feat: permission before tool call #1313
Conversation
| } | ||
| } | ||
| }, | ||
| "chat" => { |
There was a problem hiding this comment.
I think instead of skipping tool calls in the loop, we really want to not advertise the tools to the agent at all? e.g. remove them from the call to provider.complete - otherwise i could see this leading to confusing dialogue
There was a problem hiding this comment.
Subtle detail, but the ToolConfirmationRequest is not added to the agent loop and is not visible to the agent. Added some context in the chat mode to allow the agent to frame the tool call requests as plans.
|
|
||
| // Wait for confirmation response through the channel | ||
| let mut rx = self.confirmation_rx.lock().await; | ||
| if let Some((req_id, confirmed)) = rx.recv().await { |
There was a problem hiding this comment.
do we need to add any waiting deadline?
There was a problem hiding this comment.
for cli, we shouldn't add a timeout imo, but for GUI, it should be considered
| // Wait for confirmation response through the channel | ||
| let mut rx = self.confirmation_rx.lock().await; | ||
| if let Some((req_id, confirmed)) = rx.recv().await { | ||
| if req_id == request.id { |
There was a problem hiding this comment.
curious, will we have the case that req_id != request.id
There was a problem hiding this comment.
no, the identical message id should be passed back.
Internal testing instructions:
pull the branch
use env var
use config
# add GOOSE_MODE=approve in ~/.config/goose/config.yaml # auto, approve,chat ./target/debug/goose session