You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Research learnings and ideas from similar GitHub Actions #42
I looked through them again and two things that stood out for me
should we consider two parameters in order to set the installation for the token: either the name of the owner login or an installation id? I think an owner login should suffice, but if the installation ID is known then we would save a request.
I was trying to migrate from the action peter-murray/workflow-application-token-action to this one, and I noticed this action does not support the private-key input as a Base64 encoded string.
Does it make sense to create an issue for that? I don't have any problem opening a PR to support that use case.
what is the use case? Maybe the decoding a base64 string could be done in a separate step before using this action?
didn't have any internal retry mechanism
For retries, @octokit has the retry plugin, it's rather heavy though, due to its dependency on bottleneck. We try to keep the code of this action, including its dependencies, to a minimum. It might make sense to implement a custom retry in this case particular case
So joining these two use cases, I have experienced failures getting the token from a GitHub app, so we start to use an extra action to retry Wandalen/wretry.action, then we have the issue with the input that is broken with multiline strings like the format of the PEM files, so we had to do encode the private key with base64.
Also, it is annoying to keep copying and pasting the ugly syntax of an action that retries another action.
So, if this action itself has the internal retry, I don't even need the base64 encoding.
So, does it make sense to implement a retry logic inside the action?
what is the use case? Maybe the decoding a base64 string could be done in a separate step before using this action?
In my case we migrate from an action expecting the private key to be Base64 encoded. One solution is to create a new secret with the plain private key, but i just tried to test how we could decode it on the fly.
It does work, but it's quite the overhead in my opinion, especially to ensure the masking is done and multiline storage is working fine:
If the action is not planned to handle the private key being potentially encoded in Base64, I think we should add this in the readme, to avoid people leaking their private keys (I can open a PR)
If the action is not planned to handle the private key being potentially encoded in Base64, I think we should add this in the readme, to avoid people leaking their private keys (I can open a PR)
Having native support for Base64 encoded keys is an advantage as it avoids needing to insert decoding steps ourselves.
Implementing something like the above always feels risky as it's quite easy to break the masking if you're doing development against a GitHub Actions workflow. If the private key gets leaked, we need to manually rotate it in the GitHub UI and then repopulate the AWS secret.
For compatibility with actions/create-github-app-token and aversion to implementing decoding ourselves, we've instead ended up storing the app's id and private key in separate secrets. See implementation below.
This feels wrong because
the app ID isn't really a secret, we just want it parameterised away so that we don't have hard coded magic numbers in various workflows
it feels wasteful (AWS charges per secret per month) to store these in two separate values when they refer to one conceptual thing
the app id and private key have a similar lifecycle (e.g. they both get populated at creation), storing in the same secret makes sense
The tibdex implementation is much neater for our use case as I'm sure you'd agree. On the other hand, it is preferable to use the official implementation from https://github.com/actions. I'd really appreciate if this was added as a feature to get the best of both worlds :)
This issue has been marked
stale
because it has been open for 180 days with no activity. Please close this issue if it is no longer needed. If this issue is still relevant and you would like it to remain open, simply update it within the next 60 days.
Activity
gr2m commentedon Sep 4, 2023
I looked through them again and two things that stood out for me
There is also a lot to learn in terms of documentation, in particular how to register a GitHub App.
[-]Research learnings and ideas from similar ideas[/-][+]Research learnings and ideas from similar GitHub Actions[/+]Moser-ss commentedon Oct 26, 2023
I was trying to migrate from the action peter-murray/workflow-application-token-action to this one, and I noticed this action does not support the private-key input as a Base64 encoded string.
Does it make sense to create an issue for that? I don't have any problem opening a PR to support that use case.
Also, an issue we have with the action peter-murray/workflow-application-token-action that didn't have any internal retry mechanism; maybe we can improve that in this action.
I will be happy to start working in some PRs to improve this action
gr2m commentedon Oct 26, 2023
what is the use case? Maybe the decoding a base64 string could be done in a separate step before using this action?
For retries, @octokit has the retry plugin, it's rather heavy though, due to its dependency on
bottleneck. We try to keep the code of this action, including its dependencies, to a minimum. It might make sense to implement a custom retry in this case particular caseMoser-ss commentedon Oct 27, 2023
So joining these two use cases, I have experienced failures getting the token from a GitHub app, so we start to use an extra action to retry Wandalen/wretry.action, then we have the issue with the input that is broken with multiline strings like the format of the PEM files, so we had to do encode the private key with base64.
Also, it is annoying to keep copying and pasting the ugly syntax of an action that retries another action.
So, if this action itself has the internal retry, I don't even need the base64 encoding.
So, does it make sense to implement a retry logic inside the action?
gr2m commentedon Oct 29, 2023
would replacing line breaks with "\n" work? That's what we do in environments where multi-line environment variables are not supported
Perfect 👍🏼
I agree we should implement a retry, could you please file a separate issue for it?
newbloke82 commentedon Nov 2, 2023
I would like to try and help implement the proxy support. Getting this working from self-hosted runners behind a corporate proxy is important for my project.
I'm interested in changing the behaviour of the app to use GITHUB_API_URL as a default but allow for an override to deal with the use case of calling GHES from GHEC. This would also support GHES to GHES requests. This is useful for services like renovatebot/dependabot that need to lookup dependencies in private repos:
https://docs.github.com/en/code-security/dependabot/working-with-dependabot/configuring-access-to-private-registries-for-dependabot
https://docs.renovatebot.com/getting-started/private-packages/
gr2m commentedon Nov 2, 2023
@newbloke82 makes sense, thank you for laying it out! Can you please first file an issue? I think
tibdex/github-app-tokenis implementing proxy support the way you suggest, right? See https://github.com/tibdex/github-app-token/blob/3eb77c7243b85c65e84acfa93fdbac02fb6bd532/action.yml#L8-L10vleon1a commentedon Jul 8, 2024
In my case we migrate from an action expecting the private key to be Base64 encoded. One solution is to create a new secret with the plain private key, but i just tried to test how we could decode it on the fly.
It does work, but it's quite the overhead in my opinion, especially to ensure the masking is done and multiline storage is working fine:
The output is a bit ugly, but at least it's concealed:
This works also, and is more clean than the previous one:
If the action is not planned to handle the private key being potentially encoded in Base64, I think we should add this in the readme, to avoid people leaking their private keys (I can open a PR)
gr2m commentedon Jul 8, 2024
that's a good idea and thank you for offering. Let's add a note here: https://github.com/actions/create-github-app-token?tab=readme-ov-file#private-key. You could also link to your comment where as an example on how to decode a base64-encoded key
If this comes up more often we can reconsider to add built-in base64 decoding, but so far I've only heard it from you.
docs(readme): document how a Base64 private key could be decoded (#155)
lewismiddleton commentedon Jul 31, 2024
To advocate for supporting Base64 encoded private keys, our use case is using AWS SecretsManager to store GitHub app credentials.
SecretsManager supports plaintext secrets or JSON objects. To store the private key (with whitespace) as a JSON key value pair we need to encode it.
We expect a secret format of:
The aws-actions/aws-secretsmanager-get-secrets action supports
parse-json-secrets: truewhich can format environment variables based on the properties.We then consume this in a GitHub Actions job like this:
Having native support for Base64 encoded keys is an advantage as it avoids needing to insert decoding steps ourselves.
Implementing something like the above always feels risky as it's quite easy to break the masking if you're doing development against a GitHub Actions workflow. If the private key gets leaked, we need to manually rotate it in the GitHub UI and then repopulate the AWS secret.
For compatibility with
actions/create-github-app-tokenand aversion to implementing decoding ourselves, we've instead ended up storing the app's id and private key in separate secrets. See implementation below.This feels wrong because
The
tibdeximplementation is much neater for our use case as I'm sure you'd agree. On the other hand, it is preferable to use the official implementation from https://github.com/actions. I'd really appreciate if this was added as a feature to get the best of both worlds :)github-actions commentedon Nov 22, 2025
This issue has been marked stale because it has been open for 180 days with no activity. Please close this issue if it is no longer needed. If this issue is still relevant and you would like it to remain open, simply update it within the next 60 days.