[AAASM-3528] 🐛 (adapters): Fix OpenAI Agents adapter to govern current API

[Chisanan232]

Description

The Python SDK's OpenAI Agents adapter (agent_assembly/adapters/openai_agents/) targeted a stale framework API, so applying it was a silent no-op — agents running on the OpenAI Agents SDK were governed only in name and actually ran ungoverned (fail-open).

Root cause: the adapter patched openai.agents + FunctionTool.__call__. Neither exists in the shipped openai-agents framework: the package is the top-level agents module, and a tool is executed by the runner calling each tool instance's on_invoke_tool(ctx, args_json) coroutine — there is no __call__. Because that class attribute did not exist, the class-level patch never applied, and no pre-execution allow/deny or event emission ever fired.

Fix:

• Detect and patch the real agents package (not openai.agents); is_available() now keys on agents.
• FunctionTool.on_invoke_tool is a per-instance callable, not a class method, so it cannot be patched once on the class. The adapter now wraps FunctionTool.__init__ (and Handoff.__init__) so every constructed instance has its on_invoke_tool / on_invoke_handoff coroutine wrapped with the governance pre-execution check, deny/pending-approval handling, audit result recording, spawn-context tracking, and delegates_to topology-edge emission.
• A denied tool now returns the framework's documented string error (sent back to the LLM) instead of executing — blocking the call without aborting the agent run. Runner.run (a real classmethod) is still wrapped for spawn context.
• The adapter's public interface (OpenAIAgentsAdapter / OpenAIAgentsPatch with apply() / revert() / register_hooks() / unregister_hooks()) is unchanged and stays consistent with the LangChain/LangGraph adapters.

Type of Change

• 🔧 Bug fix

Breaking Changes

• No

Related Issues

• Related JIRA ticket: AAASM-3528

Testing

• Unit tests added/updated
• Integration tests added/updated
• Manual testing performed

The unit and integration tests now drive the real per-instance on_invoke_tool contract — the integration suite installs the real agents framework (importorskip-guarded; openai-agents added as a dev/test dep) and drives a genuine function_tool exactly as the runner does. The tests assert governance actually fires (a denied tool returns a policy error and its body does not run; an allowed tool runs and is recorded), and they fail if the patch is reverted to a no-op — verified locally by stubbing the wrapper to a no-op and confirming both the unit and integration governance tests go red.

Local checks (all clean for the change):

• uv run pytest — 584 passed, 13 skipped.
• uv run ruff check / ruff format --check — clean for changed files (pre-existing ARG002 noise in test/bench/** predates this change and was reduced, not added).
• uv run mypy agent_assembly — adapter clean; the 4 remaining package-wide errors are the pre-existing _core native-shim / grpc-stub baseline, unrelated to this change.
• pre-commit run (isort/autoflake/black/mypy) — all pass.

Checklist

• Code follows project style guidelines
• Self-review completed
• Comments added for complex logic
• All tests passing

Closes AAASM-3528

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 94.52055% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
agent_assembly/adapters/openai_agents/patch.py	94.44%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
94.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[Chisanan232]

🤖 Claude Code — PR review (record)

CI: ✅ all green/skipped.

Scope vs AAASM-3528: ✅ fully covers it — and it's a real security fix (silent fail-open → governed). The adapter targeted a stale API (openai.agents + FunctionTool.__call__); the shipped framework is the top-level agents package whose tools run via per-instance on_invoke_tool (no __call__), so patch.apply() was a no-op and OpenAI-Agents agents ran ungoverned. Fix wraps FunctionTool.__init__ / Handoff.__init__ so every constructed instance's invocation hook is governed; is_available() keys on agents. Proven not-a-no-op: stubbing the wrapper to a no-op turns the new unit + integration tests red; restored, they pass. Full suite 584 passed / 13 skipped; ruff + mypy clean on changed files.

Verdict: ✅ Ready to approve & merge. Closes AAASM-3528. Unblocks the OpenAI-Agents live framework smoke (now landed real in #94).

[Chisanan232]

This is the ref:
https://github.com/openai/openai-agents-python