What is an experimental runbook
A runbook is the design idea of expressing an operation as versioned structured intent instead of a free-form prompt. A mature implementation could make required inputs, assumptions, approvals, and expected receipts reviewable. The legacy lyth-mcp templates are only fixtures for that design.
Today there is no released contract for drafting, validating, preparing, signing, or submitting an economic runbook. Legacy tool names, status fields, hashes, local receipts, and client-side policy checks are not consensus guarantees. In particular, no supported runbook may pay a vendor, request funds, book a service, open escrow, trade, change spending policy, or revoke economic authority.
Intended safety properties
Section titled “Intended safety properties”A future release must bind one reviewed intent to exact transaction bytes; reread and simulate current state; enforce limits in protocol state; use a dedicated key rather than a primary wallet; expire and revoke authority safely; prevent replay and concurrent cap races; and reconcile a durable receipt. Human approval alone does not fill those gaps.
Current safe use
Section titled “Current safe use”Treat runbook files as non-executable design examples. Agents may use read-only tools and prepare human-readable context, but must stop before transaction construction, signing, submission, booking, payment, escrow, dispute, or review.