Add attenu-guard integration page - #585
Conversation
attenu-guard gives each Haystack sub-agent only the authority its task needs, checked on every Tool.invoke / AgentTool call, with a hash-chained offline-verifiable audit log.
|
@rafaelasor is attempting to deploy a commit to the deepset Team on Vercel. A member of the Team first needs to authorize it. |
|
The Vercel preview is waiting on a team member to authorize the deploy, so the red check is not about the page itself. Can someone on the team approve it, or is the preview not needed for an integration page? |
|
@kacperlukawski Ignore my note about the Vercel check. #584, #557, #547 and #567 all merged with it red, so it gates the fork preview and not the merge. On need: a Haystack sub-agent reached through I can add |
Adds
integrations/attenu-guard.mdfor attenu-guard, an Apache-2.0 librarythat gives each Haystack
Agent/AgentToolsub-agent only the authorityits task needs and checks every tool call against it before the tool body
runs. It hooks
Tool.invoke/invoke_asyncand thebefore_toolConfirmationStrategyprotocol, using only public Haystack extensionpoints.
The usage snippet is a self-contained, offline scripted-model example: a
coordinator delegates to a sub-agent with a narrower authority, the
sub-agent's model tries to export data it was never granted, and the
export is denied before the tool body runs — verified runnable against
haystack-ai 3.1.0 and attenu-guard 0.8.0.
Disclosure: I maintain attenu-guard.