chat-completions: stream conversational turns (SSE token relay) for realtime voice

Adds stream=true support to POST /v1/chat/completions (it previously 400'd).
When streaming, it runs the no-tools `conversational` agent via
`claude -p --output-format stream-json --include-partial-messages --verbose`
and relays each content_block_delta as an OpenAI chat.completion.chunk SSE
event, ending with finish_reason=stop + [DONE]. Free CLI/subscription auth, no
tools, no API key.

Stateless by design: the full message history is flattened into the prompt
(prior assistant turns kept), so an OpenAI-style client that re-sends history
each turn — e.g. Pipecat's OpenAILLMService — can stream from us directly. The
non-streaming path (recruiter-triage workspace agent) is unchanged.

This is phase 1 of the Pipecat realtime full-duplex voice-agent rebuild for
portal-assistant (continuous audio, VAD endpointing, barge-in, ~seconds to
first words). New pure helpers (stream_argv/delta_text/openai_chunk/
synthesise_chat_prompt) are unit-tested; the SSE endpoint has a mocked-subprocess
integration test. 429 passing.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/test_conversational.py

@@ -171,3 +171,79 @@ async def test_conversational_returns_503_on_failure(auth_header):
             )
     assert r.status_code == 503
     assert r.json()["error"] == "execution failed"
+
+
+# --------------------------------------------------------------------------- #
+# streaming helpers (OpenAI-compatible token relay for the realtime voice agent)
+# --------------------------------------------------------------------------- #
+from collections import namedtuple  # noqa: E402
+
+_Msg = namedtuple("_Msg", "role content")
+
+
+def test_stream_argv_uses_stream_json_and_is_stateless():
+    argv = conversational.stream_argv("hello", "sonnet")
+    assert argv[:2] == ["claude", "-p"]
+    assert "--agent" in argv and "conversational" in argv
+    assert "stream-json" in argv
+    assert "--include-partial-messages" in argv
+    assert "--verbose" in argv
+    assert "--model" in argv and "sonnet" in argv
+    assert argv[-1] == "hello"
+    # stateless + no tools
+    assert "--resume" not in argv and "--session-id" not in argv
+    assert "--dangerously-skip-permissions" not in argv
+
+
+def test_delta_text_extracts_content_block_delta():
+    line = json.dumps({
+        "type": "stream_event",
+        "event": {"type": "content_block_delta",
+                  "delta": {"type": "text_delta", "text": "Слон"}},
+    })
+    assert conversational.delta_text(line) == "Слон"
+
+
+def test_delta_text_ignores_non_text_events():
+    for ev in [
+        {"type": "system"},
+        {"type": "stream_event", "event": {"type": "message_start"}},
+        {"type": "stream_event", "event": {"type": "content_block_delta",
+            "delta": {"type": "input_json_delta", "partial_json": "{"}}},
+        {"type": "result"},
+    ]:
+        assert conversational.delta_text(json.dumps(ev)) is None
+    assert conversational.delta_text("") is None
+    assert conversational.delta_text("not json") is None
+
+
+def test_openai_chunk_valid_sse_and_keeps_cyrillic():
+    s = conversational.openai_chunk("chatcmpl-x", "sonnet", 123, content="две")
+    assert s.startswith("data: ") and s.endswith("\n\n")
+    payload = json.loads(s[len("data: "):].strip())
+    assert payload["object"] == "chat.completion.chunk"
+    assert payload["choices"][0]["delta"]["content"] == "две"
+    assert payload["choices"][0]["finish_reason"] is None
+    assert "две" in s  # not unicode-escaped
+
+
+def test_openai_chunk_role_and_finish():
+    role = conversational.openai_chunk("id", "m", 1, role="assistant")
+    assert json.loads(role[6:].strip())["choices"][0]["delta"] == {"role": "assistant"}
+    stop = conversational.openai_chunk("id", "m", 1, finish_reason="stop")
+    c = json.loads(stop[6:].strip())["choices"][0]
+    assert c["finish_reason"] == "stop" and c["delta"] == {}
+
+
+def test_synthesise_chat_prompt_keeps_assistant_turns():
+    msgs = [
+        _Msg("system", "Be brief."),
+        _Msg("user", "Здравей"),
+        _Msg("assistant", "Здравей! Как си?"),
+        _Msg("user", "Добре, ти?"),
+    ]
+    p = conversational.synthesise_chat_prompt(msgs)
+    assert "Be brief." in p
+    assert "User: Здравей" in p
+    assert "Assistant: Здравей! Как си?" in p
+    assert p.strip().endswith("User: Добре, ти?")

tests/test_openai_compat.py

@@ -98,14 +98,15 @@ async def test_chat_completions_happy_path(auth_header):
 
 
 @pytest.mark.asyncio
-async def test_chat_completions_rejects_streaming(auth_header):
-    """stream=true is not supported and must 400 with a clear message."""
+async def test_chat_completions_streaming_rejects_unsupported_model(auth_header):
+    """Streaming is supported now; model validation still runs first, so an
+    unsupported model 400s before any CLI is spawned."""
     transport = ASGITransport(app=app)
     async with AsyncClient(transport=transport, base_url="http://test") as client:
         response = await client.post(
             "/v1/chat/completions",
             json={
-                "model": "haiku",
+                "model": "gpt-4",
                 "messages": [{"role": "user", "content": "hi"}],
                 "stream": True,
             },
@@ -113,7 +114,7 @@ async def test_chat_completions_rejects_streaming(auth_header):
         )
     assert response.status_code == 400
     body = response.json()
-    assert "streaming not supported" in json.dumps(body).lower()
+    assert "unsupported model" in json.dumps(body).lower()
 
 
 @pytest.mark.asyncio
@@ -370,3 +371,58 @@ async def test_chat_completions_response_model_echoes_default_when_missing(auth_
     )
     assert status == 200
     assert body["model"] == "sonnet"
+
+
+def _delta_line(text: str) -> str:
+    return json.dumps({
+        "type": "stream_event",
+        "event": {"type": "content_block_delta",
+                  "delta": {"type": "text_delta", "text": text}},
+    })
+
+
+@pytest.mark.asyncio
+async def test_chat_completions_streaming_relays_token_sse(auth_header):
+    """stream=true relays CLI stream-json token deltas as OpenAI SSE chunks."""
+    cli_output = "\n".join([
+        json.dumps({"type": "system"}),
+        json.dumps({"type": "stream_event", "event": {"type": "message_start"}}),
+        _delta_line("Две"),
+        _delta_line(" точки."),
+        json.dumps({"type": "result", "subtype": "success"}),
+    ]).encode()
+    mock_proc = _mock_subprocess_returning(cli_output, returncode=0)
+
+    with patch("app.main.asyncio.create_subprocess_exec", return_value=mock_proc):
+        transport = ASGITransport(app=app)
+        async with AsyncClient(transport=transport, base_url="http://test") as client:
+            response = await client.post(
+                "/v1/chat/completions",
+                json={
+                    "model": "sonnet",
+                    "stream": True,
+                    "messages": [{"role": "user", "content": "Колко е?"}],
+                },
+                headers=auth_header,
+            )
+
+    assert response.status_code == 200, response.text
+    assert response.headers["content-type"].startswith("text/event-stream")
+    body = response.text
+    assert "chat.completion.chunk" in body
+    assert body.rstrip().endswith("data: [DONE]")
+
+    # Reassemble the streamed assistant content from the delta chunks.
+    content = ""
+    saw_role = False
+    for line in body.splitlines():
+        if not line.startswith("data: ") or line.strip() == "data: [DONE]":
+            continue
+        payload = json.loads(line[len("data: "):])
+        assert payload["object"] == "chat.completion.chunk"
+        delta = payload["choices"][0]["delta"]
+        if delta.get("role") == "assistant":
+            saw_role = True
+        content += delta.get("content", "")
+    assert saw_role
+    assert content == "Две точки."