Reduce Premium Copilot requests

[Arithmomaniac]

As mentioned in #47, the "X-Initiator" header should be set for non-user calls to not unnecessarily blow through premium requests. See sst/opencode#595 for discussion and implementation.

[ericc-ch]

Hey I've added the header logic, directly copied and pasted from opencode lol
Feel free to reopen if it persists

[cuipengfei]

Hey I've added the header logic, directly copied and pasted from opencode lol Feel free to reopen if it persists

I noticed that when using claude code + copilot api, the premium quota goes down slower than when using copilot itself directly.

When using copilot itself directly, pretty much every message I send in the chat box will reduce quota by 1. But when using claude code + copilot api, I can send multiple rounds of messages and the quota does not seem to go down nearly as quickly.

Hope this does not trigger any alerts or be marked as suspecious behavior by github copilot backend.

[cuipengfei]

this is interesting, i sent quite a few messages in claude code

then copilot api correctly set the header as user for each message i sent, there is no problem here

but then i checked my quota, it did not go down

i sent a few more after the screenshot was taken, the quota still not go down

do you see any potential of this being flagged as abusing by github?

since when i use copilot itself, it pretty much just charge me one premium quota for each message sent

any ideas would be appreciated @ericc-ch @Arithmomaniac

[Arithmomaniac]

Sorry, I really don't have further insight here.

[cuipengfei]

so i plugged localhost:4141 right back into vs code copilot plugin itself

i had been using gpt-5 for more than 20 agent mode and chat mode prompts

and it is only costing me one premium quota

and if i were using copilot as is in vs code, it would have cost me 20 premium requests

i will probably do more digging about this

were you using any particular tools for reverse engineering? @ericc-ch like wireshark etc?

or would i have better luck now that we have this: https://github.com/microsoft/vscode-copilot-chat?

[ericc-ch]

This is actually nothing special, as you can see in the linked opencode PR
I just marked requests with assistant / tool as "agent" call
That's why they don't cost premium requests, because they are counted as a single session

For reverse engineering I used https://www.mitmproxy.org since I find it easy to use and the web UI is very neat

@cuipengfei

[cuipengfei]

i am only guessing here, for the copilot plugin itself

if the last message in the array was from the user, it would set the header as user, and the quota will go down

while in copilot api code, if the whole array contains as least one assistant or tool message, it would set the header as assistant, thus saving the quota

sounds about right?

[ericc-ch]

i am only guessing here, for the copilot plugin itself

if the last message in the array was from the user, it would set the header as user, and the quota will go down

while in copilot api code, if the whole array contains as least one assistant or tool message, it would set the header as assistant, thus saving the quota

sounds about right?

Yes that's correct. This is the exact same behavior as opencode