Pair-safe prompt-too-long fallback — ptl

Drops complete historical API rounds until the estimated request releases the provider-reported token gap. Without an actual/limit pair, targets a 20% reduction. The newest round and pending turn are always retained.

Usage

ptl_fallback(chat, drop_turns = 3L, error_msg = NULL, pending_turn = NULL)

Arguments

chat: An ellmer::Chat object.
drop_turns: Deprecated compatibility argument; raw turns are never dropped independently.
error_msg: Character or NULL. Provider PTL message.
pending_turn: Optional pending turn, retained outside persisted history.

Value

A sanitized internal recovery decision.