Idle token consumption / idle-state misclassification affects Claude and Codex

[voyax]

Summary

There are reports that token consumption or apparent token consumption can happen during supposed idle periods, and this is not limited to Codex. Claude and Codex should both be included in scope.

Scope correction

The initial filing was too narrow. This issue should track the broader runtime-level problem instead of treating it as Codex-only.

Included runtimes:

• Claude
• Codex

Already ruled out as complete explanations

• Old Claude /usage interruption bug (PR fix: guard /usage check against active prompts using in_prompt signal #343)
• Old Codex total_token_usage vs last_token_usage accounting bug

Those historical bugs should not be treated as the full explanation for the current report.

Current observed behavior

• There are reports of token consumption or apparent token consumption during idle windows
• The behavior may be one of two things, and the issue should allow for both:
  • real background token burn during idle
  • idle-state / busy-state / usage-accounting misclassification that makes the runtime look like it is consuming tokens while idle
• This needs to be checked across both Claude and Codex paths, not only one runtime

Investigation scope

• Verify whether usage counters, rate-limit usage, or cost actually move during clean idle windows
• Separate true idle token burn from misleading monitor/runtime state transitions
• Compare Claude and Codex behavior instead of assuming the issue belongs to only one runtime
• Focus on current production behavior

Notes

This issue was filed directly on request first, then corrected to widen the scope from Codex-only to Claude + Codex.

[eric-sss89]

Update related to #470:

The usage-monitor implementation has now been changed to a local file-based path, and this matters for the idle-token investigation here.

Current behavior after #470:

• The old sidecar probe path was removed.
• We no longer start an extra Claude/Codex side session just to collect usage data.
• Claude usage is read from local activity-monitor files (statusline.json, with usage.json only as compatibility state if needed).
• Codex usage is read from the local usage-codex.json file.
• Usage monitoring remains behind an explicit config flag and is disabled by default.

Impact on token usage:

• The current usage-monitor path itself should not consume model tokens.
• It is reading already-written local status files, not sending a prompt, not calling /usage through a sidecar session, and not creating a second runtime session.
• So if there is still token usage or apparent token usage during idle windows, the usage-monitor implementation introduced by Usage monitoring: remove sidecar and read usage from activity-monitor status files #470 should no longer be treated as the likely source.

In other words: #470 narrows the scope of this issue. It removes one concrete background-noise source from the investigation. If idle-period token movement is still observed, the remaining suspects are more likely to be heartbeat/probe behavior, idle/busy-state misclassification, or some other runtime-level path outside the new local file-based usage reader.