Scripting

Custom scripts on AI Dungeon scenarios allow you to modify the memory, input, and output as well as keep track of custom state objects that might be relevant for your adventure. You can write custom scripts in Javascript by going to the "Scripts" section while on web on the edit scenario page. For security reasons some Javascript functionality is locked down however. Submit a request if there is functionality you would like opened up and we can look into it.

Examples

You can check out some examples for how to use scripting here. We also have user contributed scripts here.

Don't Know How to Code?

Then now's a great time to learn! A good resource to learn javascript from scratch is this free course.

https://www.codecademy.com/courses/introduction-to-javascript

History

You have access to (but can't modify) the history object which is a list of the previous actions of the player and of the AI, including the action type.

Memory

You have access to (but can't modify) the memory object which is the current user defined memory. You can modify the memory the game uses by settings the state.memory.context value. This will replace the user defined memory. You can also set state.memory.frontMemory, which will include whatever is there in front of even the last action when it's fed into the model, but still not display it to the user.

Author's Note

You can set state.memory.authorsNote to provide a piece of text that will always be injected three lines back in the game history. This will not be shown to the user, but the AI will see it.

As an example, if you set state.memory.authorsNote to the following paragraphs are scary., the AI will see [Author's note: the following paragraphs are scary.] three lines back, causing it to be more likely to generate scary text. Another example could be a dragon will show up soon or the player will soon receive a quest.

Modifiers

Shared Library

Prepended to the start of the other three scripts before execution so that you can share code between all three.

Input Modifier

Called each time the player gives an input and has the opportunity to modify that input. When inside of an Input Modifier, you can return stop: true in order to stop processing——see examples/commandParser.js.

Context Modifier

Called each time the AI model is about to receive input and has the opportunity to modify that input (by up to a 75% edit distance change). When inside of a Context Modifier, you can return stop: true in order to stop processing.

Output Modifier

Called each time the model generates an output and has the opportunity to modify that output.

World Entries

You can read from the worldEntries parameter (same as world info that you can set on the scenario)

You can modify worldEntries with the below functions

• addWorldEntry(keys, entry)
• removeWorldEntry(index)
• updateWorldEntry(index, keys, entry)

State

The state variable can be used to store information that's persistent across function calls/modifiers.

• The state.memory.context value will replace the user defined memory if it exists
• The state.message value will be displayed as a extra message in the game (if it exists)
• You can set any variable on state to store and modify adventures throughout an adventure.

Console

console.log("Some message") will log messages that you can see in the scripting console

Info

info contains some useful values, depending on which modifier you're in. All modifiers have access to info.actionCount, the number of actions in the adventure so far.

When in a Context Modifier, info.memoryLength and info.maxChars are also set, indicating the length of the memory portion of text (if any), and the total allowed length of the context after which it will be truncated.

Last Model Input (LMI)

Clicking on the brain icon in the scripting interface will open LMI, in which you can see the last context the AI was provided, the console, and the state.