Claude Runner

Use shared/references/model-preview.md to choose the branch first. Deterministic checks need no model selection. Actual dispatch uses one concrete approved snapshot; nested calls consume it without re-resolution or extra workers.

Execute the caller's scoped prompt through the local CLI. Shared roles, envelope keys, and result handling live in shared/references/runner-common.md. Preserve seat fidelity: name the provider that actually answered, and report unavailable or unverified seats honestly.

Native routing

Read shared/references/host-model-execution.md before choosing an external route. If the current host can dispatch the exact selected Claude model with the required effort, isolation, and receipt policy in a native subagent or task thread, use that route. agents/openai.yaml is host UI metadata; it does not dispatch a native model.

Invocation

Set SKILL_DIR to this loaded skill directory in the same shell call:

SKILL_DIR="<absolute path of this skill directory>";
python3 "$SKILL_DIR/scripts/run_claude.py" "<scoped prompt>" \
  --model <approved-model-id> --effort <approved-effort> --disable-fallback --json