Decisions
Why agent actions use one contract across command, tool and route surfaces.
Actions
Why an agent's commands, tools, routes and CLI come from one registry of namespaced actions, and how the registry is shaped.
Why is there one kind of action?
An action is a typed request from a caller, answered outside the
caller's walls. A part that relays a protocol it does not type, such as
MCP or inference, is the same action with carries in place of
answer, and one of the two is required. A scheduled run, a webhook and
a connection's message are callers, not actions: the caller is Trigger,
which acts under the grant it was set up with, since it acts later with
no owner there.
What it costs
One registry, one way of proving a caller and one table for the proxy
replace a separate kind with its own plumbing for each. The server
branches once, between a typed answer and relayed bytes. Every place
typing cannot hold is one search for carries.
Why are the names two parts?
A name is <namespace>:<action> as a command, <namespace>__<action>
as a tool and /_link/<namespace>/<action> as a route, so one action is
found by the same words everywhere. Another harness's familiar spellings
are aliases that resolve to an action and add no handler or permission.
Why these namespaces?
Each holds one subject and is named in the singular. The ways a message
goes into a conversation, such as goal, are chat's
rather than namespaces of their own. memory stays apart from chat
because it outlives conversations and is a store of its own, and
browser and desktop are separate from web so that two-part names
can hold each.