Skip to content

Repository files navigation

everywhere — GPU content-grep

everything finds files by name. everywhere finds what's inside them.

A stateless C++20/CUDA multi-pattern scanner for Windows: streams bytes from NVMe through a GPU-resident failureless automaton at near-drive speed, so pattern count is nearly free — 10,000 literals cost about what 10 cost. No index, no daemon, no staleness: every run reads the real bytes, so every answer is true at the moment it prints.

Status: spec-first, pre-implementation (groundwork laid 2026-07-01). Nothing builds yet. Read EVERYWHERE_ARCHITECTURE.md (the spec) and CLAUDE.md (the working rules) before touching anything.

Where it sits in the organ family

organ finds
C:\everything files by name/path/date/size
C:\everywhere files by raw content (exact/multi-pattern/regex)
C:\chunker structure — splits what's too big to read
C:\imguard / C:\earshot modalities — images, audio
C:\cortex meaning — semantic memory over C:\Data

Chains: es.exe -ext md | everywhere "pattern" (narrow by name, scan by content) · everywhere --patterns groups.txt --jsonl (CORTEX's L0 exact-match pre-filter — zero tokens).

The contract (once built)

  • rg-compatible semantics on the supported flag subset — ripgrep is the pinned oracle; results diff byte-for-byte.
  • Targets: ≥80% of drive sequential cold @1k patterns · ≥15 GB/s warm (PCIe-4) · ≥500k small files/min · first hit <100 ms.
  • Streams results; exit codes 0/1/2 like rg; --jsonl for machines; a live throughput meter for humans.

Layout

  • EVERYWHERE_ARCHITECTURE.md — the spec (thesis, physics, anatomy, oracle, stages, refusals).
  • CLAUDE.md — session rules: oracle-before-engine, frozen goldens, language/platform discipline, falsifiers.
  • goldens/ — operator-frozen golden corpora + expected results (never agent-edited).
  • everywhere.lock — pinned oracle (rg version), toolchain, reference-corpus hash, perf baselines.
  • _run_state/STATE.md — live build state; trust it + git over memory.

Build & run (Stage 0 — CPU skeleton + ripgrep oracle)

Stage 0 is CPU-only (no CUDA yet): a plain multi-literal scanner graded byte-for-byte against pinned ripgrep. Prereqs: VS 2022 (MSVC 14.4x), CMake ≥3.28, ripgrep 15.1.0 on PATH (winget install BurntSushi.ripgrep.MSVC).

# configure + build the three targets (everywhere, ew-plant, ew-harness)
cmake -B build -S . -G "Visual Studio 17 2022" -A x64
cmake --build build --config Release

# generate the planted golden corpus (deterministic, seed 20260701)
build\Release\ew-plant.exe --out goldens\planted        # add --small 2000 for a fast corpus

# run the ripgrep-differential + property harness (green == exit 0)
build\Release\ew-harness.exe --corpus goldens\planted

The harness runs pinned rg and everywhere over every planted case across the supported flag subset, canonicalizes both to (file, line, byte-offset) sets, and fails red on any divergence with a minimal repro. It also runs the property tests (chunk-size / stream-count invariance, concat-union, idempotence) and a report-only perf smoke (the CPU baseline Stage 1's GPU must beat ≥10×). everywhere itself is an ordinary grep:

build\Release\everywhere.exe -n -i -e foo -e bar src\        # walk + literal search
es.exe -ext md | build\Release\everywhere.exe "pattern"      # organ chain: names -> contents

Groundwork by Claude Fable 5, 2026-07-01. Stage-0 CPU skeleton + oracle harness by Claude Opus 4.8, 2026-07-08. Build the harness first.

About

A stateless C++20/CUDA GPU multi-pattern content grep for Windows. 'everything' finds files by name; 'everywhere' finds what's inside them — at drive speed, pattern count nearly free (GPU PFAC automaton), byte-for-byte ripgrep-compatible, no index, no daemon.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages