What feature would you like to see?
Suggestion
Include the rawInput in the permission request, making the message self-contained so clients can potentially auto-answer without consulting state saved when handling earlier messages relating to the tool call.
Background
Tool execute permission requests include a human readable description of the command to run (content.text), trimmed to 50 characters, but not the actual command in a field of its own.
If the client wants examine the exact command in order to possibly auto-approve the request based on client side config, it needs to either parse the human readable description (which includes extra text and won't include the whole command if it is long) or remember what was in the content field of an earlier message with the same toolCallId.
Additional information
For my particular use case is a custom agent-shell-permission-responder-function for use with the the agent-shell emacs package. Looking up the content of a previous message by toolCallId cannot be done there due to a limitation in agent-shell.el. The toolCallId is not included in the arguments to the responder function. Fixing that in agent-shell would allow me to track the tool call requests and get the command string from an earlier message, but that would still be more complicated than just including the command in the permission request.
I have a patch for this, but It is 100% vibe coded and I don't know the code well enough to vouch for it beyond "seems to work on my machine", so I'm not opening a PR. If someone who knows what they are doing wants to use it for a PR, I won't object.
What feature would you like to see?
Suggestion
Include the rawInput in the permission request, making the message self-contained so clients can potentially auto-answer without consulting state saved when handling earlier messages relating to the tool call.
Background
Tool execute permission requests include a human readable description of the command to run (content.text), trimmed to 50 characters, but not the actual command in a field of its own.
If the client wants examine the exact command in order to possibly auto-approve the request based on client side config, it needs to either parse the human readable description (which includes extra text and won't include the whole command if it is long) or remember what was in the content field of an earlier message with the same toolCallId.
Additional information
For my particular use case is a custom agent-shell-permission-responder-function for use with the the agent-shell emacs package. Looking up the content of a previous message by toolCallId cannot be done there due to a limitation in agent-shell.el. The toolCallId is not included in the arguments to the responder function. Fixing that in agent-shell would allow me to track the tool call requests and get the command string from an earlier message, but that would still be more complicated than just including the command in the permission request.
I have a patch for this, but It is 100% vibe coded and I don't know the code well enough to vouch for it beyond "seems to work on my machine", so I'm not opening a PR. If someone who knows what they are doing wants to use it for a PR, I won't object.