Token count preferring real usage over the char heuristic

Mirrors Claude Code tokenCountWithEstimation (src/utils/tokens.ts): use the real token usage from the last API exchange when available, otherwise fall back to the char/3.5 estimate. This makes the compaction trigger fire on actual model token counts rather than a rough character approximation.

Usage

token_count_with_estimation(chat, allow_network = FALSE)

Arguments

chat

An ellmer::Chat object.

allow_network

Logical. If TRUE, explicitly call chat$token_count(include = "complete"); defaults to FALSE so UI and compaction paths never perform implicit token-count HTTP requests.

Value

Integer token count.