Ctxfw – In-memory Tree-sitter AST compactor that cuts coding tokens by 72%

Ctxfw is an open‑core, in‑memory Tree‑sitter AST compactor and Model Context Protocol (MCP) gatekeeper designed to cut LLM coding context bloat. It prunes syntax trees polyglotically (Python, TypeScript/JavaScript, Go, Java) by preserving the target file in full, reducing distance‑1 dependencies to signatures and docstrings, and collapsing distance‑2+ dependencies into compact type stubs. Benchmarks claim a reduction in token consumption from 100% to 27.6% (72.4% bloat eliminated). A deterministic axiomatic intake sieve enforces specification rigor (aiming for ACI ≥ 0.9000), rejecting briefs missing negative invariants (N ≥ 5), bounded variable domains, deterministic state machines, or a 4‑class error taxonomy. Operation is local by default with zero telemetry egress.

Core subsystems include a native in‑memory parser and a SQLite WAL high‑concurrency semantic cache (PRAGMA journal_mode=WAL, busy_timeout=5000ms) for sub‑millisecond warm hits. Deployment features include zero‑touch provisioning and IDE/global init commands, repo pre‑commit sentinels to block uncertified commits, and an MCP server. Enterprise controls add team token/dollar circuit breakers, multi‑node reverse‑proxy governance for streaming endpoints, a finops telemetry ledger with tamper‑evident audit trails, and a control plane/enterprise licensing option. Failure cases produce quarantine, execution halts, forensic reports, or verified SHA‑256 seals.