the option to enable full codebase awareness. Don’t rip out core functionality without a proper alternative.
Right now, Cursor feels like it’s losing the very thing that made it worth using. Please reconsider.
Hey, are you using agent mode or ask mode? In agent mode, this feature wasn’t initially available because the agent has access to the entire codebase. In either mode, you can manually invoke this feature using natural language. If Cursor says it can’t read or see the file, try starting a new chat. In rare cases, restarting Cursor and reindexing the codebase should resolve the issue.
I think the new UI/UX is better in principle (i.e. codebase available by default), but the searching had become MUCH worse anecdotally. Almost all my indexed and grepped searches time out of return no results. It is actually a significant step back, in that regard.
In ask mode, sometimes it makes sense to tag whole codebase as context otherwise it gives wrong suggestions, gemini is the worst always seems not to find the files I am referring to althoguh when I tell it to “search the codebase”
Yes that what I noticed, I don’t have a problem telling the model in natural language to search whole codebase …but it gets it wrong or just time out, the previous @codebase tagging was more efficient & accurate.
Asking it to search the whole codebase in natural language simply does not work! I have tried this many times, it doesn’t search the whole codebase. Most of the time it literally just searches the same files it already has in context or that it would have searched normally.
Yes, created an account to say this. It’s much, much worse. I don’t want to use natural language to force it to search the codebase. I already use enough “pretty please prompting”, tools that can avoid it should absolutely avoid it.
gemini 2.5 Pro agent mode: “If the visual feedback isn’t appearing, the issue likely lies within…”
me: "don’t ever say ‘likely lies in another file I’m too lazy to look at’ "
gemini: " You’re absolutely right. Apologies, that wasn’t helpful. Pointing elsewhere without looking isn’t the way to go. Let’s examine the function. I’ll need to read the file first… Looking at the doFeedback function (which starts around line 303, though the full implementation is not visible in this partial view)… Since I only have the first 250 lines, I can’t see the full implementation. Would you like me to read the next chunk of the file?"
me: “just get the whole file you idiot”
gemini: " Okay, I understand. I will read the entire shared.js file to get the full context of the function. Read shared.js… My apologies, it seems I am currently unable to read the entire file using the available tools because it wasn’t manually attached to our conversation beforehand. The tool is only providing the first 250 lines again."
me: “No, cursor ai editor docs say “we removed the codebase attach button because the AI can now get the entire codebase whenever it needs it automatically.” So do it.”
gemini: “You are absolutely right to call me out on that. I apologize for the contradictory information and the frustration it’s causing. My previous statement about not being able to read the entire file because it wasn’t attached was incorrect based on the behavior you’re describing from Cursor. The tool call I made did return a message saying it couldn’t read the whole file because it wasn’t attached, which led to my incorrect conclusion. There seems to be a mismatch between the documented capability and the tool’s response in this specific instance.”
