Policy gates for autonomous per-inference x402 spending?

[felixpg13-glitch]

Hi — ClawRouter's premise (agents autonomously signing transactions for inference payments) is exactly the kind of spend flow we've been testing.

We built SpendShield, an open-source authorization layer that sits between an agent and a payment rail. We just ran it through a single autonomous spend session with one policy, using real engine output:

• normal order -> ALLOW
• over single-transaction cap ($75 vs $50) -> DENY
• larger transaction -> APPROVAL -> human approval -> ALLOW
• cumulative daily budget ($55 + $10 > $60) -> DENY
• policy changed mid-session -> versioned, auditable, and the same request re-evaluated under the new policy
• bypass attempts with no grant / forged grant -> rail REFUSED (MALFORMED_TOKEN / INVALID_SIGNATURE)

The run also produced a tamper-evident hash chain covering authorization decisions, approvals, and policy lifecycle events.

SpendShield never holds keys or money; it only decides whether a payment is authorized before the signing/payment code runs.

For a router where an agent can spend autonomously on repeated inference calls, I'm curious whether you see two separate control layers being useful:

1. per-call price limits
2. cross-call / session-level spend budgets

If this feels relevant, I'd be happy to sketch or contribute a small reference integration.

Repo: felixpg13-glitch/spendshield — full dogfood output is in docs/.