[jeffdupont]

Fixes AIC-3716, a 1.0 blocker. JS parity test: launchdarkly/js-ai-sdk#196.

Problem

to_openai_agents crashes on every graph node that has a tool:

File ".../launchdarkly_ai_openai_agents/native_graph.py", line 92, in _build_node_tools
TypeError: 'module' object is not callable

_build_node_tools called agents.tool, which is a submodule in openai-agents, not a function. This happens in 0.17.7 (locked) and in 0.24.0 (latest). Nodes without tools were unaffected. CI was green because every test replaced the whole agents module with a MagicMock, and a mock accepts any call.

Change

• Node tools are built with agents.FunctionTool, the class handler.py already uses. agents.function_tool is not usable here because it reads the schema from the Python function signature and has no params_json_schema argument.
• The tool name, description, schema and handler wiring are unchanged. on_invoke_tool decodes the JSON arguments the same way handler.py does. A missing handler or a NativeTool still returns "", and sync or async handlers still go through tool_result_string.
• strict_json_schema=False, so the config schema reaches the model as given. This matches the TypeScript graph's strict: false. Strict mode would make every property required, add additionalProperties: false, and reject additionalProperties: true.
• Handler errors propagate, matching the TypeScript graph's errorFunction: null. A directly built FunctionTool does not wrap on_invoke_tool.
• openai-agents>=0.17,<0.25. openai-agents makes breaking changes in minor (0.Y) releases (its docs/release.md), and 0.24.0 is the latest, so the bound stops at 0.25. uv lock --check passes; the lockfile changes only the specifier.

Tests

• TestBuildNodeToolsSyncHandlers now builds tools with the real FunctionTool instead of patching importlib.import_module.
• New TestNativeGraphRunsANodeWithAToolOnTheRealSdk runs to_openai_agents with the real agents package and fakes only the model (MultiProvider.get_model). It asserts that the tool call reaches the handler with the parsed arguments, that the run returns the model's final answer, and that the model receives a FunctionTool with the config's name, description and schema unchanged. The schema has an optional property, so a strict-mode rewrite would fail the test.
• On main 8218626, all 4 of these tests fail with the ticket's TypeError. With this change they pass on 0.17.7, and the native graph suite passes on 0.24.0 (uv run --with openai-agents==0.24.0).
• The stale agents_mock.tool stubs are removed.
• The real-SDK test turns off the SDK's trace export for its run, so a developer with OPENAI_API_KEY set sends nothing to OpenAI. I checked this with OPENAI_API_KEY=sk-fake: without the patch the run logs a 401 from the trace exporter, and with it there is no request.

Spec

No change needed. TESTING.md §2.2 "Tool forwarding" already requires node tools "in that provider's native tool format". Python broke that rule, and this PR fixes it.

Checks

make lint format-check typecheck api-report pass and uv lock --check passes. make test: 3714 passed, 18 skipped, and 2 failed in packages/client/tests/test_lifecycle.py::TestInitClientSDKKeyPath. Those 2 fail the same way locally on clean main 8218626, pass when that file runs alone, and this PR does not touch the client package.

Notes

• Known difference from JS, kept on purpose: tool arguments that are not valid JSON become {} and the handler still runs, as in Python's handler.py. JS tool() raises InvalidToolInputError and fails the run. Changing this would make Python's handler and native graph behave differently, so it is left for a follow-up that aligns both Python paths with JS.

• fix(AIC-3573): never send a NativeTool to the model as a function tool #248 (AIC-3573) edits the same loop in _build_node_tools, adding an is_function_tool skip. Whichever merges second gets a small conflict.

• On openai-agents 0.24.0, test_parameter_forwarding.py::test_every_field_is_classified_exactly_once fails because ModelSettings gained preserve_raw_usage, prompt_cache_options and timeout. This is not caused by this PR, and CI is locked to 0.17.7. It needs a separate ticket if the lock moves.

🤖 Generated with Claude Code

---

Note

Overview
Fixes a 1.0 blocker where to_openai_agents crashed on any graph node that declared tools: _build_node_tools called agents.tool, which is a submodule, not a decorator—raising TypeError: 'module' object is not callable.

Node tools are now built with agents.FunctionTool, matching the single-node handler. Tool callbacks use on_invoke_tool with JSON argument decoding, default empty-object schemas when missing, and strict_json_schema=False so LaunchDarkly config schemas reach the model unchanged (TypeScript strict: false parity).

The openai-agents dependency is capped at >=0.17,<0.25 because the SDK ships breaking changes in minor releases.

Tests stop mocking a fake agents.tool and instead exercise the real FunctionTool; a new integration test runs the native graph through the real Agents SDK with only the model faked, asserting tool calls reach handlers and schemas are not strict-rewritten.

^{Reviewed by Cursor Bugbot for commit 1969deb. Bugbot is set up for automated code reviews on this repo. Configure here.}