Workspaces
A workspace is the tenancy boundary. It scopes Actors, certificates, model access, confinement templates, and — since they can be set per workspace — trust policy and the kill switch.
Membership is not a membership table: it is
certificates scoped to workspace:<id>.
The tabs
| Tab | What it holds |
|---|---|
| Overview | Counts, and the effective trust policy — read-only |
| Actors | Assign or remove Actors |
| Access | Humans holding a workspace:<id> grant, with a deep link to issue one |
| Confinement | The confinement templates attached to this workspace |
| Settings | Every mutation: kill switch, trust policy, identity, delete |
| Budgets & Model Access | Not built yet |
Overview shows the state; Settings changes it. That split is why the Overview trust-policy card is read-only and links one tab across rather than duplicating the form.
Per-workspace trust policy
Both trust policy knobs — fail mode and staple TTL — have an org-wide default under Settings, and a workspace may override either of them.
Inheritance is per field. A workspace can pin its own fail mode and keep following the org on staleness. Each field on the Overview tab carries an inherited or overridden badge saying which level the effective value came from.
Leaving a field empty means inherit. A stored 0 staple TTL is a real value, and
it is the strictest one — "no cached status is acceptable". Do not clear a
field expecting to loosen it; clearing it hands the decision back to the org-wide
setting, which may itself be 0.
The most specific scope wins here, which is the opposite of the kill switch, where an armed global switch short-circuits every workspace. Trust policy is configuration; a kill switch is an emergency brake.
Saving re-pushes immediately. Changing a workspace's policy pushes new configuration to that workspace's connected Actors; changing the org-wide default pushes to all of them. Offline Actors pick it up on reconnect. A recipient never learns the word "workspace" — it receives one already-resolved pair.
Confinement templates
A template is a reusable set of tier-B confinement settings — the sandbox settings a supervised harness launches under. Manage the templates themselves under Certificates → Confinement templates; attach them to a workspace from its Confinement tab.
- A workspace can have several templates attached.
- At most one is marked Default, and that is the one pre-selected on the issuance form for an Actor in that workspace.
- On the issuance form, a workspace's templates are sorted first, never filtered — every template stays selectable, because an admin may legitimately want one from elsewhere.
It only pre-fills the issuance form. The signed certificate records what the admin actually submitted, which may differ from the template, and editing a template afterwards changes no certificate already issued.
Deleting a workspace
The default workspace cannot be deleted. Deleting any other one revokes the certificates scoped to it first, then removes the workspace; its Actors survive, simply unassigned. Nothing about the Actors themselves is deleted — "delete this workspace" must not quietly mean "delete the agents in it".
Deletion matches scoped certificates narrowly — the certificate's own
workspace and a workspace:<id> scope, but not an Actor's mere workspace
assignment. A kill switch deliberately matches more widely.
The asymmetry is the point: one is reversible, the other revokes for good.