What problem does it solve? When orchestrating Claude Code with multiple AI providers (Codex, Grok, Claude, Qwen), it is hard to know which provider has remaining quota and which model to delegate work to. This Skill turns the injected routing context into an authoritative capacity snapshot so the orchestrator always delegates to the provider with the most headroom instead of guessing or exhausting a single quota. ## Core Features & Use Cases - Capacity-aware delegation: Ranks selected_workers by weekly and five-hour remaining quota, excludes exhausted providers, and enforces exact-model concurrency limits before launching Agent/Task subagents. - Dynamic model selection: Honors explicitly requested models via model_prefixes while enforcing a merged denylist (disabled_subagent_models) that overrides even explicit requests. - Worker lifecycle management: Defines rules for background fan-out, recipient reuse via SendMessage, independent verification of subagent results, and RAM-aware parallel caps on macOS. - Use Case: In a claudex orchestrator session, a user asks for a multi-file refactor; the Skill routes independent scopes to the top-ranked Codex and Grok workers in parallel, keeps the custom-advisor on a separate capacity channel, and verifies each worker's diff before reporting completion. ## Quick Start Ask the claudex orchestrator to implement a feature and let it automatically delegate the work to the highest-headroom provider subagent.
