Context & Compression

This content describes a system that no longer exists

Context compression was a feature of the old external-LLM routing layer — sliding-window pruning, summarization, hybrid strategies, history-snip, max-output recovery. None of that applies to the Divinity Engine handlers that replaced the LLM layer in 2026-04.

If you arrived here from an older link, here’s what’s true now:

• Every Ishvana agent feature runs on local handlers with deterministic inputs and outputs. There’s no token budget to compress around because there’s no token billing.
• Long-running analyses (manuscript scans, full-project consistency checks, Lore ML training) handle scale through streaming progress chunks, not through context pruning. See Etherforce overview → Streaming results.
• Per-content-hash caching means re-running the same analysis on unchanged content is free, regardless of size. See Etherforce overview → Handler-result cache.

Etherforce overview The dispatch layer that replaced the external-LLM routing layer.

Engine overview How the Divinity Engine fits with the rest of the backend.

This page stays here as a redirect. The conceptual content from the original Context & Compression page has been retired because the system it described is no longer in the product.