OpenAI Agents Python conditional approval and Sume dry_run

The v0.22.3 release of openai-agents-python lists a fix to "align conditional approvals with validated tool arguments". For Sume tools the practical rule is the same: decide whether to ask a human using the arguments the tool contract accepts, and use dry_run to get the cost before the question is asked.

The release notes only give the one-line title of that fix (PR 5066), so this post does not describe how the SDK implements it. The Sume side is from MCP tools and gates, read 2026-10-01.

What does the v0.22.3 release say?

Under "What's Changed" the first entry is "fix(core): align conditional approvals with validated tool arguments". That is the whole description in the notes. If your approval predicate runs on tool arguments, read the linked pull request before relying on a specific behavior.

Which Sume fields should a conditional approval read?

Sume publishes one contract per tool through tools_schema (fetch by name), so a predicate can be written against the same fields the server validates. The gates that matter for a paid call are below.

Does idempotency_key count as approval?

No. The docs describe it as a key for transport and dedup, not human approval. An approval step in your agent is a separate decision, and the key should be a stable value you generate once per intended create so a retry does not become a second paid job.

How do I show a cost before asking?

Call the create tool with dry_run=true first. It returns an admission and cost preview without submitting. The docs also say to prefer generation_admission_preview and/or dry_run before expensive bursts, and that ordinary single creates do not need that extra step. So a reasonable condition is: ask only when the call is a paid write and you are about to fan out several of them.

For the broader approval pattern, see OpenAI Agents SDK `require_approval` for Sume write tools.

What should I do after upgrading?

Re-run any approval tests that pass arguments the model might get wrong, such as a missing field or a wrong type. Check that the predicate sees the arguments in the shape the schema defines, and keep the idempotency_key out of the decision so retries behave the same as first attempts.