majeedkazemi / coding-steps Goto Github PK

using:

• python-shell
• websocket

this list could help them utilize codex better

• so that the instrumentation states would be completely discrete

• needs to add a \n at the end of it

• type: write code, modify small piece of code, multiple choice, answer question, debug code
• id

hierarchical mongo class (using polymorphism)

• press ctrl+space -> open input + enter description -> place generated code at cursor -> highlight code + ask to accept/reject
• in a completely different box -> allow the user to enter description -> generate code below -> press insert to

change # Prompt: to # Instructions:

show a message: you've skipped this task on top so that they would not do it again.

such as what they are missing.

collecting data is easy, but an important and difficult thing is to reduce data-loss. how can we do that?

will also be used in the blocks-to-code project.

• mongodb
• express-session or cookie

what is a string
what does str do
what does int do

when it's too long.

• finished task ids : array of task ids
• finished task performances: array of task_id + user_id performance metrics (data collected for a particular task)

performance metrics could be different based on task type

a simple rate limiter algorirthm:

• std:output per second should be less than a particular value -> kill it if it becomes more than something.

see these:

• https://beta.openai.com/codex-javascript-sandbox
• https://blog.scottlogic.com/2021/09/24/how-the-openai-codex-javascript-sandbox-works.html

implementation:

• add api-key to .env
• set stream = true add stream using fetch-event-source -> don't use the codex-router
• add <|endoftext|> at the beginning of the prompt
• start each command with # Command:
• add a few shot examples -> could even come from the user study (just grab from the logs in the db)

should check if the socket is open before communicating or something like that.

using:

• pylsp
• monaco-languageclient

code changed, do you want to run it again?

so that the user could login and then start the process

• combine input and randoms
• instead of asking them to add dice3 to check that

• int
• str
• input
• range
  ...

to help easily identify them

• changes to the code:
  • before
  • after
  • a sequence of edits
    • insert character
    • replace character
    • delete char
• executed code
• usage of documentation
• time between each task
• natural language prompts

most important design goal is to collect interactions and times between each. therefore:

• click on a button to show a big documentation covering most of the screen
• inside the documentation, there could be a search box and lots of boxes to click on.
• clicking on each box would show a big box below it (clicking on another box would close all other open ones)
• each box could have a few internal boxes similar to the python-docs