Issue in claude code

[WayneEcon]

Since Aug 6th, when using Claude Code CLI → copilot-api → GitHub Copilot, a single short user message causes multiple POST /v1/messages?beta=true calls and increments my Premium requests by 2–3 (or more).
From Aug 1–5, the exact same setup behaved like an “agent session”: chatting for a long time in one terminal typically consumed only 1 Premium request overall.
I’d like to understand whether this is:

an intended change in copilot-api behavior,
an incompatibility with newer Claude Code releases, or
something on the GitHub Copilot side that copilot-api can’t (or shouldn’t) aggregate.

If there’s a way for copilot-api to aggregate multi-turn tool/diagnostic subcalls into a single Coding Agent session (or a “single-shot per user message”) I’d love to enable it.

[ericc-ch]

We have the code to determine if something is an agent request (basically if messages contains assistant or tool role, it'll count as an agent request)
But the request is completely controlled by the client (in this case claude code)
So nothing has been changed about premium request usage from our side

That said, have you tried disabling non essential model calls?
The usage might be from there

DISABLE_NON_ESSENTIAL_MODEL_CALLS=1

https://docs.anthropic.com/en/docs/claude-code/settings#environment-variables

[WayneEcon]

Thanks for the reminder! Seeing your suggestion, I went to check the JSON.
Although it wasn't the line of code you mentioned, it helped me find the real problem.

My "ANTHROPIC_SMALL_FAST_MODEL" was automatically completed to sonnet-4 by Copilot... each bash costs at least 1 premium that's insane.

This is probably why my "hello" command was deducting heavily from premium😂. Thank you very much! 🙏