src/compact/truncating.rs

This file holds your default shrink strategy, which is dependency-free and makes no LLM calls. It also holds the exported splitter you use in custom compactors.

Key items

TruncatingCompactor { preserve_recent: 4, min_messages: 6 } comes with builders clamped to minimums 1 and 2.

TokenSplitter takes the same knobs. Rules for the split come from the same pair. Your split(&messages) -> SplitResult { to_compact, preserved, token counts, split_index } call returns what to compact and what to keep. The budgeted variant split_within_budget, which both LLM summarizers use, joined in 0.4.1.

Tool-pair machinery stays private: PartMate and ToolPairing give you per-occurrence call and result pairing, orphan-safe split adjustment, and first-message result reattachment.

Behavior notes

Conversations under min_messages pass through untouched, and the first message is always kept for you.

Tool call and result pairs are never orphaned, because the split point moves and a fixpoint loop rechecks newly admitted messages. Your compactor never returns an empty list.

The target token budget is a hint here, since the truncator sheds by message count and your manager’s fit check decides adequacy.

TokenSplitter splits only at assistant-to-user turn boundaries that are pair-safe. If there is no safe boundary, everything is preserved with split_index: 0, which is the safe direction.

The budgeted variant falls back to the deepest pair-safe cut when no tail fits, never the preserve-almost-everything shape (0.4.1). A history with sparse boundaries used to compact to a no-op every pass and drift into the emergency stall.

Only a history with no pair-safe cut at all keeps the count-based default.

The budgeted variant also pulls Durable receipts within the landing’s budget, through durable_pull_indices_within_budget (0.4.1). Newest pairs ride first, and a pair too costly to fit is skipped. The walk is occurrence-aware, so a reused call id pairs with its own exchange, and the truncating terminal and both summarizers share it.

Even your no-change paths filter orphaned lone results and report the savings honestly.

For the deep dive, see Compaction.