fable5-model-routing
Fable 5 Model Routing
Assign a model and a reasoning effort to every subagent a Claude Fable lead spawns. Fable means Claude Fable 5 or Claude Fable 5.1; Mythos 5 and 5.1 behave the same for this purpose. The scope is the lead model, not the harness: the same policy applies wherever a Fable lead can delegate, and the harness only changes how a choice is expressed. references/harness-adapters.md holds those mechanics, with Claude Code, Cursor, Hermes, and proxy-routed subagents as worked examples and a procedure for any other harness.
When It Applies
- Fable leads. The policy stands for the whole session. Each time the lead decides to delegate, the task gets a model and an effort chosen for its tier, stated before the spawn. If a worker may delegate further, pass the applicable family policy and the same authority and resource limits to its children.
- Another model leads. Apply on request, when the user names this skill or asks to put Fable on the judgment. Bring Fable in as a judge only if the harness can actually run Fable as a subagent and the problem has a consequential judgment core: a material trade-off, hidden premise, conflicting evidence, or a recommendation someone must defend. A hard-looking task is not that.
- A harness that cannot vary subagent model or effort. The skill still applies. Decide whether the inherited settings are acceptable for the tier, and say so. Do not pretend a setting was applied.
Read the lead's identity from what the harness states. If it cannot be established, treat the session as non-Fable.
This skill does not decide whether to delegate; orchestrate-subagents and the lead's own judgment do. An explicit user instruction about delegation, such as not delegating, a cap, or a model, overrides the tier table; without one, delegating on the lead's own judgment is expected. It also does not write the worker's prompt; small-model-handoff does that for a weaker executor when a pack skill calls it.
Anthropic Routing
This skill owns Anthropic model choices. GPT alternatives follow gpt6-astra-model-routing: Astra for judgment-bearing work, Sol 6.1 only for very easy deterministic execution. The two policies share orchestration principles, not mirrored model tables.