feat(codex): 支持compact端点非流式请求，更新请求头与字段清理

- Codex适配器支持compact端点：非流式请求使用application/json Accept头，
  stream_policy根据compact上下文返回FORCE_NON_STREAM
- 更新Codex请求头格式：添加Version/Connection头，header key首字母大写
- 简化include列表处理：normalizer和request_patching统一强制为固定列表
- 清理Codex不支持的字段：truncation、context_management、user
- 默认instructions改为空字符串
- 前端用量页面：用户页面使用前端筛选后总数，避免不必要的后端分页请求

frontend/src/features/usage/composables/useUsageData.ts

@@ -172,7 +172,7 @@ export function useUsageData(options: UseUsageDataOptions) {
         // 使用 mergeRecordStatus 保护已有的活跃状态，避免轮询更新被覆盖
         const nextRecords = (userData.records || []) as UsageRecord[]
         currentRecords.value = mergeRecordStatus(currentRecords.value, nextRecords)
-        totalRecords.value = currentRecords.value.length
+        totalRecords.value = userData.pagination?.total ?? currentRecords.value.length
 
         // 从记录中提取筛选选项和 API 格式统计
         const models = new Set<string>()

frontend/src/views/shared/Usage.vue

@@ -67,7 +67,7 @@
       :available-providers="availableProviders"
       :current-page="currentPage"
       :page-size="pageSize"
-      :total-records="totalRecords"
+      :total-records="effectiveTotalRecords"
       :page-size-options="pageSizeOptions"
       :auto-refresh="globalAutoRefresh"
       @update:time-range="handleTimeRangeChange"
@@ -391,7 +391,7 @@ onUnmounted(() => {
   stopGlobalAutoRefresh()
 })
 
-// 用户页面的前端分页
+// 用户页面的前端分页（后端一次性返回所有记录，前端分页+筛选）
 const paginatedRecords = computed(() => {
   if (!isAdminPage.value) {
     const start = (currentPage.value - 1) * pageSize.value
@@ -401,6 +401,14 @@ const paginatedRecords = computed(() => {
   return currentRecords.value
 })
 
+// 用户页面使用前端筛选后的总数，管理员页面使用后端返回的总数
+const effectiveTotalRecords = computed(() => {
+  if (!isAdminPage.value) {
+    return filteredRecords.value.length
+  }
+  return totalRecords.value
+})
+
 // 显示的记录
 const displayRecords = computed(() => paginatedRecords.value)
 
@@ -419,20 +427,22 @@ onMounted(async () => {
   const heatmapTask = loadHeatmapData().catch(err => {
     log.error('加载热力图数据失败:', err)
   })
-  const recordsTask = loadRecords(
-    { page: currentPage.value, pageSize: pageSize.value },
-    getCurrentFilters()
-  )
 
-  const tasks: Promise<unknown>[] = [statsTask, heatmapTask, recordsTask]
+  const tasks: Promise<unknown>[] = [statsTask, heatmapTask]
 
   if (isAdminPage.value) {
+    // 管理员页面：stats 和 records 分开加载（后端分页）
+    tasks.push(loadRecords(
+      { page: currentPage.value, pageSize: pageSize.value },
+      getCurrentFilters()
+    ))
     tasks.push(
       usersApi.getAllUsers().then(users => {
         availableUsers.value = users.map(u => ({ id: u.id, username: u.username, email: u.email }))
       })
     )
   }
+  // 用户页面：loadStats 已包含记录加载，不需要单独调用 loadRecords
 
   await Promise.allSettled(tasks)
 })
@@ -442,20 +452,29 @@ async function handleTimeRangeChange(value: DateRangeParams) {
   timeRange.value = value
   currentPage.value = 1 // 重置到第一页
   await loadStats(timeRange.value)
+  if (isAdminPage.value) {
     await loadRecords({ page: 1, pageSize: pageSize.value }, getCurrentFilters())
+  }
+  // 用户页面：loadStats 已包含记录加载
 }
 
 // 处理分页变化
 async function handlePageChange(page: number) {
   currentPage.value = page
+  if (isAdminPage.value) {
     await loadRecords({ page, pageSize: pageSize.value }, getCurrentFilters())
+  }
+  // 用户页面使用前端分页，无需重新请求
 }
 
 // 处理每页大小变化
 async function handlePageSizeChange(size: number) {
   pageSize.value = size
   currentPage.value = 1  // 重置到第一页
+  if (isAdminPage.value) {
     await loadRecords({ page: 1, pageSize: size }, getCurrentFilters())
+  }
+  // 用户页面使用前端分页，无需重新请求
 }
 
 // 获取当前筛选参数
@@ -475,7 +494,11 @@ async function handleFilterSearchChange(value: string) {
   filterSearch.value = value
   currentPage.value = 1
 
+  if (isAdminPage.value) {
     await loadRecords({ page: 1, pageSize: pageSize.value }, getCurrentFilters())
+  }
+  // 用户页面：search 需要重新从后端拉取数据（后端支持 search 参数）
+  // 但通过 filteredRecords 做前端过滤已覆盖，无需额外请求
 }
 
 async function handleFilterUserChange(value: string) {
@@ -526,7 +549,10 @@ async function handleFilterStatusChange(value: string) {
 // 刷新数据
 async function refreshData() {
   await loadStats(timeRange.value)
+  if (isAdminPage.value) {
     await loadRecords({ page: currentPage.value, pageSize: pageSize.value }, getCurrentFilters())
+  }
+  // 用户页面：loadStats 已包含记录加载
 }
 
 // 显示请求详情

src/core/api_format/conversion/normalizers/openai_cli.py

@@ -315,18 +315,16 @@ class OpenAICliNormalizer(FormatNormalizer):
         # Codex 特定设置（覆盖/删除不支持的字段）
         if is_codex:
             result["parallel_tool_calls"] = True
-            # 添加 reasoning.encrypted_content 到 include
+            # 和 codex passthrough patch 保持一致：固定 include 列表
-            include = result.get("include", [])
+            result["include"] = [self._CODEX_REQUIRED_INCLUDE]
-            if not isinstance(include, list):
-                include = []
-            if self._CODEX_REQUIRED_INCLUDE not in include:
-                include.append(self._CODEX_REQUIRED_INCLUDE)
-            result["include"] = include
             # 删除 Codex 不支持的字段
             for key in (
                 "previous_response_id",
                 "service_tier",
                 "max_completion_tokens",
+                "truncation",
+                "context_management",
+                "user",
             ):
                 result.pop(key, None)
 

src/services/provider/adapters/codex/envelope.py

@@ -30,31 +30,34 @@ class CodexOAuthEnvelope:
     """Provider envelope hooks for Codex OAuth upstream."""
 
     name = "codex:oauth"
+    _CODEX_VERSION = "0.101.0"
+    _CODEX_ORIGINATOR = "codex_cli_rs"
 
     def extra_headers(self) -> dict[str, str] | None:
-        # These headers are best-effort: Codex upstream is stricter than public OpenAI API.
+        # Keep these headers provider-scoped to avoid leaking to other upstreams.
-        # Keep them provider-scoped (via ProviderEnvelope) to avoid leaking to other upstreams.
         headers: dict[str, str] = {
-            "OpenAI-Beta": "responses=experimental",
             # Codex upstream is strict about Content-Type; variants like
             # "application/json; charset=utf-8" are rejected.
             "Content-Type": "application/json",
-            "x-oai-web-search-eligible": "true",
+            "Version": self._CODEX_VERSION,
-            "session_id": str(uuid.uuid4()),
+            "Session_id": str(uuid.uuid4()),
-            "originator": "codex_cli_rs",
+            "Connection": "Keep-Alive",
-            # Ensure SSE is returned when upstream is forced to streaming mode.
+            "Originator": self._CODEX_ORIGINATOR,
-            "Accept": "text/event-stream",
         }
 
+        # Compact endpoint is non-stream; normal responses endpoint expects SSE.
+        ctx = get_codex_request_context()
+        is_compact = bool(ctx.is_compact) if ctx else False
+        headers["Accept"] = "application/json" if is_compact else "text/event-stream"
+
         ua = str(getattr(config, "internal_user_agent_openai_cli", "") or "").strip()
         if ua:
             headers["User-Agent"] = ua
 
         # Add chatgpt-account-id from context (set by wrap_request).
         # Context is NOT cleared here — build_codex_url reads is_compact from it later.
-        ctx = get_codex_request_context()
         if ctx and ctx.account_id:
-            headers["chatgpt-account-id"] = ctx.account_id
+            headers["Chatgpt-Account-Id"] = ctx.account_id
 
         return headers
 

src/services/provider/adapters/codex/request_patching.py

@@ -8,14 +8,14 @@ round-trip, so every field the client sent is preserved as-is unless explicitly
 modified here.
 
 Transformations applied:
-- Force ``store=false`` (avoid persistence features not supported by some gateways).
+- Force ``store=false``.
-- Force ``stream=true`` (Codex gateways require streaming).
+- Force ``stream=true`` (except compact requests).
 - Force ``parallel_tool_calls=true``.
-- Ensure ``instructions`` exists (Codex expects it in some deployments).
+- Ensure ``instructions`` exists (empty string when absent).
-- Convert ``role=system`` messages to ``role=developer`` (Codex may not accept ``system``).
+- Convert ``role=system`` messages to ``role=developer``.
 - Drop request parameters known to be rejected by Codex gateways.
-- Ensure ``include`` contains ``"reasoning.encrypted_content"`` for parity with CLI behavior.
+- Force ``include`` to ``["reasoning.encrypted_content"]``.
-- Remove ``previous_response_id`` (not supported by Codex gateways).
+- Drop compatibility-problematic fields (``context_management`` / ``user``).
 """
 
 from __future__ import annotations
@@ -28,11 +28,11 @@ _REJECTED_PARAMS: frozenset[str] = frozenset(
     {
         "max_output_tokens",
         "max_completion_tokens",
-        "max_tokens",
         "temperature",
         "top_p",
         "service_tier",
         "previous_response_id",
+        "truncation",
     }
 )
 
@@ -53,7 +53,11 @@ def patch_openai_cli_request_for_codex(request_body: dict[str, Any]) -> dict[str
     # Codex gateways often reject/ignore persistence; be explicit.
     out["store"] = False
 
-    # Codex gateways require streaming.
+    # Codex compact endpoint is non-streaming; normal responses requires stream=true.
+    is_compact = bool(out.pop("_aether_compact", False))
+    if is_compact:
+        out.pop("stream", None)
+    else:
         out["stream"] = True
 
     # Codex expects parallel tool calls enabled.
@@ -62,7 +66,7 @@ def patch_openai_cli_request_for_codex(request_body: dict[str, Any]) -> dict[str
     # Ensure instructions exists (some gateways require it even if empty).
     instructions = out.get("instructions")
     if not isinstance(instructions, str):
-        out["instructions"] = "You are a helpful coding assistant."
+        out["instructions"] = ""
 
     # Convert "system" role to "developer" (Codex behavior).
     input_items = out.get("input")
@@ -78,23 +82,13 @@ def patch_openai_cli_request_for_codex(request_body: dict[str, Any]) -> dict[str
                 patched_items.append(item)
         out["input"] = patched_items
 
-    # Ensure required include item exists.
+    # Keep codex behavior deterministic: force the exact include list.
-    include = out.get("include")
-    if include is None:
-        out["include"] = [_REQUIRED_INCLUDE_ITEM]
-    elif isinstance(include, str):
-        out["include"] = (
-            [include] if include == _REQUIRED_INCLUDE_ITEM else [include, _REQUIRED_INCLUDE_ITEM]
-        )
-    elif isinstance(include, (list, tuple, set)):
-        include_list = list(include)
-        if _REQUIRED_INCLUDE_ITEM not in include_list:
-            include_list.append(_REQUIRED_INCLUDE_ITEM)
-        out["include"] = include_list
-    else:
-        # Unknown type; overwrite to keep behavior deterministic.
     out["include"] = [_REQUIRED_INCLUDE_ITEM]
 
+    # Codex upstream currently rejects these fields.
+    out.pop("context_management", None)
+    out.pop("user", None)
+
     return out
 
 

src/services/provider/stream_policy.py

@@ -60,6 +60,15 @@ def get_upstream_stream_policy(
     provider_obj = getattr(endpoint, "provider", None)
     pt = str(provider_type or getattr(provider_obj, "provider_type", "") or "").strip().lower()
     sig = str(endpoint_sig or getattr(endpoint, "api_format", "") or "").strip().lower()
+    is_codex_compact = False
+    if pt == ProviderType.CODEX and sig == "openai:cli":
+        try:
+            from src.services.provider.adapters.codex.context import get_codex_request_context
+
+            ctx = get_codex_request_context()
+            is_codex_compact = bool(ctx and ctx.is_compact)
+        except Exception:
+            is_codex_compact = False
 
     # Explicit config wins (unless upstream has a hard constraint).
     cfg = getattr(endpoint, "config", None)
@@ -76,6 +85,7 @@ def get_upstream_stream_policy(
                 pt == ProviderType.CODEX
                 and sig == "openai:cli"
                 and parsed == UpstreamStreamPolicy.FORCE_NON_STREAM
+                and not is_codex_compact
             ):
                 return UpstreamStreamPolicy.FORCE_STREAM
             if pt == ProviderType.KIRO and parsed == UpstreamStreamPolicy.FORCE_NON_STREAM:
@@ -84,7 +94,11 @@ def get_upstream_stream_policy(
 
     # Safe-by-default: Codex Responses OAuth behaves like SSE-only.
     if pt == ProviderType.CODEX and sig == "openai:cli":
-        return UpstreamStreamPolicy.FORCE_STREAM
+        return (
+            UpstreamStreamPolicy.FORCE_NON_STREAM
+            if is_codex_compact
+            else UpstreamStreamPolicy.FORCE_STREAM
+        )
 
     # Kiro upstream streams binary AWS Event Stream; treat as stream-only.
     if pt == ProviderType.KIRO:

tests/services/test_codex_request.py

@@ -12,8 +12,8 @@ def test_patch_openai_cli_request_for_codex_sets_store_and_instructions() -> Non
 
     assert out is not req
     assert out["store"] is False
-    assert "instructions" in out
+    assert out["stream"] is True
-    assert isinstance(out["instructions"], str)
+    assert out["instructions"] == ""
 
 
 def test_patch_openai_cli_request_for_codex_strips_rejected_params() -> None:
@@ -22,20 +22,24 @@ def test_patch_openai_cli_request_for_codex_strips_rejected_params() -> None:
         "input": [],
         "max_output_tokens": 123,
         "max_completion_tokens": 456,
-        "max_tokens": 789,
         "temperature": 0.5,
         "top_p": 0.9,
         "service_tier": "default",
+        "truncation": "auto",
+        "context_management": {"compaction": {"type": "summary"}},
+        "user": "u_123",
     }
     out = patch_openai_cli_request_for_codex(req)
 
     for key in (
         "max_output_tokens",
         "max_completion_tokens",
-        "max_tokens",
         "temperature",
         "top_p",
         "service_tier",
+        "truncation",
+        "context_management",
+        "user",
     ):
         assert key not in out
 
@@ -68,8 +72,28 @@ def test_patch_openai_cli_request_for_codex_adds_required_include_item() -> None
     req = {"model": "gpt-test", "input": []}
     out = patch_openai_cli_request_for_codex(req)
 
-    assert "include" in out
+    assert out["include"] == ["reasoning.encrypted_content"]
-    assert "reasoning.encrypted_content" in out["include"]
+
+
+def test_patch_openai_cli_request_for_codex_overrides_include() -> None:
+    req = {
+        "model": "gpt-test",
+        "input": [],
+        "include": ["foo", "bar"],
+    }
+    out = patch_openai_cli_request_for_codex(req)
+    assert out["include"] == ["reasoning.encrypted_content"]
+
+
+def test_patch_openai_cli_request_for_codex_compact_drops_stream() -> None:
+    req = {
+        "model": "gpt-test",
+        "input": [],
+        "_aether_compact": True,
+        "stream": True,
+    }
+    out = patch_openai_cli_request_for_codex(req)
+    assert "stream" not in out
 
 
 def test_maybe_patch_request_for_codex_is_noop_for_non_codex() -> None:
@@ -110,7 +134,34 @@ def test_codex_envelope_extra_headers_includes_sse_accept_and_session() -> None:
 
     headers = codex_oauth_envelope.extra_headers() or {}
     assert headers.get("Accept") == "text/event-stream"
-    assert headers.get("x-oai-web-search-eligible") == "true"
+    assert headers.get("Originator") == "codex_cli_rs"
-    assert headers.get("originator") == "codex_cli_rs"
+    assert headers.get("Version") == "0.101.0"
-    assert isinstance(headers.get("session_id"), str)
+    assert headers.get("Connection") == "Keep-Alive"
-    assert headers.get("session_id")
+    assert isinstance(headers.get("Session_id"), str)
+    assert headers.get("Session_id")
+
+
+def test_codex_envelope_extra_headers_compact_uses_json_accept() -> None:
+    from src.services.provider.adapters.codex.context import (
+        CodexRequestContext,
+        set_codex_request_context,
+    )
+    from src.services.provider.adapters.codex.envelope import codex_oauth_envelope
+
+    set_codex_request_context(CodexRequestContext(is_compact=True))
+    headers = codex_oauth_envelope.extra_headers() or {}
+    assert headers.get("Accept") == "application/json"
+    set_codex_request_context(None)
+
+
+def test_codex_envelope_extra_headers_uses_account_id_header() -> None:
+    from src.services.provider.adapters.codex.context import (
+        CodexRequestContext,
+        set_codex_request_context,
+    )
+    from src.services.provider.adapters.codex.envelope import codex_oauth_envelope
+
+    set_codex_request_context(CodexRequestContext(account_id="acc_123"))
+    headers = codex_oauth_envelope.extra_headers() or {}
+    assert headers.get("Chatgpt-Account-Id") == "acc_123"
+    set_codex_request_context(None)

tests/services/test_provider_transport_codex.py

@@ -3,6 +3,10 @@ from __future__ import annotations
 from dataclasses import dataclass
 from types import SimpleNamespace
 
+from src.services.provider.adapters.codex.context import (
+    CodexRequestContext,
+    set_codex_request_context,
+)
 from src.services.provider.transport import build_provider_url
 
 
@@ -44,3 +48,19 @@ def test_codex_openai_cli_does_not_duplicate_responses_suffix() -> None:
     )
 
     assert url == "https://chatgpt.com/backend-api/codex/responses"
+
+
+def test_codex_openai_cli_uses_compact_suffix_when_context_marked_compact() -> None:
+    endpoint = _DummyEndpoint(
+        base_url="https://chatgpt.com/backend-api/codex",
+        api_format="openai:cli",
+        provider=SimpleNamespace(provider_type="codex"),
+    )
+    set_codex_request_context(CodexRequestContext(is_compact=True))
+    url = build_provider_url(
+        endpoint,  # type: ignore[arg-type]
+        path_params={"model": "ignored"},
+        is_stream=False,
+    )
+    assert url == "https://chatgpt.com/backend-api/codex/responses/compact"
+    set_codex_request_context(None)

tests/services/test_upstream_stream_policy.py

@@ -3,6 +3,10 @@ from __future__ import annotations
 from dataclasses import dataclass
 from types import SimpleNamespace
 
+from src.services.provider.adapters.codex.context import (
+    CodexRequestContext,
+    set_codex_request_context,
+)
 from src.services.provider.stream_policy import (
     UpstreamStreamPolicy,
     enforce_stream_mode_for_upstream,
@@ -42,6 +46,19 @@ def test_get_upstream_stream_policy_codex_ignores_force_non_stream_config() -> N
     assert get_upstream_stream_policy(ep) == UpstreamStreamPolicy.FORCE_STREAM
 
 
+def test_get_upstream_stream_policy_codex_compact_forces_non_stream() -> None:
+    ep = _DummyEndpoint(
+        api_format="openai:cli",
+        config=None,
+        provider=SimpleNamespace(provider_type="codex"),
+    )
+    try:
+        set_codex_request_context(CodexRequestContext(is_compact=True))
+        assert get_upstream_stream_policy(ep) == UpstreamStreamPolicy.FORCE_NON_STREAM
+    finally:
+        set_codex_request_context(None)
+
+
 def test_enforce_stream_mode_for_upstream_openai_chat_sets_stream_options_usage() -> None:
     body = {"stream": False}
     out = enforce_stream_mode_for_upstream(