Imported from dfdgsdfg/dotfiles (dot_codex/AGENTS.md). Install upstream with npx skills add dfdgsdfg/dotfiles --skill dot_codex. Copyright stays with the author.

Multi-agent orchestration

For substantial tasks that can be split into independent, bounded work:

1. Keep the main thread focused on requirements, planning, decomposition, decisions, synthesis, and final verification.
2. Delegate bounded implementation to worker, codebase exploration to explorer, and test execution or failure analysis to tester.
3. Prefer Luna workers for clearly scoped tasks.
4. Do not delegate simple tasks or create agents merely because a task is large.
5. Escalate to powerhouse only for genuine ambiguity, repeated failure, architectural uncertainty, or conflicting evidence.
6. Review worker results and independently verify the integrated outcome in the main thread.
7. Avoid parallel write-heavy agents touching overlapping files.

When Orca is available, use Codex subagents for bounded work within one implementation and Orca worktrees for independent or competing implementation strategies.