I'm in favor of this approach for the community MVP 👍🏼

This ensures the workspace lifecycle is fully-managed at the template level, but also does not remove support for dotfiles (which has first class support in other remote development platforms).

I really like this approach, especially that we can bake it into thecoder binary itself and avoid shipping a shell script.

Check out repo, deal with updating if already exists

What do we do if something "weird" happens and we can't fast-forward to the remote branch? Bail? Force-reset?

I really like this approach, especially that we can bake it into thecoder binary itself and avoid shipping a shell script.

Check out repo, deal with updating if already exists

What do we do if something "weird" happens and we can't fast-forward to the remote branch? Bail? Force-reset?

My default plan is to fallback to whatever the current personalize script from V1 does, which I think is just present the git error. Very much open to input on handling cases.

I really like this approach, especially that we can bake it into thecoder binary itself and avoid shipping a shell script.

Check out repo, deal with updating if already exists

What do we do if something "weird" happens and we can't fast-forward to the remote branch? Bail? Force-reset?

My default plan is to fallback to whatever the current personalize script from V1 does, which I think is just present the git error. Very much open to input on handling cases.

I think just keeping v1 behaviour is legit for starts 👍

Down the line, I could imagine some additional fanciness, such as:

• If the checked out repo is dirty, stash all the local changes (maybe also add all untracked files in case they conflict with upstream changes)
• If the local and remote branches have diverged, rename the local branch with a suffixcoder-$(date) and set the original local branch to upstream

In both of those cases, we'd want to warn the user.

I had no idea that GitHub documented a list of places to check for dotfiles init scripts, but that's really cool. Ireally like the simple addition of a command, and this integrates ✨ beautifully ✨ withgitssh! 👏

It's simple UX, market-driven, easy to implement, etc. I did a cursory glance but I like the proposed solution. I defer to the other smart people in this thread to make the right call.

I may be misunderstanding this, but it sounds like the above implementation is Terraform-specific. I also am not thrilled about this being under the project owner's control rather than the user's.

I may be misunderstanding this, but it sounds like the above implementation is Terraform-specific. I also am not thrilled about this being under the project owner's control rather than the user's.

It is not terraform specific, thecoder dotfiles command can be run anytime. For example, if we didn't add this to our dogfood template I could just runcoder dotfiles git@github.com:f0ssel/coder-dotfiles.git every now and then in my workspace and I'll get the same experience in a manual fashion.

It is up to the template owners to implement the call tocoder dotfiles in the startup script, but the url can be easily made a variable that can be filled by the user at workspace build. The terraform changes here are an example of what the real diff would look like if we added added dotfiles to the example local docker template.

I do understand that there's a use case where customers want the same dotfiles for everyone, so this serves that. I guess it doesn't get in the way of something user-driven.

That just leaves me with the concern about it being Terraform specific. I think this is a good opportunity to consider the workspace lifecycle hooks idea. I think it's not too different in complexity, especially if we're just doing a start event initially, but it's more flexible and more general.

Yes it will support any kind of commands for the startup script, so you could have it be:
coder dotfiles git@github.com:coder/org-dotfiles.git
or it could be
coder dotfiles ${var.dotfiles}

Up to the template owner.

Thestartup_script is essentially a one-off workspace lifecycle hook, it always runs if specified after the agent has started successfully. I believe there's value in a formal design for workspace lifecycle hooks, but right now I have what I need to implement this today withstartup_script, so I think that can be discussed in a separate issue. Whatever we land on there, I'm sure we can hook in the dotfiles command somewhere that makes sense.

I'm in favor of this approach as a first-step.

Hook into templatestartup_script variable to add a script that callscoder dotfiles <url>

This seems like a possible point of incompatibility.

I'm not sure about Codespaces, but I just tested how dotfiles work on Gitpod, and it looks like they preserve the/workspace working directory butdon't preserve the contents of the/home/gitpod home directory across workspace restarts. So if a dotfiles repo has an install script, we could end up running it multiple times in the same homedir, even though Gitpod only runs it once. This could break things if the script isn't idempotent.

I guess ultimately, the decision of how to handle this is up to the template author. But if we plan to ship templates where the home directory is persistent, then to maximize compatibility, maybe we should provide afirst_run_script in addition tostartup_script, and use that to invoke the dotfiles command?

Or we could just punt on it for now, and warn users that the script may be run an undefined number of times.

@dwahler that is a great point. I appreciate you doing this research into how others have solved similar problems.

We currently have the expectation of idempotent install scripts in V1 (not that it's a total authority, but an existing expectation), and the coder dotfiles command itself is also idempotent. I feel pretty okay with that requirement today but I do see the divergence from the way others handle it.

Another point is we don't really have an opinion on what persists, it's completely up to the template. I can see this being a good case for letting the command take in an argument for where to store the repo, so the caller can place it somewhere ephemeral if they'd like to.

If we see this as the first lifecycle hook, I think we should reconsider the namestartup_script and instead use language consistent with the internal event name and event handlers in general, e.g.onStart.

If we see this as the first lifecycle hook, I think we should reconsider the namestartup_script and instead use language consistent with the internal event name and event handlers in general, e.g.onStart.

I agree, but we should discuss in a dedicated issue so we can prioritize it correctly. I think it's doable to make these changes before community if we find it valuable.

Closing now that the implementation is merged.