bld-runtime-agentslisted

# bld-runtime-agents — Claude is the boss, Codex is the executor The user's request comes after `/bld-runtime-agents`. **Once this skill is invoked, Codex writes the code — all of it.** Claude is the manager: it holds the taste, the context, the design, and the requirements, and spends its window on those. Every token Claude spends typing implementation is a token wasted twice — once on the typing, once on the context it displaces. ## ⛔ "Agents" here means OTHER companies' agents — never Claude subagents This skill exists to spend **someone else's quota**. Codex bills ChatGPT Plus; Gemini bills its own free tier. **That asymmetry is the entire point.** **Spawning Claude subagents is NOT delegation** — it spends the same budget the main window does, except each subagent carries its own full context, so it costs *more*, not less. If this skill is invoked and the work goes to `Agent`/`Workflow` instead of the `codex` CLI, the skill has been misused. Hard rules (full rationale + the 939k-token incident that produced them is in the `~/.claude/CLAUDE.md` subagent-cap section): - **Max 2–3 Claude subagents at a time, only when genuinely necessary.** - **Never tell a subagent to invoke a heavy skill** (`impeccable`, `ui-ux-pro-max`, `emil-design-eng`, `gstack-spec`) — that multiplies a large document across every agent's context. - **Never fan out one verifier agent per finding.** Review the list yourself against the files; it costs a fraction. - Order of preference when