Tool permissions
The assistant can’t do more than the person using it. Tools they can’t use don’t exist.
The problem
The easy way to gate an assistant’s tools is inside each one: check the permission, return “you can’t do that.” But the model still sees the tool. It tries it, wastes a round, tries again with different arguments, and the assistant becomes a way to poke at a gate the UI already hides.
The same thing happens with identity. A request often carries two ideas of who’s calling: the verified one from the auth check, and a convenient header set somewhere upstream. Code that falls back from one to the other works fine for logs. It’s a hole when it decides whose data gets read.
What I concluded
Make the tool list be the permission. One function builds the tools for each request from the person’s permissions and what’s on screen. A tool the person can’t use is never built.
function toolsFor(request) {
return [
searchDocs, findRecord,
...(can(request.user, "reports:read") ? [auditReport] : []),
...(request.screen.filters ? [filterTool(request.screen.filters)] : []),
];
}
That one function feeds every place tools are listed: the model’s turn, the manifest the client reads, and the debug panel. They can’t disagree.
Rules
- Absent, not refused. The model never sees a tool it can’t use, so it never wastes a turn on a wall.
- Verified identity only, for anything personal. Logs can use whatever identity is handy. Storing a conversation, writing a profile or acting with someone’s connection uses the verified user or nothing.
- Still check inside. A tool that exists because something was on screen checks again when it runs, because the screen can change between request and call.
- Versions are access too. A new assistant ships as a whole new version on an access list, not as flags scattered through the old one. A request that names a version runs exactly that one, and debug is extra visibility on top of a version, never a different assistant.
The trade-off
The model can’t say “you’d need admin for that,” because it doesn’t know the tool exists. I think that’s the right trade. The UI already explains permissions, and the assistant shouldn’t be a second way to probe them.
What you get
Nothing is offered that will be refused, and the assistant can never do more than the person in front of it. Adding a gate is one condition in one function, and every surface that lists tools picks it up.
Part of Agent as broker.