feat(codex): 拆分openai:compact为独立端点，简化Codex请求为透传模式

- 新增openai:compact端点类型(EndpointKind.COMPACT)，独立于openai:cli - OpenAICompactAdapter继承OpenAICliAdapter，自动标记compact模式 - Codex请求补丁改为纯透传：仅清理内部标记，不再修改客户端payload - stream_policy支持openai:compact独立策略，compact端点移除stream字段 - candidate_builder支持compact回退到cli端点 - auth_type: vertex_ai重命名为service_account，保持向后兼容 - Vertex Provider新增api_formats与auth_type组合校验 - KeyAllowedModels对话框改为从Provider获取模型，展示provider_model_name - Dialog内Select组件自动禁用Portal，修复层级遮挡问题 - 新增Codex compact端点回填迁移脚本
2026-09-01 17:00:21 +08:00 · 2026-03-01 23:55:26 +08:00
parent 4bf3a453e7
commit 97d42703da
37 changed files with 650 additions and 286 deletions
--- a/alembic/versions/20260301_1700_f0c3a7b9d1e2_backfill_codex_compact_endpoint.py
+++ b/alembic/versions/20260301_1700_f0c3a7b9d1e2_backfill_codex_compact_endpoint.py
@@ -0,0 +1,199 @@
+"""backfill_codex_compact_endpoint
+
+Backfill Codex reverse-proxy endpoints:
+- ensure `openai:cli` endpoint is pinned to force_stream
+- ensure `openai:compact` endpoint exists
+
+Revision ID: f0c3a7b9d1e2
+Revises: 2a624af8dd3a
+Create Date: 2026-03-01 17:00:00.000000+00:00
+"""
+
+from __future__ import annotations
+
+import json
+import uuid
+from typing import Any
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "f0c3a7b9d1e2"
+down_revision = "2a624af8dd3a"
+branch_labels = None
+depends_on = None
+
+_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
+_COMPACT_FORMAT = "openai:compact"
+_CLI_FORMAT = "openai:cli"
+_FORCE_STREAM = "force_stream"
+
+
+def _find_codex_provider_ids(conn: sa.Connection) -> list[str]:
+    """Find Codex providers (by provider_type or legacy base_url pattern)."""
+    rows = conn.execute(sa.text("""
+        SELECT DISTINCT p.id
+        FROM providers p
+        LEFT JOIN provider_endpoints pe ON pe.provider_id = p.id
+        WHERE lower(COALESCE(p.provider_type, '')) = 'codex'
+           OR (
+                lower(COALESCE(pe.api_format, '')) = 'openai:cli'
+            AND lower(COALESCE(pe.base_url, '')) LIKE '%/backend-api/codex%'
+           )
+    """))
+    return [str(r[0]) for r in rows if r[0]]
+
+
+def _get_cli_endpoint(conn: sa.Connection, provider_id: str) -> dict[str, Any] | None:
+    """Load existing openai:cli endpoint for the provider."""
+    row = (
+        conn.execute(
+            sa.text("""
+                SELECT base_url, header_rules, body_rules, max_retries, proxy, config
+                FROM provider_endpoints
+                WHERE provider_id = :pid AND api_format = :fmt
+                LIMIT 1
+            """),
+            {"pid": provider_id, "fmt": _CLI_FORMAT},
+        )
+        .mappings()
+        .first()
+    )
+    return dict(row) if row else None
+
+
+def _pin_cli_force_stream(conn: sa.Connection, provider_id: str, cli: dict[str, Any]) -> None:
+    """Set upstream_stream_policy=force_stream on existing cli endpoint."""
+    cfg = dict(cli.get("config") or {}) if isinstance(cli.get("config"), dict) else {}
+    cfg.pop("upstreamStreamPolicy", None)
+    cfg.pop("upstream_stream", None)
+    cfg["upstream_stream_policy"] = _FORCE_STREAM
+    conn.execute(
+        sa.text("""
+            UPDATE provider_endpoints
+            SET api_family = 'openai',
+                endpoint_kind = 'cli',
+                config = CAST(:config AS json),
+                updated_at = CURRENT_TIMESTAMP
+            WHERE provider_id = :pid AND api_format = :fmt
+        """),
+        {
+            "pid": provider_id,
+            "fmt": _CLI_FORMAT,
+            "config": json.dumps(cfg, ensure_ascii=False),
+        },
+    )
+
+
+def _ensure_compact_endpoint(conn: sa.Connection, provider_id: str, cli: dict[str, Any]) -> None:
+    """Create openai:compact endpoint if missing (clone from cli)."""
+    exists = conn.execute(
+        sa.text(
+            "SELECT 1 FROM provider_endpoints WHERE provider_id = :pid AND api_format = :fmt LIMIT 1"
+        ),
+        {"pid": provider_id, "fmt": _COMPACT_FORMAT},
+    ).first()
+    if exists:
+        # Already exists, just ensure api_family/endpoint_kind are set.
+        conn.execute(
+            sa.text("""
+                UPDATE provider_endpoints
+                SET api_family = 'openai', endpoint_kind = 'compact',
+                    updated_at = CURRENT_TIMESTAMP
+                WHERE provider_id = :pid AND api_format = :fmt
+            """),
+            {"pid": provider_id, "fmt": _COMPACT_FORMAT},
+        )
+        return
+
+    # Clone from cli endpoint, strip stream policy.
+    cfg = dict(cli.get("config") or {}) if isinstance(cli.get("config"), dict) else {}
+    for k in ("upstream_stream_policy", "upstreamStreamPolicy", "upstream_stream"):
+        cfg.pop(k, None)
+
+    def _json(val: Any) -> str | None:
+        return json.dumps(val, ensure_ascii=False) if val is not None else None
+
+    conn.execute(
+        sa.text("""
+            INSERT INTO provider_endpoints (
+                id, provider_id, api_format, api_family, endpoint_kind,
+                base_url, custom_path, header_rules, body_rules,
+                max_retries, is_active, config, format_acceptance_config,
+                proxy, created_at, updated_at
+            ) VALUES (
+                :id, :pid, :fmt, 'openai', 'compact',
+                :base_url, NULL, CAST(:header_rules AS json), CAST(:body_rules AS json),
+                :max_retries, TRUE, CAST(:config AS json), NULL,
+                CAST(:proxy AS jsonb), CURRENT_TIMESTAMP, CURRENT_TIMESTAMP
+            )
+        """),
+        {
+            "id": str(uuid.uuid4()),
+            "pid": provider_id,
+            "fmt": _COMPACT_FORMAT,
+            "base_url": cli.get("base_url") or _CODEX_BASE_URL,
+            "header_rules": _json(cli.get("header_rules")),
+            "body_rules": _json(cli.get("body_rules")),
+            "max_retries": cli.get("max_retries") or 2,
+            "config": _json(cfg or None),
+            "proxy": _json(cli.get("proxy")),
+        },
+    )
+
+
+def _add_compact_to_key_formats(conn: sa.Connection, provider_id: str) -> None:
+    """Ensure provider keys include openai:compact in api_formats."""
+    rows = (
+        conn.execute(
+            sa.text("SELECT id, api_formats FROM provider_api_keys WHERE provider_id = :pid"),
+            {"pid": provider_id},
+        )
+        .mappings()
+        .all()
+    )
+    for row in rows:
+        raw = row["api_formats"]
+        formats: list[str] = []
+        if isinstance(raw, list):
+            for item in raw:
+                v = str(item or "").strip().lower()
+                if v and v not in formats:
+                    formats.append(v)
+
+        if _COMPACT_FORMAT in formats:
+            continue
+
+        # Insert compact right after cli, or at end.
+        if _CLI_FORMAT in formats:
+            idx = formats.index(_CLI_FORMAT) + 1
+            formats.insert(idx, _COMPACT_FORMAT)
+        else:
+            formats.append(_COMPACT_FORMAT)
+
+        conn.execute(
+            sa.text("""
+                UPDATE provider_api_keys
+                SET api_formats = CAST(:fmts AS json), updated_at = CURRENT_TIMESTAMP
+                WHERE id = :id
+            """),
+            {"id": row["id"], "fmts": json.dumps(formats, ensure_ascii=False)},
+        )
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    for provider_id in _find_codex_provider_ids(conn):
+        cli = _get_cli_endpoint(conn, provider_id)
+        if not cli:
+            continue  # No cli endpoint to clone from; skip.
+        _pin_cli_force_stream(conn, provider_id, cli)
+        _ensure_compact_endpoint(conn, provider_id, cli)
+        _add_compact_to_key_formats(conn, provider_id)
+
+
+def downgrade() -> None:
+    # Data backfill: no-op to avoid deleting user-managed data.
+    return
--- a/frontend/src/api/endpoints/types/api-format.ts
+++ b/frontend/src/api/endpoints/types/api-format.ts
@@ -5,6 +5,7 @@ export const API_FORMATS = {
  CLAUDE_CLI: 'claude:cli',
  OPENAI: 'openai:chat',
  OPENAI_CLI: 'openai:cli',
+  OPENAI_COMPACT: 'openai:compact',
  OPENAI_VIDEO: 'openai:video',
  GEMINI: 'gemini:chat',
  GEMINI_CLI: 'gemini:cli',
@@ -19,6 +20,7 @@ export const API_FORMAT_LABELS: Record<string, string> = {
  [API_FORMATS.CLAUDE_CLI]: 'Claude CLI',
  [API_FORMATS.OPENAI]: 'OpenAI Chat',
  [API_FORMATS.OPENAI_CLI]: 'OpenAI CLI',
+  [API_FORMATS.OPENAI_COMPACT]: 'OpenAI Compact',
  [API_FORMATS.OPENAI_VIDEO]: 'OpenAI Video',
  [API_FORMATS.GEMINI]: 'Gemini Chat',
  [API_FORMATS.GEMINI_CLI]: 'Gemini CLI',
@@ -28,6 +30,7 @@ export const API_FORMAT_LABELS: Record<string, string> = {
  CLAUDE_CLI: 'Claude CLI',
  OPENAI: 'OpenAI Chat',
  OPENAI_CLI: 'OpenAI CLI',
+  OPENAI_COMPACT: 'OpenAI Compact',
  OPENAI_VIDEO: 'OpenAI Video',
  GEMINI: 'Gemini Chat',
  GEMINI_CLI: 'Gemini CLI',
@@ -38,6 +41,7 @@ export const API_FORMAT_LABELS: Record<string, string> = {
 export const API_FORMAT_SHORT: Record<string, string> = {
  [API_FORMATS.OPENAI]: 'O',
  [API_FORMATS.OPENAI_CLI]: 'OC',
+  [API_FORMATS.OPENAI_COMPACT]: 'OCP',
  [API_FORMATS.OPENAI_VIDEO]: 'OV',
  [API_FORMATS.CLAUDE]: 'C',
  [API_FORMATS.CLAUDE_CLI]: 'CC',
@@ -47,6 +51,7 @@ export const API_FORMAT_SHORT: Record<string, string> = {
  // legacy 兼容（仅用于展示历史数据）
  OPENAI: 'O',
  OPENAI_CLI: 'OC',
+  OPENAI_COMPACT: 'OCP',
  OPENAI_VIDEO: 'OV',
  CLAUDE: 'C',
  CLAUDE_CLI: 'CC',
@@ -59,6 +64,7 @@ export const API_FORMAT_SHORT: Record<string, string> = {
 export const API_FORMAT_ORDER: string[] = [
  API_FORMATS.OPENAI,
  API_FORMATS.OPENAI_CLI,
+  API_FORMATS.OPENAI_COMPACT,
  API_FORMATS.OPENAI_VIDEO,
  API_FORMATS.CLAUDE,
  API_FORMATS.CLAUDE_CLI,
--- a/frontend/src/components/ui/dialog/Dialog.vue
+++ b/frontend/src/components/ui/dialog/Dialog.vue
@@ -92,8 +92,9 @@
 </template>

 <script setup lang="ts">
-import { computed, useSlots, type Component } from 'vue'
+import { computed, provide, useSlots, type Component } from 'vue'
 import { useEscapeKey } from '@/composables/useEscapeKey'
+import { DIALOG_CONTEXT_KEY } from './context'

 // Props 定义
 const props = defineProps<{
@@ -116,6 +117,8 @@ const emit = defineEmits<{
  'update:modelValue': [value: boolean]
 }>()

+provide(DIALOG_CONTEXT_KEY, true)
+
 // 获取 slots 以便在模板中使用
 const slots = useSlots()

--- a/frontend/src/components/ui/dialog/context.ts
+++ b/frontend/src/components/ui/dialog/context.ts
@@ -0,0 +1,3 @@
+import type { InjectionKey } from 'vue'
+
+export const DIALOG_CONTEXT_KEY: InjectionKey<boolean> = Symbol('dialog-context')
--- a/frontend/src/components/ui/select-content.vue
+++ b/frontend/src/components/ui/select-content.vue
@@ -1,5 +1,5 @@
 <template>
-  <SelectPortal>
+  <SelectPortal :disabled="shouldDisablePortal">
    <SelectContentPrimitive
      v-bind="$attrs"
      :class="contentClass"
@@ -23,7 +23,8 @@ import {
  SelectViewport,
 } from 'radix-vue'
 import { cn } from '@/lib/utils'
-import { computed } from 'vue'
+import { computed, inject } from 'vue'
+import { DIALOG_CONTEXT_KEY } from './dialog/context'

 interface Props {
  class?: string
@@ -32,6 +33,7 @@ interface Props {
  sideOffset?: number
  align?: 'start' | 'center' | 'end'
  alignOffset?: number
+  disablePortal?: boolean
 }

 const props = withDefaults(defineProps<Props>(), {
@@ -41,8 +43,12 @@ const props = withDefaults(defineProps<Props>(), {
  sideOffset: 4,
  align: undefined,
  alignOffset: undefined,
+  disablePortal: false,
 })

+const isInsideDialog = inject(DIALOG_CONTEXT_KEY, false)
+const shouldDisablePortal = computed(() => props.disablePortal || isInsideDialog)
+
 const contentClass = computed(() =>
  cn(
    'z-[200] max-h-96 min-w-[8rem] overflow-hidden rounded-2xl border border-border bg-card text-foreground shadow-2xl backdrop-blur-xl pointer-events-auto',
--- a/frontend/src/features/providers/components/EndpointFormDialog.vue
+++ b/frontend/src/features/providers/components/EndpointFormDialog.vue
@@ -57,7 +57,7 @@
                  size="icon"
                  :class="getUpstreamStreamButtonClass(endpoint)"
                  :title="getUpstreamStreamTooltip(endpoint)"
-                  :disabled="savingEndpointId === endpoint.id"
+                  :disabled="savingEndpointId === endpoint.id || isUpstreamStreamPolicyLocked(endpoint)"
                  @click="handleCycleUpstreamStream(endpoint)"
                >
                  <Radio class="w-3.5 h-3.5" />
@@ -2029,12 +2029,21 @@ async function handleToggleFormatConversion(endpoint: ProviderEndpoint) {

 // 获取上游流式按钮的当前状态（优先使用编辑状态）
 function getCurrentUpstreamStreamPolicy(endpoint: ProviderEndpoint): string {
+  if (isUpstreamStreamPolicyLocked(endpoint)) return 'force_stream'
  const state = endpointEditStates.value[endpoint.id]
  return state?.upstreamStreamPolicy ?? getEndpointUpstreamStreamPolicy(endpoint)
 }

+function isUpstreamStreamPolicyLocked(endpoint: ProviderEndpoint): boolean {
+  return (props.provider?.provider_type || '').toLowerCase() === 'codex'
+    && endpoint.api_format === 'openai:cli'
+}
+
 // 获取上游流式按钮的样式类
 function getUpstreamStreamButtonClass(endpoint: ProviderEndpoint): string {
+  if (isUpstreamStreamPolicyLocked(endpoint)) {
+    return 'h-7 w-7 text-primary/70 cursor-not-allowed'
+  }
  const policy = getCurrentUpstreamStreamPolicy(endpoint)
  const base = 'h-7 w-7'
  if (policy === 'force_stream') return `${base} text-primary`
@@ -2044,6 +2053,7 @@ function getUpstreamStreamButtonClass(endpoint: ProviderEndpoint): string {

 // 获取上游流式按钮的提示文字
 function getUpstreamStreamTooltip(endpoint: ProviderEndpoint): string {
+  if (isUpstreamStreamPolicyLocked(endpoint)) return '固定流式（Codex OpenAI CLI，已锁定）'
  const policy = getCurrentUpstreamStreamPolicy(endpoint)
  if (policy === 'force_stream') return '固定流式（点击切换为固定非流）'
  if (policy === 'force_non_stream') return '固定非流（点击切换为跟随请求）'
@@ -2052,6 +2062,8 @@ function getUpstreamStreamTooltip(endpoint: ProviderEndpoint): string {

 // 循环切换上游流式策略并直接保存
 async function handleCycleUpstreamStream(endpoint: ProviderEndpoint) {
+  if (isUpstreamStreamPolicyLocked(endpoint)) return
+
  const currentPolicy = getCurrentUpstreamStreamPolicy(endpoint)
  let nextPolicy: string
  let nextLabel: string
--- a/frontend/src/features/providers/components/KeyAllowedModelsEditDialog.vue
+++ b/frontend/src/features/providers/components/KeyAllowedModelsEditDialog.vue
@@ -72,7 +72,7 @@
          <div class="max-h-96 overflow-y-auto">
            <!-- 加载中 -->
            <div
-              v-if="loadingGlobalModels"
+              v-if="loadingProviderModels"
              class="flex items-center justify-center py-12"
            >
              <Loader2 class="w-6 h-6 animate-spin text-primary" />
@@ -152,7 +152,7 @@
              </div>

              <!-- 提供商模型 -->
-              <template v-if="filteredGlobalModels.length > 0">
+              <template v-if="filteredProviderModels.length > 0">
                <!-- 标题 sticky top -->
                <div
                  class="flex items-center justify-between px-3 py-2 bg-muted sticky top-0 z-20 cursor-pointer hover:bg-muted/80 transition-colors"
@@ -164,14 +164,14 @@
                      :class="collapsedGroups.has('global') ? '-rotate-90' : ''"
                    />
                    <span class="text-xs font-medium">提供商模型</span>
-                    <span class="text-xs text-muted-foreground">({{ filteredGlobalModels.length }})</span>
+                    <span class="text-xs text-muted-foreground">({{ filteredProviderModels.length }})</span>
                  </div>
                  <button
                    type="button"
                    class="text-xs text-primary hover:underline"
-                    @click.stop="toggleAllGlobalModels"
+                    @click.stop="toggleAllProviderModels"
                  >
-                    {{ isAllGlobalModelsSelected ? '取消全选' : '全选' }}
+                    {{ isAllProviderModelsSelected ? '取消全选' : '全选' }}
                  </button>
                </div>
                <!-- 内容 -->
@@ -180,7 +180,7 @@
                  class="space-y-1 p-2"
                >
                  <div
-                    v-for="model in filteredGlobalModels"
+                    v-for="model in filteredProviderModels"
                    :key="model.name"
                    class="flex items-center gap-2 px-2 py-1.5 rounded hover:bg-muted cursor-pointer"
                    @click="toggleModel(model.name)"
@@ -196,11 +196,14 @@
                    </div>
                    <div class="flex-1 min-w-0">
                      <p class="text-sm font-medium truncate">
-                        {{ model.display_name }}
-                      </p>
-                      <p class="text-xs text-muted-foreground truncate font-mono">
                        {{ model.name }}
                      </p>
+                      <p
+                        v-if="model.global_model_display_name || model.global_model_name"
+                        class="text-xs text-muted-foreground truncate font-mono"
+                      >
+                        {{ model.global_model_display_name || model.global_model_name }}
+                      </p>
                    </div>
                    <button
                      v-if="selectedModels.includes(model.name)"
@@ -366,17 +369,19 @@ import { useToast } from '@/composables/useToast'
 import { useConfirm } from '@/composables/useConfirm'
 import { parseApiError } from '@/utils/errorParser'
 import {
+  getProviderModels,
  updateProviderKey,
  type EndpointAPIKey,
  type AllowedModels,
 } from '@/api/endpoints'
-import { getGlobalModels, type GlobalModelResponse } from '@/api/global-models'
 import { useUpstreamModelsCache } from '../composables/useUpstreamModelsCache'
 import { API_FORMAT_SHORT, type UpstreamModel } from '@/api/endpoints/types'

 interface AvailableModel {
  name: string
  display_name: string
+  global_model_name?: string
+  global_model_display_name?: string
 }

 const props = defineProps<{
@@ -396,7 +401,7 @@ const { fetchModels: fetchCachedModels } = useUpstreamModelsCache()

 const isOpen = computed(() => props.open)
 const saving = ref(false)
-const loadingGlobalModels = ref(false)
+const loadingProviderModels = ref(false)
 const fetchingUpstreamModels = ref(false)
 const upstreamModelsLoaded = ref(false)

@@ -406,8 +411,8 @@ let loadingCancelled = false
 // 搜索
 const searchQuery = ref('')

-// 可用模型列表（全局模型）
-const allGlobalModels = ref<AvailableModel[]>([])
+// 可用模型列表（该 Provider 已关联的模型）
+const allProviderModels = ref<AvailableModel[]>([])
 // 上游模型列表（从 API 查询获取）
 const upstreamModels = ref<UpstreamModel[]>([])

@@ -427,7 +432,7 @@ const isAutoFetchMode = computed(() => props.apiKey?.auto_fetch_models ?? false)

 // 空状态判断
 const showEmptyState = computed(() => {
-  return filteredGlobalModels.value.length === 0 &&
+  return filteredProviderModels.value.length === 0 &&
         filteredUpstreamModels.value.length === 0 &&
         customModels.value.length === 0
 })
@@ -450,22 +455,22 @@ const hasChanges = computed(() => {
  return sortedLocked1.some((v, i) => v !== sortedLocked2[i])
 })

-// 所有已知模型的集合（全局 + 上游模型）
+// 所有已知模型的集合（提供商模型 + 上游模型）
 const allKnownModels = computed(() => {
  const set = new Set<string>()
-  allGlobalModels.value.forEach(m => set.add(m.name))
+  allProviderModels.value.forEach(m => set.add(m.name))
  upstreamModels.value.forEach(m => set.add(m.id))
  return set
 })

-// 全局模型名称集合（用于判断模型是否为全局模型）
-const globalModelNamesSet = computed(() => {
-  return new Set(allGlobalModels.value.map(m => m.name))
+// 提供商模型名称集合（用于判断模型是否为“提供商模型”分组项）
+const providerModelNamesSet = computed(() => {
+  return new Set(allProviderModels.value.map(m => m.name))
 })

-// 判断模型是否为全局模型（提供商模型）
-function isGlobalModel(modelId: string): boolean {
-  return globalModelNamesSet.value.has(modelId)
+// 判断模型是否为“提供商模型”分组项
+function isProviderModel(modelId: string): boolean {
+  return providerModelNamesSet.value.has(modelId)
 }

 // 上游模型列表（后端已按 id 聚合，包含 api_formats 数组）
@@ -541,27 +546,29 @@ const canAddAsCustom = computed(() => {
  if (selectedModels.value.includes(search)) return false
  // 已经在自定义模型列表中就不显示
  if (allCustomModels.value.includes(search)) return false
-  // 精确匹配全局模型就不显示
-  if (allGlobalModels.value.some(m => m.name === search)) return false
+  // 精确匹配提供商模型就不显示
+  if (allProviderModels.value.some(m => m.name === search)) return false
  // 精确匹配上游模型就不显示
  if (upstreamModelNames.value.includes(search)) return false
  return true
 })

-// 搜索过滤后的全局模型
-const filteredGlobalModels = computed(() => {
-  if (!searchQuery.value.trim()) return allGlobalModels.value
+// 搜索过滤后的提供商模型
+const filteredProviderModels = computed(() => {
+  if (!searchQuery.value.trim()) return allProviderModels.value
  const query = searchQuery.value.toLowerCase()
-  return allGlobalModels.value.filter(m =>
+  return allProviderModels.value.filter(m =>
    m.name.toLowerCase().includes(query) ||
-    m.display_name.toLowerCase().includes(query)
+    m.display_name.toLowerCase().includes(query) ||
+    (m.global_model_name || '').toLowerCase().includes(query) ||
+    (m.global_model_display_name || '').toLowerCase().includes(query)
  )
 })

-// 全局模型是否全选
-const isAllGlobalModelsSelected = computed(() => {
-  if (filteredGlobalModels.value.length === 0) return false
-  return filteredGlobalModels.value.every(m => selectedModels.value.includes(m.name))
+// 提供商模型是否全选
+const isAllProviderModelsSelected = computed(() => {
+  if (filteredProviderModels.value.length === 0) return false
+  return filteredProviderModels.value.every(m => selectedModels.value.includes(m.name))
 })

 // 切换模型选中状态
@@ -569,9 +576,9 @@ function toggleModel(modelId: string) {
  const idx = selectedModels.value.indexOf(modelId)
  if (idx === -1) {
    selectedModels.value.push(modelId)
-    // 自动获取模式下，勾选全局模型时自动锁定
+    // 自动获取模式下，勾选提供商模型时自动锁定
    // 防止下次刷新时被覆盖（即使全局模型与上游模型同名）
-    if (isAutoFetchMode.value && isGlobalModel(modelId)) {
+    if (isAutoFetchMode.value && isProviderModel(modelId)) {
      if (!lockedModels.value.includes(modelId)) {
        lockedModels.value.push(modelId)
      }
@@ -618,10 +625,10 @@ function addCustomModel() {
  }
 }

-// 全选/取消全选全局模型
-function toggleAllGlobalModels() {
-  const allNames = filteredGlobalModels.value.map(m => m.name)
-  if (isAllGlobalModelsSelected.value) {
+// 全选/取消全选提供商模型
+function toggleAllProviderModels() {
+  const allNames = filteredProviderModels.value.map(m => m.name)
+  if (isAllProviderModelsSelected.value) {
    // 取消全选
    selectedModels.value = selectedModels.value.filter(id => !allNames.includes(id))
    // 同时取消锁定
@@ -631,7 +638,7 @@ function toggleAllGlobalModels() {
    allNames.forEach(name => {
      if (!selectedModels.value.includes(name)) {
        selectedModels.value.push(name)
-        // 自动获取模式下，勾选全局模型时自动锁定
+        // 自动获取模式下，勾选提供商模型时自动锁定
        if (isAutoFetchMode.value && !lockedModels.value.includes(name)) {
          lockedModels.value.push(name)
        }
@@ -650,21 +657,23 @@ function toggleGroupCollapse(group: string) {
  collapsedGroups.value = new Set(collapsedGroups.value)
 }

-// 加载全局模型
-async function loadGlobalModels() {
-  loadingGlobalModels.value = true
+// 加载该 Provider 已关联的模型
+async function loadProviderModels() {
+  loadingProviderModels.value = true
  try {
-    const response = await getGlobalModels({ limit: 1000 })
+    const response = await getProviderModels(props.providerId, { limit: 1000 })
    if (loadingCancelled) return
-    allGlobalModels.value = response.models.map((m: GlobalModelResponse) => ({
-      name: m.name,
-      display_name: m.display_name
+    allProviderModels.value = response.map(m => ({
+      name: m.provider_model_name,
+      display_name: m.global_model_display_name || m.global_model_name || m.provider_model_name,
+      global_model_name: m.global_model_name,
+      global_model_display_name: m.global_model_display_name
    }))
  } catch {
    if (loadingCancelled) return
-    showError('加载全局模型失败', '错误')
+    showError('加载提供商模型失败', '错误')
  } finally {
-    loadingGlobalModels.value = false
+    loadingProviderModels.value = false
  }
 }

@@ -734,8 +743,8 @@ watch(() => props.open, async (open) => {
      collapsedGroups.value = new Set()
    }

-    // 加载全局模型
-    await loadGlobalModels()
+    // 加载该 Provider 已关联模型
+    await loadProviderModels()

    // 自动获取模式下，获取上游模型用于显示（但选中状态使用已保存的 allowed_models）
    if (props.apiKey.auto_fetch_models) {
@@ -745,11 +754,11 @@ watch(() => props.open, async (open) => {
      // selectedModels 已在上面从 props.apiKey.allowed_models 初始化
    }

-    // 提取自定义模型（不在全局模型和上游模型中的）
+    // 提取自定义模型（不在提供商模型和上游模型中的）
    const upstreamModelIdsSet = new Set(upstreamModels.value.map(m => m.id))
    // 自定义模型是用户手动添加的、不在已知模型列表中的
    allCustomModels.value = selectedModels.value.filter(m =>
-      !globalModelNamesSet.value.has(m) && !upstreamModelIdsSet.has(m)
+      !providerModelNamesSet.value.has(m) && !upstreamModelIdsSet.has(m)
    )
  } else {
    loadingCancelled = true
--- a/frontend/src/features/providers/composables/useEndpointStatus.ts
+++ b/frontend/src/features/providers/composables/useEndpointStatus.ts
@@ -8,6 +8,7 @@ const ENDPOINT_SORT_ORDER = [
  'claude:cli',
  'openai:chat',
  'openai:cli',
+  'openai:compact',
  'gemini:chat',
  'gemini:cli',
  'openai:video',
--- a/frontend/src/features/providers/composables/useProviderFilters.ts
+++ b/frontend/src/features/providers/composables/useProviderFilters.ts
@@ -28,6 +28,7 @@ export function useProviderFilters(
    { value: 'claude:cli', label: 'Claude CLI' },
    { value: 'openai:chat', label: 'OpenAI Chat' },
    { value: 'openai:cli', label: 'OpenAI CLI' },
+    { value: 'openai:compact', label: 'OpenAI Compact' },
    { value: 'gemini:chat', label: 'Gemini Chat' },
    { value: 'gemini:cli', label: 'Gemini CLI' },
  ]
--- a/frontend/src/features/providers/utils/providerTypeUtils.ts
+++ b/frontend/src/features/providers/utils/providerTypeUtils.ts
@@ -0,0 +1,19 @@
+/**
+ * Provider 类型判断工具函数。
+ *
+ * 区分"密钥型"和"OAuth 账号型"两类 Provider，影响前端显示标签和操作入口。
+ */
+
+const oauthAccountProviderTypes = new Set([
+  'claude_code',
+  'codex',
+  'gemini_cli',
+  'antigravity',
+  'kiro',
+])
+
+export const isOAuthAccountProviderType = (providerType?: string | null): boolean =>
+  oauthAccountProviderTypes.has((providerType || '').toLowerCase())
+
+export const isKeyManagedProviderType = (providerType?: string | null): boolean =>
+  !isOAuthAccountProviderType(providerType)
--- a/frontend/src/features/usage/components/UsageRecordsTable.vue
+++ b/frontend/src/features/usage/components/UsageRecordsTable.vue
@@ -723,6 +723,7 @@ const emit = defineEmits<{
 const AVAILABLE_API_FORMATS = [
  { value: 'openai:chat', label: 'OpenAI Chat' },
  { value: 'openai:cli', label: 'OpenAI CLI' },
+  { value: 'openai:compact', label: 'OpenAI Compact' },
  { value: 'openai:video', label: 'OpenAI Video' },
  { value: 'claude:chat', label: 'Claude Chat' },
  { value: 'claude:cli', label: 'Claude CLI' },
--- a/frontend/src/features/usage/conversation/tests/gemini.spec.ts
+++ b/frontend/src/features/usage/conversation/tests/gemini.spec.ts
@@ -0,0 +1,52 @@
+import { describe, it, expect } from 'vitest'
+import { parseResponse, renderResponse } from '../registry'
+
+describe('Gemini conversation parser', () => {
+  const requestBody = { model: 'gemini-3-pro-preview' }
+  const normalizedResponse = {
+    status: 'completed',
+    output: [
+      {
+        type: 'message',
+        role: 'assistant',
+        content: [
+          {
+            type: 'output_text',
+            text: 'Hello!',
+          },
+        ],
+      },
+    ],
+  }
+
+  it('parses normalized output payload when hint is gemini', () => {
+    const parsed = parseResponse(normalizedResponse, requestBody, 'gemini:chat')
+    expect(parsed.apiFormat).toBe('gemini')
+    expect(parsed.messages).toHaveLength(1)
+    expect(parsed.messages[0]?.role).toBe('assistant')
+    expect(parsed.messages[0]?.content[0]).toMatchObject({
+      type: 'text',
+      text: 'Hello!',
+    })
+  })
+
+  it('renders normalized output payload when hint is gemini', () => {
+    const rendered = renderResponse(normalizedResponse, requestBody, 'gemini:chat')
+    expect(rendered.error).toBeUndefined()
+    expect(rendered.blocks).toHaveLength(1)
+    expect(rendered.blocks[0]).toMatchObject({
+      type: 'message',
+      role: 'assistant',
+    })
+
+    const firstBlock = rendered.blocks[0]
+    if (!firstBlock || firstBlock.type !== 'message') {
+      throw new Error('expected first render block to be message')
+    }
+
+    expect(firstBlock.content[0]).toMatchObject({
+      type: 'text',
+      content: 'Hello!',
+    })
+  })
+})
--- a/src/api/admin/models/routing.py
+++ b/src/api/admin/models/routing.py
@@ -422,6 +422,7 @@ class AdminGetModelRoutingPreviewAdapter(AdminApiAdapter):
            preferred_order = [
                "openai:chat",
                "openai:cli",
+                "openai:compact",
                "openai:video",
                "claude:chat",
                "claude:cli",
--- a/src/api/admin/system.py
+++ b/src/api/admin/system.py
@@ -838,15 +838,20 @@ class AdminGetApiFormatsAdapter(AdminApiAdapter):
        def _label_for(sig: str) -> str:
            fam, kind = (sig.split(":", 1) + [""])[:2]
            fam_title = {"claude": "Claude", "openai": "OpenAI", "gemini": "Gemini"}.get(fam, fam)
-            kind_title = {"chat": "Chat", "cli": "CLI", "video": "Video", "image": "Image"}.get(
-                kind, kind
-            )
+            kind_title = {
+                "chat": "Chat",
+                "cli": "CLI",
+                "compact": "Compact",
+                "video": "Video",
+                "image": "Image",
+            }.get(kind, kind)
            return f"{fam_title} {kind_title}".strip()

        endpoint_defs = list_endpoint_definitions()
        preferred_order = [
            "openai:chat",
            "openai:cli",
+            "openai:compact",
            "openai:video",
            "claude:chat",
            "claude:cli",
--- a/src/api/handlers/openai_cli/init.py
+++ b/src/api/handlers/openai_cli/init.py
@@ -2,10 +2,11 @@
 OpenAI CLI 透传处理器
 """

-from src.api.handlers.openai_cli.adapter import OpenAICliAdapter
+from src.api.handlers.openai_cli.adapter import OpenAICliAdapter, OpenAICompactAdapter
 from src.api.handlers.openai_cli.handler import OpenAICliMessageHandler

 __all__ = [
    "OpenAICliAdapter",
+    "OpenAICompactAdapter",
    "OpenAICliMessageHandler",
 ]
--- a/src/api/handlers/openai_cli/adapter.py
+++ b/src/api/handlers/openai_cli/adapter.py
@@ -15,7 +15,7 @@ from src.api.handlers.base.cli_adapter_base import CliAdapterBase, register_cli_
 from src.api.handlers.base.cli_handler_base import CliMessageHandlerBase
 from src.api.handlers.openai.adapter import OpenAIChatAdapter
 from src.config.settings import config
-from src.core.api_format import ApiFamily
+from src.core.api_format import ApiFamily, EndpointKind
 from src.utils.url_utils import is_codex_url


@@ -166,3 +166,18 @@ class OpenAICliAdapter(CliAdapterBase):


 __all__ = ["OpenAICliAdapter"]
+
+
+@register_cli_adapter
+class OpenAICompactAdapter(OpenAICliAdapter):
+    """OpenAI Compact Responses adapter (/v1/responses/compact)."""
+
+    FORMAT_ID = "openai:compact"
+    ENDPOINT_KIND = EndpointKind.COMPACT
+    name = "openai.compact"
+
+    def __init__(self, allowed_api_formats: list[str] | None = None):
+        super().__init__(allowed_api_formats=allowed_api_formats, compact=True)
+
+
+__all__.append("OpenAICompactAdapter")
--- a/src/api/public/models.py
+++ b/src/api/public/models.py
@@ -36,7 +36,7 @@ router = APIRouter(tags=["System Catalog"])

 # 各格式对应的 API 格式列表（包括对应的 CLI 格式）
 _CLAUDE_FORMATS = ["claude:chat", "claude:cli"]
-_OPENAI_FORMATS = ["openai:chat", "openai:cli"]
+_OPENAI_FORMATS = ["openai:chat", "openai:cli", "openai:compact"]
 _GEMINI_FORMATS = ["gemini:chat", "gemini:cli"]

 # 所有格式（用于格式转换时的查询）
--- a/src/api/public/openai.py
+++ b/src/api/public/openai.py
@@ -15,7 +15,7 @@ from sqlalchemy.orm import Session

 from src.api.base.pipeline import ApiRequestPipeline
 from src.api.handlers.openai import OpenAIChatAdapter
-from src.api.handlers.openai_cli import OpenAICliAdapter
+from src.api.handlers.openai_cli import OpenAICliAdapter, OpenAICompactAdapter
 from src.database import get_db

 router = APIRouter(tags=["OpenAI API"])
@@ -68,7 +68,7 @@ async def create_responses_compact(

    **认证方式**: Bearer Token（API Key 或 JWT Token）
    """
-    adapter = OpenAICliAdapter(compact=True)
+    adapter = OpenAICompactAdapter()
    return await pipeline.run(
        adapter=adapter,
        http_request=http_request,
--- a/src/api/user_me/routes.py
+++ b/src/api/user_me/routes.py
@@ -1263,6 +1263,7 @@ class ListAvailableModelsAdapter(AuthenticatedApiAdapter):
        all_formats = [
            "openai:chat",
            "openai:cli",
+            "openai:compact",
            "claude:chat",
            "claude:cli",
            "gemini:chat",
--- a/src/core/api_format/conversion/normalizers/openai_cli.py
+++ b/src/core/api_format/conversion/normalizers/openai_cli.py
@@ -199,24 +199,18 @@ class OpenAICliNormalizer(FormatNormalizer):

        return internal

-    # Codex 需要的 include 项
-    _CODEX_REQUIRED_INCLUDE = "reasoning.encrypted_content"
-
    def request_from_internal(
        self,
        internal: InternalRequest,
        *,
        target_variant: str | None = None,
    ) -> dict[str, Any]:
+        _ = target_variant
        openai_cli_extra = internal.extra.get("openai_cli", {})
-        is_compact = bool(openai_cli_extra.get("_aether_compact"))
-        is_codex = str(target_variant or "").lower() == "codex" and not is_compact

        result: dict[str, Any] = {
            "model": internal.model,
-            "input": self._internal_messages_to_input(
-                internal.messages, system_to_developer=is_codex
-            ),
+            "input": self._internal_messages_to_input(internal.messages, system_to_developer=False),
        }

        # 合并 instructions，如果没有则使用 system
@@ -229,20 +223,17 @@ class OpenAICliNormalizer(FormatNormalizer):
        # 统一添加该字段以确保兼容性
        result["instructions"] = instructions_text or ""

-        # max_output_tokens/temperature/top_p: Codex 不支持，标准 API 可选
-        if not is_codex:
-            if internal.max_tokens is not None:
-                # Responses API 使用 max_output_tokens
-                result["max_output_tokens"] = internal.max_tokens
-            if internal.temperature is not None:
-                result["temperature"] = internal.temperature
-            if internal.top_p is not None:
-                result["top_p"] = internal.top_p
+        if internal.max_tokens is not None:
+            # Responses API 使用 max_output_tokens
+            result["max_output_tokens"] = internal.max_tokens
+        if internal.temperature is not None:
+            result["temperature"] = internal.temperature
+        if internal.top_p is not None:
+            result["top_p"] = internal.top_p

        if internal.stop_sequences:
            result["stop"] = list(internal.stop_sequences)
-        # Codex 强制要求 stream=true；其他情况尊重客户端请求
-        result["stream"] = True if is_codex else bool(internal.stream)
+        result["stream"] = bool(internal.stream)

        if internal.tools:
            # Responses API 使用扁平结构: {type, name, description, parameters}
@@ -308,26 +299,10 @@ class OpenAICliNormalizer(FormatNormalizer):
            if key not in handled_keys and key not in result:
                result[key] = value

-        # 统一设置 store=false（Codex 强制要求，标准 API 兼容）
+        # 标准 Responses API 默认设置 store=false
        if "store" not in result:
            result["store"] = False

-        # Codex 特定设置（覆盖/删除不支持的字段）
-        if is_codex:
-            result["parallel_tool_calls"] = True
-            # 和 codex passthrough patch 保持一致：固定 include 列表
-            result["include"] = [self._CODEX_REQUIRED_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)
-
        return result

    # =========================
--- a/src/core/api_format/detection.py
+++ b/src/core/api_format/detection.py
@@ -62,6 +62,10 @@ def _detect_data_format(
            return EndpointSignature(api_family=ApiFamily.CLAUDE, endpoint_kind=EndpointKind.CLI)
        return EndpointSignature(api_family=ApiFamily.CLAUDE, endpoint_kind=EndpointKind.CHAT)

+    # OpenAI compact: /responses/compact
+    if "/responses/compact" in normalized:
+        return EndpointSignature(api_family=ApiFamily.OPENAI, endpoint_kind=EndpointKind.COMPACT)
+
    # OpenAI CLI: /responses
    if "/responses" in normalized:
        return EndpointSignature(api_family=ApiFamily.OPENAI, endpoint_kind=EndpointKind.CLI)
--- a/src/core/api_format/enums.py
+++ b/src/core/api_format/enums.py
@@ -36,6 +36,7 @@ class EndpointKind(str, Enum):

    CHAT = "chat"
    CLI = "cli"
+    COMPACT = "compact"
    VIDEO = "video"
    IMAGE = "image"

--- a/src/core/api_format/metadata.py
+++ b/src/core/api_format/metadata.py
@@ -123,6 +123,19 @@ _ENDPOINT_DEFINITIONS: dict[tuple[ApiFamily, EndpointKind], EndpointDefinition]
        protected_keys=frozenset({"authorization", "content-type"}),
        data_format_id="openai_responses",
    ),
+    (ApiFamily.OPENAI, EndpointKind.COMPACT): EndpointDefinition(
+        api_family=ApiFamily.OPENAI,
+        endpoint_kind=EndpointKind.COMPACT,
+        aliases=("openai_compact", "responses_compact"),
+        default_path="/v1/responses/compact",
+        auth_method=AuthMethod.BEARER,
+        auth_header="Authorization",
+        auth_type="bearer",
+        protected_keys=frozenset({"authorization", "content-type"}),
+        # compact endpoint is non-streaming by design.
+        stream_in_body=False,
+        data_format_id="openai_responses",
+    ),
    (ApiFamily.OPENAI, EndpointKind.VIDEO): EndpointDefinition(
        api_family=ApiFamily.OPENAI,
        endpoint_kind=EndpointKind.VIDEO,
--- a/src/services/health/endpoint.py
+++ b/src/services/health/endpoint.py
@@ -481,6 +481,7 @@ class EndpointHealthService:
        kind_label = {
            "chat": "Chat",
            "cli": "CLI",
+            "compact": "Compact",
            "video": "Video",
            "image": "Image",
        }.get(kind, kind)
--- a/src/services/model/upstream_fetcher.py
+++ b/src/services/model/upstream_fetcher.py
@@ -21,7 +21,7 @@ MAX_CONCURRENT_REQUESTS = 5

 # 模型获取格式优先级：同族内优先使用 chat 端点，若无则回退到 cli 端点
 MODEL_FETCH_FORMAT_PRIORITY: list[tuple[str, ...]] = [
-    ("openai:chat", "openai:cli"),
+    ("openai:chat", "openai:cli", "openai:compact"),
    ("claude:chat", "claude:cli"),
    ("gemini:chat", "gemini:cli"),
 ]
--- a/src/services/provider/adapters/codex/request_patching.py
+++ b/src/services/provider/adapters/codex/request_patching.py
@@ -1,21 +1,8 @@
-"""
-Codex provider request patching helpers (passthrough path).
+"""Codex provider request patching helpers (passthrough path).

-This is the **primary** Codex request transformation used by the normalizer's
-``patch_for_variant("codex")`` fast path.  It applies minimal, non-destructive
-patches directly on the original request dict -- no internal representation
-round-trip, so every field the client sent is preserved as-is unless explicitly
-modified here.
-
-Transformations applied:
- Force ``store=false``.
- Force ``stream=true`` (except compact requests).
- Force ``parallel_tool_calls=true``.
- Ensure ``instructions`` exists (empty string when absent).
- Convert ``role=system`` messages to ``role=developer``.
- Drop request parameters known to be rejected by Codex gateways.
- Force ``include`` to ``["reasoning.encrypted_content"]``.
- Drop compatibility-problematic fields (``context_management`` / ``user``).
+Codex requests are now treated as passthrough:
+- Do not mutate client payload fields.
+- Only strip internal sentinel fields that must never reach upstream.
 """

 from __future__ import annotations
@@ -24,20 +11,6 @@ from typing import Any

 from src.core.provider_types import ProviderType

-_REJECTED_PARAMS: frozenset[str] = frozenset(
-    {
-        "max_output_tokens",
-        "max_completion_tokens",
-        "temperature",
-        "top_p",
-        "service_tier",
-        "previous_response_id",
-        "truncation",
-    }
-)
-
-_REQUIRED_INCLUDE_ITEM = "reasoning.encrypted_content"
-

 def patch_openai_cli_request_for_codex(request_body: dict[str, Any]) -> dict[str, Any]:
    """
@@ -46,49 +19,8 @@ def patch_openai_cli_request_for_codex(request_body: dict[str, Any]) -> dict[str
    This function never mutates the input object.
    """
    out: dict[str, Any] = dict(request_body)
-
-    for k in _REJECTED_PARAMS:
-        out.pop(k, None)
-
-    # Codex gateways often reject/ignore persistence; be explicit.
-    out["store"] = False
-
-    # 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.
-    out["parallel_tool_calls"] = True
-
-    # Ensure instructions exists (some gateways require it even if empty).
-    instructions = out.get("instructions")
-    if not isinstance(instructions, str):
-        out["instructions"] = ""
-
-    # Convert "system" role to "developer" (Codex behavior).
-    input_items = out.get("input")
-    if isinstance(input_items, list):
-        patched_items: list[Any] = []
-        for item in input_items:
-            if isinstance(item, dict):
-                patched = dict(item)
-                if patched.get("role") == "system":
-                    patched["role"] = "developer"
-                patched_items.append(patched)
-            else:
-                patched_items.append(item)
-        out["input"] = patched_items
-
-    # Keep codex behavior deterministic: force the exact include list.
-    out["include"] = [_REQUIRED_INCLUDE_ITEM]
-
-    # Codex upstream currently rejects these fields.
-    out.pop("context_management", None)
-    out.pop("user", None)
-
+    # Internal routing marker; never send upstream.
+    out.pop("_aether_compact", None)
    return out


@@ -108,7 +40,7 @@ def maybe_patch_request_for_codex(
    """
    if (provider_type or "").lower() != ProviderType.CODEX:
        return request_body
-    if (provider_api_format or "").lower() != "openai:cli":
+    if (provider_api_format or "").lower() not in {"openai:cli", "openai:compact"}:
        return request_body
    if not isinstance(request_body, dict):
        return request_body
--- a/src/services/provider/stream_policy.py
+++ b/src/services/provider/stream_policy.py
@@ -55,13 +55,15 @@ def get_upstream_stream_policy(

    Defaults:
    - Codex + openai:cli: FORCE_STREAM (Codex upstream requires stream=true).
+    - Codex + openai:compact: follow endpoint/client policy (no hard force).
    """

    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":
+    is_codex_cli = pt == ProviderType.CODEX and sig == "openai:cli"
+    is_codex_compact = pt == ProviderType.CODEX and sig == "openai:compact"
+    if is_codex_cli:
        try:
            from src.services.provider.adapters.codex.context import get_codex_request_context

@@ -82,8 +84,7 @@ def get_upstream_stream_policy(
        if parsed != UpstreamStreamPolicy.AUTO:
            # Codex upstream requires streaming; do not allow forcing non-stream.
            if (
-                pt == ProviderType.CODEX
-                and sig == "openai:cli"
+                is_codex_cli
                and parsed == UpstreamStreamPolicy.FORCE_NON_STREAM
                and not is_codex_compact
            ):
@@ -93,7 +94,7 @@ def get_upstream_stream_policy(
            return parsed

    # Safe-by-default: Codex Responses OAuth behaves like SSE-only.
-    if pt == ProviderType.CODEX and sig == "openai:cli":
+    if is_codex_cli:
        return (
            UpstreamStreamPolicy.FORCE_NON_STREAM
            if is_codex_compact
@@ -134,13 +135,30 @@ def enforce_stream_mode_for_upstream(
    meta = resolve_endpoint_definition(provider_api_format)
    provider_uses_stream = meta.stream_in_body if meta is not None else True

+    provider_fmt = str(provider_api_format or "").strip().lower()
+    # OpenAI compact endpoint: keep request body stream field absent.
+    if provider_fmt == "openai:compact":
+        request_body.pop("stream", None)
+        return request_body
+
+    # Backward compatibility: Codex compact routed through openai:cli + context marker.
+    if provider_fmt == "openai:cli":
+        try:
+            from src.services.provider.adapters.codex.context import get_codex_request_context
+
+            ctx = get_codex_request_context()
+            if ctx and ctx.is_compact:
+                request_body.pop("stream", None)
+                return request_body
+        except Exception:
+            pass
+
    if provider_uses_stream:
        request_body["stream"] = bool(upstream_is_stream)
    else:
        request_body.pop("stream", None)

    # OpenAI Chat Completions: request usage in streaming mode.
-    provider_fmt = str(provider_api_format or "").strip().lower()
    if upstream_is_stream and provider_fmt == "openai:chat":
        stream_options = request_body.get("stream_options")
        if not isinstance(stream_options, dict):
--- a/src/services/provider_keys/auth_type.py
+++ b/src/services/provider_keys/auth_type.py
@@ -6,7 +6,12 @@ Provider Key 认证类型相关规则。
 def normalize_auth_type(raw: str) -> str:
    """将数据库中的 auth_type 归一化为逻辑类型。

-    Kiro 在数据库中存储为 ``"kiro"`` 或 ``"oauth"``，统一映射为 ``"oauth"``。
+    - ``"kiro"`` -> ``"oauth"``  (Kiro 使用 OAuth 流程)
+    - ``"vertex_ai"`` -> ``"service_account"``  (旧的 Vertex AI auth_type 已重命名)
    """
    t = str(raw or "api_key").strip() or "api_key"
-    return "oauth" if t == "kiro" else t
+    if t == "kiro":  # TODO: 迁移稳定后清理，同步清理各处 in ("...", "kiro") 兼容检查
+        return "oauth"
+    if t == "vertex_ai":  # TODO: 迁移稳定后清理，同步清理各处 in ("...", "vertex_ai") 兼容检查
+        return "service_account"
+    return t
--- a/src/services/provider_keys/duplicate_check.py
+++ b/src/services/provider_keys/duplicate_check.py
@@ -26,14 +26,14 @@ def check_duplicate_key(

    对于不同的认证类型，使用不同的比较方式：
    - api_key: 比较 API Key 的哈希值
-    - vertex_ai: 比较 Service Account 的 client_email
+    - service_account: 比较 Service Account 的 client_email

    Args:
        db: 数据库会话
        provider_id: Provider ID
-        auth_type: 认证类型 (api_key, vertex_ai, oauth)
+        auth_type: 认证类型 (api_key, service_account, oauth)
        new_api_key: 新的 API Key（用于 api_key 类型）
-        new_auth_config: 新的认证配置（用于 vertex_ai 类型）
+        new_auth_config: 新的认证配置（用于 service_account 类型）
        exclude_key_id: 要排除的 Key ID（用于更新场景）
    """
    if auth_type == "api_key" and new_api_key:
@@ -66,7 +66,7 @@ def check_duplicate_key(
                # 解密失败时跳过该 Key
                continue

-    elif auth_type == "vertex_ai" and new_auth_config:
+    elif auth_type in ("service_account", "vertex_ai") and new_auth_config:
        new_client_email = (
            new_auth_config.get("client_email") if isinstance(new_auth_config, dict) else None
        )
@@ -76,7 +76,7 @@ def check_duplicate_key(
        # 仅查询同 auth_type 且有 auth_config 的 Keys
        query = db.query(ProviderAPIKey).filter(
            ProviderAPIKey.provider_id == provider_id,
-            ProviderAPIKey.auth_type == "vertex_ai",
+            ProviderAPIKey.auth_type.in_(["service_account", "vertex_ai"]),
            ProviderAPIKey.auth_config.isnot(None),
        )
        if exclude_key_id:
--- a/src/services/provider_keys/key_command_service.py
+++ b/src/services/provider_keys/key_command_service.py
@@ -16,6 +16,7 @@ from sqlalchemy.orm import Session
 from src.core.crypto import crypto_service
 from src.core.exceptions import InvalidRequestException, NotFoundException
 from src.core.logger import logger
+from src.core.provider_types import ProviderType
 from src.models.database import Provider, ProviderAPIKey
 from src.models.endpoint_models import (
    EndpointAPIKeyCreate,
@@ -32,6 +33,37 @@ from src.services.provider_keys.key_side_effects import (
 from src.services.provider_keys.response_builder import build_key_response


+def _validate_vertex_api_formats(
+    provider_type: str | None,
+    auth_type: str,
+    api_formats: list[str] | None,
+) -> None:
+    """校验 Vertex Provider 的 key.api_formats 与 auth_type 是否匹配。"""
+    if str(provider_type or "").strip().lower() != ProviderType.VERTEX_AI.value:
+        return
+
+    formats = [
+        str(fmt or "").strip().lower() for fmt in (api_formats or []) if str(fmt or "").strip()
+    ]
+    if not formats:
+        return
+
+    if auth_type == "api_key":
+        allowed = {"gemini:chat"}
+    elif auth_type in {"service_account", "vertex_ai"}:
+        allowed = {"gemini:chat", "claude:chat"}
+    else:
+        return
+
+    invalid = sorted({fmt for fmt in formats if fmt not in allowed})
+    if invalid:
+        allowed_text = ", ".join(sorted(allowed))
+        invalid_text = ", ".join(invalid)
+        raise InvalidRequestException(
+            f"Vertex {auth_type} 不支持以下 API 格式: {invalid_text}；允许: {allowed_text}"
+        )
+
+
@dataclass
 class _UpdateKeyPreparation:
    """更新 Key 前置准备结果。"""
@@ -154,14 +186,14 @@ def _prepare_update_key_payload(
            raise InvalidRequestException("API Key 认证模式下 api_key 不能为空")
        # 切换回 API Key：清理非本模式配置
        update_data["auth_config"] = None
-    elif target_auth_type == "vertex_ai":
+    elif target_auth_type == "service_account":
        if is_auth_type_switch and not update_data.get("auth_config"):
            raise InvalidRequestException(
-                "从 API Key 切换到 Vertex AI 认证模式时，必须提供 Service Account JSON"
+                "切换到 Service Account 认证模式时，必须提供 Service Account JSON"
            )
-        # Vertex AI 不允许手工写入 api_key，仅保留占位符
+        # Service Account 不允许手工写入 api_key，仅保留占位符
        if api_key_in_payload and api_key_value not in {None, "__placeholder__"}:
-            raise InvalidRequestException("Vertex AI 认证模式下不允许直接填写 api_key")
+            raise InvalidRequestException("Service Account 认证模式下不允许直接填写 api_key")
        if is_auth_type_switch or api_key_in_payload:
            update_data["api_key"] = "__placeholder__"
    elif target_auth_type == "oauth":
@@ -186,6 +218,15 @@ def _prepare_update_key_payload(
        exclude_key_id=key_id,
    )

+    # Vertex Provider: auth_type 与 api_formats 的组合必须合法
+    provider = getattr(key, "provider", None)
+    effective_api_formats = update_data.get("api_formats", key.api_formats)
+    _validate_vertex_api_formats(
+        getattr(provider, "provider_type", None),
+        target_auth_type,
+        effective_api_formats,
+    )
+
    if "api_key" in update_data:
        api_key_raw = update_data["api_key"]
        if api_key_raw is None:
@@ -261,7 +302,7 @@ def _prepare_create_key_payload(
    if auth_type == "api_key":
        if not key_data.api_key:
            raise InvalidRequestException("API Key 认证模式下 api_key 为必填字段")
-    elif auth_type == "vertex_ai":
+    elif auth_type == "service_account":
        if not key_data.auth_config:
            raise InvalidRequestException("Service Account 认证模式下 auth_config 为必填字段")
    elif auth_type == "oauth":
@@ -384,12 +425,19 @@ async def create_provider_key_response(
    if not key_data.api_formats:
        raise InvalidRequestException("api_formats 为必填字段")

-    _, new_key = _prepare_create_key_payload(
+    auth_type, new_key = _prepare_create_key_payload(
        db=db,
        provider_id=provider_id,
        key_data=key_data,
    )

+    # Vertex Provider: auth_type 与 api_formats 的组合必须合法
+    _validate_vertex_api_formats(
+        getattr(provider, "provider_type", None),
+        auth_type,
+        key_data.api_formats,
+    )
+
    db.add(new_key)
    db.commit()
    db.refresh(new_key)
--- a/src/services/provider_keys/key_query_service.py
+++ b/src/services/provider_keys/key_query_service.py
@@ -59,7 +59,7 @@ def get_keys_grouped_by_format(db: Session) -> dict:
            continue  # 跳过没有 API 格式的 Key

        auth_type = normalize_auth_type(getattr(key, "auth_type", "api_key"))
-        if auth_type == "vertex_ai":
+        if auth_type in ("service_account", "vertex_ai"):
            masked_key = "[Service Account]"
        elif auth_type == "oauth":
            masked_key = "[OAuth Token]"
@@ -166,15 +166,15 @@ def reveal_endpoint_key_payload(

    auth_type = normalize_auth_type(getattr(key, "auth_type", "api_key"))

-    # Vertex AI 类型返回 auth_config（需要解密）
-    if auth_type == "vertex_ai":
+    # Service Account 类型返回 auth_config（需要解密）
+    if auth_type in ("service_account", "vertex_ai"):
        encrypted_auth_config = getattr(key, "auth_config", None)
        if encrypted_auth_config:
            try:
                decrypted_config = crypto_service.decrypt(encrypted_auth_config)
                auth_config = json.loads(decrypted_config)
                logger.info(f"[REVEAL] 查看 Auth Config: ID={key_id}, Name={key.name}")
-                return {"auth_type": "vertex_ai", "auth_config": auth_config}
+                return {"auth_type": auth_type, "auth_config": auth_config}
            except Exception as e:
                logger.error(f"解密 Auth Config 失败: ID={key_id}, Error={e}")
                raise InvalidRequestException(
@@ -184,12 +184,11 @@ def reveal_endpoint_key_payload(
        # 兼容：auth_config 为空时尝试从 api_key 解密（仅对迁移前的旧数据有效）
        try:
            decrypted_key = crypto_service.decrypt(key.api_key)
-            # 检查是否是新格式的占位符（表示 auth_config 丢失）
            if decrypted_key == "__placeholder__":
-                logger.error(f"Vertex AI Key 缺少 auth_config: ID={key_id}")
+                logger.error(f"Service Account Key 缺少 auth_config: ID={key_id}")
                raise InvalidRequestException("认证配置丢失，请重新添加该密钥。")
-            logger.info(f"[REVEAL] 查看完整 Key (legacy vertex_ai): ID={key_id}, Name={key.name}")
-            return {"auth_type": "vertex_ai", "auth_config": decrypted_key}
+            logger.info(f"[REVEAL] 查看完整 Key (legacy SA): ID={key_id}, Name={key.name}")
+            return {"auth_type": auth_type, "auth_config": decrypted_key}
        except InvalidRequestException:
            raise
        except Exception as e:
--- a/src/services/provider_keys/response_builder.py
+++ b/src/services/provider_keys/response_builder.py
@@ -19,8 +19,8 @@ def build_key_response(
    """构建 Key 响应对象。"""
    auth_type = normalize_auth_type(getattr(key, "auth_type", "api_key"))

-    if auth_type == "vertex_ai":
-        # Vertex AI 使用 Service Account，不显示占位符
+    if auth_type in ("service_account", "vertex_ai"):
+        # Service Account 不显示占位符
        masked_key = "[Service Account]"
    elif auth_type == "oauth":
        masked_key = "[OAuth Token]"
--- a/src/services/scheduling/candidate_builder.py
+++ b/src/services/scheduling/candidate_builder.py
@@ -415,9 +415,12 @@ class CandidateBuilder:
                if isinstance(raw, int) and raw > 0:
                    output_limit = raw

-        # chat/cli 互相可回退（用于同协议族下的端点变体），video/image 等不跨类回退
+        # chat/cli 互相可回退（用于同协议族下的端点变体），compact 可回退到 cli。
+        # video/image 等不跨类回退。
        if client_kind in {EndpointKind.CHAT, EndpointKind.CLI}:
            allowed_kinds = {EndpointKind.CHAT, EndpointKind.CLI}
+        elif client_kind == EndpointKind.COMPACT:
+            allowed_kinds = {EndpointKind.COMPACT, EndpointKind.CLI}
        else:
            allowed_kinds = {client_kind}

--- a/tests/services/test_codex_request.py
+++ b/tests/services/test_codex_request.py
@@ -6,94 +6,39 @@ from src.services.provider.adapters.codex.request_patching import (
 )


-def test_patch_openai_cli_request_for_codex_sets_store_and_instructions() -> None:
-    req = {"model": "gpt-test", "input": []}
-    out = patch_openai_cli_request_for_codex(req)
-
-    assert out is not req
-    assert out["store"] is False
-    assert out["stream"] is True
-    assert out["instructions"] == ""
-
-
-def test_patch_openai_cli_request_for_codex_strips_rejected_params() -> None:
+def test_patch_openai_cli_request_for_codex_is_passthrough_except_internal_sentinel() -> None:
    req = {
        "model": "gpt-test",
-        "input": [],
-        "max_output_tokens": 123,
-        "max_completion_tokens": 456,
-        "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",
-        "temperature",
-        "top_p",
-        "service_tier",
-        "truncation",
-        "context_management",
-        "user",
-    ):
-        assert key not in out
-
-
-def test_patch_openai_cli_request_for_codex_converts_system_role_to_developer() -> None:
-    req = {
-        "model": "gpt-test",
-        "instructions": "ignored",
        "input": [
            {
                "type": "message",
                "role": "system",
-                "content": [{"type": "input_text", "text": "You are a pirate."}],
-            },
-            {
-                "type": "message",
-                "role": "user",
                "content": [{"type": "input_text", "text": "Hello"}],
-            },
+            }
        ],
-    }
-    out = patch_openai_cli_request_for_codex(req)
-
-    assert isinstance(out.get("input"), list)
-    assert out["input"][0]["role"] == "developer"
-    assert out["input"][1]["role"] == "user"
-
-
-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 out["include"] == ["reasoning.encrypted_content"]
-
-
-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": [],
+        "store": True,
+        "stream": False,
+        "instructions": "keep",
+        "include": ["foo"],
+        "parallel_tool_calls": False,
+        "temperature": 0.7,
+        "context_management": {"compaction": {"type": "summary"}},
+        "user": "u_123",
        "_aether_compact": True,
-        "stream": True,
    }
    out = patch_openai_cli_request_for_codex(req)
-    assert "stream" not in out
+
+    assert out is not req
+    assert "_aether_compact" not in out
+    assert out["store"] is True
+    assert out["stream"] is False
+    assert out["instructions"] == "keep"
+    assert out["include"] == ["foo"]
+    assert out["parallel_tool_calls"] is False
+    assert out["temperature"] == 0.7
+    assert out["context_management"] == {"compaction": {"type": "summary"}}
+    assert out["user"] == "u_123"
+    assert out["input"][0]["role"] == "system"


 def test_maybe_patch_request_for_codex_is_noop_for_non_codex() -> None:
@@ -117,7 +62,7 @@ def test_maybe_patch_request_for_codex_is_noop_for_non_openai_cli() -> None:


 def test_maybe_patch_request_for_codex_patches_for_codex_openai_cli() -> None:
-    req = {"model": "gpt-test", "input": []}
+    req = {"model": "gpt-test", "input": [], "_aether_compact": True, "store": True}
    out = maybe_patch_request_for_codex(
        provider_type="codex",
        provider_api_format="openai:cli",
@@ -125,8 +70,41 @@ def test_maybe_patch_request_for_codex_patches_for_codex_openai_cli() -> None:
    )

    assert out is not req
+    assert out["store"] is True
+    assert "_aether_compact" not in out
+
+
+def test_maybe_patch_request_for_codex_patches_for_codex_openai_compact() -> None:
+    req = {"model": "gpt-test", "input": [], "_aether_compact": True, "store": True}
+    out = maybe_patch_request_for_codex(
+        provider_type="codex",
+        provider_api_format="openai:compact",
+        request_body=req,
+    )
+
+    assert out is not req
+    assert out["store"] is True
+    assert "_aether_compact" not in out
+
+
+def test_openai_cli_normalizer_request_from_internal_codex_variant_preserves_store() -> None:
+    from src.core.api_format.conversion.normalizers.openai_cli import OpenAICliNormalizer
+
+    normalizer = OpenAICliNormalizer()
+    internal = normalizer.request_to_internal({"model": "gpt-test", "input": [], "store": True})
+    out = normalizer.request_from_internal(internal, target_variant="codex")
+
+    assert out["store"] is True
+
+
+def test_openai_cli_normalizer_request_from_internal_codex_variant_defaults_store_false() -> None:
+    from src.core.api_format.conversion.normalizers.openai_cli import OpenAICliNormalizer
+
+    normalizer = OpenAICliNormalizer()
+    internal = normalizer.request_to_internal({"model": "gpt-test", "input": []})
+    out = normalizer.request_from_internal(internal, target_variant="codex")
+
    assert out["store"] is False
-    assert "instructions" in out


 def test_codex_envelope_extra_headers_includes_sse_accept_and_session() -> None:
--- a/tests/services/test_provider_keys_key_command_service.py
+++ b/tests/services/test_provider_keys_key_command_service.py
@@ -65,6 +65,8 @@ def _build_key(**overrides: Any) -> SimpleNamespace:
        "model_exclude_patterns": None,
        "provider_id": "provider-1",
        "auth_type": "api_key",
+        "api_formats": [],
+        "provider": None,
    }
    base.update(overrides)
    return SimpleNamespace(**base)
--- a/tests/services/test_provider_transport_codex.py
+++ b/tests/services/test_provider_transport_codex.py
@@ -64,3 +64,17 @@ def test_codex_openai_cli_uses_compact_suffix_when_context_marked_compact() -> N
    )
    assert url == "https://chatgpt.com/backend-api/codex/responses/compact"
    set_codex_request_context(None)
+
+
+def test_codex_openai_compact_uses_compact_path_without_v1_prefix() -> None:
+    endpoint = _DummyEndpoint(
+        base_url="https://chatgpt.com/backend-api/codex",
+        api_format="openai:compact",
+        provider=SimpleNamespace(provider_type="codex"),
+    )
+    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"
--- a/tests/services/test_upstream_stream_policy.py
+++ b/tests/services/test_upstream_stream_policy.py
@@ -59,6 +59,15 @@ def test_get_upstream_stream_policy_codex_compact_forces_non_stream() -> None:
        set_codex_request_context(None)


+def test_get_upstream_stream_policy_codex_openai_compact_defaults_to_auto() -> None:
+    ep = _DummyEndpoint(
+        api_format="openai:compact",
+        config=None,
+        provider=SimpleNamespace(provider_type="codex"),
+    )
+    assert get_upstream_stream_policy(ep) == UpstreamStreamPolicy.AUTO
+
+
 def test_enforce_stream_mode_for_upstream_openai_chat_sets_stream_options_usage() -> None:
    body = {"stream": False}
    out = enforce_stream_mode_for_upstream(
@@ -79,3 +88,30 @@ def test_enforce_stream_mode_for_upstream_gemini_drops_stream_field() -> None:
    )
    assert "stream" not in out
    assert out["foo"] == "bar"
+
+
+def test_enforce_stream_mode_for_upstream_openai_compact_drops_stream_field() -> None:
+    body = {"stream": True, "foo": "bar"}
+    out = enforce_stream_mode_for_upstream(
+        body,
+        provider_api_format="openai:compact",
+        upstream_is_stream=True,
+    )
+    assert "stream" not in out
+    assert out["foo"] == "bar"
+
+
+def test_enforce_stream_mode_for_upstream_codex_compact_keeps_stream_absent() -> None:
+    body = {"stream": True, "foo": "bar"}
+    try:
+        set_codex_request_context(CodexRequestContext(is_compact=True))
+        out = enforce_stream_mode_for_upstream(
+            body,
+            provider_api_format="openai:cli",
+            upstream_is_stream=False,
+        )
+    finally:
+        set_codex_request_context(None)
+
+    assert "stream" not in out
+    assert out["foo"] == "bar"