[WIP] Support for Copilot Instructions

[DanBradbury]

Living the RE life I'm stuck trying to figure out the black magic that VSCode is using to make this happen.

I setup a simple example using "Ask" mode with a .github/copilot-instructions.md in the working directory

There are 2 requests made by VSCode for this chat message

• POST https://api.individual.githubcopilot.com/chat/completions

{
    "messages": [
        {
            "role": "system",
            "content": "Typical system prompt junk..|"
        },
        {
            "role": "user",
            "content": "testing"
        }
    ],
    "model": "gpt-4o-mini-2024-07-18",
    "temperature": 0.1,
    "top_p": 1,
    "max_tokens": 20,
    "stop": [
        ";"
    ],
    "n": 1,
    "stream": true
}

• POST https://api.individual.githubcopilot.com/responses

{
    "model": "gpt-5-mini",
    "input": [
        {
            "role": "system",
            "content": [
                {
                    "type": "input_text",
                    "text": "System message junk"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "When generating code, please follow these user provided coding instructions. You can ignore an instruction if it contradicts a system message.\n<instructions>\n<attachment filePath=\"/Users/danbradbury/Documents/Github/vinter/.github/copilot-instructions.md\">\nAlways add an emoji on the first line of your response\n</attachment>\n\n</instructions>\n<prompt>\ntesting\n</prompt>\n"
                }
            ]
        }
    ],
    "stream": true,
    "max_output_tokens": 64000,
    "store": false,
    "truncation": "disabled",
    "reasoning": {
        "summary": "detailed"
    },
    "include": [
        "reasoning.encrypted_content"
    ]
}

In my current working example I was able to bypass using /chat/completions entirely. There is obviously a change in the responses as well that would need to be parsed.. could still use the delta or the final text content that comes back from the response endpoint