`/auto` decides judgment calls instead of asking you (hook-enforced, per session). `/auto away` (alias `/afk`) adds the unattended ladder and loop. `/carryon` one next action, stops on a decision. `/oneshot` idea to green PR. Everything else native: `/loop`, `/goal`, Workflow, agent teams. Say "afk" to start it, "deslop" for ai-slop-cleaner. Default process layer, here and in Codex. Let it lead: brainstorming before creative work, TDD, systematic-debugging, requesting-code-review, verification-before-completion. Follow the skill it picks, do not route around it. Announce one once when it fires, not every turn. The user's skills are on demand only: `/prdforge` for a written PRD instead of a conversation, `/done` for a full gate run, `/multiask` and `/codex:rescue` to escalate. Bare: architect, code-reviewer, critic, debugger, executor, planner, qa-tester, security-reviewer, verifier, Explore. Delegate multi-file work, refactors, debugging, reviews, planning, research. Do trivial ops and single commands yourself. Need a decision from the user? AskUserQuestion, one question, 2 to 4 options. Prose only for free-form answers. Placeholder TODOs, `test.skip`, `.only`, stub tests and unimplemented branches are blockers, not evidence. Check changed files for them before claiming done. Per-project agent state (handoffs, plans, research) lives in `.omc/`. Historical name, live directory, the HUD reads it. `CLAUDE_SKIP_HOOKS` disables one hook for one command.
In a repo with a .codegraph/ directory, run codegraph explore "<symbols or question>" before grep or reading files. It returns the relevant source plus the call paths between symbols, including dispatch hops grep cannot follow. No .codegraph/ directory means skip it entirely; indexing is the user’s call.