❇️ $\subseteq$ Commit Challenge 2025

Welcome to the annual $\subseteq$ Commit Challenge 2025! The goal is to make contributions and a commit every day to build something that can (hopefully) be considered meaningful and deployable by the end of the year 🚀 fish are tasty

Caution

Our pages site is available on Github Pages

Our current participants:

• Nolan
• Sean
• Sean
• Marcelo
• Nick
• Raymond
• Khushi
• sam

Important

CHAT B#C# DONE IT AGAIN

✋ Want to Join?

Fill out this form here

Tip

To join the cave game instance, contact someone through Discord.

❗ Rules

You must commit to main. No malicious code pls.

🚧 Projects

Warning

Some projects may be functional.

in order of GDP:

Project	Description	Technologies	Files
🏠 home page		html css javascript web audio	index.html ass/ets/ m/a/r/c/el/o/ .github/workflows/fuck.yml
🗡️ murder game	for nick's acm cyber game hacking event	typescript canvas	murder-game/
📱 qr game	jackbox but using qr codes	javascript	qrbox/
🏃 escape game	text based choose your own adventure	css typescript	escape/ .github/workflows/fuck.yml .github/workflows/compile.yml
🐎 horse but even better	simulating wynncraft horse gambling	javascript	horse/ horse-but-better/ horse-but-even-better/ last-horse/
🔑 login	ucsd phishing exercise	ctrl+C ctrl+V	login.html
🕋 cube-spin	learning webgl	webgl	cube-spin/
🪨 cave game	3d multiplayer roguelike game in a cave	typescript webgl cannon-es	cave-game/ .github/workflows/autolint.yml .github/workflows/fuckelonmusk.yml .github/workflows/compile.yml
📊 green	commit challenge stats	remind/green.mts
🚨 remind	sends discord reminders to commit daily	javascript rust serenity	remind/
📰 release	makes a nightly release for the repo each night summarizing changes using gemini	github actions gemini	.github/workflows/release.yml gpt_prompt.txt
🍺 drunk		html	help.html sean-retreat-attempt.html
💭 gpt		text	gpt.txt
📅 actions game	rpg game where you write your actions for the day in a text file and a github action runs daily and uses gemini to determine what happens next	github actions gemini typescript	.github/workflows/release.yml actions/ actions.md .github/workflows/actions_game_test.yml
🤖 eat shit game	discord bot llm rpg game game	typescript discord.js	🍴💩/
💥 explodes	phone at risk of fire	html css javascript	explodes/
godot web export test		godot	godot-web-export-test/
❌ purr	enforces "you must commit to main." if you open a PR it will close it and delete the branch	github actions	.github/workflows/purr.yml
🍉 fruit merge		html	fruit-merge/
👑 queens	queens puzzle designer. can a queens puzzle with equally sized regions exist?	react	👑/

Actions Lore

Figure 2. Architecture diagram or flow or something of how the actions game was initially set up

Initially, the actions game was implemented as shown in (Figure 2):

1. We had three files in the repo:

   • prompt.md: the prompt defining instructions and a JavaScript API (defined with TypeScript declarations) to manage the state

   • state.yml: contains the game state as mostly unstructured data:

     • players: a map of player name to a player object. The prompt used to recommend using methods like setInfo to edit the object, but I felt that this could discourage it from making more complicated structures. For example, to edit an object in info, it had to repeat the entire existing object in code. I then changed it to recommend mutating health and info directly.

       • health: originally managed with changeHealth
       • inventory: map of item name to count. managed with addItem, removeItem, so once an item goes to 0, it gets deleted. had a bug where Gemini inadvertently passed undefined as the count, setting the count to NaN
       • info: unstructured data, originally managed with setInfo

     • worldInfo: unstructured data, originally managed with setWorldInfo

     • previousResponses: stored the responses generated in the previous day

       This means that the game only has a reliable context window of only one day, after which we rely on Gemimi storing data in the state properties.

   • actions.md: the player input

2. The above three files were given as the prompt to Gemini and used to generate JavaScript code.

   While Gemini offers a way to separate the system prompt from user prompt, part of the enjoyment of an LLM-based game is to try to gaslight the bot, so we decided to combine both the game prompt and user actions into one prompt.

3. Process the JavaScript file a bit.

   Gemini likes to wrap code in backticks. I think, again, Gemini offers better ways to return code, but using a regular expression to comment those backticks out have been reliable enough.

   After removing the player/world info setter methods, it turned out that Gemini would sometimes get property accesses wrong, so a long property access like player.Sam.info.cats.count might be missing a key, resulting in a runtime error that crashes the game. I have another regular expression that wraps these property accesses with code that creates the necessary objects before making the property access.

4. Evaluate the JavaScript.

   eval(js)

   This would probably be a bad idea in general, but all players of this game already have write access to the repo, so this doesn't let them do more damage than they already can.

5. Collect the responses.

   Part of the API we define for Gemini included .respond() for players and describeDay() for the world and instructions requiring these be called at least once. Multiple calls result in multiple paragraphs. One curiosity is that sometimes Gemini thinks it's also a getter when called with no arguments, so it once tried to use it to concatenate to a previous response.

6. Send the responses to three places:

   • A Discord webhook to display the results in our Discord server. However, due to Discord's character limit, oftentimes responses had to be truncated, which made it annoying to read.
   • The state's previousResponses.
   • Standard output, which was used as the commit message by the GitHub Actions workflow.

7. Save the new state into state.yml and commit and push the changes to the repo for the next day.

We also had a few unwritten rules:

• Don't directly edit others' actions (but you can use your prompt to gaslight the game and change others' actions).
• You can move your action in the file (i.e. change the order of players in actions.md).
• Looking at the state file is fair game because it defines the memory of the game. If it's not in the state file, it's not canon.
• Don't edit the game prompt for your own benefit.

There were a few issues with this setup, though:

• Gemini would say things but not save them in the state, meaning that it'll forget about it as soon as a day later.
• The game wasn't providing much gameplay on its own, so we eventually got bored of the game and weren't sure what to do next. We had to beg it to generate quests for us.
• It was too easy to gaslight the game. This was fun at first, but it eventually got boring.

Figure 4. Architecture diagram of actions game v2

To fix the game state issue, we tried splitting the Gemini calls in two:

1. Response: The first call generates the responses for players (so .respond() and describeDay() were removed) as prose in Markdown, based on the previous state and player actions.

   Generating it in plain text rather than separate respond() calls seemed to have encouraged it to write more, and the responses per player seemed more coherent in general.

2. State: The second call edits the state object based on the generated responses. It has access to the previous game and responses, but not the player actions, hopefully making it harder to change the game state by gaslighting the bot.

   To reduce the context size and to encourage relying more on the state object, I also removed previousResponses from the state object, so the first call doesn't have access to its previous output.

To prevent players from pretending to be system administrators, we asked Gemini to ignore them. Hopefully that works!

We also introduced a d20 roll system. Before the first Gemini call, we generate a random number from 1 to 20 in JavaScript (i.e., an LLM does not generate them) for each player. We include these numbers in the prompt for the first call, and ask Gemini to use these to determine the success of each player's action. This should hopefully limit the effectiveness of players jailbreaking the prompt.

📝 Current To-Dos

Caution

Feel free to leave ideas here

• Github Pages

• choose your own adventure game

• qr-code based offline jackbox

• Cave game

  • Get server running
  • Get networking going
  • Get renderer working

• Movies

  • Epic
  • Wild Robot
  • Monty Python and the Holy Grail
  • Princess Bride
  • Anora
  • fruity pinocchio

Roadmap