OpenAI Agents SDK 0.19.4 Redacts Leaked Tool Arguments

OpenAI shipped 17 fixes to its Agents SDK on August 5, one day after version 0.19.3, per the release notes on GitHub.

Invalid tool-call arguments used to echo straight into error logs, meaning attacker-controlled arguments could land in your logs unredacted. Version 0.19.4 fixes that.

Anyone running the SDK with output guardrails and logging on should pull this patch before the next deploy.

Four other fixes untangle races between guardrails and session handling, the kind of bug that only shows up under real concurrent load.

The SDK now defers non-streaming session saves until output guardrails finish running, so a session can't get persisted before it's actually been checked. It also preserves completed tool guardrail results instead of discarding them mid-run.

It honors a request's resolved status before running policy checks. And it cancels sibling tasks after a concurrent failure, instead of letting them keep executing.

Landing this many ordering fixes a single day after 0.19.3 reads like OpenAI found these guardrail-session races in production, not in a review pass.

0.19.4 shipped less than 24 hours after 0.19.3, so if you're already running the SDK, you're likely still on the version with the logging gap.