sdk/voiceagents/azure-ai-voiceagents/tsp-location.yaml:1

• The PR description is still the template (missing a real summary of changes, links to relevant issues/spec PRs, and checklist updates). Please replace the placeholder description with a concise overview of what this PR adds (new SDK package, realtime WebSocket client, samples/tests) and link the spec change PR/commit referenced in tsp-location.yaml.
  sdk/voiceagents/azure-ai-voiceagents/tests/unit/test_configuration.py:1
• This test doesn’t actually validate overriding behavior because it passes the same value as the default (\"v1\"). Consider changing it to pass a different value (or None if supported by the client) so the test can detect regressions where the override is ignored.
  sdk/voiceagents/azure-ai-voiceagents/samples/management/sample_create_and_manage_voice_agent.py:1
• Using a fixed, non-namespaced agent_name risks collisions across repeated runs and (more importantly) can inadvertently delete/modify a real agent if a user already has an agent with that name. Safer pattern for samples is to generate a unique name (e.g., including a UUID suffix) and only delete the resource that the sample created.

sdk/voiceagents/ci.yml:1

• The PR description is still the template text (it asks to 'add an informative description' and link issues), but the change set is substantial (new SDK + CI + tests + samples). Please update the PR description to summarize the intent/scope and link any relevant spec/issue/PR references so reviewers can validate the changes against requirements.
  sdk/voiceagents/azure-ai-voiceagents/pyproject.toml:1
• Excluding the azure and azure.ai packages from package discovery is likely to produce an invalid wheel where the namespace package initializers (azure/__init__.py, azure/ai/__init__.py) are not installed, breaking imports in environments that rely on pkgutil-style namespace packaging. Recommended fix: remove \"azure\" and \"azure.ai\" from the exclude list (or switch to an explicit include pattern for azure* packages consistent with other Azure SDK Python packages).
  sdk/voiceagents/azure-ai-voiceagents/samples/live/sample_live_audio_conversation_async.py:1
• The callback can return fewer than wanted bytes when it hits the end-of-stream marker (if not data: break). pyaudio callbacks must return exactly the requested frame size; otherwise playback can glitch or error. Instead of break, pad the remainder with silence (or return paComplete with a correctly-sized buffer) to satisfy pyaudio’s size contract.
  sdk/voiceagents/azure-ai-voiceagents/tests/unit/test_configuration.py:1
• This test name claims the API version is 'overridden', but it sets the value to the default (\"v1\"), so it doesn’t actually verify override behavior. Either rename the test to reflect what it asserts (e.g., that api_version is stored), or set api_version to a non-default value (if supported by the client) to validate that overriding works.
  sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_realtime.py:648
• Passing the WebSocket subprotocol by manually setting Sec-WebSocket-Protocol is fragile because many WS clients (including aiohttp) expect to manage/validate this header via the protocols=[...] argument. Prefer ws_connect(..., protocols=[\"realtime\"], ...) and remove the manual header to avoid handshake issues and ensure the negotiated subprotocol is tracked correctly by the client.

        headers: Dict[str, str] = {
            "Authorization": f"Bearer {token.token}",
            "Foundry-Features": str(foundry_features),
            "Sec-WebSocket-Protocol": "realtime",
        }

sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_realtime.py:655

• Passing the WebSocket subprotocol by manually setting Sec-WebSocket-Protocol is fragile because many WS clients (including aiohttp) expect to manage/validate this header via the protocols=[...] argument. Prefer ws_connect(..., protocols=[\"realtime\"], ...) and remove the manual header to avoid handshake issues and ensure the negotiated subprotocol is tracked correctly by the client.

            connection = await session.ws_connect(url, headers=headers, params=params, **self._kwargs)

sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/_configuration.py:36

• The docstring says the default api_version is None, but the implementation defaults it to \"v1\". Please align the docstring with the actual behavior (and consider applying the same fix in the async configuration file as well) to avoid confusing SDK consumers.

    :keyword api_version: The API version to use for this operation. Known values are "v1" and
     None. Default value is None. If not set, the operation's default API version will be used. Note
     that overriding this default value may result in unsupported behavior.

sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/_configuration.py:41

• The docstring says the default api_version is None, but the implementation defaults it to \"v1\". Please align the docstring with the actual behavior (and consider applying the same fix in the async configuration file as well) to avoid confusing SDK consumers.

    def __init__(self, endpoint: str, credential: "TokenCredential", **kwargs: Any) -> None:
        api_version: str = kwargs.pop("api_version", "v1")

sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_realtime.py:51

• The exported connection's public methods and attributes use ClientEvent, ServerEvent, ConversationItem, and the resource helper classes in their annotations, but this module and aio/_patch.py export only the three connection classes. Consumers therefore see public annotations that can only be imported from the private _realtime module. Re-export all types used by this public surface from azure.ai.voiceagents.aio.

__all__ = [
    "AsyncRealtime",
    "AsyncRealtimeConnection",
    "AsyncRealtimeConnectionManager",
]

sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_patch.py:45

• Constructing aiohttp.ClientSession here eagerly requires a running event loop and binds the client to that loop. This makes otherwise valid async-client construction outside asyncio.run() fail before any request. Set the default Accept-Encoding through the headers policy instead so AioHttpTransport can retain its lazy session creation.

                kwargs["transport"] = AioHttpTransport(
                    session=aiohttp.ClientSession(auto_decompress=False, headers={"Accept-Encoding": "gzip, deflate"})
                )

sdk/voiceagents/azure-ai-voiceagents/samples/live/sample_live_audio_conversation_async.py:138

• Barge-in does not discard a partially consumed chunk: each callback copies bytes from remaining before checking that chunk's sequence against _playback_base. A large delta can therefore continue playing across callbacks after skip_pending_audio(). Track the sequence associated with remaining and invalidate it before slicing whenever it is older than the playback base.
  sdk/voiceagents/azure-ai-voiceagents/samples/quickstart/sample_quickstart_async.py:100
• Barge-in does not discard a partially consumed chunk: bytes are copied from remaining before the callback checks sequence validity. If a delta spans multiple callback buffers, stale audio continues playing after skip_pending_audio(). Associate remaining with its sequence and clear it before slicing when that sequence predates _playback_base.
  sdk/voiceagents/azure-ai-voiceagents/tests/unit/test_configuration.py:47
• This test passes the same value as the default, so it still passes if the api_version override is ignored. Use a distinct sentinel value to verify that the keyword is actually honored.
  sdk/voiceagents/azure-ai-voiceagents/samples/live/sample_live_text_conversation_async.py:179
• After a timeout, the response remains active and its eventual events stay queued. On the next turn, pump() can consume the previous response's response.done and return before the new response completes, or the service can reject the new response while the old one is active. Cancel the timed-out response and drain/match its terminal event before accepting another prompt.

sdk/voiceagents/azure-ai-voiceagents/pyproject.toml:13

• This new package is missing both api.md and api.metadata.yml. The repository API consistency gate explicitly marks an affected package missing either file as a failure (.github/workflows/src/api-md-consistency/find_mismatches.js:76-92). Generate and commit both files with azpysdk apistub azure-ai-voiceagents.
  sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_patch.py:43
• Creating aiohttp.ClientSession in the client constructor makes VoiceAgentsClient(...) require a running event loop and binds the transport to that loop. This breaks the normal pattern of constructing an async SDK client before asyncio.run(...) (the new test even documents this limitation). Create/configure the session lazily when the transport opens instead.

                kwargs["transport"] = AioHttpTransport(
                    session=aiohttp.ClientSession(auto_decompress=False, headers={"Accept-Encoding": "gzip, deflate"})
                )

sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_realtime.py:242

• agent_name is inserted into the URL as raw path text, unlike the generated operation which URL-serializes this path parameter. Reserved characters can therefore change the route instead of identifying the requested agent. Percent-encode the path segment before interpolation.

    return f"{base}/agents/{agent_name}/endpoint/protocols/voice"

sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_patch.py:65

• The public annotations/docstrings refer to azure.ai.voiceagents.aio.ClientEvent and ServerEvent, but these aliases are defined only in the private _realtime module and are not re-exported here. Those documented imports therefore fail; include both aliases in the public aio surface.

__all__: list[str] = [
    "VoiceAgentsClient",
    "AsyncRealtime",
    "AsyncRealtimeConnection",
    "AsyncRealtimeConnectionManager",
]  # Add all objects you want publicly available to users at this package level

sdk/voiceagents/azure-ai-voiceagents/samples/quickstart/sample_quickstart_async.py:115

• The callback consumes remaining before checking its sequence against _playback_base. After skip_pending_audio() advances the base during barge-in, bytes left from the old response are still played on the next callback, so interruption does not actually drop all pending audio. Track the sequence associated with remaining and clear it before consuming stale bytes.
  sdk/voiceagents/azure-ai-voiceagents/samples/live/sample_live_audio_conversation_async.py:153
• The callback consumes remaining before checking its sequence against _playback_base. After barge-in advances the base, bytes left from the previous response are still played on the next callback, contradicting the promised immediate drop of pending audio. Track the sequence associated with remaining and discard stale bytes before playback.
  sdk/voiceagents/azure-ai-voiceagents/tests/unit/test_configuration.py:47
• This test supplies the same v1 value as the default, so it still passes if the override is ignored. Use a non-default sentinel to exercise the behavior named by the test.
  sdk/voiceagents/azure-ai-voiceagents/azure/ai/voiceagents/aio/_realtime.py:522
• The hand-written realtime implementation has no deterministic unit coverage for frame serialization/deserialization, typed-event mapping and unknown-event fallback, close/error handling, or resource helpers; the live smoke test covers only one service path. Add mocked WebSocket/session tests for these public behaviors so regressions do not require a gated live environment to detect.

    async def recv(self) -> ServerEvent:

Suppressed comments (16)

sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_basic.py:24

• Voice-agent lifecycle APIs are introduced in 2.6.0, but this command permits 2.0.0. Raise the minimum so the models and operations used by the sample are present.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_basic_async.py:18
• Voice-agent lifecycle APIs are introduced in 2.6.0, but this command permits 2.0.0. Raise the minimum so the async sample installs a compatible SDK.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_generate.py:20
• generate_agent for voice agents is added in 2.6.0, so the current minimum can install an SDK that lacks the imported request type and method. Require 2.6.0.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_function_tool.py:23
• The realtime voice-agent API used by this sample is introduced in 2.6.0, but this command permits 2.0.0. Raise the minimum so client.realtime and the event models are available.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation.py:30
• The realtime extra and client.realtime API are introduced in 2.6.0, but this command accepts 2.0.0. Users following it can install an older SDK with no realtime namespace; require 2.6.0.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation_async.py:27
• The realtime extra and async realtime API are introduced in 2.6.0, but this command accepts 2.0.0. Require 2.6.0 so the documented install includes the APIs used below.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation.py:22
• The beta voice-conversation operation group is added in 2.6.0, so allowing 2.0.0 makes these instructions install versions without beta.agent_endpoint_conversations. Raise the minimum to 2.6.0.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation_audio.py:26
• The beta voice-conversation audio APIs are added in 2.6.0, so allowing 2.0.0 can install an SDK without these methods. Raise the documented minimum to 2.6.0.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_versions.py:19
• Voice-agent APIs are introduced in 2.6.0 (CHANGELOG.md:3-17), but this command permits 2.0.0, where VoiceAgentDefinition does not provide this surface. Require 2.6.0 in the setup command.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_with_tools.py:25
• Voice-agent APIs are introduced in 2.6.0 (CHANGELOG.md:3-17), but this command permits 2.0.0, where these models are unavailable. Require 2.6.0 so the documented setup installs a compatible SDK.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_crud.py:33
• This limitation contradicts the new test_voice_agent_conversations.py, which creates a persisted realtime conversation ID and then reads it through beta.agent_endpoint_conversations. Update this note now that read-back is covered separately.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_crud_async.py:34
• This limitation contradicts the new async conversation test, which creates a persisted realtime conversation ID and reads it through this REST operation group. Replace the stale service-bug note with a reference to the dedicated coverage.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:34
• Every test in this new sync test module is unconditionally skipped, leaving the telephony and generated-audio operation behavior without executable functional coverage; the feature-header tests only validate header injection. Add transport-mocked tests or enable these scenarios before the API is considered covered.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:62
• This says realtime-generated conversation IDs are rejected, but the new test_voice_agent_conversations.py exercises exactly that setup and read-back path. Update this bullet to point to the dedicated test instead of documenting a conflicting limitation.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony_async.py:34
• Every test in this new async test module is unconditionally skipped, leaving the async telephony and generated-audio operation behavior without executable functional coverage; the feature-header tests only validate header injection. Add transport-mocked tests or enable these scenarios before the API is considered covered.
  sdk/ai/azure-ai-projects/tsp-location.yaml:2
• This regeneration pin changes the source specification, but the PR description is still the unfilled template and does not link the corresponding azure-rest-api-specs PR. Replace the template with the voice-agent scope and the source-spec PR so reviewers can validate these generated API changes.

• Files reviewed: 75/79 changed files
• Comments generated: 6
• Review effort level: Balanced

Suppressed comments (9)

eng/tools/azure-sdk-tools/azpysdk/samples.py:96

• This ignore list still allows both read-conversation samples to run, although they unconditionally read FOUNDRY_VOICE_CONVERSATION_ID and the new test resources cannot provide a pre-existing conversation ID. Because azpysdk samples executes every unignored Python sample, the samples check will fail with KeyError; add both read samples to this list as the package-level sample tests already do.

        "sample_voice_agent_live_audio_conversation_async.py",

scripts/devops_tasks/test_run_samples.py:98

• This legacy runner also executes every unignored Python sample, so the two read-conversation samples will fail when FOUNDRY_VOICE_CONVERSATION_ID is absent. Keep this duplicate ignore table aligned with the azpysdk table and exclude both state-dependent samples.

        "sample_voice_agent_live_audio_conversation_async.py",

sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation.py:29

• The included live conversation samples obtain the persisted ID from the session.created event, not conversation.created. Referencing the latter sends users looking for an event this client intentionally leaves untyped; update the instruction to match the demonstrated API.
  sdk/ai/azure-ai-projects/tests.yml:11
• This switches the entire azure-ai-projects test run to live mode, but the accompanying Bicep template explicitly provisions only the voice-agent surface and omits resources needed by the package's existing datasets, evaluations, fine-tuning, storage, and connection tests. The manual pipeline will therefore run unrelated recorded tests against an incomplete live environment; filter the test job to dedicated voice markers/tests or provision the complete package resource set.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_realtime_live.py:222
• This live test bypasses the configured FOUNDRY_VOICE_MODEL_NAME used by the other tests in this class. Any test deployment whose resource name is not literally gpt-realtime will fail before reaching the WebSocket behavior being tested. Use the prepared deployment name instead.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_realtime_live_async.py:219
• The async function-tool test also hardcodes the model resource name instead of using the deployment supplied by servicePreparer. This makes the test fail in otherwise valid environments that deploy the realtime model under another name.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:82
• Every test in this new module is unconditionally skipped, so none of the telephony request paths, response deserialization, or generated-audio operations execute in CI. Header-proxy tests alone cannot validate those contracts; add transport-mocked/recorded coverage that can run now, and keep only genuinely service-dependent success scenarios skipped.
  sdk/ai/azure-ai-projects/tests/samples/test_samples.py:409
• After get_sample_paths filters out _async.py, this blacklist excludes every remaining voice sample, so this parametrized test has zero runnable cases. Commit recordings for at least the REST-only samples (or omit this no-op test until they are available) so the newly added samples receive actual execution coverage.
  sdk/ai/azure-ai-projects/tests/samples/test_samples_async.py:328
• These three entries are all of the async voice samples discovered by get_async_sample_paths, so this new test parametrizes no runnable case. Add a recording for the REST-only async sample (or defer this no-op test) rather than presenting an empty sample test suite.

• Files reviewed: 74/78 changed files
• Comments generated: 3
• Review effort level: Balanced

Suppressed comments (5)

Previously missed (2) — in code that hasn't changed since the last review.

sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_conversations.py:246

• This loop can fall through after every item returns 404, so the test passes without finding any item audio even though the preceding comment says per-item audio is a hard requirement. Use the loop's else branch to fail when no iteration reaches break.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_conversations_async.py:239
• This loop can exhaust all items after 404 responses and still let the test pass, leaving both per-item audio methods effectively unverified. Fail via a for/else when no item reaches this break.

sdk/ai/azure-ai-projects/CHANGELOG.md:9

• agents has no public update method (the generated surface exposes create_version and update_details), and voice-agent definitions are updated by creating a new immutable version. Referring users to create_version/update therefore points them to a nonexistent API.

  * Added guided authoring via `project_client.agents.generate_agent(GenerateVoiceAgentRequest(kind=AgentKind.VOICE, ...))`, which returns a service-generated starter definition that can be edited afterward through the standard `create_version`/`update` flow.

sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation.py:29

• The referenced live-audio sample obtains the persisted ID from RealtimeServerEventSessionCreated, not from a conversation.created event. The current instruction sends users looking for an event that the sample never handles.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:82
• All three tests in this new module are unconditionally skipped, so CI never exercises telephony response deserialization, service error handling, or generated-audio error mapping. The header-capture suite stops at request transmission; please add transport-mocked or recorded response tests that can run until the service routes are available.

• Files reviewed: 75/79 changed files
• Comments generated: 2
• Review effort level: Balanced

Suppressed comments (2)

sdk/ai/azure-ai-projects/azure/ai/projects/_patch.py:265

• This top-level preview surface bypasses the client's documented allow_preview opt-in: Realtime.connect() always sends VoiceAgents=V1Preview, even when the client was constructed with the default allow_preview=False. Unlike .beta, client.realtime is a top-level API, and every added sample passes allow_preview=True, so accessing it without that consent should fail rather than silently enabling preview behavior. Apply the same guard to the async property.

    @property
    def realtime(self) -> Realtime:
        """Realtime streaming entry point for voice agents.

        :return: The realtime namespace, exposing ``connect(...)``.
        :rtype: ~azure.ai.projects.Realtime
        """
        if self._realtime is None:
            self._realtime = Realtime(self)

sdk/ai/azure-ai-projects/azure/ai/projects/aio/_patch.py:210

• The async top-level realtime surface also silently opts callers into VoiceAgents=V1Preview even when allow_preview is left at its documented default of False. This differs from other top-level preview APIs and makes the explicit opt-in ineffective; mirror the sync guard here.

    @property
    def realtime(self) -> AsyncRealtime:
        """Realtime streaming entry point for voice agents.

        :return: The realtime namespace, exposing ``connect(...)``.
        :rtype: ~azure.ai.projects.aio.AsyncRealtime
        """
        if self._realtime is None:
            self._realtime = AsyncRealtime(self)

• Files reviewed: 75/79 changed files
• Comments generated: 2
• Review effort level: Balanced

Suppressed comments (6)

Previously missed (4) — in code that hasn't changed since the last review.

sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_conversations.py:226

• If every item returns 404, this loop falls through and the test passes without ever exercising get_agent_conversation_item_audio_content, despite the preceding comment making per-item audio coverage a hard requirement. Track whether an item with audio was found and fail after the loop when none was found, so a cassette cannot silently omit this API coverage.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_conversations_async.py:218
• If every item returns 404, this loop falls through and the test passes without ever exercising get_agent_conversation_item_audio_content, despite the preceding comment making per-item audio coverage a hard requirement. Track whether an item with audio was found and fail after the loop when none was found, so a cassette cannot silently omit this API coverage.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_realtime_live.py:267
• The follow-up response is created before the current function-call response emits response.done. The realtime protocol can reject this as a concurrent active response; the existing Voice Live test waits for response.done before adding the tool output and calling response.create (sdk/voicelive/azure-ai-voicelive/tests/live/test_live_realtime_tools.py:271-281). Defer these operations until the function-call response completes.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_realtime_live_async.py:265
• The follow-up response is created before the current function-call response emits response.done. The realtime protocol can reject this as a concurrent active response; the existing Voice Live test waits for response.done before adding the tool output and calling response.create (sdk/voicelive/azure-ai-voicelive/tests/live/test_live_realtime_tools.py:271-281). Defer these operations until the function-call response completes.

sdk/ai/azure-ai-projects/tests/samples/test_samples.py:390

• This parameterization currently discovers no tests: every synchronous sample_*.py in samples/agents/voice is present in samples_to_skip. Pytest therefore reports only an empty-parameter skip, so this method provides no sample execution coverage. Add a recording for at least the REST-only samples (or remove the placeholder test until one is runnable).
  sdk/ai/azure-ai-projects/tests/samples/test_samples_async.py:320
• This parameterization currently discovers no tests: all three asynchronous sample_*.py files in samples/agents/voice are listed in samples_to_skip. The new test is therefore always an empty-parameter skip and exercises no sample. Add a recording for the REST-only basic sample (or remove the placeholder until it is runnable).

• Files reviewed: 75/79 changed files
• Comments generated: 4
• Review effort level: Balanced

Suppressed comments (8)

sdk/ai/azure-ai-projects/CHANGELOG.md:9

• .agents.update() was removed from this package; the existing changelog explicitly directs users to create_version() instead (CHANGELOG.md:406-407). Referring to an update flow here sends users to a nonexistent API.

  * Added guided authoring via `project_client.agents.generate_agent(GenerateVoiceAgentRequest(kind=AgentKind.VOICE, ...))`, which returns a service-generated starter definition that can be edited afterward through the standard `create_version`/`update` flow.

sdk/ai/azure-ai-projects/azure/ai/projects/_patch.py:269

• This top-level preview entry point bypasses the client's explicit preview opt-in. allow_preview defaults to False, but Realtime.connect() defaults foundry_features to VoiceAgents=V1Preview and the connection manager always sends that header, so client.realtime.connect(...) opts in even when the caller did not. Gate this property/connection on self._config.allow_preview, or omit the preview header unless the client was opted in.

    @property
    def realtime(self) -> Realtime:
        """Realtime streaming entry point for voice agents.

        :return: The realtime namespace, exposing ``connect(...)``.
        :rtype: ~azure.ai.projects.Realtime
        """
        if self._realtime is None:
            self._realtime = Realtime(self)
        return self._realtime

sdk/ai/azure-ai-projects/azure/ai/projects/aio/_patch.py:215

• The async realtime entry point also bypasses the client's explicit preview opt-in: AsyncRealtime.connect() supplies VoiceAgents=V1Preview unconditionally, even when allow_preview=False. Make the async path enforce the same opt-in contract as the other top-level preview APIs.

    @property
    def realtime(self) -> AsyncRealtime:
        """Realtime streaming entry point for voice agents.

        :return: The realtime namespace, exposing ``connect(...)``.
        :rtype: ~azure.ai.projects.aio.AsyncRealtime
        """
        if self._realtime is None:
            self._realtime = AsyncRealtime(self)
        return self._realtime

sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_function_tool.py:129

• response.function_call_arguments.done completes the arguments, but the current response is still active until its subsequent response.done. Starting response.create() here can be rejected as a concurrent response; the repository's VoiceLive flow waits for response.done before adding the tool output and creating the follow-up response (sdk/voicelive/azure-ai-voicelive/tests/live/test_live_realtime_tools.py:271-281). Defer the output/follow-up until that event.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:33
• Every test in this new module is unconditionally skipped, so none of the advertised telephony, transfer-target, or generated-audio behavior is exercised in CI. Please add transport-mocked/recorded coverage for request serialization and response handling, or defer these tests until they can run.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony_async.py:33
• Every async test in this new module is unconditionally skipped, so the async telephony and generated-audio paths receive no executable coverage. Please add transport-mocked/recorded cases or defer these placeholders until they can run.
  sdk/ai/azure-ai-projects/tests/samples/test_samples.py:409
• This blacklist excludes every synchronous sample_*.py currently present under samples/agents/voice, so parametrization produces no runnable cases and this test adds no sample coverage. Record the REST-only samples (or otherwise provide executable fixtures) and remove them from this list.
  sdk/ai/azure-ai-projects/tests/samples/test_samples_async.py:328
• These entries, together with the WebSocket skip above, exclude all three asynchronous voice samples, leaving this parametrized test with zero runnable cases. Add a recording/fixture for the REST-only sample and stop skipping it so the async sample suite exercises at least one new sample.

• Files reviewed: 77/81 changed files
• Comments generated: 3
• Review effort level: Balanced

Suppressed comments (8)

eng/tools/azure-sdk-tools/azpysdk/samples.py:93

• This ignore list omits sample_voice_agent_live_function_tool.py. That sample also opens a real client.realtime WebSocket (and the package sample suite skips it for that reason), so the azpysdk sample check will execute it non-interactively and attempt a live connection. Add it beside the other realtime samples.

        "sample_voice_agent_live_text_conversation.py",
        "sample_voice_agent_live_text_conversation_async.py",

scripts/devops_tasks/test_run_samples.py:95

• This runner has the same missing exclusion for sample_voice_agent_live_function_tool.py. Since that sample requires an uncaptured live WebSocket, running it in this non-interactive sample job will fail instead of exercising a replayable sample.

        "sample_voice_agent_live_text_conversation.py",
        "sample_voice_agent_live_text_conversation_async.py",

sdk/ai/azure-ai-projects/.env.template:27

• This comment is inaccurate: foundry_voice_model_name is also consumed by the conversation setup, telephony tests, and realtime live tests, not only the CRUD tests. Update it so contributors know which tests require this environment variable.

# Read by the recorded voice-agent CRUD tests only (tests/test_base.py), not by any sample.
FOUNDRY_VOICE_MODEL_NAME=

sdk/ai/azure-ai-projects/azure/ai/projects/aio/operations/init.py:22

• The async VoiceAgentWebSocketOperations export has no corresponding AIProjectClient.voice_agent_web_socket attribute either, despite the generated class directing users there. Remove this unusable public surface or expose a functional async operation group consistently with the intended async_client.realtime API.

from ._operations import VoiceAgentWebSocketOperations  # type: ignore

sdk/ai/azure-ai-projects/azure/ai/projects/operations/init.py:22

• VoiceAgentWebSocketOperations is exported as public API, but AIProjectClient never creates the documented voice_agent_web_socket attribute; the supported implementation is client.realtime. Users following this generated operation class's documentation therefore get AttributeError. Hide this generated WebSocket operation group or wire it to a usable client surface consistently.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation.py:29
• This instruction names a conversation.created event, but the referenced live sample obtains the ID from RealtimeServerEventSessionCreated (session.created). Following the current text leaves users waiting for an event this client does not handle; point them to session.created instead.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:37
• Every test in this newly added module is unconditionally skipped, so the sync telephony and generated-audio APIs receive no executable functional coverage. Before this feature is merged, add transport-mocked/recorded coverage that runs in CI (or enable these tests once the routes are available); placeholder tests cannot detect regressions.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony_async.py:37
• Every test in this newly added async module is unconditionally skipped, so the async telephony and generated-audio APIs receive no executable functional coverage. Before this feature is merged, add transport-mocked/recorded coverage that runs in CI (or enable these tests once the routes are available); placeholder tests cannot detect regressions.

• Files reviewed: 77/81 changed files
• Comments generated: 0 new
• Review effort level: Balanced

Suppressed comments (4)

sdk/ai/azure-ai-projects/docs/public-methods.md:15

• The new agent_endpoint_conversations and agent_telephony methods are counted as stable solely because they are top-level, but this PR identifies voice agents as preview and gates these calls on VoiceAgents=V1Preview/allow_preview. Calling all 97 methods “stable” misstates their support status; update GeneratePublicMethods.ps1 to distinguish top-level placement from API stability, then regenerate this document.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation.py:30
• The setup command omits python-dotenv, although this sample imports dotenv below. In a clean environment following these instructions, the sample fails before it can run.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation_async.py:27
• The setup command omits python-dotenv, although this sample imports dotenv below. In a clean environment following these instructions, the sample fails before it can run.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation.py:29
• The referenced live-audio sample reads this ID from RealtimeServerEventSessionCreated, whose wire event is session.created; it does not capture a conversation.created event. Naming the wrong event sends users looking for an event the example never handles.

• Files reviewed: 79/83 changed files
• Comments generated: 5
• Review effort level: Balanced

Suppressed comments (15)

Previously missed (4) — in code that hasn't changed since the last review.

sdk/ai/azure-ai-projects/api.metadata.yml:2

• The updated API snapshot omits the new agent_endpoint_conversations and agent_telephony attributes from both sync and async AIProjectClient, even though _client.py now exposes them. Regenerate api.md and this metadata together so API review and consistency checks cover the actual client surface.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_conversations.py:236
• If every item returns 404 for audio, this loop falls through and the test passes without ever exercising get_agent_conversation_item_audio_content, despite the hard coverage requirement stated immediately above. Make absence of an audio-bearing item fail the test, for example with a for ... else assertion.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_conversations_async.py:239
• The async test also succeeds when every item has no audio, so get_agent_conversation_item_audio_content can remain completely unexercised. Add a failure on loop exhaustion to ensure the claimed per-item audio coverage is real.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_realtime_live_async.py:62
• These async live tests are not marked with pytest.mark.asyncio, and this package does not enable pytest-asyncio auto mode. Pytest will therefore skip the coroutine tests instead of running them, so the new live pipeline only exercises the sync client. Add the asyncio marker at module or class level.

sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_basic.py:24

• This voice-agent API is introduced in 2.7.0b1, but >=2.0.0 can install the 2.6.0 stable release, where these imports do not exist. Set the sample's minimum to 2.7.0b1.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_basic_async.py:18
• This voice-agent API is introduced in 2.7.0b1, but >=2.0.0 can install the 2.6.0 stable release, where these imports do not exist. Set the sample's minimum to 2.7.0b1.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_generate.py:20
• This voice-agent API is introduced in 2.7.0b1, but >=2.0.0 can install the 2.6.0 stable release, where these imports do not exist. Set the sample's minimum to 2.7.0b1.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_function_tool.py:23
• The realtime extra and voice APIs are introduced in 2.7.0b1, but >=2.0.0 can install 2.6.0 stable, where this sample cannot run. Raise the minimum preview version.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation.py:30
• The realtime extra and voice APIs are introduced in 2.7.0b1, but >=2.0.0 can install 2.6.0 stable, where this sample cannot run. Raise the minimum preview version.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation_async.py:27
• The async realtime and voice APIs are introduced in 2.7.0b1, but >=2.0.0 can install 2.6.0 stable, where this sample cannot run. Raise the minimum preview version.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation.py:22
• This voice-agent API is introduced in 2.7.0b1, but >=2.0.0 can install the 2.6.0 stable release, where these methods do not exist. Set the sample's minimum to 2.7.0b1.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation_audio.py:26
• This voice-agent API is introduced in 2.7.0b1, but >=2.0.0 can install the 2.6.0 stable release, where these methods do not exist. Set the sample's minimum to 2.7.0b1.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_versions.py:19
• This voice-agent API is introduced in 2.7.0b1, but >=2.0.0 can install the 2.6.0 stable release, where these imports do not exist. Set the sample's minimum to 2.7.0b1.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_with_tools.py:25
• This voice-agent API is introduced in 2.7.0b1, but >=2.0.0 can install the 2.6.0 stable release, where these imports do not exist. Set the sample's minimum to 2.7.0b1.
  sdk/ai/azure-ai-projects/tsp-location.yaml.saved:2
• This regeneration is pinned to the head of the still-open, blocked azure-rest-api-specs PR Azure/azure-rest-api-specs#45852, but this PR's description is still the untouched template and omits the required spec link and feature scope. Add that link and summarize the voice-agent batch-2 contract before this can be reviewed or merged as an SDK regeneration.

• Files reviewed: 78/82 changed files
• Comments generated: 2
• Review effort level: Balanced

Suppressed comments (22)

Previously missed (3) — in code that hasn't changed since the last review.

sdk/ai/azure-ai-projects/azure/ai/projects/_realtime.py:863

• This preview API unconditionally defaults and sends VoiceAgents=V1Preview, even when the parent client was constructed with the default allow_preview=False. That bypasses the explicit preview opt-in contract documented by AIProjectClient and README.md:45. Reject connect() unless self._config.allow_preview is true (or otherwise avoid adding the preview header) and mirror the behavior in the async implementation.
  sdk/ai/azure-ai-projects/azure/ai/projects/aio/_realtime.py:865
• The async realtime API has the same preview-opt-in bypass as the sync API: it always supplies VoiceAgents=V1Preview even when the parent client has allow_preview=False. Enforce the parent's preview setting before returning the connection manager so the documented stable-client contract is consistent across both clients.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_realtime_live_async.py:62
• These coroutine tests are not marked for an async pytest runner. Unlike tests/agents/test_realtime_client_async.py, which sets pytestmark = pytest.mark.asyncio, this live-only class has neither an asyncio marker nor recorded_by_proxy_async; in standard pytest-asyncio strict mode the methods are skipped instead of executed. Add the class-level marker so the new live coverage actually runs.

sdk/ai/azure-ai-projects/azure/ai/projects/_realtime.py:256

• agent_name is interpolated as a raw path segment, unlike the generated REST operations that percent-encode this parameter. Names containing reserved URL characters will alter the route or query rather than identify the intended agent. Encode the segment with urllib.parse.quote(agent_name, safe="") and add a reserved-character URL test.

    base = endpoint.rstrip("/")
    if base.startswith("https://"):
        base = "wss://" + base[len("https://") :]
    return f"{base}/agents/{agent_name}/endpoint/protocols/voice"

sdk/ai/azure-ai-projects/azure/ai/projects/aio/_realtime.py:270

• The async URL builder repeats the raw path interpolation from the sync client, so reserved characters in an agent name produce the wrong WebSocket route. Percent-encode agent_name as a single path segment and mirror the regression test in the async URL-helper tests.

    base = endpoint.rstrip("/")
    if base.startswith("https://"):
        base = "wss://" + base[len("https://") :]
    return f"{base}/agents/{agent_name}/endpoint/protocols/voice"

sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_basic.py:24

• This command can install stable 2.6.x, which predates the voice-agent API. Require the preview version that introduces the sample.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_basic_async.py:18
• This command can install stable 2.6.x, which predates the voice-agent API. Require the preview version that introduces the sample.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_generate.py:20
• This command may install stable 2.6.x, which lacks generate_agent and the voice-agent models. Require the first release containing this API.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_audio_conversation_async.py:32
• >=2.0.0 can resolve stable 2.6.x and leave users without the new voice/realtime classes. Require 2.7.0b1 or newer explicitly.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_function_tool.py:23
• >=2.0.0 does not opt pip into prereleases, so this can install stable 2.6.x without client.realtime. Require 2.7.0b1 or newer explicitly.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation.py:30
• >=2.0.0 does not opt pip into prereleases, so this can install stable 2.6.x without client.realtime. Require 2.7.0b1 or newer explicitly.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_live_text_conversation_async.py:27
• This can resolve stable 2.6.x, where the new async realtime surface is unavailable. Require the preview release that contains the sample.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation.py:22
• This command may install stable 2.6.x, which lacks agent_endpoint_conversations. Require the first release containing this API.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_read_conversation_audio.py:26
• This install command can resolve the current stable 2.6.x release, which does not contain the voice-agent APIs introduced in this PR, because pip excludes prereleases by default. Require the first release containing this sample so users do not immediately get import/attribute errors.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_versions.py:19
• This command can install stable 2.6.x, which predates voice-agent versioning. Require the preview version that introduces these APIs.
  sdk/ai/azure-ai-projects/samples/agents/voice/sample_voice_agent_with_tools.py:25
• This command may install stable 2.6.x, where the voice-agent models used below do not exist. Pin the minimum to the preview release that introduces this sample.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_crud.py:31
• This note is contradicted by test_voice_agent_conversations.py: that test obtains the ID from a live realtime session.created event and successfully uses it with agent_endpoint_conversations. Remove the obsolete service-limitation claim so it does not discourage valid coverage or usage.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_crud_async.py:32
• This note is contradicted by the async conversation test, which creates an ID through the realtime WebSocket and reads it through the REST conversation client. Remove this obsolete limitation.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:62
• The new conversation tests now prove that a realtime session.created conversation ID works with agent_endpoint_conversations, so this stated limitation is stale and incorrectly justifies missing coverage here.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony.py:79
• All three tests in this new file are unconditionally skipped, so none of the added telephony binding/call/transfer behavior executes in CI. Add transport-mocked contract tests now (route, body, response/error handling), or defer exposing these APIs until runnable service coverage is available; header-only tests do not exercise this behavior.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony_async.py:62
• The async conversation test now obtains a realtime session.created ID and successfully reads it through this REST operation group. Remove this obsolete service-limitation claim.
  sdk/ai/azure-ai-projects/tests/agents/test_voice_agent_telephony_async.py:79
• Every async telephony test in this file is permanently skipped, leaving the newly added async binding/call/transfer paths unexecuted. Add transport-mocked contract tests or wait to expose the surface until runnable service coverage exists.

• Files reviewed: 80/84 changed files
• Comments generated: 2
• Review effort level: Balanced