fix(ai): route async OpenAI streaming through capture_streaming_event

[ckarnell]

💡 Motivation and Context

The two async OpenAI streaming paths (openai_async.py:242 and :567) build the $ai_generation property dict inline. Sync OpenAI, both Anthropic twins and both Gemini twins all delegate to posthog.ai.utils.capture_streaming_event. Async OpenAI streaming therefore drops $ai_usage, which #411 added "for backend cost calculations", along with $ai_tokens_source, $ai_instructions on the Responses path, and the AI Gateway double-billing warning.

The history suggests drift rather than a decision. #411 and #444 changed only utils.py and missed both async paths, while #499 edited openai_async.py directly and the property it added is on both twins today. $ai_tokens_source has been absent since February.

This routes both async paths through the shared helper instead of adding the missing keys to the inline dicts, since the duplication is what let them drift apart.

💚 How did you test it?

posthog/test/ai/openai/test_async_parity.py compares the properties captured by sync and async streaming and fails on anything the async side omits. On main it fails with drops ['$ai_tokens_source', '$ai_usage'], and it passes with this change. Anthropic runs as a control in the same test and its twins already agree, so the OpenAI difference is not an artifact of the harness.

posthog/test/ai/ is 420 passed, 88 skipped, with otel/ excluded for a dependency missing locally. ruff format is clean.

I have no live PostHog project and made no real OpenAI calls, so neither event has been watched landing in ingestion. mypy was not run.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran sampo add to generate a changeset file

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

I direct a small studio and set this work going. The submission is mine. The finding came from a scanner we run across Python SDKs looking for sync and async twins that disagree, and Claude Code (Opus) wrote both the patch and the parity test.

Verification was deliberately not left to the same session that produced the patch. A second session re-ran the test against main and against the change, confirmed the failure text names the two dropped properties, and checked the ruff delta separately. The commit history behind the drift claim above was re-read at the commits, not carried over from the first summary.

One alternative was considered and dropped. Adding the four missing keys to the two inline dicts is a smaller diff, and it leaves in place the duplication that caused this.

The template asks for the DRI to be set as assignee. GitHub does not let an outside contributor assign anyone, so it is left blank and I am the DRI.

[marandaneto]

Inline review finding.

[marandaneto]

moving to draft until comments are resolved

[ckarnell]

Done. You were right to push back on the suppression, since it would have left the same import doing the same thing with the warning turned off.

Both fixtures now live in posthog/test/ai/openai/conftest.py, so pytest supplies them by discovery and no test module imports or rebinds the names. The noqa comments are gone. mock_client and streaming_tool_call_chunks moved out of test_openai.py unchanged, along with the two chunk-type imports only they used.

ruff check passes on those files. posthog/test/ai is 420 passed and 88 skipped, unchanged.

The API snapshot was also out of date and is regenerated. Routing the async streaming path through capture_streaming_event dropped the get_model_params alias from openai_async, which is the one line that changed.

[marandaneto]

Inline review findings.

[marandaneto]

pushed some changes

[ckarnell]

Thanks for cleaning it up, the re-export restore and the zero value assertions both look right to me.