Usage-limit wrap-up messages can be tuned per model and now cover the near-limit notice

What

The hidden messages Claude Code can send to the model around a usage limit's grace window (a short period near hitting the limit) changed in two ways.

The text can now be set per model through remote settings:

• tengu_lantern_wick_text_by_model, tengu_lantern_wick_release_by_model and tengu_lantern_wick_resume_by_model are maps looked up by model id, then its normalized form or model family, then "default".
• If no entry matches, they fall back to the existing single-text settings tengu_lantern_wick_text and tengu_lantern_wick_release.
• New tengu_lantern_wick_braid_text and tengu_lantern_wick_braid_release are read for agents other than the main conversation, such as subagents (helper agents Claude starts for part of a task).

The near-limit notice is handled differently:

• The near-limit message is now tagged usageLimitNote: "near_limit", where before it carried no tag.
• The code that finds the last usage-limit note now returns which note it was, instead of only whether it was wrap_up.
• A near_limit note is now treated like wrap_up: it prevents a reset and can trigger the "release" message. That release step is still controlled by tengu_lantern_wick_braid and related settings such as tengu_lantern_wick_mode.

Nothing has been read about any of these settings, so whether they are active is unknown.

Why

This changes which grace-window messages the model receives when you are close to a usage limit, and lets the wording differ between models and between the main conversation and subagents.