core/lmproto+agent: stream via chat_completions(stream=True)

Keep library async-def→async-iterator shape; remove raw-lines path;
typed StreamToolCallDelta merge; scripted clients emit real chunks.

src/plyngent/agent/client.py

@@ -15,7 +15,9 @@ if TYPE_CHECKING:
 class ChatClient(Protocol):
     """Structural protocol for OpenAI-compatible chat completion clients.
 
-    ``chat_completions_raw_lines`` is optional (checked dynamically in the loop).
+    When ``stream=True``, implementations may return an async iterator from an
+    async method (``await client.chat_completions(..., stream=True)`` then
+    ``async for chunk in stream``). That library shape is intentional.
     """
 
     @overload

src/plyngent/agent/loop.py

@@ -4,11 +4,13 @@ from typing import TYPE_CHECKING, cast
 
 from msgspec import UNSET
 
+from plyngent.lmproto.openai_compatible.client import merge_stream_tool_calls
 from plyngent.lmproto.openai_compatible.model import (
     AnyAssistantToolCall,
     AssistantChatMessage,
     AssistantFunctionToolCall,
     ChatCompletionsParam,
+    StreamToolCallDelta,
     ToolChatMessage,
 )
 from plyngent.typedef import Unset  # noqa: TC001
@@ -79,44 +81,32 @@ async def _stream_and_build_assistant(
     client: ChatClient,
     param: ChatCompletionsParam,
 ) -> tuple[AssistantChatMessage, list[TextDeltaEvent]]:
-    """Stream a round, return the assistant message and any text deltas.
+    """Stream one completion via the normal client API.
 
-    Requires ``client.chat_completions_raw_lines`` (async generator of raw SSE
-    payload bytes). Falls back to non-streaming if unavailable.
+    Pattern: ``stream = await client.chat_completions(..., stream=True)`` then
+    ``async for chunk in stream``. That return type (async function → async
+    iterator) is accepted as the library interface.
     """
-    raw_lines_attr = getattr(client, "chat_completions_raw_lines", None)
-    if raw_lines_attr is None:
-        return await _non_stream_assistant(client, param)
-
-    # Async generators are not awaitable — call to get the iterator.
-    stream_iter = raw_lines_attr(param)
-    raw_lines: list[bytes] = []
+    stream = await client.chat_completions(param, stream=True)
     content_parts: list[str] = []
     text_events: list[TextDeltaEvent] = []
-    stream_decoder = getattr(client, "stream_decoder", None)
+    tool_deltas: list[StreamToolCallDelta] = []
 
-    async for raw in stream_iter:
-        raw_lines.append(raw)
-        if stream_decoder is None:
-            continue
-        try:
-            chunk = stream_decoder.decode(raw)
-        except Exception:  # noqa: BLE001 — skip malformed SSE payloads
-            continue
+    async for chunk in stream:
         if not chunk.choices:
             continue
-        delta_text = chunk.choices[0].delta.content
-        if isinstance(delta_text, str) and delta_text:
-            content_parts.append(delta_text)
-            text_events.append(TextDeltaEvent(content=delta_text))
+        choice = chunk.choices[0]
+        delta = choice.delta
+        if isinstance(delta.content, str) and delta.content:
+            content_parts.append(delta.content)
+            text_events.append(TextDeltaEvent(content=delta.content))
+        if delta.tool_calls is not UNSET and delta.tool_calls:
+            tool_deltas.extend(delta.tool_calls)
 
     full_content = "".join(content_parts)
     tool_calls: list[AnyAssistantToolCall] | Unset = UNSET
-    # Reconstruct tool calls from raw lines whenever present (do not rely solely
-    # on finish_reason — some providers omit it on intermediate chunks).
-    from plyngent.lmproto.openai_compatible.client import merge_stream_tool_calls
-
-    calls = merge_stream_tool_calls(raw_lines)
+    if tool_deltas:
+        calls = merge_stream_tool_calls(tool_deltas)
         if calls:
             tool_calls = cast("list[AnyAssistantToolCall]", calls)
 
@@ -140,10 +130,8 @@ async def run_chat_loop(
 ) -> AsyncIterator[AgentEvent]:
     """Multi-round chat/tool loop; mutates ``messages`` in place and yields events.
 
-    When ``stream=True``, text tokens yield as they arrive; tool calls are
-    reconstructed from raw SSE payloads when the client supports raw streaming.
-
-    Continues until the model returns no tool calls, or ``max_rounds`` is hit.
+    When ``stream=True``, uses ``chat_completions(..., stream=True)`` and yields
+    text deltas as chunks arrive; tool calls are merged from stream deltas.
     """
     tool_items: Sequence[AnyToolItem] | None = None
     if tools is not None and len(tools) > 0:

src/plyngent/lmproto/openai_compatible/client.py

@@ -1,9 +1,10 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any, Literal, cast, overload
+from typing import TYPE_CHECKING, Literal, overload
 
 import msgspec
 import niquests
+from msgspec import UNSET
 from niquests.auth import BearerTokenAuth
 
 from .config import OpenAIConfig  # noqa: TC001
@@ -13,11 +14,11 @@ from .model import (
     ChatCompletionChunk,
     ChatCompletionResponse,
     ChatCompletionsParam,
-    StreamChatCompletionChunk,
+    StreamToolCallDelta,
 )
 
 if TYPE_CHECKING:
-    from collections.abc import AsyncIterator, Mapping
+    from collections.abc import AsyncIterator
 
     from niquests.async_session import AsyncSession
     from niquests.models import AsyncResponse
@@ -28,7 +29,6 @@ class BaseOpenAIClient:
     encoder: msgspec.json.Encoder
     decoder: msgspec.json.Decoder[ChatCompletionResponse]
     chunk_decoder: msgspec.json.Decoder[ChatCompletionChunk]
-    stream_decoder: msgspec.json.Decoder[StreamChatCompletionChunk]
 
     def __init__(self, config: OpenAIConfig) -> None:
         self.session = niquests.AsyncSession(
@@ -38,38 +38,15 @@ class BaseOpenAIClient:
         self.encoder = msgspec.json.Encoder()
         self.decoder = msgspec.json.Decoder(ChatCompletionResponse)
         self.chunk_decoder = msgspec.json.Decoder(ChatCompletionChunk)
-        self.stream_decoder = msgspec.json.Decoder(StreamChatCompletionChunk)
 
     async def _parse_sse(self, resp: AsyncResponse) -> AsyncIterator[ChatCompletionChunk]:
-        lines = resp.iter_lines()
-        async for line in lines:
-            if not line or line == b"data: [DONE]":
-                continue
-            if line.startswith(b"data: "):
-                yield self.chunk_decoder.decode(line[6:])
-
-    async def chat_completions_raw_lines(self, param: ChatCompletionsParam) -> AsyncIterator[bytes]:
-        """Yield raw SSE ``data: `` payload bytes for manual accumulation.
-
-        Each yielded bytes object is a complete JSON object (no ``data: `` prefix).
-        The HTTP response is closed when the iterator finishes or is closed early
-        (e.g. task cancellation during streaming).
-        """
-        data = self.encoder.encode(msgspec.structs.replace(param, stream=True))
-        resp = await self.session.post(
-            "/chat/completions",
-            data=data,
-            headers={"Content-Type": "application/json"},
-            stream=True,
-        )
         try:
             async for line in resp.iter_lines():
                 if not line or line == b"data: [DONE]":
                     continue
                 if line.startswith(b"data: "):
-                    yield line[6:]
+                    yield self.chunk_decoder.decode(line[6:])
         finally:
-            # Best-effort abort of the in-flight stream (level-2 cancel toward HTTP).
             aclose = getattr(resp, "aclose", None)
             if callable(aclose):
                 await aclose()  # pyright: ignore[reportGeneralTypeIssues]
@@ -96,6 +73,7 @@ class OpenAIClient(BaseOpenAIClient):
     async def chat_completions(
         self, param: ChatCompletionsParam, *, stream: bool = False
     ) -> ChatCompletionResponse | AsyncIterator[ChatCompletionChunk]:
+        # Library pattern: async def returns AsyncIterator when stream=True.
         param = msgspec.structs.replace(param, stream=stream)
         data = self.encoder.encode(param)
         if stream:
@@ -116,92 +94,31 @@ class OpenAIClient(BaseOpenAIClient):
         return self.decoder.decode(resp.content)
 
 
-def _as_mapping(value: object) -> Mapping[str, Any] | None:
-    if isinstance(value, dict):
-        return cast("Mapping[str, Any]", value)
-    return None
-
-
-def _as_list(value: object) -> list[object] | None:
-    if isinstance(value, list):
-        return cast("list[object]", value)
-    return None
-
-
-def _dstr(d: Mapping[str, Any], key: str, default: str = "") -> str:
-    v = d.get(key)
-    return v if isinstance(v, str) else default
-
-
-def _merge_tool_entry(merge: dict[int, dict[str, str]], tc: Mapping[str, Any]) -> None:
-    idx_raw = tc.get("index", 0)
-    idx = idx_raw if isinstance(idx_raw, int) else 0
-    if idx not in merge:
-        merge[idx] = {"id": "", "name": "", "arguments": ""}
-    entry = merge[idx]
-    raw_id = tc.get("id")
-    if isinstance(raw_id, str) and raw_id:
-        entry["id"] = raw_id
-    fn = _as_mapping(tc.get("function", {}))
-    if fn is None:
-        return
-    name = _dstr(fn, "name")
-    if name:
-        entry["name"] = name
-    args = _dstr(fn, "arguments")
-    if args:
-        entry["arguments"] = entry["arguments"] + args
-
-
-def _stream_choices(raw_line: bytes) -> list[Mapping[str, Any]]:
-    """Extract ``choices`` list from a raw SSE payload byte string, or empty list."""
-    import json as _json
-
-    try:
-        raw_data: object = _json.loads(raw_line)
-    except _json.JSONDecodeError:
-        return []
-    data = _as_mapping(raw_data)
-    if data is None:
-        return []
-    choices = _as_list(data.get("choices", []))
-    if choices is None:
-        return []
-    result: list[Mapping[str, Any]] = []
-    for item in choices:
-        mapping = _as_mapping(item)
-        if mapping is not None:
-            result.append(mapping)
-    return result
-
-
-def merge_stream_tool_calls(raw_lines: list[bytes]) -> list[AssistantFunctionToolCall]:
-    """Accumulate streaming tool-call deltas by index across raw SSE payload bytes."""
+def merge_stream_tool_calls(deltas: list[StreamToolCallDelta]) -> list[AssistantFunctionToolCall]:
+    """Accumulate streaming tool-call deltas by index into complete tool calls."""
     merge: dict[int, dict[str, str]] = {}
-    for raw in raw_lines:
-        for choice in _stream_choices(raw):
-            delta = _as_mapping(choice.get("delta", {}))
-            if delta is None:
-                continue
-            calls = _as_list(delta.get("tool_calls", []))
-            if calls is None:
-                continue
-            for tc in calls:
-                mapping = _as_mapping(tc)
-                if mapping is not None:
-                    _merge_tool_entry(merge, mapping)
+    for delta in deltas:
+        if delta.index not in merge:
+            merge[delta.index] = {"id": "", "name": "", "arguments": ""}
+        entry = merge[delta.index]
+        if isinstance(delta.id, str) and delta.id:
+            entry["id"] = delta.id
+        if delta.function is not UNSET:
+            fn = delta.function
+            if isinstance(fn.name, str) and fn.name:
+                entry["name"] = fn.name
+            if isinstance(fn.arguments, str) and fn.arguments:
+                entry["arguments"] = entry["arguments"] + fn.arguments
 
     result: list[AssistantFunctionToolCall] = []
     for idx in sorted(merge):
         entry = merge[idx]
-        entry_id = entry["id"]
-        fn_name = entry["name"]
-        if not entry_id or not fn_name:
+        if not entry["id"] or not entry["name"]:
             continue
         result.append(
             AssistantFunctionToolCall(
-                id=entry_id,
-                function=AssistantFunctionTool(name=fn_name, arguments=entry["arguments"]),
+                id=entry["id"],
+                function=AssistantFunctionTool(name=entry["name"], arguments=entry["arguments"]),
             )
         )
     return result

src/plyngent/lmproto/openai_compatible/model.py

@@ -219,10 +219,23 @@ class ChatCompletionResponse(Struct):
     service_tier: ServiceTier | Unset = UNSET
 
 
+# Streaming deltas may be partial (id/name on first chunk, arguments later).
+class StreamFunctionDelta(Struct):
+    name: str | Unset = UNSET
+    arguments: str | Unset = UNSET
+
+
+class StreamToolCallDelta(Struct):
+    index: int = 0
+    id: str | Unset = UNSET
+    type: Literal["function"] | Unset = UNSET
+    function: StreamFunctionDelta | Unset = UNSET
+
+
 class DeltaMessage(Struct):
     role: RoleAssistant | Unset = UNSET
     content: str | Unset = UNSET
-    tool_calls: list[AnyAssistantToolCall] | Unset = UNSET
+    tool_calls: list[StreamToolCallDelta] | Unset = UNSET
 
 
 class ChunkChoice(Struct):
@@ -239,25 +252,3 @@ class ChatCompletionChunk(Struct):
     model: str
     choices: list[ChunkChoice]
     usage: dict[str, Any] | Unset = UNSET
-
-
-# -- Streaming-tolerant models (ignore partial tool calls in stream) --
-
-class StreamDelta(Struct):
-    """Delta that accepts any extra fields (partial tool calls), ignoring them."""
-    content: str | Unset = UNSET
-
-
-class StreamChunkChoice(Struct):
-    index: int
-    delta: StreamDelta
-    finish_reason: FinishReason | None | Unset = UNSET
-
-
-class StreamChatCompletionChunk(Struct):
-    id: str
-    object: Literal["chat.completion.chunk"]
-    created: int
-    model: str
-    choices: list[StreamChunkChoice]
-    usage: dict[str, Any] | Unset = UNSET

tests/test_agent/test_loop.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 from typing import TYPE_CHECKING, Literal, overload
 
 import pytest
+from msgspec import UNSET
 
 from plyngent.agent import (
     AssistantMessageEvent,
@@ -25,6 +26,10 @@ from plyngent.lmproto.openai_compatible.model import (
     ChatCompletionChunk,
     ChatCompletionResponse,
     ChatCompletionsParam,
+    ChunkChoice,
+    DeltaMessage,
+    StreamFunctionDelta,
+    StreamToolCallDelta,
     UserChatMessage,
 )
 from plyngent.memory import MemoryStore
@@ -33,8 +38,79 @@ if TYPE_CHECKING:
     from collections.abc import AsyncIterator
 
 
+def _chunks_from_response(response: ChatCompletionResponse) -> list[ChatCompletionChunk]:
+    """Turn a full response into stream chunks (library-style stream=True path)."""
+    message = response.choices[0].message
+    chunks: list[ChatCompletionChunk] = []
+    if isinstance(message.content, str) and message.content:
+        chunks.append(
+            ChatCompletionChunk(
+                id=response.id,
+                object="chat.completion.chunk",
+                created=response.created,
+                model=response.model,
+                choices=[
+                    ChunkChoice(
+                        index=0,
+                        delta=DeltaMessage(content=message.content),
+                        finish_reason=None,
+                    )
+                ],
+            )
+        )
+    tool_calls = message.tool_calls
+    if tool_calls is not UNSET and tool_calls:
+        deltas: list[StreamToolCallDelta] = []
+        for i, call in enumerate(tool_calls):
+            if isinstance(call, AssistantFunctionToolCall):
+                deltas.append(
+                    StreamToolCallDelta(
+                        index=i,
+                        id=call.id,
+                        type="function",
+                        function=StreamFunctionDelta(
+                            name=call.function.name,
+                            arguments=call.function.arguments,
+                        ),
+                    )
+                )
+        if deltas:
+            chunks.append(
+                ChatCompletionChunk(
+                    id=response.id,
+                    object="chat.completion.chunk",
+                    created=response.created,
+                    model=response.model,
+                    choices=[
+                        ChunkChoice(
+                            index=0,
+                            delta=DeltaMessage(tool_calls=deltas),
+                            finish_reason="tool_calls",
+                        )
+                    ],
+                )
+            )
+    if not chunks:
+        chunks.append(
+            ChatCompletionChunk(
+                id=response.id,
+                object="chat.completion.chunk",
+                created=response.created,
+                model=response.model,
+                choices=[
+                    ChunkChoice(
+                        index=0,
+                        delta=DeltaMessage(),
+                        finish_reason=response.choices[0].finish_reason or "stop",
+                    )
+                ],
+            )
+        )
+    return chunks
+
+
 class ScriptedClient:
-    """Returns scripted non-streaming chat completions in order."""
+    """Scripted chat completions; supports stream=True via chunked responses."""
 
     _responses: list[ChatCompletionResponse]
     calls: list[ChatCompletionsParam]
@@ -57,16 +133,16 @@ class ScriptedClient:
         self, param: ChatCompletionsParam, *, stream: bool = False
     ) -> ChatCompletionResponse | AsyncIterator[ChatCompletionChunk]:
         self.calls.append(param)
-        if stream:
-            return self._empty_stream()
         if not self._responses:
             msg = "no more scripted responses"
             raise RuntimeError(msg)
-        return self._responses.pop(0)
+        response = self._responses.pop(0)
+        if stream:
+            return self._as_stream(response)
+        return response
 
-    async def _empty_stream(self) -> AsyncIterator[ChatCompletionChunk]:
-        empty: list[ChatCompletionChunk] = []
-        for chunk in empty:
+    async def _as_stream(self, response: ChatCompletionResponse) -> AsyncIterator[ChatCompletionChunk]:
+        for chunk in _chunks_from_response(response):
             yield chunk
 
 
@@ -290,15 +366,15 @@ async def test_chat_agent_retry_after_failure() -> None:
             if self.calls == 1:
                 msg = "temporary"
                 raise RuntimeError(msg)
+            response = _response(AssistantChatMessage(content="recovered"))
             if stream:
 
-                async def empty() -> AsyncIterator[ChatCompletionChunk]:
-                    empty_chunks: list[ChatCompletionChunk] = []
-                    for chunk in empty_chunks:
+                async def as_stream() -> AsyncIterator[ChatCompletionChunk]:
+                    for chunk in _chunks_from_response(response):
                         yield chunk
 
-                return empty()
-            return _response(AssistantChatMessage(content="recovered"))
+                return as_stream()
+            return response
 
     client = FlakyClient()
     agent = ChatAgent(client, model="m", memory=store, session_id=session.sid)

tests/test_cli/test_retry.py

@@ -13,6 +13,8 @@ from plyngent.lmproto.openai_compatible.model import (
     ChatCompletionChunk,
     ChatCompletionResponse,
     ChatCompletionsParam,
+    ChunkChoice,
+    DeltaMessage,
     UserChatMessage,
 )
 from plyngent.memory import MemoryStore
@@ -40,6 +42,24 @@ def _response(message: AssistantChatMessage) -> ChatCompletionResponse:
     )
 
 
+async def _as_stream(response: ChatCompletionResponse) -> AsyncIterator[ChatCompletionChunk]:
+    message = response.choices[0].message
+    content = message.content if isinstance(message.content, str) else ""
+    yield ChatCompletionChunk(
+        id=response.id,
+        object="chat.completion.chunk",
+        created=response.created,
+        model=response.model,
+        choices=[
+            ChunkChoice(
+                index=0,
+                delta=DeltaMessage(content=content),
+                finish_reason="stop",
+            )
+        ],
+    )
+
+
 class FlakyClient:
     calls: int
     fail_times: int
@@ -66,15 +86,10 @@ class FlakyClient:
         if self.calls <= self.fail_times:
             msg = f"fail-{self.calls}"
             raise RuntimeError(msg)
+        response = _response(AssistantChatMessage(content="ok"))
         if stream:
-
-            async def empty() -> AsyncIterator[ChatCompletionChunk]:
-                empty_chunks: list[ChatCompletionChunk] = []
-                for chunk in empty_chunks:
-                    yield chunk
-
-            return empty()
-        return _response(AssistantChatMessage(content="ok"))
+            return _as_stream(response)
+        return response
 
 
 async def test_sleep_cancellable_completes() -> None: