After I get a full answer to question A in the parent chat, I Fork the conversation and ask a new question B only in the forked chat.
The agent’s reply in the fork re-answers A (often a shorter regenerated version), then answers B.
The fork should already include context up to the fork point (including the assistant’s answer to A). Instead, the model behaves as if A was never answered.
This makes Fork Chat unreliable for branching into a side question without redoing the previous turn.
Steps to Reproduce
Open Agent chat (local).
Ask question A, e.g. “Give me a detailed comparison of Option 1 vs Option 2 for this project.”
Wait for a complete assistant reply to A.
Fork the chat.
In the forked chat only, ask question B — a different topic, e.g. “What do you think of this email? Should I reply?” and paste unrelated content.
Observe: the forked reply re-summarizes or re-answers A before answering B.
Expected Behavior
The fork keeps context up to the fork point (per Cursor docs: “all context up to the selected point”).
If I fork after the assistant has answered A, the first reply in the fork to question B should address B only (or briefly say “covered above”), not regenerate the full answer to A.
Hey, thanks for the detailed report and steps. This is a bug on our side, and your expectation is correct. A fork should keep all context up to the fork point, including the assistant’s last reply.
It only reproduces when you fork from the last turn using the button on that message. In that case, the previous reply gets dropped, and the model answers as if A never got a response.
Workaround for now: don’t fork via the action on the specific message. Use the /fork slash command in the input box, or use Fork Chat from the ⋯ menu in the chat list. Both options clone the full checkpoint and keep the last reply, so your new request B will be answered as just B.
We’ve logged the bug, but I don’t have an exact ETA for the fix yet. I’ll post an update in the thread as soon as I have one.
Your workaround (/fork / chat-list Fork) clones the current end of the chat. That doesn’t help when I want to branch from a specific earlier reply after the parent chat has moved on.
Is that use case possible right now, or not until the bug is fixed?
Good question, and the honest answer is that you can’t reliably do this right now.
Branching from a specific earlier reply is done via the fork button on that message. That’s exactly what’s currently bugged and loses context. The reliable workarounds (/fork and Fork Chat from the ⋯ menu) only clone the current end of the chat, so they don’t cover your case. You want to branch from an earlier point after the parent chat has already moved on.
So proper branching from any earlier point will work once the bug is fixed. I can’t share an ETA yet, but I’ll post an update in the thread as soon as I have one.
As a temporary workaround, if you specifically need to go back to an earlier point, roll back to the right checkpoint in the original chat and run /fork from there. The fork will copy the state from that point. It’s not as convenient as forking from the message, but it should keep context correctly.
The fix for the fork that was replying to the previous request is included in Cursor 3.14. Once you update to that version, please try again. Forking from a specific earlier reply should also correctly keep context after this update.
Until 3.14 is out, the previous workarounds still apply: use /fork or Fork Chat from the ⋯ menu to fork from the current end of the chat. For an earlier point, roll back to the right checkpoint in the original chat, then run /fork from there.
If, after updating, any scenario still doesn’t behave as expected, post here and I’ll take a look.
I’m on the most recent version of Cursor and encountered the same issue @gakiwo originally described. I forked a chat using the Fork Chat button that appears at the end of each reply. I clicked the button that corresponds with a reply earlier than the chat’s last reply, but after further prompting, the forked chat’s reply contained context from later replies that should not have been included in that fork.
That extra unwanted context also comes with other negative side effects: it increases the cost of the chat, and leads to the chat context being summarized sooner.
Hey @Leland_Hepworth, thanks for the details. What you’re describing is different from the original issue in this thread. In the original case, a fork from the latest reply would lose context. In your case, it’s the opposite. Forking from an earlier reply pulls in context from later messages. Also, you’re on 3.20.17 and using the IDE layout, so this is after that fix. This looks like a separate issue, so let’s treat it as its own case.
To narrow it down, can you share a couple things:
Which exact UI are you using to fork. The button at the end of a message, or Fork Chat from the ⋯ menu? I want to make sure we’re talking about the same action.
Short repro steps. How many turns were in the chat, which turn number did you fork from, and what exactly got pulled in from the later replies.
The Request ID for the first request in the forked chat. Chat menu in the top right > Copy Request ID. If you have Privacy Mode on, the Request ID will be limited, that’s fine. In that case, a short screen recording or screenshots are enough.
One more ask. Since this thread is tied to the original issue that’s already fixed, please start a new thread for your case. You can copy your post from above and add the details from the points above. That’ll make it easier to track separately, and it won’t get mixed up with the 3.14 fix. Drop the link here and I’ll take it from there.
When I went back to check the actual chat that had the issue, the forked chat now lists messages from the original chat that appeared after the forked chat. Other aspects of the chat are weird too; if I click a Thought dropdown it is now empty. Thoughts for the most recent reply in the forked chat do appear, but there are some earlier replies from prompts made in the forked chat and not in the original chat that also have Thought content that is missing.
@deanrie Does this behavior make sense given the update from version 3.20.17 to 3.20.21? Knowing that would give me some peace of mind regarding forking a chat at an earlier reply, without needing to perform cumbersome workarounds.
Regardless, I won’t put in the separate bug report unless I see the issue again on the most recent version. And as you can see in the simple test screenshots above, thoughts on new chats after updating are correctly appearing.
Hey, yeah, that behavior makes sense, and your plan is reasonable.
The fork context is locked in at the moment the fork is created. It takes a checkpoint at the point where you forked. The forks you made back on 3.20.17 already include that incorrect state with later messages pulled in. Updating the client doesn’t rebuild existing forks retroactively, so the old forked chat still shows those later messages. Forks you create now on 3.20.21 use the fixed checkpoint selection. That’s why your new banana test is clean. From your screenshots, the two forks are independent, each only keeps its own context (banana + cactus and banana + telephone), and thoughts render correctly.
The empty Thought dropdowns on earlier replies in that old chat are a separate issue. It’s about how already-saved data in that specific chat is displayed, not the fork-context fix. On new chats, thoughts show correctly, so it doesn’t affect new work.
So forking from an earlier reply on the current version should keep the context correctly, with no workarounds. I agree with your approach. There’s no point filing a separate report unless you see it again on a fresh fork on latest. If it does happen again, start a new thread with repro steps and the Request ID from the first request in the fork (top right of the chat > Copy Request ID), and I’ll take a look.