src/engine/bare/prune.rs

This file holds the opt-in prune, the no-LLM reclamation pass. Old tool results beyond a protected tail are cleared to [cleared: prior tool output] markers without any summarizer call, so you free context cheaply.

It runs at the pre-request site, before the compaction deferral.

Key items

PruneConfig carries your protected tail token budget, which defaults to 2,000, and the minimum-free gate, which is a fraction of the window defaulting to 10%.

That gate is evaluated in the counter’s raw units, so a settled calibration ratio cannot distort the verdict you get.

BareLoop::set_prune(config) is opt-in, and default absent means off. Your prune line must sit strictly below the compaction threshold, which is validated with InvalidInput and never clamped.

Behavior notes

Clearing order goes untagged results first, then Requery only when the untagged pool cannot satisfy the minimum-free gate, and Durable never.

Identity survives the marker, meaning call id, tool name, error flag, and retention class, and clearing runs oldest first and deterministically.

The originals hand off to your demotion sink under CompactReason::Prune, so pruned is not destroyed.

A prune that frees enough can avert the compaction it would otherwise defer to. The machine then adopts the rewritten conversation through LoopMachine::adopt_pruned, which preserves your history and pending split and performs no state transition.

For the deep dive, see Context accounting and prune.