It shows what it makes.
Diagrams and pages, drawn live beside the chat.
What happens between the model and your machine.
ValidatePoliceApproveExecuteObserve
Is this a real tool, called the right way?
Does it stay inside your folder and your rules?
If it matters, Medha stops and asks you.
It runs in a sandbox, not loose on your machine.
The result goes into a record you can rewind.
the model proposes
edit_file({
path:"crates/store/src/pool.rs",
find:"Pool::new(1)",
replace:"Pool::new(4)"
})
edit_file({
path:
find:
replace:
})
edit_file is a known tool
its arguments match what it expects
passed on to the next check
a command that reaches outside your project
rm -rf ~/Documents
rm -rf ~/Documents
outside the project folder
matches a destructive pattern
refused before you are even asked
Medha wants to use excalidraw.create_viewDraw the text-to-SQL pipeline on a new canvas.
Allow onceAlways allowDeny
nothing runs until you choose
cargo test -p store
Seatbeltsandbox
this foldercan write
offnetwork
exit 0 in 1.2 s, 42 tests pass
a41fasked to edit pool.rskept
9c02checked and allowedkept
d510approved by you, oncekept
e2f942 tests passkept
each entry is sealed to the one before it
