Compaction
Sign in
Documentation

Hybrid optimization

Hybrid optimization is the input side: it shrinks a live session’s conversation history on-device, before the request leaves your machine. Long sessions stay under the context ceiling instead of dying at it.

This is what Community adds on top of Open. It is free, and it needs three things: a free Community account, the adaptive engine (delivered separately and installed for you when you activate Community), and compaction mode full.

How it works

  • Deterministic protection first. Code blocks, shell commands, and file paths are locked verbatim — protected spans are preserved byte-for-byte, structurally, not by model judgment.
  • An adaptive engine for the rest. Conversation history that has gone stale is summarized by an engine that runs on your machine. Nothing is sent anywhere to be optimized.
  • Live, not static. It operates on the live session through the gateway — which is what extends usable session length — not only on exported trace files.

Measured effect

Measured on real end-to-end gateway sessions, provider-reported: ~48–50% less billed input on uncached API sessions, and ~5–10% less input on cached sessions (a cached session is already cheap on the input side, so the headroom effect — a longer usable session — matters more there). See Proof and receipts for how each figure is labeled.

Turning it on

Run compaction and choose Community, or from an existing install:

compaction login      # free Community account, 1 device
compaction mode full  # adaptive input + output optimization

Shipping the capability never activates it — observe and basic never touch the input side, and full needs a valid entitlement on this device.

  • record mode (the default) never changes a request
  • whenever compaction changes a request, the original body is retained locally — compaction gateway recover <id> returns the exact bytes
  • every change is receipted with before/after counts and the evidence label for each figure

The Community allowance

Community includes an optimized-input allowance per period. Check what this device has used with compaction usage — it reads local disk only. When the allowance for the period is used up, input optimization pauses and says so on the per-turn line; output shaping keeps running and your requests keep going through unchanged otherwise. Nothing breaks and nothing is billed.

Boundaries

  • on-device only — the engine runs locally and your history is never sent anywhere to be optimized
  • protected spans (code / commands / paths) are never summarized
  • recoverability is the guarantee (byte-exact); whether removed context mattered stays a human call