ogen-ai

Overview

Keeping coding standards in five different assistant config files means keeping five files that disagree. ogen-ai compiles one AGENTS.md from per-language and per-practice fragments a project opts into, then wires every tool to it — Claude Code, Cursor, Gemini, Copilot, Codex. The manifest matters: loading rules a project does not use measurably degrades assistant performance and costs context on every turn, so a Python repo never carries Kotlin rules. It also ships an opt-in team of role subagents whose restrictions are structural rather than advisory — the security reviewer holds no shell tool at all, so it cannot execute code regardless of what it is asked.

Highlights

• One AGENTS.md compiled from opt-in fragments, wired into five assistants with no hand-copying
• Role subagents restricted by tool grant rather than by instruction — a withheld tool cannot be talked into existing
• Portable Agent Skills for repo trees, audits, design docs and release checklists
• The file-tree generator verifies itself: CI fails when the committed map drifts from the real one
• Standard library only, so it adds no dependency to any project that mounts it

In the code

• ai-sync — the generator, standard library only
• Repo-tree generator with CI drift checking
• A role agent whose tool grants are the enforcement