[jeffdupont]

Fixes AIC-3586.

Problem

openai-agents builds each tool with @openai/agents' tool() and leaves errorFunction unset. When it's unset, the SDK catches an error thrown by execute() and returns "An error occurred while running the tool. Please try again. Error: ..." to the model as the tool result. So a failing tool handler never fails the run: the model gets the error text and keeps going, and the caller never sees the error.

The shared spec says the opposite: TESTING.md §1.4, "The handler must propagate the error (not swallow it silently)." Python and Go already propagate.

Fix

• errorFunction: null on both tool() calls: src/handler.ts and src/native-graph.ts.
• New test in handler.test.ts and in native-graph.test.ts. Each passes the handler's arguments to the real tool() and checks that invoke() rejects. The existing test mocks tool() and calls the raw execute, which skips the SDK's error wrapper, so it couldn't catch this.
• agents.md: the tool example shows errorFunction: null, plus a pitfall note on why it's needed.

Testing

• All 126 openai-agents tests pass; tsc and biome are clean.
• With only the fix reverted, exactly the 2 new tests fail: promise resolved "An error occurred while running the t…" instead of rejecting.

Not covered: a full run() against a fake model. The tests stop at tool.invoke.

🤖 Generated with Claude Code

---

Note

Overview
Sets errorFunction: null on every @openai/agents tool() built in the single-agent handler (buildAgentTools) and the native graph adapter (buildNodeTools). Without it, the SDK turns handler failures into a tool result string for the model and the run continues; with it, errors reject per TESTING.md §1.4 (aligned with Python/Go).

Adds integration-style tests in handler.test.ts and native-graph.test.ts that rebuild tools with the real SDK tool() and assert invoke() rejects—mock-only execute tests miss the SDK wrapper. agents.md documents the option in the wiring example and a common pitfall.

^{Reviewed by Cursor Bugbot for commit 82d1fc3. Bugbot is set up for automated code reviews on this repo. Configure here.}