Design or review agent-friendly command-line interfaces for tools, wrappers, and internal platforms. Use when Codex needs to create a new CLI, simplify an existing one, map real workflows to commands and flags, improve `--help` and error behavior, define stdout/stderr and file output contracts, add safety rails for destructive actions, or turn an API/service/browser workflow into a shell-first interface that humans and agents can drive reliably.
Review a codebase, directory, or PR diff using John Ousterhout’s “A Philosophy of Software Design” principles. Use when the user wants design/maintainability feedback: complexity (dependencies/obscurity), deep modules, information hiding/leakage, interface design (common case), layering, error-handling strategy, comments and naming/obviousness. Do NOT use for formatting-only or lint-only requests, or when the user is explicitly asking for performance benchmarking, a pure security audit, or purely stylistic rewrites.
Delegate coding tasks to a nested Codex CLI run using `codex exec --yolo -m gpt-5.3-codex -c model_reasoning_effort=xhigh`. Use when you want the current Codex agent to act as an orchestrator (plan/review/verify) while a separate worker session performs edits and runs tests.
Lightweight code review for diffs/PRs focusing on high-signal simplifications, code smells, security, performance, and whether new tests are needed; focus on recently modified code unless instructed otherwise.