Quota draining? Claude Code re-bills a finished sub-agent's whole context on a follow-up

A user discovered a cache-resume issue in Claude Code, potentially contributing to widespread usage limit issues. On Claude Code 2.1.273, a resumed request was significantly smaller than the previous live one, by the exact amount of "thinking" tokens produced by the agent. The API reported no removals, and Claude Code requested all thinking be kept, indicating the thinking was not sent. This suggests a re-billing of the sub-agent's entire context on follow-up, an issue that has been filed upstream.