test: cover direct browser stream cancellation

[rgarcia]

summary

• add an async telemetry stream that blocks until its consumer is canceled
• verify cancellation stops the consumer and closes the direct-routed response body
• assert the request reaches the browser VM telemetry path

tests

• pytest -q — 671 passed
• ruff check .
• pyright --pythonpath /tmp/kernel-python-venv/bin/python tests/test_browser_routing.py

---

Note

Low Risk
Test-only addition with no runtime code changes; low risk beyond CI/test maintenance.

Overview
Adds test_async_telemetry_stream_cancellation_survives_direct_routing, which exercises AsyncKernel when telemetry/stream is allowlisted for direct browser VM routing.

The test uses a mock httpx.AsyncByteStream that blocks until the stream consumer is cancelled, then asserts the consumer raises CancelledError and the underlying response read/teardown completes (via read_stopped). It also pins that the request hits /browser/kernel/telemetry/stream on the cached browser route.

No production or SDK behavior changes—only regression coverage for cancelling long-lived SSE reads on direct-routed telemetry streams.

^{Reviewed by Cursor Bugbot for commit b136c50. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Sayan-]

Post-merge verification. Passes cleanly (24 tests in this file), no objection to it landing. Flagging that it guards less than the title suggests, so it is not counted as coverage we have when we do not.

Two things I checked:

1. The routing assertion is redundant. I disabled the rewrite in rewrite_direct_vm_options and this test failed, but so did three pre-existing tests, including test_telemetry_stream_routes_directly_to_vm immediately above it. So the /browser/kernel/telemetry/stream path assertion adds nothing new.

2. The cancellation assertions are close to tautological. pytest.raises(asyncio.CancelledError) after consumer.cancel() is a property of asyncio, not of this SDK. read_stopped is set by the test's own mock in both __aiter__'s finally and aclose(), so either path satisfies it and the assertion cannot tell them apart.

More to the point, there is no analogous defect available to regress. The whole routing surface is two hooks in _client.py (lines 337 and 341): rewrite_direct_vm_options, which model_copys FinalRequestOptions and edits only url and params, and strip_direct_vm_auth, which pops a header. Neither touches request construction, the transport, or the cancel scope, which is exactly why Python was never affected by the Node defect.

If you want this to earn its name, the assertion that would actually bite is that cancellation reaches the transport, for example by asserting the mock transport observed the disconnect rather than that the mock's own bookkeeping flag was set.

Nit: yield b"" after await asyncio.Event().wait() is unreachable.

[Sayan-]

Verified post-merge: passes, no objection. Approving; see the note above on what it does and does not guard.

[rgarcia]

addressed in #151. The test now records CancelledError inside the transport byte stream, so it proves consumer cancellation reaches the transport. It also removes the duplicate route-path assertion and unreachable mock yield.