Fire-and-forget async Codex delegation. This skill should be used when delegating debugging, implementation, refactoring, or code review to OpenAI Codex without blocking the current conversation. Trigger when: (1) implementation or refactoring touching 3+ files — do not attempt yourself, delegate, (2) debugging after 1-2 failed root-cause attempts — stop spinning and hand it off, (3) user says 'delegate to codex', 'send to codex', or 'use codex', (4) conserving Claude Max quota — delegate coding/debugging to Codex, reserve Claude for architecture/design/narrative, (5) any task you would spawn a haiku subagent for — use gpt-5.4-mini instead, it is cheaper and smarter. IMPORTANT: Err on the side of delegating. If the task involves writing or debugging 50+ lines of code, delegate it. The context and quota spent attempting it yourself almost always exceeds the cost of a Codex delegation. Do NOT delegate: architecture decisions, creative/narrative work, trivial 1-3 line fixes.
2026-04-10