Comparing changes

…334) Four pre-existing bugs left this entire test file unrunnable on main (4 failures + 24 errors); fixing them here so the suite actually exercises TemporalStreamingModel and protects against regressions. Bug 1 (24 errors): `conftest.py` defines fixture `mock_adk_streaming` (no underscore) but every test in TestStreamingModelSettings and TestStreamingModelTools requested it as `_mock_adk_streaming`, so pytest failed to resolve the fixture before the body ever ran. The fixture is ``autouse=True`` and the param value was never used in any test body, so the parameter was vestigial — replaced with `_streaming_context_vars`, which provides the ContextVar setup these tests now actually need. Bug 2 (4 failures): `TemporalStreamingModel.get_response()` reads `task_id`, `trace_id`, and `parent_span_id` from ContextVars populated by `ContextInterceptor` from request headers in real Temporal flows. Tests had been passing `task_id=...` as a kwarg, which is silently swallowed by `**kwargs` and ignored, so all three ContextVars stayed at their defaults and the validation at the top of `get_response` raised before any work happened. New `_streaming_context_vars` fixture in conftest sets all three vars (and resets them on teardown), simulating what `ContextInterceptor` does in production. Bug 3 (test_computer_tool): A recent commit narrowed `ComputerTool` serialization to require an actual `Computer`/`AsyncComputer` instance, but `sample_computer_tool` still built a bare `MagicMock`. Switched to `MagicMock(spec=Computer)` so the production isinstance check passes. Bug 4 (3 streaming-context tests): The 3 tests in TestStreamingModelBasics that assert on `streaming_task_message_context` calls built event sequences with raw `MagicMock(type="...")`. Production dispatches via `isinstance(event, ResponseOutputItemAddedEvent)` etc., which `MagicMock` without `spec` never satisfies, so dispatch was silently skipped and the assertions failed. Switched to `MagicMock(spec=...)` for each event type — passes isinstance without triggering pydantic validation on the event's required fields. Also fixed `test_task_id_threading` which had been asserting against a hardcoded `task_id="test_task_12345"` that was never actually threaded anywhere (the kwarg was ignored, just like in Bug 2); it now asserts against the value yielded by the fixture, which is the value production reads from the ContextVar. After all four fixes: 28/28 pass, ruff clean, pyright clean.

* feat(api): api update * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * chore(internal): more robust bootstrap script * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * fix: use correct field name format for multipart file arrays * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * feat: support setting headers via env * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * codegen metadata * fix: allow litellm security patch (#336) * fix(adk): Always inject headers on execute activity (#337) * perf(streaming): coalesce per-token publishes to Redis (50ms / 128-char window) (#333) * perf(streaming): coalesce per-token publishes to Redis (50ms / 128-char window) Per-token Redis publishes from TemporalStreamingModel were adding ~45s (56-62%) overhead to agent response latency, mostly from head-of-line blocking on the model's event loop: each `await streaming_context.stream_update(...)` inside the OpenAI stream `async for` paused token consumption until the publish round-trip completed. This change introduces a `CoalescingBuffer` driven by an `asyncio.Event`, so the producer never awaits on Redis. Deltas are merged consecutive-only (preserving character order in every (type, index) channel) and flushed on a 50ms timer, on a 128-char size threshold, or immediately for the first delta to keep perceived responsiveness high. The buffer's `close()` drains remaining deltas before the DONE event, so consumers see the full sequence in order. A new `StreamingMode = Literal["off", "per_token", "coalesced"]` lives in `streaming.py` as the single source of truth and is plumbed through the adk streaming module, `StreamingService.streaming_task_message_context`, and `StreamingTaskMessageContext`. Default is `"coalesced"` everywhere, so all 13+ existing context callers (claude_agents, langgraph, litellm provider, openai sync provider, etc.) benefit automatically. * chore(streaming): fix import ordering (ruff I001) * fix(streaming): address greptile review findings - _run: when CancelledError is raised mid-flush in the for-loop, re-enqueue the in-flight item plus any remaining items in the local `drained` list back into self._buf so close()'s final drain can recover them. Previously the local `drained` list was unreachable after CancelledError exited the for-loop, causing the last coalesced batch to be silently dropped on close-during-flush races. Trade-off: the in-flight item may be duplicated on the consumer side (Redis pub may have completed before cancel was delivered), which is preferable to silent loss for streaming UX. - _merge_pair: replace `return b` fallback with AssertionError. All six current TaskMessageDelta variants have explicit isinstance branches, so the fallback is unreachable today. But _can_merge returns True for any same-type pair, so adding a 7th delta variant without updating _merge_pair would silently drop `a`'s accumulated content. Asserting turns a future silent data-loss into an immediate, diagnosable crash. * test(streaming): add coalescing-layer tests; loosen one model assertion After merging the test-suite repair from main (#334) into this branch, one model test (test_responses_api_streaming) regressed because its assert_called_with strict-matched all kwargs of streaming_task_message_context and didn't tolerate the new `streaming_mode='coalesced'` kwarg this PR adds. Switched to assert_called() + targeted kwarg checks so the test verifies what it cares about (task_id threading) without locking in implementation details. Replaced the ad-hoc smoke scripts that lived in conversation with a real pytest module at tests/lib/core/services/adk/test_streaming.py covering: - _delta_char_len, _can_merge, _merge_pair: per-channel correctness + None-handling - _merge_consecutive: pure-text collapse, cross-channel order preservation, per-channel reconstruction matches per-token semantics - CoalescingBuffer: first-delta-immediate flush within ~20ms, size-threshold flush before timer fires, multi-delta coalescing within one window, idle close, add-after-close no-op - CoalescingBuffer cancel-during-flush regression test for the P1 fix: five queued chunks must all surface across publishes when close() cancels mid-flush (asserts substring presence rather than exact ordering, since the documented trade-off allows duplicates of the in-flight item) - StreamingTaskMessageContext mode dispatch: "off" suppresses publishes but persists full content, "per_token" publishes each delta synchronously, "coalesced" batches and persists full content * chore(streaming): route TemporalStreamingModel logger through make_logger The model file used raw ``logging.getLogger("agentex.temporal.streaming")``, which returns a logger with no handler attached and no level configured — so the existing ``[TemporalStreamingModel] Initialized ... streaming_mode=...`` INFO log was silently dropped, making it impossible to verify at runtime that a coalesced (or any) streaming mode was actually wired. Switch to the SDK's ``make_logger`` helper (level=INFO, RichHandler in local mode, StreamHandler otherwise) used everywhere else in the SDK. The explicit logger name ``agentex.temporal.streaming`` is preserved so any external logging configuration targeting that name keeps working. * codegen metadata * feat(api): api update * release: 0.10.3 --------- Co-authored-by: stainless-app[bot] <142633134+stainless-app[bot]@users.noreply.github.com> Co-authored-by: Brandon Allen <brandon.allen@scale.com> Co-authored-by: Declan Brady <declan.brady@scale.com> Co-authored-by: Stas Moreinis <stas.moreinis@scale.com>

Navigation Menu

Uh oh!

Comparing changes

Open a pull request

Commits on Apr 30, 2026

This comparison is taking too long to generate.

Uh oh!