chore(deps): bump openai-agents to 0.18.3

[mendral-app]

⚠️ Reviewer attention

• Massive pre-1.0 jump: openai-agents 0.0.19 → 0.18.3
• Import paths unchanged (from agents import ... still works)
• FunctionTool constructor gains optional fields (timeout, guardrails, approval); existing kwargs still valid
• ToolContext is used as a type annotation in our callback — we don't construct it, so constructor changes don't affect us
• Runner.run() gains optional parameters; existing call signatures unchanged
• New behavior: ModelRefusalError may now be raised during Runner.run() (previously returned empty output). This affects downstream users, not our SDK integration code.
• Suggested manual checks: Run openai integration tests to verify FunctionTool callback execution works correctly

Packages bumped

Package	Group	Old	New
openai-agents	openai	>=0.0.19	>=0.18.3

Per-package details

openai-agents (>=0.0.19 → >=0.18.3)

• Complete architectural evolution: added sessions, tool guardrails, retry policies, sandbox agents, WebSocket support.
• ToolContext constructor now requires more params (we don't construct it — framework injects it into callbacks).
• FunctionTool adds optional tool_input_guardrails, needs_approval, timeout_seconds fields (all optional, backward-compatible).
• on_invoke_tool callback can now return structured types (ToolOutputText, etc.) in addition to strings.
• ModelRefusalError raised instead of empty output on refusals.
• Impact: Our code in src/blaxel/openai/ creates FunctionTool with name, description, params_json_schema, on_invoke_tool — all still valid kwargs. ToolContext is type-annotated on the callback parameter, not constructed by us.

Files modified

• pyproject.toml

---

Note

Created by Mendral. Tag @mendral-app with feedback or questions.

• View implementation session