Model-agnostic operating config for AI coding agents.

One canonical rule set, shared across Claude Code, opencode, and any other agent that reads an instruction file at session start. Symlinked live from my NixOS config (universe, modules/home/dotagents.nix) via mkOutOfStoreSymlink, so edits apply instantly with no rebuild.

• AGENTS.md - the canonical rules, including the always-on caveman and ponytail modes; CLAUDE.md is a symlink to it so every agent shares one source
• claude/ - Claude Code tooling: settings.json, hooks, statusline, usage script
• opencode/ - opencode config (opencode.json)
• skills/manifest.txt - the skill set, installed for both agents by universe's skills-sync timer via bunx skills

Plugin-free by design: no Claude-only plugins. Behavior rules live in AGENTS.md (both agents read it), on-demand tooling comes from skills. Nothing is tied to one agent.

Agent tools auto-load one instruction file at session start, by filename: CLAUDE.md (Claude Code), AGENTS.md (opencode, Copilot, Cursor, Zed; the cross-vendor standard). Canonical rules live in AGENTS.md for the broadest support, with CLAUDE.md symlinked to it.

Global locations, all symlinked to this AGENTS.md by dotagents.nix: ~/.claude/CLAUDE.md, ~/.config/opencode/AGENTS.md. Same filenames at a project root override these globals.

MIT, see LICENSE. This is a personal repo: read it, fork it, open an issue. Pull requests are not accepted, see CONTRIBUTING.md.