Features
The 22 features the admin package declares, in two tiers: project.* inside a project, platform.* across projects.
Every admin route names the feature it requires, and the server checks it against the caller's real grants before the handler runs (the mechanism is on package-system Features). The Admin widget hides a tab the caller cannot use, but hiding is presentation; the server check is the gate. The routes each feature reaches are listed on Routes.
Project tier
A project.* feature acts inside the caller's current project.
| Feature | Grants |
|---|---|
project.dashboard | The Admin widget itself, dashboard stats, usage charts, health, package and stream lists, and the projects the caller belongs to |
project.members | This project's member and role inventory, and the exact-email lookup for adding an existing user |
project.members.invite | Inviting a new user into this project and issuing, listing and revoking one-time invite links. The role's invite flag is also required, and the assigned role may not be stronger than the caller's |
project.agents.assign | Assigning or unassigning a project agent to a member — who may manage it; who may use it is the Agents tab's attach. The role's role-management flag is also required |
project.audit | The project-scoped logs: a project-installed package's log and the per-agent run logs |
project.roles | Reading this project's role map. Rewriting it also needs the role-management flag, which is a role property, not a grantable feature |
project.sources | This project's source configurations, with live health, and their permission and sync edits |
project.limits | This project's spend-limit rules, including other members' per-user caps. Your own cap is always visible; editing limits needs the role-management flag |
project.credentials | The credential catalog at your own scope: ids and whether a value is set, never values |
project.credentials.write | Setting or deleting a credential value at your own scope |
project.canvas | The organization canvas and its per-member and per-agent detail |
packages.manage | The project's packages (trust, build, the feature that gates one, withheld tools), cache clear and rescan. The host re-checks the live package-management grant on every call |
Platform tier
A platform.* feature crosses the project boundary. None of them is
granted to any role by default; an owner grants them explicitly.
| Feature | Grants |
|---|---|
platform.projects | Projects you are not a member of: the cross-project list, platform-scope usage and the per-project grouping |
platform.users | Platform-wide user records and cross-user spend, acting on a user in no project, deleting a user (with authority over them in every project), and the platform Devices tab |
platform.scope | Acting on a scope outside your session: another project, member or agent, or the platform-wide credential scope, and only on principals inside your own projects |
platform.audit | Every platform-global log: the audit log, the route log and the native packages' own logs. None of them carries a project id |
platform.config | Platform settings, the custom endpoint lists, the host CLI import, and the release line with its notification |
platform.vector | Vector store status, collections and the active embedding model |
platform.vector.reset | The destructive, platform-wide vector reset |
platform.packages | Every package the host knows, and asking for one to be installed, updated or removed while the host runs |
platform.packages.approve | Approving a package request and restarting the host to apply it. What it installs runs inside the host on every project, so approving is held apart from asking |
platform.projects.create | Reaching project creation. Creating a project also needs owner strength (priority 1 in an existing project), so this feature alone is never enough |
Default grants
| Role | Default admin grants |
|---|---|
owner | every feature |
admin | every project-tier feature above |
manager, member, viewer | project.dashboard |
Two features named by admin routes belong to other packages: core.agents
(agent-core, granted to every role) gates the Agents tab's delete, whose real
authority stays agent-core's per-agent check; platform.vector.maintain
(brain-core, no default grant) gates the embedding-model rebuild.
Escalation inside one route
Some routes serve a project answer to a baseline caller and a wider answer to a caller who holds the matching platform feature:
- the usage series needs
platform.projectsfor the platform scope andplatform.usersfor per-user grouping, and the Codex savings card at platform scope needs both; - dashboard stats fill the platform user and project counts only for holders
of
platform.usersandplatform.projects, and read 0 otherwise; - a project you do not belong to answers
404unless you holdplatform.projects; - in a project record, other members' emails, other roles' grant lists and
other members' spend caps each appear only for holders of
project.members,project.rolesandproject.limits; - a log source rooted in the platform zone needs
platform.auditand is not even listed without it.
A missing escalation feature drops the field or answers 403; it never widens
the answer.