Claude Code tool.check ceiling and agentId: gate paid Sume tool calls

If you run Claude Code with Sume's MCP tools, you can now write a plugin hook that treats a subagent's paid call differently from the main session's. Version 2.1.290, released October 5, added agentId to the tool.check event of plugin hooks, and added ceiling to the question and verdict a mod's tool.check hook reads, naming the approval an organization requires for a tool. This gives a hook the two facts a spend rule needs: who is asking, and what the policy ceiling is.

What changed in 2.1.290

The changelog also lists 2.1.291, released October 6, as a fix release for cloud sessions and session-end message loss.

• agentId on tool.check, so a hook can tell a subagent's permission check from the main session's.
• ceiling in the question and verdict, naming the approval an organization requires.
• claude plugin validate now lists each hook registered at a gating site and whether it has a .catch.

The Sume side of the rule

A sensible policy

Let the main session call read tools freely. For paid or write tools, have the hook require an approval whenever the caller is a subagent, and require that the call carries an idempotency_key and a max_spend_usd. A subagent loop is the usual way a bill grows unnoticed, because nobody watches its individual calls.

The hook cannot replace Sume's own limits. A permission check is a client-side decision, and it only covers sessions that run your plugin. Server-side, an OAuth session without write cannot spend at all, and an Agent Completion needs generation_spend_cap_usd or it fails. Treat the hook as the first fence and the server as the last one.

I have not tested a specific hook implementation here. The changelog gives the event fields, and the exact hook API is in Claude Code's plugin docs.