diff --git a/gateway/model_pool.py b/gateway/model_pool.py
index 9c73b75..c5508b2 100644
--- a/gateway/model_pool.py
+++ b/gateway/model_pool.py
@@ -21,11 +21,13 @@ _POOL_PATH = Path(__file__).resolve().parent.parent / "config" / "model_pool.jso
 TIERS = ("local", "budget", "premium")
 BACKENDS = ("mock", "llama_server", "openai")
 ROLES = ("architect", "worker", "agent")
+PROVIDERS = ("deepseek", "openai", "anthropic")   # 代理层 usage 归一化用（D-P3）
 
 # 池条目允许的字段（其余字段拒绝写入）
 ENTRY_FIELDS = {
     "id", "name", "tier", "backend", "base_url", "model", "api_key",
     "price_in", "price_out", "temperature", "max_tokens", "enabled",
+    "provider", "in_hit_price",   # 代理层扩展（D-P3）：usage 归一化 / 按命中价选上游
 }
 
 # 单价默认值（$/1M tokens）；local 档为 0
@@ -177,6 +179,8 @@ class PoolStore:
         backend = entry.get("backend", "openai")
         if backend not in BACKENDS:
             raise PoolError(f"backend 必须是 {BACKENDS} 之一")
+        if entry.get("provider") and entry["provider"] not in PROVIDERS:
+            raise PoolError(f"provider 必须是 {PROVIDERS} 之一")
         tier = entry.get("tier", "budget")
         if tier not in TIERS:
             raise PoolError(f"tier 必须是 {TIERS} 之一")
@@ -212,6 +216,10 @@ class PoolStore:
             "temperature": temperature,
             "max_tokens": max_tokens,
             "enabled": bool(entry.get("enabled", True)),
+            # 代理层扩展（D-P3）：provider 缺省 openai；in_hit_price 缺省 = price_in × 1/30
+            "provider": (str(entry["provider"]) if entry.get("provider") else "openai"),
+            "in_hit_price": (float(entry["in_hit_price"]) if entry.get("in_hit_price") is not None
+                             else round(price_in / 30.0, 6)),
         }
 
     @staticmethod
diff --git a/gateway/proxy/upstream.py b/gateway/proxy/upstream.py
index c6370cc..3ce91f3 100644
--- a/gateway/proxy/upstream.py
+++ b/gateway/proxy/upstream.py
@@ -1,21 +1,139 @@
-"""上游客户端（T-P2 落地；本文件先立签名）。
+"""上游客户端（T-P2）：流式派发 + usage 注入/过滤 + 三家归一化 + 首 token 前 failover。
 
-规格（§6）：模块级 httpx.AsyncClient 单例（keepalive，max_connections=100），
-超时 connect=10s/read=120s/write=10s/pool=30s；始终注入
-stream_options.include_usage（客户端未要求 usage 时过滤该 chunk 不下发）；
-首 token 前 failover（D-P4）。
+锁定决策：
+- D-P4：仅在上游**首 token 返回前**允许切换备用条目；流中失败 = aborted（由调用方
+  按已收 usage 计费），本模块以 UpstreamAborted 标记。
+- §5.1：代理向上游始终注入 stream_options.include_usage（计量不依赖客户端行为）；
+  客户端未要求 usage 时，透传层过滤该 chunk 不下发（routes 的 tee 负责，见 filter_usage_chunk）。
+- httpx.AsyncClient 模块级单例（keepalive；limits.max_connections=100），
+  超时 connect=10s / read=120s / write=10s / pool=30s。
 """
 from __future__ import annotations
 
-from typing import Any, AsyncIterator, Dict
+import json
+from typing import Any, AsyncIterator, Dict, List, Optional
+
+import httpx
+
+from gateway.proxy.errors import UpstreamError
+
+_TIMEOUT = httpx.Timeout(connect=10.0, read=120.0, write=10.0, pool=30.0)
+_LIMITS = httpx.Limits(max_connections=100, max_keepalive_connections=20)
+
+_client: Optional[httpx.AsyncClient] = None
 
 
-async def stream(body: dict, entry: Dict[str, Any], usage_sink) -> AsyncIterator[bytes]:
-    """流式派发到上游，逐块 yield（T-P2 实现）。"""
-    raise NotImplementedError("T-P2")
-    yield b""  # pragma: no cover
+def get_client() -> httpx.AsyncClient:
+    """模块级单例（keepalive 连接池）。"""
+    global _client
+    if _client is None:
+        _client = httpx.AsyncClient(timeout=_TIMEOUT, limits=_LIMITS)
+    return _client
+
+
+async def close_client() -> None:
+    global _client
+    if _client is not None:
+        await _client.aclose()
+        _client = None
+
+
+class UpstreamAborted(Exception):
+    """首 token 已下发后上游失败（D-P4：不可 failover，按已收 usage 计费）。"""
 
 
 def normalize_usage(provider: str, usage_dict: Dict[str, Any]) -> Dict[str, int]:
-    """三家 usage 字段 -> 统一 {in_miss, in_hit, out}（T-P2 实现）。"""
-    raise NotImplementedError("T-P2")
+    """三家 usage 字段 -> 统一 {in_miss, in_hit, out}。
+
+    - deepseek：prompt_cache_hit_tokens / prompt_tokens
+    - openai 兼容：prompt_tokens_details.cached_tokens
+    - anthropic：cache_read_input_tokens / input_tokens
+    """
+    usage = usage_dict or {}
+    out = int(usage.get("completion_tokens") or usage.get("output_tokens") or 0)
+    total_in = int(usage.get("prompt_tokens") or usage.get("input_tokens") or 0)
+    if provider == "deepseek":
+        hit = int(usage.get("prompt_cache_hit_tokens") or 0)
+    elif provider == "anthropic":
+        hit = int(usage.get("cache_read_input_tokens") or 0)
+    else:  # openai / 兼容端点
+        details = usage.get("prompt_tokens_details") or {}
+        hit = int(details.get("cached_tokens") or 0) if isinstance(details, dict) else 0
+    hit = min(hit, total_in)
+    return {"in_miss": total_in - hit, "in_hit": hit, "out": out}
+
+
+def _inject_usage_option(body: dict) -> dict:
+    """浅拷贝注入 stream_options.include_usage（§5.1 计量不依赖客户端行为）。"""
+    shaped = dict(body)
+    shaped["stream_options"] = {"include_usage": True}
+    return shaped
+
+
+def filter_usage_chunk(raw_line: str, client_wants_usage: bool) -> Optional[str]:
+    """透传过滤：客户端未要求 usage 时剥除 usage 字段所在 chunk（返回 None = 丢弃）。"""
+    if client_wants_usage or not raw_line.startswith("data:"):
+        return raw_line
+    payload = raw_line[5:].strip()
+    if payload == "[DONE]":
+        return raw_line
+    try:
+        obj = json.loads(payload)
+    except json.JSONDecodeError:
+        return raw_line
+    if obj.get("usage"):
+        return None
+    return raw_line
+
+
+async def stream(body: dict, entry: Dict[str, Any], usage_sink: Dict[str, Any],
+                 fallback_entries: Optional[List[Dict[str, Any]]] = None
+                 ) -> AsyncIterator[bytes]:
+    """流式派发（§6 签名扩展 failover 链）。
+
+    - 始终注入 include_usage；usage chunk 交给 usage_sink（透传过滤由调用方用
+      filter_usage_chunk 决定）。
+    - D-P4：首 token 前（未 yield 任何字节）失败 -> 切换 fallback_entries；
+      已 yield 后失败 -> 抛 UpstreamAborted。
+    - usage_sink["usage"] 收敛为归一化 dict；sink["ttfb_ms"] 记录首字节耗时。
+    """
+    import time as _time
+    candidates: List[Dict[str, Any]] = [entry] + list(fallback_entries or [])
+    last_err: Optional[Exception] = None
+
+    for cand in candidates:
+        provider = str(cand.get("provider") or "openai")
+        url = cand["base_url"].rstrip("/") + "/chat/completions"
+        headers = {"Authorization": f"Bearer {cand['api_key']}"} if cand.get("api_key") else {}
+        payload = _inject_usage_option(body)
+        t0 = _time.perf_counter()
+        first = False
+        try:
+            client = get_client()
+            async with client.stream("POST", url, headers=headers, json=payload) as resp:
+                resp.raise_for_status()
+                async for line in resp.aiter_lines():
+                    if not line.startswith("data:"):
+                        continue
+                    raw = line
+                    payload_txt = line[5:].strip()
+                    if payload_txt == "[DONE]":
+                        yield (raw + "\n\n").encode("utf-8")
+                        continue
+                    try:
+                        obj = json.loads(payload_txt)
+                    except json.JSONDecodeError:
+                        continue
+                    if obj.get("usage"):
+                        usage_sink["usage"] = normalize_usage(provider, obj["usage"])
+                    if not first:
+                        first = True
+                        usage_sink["ttfb_ms"] = int((_time.perf_counter() - t0) * 1000)
+                    yield (raw + "\n\n").encode("utf-8")
+            return  # 正常结束
+        except Exception as exc:  # noqa: BLE001
+            if first:
+                raise UpstreamAborted(str(exc)) from exc
+            last_err = exc
+            continue  # 首 token 前失败 -> failover
+    raise UpstreamError(f"上游均不可用: {type(last_err).__name__}: {last_err}")
diff --git a/tests/test_proxy_upstream.py b/tests/test_proxy_upstream.py
new file mode 100644
index 0000000..591ad20
--- /dev/null
+++ b/tests/test_proxy_upstream.py
@@ -0,0 +1,206 @@
+"""上游客户端测试（T-P2）：SSE 透传 / usage 注入过滤 / 三家归一化 / failover / aborted。"""
+import asyncio
+import json
+
+import httpx
+import pytest
+
+from gateway.proxy.errors import UpstreamError
+from gateway.proxy.upstream import (
+    UpstreamAborted,
+    filter_usage_chunk,
+    normalize_usage,
+    stream,
+)
+
+
+def asyncio_run(coro):
+    return asyncio.run(coremaybe(coro))
+
+
+async def coremaybe(coro):
+    return await coro
+
+
+def _sse(lines) -> bytes:
+    return ("\n\n".join(lines) + "\n\n").encode("utf-8")
+
+
+def test_stream_passthrough_and_usage_sink():
+    """SSE 逐块 yield；usage chunk 进 sink；始终注入 include_usage。"""
+    body = _sse([
+        'data: {"choices":[{"delta":{"content":"你"}}]}',
+        'data: {"choices":[{"delta":{"content":"好"}}]}',
+        'data: {"choices":[{"delta":{},"finish_reason":"stop"}],'
+        '"usage":{"prompt_tokens":100,"prompt_cache_hit_tokens":60,"completion_tokens":20}}',
+        "data: [DONE]",
+    ])
+    seen_bodies = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        seen_bodies.append(json.loads(request.read()))
+        return httpx.Response(200, content=body)
+
+    transport = httpx.MockTransport(handler)
+    import gateway.proxy.upstream as up
+    client = httpx.AsyncClient(transport=transport)
+    orig = up._client
+    up._client = client
+    try:
+        entry = {"base_url": "http://up", "api_key": "k", "provider": "deepseek"}
+
+        async def run():
+            sink = {}
+            chunks = []
+            async for b in stream({"model": "deepseek-chat", "messages": []}, entry, sink):
+                chunks.append(b)
+            return sink, chunks
+
+        sink, chunks = asyncio_run(run())
+    finally:
+        up._client = orig
+        asyncio.get_event_loop_policy()
+        client.__asyncio_run = None
+        try:
+            asyncio.get_event_loop().run_until_complete(client.aclose())
+        except Exception:
+            pass
+    text = b"".join(chunks).decode("utf-8")
+    assert "\u4f60" in text and "\u597d" in text and "data: [DONE]" in text
+    assert sink["usage"] == {"in_miss": 40, "in_hit": 60, "out": 20}
+    assert isinstance(sink["ttfb_ms"], int)
+    assert seen_bodies[0]["stream_options"] == {"include_usage": True}
+
+
+def test_failover_before_first_token():
+    """首条目连不上 -> 切换备用 -> 正常输出。"""
+    ok_body = _sse(['data: {"choices":[{"delta":{"content":"ok"}}]}', "data: [DONE]"])
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        if request.url.host == "bad":
+            raise httpx.ConnectError("拒连")
+        return httpx.Response(200, content=ok_body)
+
+    transport = httpx.MockTransport(handler)
+    import gateway.proxy.upstream as up
+    client = httpx.AsyncClient(transport=transport)
+    orig = up._client
+    up._client = client
+    try:
+        async def run():
+            sink = {}
+            return [b async for b in stream(
+                {"model": "m", "messages": []},
+                {"base_url": "http://bad", "api_key": "", "provider": "openai"},
+                sink, [{"base_url": "http://good", "api_key": "", "provider": "openai"}])]
+
+        chunks = asyncio_run(run())
+        assert b"ok" in b"".join(chunks)
+    finally:
+        up._client = orig
+        try:
+            asyncio.run(client.aclose())
+        except Exception:
+            pass
+
+
+def test_abort_after_first_token_no_failover():
+    """首 token 已 yield 后断流 -> UpstreamAborted（不切换，D-P4）。"""
+    class StreamResp:
+        def raise_for_status(self):
+            pass
+
+        async def aiter_lines(self):
+            yield 'data: {"choices":[{"delta":{"content":"前半"}}]}'
+            raise httpx.ReadError("中途断")
+
+    class CM:
+        async def __aenter__(self):
+            return StreamResp()
+
+        async def __aexit__(self, *a):
+            return False
+
+    class FakeClient:
+        def stream(self, *a, **k):
+            return CM()
+
+    import gateway.proxy.upstream as up
+    orig = up._client
+    up._client = FakeClient()
+    try:
+        async def run():
+            sink = {}
+            gen = stream({"model": "m", "messages": []},
+                         {"base_url": "http://x", "api_key": "", "provider": "openai"}, sink)
+            first = await gen.__anext__()
+            assert "\u524d\u534a".encode("utf-8") in first   # "前半"
+            rest = []
+            async for b in gen:
+                rest.append(b)
+            return rest
+
+        try:
+            asyncio_run(run())
+            raised = False
+        except UpstreamAborted:
+            raised = True
+        assert raised, "应抛 UpstreamAborted"
+    finally:
+        up._client = orig
+
+
+def test_all_candidates_fail_raises_upstream_error():
+    """全部候选首 token 前失败 -> UpstreamError(502)。"""
+    import gateway.proxy.upstream as up
+
+    class FakeClient:
+        def stream(self, *a, **k):
+            raise httpx.ConnectError("全挂")
+
+    orig = up._client
+    up._client = FakeClient()
+    try:
+        async def run():
+            async for _ in stream({"model": "m", "messages": []},
+                                  {"base_url": "http://a", "api_key": "", "provider": "openai"},
+                                  {}, [{"base_url": "http://b", "api_key": "",
+                                        "provider": "openai"}]):
+                pass
+
+        try:
+            asyncio_run(run())
+            raised = False
+        except UpstreamError:
+            raised = True
+        assert raised
+    finally:
+        up._client = orig
+
+
+def test_normalize_usage_three_providers():
+    """deepseek / openai 兼容 / anthropic 三家字段归一；命中数钳制。"""
+    assert normalize_usage("deepseek", {
+        "prompt_tokens": 100, "prompt_cache_hit_tokens": 60, "completion_tokens": 20
+    }) == {"in_miss": 40, "in_hit": 60, "out": 20}
+    assert normalize_usage("openai", {
+        "prompt_tokens": 100,
+        "prompt_tokens_details": {"cached_tokens": 30},
+        "completion_tokens": 20,
+    }) == {"in_miss": 70, "in_hit": 30, "out": 20}
+    assert normalize_usage("anthropic", {
+        "input_tokens": 100, "cache_read_input_tokens": 90, "output_tokens": 5
+    }) == {"in_miss": 10, "in_hit": 90, "out": 5}
+    assert normalize_usage("deepseek", {
+        "prompt_tokens": 10, "prompt_cache_hit_tokens": 99, "completion_tokens": 1
+    })["in_miss"] == 0
+
+
+def test_filter_usage_chunk():
+    """客户端未要求 usage 时剥除 usage chunk；要求时透传；[DONE] 恒透传。"""
+    usage_line = 'data: {"choices":[],"usage":{"prompt_tokens":1}}'
+    normal_line = 'data: {"choices":[{"delta":{"content":"x"}}]}'
+    assert filter_usage_chunk(usage_line, client_wants_usage=False) is None
+    assert filter_usage_chunk(usage_line, client_wants_usage=True) == usage_line
+    assert filter_usage_chunk(normal_line, client_wants_usage=False) == normal_line
+    assert filter_usage_chunk("data: [DONE]", False) == "data: [DONE]"
diff --git a/任务拆解与执行计划.md b/任务拆解与执行计划.md
index 0ee0c68..403e820 100644
--- a/任务拆解与执行计划.md
+++ b/任务拆解与执行计划.md
@@ -137,7 +137,7 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
 |---|------|------|--------|
 | T-P0 | 骨架：gateway/proxy/ 包 + SQLite DDL + enabled 门控挂路由 | ✅ 完成 | T-P0 |
 | T-P1 | 鉴权+账本：key 签发/令牌桶/四表/request_id 幂等/日限额 | ✅ 完成 | T-P1 |
-| T-P2 | 上游客户端：流式派发+三家 usage 归一化+首 token 前 failover | ⬜ 待办 | |
+| T-P2 | 上游客户端：流式派发+三家 usage 归一化+首 token 前 failover | ✅ 完成 | T-P2 |
 | T-P3 | 计价+结算：峰谷窗口/毫元整数/黄金用例 ≥10 组 | ⬜ 待办 | |
 | T-P4 | 路由端到端：/proxy/v1 非流式+流式+402/429/413 语义 | ⬜ 待办 | |
 | T-P5 | 规范化+桶：稳定哈希/整形顺序/doc_version 失效/多轮不缓存 | ⬜ 待办 | |