with GPT-5.4 if I switch to ask mode form agent mode and then back. the agent still think it’s in ask mode and I have to specifically mention its in agent mode for it to start coding.
Steps to Reproduce
use GPT-5.4 and switch from agent to ask mode ask some questions and then switch back to agent mode. the agent will not update any code, sometimes even says its unable to update code in ask mode
Hey, this is a known issue. The mode state can get stuck if you switch between Ask and Agent in the same chat session. It’s not specific to GPT-5.4. It happens with other models too.
A workaround that works reliably is to open a new chat CMD + I and choose Agent mode from the start instead of switching modes in the same chat.
If you want to keep the context from your Ask session, you can also tell the model directly in your request: “You are in Agent mode. Edit the files directly.” Users say this usually unsticks it.
Seems that I’m burning token because, when in “Agent Mode”, the agent tells me he is read-only and doesn’t write changes.
When I tell him it’s Agent Mode, he says “no I’m in Ask Mode”.
Switching to Ask mode then Agent mode back doesn’t fix the issue.
Steps to Reproduce
Don’t exactly know how I entered this state. Maybe by asking a question in “Agent Mode”.
But now, when I’m in agent mode, cursor answers as if I was in ask mode.
“I’m in read-only mode here, so you’ll need to paste this into”
When I tell “we’re in agent mode”:
“You’re seeing conflicting messages, but the current system state is Ask mode, which I’m required to obey even if you say “agent” — so I can’t directly modify files, only show you exactly what to paste.”
Had to create a new chat for it to be able to work in Agent Mode again, which made me lost my context.
I mainly use two modes in the chat i.e. Agent and Ask
Ask to ask anything related to codebase or discuss stuff before making any changes to the codebase.
Agent to make subsequent changes decided in ask mode.
Problem: Switching between the two modes within the same chat can sometimes be buggy, if i go from ask to agent, sometimes the model still thinks i am in ask mode and doesnt make changes in code. I can’t use different chats because i use ask mode to build up context.
Steps to Reproduce
Open a chat in any codebase. Try using ask first, then switch to agent mode and see if it works. Somtimes it will other times it will remain in ask mode.
Saw this issue today. The first time I told the agent it was NOT in fact in ask mode it complied. Later in the chat it would not and kept parroting that its system instruction took higher priority.
The chat agent thinks that it is in Ask mode when it actually is in Agent mode (it also often does the opposite where it tries to edit files and fails while in Ask mode but that’s a separate issue). This happens to me I would say every few days and it’s very aggravating.
I can tell it it’s in Agent mode and sometimes this will work but much of the time it will argue with me, maintaining that it’s in Ask mode when its not. When that happens the only workaround I’ve found is to actually put it in Ask mode, ask it what mode its in, then switch it to Agent mode and ask it what mode its in, but this is annoying and all of it wastes my prompts.
I had the agent actively working on changes, when all of a sudden it stopped and “I’m currently constrained by the environment being Ask mode right now—so I can’t continue applying code” …. it took about 4 messages back and forth where I told it repeatedly if WAS in agent mode, but it just kept telling me that it was “constrained by an active Ask-mode guard from the system”. Eventually I had to switch it to ask mode, tell it I was doing so, then switch it to agent mode and finally it worked and was able to proceed.
This then happened again in a totally separate chat thread. Even when I say “You are in agent mode now” the agent doesn’t think so … it says “In this thread, I’m still constrained by an active Ask-mode guard from the system”.
