[longcw]

Problem: The unit-tests CI job grew from about 1 min in July to 8 min, because most new async tests wait on the real clock through sleeps, timeouts, retry backoff and close delays. The suite also grew from about 90 to 357 files, mostly one file per PR, with copied fakes and tests that repeat what another test already covers.

Fix: Timing-driven tests now run under virtual_time, real servers give way to in-memory fakes where possible, and the tests that still need real time use short timeouts and events. Per-PR files are folded into one module per unit under test, near-duplicates become parametrized tests, and a test is deleted only when another test still covers its behavior.

Context for reviewing and coding agents

Where the numbers come from

Before: CI run 42965670118 on main spent 450s in pytest for 7262 tests. A local per-test timing pass showed 30 files took 389s of the 432s summed test time, and the other ~225 files took under 1s each. After: the local --unit --audio_eot gate runs in about 44s wall (6115 passed, 2 runs, no flakes), against about 320s on main on the same machine. Test files go from 357 to 236.

What made it fast, by size of effect

• virtual_time on timing-driven modules: voice core 65s → 3s, realtime 75s → 4.5s, telemetry 59s → 1.7s.
• tests/virtual_time.py now patches time.time_ns and the time_ns that opentelemetry.sdk.trace captures at import. Span tests needed real_time before this, because the SDK stamped spans on the real clock while the agent passed virtual start times.
• shutdown_default_executor in the virtual loop waits on real time. Its 300s join timeout fired at once under virtual time, which printed "The executor did not finish joining its threads within 300 seconds".
• In-memory fakes replace loopback servers in test_inference_stt.py and test_plugin_baseten_stt.py, so those run under virtual time.
• The AWS STT fake now ends its output stream when the input closes, as AWS does; every test used to wait out the plugin's hard-coded 3s drain (18s → 0.5s).
• test_ipc.py spawns about 11 processes instead of about 30.

What was deleted

• test_llm.py was entirely commented out.
• test_chat_ctx.py::test_summarize made a real inference.LLM call when credentials were set, and asserted nothing.
• test_tokenizer_xml_markup.py downloaded a tiktoken encoding from openaipublic.blob.core.windows.net; it now splits word pieces locally. That was the only network call left in the unit gate.
• Print-only tests, getter-only tests, and tests that read private fields when a wire-level test covers the outcome.
• Tests whose behavior another remaining test covers. Each one was mapped to its covering test during the cleanup, for example test_trace_schema.py::test_full_fake_session_is_well_formed → test_trace_spans.py::test_plain_reply_is_one_generation.

Most of the drop from about 7250 to about 6130 collected tests is parametrization, not lost cases: test_transcription_filter.py checks all 8 chunk sizes inside one test per case (430 → 89 items).

What stays slow, and why

test_ipc.py (about 6s wall, real process spawns, concurrent group) and the Rime, Reson8 and Zoom plugin tests (about 8s summed, real loopback websocket servers). A shared in-memory websocket fake that aiohttp ws_connect code accepts would let those three run under virtual time.

Scope boundaries

Modules in the plugin, stt, tts, realtime, evals and docs categories are unchanged, and they still collect. test_plugin_liveavatar.py and test_plugin_synthesia.py carried both unit and plugin(...) markers; both are hermetic and keep only unit. No source outside tests/ changes, apart from dropping the unused OPENAI_API_KEY from the unit job, --durations=15 in make unit-tests, and the AGENTS.md rule on where a new test goes.

Product issues found, not fixed here

• APIConnectOptions._interval_for_retry in livekit-agents/livekit/agents/types.py documents an immediate first retry but returns 0.1s.
• ThreadJobExecutor.initialize() ignores InitializeResponse.error, so a failing initialize_process_fnc passes on THREAD and fails on PROCESS.
• utils/aio/interval.py Interval.tick treats _last_sleep == 0.0 as "not yet ticked", so it spins when loop.time() is 0.0.
• AgentSession.aclose() waits the full session_close_transcript_timeout (2s) even for an idle session.