A cost-aware Agent Skill for routing GPT-6 Astra and smaller Codex models on ChatGPT Plus or Pro. It helps you choose between direct Astra, a lightweight coordinator with an Astra implementer, and Astra-led planning for ambitious work.
The attached “Idiot Boss” experiment motivates the main change: on two visual builds, a small coordinator that browser-tested one continuing Astra implementation worker was cheaper and faster by API-equivalent estimate than the tested alternatives, with much less quality loss than asking Luna to implement everything. This is not a verified Plus/Pro quota conversion or a general parity claim. Your own tasks may favor an Astra xhigh root.
npx skills@latest add Enixes/astra-efficient-orchestrationOr clone into ~/.codex/skills/astra-efficient-orchestration and start a new Codex session.
| Situation | Suggested route |
|---|---|
| Small, coupled, quality-first, or hard to evaluate | Direct Astra with required checks |
| Bounded visual/product build with browser-testable acceptance | Luna xhigh coordinator → one Astra low implementation worker; reuse its thread for focused repairs |
| Product QA too demanding for Luna | Terra/Sol coordinator → Astra worker |
| Ambitious multi-step architecture or consequential integration | Astra xhigh plans and resolves key decisions; delegate bounded execution only when useful |
Model selection and delegation depend on your Codex environment and permissions. A coordinator should pass the original brief and evidence to the capable worker, not prescribe technical design. It owns browser QA and the final handoff; essential verification remains assigned and completed.
For a bounded frontend build:
Use $astra-efficient-orchestration. If delegation is permitted, let a small coordinator give one Astra worker the original brief for the complete implementation. Reuse that worker for fixes; test in the browser and send screenshots plus concrete defects. Preserve required tests and stop when acceptance passes.
For quality-first or hard-to-test work:
Use $astra-efficient-orchestration. Work directly with Astra unless a bounded independent support task clearly helps. Complete required verification.
For your existing xhigh-root baseline:
Use $astra-efficient-orchestration with Astra xhigh at planning and integration. Delegate only bounded useful work; compare quality and visible usage before changing the route.
See model routing, usage profiles, handoff templates, prompt recipes, and evaluation guidance. Research basis separates experiment findings from quota assumptions.
Pair with Astra Frontend Design when visible product quality is a primary deliverable.