Agent Setup
cognitiveOS is agent-agnostic. init generates both CLAUDE.md and AGENTS.md (identical routing tables) plus an agent skill for whichever agents you pick, so your agent reads your context automatically. Nothing here is manual: you select agents in the 3-question wizard and init writes the right files.
What each agent loads
| Agent | Skill / rule it loads | Session-start hook |
|---|---|---|
| Claude Code | .claude/skills/cognitiveos/SKILL.md + slash hooks /start-session /end-session /dump | ✅ auto-loads STATE.md |
| Codex CLI | .codex/skills/cognitiveos/SKILL.md (+ AGENTS.md) | not yet |
| Antigravity | .agents/skills/cognitiveos/SKILL.md (+ AGENTS.md) | ✅ auto-loads STATE.md |
| Cursor / Windsurf | .cursor/rules/cognitiveos.mdc (+ AGENTS.md) | n/a |
Claude Code and Antigravity also get the deterministic session-start hook that loads STATE.md with zero typing. Codex and Cursor load the skill and AGENTS.md, so they pick up context on read, the automatic hook for those is an open community target.
The skill is the operating manual
The generated SKILL.md (or .mdc rule for Cursor) is what your agent triggers automatically. It carries the zone routing, the one-task and ADHD response rules, and a silent work loop: read state, take one action, update, re-read. It's why the agent behaves the same across every supported tool.
CLAUDE.md = AGENTS.md
The two routing files are byte-identical on purpose. cognitiveos check detects drift between them, and check --fix regenerates CLAUDE.md from AGENTS.md (the source of truth). --fix never touches STATE.md or your own content.
initnever overwrites or deletes your files. Existing configs and skill files are kept and merged, not clobbered.
See also: Hooks · Getting Started.