Operation names are namespaced. The dossier that follows recommended the opposite — flat names, with the decision spent instead on recording who declared what. The owner chose namespaces. That is the answer, and the rest of this page is now the record of how the question was argued rather than an open question.
The shape, settled so that nothing about it has to be asked again:
namespace:name, with at most one colon. Two or more is
refused.:orders, orders: — is refused.-, _ and ., and
nothing else. The 128-byte limit applies to the whole name, separator included: a public limit does
not get quietly raised to make room for a new feature.
This is the part that turned a preference into a deadline. internal/store/collision_test.go
holds TestAnOperationNameMayHoldAnySeparatorAName, which measures that
every separator a convention might pick is already a legal operation name —
orders/recent, @acme/orders.recent and acme:orders.recent all
declare successfully today. COMPATIBILITY.md section 3 forbids tightening a rule inside
1.x.
Put those two together: a 1.1 that refused a:b:c, or that reinterpreted the colon in a
name somebody had already declared, would be breaking a promise 1.0.0 made. A namespace that
misses the tag can never be added at all. That test was written to record a curiosity about
how loose the name rule was. It turned out to be the clock.
Decide whether operation names are namespaced, and write the decision down with its reasoning. There are two answers and this ticket is done when one of them is chosen: namespaces exist, or names are flat forever and the collision behaviour is documented as intended. The decision has now been made; the ticket stays open until the behaviour it chose is in the code and under test.
Naming is the flat-out worst thing to leave undecided across a 1.0 boundary. Adding namespaces later changes what an existing name means, and every operation anybody declared is addressed by one. Deciding "no namespaces" later is equally impossible if somebody has already been given a way to write a qualified name. Both doors close at the tag; this ticket picks which one closes.
Measured 2026-09-20. There is no namespace. Searching every.gofile fornamespacefinds two hits, both insapedb_test.go(lines 32 and 34) and both about Go's own package scope, not about operation names. The collision behaviour is ininternal/store/ops.go:367,DeclareOperation:operation.Version = 1, thenif found { operation.Version = latest.Version + 1 }. Declaring a name that already exists does not fail and does not warn — it creates version N+1 of the existing name, keeps the older versions, and returns success. Two unrelated authors choosingorders.listproduce one operation with two versions, and the second author's declaration silently becomes the oneInvokereaches, because Invoke takes the newest.
The sharpest form of that measurement, from the dossier: the first author's application did not change
one line. It called the name it was given, with the argument it was given, and deleted a
document — because the second declaration of that name was a delete where the
first had been a get. Nothing about a redeclaration has to resemble what it declares over:
not the action, not the collection, not the declared scopes.
pipelines/tasks/0071 adds a reason this decision cannot wait for the external-operations
group to land first: the recommended v1 shape for an external operation is an OperationBundle
— one or more store.Operations declared under a signature — and every operation inside a
bundle is still just a name in the same flat table DeclareOperation already writes to. A
stranger's bundle and a name already declared on the target server collide exactly the way two
unrelated authors do in the blockquote above, silently, by this ticket's own measurement — there is
nothing about "external" that changes the collision rule. Whichever way this ticket decides, that
decision is what an OperationBundle author has to be told before they publish one.
DeclareOperation — not by comparing name strings.InvokeVersion addresses older versions. That works and is shipped; it
simply takes the full namespace:name now.