Tool usages in Copilot sometimes consumes premium request

[aklajnert]

Description

According to GitHub's changelog GitHub Copilot coding agent now uses one premium request per session, Copilot should consume only one premium request per session, regardless of length, complexity, or tool calls. This implies tool usages (like question) shouldn't trigger additional requests.

In OpenCode, this holds mostly true, except for the question tool. I encourage the model to clarify uncertainties by this tool. During sessions, I monitor the usage meter at github.com/settings/copilot/features.

Observation: Answering the model's question (via dedicated UI, without sending a new main prompt) sometimes increases the premium request counter. It's inconsistent – happens sporadically, not every time. I can't pinpoint exact triggers (model? session state?). It seems that it happens more often with Opus, which is annoying as this is the most expensive model.

I'm using OpenCode Desktop.

Plugins

No response

OpenCode version

v1.2.10

Steps to reproduce

1. Start Copilot session in OpenCode in agent mode with some premium model like Claude Sonnet 4.6
2. Check the usage meter value
3. Send a following prompt:

Ask me 5 different questions, one by one. Only ask one question at a time, and once I answer it, ask another one.

4. Observe the usage meter after sending the prompt and before answering any question - the meter should increase.
5. Answer in UI → watch meter → sometimes +1 premium request.

Note: I just did the steps, and now the meter didn't increase, even when I waited 30 minutes before sending answer to question number 4. Perhaps this was too simple, and the additional usages are counted when the session gets too big?

Operating System

Windows 11

Terminal

OpenCode Desktop

[pushad]

I'm pretty sure it consumes a premium request when it compacts. You can setup OpenRouter (or any other provider) specifically for compacting so that it does not go via Copilot. By default it will use a lower premium costing model though, so it should not be +1 premium request for compacting.

[aklajnert]

Oh, that makes sense. I don't think I can use OpenRouter in my company, but I could use a free model from Copilot for that.

Any hints on how to trigger automatic compaction so I can confirm that it was the cause of the problem?

[shekohex]

to trigger a compact (manually) you run /compact

btw, how do you find the copilot premium request value? and how much is the context window for say, opus?

[aklajnert]

I would need the autocompaction to trigger without my intervention to confirm that this is the case. Or a way to confirm that the autocompaction occurred in my session.

The context window for each model is presented in the VS Code when you click "Manage models":

[buz-ch]

I see it consuming multiple requests when using GUI (codenomad, openchamber) but not in the TUI. Very strange.

[georgeharker]

It uses a premium request to compact (unless you set the compaction model) but also the resume after compaction is currently a premium request also

[riosengineer]

It uses a premium request to compact (unless you set the compaction model) but also the resume after compaction is currently a premium request also

I didn't know this, how can I tell when it consumes a premium request? I noticed the last few updates auto compaction for me which looks like it's contributed to my consumption on requests.

[krasen1234]

It uses a premium request to compact (unless you set the compaction model) but also the resume after compaction is currently a premium request also

Hmm im using compaction model as free but the resume seem to be counted still?

[real-or-random]

This could be the same underlying issue as karthink/gptel#1224. Copilot requires special headers when doing tool calls to get the billing right.