Pydantic AI ToolCallJudge: check a Sume paid call before it runs

Add ToolCallJudge from Pydantic AI v2.53.0 (2026-10-02) in front of Sume's paid tools and make the judge check that the call carries a spend ceiling. The release describes it as a way to assess tool calls before execution, which is the right place to catch a missing max_spend_usd.

The judge reduces mistakes; it is not a guarantee, so Sume's own checks stay in the loop.

What 2.53.0 includes

The release adds ToolCallJudge, an AskUser capability that defers to the host with a timeout, and managed subagents in CLAI2.

At a glance

What the judge should look for

For Sume, the useful checks are mechanical: the call has an idempotency_key, max_spend_usd is set and within your budget, and dry_run was used first for anything new. Sume will reject a write or paid call that lacks idempotency_key, but it does not require the ceiling.

Wire the Sume server as an MCP toolset, then let the judge see the tool name and arguments.

• Reject paid calls without max_spend_usd.
• Allow read tools without review.
• Use AskUser for anything the judge flags.

Limits and what is not verified

I did not run ToolCallJudge against a live Sume server, and the release notes do not describe the judge's interface in detail. Use Pydantic AI's docs for the code.