BEYOND THE WINDOW
Every model has a fixed attention window. Your codebase does not fit inside it. Everything below is how we stopped pretending it did — and how a repo of any size became addressable in a single turn.
What runs
every turn
Addressable Index
Your repository stops being a paste and becomes a live vector space. Indexing is incremental — a one-file edit costs one file, not a re-read of the tree.
Sub-Quadratic Retrieval
Attending over a whole codebase is O(n²) and collapses at scale. Ranked retrieval reaches the same corpus for a fraction of the compute, on every single turn.
Context Compilation
A per-turn token budget assembles only the slices that matter. The window carries signal instead of bulk, so the model spends its attention where it counts.
Enforced Grounding
Rules that live on the server and cannot be switched off by the client: cite a real file:line, or abstain. The agent does not get to guess.
CONTEXT
IS NOT A WINDOW
IT IS AN INDEX
THE LAYERS
Four layers sit between your prompt and the answer. Every one of them runs on every turn, server-side, whether or not the client knows they exist.
Layer INDEX
Your repository, chunked and embedded into a live vector space that only your key can read.
STATUS: ONLINE
SCOPE per-owner · SYNC continuous
Layer RETRIEVAL
Semantic and lexical ranking, fused, across the whole corpus — not the fraction that fits.
STATUS: ONLINE
ACCESS sub-quadratic · MISS FAILS OPEN
Layer CONTEXT PACK
A token budget compiles the retrieved slices into the window the model actually sees.
STATUS: ONLINE
BUDGET per-turn · PAYLOAD signal only
Layer GROUNDING
Server-enforced citation. A claim carries a real file:line or it does not ship.
STATUS: ENFORCED
OVERRIDE none · CLIENT cannot disable