FR: Make it an error to use a divergent change id?

[martinvonz]

Is your feature request related to a problem? Please describe.

When there is a divergent change, one usually doesn't want to use the change id to refer to all of the commits. A recent Discord discussion had an example where one side of the divergence was not shown in the log output. In such cases, it's easy to think that e.g. jj abandon <change id> would just abandon the visible commit.

Describe the solution you'd like

If users rarely want to refer to all the sides of a divergent change, would could just make it an error. We can provide a revset function for when the user actually does want all sides of the divergence.

Describe alternatives you've considered

Keep the current behavior.

[ilyagr]

A few inconclusive thoughts:

The most useful command to run on divergent changes is probably jj log -r divergent_change and such.

Another option would be to make this a warning, and rely on users to undo if this is not what they meant.

Alternatively, we could make jj abandon require the all: prefix like many other commands do. Then, jj abandon changeid will fail if changeid is divergent. I initially thought that jj abandon main..@ is too convenient, but I don't think it's a huge deal to require jj abandon all:main..@.

Originally, divergent changes were part of my motivation for requiring the all: prefix for many commands (well, actually for the predecessor of that feature, but that's not very relevant). At the time, I was wondering about what would be something like an allow_divergent_expansion: prefix (with a shorter name), but I preferred all: because @- is another operation with similar properties. People often expect @- to result in a single revision, but that expectation can be dangerous.

[ilyagr]

I saw on Discord that you were also thinking about removing all:. Perhaps these issues should be considered interconnected; we could have an overarching issue for both of them. I don't thing we should change these things one at a time.