Idle token consumption

[yvyw]

Description

I left OpenCode running overnight, and in the morning, I noticed that claude -p . --model haiku is consuming about 25% of the five-hour session limit of the Pro plan. I think this should be mentioned in the readme.
Probably related #99 and #93 (comment)

Steps to reproduce

• Start opencode
• Open https://claude.ai/settings/usage
• Watch your tokens slowly disappear

Operating system

Linux

Plugin version

1.3.3

OpenCode version

1.3.5

Debug logs

[troy-johnson]

I'm not sure if this is related or not, but I was running OC overnight using a gpt model. I woke up to 100% 5-hour session usage consumed on my claude pro plan and ~25% of my weekly usage consumed. The auth token history under https://claude.ai/settings/claude-code shows a ton of user:ccr_inference lines during the past 12 hours as well.

[Johannbr]

@yvyw , are you using https://github.com/hamed-elfayome/Claude-Usage-Tracker ?
This app seems to me to be responsible for idle tokens consumption.

[yvyw]

@Johannbr no, I don't use it. My setup is quite basic. I investigated the issue before reporting it and it's claude -p . --model haiku run by opencode-claude-auth to refresh the authorization tokens in my case. At first, I thought it was malware, but it wasn't. Not this time.

[0xRake]

The core problem here is that auth token refresh is implemented via inference calls (ccr_inference). This is an architectural issue, not a configuration one — refreshing an auth token should never consume session quota.

Two separate concerns:

• Auth refresh should be a zero-cost operation. If opencode-claude-auth is validating tokens by making lightweight inference calls, that's a design flaw. Token validation endpoints exist for this purpose.
• The refresh loop runs unconditionally. There's no idle detection — if OpenCode is open, the loop runs whether you're actively coding or not.

This should be documented in the README as a known limitation until fixed. Users leaving sessions open overnight are silently burning 25-100% of their Pro plan session allocation.

[griffinmartin]

Thanks for bringing this to attention. I'm currently testing a direct oauth refresh, with CLI as a fallback, which should eliminate the token burn here.

[griffinmartin]

Fix went out in v1.3.5.