-
Notifications
You must be signed in to change notification settings - Fork 2.1k
feat: Rich Approval Mechanism for Human in the loop Tool Approvals #711
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
4bcc449 to
c4305eb
Compare
This PR is now significantly updated from its initial version - we are using it day to day with our UI and it is working well. I have simplified the design, but it still requires changes to the core logic - implementing it using only callbacks would be challenging as it relies on logic in the content preprocessor before the before_model_callback and it would be very hard to replicate this entirely after that preprocessor has run.
4afd578 to
bc77db3
Compare
@calvingiles thanks for your contribution. We will evaluate it in the upcoming weeks.
e272c56 to
d0d1865
Compare
We have released a new plugin framework. Could you try if this feature can be implemented as a plugin? 162228d
Uh oh!
There was an error while loading. Please reload this page.
This is an initial proposal seeking feedback adds an Approval Mechanism.
This is a fully working implementation that we have begun using in our internal agents.
This mechanism allows tools to be annotated like:
when you then try to use the agent, you will receive an approval request in response with the challenges that need to be granted in order to call the tool:
{ "id": "adk-38e8d3c6-a14c-4673-b4e8-cc2e2908d64a", "args": { "grantee": { "id": "tool:list_directory:adk-6b26eed9-2fa7-496e-ba8f-9d75362d0e0a", "type": "tool", "on_behalf_of": { "id": "3a0400f2-d9d9-4909-8ec3-41f8f28c76fe", "type": "agent", "on_behalf_of": { "id": "calvingiles", "type": "user" } } }, "challenges": [ { "actions": [ "tool:local_file:read" ], "resources": [ "tool:local_file:/Users/calvingiles/eng/repos/adk-python" ] } ], "function_call": { "id": "adk-6b26eed9-2fa7-496e-ba8f-9d75362d0e0a", "args": { "dir_path": "/Users/calvingiles/eng/repos/adk-python" }, "name": "list_directory" } }, "name": "adk_request_approval" }This can then be granted by sending a message like:
{ "name": "adk_request_approval", "response": { "grants": [ { "effect": "allow", "actions": [ "tool:local_file:read" ], "grantee": { "id": "tool:list_directory:*", "type": "tool", "on_behalf_of": { "id": "3a0400f2-d9d9-4909-8ec3-41f8f28c76fe", "type": "agent", "on_behalf_of": { "id": "calvingiles", "type": "user" } } }, "grantor": { "id": "calvingiles", "type": "user" }, "resources": [ "tool:local_file:/Users/calvingiles/eng/repos/adk-python" ] } ] } }Where the grant can include wildcards so that )for example, in this case) any function call that matches these actions and resources called by this agent will be approved in the future.
This is working in our internal service, but I am interested in whether this would be suitable to include in adk directly.
Relates to the issue #640