Autonomous advent of code

• Each day during advent of code this repo solves the latest puzzle and submits the solution - all without human intervention
• All code in this repo was written by sourcery-ai[bot]
• Everything was set up by manually writing five issues and asking sourcery-ai[bot] to implement them
• Since then everything runs fully autonomously

There are two main processes that work together to solve the puzzles:

• Four GitHub actions workflows download the puzzles, create issues for sourcery-ai[bot] to solve the puzzles, merge PRs, and submit solutions. All actions they perform are carried out by @SourceryAI which is a GitHub account not associated with a real person
• sourcery-ai[bot] solves the puzzles and opens pull requests with the solutions and answers

How does it work?

Here are the steps:

1. Twelve hours after each puzzle is published a GitHub actions workflow ¹:
   • Download the puzzle and input
   • ROT13 encode the puzzle ²
   • Commit both to the puzzle directory and push to main
   • This triggers step 2
2. A GitHub action workflow is triggered when puzzles are pushed to main:
   • Create a GitHub issue to solve the part
   • Post an issue comment with content @sourcery-ai develop
   • This triggers step 3
3. sourcery-ai[bot] listens to the comment:
   • Runs a script to ROT13 decode the puzzle
   • Posts a plan to the issue to solve the puzzle
   • Opens a pull request to implement the solution and write the solution
   • This triggers step 4
4. A GitHub action workflow is triggered when pull requests are opened
   • It merges pull requests only if the author is sourcery-ai[bot]
   • This triggers step 5
5. A GitHub action workflow is triggered when answers are pushed to main
   • It submits the answer to advent of code
   • It updates the results below
   • If the answer is correct and it was for part 1, it downloads the second part of the puzzle and commits it
   • This triggers step 2

How sourcery-ai[bot] works:

• When a member of a repo comments on an issue with @sourcery-ai develop it will post a plan to the issue, open a pull request with the solution, and request a review from the commenter
• If any review comments are left on its pull request it will address them and push new commits

Results

Here are the puzzles attempted and whether they were successfully solved or not.

Try out Sourcery

sourcery-ai[bot] is an GitHub app with two main functions:

• Issue agent
  • Comment an issue with @sourcery-ai plan to post a comment with a planned implementation
  • Comment and issue with @sourcery-ai develop to open a pull request to resolve the issue 👈 the functionality on display in this repo
• Pull request agent
  • Automatically reviews pull requests
  • Generates pull request titles / bodies
  • Generates review guides with diagrams explaining the changes

The functionality in this repo (@sourcery-ai develop) is currently in alpha testing and will be generally available in the new year. As well as the functionality shown here it will:

• Run in GitHub / GitLab - cloud and self-hosted
• Integrate with issue trackers - Linear, Jira
• Resolve errors in production - Sentry, etc.

It will be most useful for handling routine maintenance and clearing your backlog by:

• Fixing bugs
• Implementing smaller features
• Automatically fixing production issues
• Writing documentation

If you like the sound of this, join our waitlist and tell us anything else you want it to do.

Footnotes

Footnotes

1. We run this twelve hours after the puzzles are published to avoid interfering with any leaderboards ↩

2. sourcery-ai[bot] cannot log into the advent of code website and read the puzzles and input, so we store them in the repository. The advent of code rules ask you not to publish the puzzle text - so we're ROT13 encoding it ↩