project-change-router-skill

In large monorepos, AI coding agents often misplace shared logic into facades or temp dirs, or spawn a second parallel implementation center next to an existing capability, all because they reason from local file names rather than true ownership boundaries. Limited context windows, attention drift, and unstable early-stage boundaries each compound the problem. project-change-router-skill intervenes before any edit: it loads a repo-local routing bundle to surface capability ownership, canonical roots, read/write boundaries, and reuse risk, so the agent guesses less and grounds more.

The skill orchestrates routing, freshness, dependency closure, public-API, reuse-scan, and governance checks through a single entry point, normalising results into traceable typed findings that a versioned policy table reduces to a single execution_gate (pass / conditional / blocked). Reuse detection runs three independent channels (intra-capability, cross-capability, new/extract/lifecycle); early and empty repos stay conservative by default. Safety envelopes, allowed/forbidden write paths, and veto reasons are projection-proof and cannot be excluded by CLI flags.

Built for teams shipping features in large multi-module, multi-owner codebases using Codex, Claude Code, or DeepSeek Harness. It does not replace reading source code or architecture review; it gives the agent an evidence-backed route, hard constraints, and calibration direction before every create, modify, or delete, catching structural drift before the first line is written.