| name | prepare |
| description | Visible Harness prepare entry. Use after Grill readiness is approved to acquire datasets, clone baselines, run baseline smoke checks, and complete prepare through workflow-supervisor. |
Prepare
Use $prepare as the human-facing alias for the Execution Supervisor
prepare action. This is not a separate Skill Contract.
Read and follow:
../../../.agents/skills/workflow-supervisor/SKILL.md../../../.agents/references/workflow-supervisor-runtime.md../../../AGENTS.md../../../CLAUDE.md
Default command:
python tooling/workflow_supervisor/scripts/workflow_ctl.py start --segment prepare --complete --json
Use Grill-produced approved readiness from .workflow_supervisor/readiness.json
when it exists. Pause only for missing inputs, policy blockers, worker
failures, or gate failures.
Before the default command, inspect Grill artifacts and readiness for a
Model Weight Ledger, model_weight_* policy rows, hf_model_access_policy,
and target_paths entries such as model_cache or model_<id>. Pull or
verify only candidate/approved weights with explicit source, access policy, and
target: exact local path, Hugging Face model id/repo via huggingface-cli download or git clone when available, or official checkpoint/release URL.
Use models/<id> only when no target is specified and the policy allows it.
Do not guess, download deferred/rejected/approval-required weights, or bypass
gated access. Report weight acquisition Gate ledger entries and artifacts
alongside the prepare command result.
