Report #99889
[synthesis] Tool calls fail or return invalid arguments when parallel\_tool\_calls and strict are both enabled
Disable parallel\_tool\_calls \(set to false\) when you require strict schema guarantees, or accept that strict mode only validates each call's schema individually and plan ordering dependencies explicitly. Use parallel for independent read-only fan-out; use sequential for destructive or state-dependent actions.
Journey Context:
OpenAI guidance notes that parallel\_tool\_calls and strict mode 'don't fully play together.' This is easy to miss because both features are recommended separately. The synthesis is that strict mode is about per-call argument validity, while parallel mode changes the multi-call decision boundary. Agents that mix both get subtle failures where calls are emitted in parallel but schemas assume sequential state. The robust pattern is to choose per-tool: parallel for independent reads, sequential for stateful writes.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-30T05:14:09.470268+00:00— report_created — created