What probably matters to youSection of the release
What
The total-tokens reminder is a line such as <total_tokens>N tokens left</total_tokens> that tells Claude how much of its token budget remains, added to the system prompt and after tool results. How its mode is chosen changed:
Otherwise the rollout value from config, falling back to "padded-countdown", applies only when the model has a new total_tokens_reminder capability and a flag check passes. If not, the mode is "off".
The setting's description changed from "Defaults to padded-countdown" to defaulting to padded-countdown only for a model with the total_tokens_reminder capability, else off.
The chosen mode is now stored once as totalTokensReminderMode, noting whether it was set explicitly, instead of being worked out on every call.
total_tokens_reminder was added to some model entries. The members point to a claude-3-5-haiku-era entry that had no capabilities, an entry with a 200000 window and 1M beta that listed only context_management, claude-3-5-sonnet, an opus entry, and one listed next to haiku_5_5_early_stopping_guidance. Several older models already had it.
In the captured prompts, the claude_3_opus model no longer shows its <total_tokens>15000000 tokens left</total_tokens> line; that link was noted but not traced.
Why
Whether Claude sees how much budget is left now depends on the model, so some models stop getting the reminder by default. Set totalTokensReminder or CLAUDE_CODE_TOTAL_TOKENS_REMINDER if you want a fixed behaviour.
