SuperAPI AI Directory

codex-model-routing-team

Agents & models

View repository

Updated 10-16 01:56

About this skill

Codex model routing team

The main Agent keeps its model and is responsible for integration and acceptance; independent execution is delegated to 1–3 Workers. With two or more Workers, first compile a TeamPlan; execution with clear boundaries and verifiable acceptance defaults to GPT-6 Luna XHigh, complex execution to Luna Max; undecided judgments use Sol Medium/High, high risk uses Sol High, critical independent review uses Sol XHigh. Before dispatch, declare task_contract: whether the decision is already made and how to verify acceptance; long prompts or many files do not by themselves trigger Sol.

Select paths as needed

  • Default native Standard, new standard team, and units that can share risk routing: read Native concise path, and use scripts/prepare_native_team.py to generate verified dispatch parameters and the initial ledger in one pass.
  • Different routing, Fast, single Worker, or restoring an existing team: read Compiler interface; for two or more Workers, validate according to TeamPlan , and for native execution read Lifecycle。
  • Persistent Task explicitly required by the user: after verifying host authorization, read Thread lifecycle and Supervision protocol. A worktree requirement by itself does not authorize creating a Task.
  • User names an exact CLI model: read CLI Surface, and verify help/catalog, authorization, and fresh context.
  • When an upstream Skill has already defined the split, phases, and artifacts, follow Adaptation protocol, and do not redo phase gates or the business ledger.

Shared by all paths

  1. Automatic dispatch requires 2+ independent deliverables and a net-positive return; a named single Worker may execute, otherwise the main Agent completes it directly.
  2. The compiler reads the registry and performs RoutePlan validation, without omitting Provider, live schema, or data boundaries. Expand only for special Provider/overrides or failure diagnosis Provider、Routing、Surface; do not read everything first every time.
  3. Before dispatch, clarify the unit, unique task_id, permissions, exact ownership, and acceptance; the complete Task package is expanded only when input/context/permissions are complex. The main Agent retains integration and final acceptance.
  4. Default standard 6/8/3, actually constrained by live child slots and stricter user limits;expanded 12/16/6 is enabled only through the TeamPlan capacity gate. Each unit has at most 2 attempts and one follow-up; failures only follow pre-declared chain。
  5. Results are adopted after verification by the main Agent; formal status confirmation closes out and runs scripts/validate_team_ledger.py. Compilation readiness does not mean it has been executed; do not treat a creation receipt as completion.

Hard gate

  • The registry determines scope; live schema only proves that the current host accepts the exact combination. Record requested/accepted/observed separately; if not echoed back as unknown。
  • The official catalog for GPT-6.1 Sol / GPT-6 Luna is V2; this Skill restricts all Workers to leaf execution roles and prohibits further dispatch. The current host must accept the exact model combination.
  • Do not add model: luna frontmatter; the orchestration entry point stays with the collaborating parent Agent, and Luna acts only as a Worker.
  • Luna minimum XHigh; Sol minimum Medium; Terra only as an explicit first item; Grok passes the gate; Gemini 3.6 blocked. DeepSeek 4.1/Gemini 3.8 only via manual/experimental CLI. Ultra is prohibited.
  • Fast means service_tier=priority;When live schema has no fields, always use Standard, and never pass off catalog or request values as observed Fast.
  • app_thread Only for worktrees, sidebar, cross-task recovery, durable supervision, or predeclared fallback, and there must be live capability and host authorization evidence.
  • Workers must not continue deriving or executing releases, sends, payments, deletions, account actions, or production changes; the main Agent does not switch models.
  • TeamPlan does not create a Planner, does not call heavy planning, and does not write persistent files; same-wave write conflicts, dependency cycles, over-budget, out-of-plan Workers, and delegated acceptance must be rejected.
  • Unconfirmed return values or pendingWorktreeId must not be treated as formal identity;UNKNOWN follow-up questions, archiving, fallback, duplicate creation, and database changes are prohibited.

Output contract

Delivery must be verified by the main Agent, including auditable Surface, model, reasoning, speed, context, Provider gate, attempts, fallback, adoption, and closing status. See the compiler at RoutePlan compiler interface; for boundaries see Validation cases and evals/。