feat(U3): G8 delta_flush_interval 调速

- ReActEngine 新增 flush_interval_ms 构造参数(默认 0 = 逐 chunk yield 向后兼容)
- execute_stream chunk 循环用 time.monotonic 节流,累积 _flush_buffer 批量 yield
- flush_interval_ms=0 条件短路为 True 逐 chunk yield 保当前行为
- 流结束 mid-interval 最终 flush 剩余 buffer 不丢字符
- ServerConfig.streaming 配置项(flush_interval_ms)
- test_delta_flush.py 覆盖 R11/R12/R14

src/agentkit/core/react.py

@@ -164,6 +164,7 @@ class ReActEngine:
         enable_tool_search: bool = True,
         middleware_chain: "MiddlewareChain | None" = None,
         prompt_cache_enable: bool = True,
+        flush_interval_ms: int = 0,
     ):
         if max_steps < 1:
             raise ValueError(f"max_steps must be >= 1, got {max_steps}")
@@ -180,6 +181,9 @@ class ReActEngine:
         # U2/G2: prompt cache 双块结构开关(True 时 Anthropic 用 cache_control blocks,
         # 其他 provider 走字符串拼接依赖自动前缀缓存)
         self._prompt_cache_enable = prompt_cache_enable
+        # U3/G8: token chunk 节流间隔(ms)。0 = 逐 chunk yield(向后兼容)。
+        # 用 time.monotonic() 不受系统时钟跳变影响。
+        self._flush_interval_ms = flush_interval_ms
         # Tiered tool description injection config
         self._core_tool_names: tuple[str, ...] | None = (
             tuple(core_tool_names) if core_tool_names is not None else None
@@ -1123,6 +1127,9 @@ class ReActEngine:
                 stream_usage = None
                 stream_tool_calls: list[Any] = []
                 stream_model = model
+                # U3/G8: delta_flush 节流 buffer,按 flush_interval_ms 批量 yield
+                _flush_buffer: list[str] = []
+                _last_flush_ts = time.monotonic()
 
                 async for chunk in _ensure_async_iterable(
                     self._llm_gateway.chat_stream(
@@ -1136,11 +1143,20 @@ class ReActEngine:
                 ):
                     if chunk.content:
                         stream_content_chunks.append(chunk.content)
-                        yield ReActEvent(
+                        _flush_buffer.append(chunk.content)
-                            event_type="token",
+                        now = time.monotonic()
-                            step=step,
+                        # flush_interval_ms=0 → 逐 chunk yield(向后兼容,条件短路为 True)
-                            data={"content": chunk.content},
+                        if (
-                        )
+                            self._flush_interval_ms == 0
+                            or now - _last_flush_ts >= self._flush_interval_ms / 1000
+                        ):
+                            yield ReActEvent(
+                                event_type="token",
+                                step=step,
+                                data={"content": "".join(_flush_buffer)},
+                            )
+                            _flush_buffer = []
+                            _last_flush_ts = now
                     if chunk.usage:
                         stream_usage = chunk.usage
                     if chunk.tool_calls:
@@ -1148,6 +1164,15 @@ class ReActEngine:
                     if chunk.model:
                         stream_model = chunk.model
 
+                # U3/G8: 流结束 mid-interval → 最终 flush 剩余 buffer(不丢字符)
+                if _flush_buffer:
+                    yield ReActEvent(
+                        event_type="token",
+                        step=step,
+                        data={"content": "".join(_flush_buffer)},
+                    )
+                    _flush_buffer = []
+
                 # Build response-like object from stream
                 stream_content = "".join(stream_content_chunks)
                 response = self._build_response_from_stream(

src/agentkit/server/config.py

@@ -117,6 +117,7 @@ class ServerConfig:
         expert_paths: list[str] | None = None,
         board: dict[str, Any] | None = None,
         prompt_cache: dict[str, Any] | None = None,
+        streaming: dict[str, Any] | None = None,
         on_change: Callable[["ServerConfig"], None] | None = None,
     ):
         self.host = host
@@ -146,6 +147,8 @@ class ServerConfig:
         self.expert_paths = expert_paths or []
         self.board = board or {}
         self.prompt_cache = prompt_cache or {}
+        # U3/G8: streaming.flush_interval_ms 控制 token chunk 节流(默认 0 = 逐 chunk yield)
+        self.streaming = streaming or {}
         self.on_change = on_change
 
         # Config watching state

tests/unit/test_delta_flush.py

@@ -0,0 +1,148 @@
+"""U3 / G8 delta_flush_interval 调速测试。
+
+覆盖 R11-R12, R14:
+- R11 chunk 按 flush_interval_ms 间隔批量 yield
+- R12 配置化(flush_interval_ms=0 退化为逐 chunk yield)
+- R14 自检:合并 content 等于原始 chunks 拼接(不丢字符)
+"""
+
+from __future__ import annotations
+
+
+
+from agentkit.core.react import ReActEngine
+from agentkit.llm.protocol import StreamChunk
+
+
+class _StubGateway:
+    """模拟 LLMGateway,yield 一串 StreamChunk 后结束。"""
+
+    def __init__(self, chunks: list[str]):
+        self._chunks = chunks
+
+    def get_provider_name_for_model(self, model: str) -> str | None:
+        return None
+
+    async def chat_stream(self, **kwargs):
+        for c in self._chunks:
+            yield StreamChunk(content=c, model="test")
+
+
+def _collect_token_events(events) -> list[str]:
+    return [e.data["content"] for e in events if e.event_type == "token"]
+
+
+# ---- R12 Config: flush_interval_ms=0 → 逐 chunk yield(向后兼容) ----
+
+
+async def test_flush_interval_zero_yields_per_chunk():
+    chunks = ["H", "e", "l", "l", "o"]
+    gw = _StubGateway(chunks)
+    engine = ReActEngine(llm_gateway=gw, flush_interval_ms=0)
+    events = []
+    async for ev in engine.execute_stream(
+        messages=[{"role": "user", "content": "hi"}],
+        tools=[],
+        model="test",
+    ):
+        events.append(ev)
+    tokens = _collect_token_events(events)
+    # 5 chunks → 5 token events(每个内容 = 单 chunk)
+    assert tokens == ["H", "e", "l", "l", "o"]
+
+
+# ---- R11 Happy path: flush_interval_ms > 0 → 批量合并 ----
+
+
+async def test_flush_interval_batches_chunks_by_interval():
+    chunks = ["a", "b", "c", "d", "e", "f"]
+    gw = _StubGateway(chunks)
+    # 间隔设很大(10s),所有 chunks 在第一个 interval 内累积,流结束后最终 flush
+    engine = ReActEngine(llm_gateway=gw, flush_interval_ms=10000)
+    events = []
+    async for ev in engine.execute_stream(
+        messages=[{"role": "user", "content": "hi"}],
+        tools=[],
+        model="test",
+    ):
+        events.append(ev)
+    tokens = _collect_token_events(events)
+    # 所有 chunk 累积到流结束,最终 flush 一次 → 1 个 token event,content = 全拼接
+    assert len(tokens) == 1
+    assert tokens[0] == "abcdef"
+
+
+# ---- R14 Self-check: 合并 content 等于原始 chunks 拼接(不丢字符) ----
+
+
+async def test_no_character_loss_after_merge():
+    chunks = ["Hello", " ", "World", "!", "你好", "世界"]
+    gw = _StubGateway(chunks)
+    engine = ReActEngine(llm_gateway=gw, flush_interval_ms=10000)
+    events = []
+    async for ev in engine.execute_stream(
+        messages=[{"role": "user", "content": "hi"}],
+        tools=[],
+        model="test",
+    ):
+        events.append(ev)
+    tokens = _collect_token_events(events)
+    # 合并所有 token events 的 content 等于原始 chunks 拼接
+    merged = "".join(tokens)
+    assert merged == "".join(chunks) == "Hello World!你好世界"
+
+
+# ---- Edge: 流结束 mid-interval → 最终 flush 剩余 buffer ----
+
+
+async def test_final_flush_on_stream_end():
+    chunks = ["x", "y", "z"]
+    gw = _StubGateway(chunks)
+    engine = ReActEngine(llm_gateway=gw, flush_interval_ms=10000)
+    events = []
+    async for ev in engine.execute_stream(
+        messages=[{"role": "user", "content": "hi"}],
+        tools=[],
+        model="test",
+    ):
+        events.append(ev)
+    tokens = _collect_token_events(events)
+    # mid-interval 累积 → 流结束最终 flush 一次
+    assert tokens == ["xyz"]
+
+
+# ---- Edge: 单个 chunk 后流结束 → 立即 flush ----
+
+
+async def test_single_chunk_immediate_flush():
+    chunks = ["only"]
+    gw = _StubGateway(chunks)
+    engine = ReActEngine(llm_gateway=gw, flush_interval_ms=10000)
+    events = []
+    async for ev in engine.execute_stream(
+        messages=[{"role": "user", "content": "hi"}],
+        tools=[],
+        model="test",
+    ):
+        events.append(ev)
+    tokens = _collect_token_events(events)
+    assert tokens == ["only"]
+
+
+# ---- Edge: chunks 含空 content(usage-only chunk)不进 buffer ----
+
+
+async def test_empty_content_chunk_not_buffered():
+    chunks = ["a", "", "b"]  # 中间 chunk 空
+    gw = _StubGateway(chunks)
+    engine = ReActEngine(llm_gateway=gw, flush_interval_ms=10000)
+    events = []
+    async for ev in engine.execute_stream(
+        messages=[{"role": "user", "content": "hi"}],
+        tools=[],
+        model="test",
+    ):
+        events.append(ev)
+    tokens = _collect_token_events(events)
+    # 空 chunk 跳过 buffer,最终 flush "ab"
+    assert tokens == ["ab"]