The root stack resource is not a change, unless it is the only work a deploy has

Decision record 0079

Amends 0007 (the Pulumi table) and 0036 (which dropped the root stack step only when it was same).

Pulumi makes one resource of its own for every stack, pulumi:pulumi:Stack, the parent of every resource of the program. The Pulumi adapter dropped its step only when it was same (0036). On a stack that was never deployed the step is create, so a program with two resources read as “3 creates”, and the row and the preview page listed create pulumi:pulumi:Stack <name> (issue 165). That is the first scan a new user sees, and the first thing they check is the number of resources they wrote.

The root stack resource is the stack coming into being, not a resource anyone wrote. So the adapter treats it like this, by the step op:

Step op of the root stack resource	Becomes
same	dropped, as before (0036)
create	dropped, unless the diff would be empty without it. Then it is the diff’s one change
update	dropped. The resource holds nothing but the stack’s outputs, and a change to outputs alone is not shown (0036)
delete, replace	kept, as a destroy. No recording holds either one, and a destroy warning too many is the safe side (0007)

The root stack resource is the resource whose URN names the type pulumi:pulumi:Stack with no parent in front of it. A resource of that type under a parent is a resource of the program.

Why the create stays when it is the only work

A stack that was never deployed and whose program holds no resource, only outputs or stack references, previews as one step: the create of the root stack resource. Dropping it would give an empty diff. The row would say in sync and have no box, while a deploy still has work: it makes the stack, its state and its outputs, which another stack may read through a stack reference. Such a stack could then never be deployed from the dashboard. So the create is kept when nothing else is left, and the row reads “1 create” of pulumi:pulumi:Stack. A step that changes nothing (same, read) does not count as something else. A tracking change does.

This is not the output change of 0036. That case is a deployed stack whose outputs changed, and the tool’s document cannot tell it. Here the stack does not exist yet, and the document says so.

Consequences

• The row of a stack that was never deployed counts the resources of its program, and so do its preview page, the summary and the result file.
• The diff hash covers the changes (0008), so the row of every stack that was never deployed gets a new hash once, on the first scan with this version. A tick made on the old hash is refused as a change that moved, like any other (0051).
• apply previews with the same adapter, so it computes the same hash, and a stack whose one change is the root stack create deploys like any other.
• The recordings hold a scenario for the one-change case, new-stack-without-resources, with both CLI versions. The update, delete and replace rows of the table come from no recording and are held by tests of written steps, in the shape of the recorded ones.
• OpenTofu, Helm and Kubernetes manifests have no such resource.