Context Compact — start thin, fetch on demand

Context windows are not a strategy.

Teams paste connector JSON into the prompt, watch quality collapse, and blame the model. HUMΛN is the trust layer between people and AI — Context Compact exists so agents start thin and fetch on demand, under the same choke point as every other governed LLM call.

The dump that ate the turn

A muscle returns five thousand invoice rows. An extension dumps a log tail. The next ctx.llm.complete ships megabytes of redundancy into a model that only needed counts, schema, and a few samples.

Context Compact sits in LLMContextImpl — before Model Registry — and crushes that bulk with deterministic strategies (JSON schema + samples, log dedupe, search top-N). Fail-open: if a crusher throws, the original messages still run.

Magic for muscles and extensions

Builders do not opt in. Fat muscle and extensions.* results get compact_text plus a CCR ccr_ref for history re-injection. Companion keeps crushers on and OutputShaper off — chat stays exquisite; bulk still shrinks.

Retrieve when it matters

Compression without retrieve is truncation with a smile. Call ctx.compact.retrieve(ref) — or the context_compact.retrieve muscle — when a high-stakes decision needs the original. Session-local CCR, TTL, no second vault of identity.

Receipts, not vibes

Every crush can leave a CompressionReceipt on provenance: tokens before/after, strategies, CCR refs, fail-open. That is Principle Seven — verifiability — applied to context itself. Autonomic Engine’s suggested_token_budget soft-caps aggressiveness when present.

Not a proxy

There is no Headroom-style baseURL rewrite. No parallel LLM summarizer path. Distillate still compresses older turns. KB ContextReceipt still governs citations. CACHE_BOUNDARY still protects provider prefix cache — Compact never rewrites the stable prefix.

Start thin. Prove the crush. Fetch when the decision needs the whole thing.

See the feature page: /context-compact/ · Developer guide: /docs/guides/context-compact/.