The sovereign, zero-retention whole-codebase intelligence engine. Ingest 2M+ tokens in volatile GPU memory in 1.4 seconds—executing non-local architectural reasoning in a single forward pass with 0 tool calls.
Plug WholeRepo into your existing pipeline with standard OpenAI-compatible endpoints, an instant zero-dependency CLI, or native TypeScript / Node SDKs.
Press "▶ Run Request" above to simulate instant whole-repo forward pass...
$ wholerepo ask "Why does the SVG layout measurement skip embedded foreignObject?"
WholeRepo discards persistent vector databases and agentic search loops. Four mathematical breakthroughs make real-time multi-million token reasoning physically possible.
Standard attention thrashing chokes the GPU memory bus by loading 2M token KV caches to and from DRAM. Our custom Triton Riemannian kernel projects attention queries along geodesics, pruning unvisited token manifold shells in SRAM before they ever hit the memory controller.
Long-horizon attention collapses as tokens drift in numerical precision. We map positional states into symplectic Hamiltonian phase-space coordinates $(q_i, p_i)$ preserving volume on invariant manifolds.
Standard 16-bit KV caches consume 2,304 bytes per token. WholeRepo segments the attention latent into 128-element micro-tiles quantized in FP8 (E4M3), compressing state by 4x without accuracy loss.
For codebases exceeding 50M tokens, WholeRepo streams raw repository chunks through an ephemeral rolling KV merge matrix. Chunks are unified into a sovereign latent representation and instantly vaporize from physical VRAM upon session completion.
Pay strictly for active GPU cycles in volatile memory. No monthly recurring vector DB fees, zero storage retainers, and 100% predictable credit consumption.
Includes 100 Frustum queries across your 3.0M token codebase.
WholeRepo tested across Linux Kernel, Ladybird Browser, PyTorch, and Chromium ASTs.