I built a context-router skill to keep AGENTS.md focused on each task

Coding-agent instruction files can turn into catchalls: old milestones, runbooks, and unrelated rules end up beside the few things every task needs. Then even a small change carries that extra context, and the important rule is easy to miss. I built agents-context-router to help split root instructions into a small always-loaded kernel and task-specific wiki pages. It includes a zero-dependency Python CLI to list, load, and check topics, plus checks for orphaned pages and size budgets. It works with Claude Code, Codex, Cursor, and Gemini CLI, with native plugins for Claude Code and Codex. I evaluated it on three repo-maintenance tasks: 33/33 checks with the skill vs 21/33 without. That is one run per task per revision and costs about 25 seconds and 6k tokens more per run, so I would treat the result as directional, not proof. Repo: github.com/runkids/agents-context-router I would love feedback from people who maintain agent instructions: would a small always-loaded kernel plus topic pages work for your repo? What would make the checks easier to trust?