HUMΛN
Context Compact
Start thin. Crush the bulk. Fetch the original when it matters — with receipts, not a proxy.
What it is
Context Compact is HUMΛN’s named compression layer. It sits on the Agent SDK LLM choke point and on fat muscle / extension results — content-aware crushers for JSON, logs, and search lists — so models see intent, not dumps.
Magic by default
Every ctx.llm.complete / stream runs Compact with zero config. Fail-open: if a crusher errors, the original messages still go through. Control mode pins strategies or turns Compact off.
Muscles & extensions
Companion and agents inherit Compact on muscle and extensions.* payloads automatically. Builders do not write compression glue — the thin view re-enters history; the full blob stays in CCR.
CCR retrieve
Compression is reversible for the session. Call ctx.compact.retrieve(ref) (or the context_compact.retrieve muscle) when a decision needs the original. Org-local store, TTL — sovereignty stays with the workspace.
Control when needed
Disable with compact: false, pin strategies, or set a soft token budget from Autonomic Engine. Companion keeps crushers on and never appends terseness shaping — Exquisite Doctrine for chat quality.
Provenance
Every crush can emit a CompressionReceipt — tokens before/after, strategies, CCR refs, fail-open flag — alongside existing cost and provenance fields. No silent prompt rewriting above the cache boundary.