According to an inspection of the released binary, the client can use a small fixed allowance from the weekly limit to finish the current step and leave up to three short bullets describing the remaining work.
The observed hidden instruction tells the model to checkpoint, avoid starting subagents or long-running work, and finish or record the current step. The binary also contains support for two rate-limit headers, anthropic-ratelimit-unified-grace-5h-utilization and anthropic-ratelimit-unified-grace-7d-utilization, plus an approaching-limit warning. The report says Claude Code saves the current todo list at the earlier warning.
The wrap-up instruction is reported to be injected only while a tool-call loop is already in progress. A new user message after a limit is reached does not receive the pending note. The behavior can work with the /config option “Continue automatically at usage limit,” which resumes after the window resets.
The source is an independent binary inspection, not an Anthropic release note. It does not establish the size of the allowance, which accounts have the server-side flags enabled, or whether the behavior is universally available.