From 89dffdd2e1915bb84f578ac3bdd97ad91ae349f3 Mon Sep 17 00:00:00 2001 From: stabey <36232531+stabey@users.noreply.github.com> Date: Wed, 27 May 2026 22:36:52 +0800 Subject: [PATCH 01/20] fix(apicompat): emit OpenAI-semantic input_tokens when converting Anthropic to Responses Anthropic Messages reports input_tokens excluding cache_read/cache_creation, but OpenAI Responses input_tokens is the total including cached tokens. The reverse converter passed Anthropic's input_tokens straight through, so client-facing prompt_tokens/input_tokens were short by the cached count and cache_creation was dropped entirely. Fix the non-stream path and the streaming state machine to add cache_read + cache_creation back into input_tokens, and track CacheCreationInputTokens on the streaming state. Six downstream paths benefit (Anthropic->Responses, Anthropic->ChatCompletions, Gemini->ChatCompletions, each sync + stream). Co-Authored-By: Claude Opus 4.7 --- .../pkg/apicompat/anthropic_responses_test.go | 136 ++++++++++++++++++ .../anthropic_to_responses_response.go | 40 ++++-- 2 files changed, 168 insertions(+), 8 deletions(-) diff --git a/backend/internal/pkg/apicompat/anthropic_responses_test.go b/backend/internal/pkg/apicompat/anthropic_responses_test.go index bb566081..8997835c 100644 --- a/backend/internal/pkg/apicompat/anthropic_responses_test.go +++ b/backend/internal/pkg/apicompat/anthropic_responses_test.go @@ -1597,3 +1597,139 @@ func TestAnthropicToResponses_TemperatureStrippedForAllGpt5Variants(t *testing.T }) } } + +// --------------------------------------------------------------------------- +// AnthropicToResponsesResponse: Anthropic input_tokens excludes cached tokens +// while OpenAI Responses input_tokens is the total including cached tokens. +// --------------------------------------------------------------------------- + +func TestAnthropicToResponsesResponse_CacheTokensUseOpenAIInputSemantics(t *testing.T) { + resp := &AnthropicResponse{ + ID: "msg_cache", + Model: "claude-sonnet-4-5-20250929", + Content: []AnthropicContentBlock{ + {Type: "text", Text: "ok"}, + }, + StopReason: "end_turn", + Usage: AnthropicUsage{ + InputTokens: 3318, + OutputTokens: 123, + CacheReadInputTokens: 50688, + CacheCreationInputTokens: 200, + }, + } + + out := AnthropicToResponsesResponse(resp) + require.NotNil(t, out.Usage) + // 3318 (uncached) + 50688 (read) + 200 (creation) = 54206 + assert.Equal(t, 54206, out.Usage.InputTokens) + assert.Equal(t, 123, out.Usage.OutputTokens) + assert.Equal(t, 54329, out.Usage.TotalTokens) + require.NotNil(t, out.Usage.InputTokensDetails) + assert.Equal(t, 50688, out.Usage.InputTokensDetails.CachedTokens) +} + +func TestAnthropicToResponsesResponse_NoCacheTokens(t *testing.T) { + resp := &AnthropicResponse{ + ID: "msg_nocache", + Model: "claude-sonnet-4-5-20250929", + Content: []AnthropicContentBlock{ + {Type: "text", Text: "ok"}, + }, + StopReason: "end_turn", + Usage: AnthropicUsage{ + InputTokens: 100, + OutputTokens: 50, + }, + } + + out := AnthropicToResponsesResponse(resp) + require.NotNil(t, out.Usage) + assert.Equal(t, 100, out.Usage.InputTokens) + assert.Equal(t, 50, out.Usage.OutputTokens) + assert.Equal(t, 150, out.Usage.TotalTokens) + assert.Nil(t, out.Usage.InputTokensDetails) +} + +func TestAnthropicEventToResponses_CacheTokensRoundTripFromMessageStart(t *testing.T) { + state := NewAnthropicEventToResponsesState() + + // message_start carries cache fields on the initial Usage object. + AnthropicEventToResponsesEvents(&AnthropicStreamEvent{ + Type: "message_start", + Message: &AnthropicResponse{ + ID: "msg_stream_cache", + Model: "claude-sonnet-4-5-20250929", + Usage: AnthropicUsage{ + InputTokens: 12, + CacheReadInputTokens: 9, + CacheCreationInputTokens: 3, + }, + }, + }, state) + + AnthropicEventToResponsesEvents(&AnthropicStreamEvent{ + Type: "message_delta", + Usage: &AnthropicUsage{ + OutputTokens: 7, + }, + }, state) + + events := AnthropicEventToResponsesEvents(&AnthropicStreamEvent{Type: "message_stop"}, state) + + // The terminal response.completed event must include OpenAI-semantic usage. + var completed *ResponsesStreamEvent + for i := range events { + if events[i].Type == "response.completed" { + completed = &events[i] + } + } + require.NotNil(t, completed, "response.completed event must be emitted") + require.NotNil(t, completed.Response) + require.NotNil(t, completed.Response.Usage) + // 12 (uncached) + 9 (read) + 3 (creation) = 24 + assert.Equal(t, 24, completed.Response.Usage.InputTokens) + assert.Equal(t, 7, completed.Response.Usage.OutputTokens) + assert.Equal(t, 31, completed.Response.Usage.TotalTokens) + require.NotNil(t, completed.Response.Usage.InputTokensDetails) + assert.Equal(t, 9, completed.Response.Usage.InputTokensDetails.CachedTokens) +} + +func TestAnthropicEventToResponses_CacheTokensFromMessageDelta(t *testing.T) { + state := NewAnthropicEventToResponsesState() + + AnthropicEventToResponsesEvents(&AnthropicStreamEvent{ + Type: "message_start", + Message: &AnthropicResponse{ + ID: "msg_delta_cache", + Model: "claude-sonnet-4-5-20250929", + Usage: AnthropicUsage{InputTokens: 20}, + }, + }, state) + + // Some upstreams only emit cache fields on the final message_delta. + AnthropicEventToResponsesEvents(&AnthropicStreamEvent{ + Type: "message_delta", + Usage: &AnthropicUsage{ + OutputTokens: 8, + CacheReadInputTokens: 11, + CacheCreationInputTokens: 4, + }, + }, state) + + events := AnthropicEventToResponsesEvents(&AnthropicStreamEvent{Type: "message_stop"}, state) + + var completed *ResponsesStreamEvent + for i := range events { + if events[i].Type == "response.completed" { + completed = &events[i] + } + } + require.NotNil(t, completed) + require.NotNil(t, completed.Response.Usage) + // 20 (uncached) + 11 (read) + 4 (creation) = 35 + assert.Equal(t, 35, completed.Response.Usage.InputTokens) + assert.Equal(t, 8, completed.Response.Usage.OutputTokens) + require.NotNil(t, completed.Response.Usage.InputTokensDetails) + assert.Equal(t, 11, completed.Response.Usage.InputTokensDetails.CachedTokens) +} diff --git a/backend/internal/pkg/apicompat/anthropic_to_responses_response.go b/backend/internal/pkg/apicompat/anthropic_to_responses_response.go index 9290e399..de8ab78d 100644 --- a/backend/internal/pkg/apicompat/anthropic_to_responses_response.go +++ b/backend/internal/pkg/apicompat/anthropic_to_responses_response.go @@ -95,10 +95,16 @@ func AnthropicToResponsesResponse(resp *AnthropicResponse) *ResponsesResponse { } // Usage + // Anthropic's input_tokens excludes cache_read/cache_creation, while OpenAI + // Responses' input_tokens is the total including cached tokens. Add them back + // when converting so downstream consumers see OpenAI semantics. + totalInputTokens := resp.Usage.InputTokens + + resp.Usage.CacheReadInputTokens + + resp.Usage.CacheCreationInputTokens out.Usage = &ResponsesUsage{ - InputTokens: resp.Usage.InputTokens, + InputTokens: totalInputTokens, OutputTokens: resp.Usage.OutputTokens, - TotalTokens: resp.Usage.InputTokens + resp.Usage.OutputTokens, + TotalTokens: totalInputTokens + resp.Usage.OutputTokens, } if resp.Usage.CacheReadInputTokens > 0 { out.Usage.InputTokensDetails = &ResponsesInputTokensDetails{ @@ -150,10 +156,13 @@ type AnthropicEventToResponsesState struct { CurrentCallID string CurrentName string - // Usage from message_delta - InputTokens int - OutputTokens int - CacheReadInputTokens int + // Usage from message_start / message_delta. InputTokens here follows + // Anthropic semantics (excludes cached tokens); they are added back when + // emitting the OpenAI Responses usage. + InputTokens int + OutputTokens int + CacheReadInputTokens int + CacheCreationInputTokens int } // NewAnthropicEventToResponsesState returns an initialised stream state. @@ -225,6 +234,12 @@ func anthToResHandleMessageStart(evt *AnthropicStreamEvent, state *AnthropicEven if evt.Message.Usage.InputTokens > 0 { state.InputTokens = evt.Message.Usage.InputTokens } + if evt.Message.Usage.CacheReadInputTokens > 0 { + state.CacheReadInputTokens = evt.Message.Usage.CacheReadInputTokens + } + if evt.Message.Usage.CacheCreationInputTokens > 0 { + state.CacheCreationInputTokens = evt.Message.Usage.CacheCreationInputTokens + } } if state.CreatedSent { @@ -392,9 +407,15 @@ func anthToResHandleMessageDelta(evt *AnthropicStreamEvent, state *AnthropicEven // Update usage if evt.Usage != nil { state.OutputTokens = evt.Usage.OutputTokens + if evt.Usage.InputTokens > 0 { + state.InputTokens = evt.Usage.InputTokens + } if evt.Usage.CacheReadInputTokens > 0 { state.CacheReadInputTokens = evt.Usage.CacheReadInputTokens } + if evt.Usage.CacheCreationInputTokens > 0 { + state.CacheCreationInputTokens = evt.Usage.CacheCreationInputTokens + } } return nil @@ -472,10 +493,13 @@ func makeResponsesCompletedEvent( seq := state.SequenceNumber state.SequenceNumber++ + // Anthropic's input_tokens excludes cache_read/cache_creation; add them + // back to match OpenAI Responses semantics where input_tokens is the total. + totalInputTokens := state.InputTokens + state.CacheReadInputTokens + state.CacheCreationInputTokens usage := &ResponsesUsage{ - InputTokens: state.InputTokens, + InputTokens: totalInputTokens, OutputTokens: state.OutputTokens, - TotalTokens: state.InputTokens + state.OutputTokens, + TotalTokens: totalInputTokens + state.OutputTokens, } if state.CacheReadInputTokens > 0 { usage.InputTokensDetails = &ResponsesInputTokensDetails{ From 56908d3c4cfc2ca6d705321e40dd15e0e32ba68e Mon Sep 17 00:00:00 2001 From: DaydreamCoding <22166516+DaydreamCoding@users.noreply.github.com> Date: Wed, 27 May 2026 19:42:35 +0800 Subject: [PATCH 02/20] =?UTF-8?q?feat(openai):=20codex=5Fcli=5Fonly=20?= =?UTF-8?q?=E6=96=B0=E5=A2=9E=E6=94=BE=E8=A1=8C=20Claude=20Code=20Codex=20?= =?UTF-8?q?=E6=8F=92=E4=BB=B6=E7=9A=84=E6=9C=BA=E5=88=B6?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 适用场景：在 Claude Code 中使用 https://github.com/openai/codex-plugin-cc 插件时，插件经官方 codex app-server 以 clientInfo.name="Claude Code" 完成 initialize 握手，请求头被设为 originator=Claude Code、User-Agent 含 "Claude Code/"，不在官方客户端白名单内，原本会被 codex_cli_only 拦截 403。 在官方客户端白名单未命中时评估两层独立放行（OR 语义）： - 按账号：account.Extra.codex_cli_only_allowed_clients 引用命名预设 （目前仅 claude_code），detector reason=allowed_client_matched - 全局开关：/admin/settings 网关服务 OpenAI 区块新增 openai_allow_claude_code_codex_plugin（默认 false），开启后对所有 codex_cli_only 账号统一放行，detector reason=global_allowed_client_matched 签名仍要求 originator=Claude Code 精确等值 + UA 含 "Claude Code/"。 上游转发保持透传不变。 Co-Authored-By: Claude Opus 4.7 (1M context) --- .../internal/handler/admin/setting_handler.go | 12 ++ backend/internal/handler/dto/settings.go | 1 + backend/internal/pkg/openai/allowed_client.go | 78 +++++++++ .../pkg/openai/allowed_client_test.go | 95 +++++++++++ backend/internal/server/api_contract_test.go | 2 + backend/internal/service/account.go | 32 ++++ ...unt_codex_cli_only_allowed_clients_test.go | 68 ++++++++ backend/internal/service/domain_constants.go | 3 + .../openai_client_restriction_detector.go | 28 +++- ...openai_client_restriction_detector_test.go | 151 +++++++++++++++++- .../service/openai_gateway_service.go | 13 +- ...nai_gateway_service_codex_cli_only_test.go | 4 +- backend/internal/service/setting_service.go | 90 +++++++++-- ...g_service_openai_allow_claude_code_test.go | 55 +++++++ backend/internal/service/settings_view.go | 1 + frontend/src/api/admin/settings.ts | 2 + .../account/BulkEditAccountModal.vue | 54 +++++++ .../components/account/CreateAccountModal.vue | 39 +++++ .../components/account/EditAccountModal.vue | 37 +++++ .../__tests__/BulkEditAccountModal.spec.ts | 19 +++ frontend/src/i18n/locales/en.ts | 6 + frontend/src/i18n/locales/zh.ts | 5 + frontend/src/views/admin/SettingsView.vue | 15 ++ 23 files changed, 787 insertions(+), 23 deletions(-) create mode 100644 backend/internal/pkg/openai/allowed_client.go create mode 100644 backend/internal/pkg/openai/allowed_client_test.go create mode 100644 backend/internal/service/account_codex_cli_only_allowed_clients_test.go create mode 100644 backend/internal/service/setting_service_openai_allow_claude_code_test.go diff --git a/backend/internal/handler/admin/setting_handler.go b/backend/internal/handler/admin/setting_handler.go index 3c7fe581..c229d340 100644 --- a/backend/internal/handler/admin/setting_handler.go +++ b/backend/internal/handler/admin/setting_handler.go @@ -256,6 +256,7 @@ func (h *SettingHandler) GetSettings(c *gin.Context) { RewriteMessageCacheControl: settings.RewriteMessageCacheControl, AntigravityUserAgentVersion: settings.AntigravityUserAgentVersion, OpenAICodexUserAgent: settings.OpenAICodexUserAgent, + OpenAIAllowClaudeCodeCodexPlugin: settings.OpenAIAllowClaudeCodeCodexPlugin, WebSearchEmulationEnabled: settings.WebSearchEmulationEnabled, PaymentVisibleMethodAlipaySource: settings.PaymentVisibleMethodAlipaySource, PaymentVisibleMethodWxpaySource: settings.PaymentVisibleMethodWxpaySource, @@ -584,6 +585,7 @@ type UpdateSettingsRequest struct { RewriteMessageCacheControl *bool `json:"rewrite_message_cache_control"` AntigravityUserAgentVersion *string `json:"antigravity_user_agent_version"` OpenAICodexUserAgent *string `json:"openai_codex_user_agent"` + OpenAIAllowClaudeCodeCodexPlugin *bool `json:"openai_allow_claude_code_codex_plugin"` // Payment visible method routing PaymentVisibleMethodAlipaySource *string `json:"payment_visible_method_alipay_source"` @@ -1655,6 +1657,12 @@ func (h *SettingHandler) UpdateSettings(c *gin.Context) { } return previousSettings.OpenAICodexUserAgent }(), + OpenAIAllowClaudeCodeCodexPlugin: func() bool { + if req.OpenAIAllowClaudeCodeCodexPlugin != nil { + return *req.OpenAIAllowClaudeCodeCodexPlugin + } + return previousSettings.OpenAIAllowClaudeCodeCodexPlugin + }(), PaymentVisibleMethodAlipaySource: func() string { if req.PaymentVisibleMethodAlipaySource != nil { return strings.TrimSpace(*req.PaymentVisibleMethodAlipaySource) @@ -2031,6 +2039,7 @@ func (h *SettingHandler) UpdateSettings(c *gin.Context) { RewriteMessageCacheControl: updatedSettings.RewriteMessageCacheControl, AntigravityUserAgentVersion: updatedSettings.AntigravityUserAgentVersion, OpenAICodexUserAgent: updatedSettings.OpenAICodexUserAgent, + OpenAIAllowClaudeCodeCodexPlugin: updatedSettings.OpenAIAllowClaudeCodeCodexPlugin, PaymentVisibleMethodAlipaySource: updatedSettings.PaymentVisibleMethodAlipaySource, PaymentVisibleMethodWxpaySource: updatedSettings.PaymentVisibleMethodWxpaySource, PaymentVisibleMethodAlipayEnabled: updatedSettings.PaymentVisibleMethodAlipayEnabled, @@ -2500,6 +2509,9 @@ func diffSettings(before *service.SystemSettings, after *service.SystemSettings, if before.OpenAICodexUserAgent != after.OpenAICodexUserAgent { changed = append(changed, "openai_codex_user_agent") } + if before.OpenAIAllowClaudeCodeCodexPlugin != after.OpenAIAllowClaudeCodeCodexPlugin { + changed = append(changed, "openai_allow_claude_code_codex_plugin") + } if before.PaymentVisibleMethodAlipaySource != after.PaymentVisibleMethodAlipaySource { changed = append(changed, "payment_visible_method_alipay_source") } diff --git a/backend/internal/handler/dto/settings.go b/backend/internal/handler/dto/settings.go index eecf98ac..17772a2e 100644 --- a/backend/internal/handler/dto/settings.go +++ b/backend/internal/handler/dto/settings.go @@ -185,6 +185,7 @@ type SystemSettings struct { RewriteMessageCacheControl bool `json:"rewrite_message_cache_control"` AntigravityUserAgentVersion string `json:"antigravity_user_agent_version"` OpenAICodexUserAgent string `json:"openai_codex_user_agent"` + OpenAIAllowClaudeCodeCodexPlugin bool `json:"openai_allow_claude_code_codex_plugin"` // Web Search Emulation WebSearchEmulationEnabled bool `json:"web_search_emulation_enabled"` diff --git a/backend/internal/pkg/openai/allowed_client.go b/backend/internal/pkg/openai/allowed_client.go new file mode 100644 index 00000000..d4ca14ee --- /dev/null +++ b/backend/internal/pkg/openai/allowed_client.go @@ -0,0 +1,78 @@ +package openai + +import "strings" + +// 命名预设 ID。账号侧 codex_cli_only_allowed_clients 只能引用这些预设键， +// 具体匹配规则固化在下方 registry 中，配置只能「选择启用哪些预设」、不能自定义规则， +// 以防该白名单退化为可任意放宽的后门。 +const ( + // AllowedClientClaudeCode 对应 Claude Code CLI 的 codex 插件。 + AllowedClientClaudeCode = "claude_code" +) + +// AllowedClientEntry 描述一个被额外放行的非官方 Codex 客户端签名。 +// Originator 必须精确等值匹配（归一化后）。 +// UAContains 为必填字段：列表为空，或列表中存在任何空白 marker，均视为非法配置， +// 整体安全失败（return false）；每一项都必须出现在 User-Agent 中。 +// 这确保双因子匹配不会因缺失 UA 声明而退化为仅凭可伪造的 originator 单因子放行。 +type AllowedClientEntry struct { + Originator string + UAContains []string +} + +// allowedClientRegistry 固化各命名预设的签名规则。 +// +// Claude Code codex 插件签名来源：插件以 clientInfo.name="Claude Code" 完成 app-server +// initialize 握手，codex 据此把 originator 设为 "Claude Code"，User-Agent 前缀同样为 +// "Claude Code/"（两者同源）。若上游 Claude Code 插件更改 clientInfo.name，此处需同步更新。 +var allowedClientRegistry = map[string]AllowedClientEntry{ + AllowedClientClaudeCode: { + Originator: "Claude Code", + UAContains: []string{"Claude Code/"}, + }, +} + +// IsAllowedClientMatch 判断请求头是否命中给定的额外客户端签名。 +// originator 必须精确等值（归一化后）；UAContains 中每一项都必须出现在 UA 中。 +// UAContains 为必填：列表为空或含任何空白 marker 均视为非法配置，整体安全失败。 +func IsAllowedClientMatch(userAgent, originator string, entry AllowedClientEntry) bool { + wantOriginator := normalizeCodexClientHeader(entry.Originator) + if wantOriginator == "" { + return false + } + if normalizeCodexClientHeader(originator) != wantOriginator { + return false + } + // 预设必须声明 UA 特征：否则将退化为仅凭可伪造的 originator 单因子匹配。 + if len(entry.UAContains) == 0 { + return false + } + ua := normalizeCodexClientHeader(userAgent) + for _, marker := range entry.UAContains { + normalizedMarker := normalizeCodexClientHeader(marker) + if normalizedMarker == "" { + // 空白 marker 让该项失去校验能力，会让双因子退化为仅 originator + // 单因子；视为非法配置，安全失败。 + return false + } + if !strings.Contains(ua, normalizedMarker) { + return false + } + } + return true +} + +// MatchAllowedClients 判断请求头是否命中 clientIDs 引用的任一预设签名。 +// 未知预设 ID 会被忽略；空列表恒不放行（默认拒绝）。 +func MatchAllowedClients(userAgent, originator string, clientIDs []string) bool { + for _, id := range clientIDs { + entry, ok := allowedClientRegistry[normalizeCodexClientHeader(id)] + if !ok { + continue + } + if IsAllowedClientMatch(userAgent, originator, entry) { + return true + } + } + return false +} diff --git a/backend/internal/pkg/openai/allowed_client_test.go b/backend/internal/pkg/openai/allowed_client_test.go new file mode 100644 index 00000000..c42aa4d5 --- /dev/null +++ b/backend/internal/pkg/openai/allowed_client_test.go @@ -0,0 +1,95 @@ +package openai + +import "testing" + +// 真实的 Claude Code codex 插件请求头：originator 与 UA 前缀同源于 clientInfo.name="Claude Code"。 +const ( + testClaudeCodeOriginator = "Claude Code" + testClaudeCodeUserAgent = "Claude Code/0.5.0 (Macos 15.5; arm64) iTerm2.app (Claude Code; 1.0.4)" +) + +func TestIsAllowedClientMatch(t *testing.T) { + entry := AllowedClientEntry{Originator: "Claude Code", UAContains: []string{"Claude Code/"}} + + tests := []struct { + name string + ua string + originator string + want bool + }{ + {name: "真实签名命中", ua: testClaudeCodeUserAgent, originator: testClaudeCodeOriginator, want: true}, + {name: "大小写不敏感", ua: "claude code/0.5.0 (macos)", originator: "claude code", want: true}, + {name: "originator 两侧空白被裁剪", ua: testClaudeCodeUserAgent, originator: " Claude Code ", want: true}, + {name: "originator 非精确（带后缀）不命中", ua: testClaudeCodeUserAgent, originator: "Claude Code Extra", want: false}, + {name: "originator 为空不命中", ua: testClaudeCodeUserAgent, originator: "", want: false}, + {name: "originator 是官方 codex 不命中", ua: testClaudeCodeUserAgent, originator: "codex_cli_rs", want: false}, + {name: "UA 缺少 Claude Code/ 标记不命中", ua: "curl/8.0", originator: testClaudeCodeOriginator, want: false}, + } + + for _, tt := range tests { + t.Run(tt.name, func(t *testing.T) { + if got := IsAllowedClientMatch(tt.ua, tt.originator, entry); got != tt.want { + t.Fatalf("IsAllowedClientMatch(%q, %q) = %v, want %v", tt.ua, tt.originator, got, tt.want) + } + }) + } +} + +func TestIsAllowedClientMatch_EmptyOriginatorEntryNeverMatches(t *testing.T) { + // registry 条目若没有配置 Originator，绝不放行，避免成为宽松后门。 + entry := AllowedClientEntry{Originator: "", UAContains: []string{"Claude Code/"}} + if IsAllowedClientMatch(testClaudeCodeUserAgent, "", entry) { + t.Fatal("空 Originator 的条目不应匹配任何请求") + } +} + +func TestIsAllowedClientMatch_EmptyUAContainsNeverMatches(t *testing.T) { + // 预设必须声明 UA 特征，否则退化为仅凭可伪造的 originator 单因子匹配，绝不放行。 + entry := AllowedClientEntry{Originator: "Claude Code", UAContains: nil} + if IsAllowedClientMatch(testClaudeCodeUserAgent, testClaudeCodeOriginator, entry) { + t.Fatal("未声明 UA 特征的预设不应匹配，避免退化为单因子 originator 匹配") + } +} + +func TestIsAllowedClientMatch_WhitespaceUAMarkerNeverMatches(t *testing.T) { + // 全空白 marker 归一化后为空，若被跳过则退化为仅 originator 单因子； + // 任何空白 marker 视为非法预设配置，必须安全失败。 + entry := AllowedClientEntry{Originator: "Claude Code", UAContains: []string{" "}} + if IsAllowedClientMatch(testClaudeCodeUserAgent, testClaudeCodeOriginator, entry) { + t.Fatal("UAContains 含全空白 marker 不应匹配，避免退化为单因子 originator 匹配") + } +} + +func TestIsAllowedClientMatch_MixedEmptyUAMarkerNeverMatches(t *testing.T) { + // 即便 UAContains 含一个真实 marker，只要其中混入任何空白 marker 也视为非法配置； + // 防止维护者只为对齐凑数而插入空字符串。 + entry := AllowedClientEntry{Originator: "Claude Code", UAContains: []string{"", "Claude Code/"}} + if IsAllowedClientMatch(testClaudeCodeUserAgent, testClaudeCodeOriginator, entry) { + t.Fatal("UAContains 混入空白 marker 不应匹配") + } +} + +func TestMatchAllowedClients(t *testing.T) { + tests := []struct { + name string + ua string + originator string + clientIDs []string + want bool + }{ + {name: "claude_code 预设命中真实签名", ua: testClaudeCodeUserAgent, originator: testClaudeCodeOriginator, clientIDs: []string{AllowedClientClaudeCode}, want: true}, + {name: "claude_code 预设 + 伪造 originator 不命中", ua: testClaudeCodeUserAgent, originator: "my_client", clientIDs: []string{AllowedClientClaudeCode}, want: false}, + {name: "空列表不放行", ua: testClaudeCodeUserAgent, originator: testClaudeCodeOriginator, clientIDs: nil, want: false}, + {name: "未知预设 ID 不放行", ua: testClaudeCodeUserAgent, originator: testClaudeCodeOriginator, clientIDs: []string{"unknown_client"}, want: false}, + {name: "ID 大小写/空白容错", ua: testClaudeCodeUserAgent, originator: testClaudeCodeOriginator, clientIDs: []string{" Claude_Code "}, want: true}, + {name: "多预设任一命中即放行", ua: testClaudeCodeUserAgent, originator: testClaudeCodeOriginator, clientIDs: []string{"unknown_client", AllowedClientClaudeCode}, want: true}, + } + + for _, tt := range tests { + t.Run(tt.name, func(t *testing.T) { + if got := MatchAllowedClients(tt.ua, tt.originator, tt.clientIDs); got != tt.want { + t.Fatalf("MatchAllowedClients(%q, %q, %v) = %v, want %v", tt.ua, tt.originator, tt.clientIDs, got, tt.want) + } + }) + } +} diff --git a/backend/internal/server/api_contract_test.go b/backend/internal/server/api_contract_test.go index 662daed1..9eea0924 100644 --- a/backend/internal/server/api_contract_test.go +++ b/backend/internal/server/api_contract_test.go @@ -843,6 +843,7 @@ func TestAPIContracts(t *testing.T) { "payment_visible_method_wxpay_enabled": false, "openai_advanced_scheduler_enabled": true, "openai_codex_user_agent": "", + "openai_allow_claude_code_codex_plugin": false, "openai_fast_policy_settings": { "rules": [] }, @@ -1079,6 +1080,7 @@ func TestAPIContracts(t *testing.T) { "payment_visible_method_wxpay_enabled": false, "openai_advanced_scheduler_enabled": false, "openai_codex_user_agent": "", + "openai_allow_claude_code_codex_plugin": false, "openai_fast_policy_settings": { "rules": [] }, diff --git a/backend/internal/service/account.go b/backend/internal/service/account.go index d488aa75..f51f0325 100644 --- a/backend/internal/service/account.go +++ b/backend/internal/service/account.go @@ -1442,6 +1442,38 @@ func (a *Account) IsCodexCLIOnlyEnabled() bool { return ok && enabled } +// GetCodexCLIOnlyAllowedClients 返回 codex_cli_only 之上额外放行的命名客户端预设 ID 列表。 +// 仅 OpenAI OAuth 账号生效；缺失或类型不符时返回空。预设 ID 的具体匹配规则由 +// openai 包的 registry 固化，配置只能引用预设键、不能自定义规则。 +func (a *Account) GetCodexCLIOnlyAllowedClients() []string { + if a == nil || !a.IsOpenAIOAuth() || a.Extra == nil { + return nil + } + raw, ok := a.Extra["codex_cli_only_allowed_clients"] + if !ok || raw == nil { + return nil + } + switch v := raw.(type) { + case []string: + result := make([]string, 0, len(v)) + for _, s := range v { + if strings.TrimSpace(s) != "" { + result = append(result, s) + } + } + return result + case []any: + result := make([]string, 0, len(v)) + for _, item := range v { + if s, ok := item.(string); ok && strings.TrimSpace(s) != "" { + result = append(result, s) + } + } + return result + } + return nil +} + // WindowCostSchedulability 窗口费用调度状态 type WindowCostSchedulability int diff --git a/backend/internal/service/account_codex_cli_only_allowed_clients_test.go b/backend/internal/service/account_codex_cli_only_allowed_clients_test.go new file mode 100644 index 00000000..c835ea27 --- /dev/null +++ b/backend/internal/service/account_codex_cli_only_allowed_clients_test.go @@ -0,0 +1,68 @@ +package service + +import ( + "testing" + + "github.com/stretchr/testify/require" +) + +func TestAccount_GetCodexCLIOnlyAllowedClients(t *testing.T) { + t.Run("OAuth 账号读取 []any 字符串列表", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only_allowed_clients": []any{"claude_code"}}, + } + require.Equal(t, []string{"claude_code"}, account.GetCodexCLIOnlyAllowedClients()) + }) + + t.Run("OAuth 账号读取 []string 列表", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only_allowed_clients": []string{"claude_code"}}, + } + require.Equal(t, []string{"claude_code"}, account.GetCodexCLIOnlyAllowedClients()) + }) + + t.Run("[]string 跳过空白元素", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only_allowed_clients": []string{"claude_code", "", " "}}, + } + require.Equal(t, []string{"claude_code"}, account.GetCodexCLIOnlyAllowedClients()) + }) + + t.Run("跳过非字符串与空白元素", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only_allowed_clients": []any{"claude_code", 123, "", " "}}, + } + require.Equal(t, []string{"claude_code"}, account.GetCodexCLIOnlyAllowedClients()) + }) + + t.Run("非 OAuth 账号返回空", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Extra: map[string]any{"codex_cli_only_allowed_clients": []any{"claude_code"}}, + } + require.Empty(t, account.GetCodexCLIOnlyAllowedClients()) + }) + + t.Run("Extra 为空返回空", func(t *testing.T) { + account := &Account{Platform: PlatformOpenAI, Type: AccountTypeOAuth} + require.Empty(t, account.GetCodexCLIOnlyAllowedClients()) + }) + + t.Run("字段缺失返回空", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{}, + } + require.Empty(t, account.GetCodexCLIOnlyAllowedClients()) + }) +} diff --git a/backend/internal/service/domain_constants.go b/backend/internal/service/domain_constants.go index 59c34eaa..b6441238 100644 --- a/backend/internal/service/domain_constants.go +++ b/backend/internal/service/domain_constants.go @@ -431,6 +431,9 @@ const ( // 当客户端 UA 被识别为浏览器（Chrome/Firefox/Safari/Edge 等）时，转发给 OpenAI 上游前会替换为此值， // 用于避免 Cloudflare 对浏览器型 UA 的质询拦截。 SettingKeyOpenAICodexUserAgent = "openai_codex_user_agent" + // SettingKeyOpenAIAllowClaudeCodeCodexPlugin 全局开关：是否额外放行 Claude Code 的 Codex 插件（默认 false）。 + // 仅在账号 codex_cli_only 开启时生效；开启后无需逐账号配置 codex_cli_only_allowed_clients。 + SettingKeyOpenAIAllowClaudeCodeCodexPlugin = "openai_allow_claude_code_codex_plugin" // 余额不足提醒 SettingKeyBalanceLowNotifyEnabled = "balance_low_notify_enabled" // 全局开关 diff --git a/backend/internal/service/openai_client_restriction_detector.go b/backend/internal/service/openai_client_restriction_detector.go index d1784e11..8589737a 100644 --- a/backend/internal/service/openai_client_restriction_detector.go +++ b/backend/internal/service/openai_client_restriction_detector.go @@ -13,6 +13,10 @@ const ( CodexClientRestrictionReasonMatchedUA = "official_client_user_agent_matched" // CodexClientRestrictionReasonMatchedOriginator 表示请求命中官方客户端 originator 白名单。 CodexClientRestrictionReasonMatchedOriginator = "official_client_originator_matched" + // CodexClientRestrictionReasonMatchedAllowedClient 表示请求命中账号级额外放行的命名客户端预设。 + CodexClientRestrictionReasonMatchedAllowedClient = "allowed_client_matched" + // CodexClientRestrictionReasonMatchedGlobalAllowedClient 表示请求命中全局额外放行的命名客户端预设。 + CodexClientRestrictionReasonMatchedGlobalAllowedClient = "global_allowed_client_matched" // CodexClientRestrictionReasonNotMatchedUA 表示请求未命中官方客户端 UA 白名单。 CodexClientRestrictionReasonNotMatchedUA = "official_client_user_agent_not_matched" // CodexClientRestrictionReasonForceCodexCLI 表示通过 ForceCodexCLI 配置兜底放行。 @@ -28,7 +32,7 @@ type CodexClientRestrictionDetectionResult struct { // CodexClientRestrictionDetector 定义 codex_cli_only 统一检测入口。 type CodexClientRestrictionDetector interface { - Detect(c *gin.Context, account *Account) CodexClientRestrictionDetectionResult + Detect(c *gin.Context, account *Account, globalAllowedClients []string) CodexClientRestrictionDetectionResult } // OpenAICodexClientRestrictionDetector 为 OpenAI OAuth codex_cli_only 的默认实现。 @@ -40,7 +44,7 @@ func NewOpenAICodexClientRestrictionDetector(cfg *config.Config) *OpenAICodexCli return &OpenAICodexClientRestrictionDetector{cfg: cfg} } -func (d *OpenAICodexClientRestrictionDetector) Detect(c *gin.Context, account *Account) CodexClientRestrictionDetectionResult { +func (d *OpenAICodexClientRestrictionDetector) Detect(c *gin.Context, account *Account, globalAllowedClients []string) CodexClientRestrictionDetectionResult { if account == nil || !account.IsCodexCLIOnlyEnabled() { return CodexClientRestrictionDetectionResult{ Enabled: false, @@ -78,6 +82,26 @@ func (d *OpenAICodexClientRestrictionDetector) Detect(c *gin.Context, account *A } } + // 官方客户端白名单未命中时，先尝试账号级额外放行的命名客户端预设（如 Claude Code codex 插件）。 + if allowed := account.GetCodexCLIOnlyAllowedClients(); len(allowed) > 0 && + openai.MatchAllowedClients(userAgent, originator, allowed) { + return CodexClientRestrictionDetectionResult{ + Enabled: true, + Matched: true, + Reason: CodexClientRestrictionReasonMatchedAllowedClient, + } + } + + // 再尝试由更高作用域（全局设置）注入的额外放行客户端列表。 + if len(globalAllowedClients) > 0 && + openai.MatchAllowedClients(userAgent, originator, globalAllowedClients) { + return CodexClientRestrictionDetectionResult{ + Enabled: true, + Matched: true, + Reason: CodexClientRestrictionReasonMatchedGlobalAllowedClient, + } + } + return CodexClientRestrictionDetectionResult{ Enabled: true, Matched: false, diff --git a/backend/internal/service/openai_client_restriction_detector_test.go b/backend/internal/service/openai_client_restriction_detector_test.go index 984b4ff6..fc115128 100644 --- a/backend/internal/service/openai_client_restriction_detector_test.go +++ b/backend/internal/service/openai_client_restriction_detector_test.go @@ -30,7 +30,7 @@ func TestOpenAICodexClientRestrictionDetector_Detect(t *testing.T) { detector := NewOpenAICodexClientRestrictionDetector(nil) account := &Account{Platform: PlatformOpenAI, Type: AccountTypeOAuth, Extra: map[string]any{}} - result := detector.Detect(newCodexDetectorTestContext("curl/8.0", ""), account) + result := detector.Detect(newCodexDetectorTestContext("curl/8.0", ""), account, nil) require.False(t, result.Enabled) require.False(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonDisabled, result.Reason) @@ -44,7 +44,7 @@ func TestOpenAICodexClientRestrictionDetector_Detect(t *testing.T) { Extra: map[string]any{"codex_cli_only": true}, } - result := detector.Detect(newCodexDetectorTestContext("codex_cli_rs/0.99.0", ""), account) + result := detector.Detect(newCodexDetectorTestContext("codex_cli_rs/0.99.0", ""), account, nil) require.True(t, result.Enabled) require.True(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonMatchedUA, result.Reason) @@ -58,7 +58,7 @@ func TestOpenAICodexClientRestrictionDetector_Detect(t *testing.T) { Extra: map[string]any{"codex_cli_only": true}, } - result := detector.Detect(newCodexDetectorTestContext("codex_vscode/1.0.0", ""), account) + result := detector.Detect(newCodexDetectorTestContext("codex_vscode/1.0.0", ""), account, nil) require.True(t, result.Enabled) require.True(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonMatchedUA, result.Reason) @@ -72,7 +72,7 @@ func TestOpenAICodexClientRestrictionDetector_Detect(t *testing.T) { Extra: map[string]any{"codex_cli_only": true}, } - result := detector.Detect(newCodexDetectorTestContext("codex_app/2.1.0", ""), account) + result := detector.Detect(newCodexDetectorTestContext("codex_app/2.1.0", ""), account, nil) require.True(t, result.Enabled) require.True(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonMatchedUA, result.Reason) @@ -86,7 +86,7 @@ func TestOpenAICodexClientRestrictionDetector_Detect(t *testing.T) { Extra: map[string]any{"codex_cli_only": true}, } - result := detector.Detect(newCodexDetectorTestContext("curl/8.0", "codex_chatgpt_desktop"), account) + result := detector.Detect(newCodexDetectorTestContext("curl/8.0", "codex_chatgpt_desktop"), account, nil) require.True(t, result.Enabled) require.True(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonMatchedOriginator, result.Reason) @@ -100,7 +100,7 @@ func TestOpenAICodexClientRestrictionDetector_Detect(t *testing.T) { Extra: map[string]any{"codex_cli_only": true}, } - result := detector.Detect(newCodexDetectorTestContext("curl/8.0", "my_client"), account) + result := detector.Detect(newCodexDetectorTestContext("curl/8.0", "my_client"), account, nil) require.True(t, result.Enabled) require.False(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonNotMatchedUA, result.Reason) @@ -116,9 +116,146 @@ func TestOpenAICodexClientRestrictionDetector_Detect(t *testing.T) { Extra: map[string]any{"codex_cli_only": true}, } - result := detector.Detect(newCodexDetectorTestContext("curl/8.0", "my_client"), account) + result := detector.Detect(newCodexDetectorTestContext("curl/8.0", "my_client"), account, nil) require.True(t, result.Enabled) require.True(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonForceCodexCLI, result.Reason) }) } + +func TestOpenAICodexClientRestrictionDetector_Detect_AllowedClients(t *testing.T) { + gin.SetMode(gin.TestMode) + + const ( + claudeCodeUA = "Claude Code/0.5.0 (Macos 15.5; arm64) iTerm2.app (Claude Code; 1.0.4)" + claudeCodeOriginator = "Claude Code" + ) + + t.Run("配置 claude_code 白名单且命中真实签名时放行", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{ + "codex_cli_only": true, + "codex_cli_only_allowed_clients": []any{"claude_code"}, + }, + } + + result := detector.Detect(newCodexDetectorTestContext(claudeCodeUA, claudeCodeOriginator), account, nil) + require.True(t, result.Enabled) + require.True(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonMatchedAllowedClient, result.Reason) + }) + + t.Run("配置白名单但伪造 originator 仍拒绝", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{ + "codex_cli_only": true, + "codex_cli_only_allowed_clients": []any{"claude_code"}, + }, + } + + result := detector.Detect(newCodexDetectorTestContext(claudeCodeUA, "my_client"), account, nil) + require.True(t, result.Enabled) + require.False(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonNotMatchedUA, result.Reason) + }) + + t.Run("未配置白名单时 Claude Code 签名仍拒绝", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only": true}, + } + + result := detector.Detect(newCodexDetectorTestContext(claudeCodeUA, claudeCodeOriginator), account, nil) + require.True(t, result.Enabled) + require.False(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonNotMatchedUA, result.Reason) + }) + + t.Run("未开启 codex_cli_only 时白名单不参与，直接绕过", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only_allowed_clients": []any{"claude_code"}}, + } + + result := detector.Detect(newCodexDetectorTestContext(claudeCodeUA, claudeCodeOriginator), account, nil) + require.False(t, result.Enabled) + require.False(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonDisabled, result.Reason) + }) + + t.Run("全局列表含 claude_code + 命中签名 → 放行(global)", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only": true}, + } + result := detector.Detect( + newCodexDetectorTestContext("Claude Code/0.5.0 (Macos 15.5; arm64) iTerm2.app (Claude Code; 1.0.4)", "Claude Code"), + account, + []string{"claude_code"}, + ) + require.True(t, result.Enabled) + require.True(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonMatchedGlobalAllowedClient, result.Reason) + }) + + t.Run("全局列表含 claude_code + 非签名 → 403", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only": true}, + } + result := detector.Detect(newCodexDetectorTestContext("curl/8.0", "my_client"), account, []string{"claude_code"}) + require.True(t, result.Enabled) + require.False(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonNotMatchedUA, result.Reason) + }) + + t.Run("全局列表为空 + 账号未配 → 403", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{"codex_cli_only": true}, + } + result := detector.Detect( + newCodexDetectorTestContext("Claude Code/0.5.0 (Macos) (Claude Code; 1.0.4)", "Claude Code"), + account, + nil, + ) + require.True(t, result.Enabled) + require.False(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonNotMatchedUA, result.Reason) + }) + + t.Run("账号白名单优先于全局列表（reason=account）", func(t *testing.T) { + detector := NewOpenAICodexClientRestrictionDetector(nil) + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Extra: map[string]any{ + "codex_cli_only": true, + "codex_cli_only_allowed_clients": []any{"claude_code"}, + }, + } + result := detector.Detect( + newCodexDetectorTestContext("Claude Code/0.5.0 (Macos) (Claude Code; 1.0.4)", "Claude Code"), + account, + []string{"claude_code"}, + ) + require.True(t, result.Matched) + require.Equal(t, CodexClientRestrictionReasonMatchedAllowedClient, result.Reason) + }) +} diff --git a/backend/internal/service/openai_gateway_service.go b/backend/internal/service/openai_gateway_service.go index f93cc221..997423b7 100644 --- a/backend/internal/service/openai_gateway_service.go +++ b/backend/internal/service/openai_gateway_service.go @@ -901,7 +901,17 @@ func SnapshotOpenAICompatibilityFallbackMetrics() OpenAICompatibilityFallbackMet } func (s *OpenAIGatewayService) detectCodexClientRestriction(c *gin.Context, account *Account) CodexClientRestrictionDetectionResult { - return s.getCodexClientRestrictionDetector().Detect(c, account) + var globalAllowedClients []string + if account != nil && account.IsCodexCLIOnlyEnabled() && s != nil && s.settingService != nil { + ctx := context.Background() + if c != nil && c.Request != nil { + ctx = c.Request.Context() + } + if s.settingService.IsOpenAIAllowClaudeCodeCodexPluginEnabled(ctx) { + globalAllowedClients = []string{openai.AllowedClientClaudeCode} + } + } + return s.getCodexClientRestrictionDetector().Detect(c, account, globalAllowedClients) } func getAPIKeyIDFromContext(c *gin.Context) int64 { @@ -959,6 +969,7 @@ func logCodexCLIOnlyDetection(ctx context.Context, c *gin.Context, account *Acco } log := logger.FromContext(ctx).With(fields...) if result.Matched { + log.Info("OpenAI codex_cli_only 放行请求") return } log.Warn("OpenAI codex_cli_only 拒绝非官方客户端请求") diff --git a/backend/internal/service/openai_gateway_service_codex_cli_only_test.go b/backend/internal/service/openai_gateway_service_codex_cli_only_test.go index 17a874ea..10d58654 100644 --- a/backend/internal/service/openai_gateway_service_codex_cli_only_test.go +++ b/backend/internal/service/openai_gateway_service_codex_cli_only_test.go @@ -18,7 +18,7 @@ type stubCodexRestrictionDetector struct { result CodexClientRestrictionDetectionResult } -func (s *stubCodexRestrictionDetector) Detect(_ *gin.Context, _ *Account) CodexClientRestrictionDetectionResult { +func (s *stubCodexRestrictionDetector) Detect(_ *gin.Context, _ *Account, _ []string) CodexClientRestrictionDetectionResult { return s.result } @@ -52,7 +52,7 @@ func TestOpenAIGatewayService_GetCodexClientRestrictionDetector(t *testing.T) { c.Request.Header.Set("User-Agent", "curl/8.0") account := &Account{Platform: PlatformOpenAI, Type: AccountTypeOAuth, Extra: map[string]any{"codex_cli_only": true}} - result := got.Detect(c, account) + result := got.Detect(c, account, nil) require.True(t, result.Enabled) require.True(t, result.Matched) require.Equal(t, CodexClientRestrictionReasonForceCodexCLI, result.Reason) diff --git a/backend/internal/service/setting_service.go b/backend/internal/service/setting_service.go index e6f0f2bc..08c0d045 100644 --- a/backend/internal/service/setting_service.go +++ b/backend/internal/service/setting_service.go @@ -141,6 +141,17 @@ const openAICodexUserAgentCacheTTL = 60 * time.Second const openAICodexUserAgentErrorTTL = 5 * time.Second const openAICodexUserAgentDBTimeout = 5 * time.Second +// cachedOpenAIAllowCodexPlugin Codex 插件放行开关缓存（进程内缓存，60s TTL）。 +// IsOpenAIAllowClaudeCodeCodexPluginEnabled 在每个 codex_cli_only 账号的网关请求热路径上被调用，避免每次访问 DB。 +type cachedOpenAIAllowCodexPlugin struct { + value bool + expiresAt int64 // unix nano +} + +const openAIAllowCodexPluginCacheTTL = 60 * time.Second +const openAIAllowCodexPluginErrorTTL = 5 * time.Second +const openAIAllowCodexPluginDBTimeout = 5 * time.Second + // DefaultSubscriptionGroupReader validates group references used by default subscriptions. type DefaultSubscriptionGroupReader interface { GetByID(ctx context.Context, id int64) (*Group, error) @@ -152,17 +163,19 @@ type WebSearchManagerBuilder func(cfg *WebSearchEmulationConfig, proxyURLs map[i // SettingService 系统设置服务 type SettingService struct { - settingRepo SettingRepository - defaultSubGroupReader DefaultSubscriptionGroupReader - proxyRepo ProxyRepository // for resolving websearch provider proxy URLs - cfg *config.Config - onUpdate func() // Callback when settings are updated (for cache invalidation) - version string // Application version - webSearchManagerBuilder WebSearchManagerBuilder - antigravityUAVersionCache atomic.Value // *cachedAntigravityUserAgentVersion - antigravityUAVersionSF singleflight.Group - openAICodexUACache atomic.Value // *cachedOpenAICodexUserAgent - openAICodexUASF singleflight.Group + settingRepo SettingRepository + defaultSubGroupReader DefaultSubscriptionGroupReader + proxyRepo ProxyRepository // for resolving websearch provider proxy URLs + cfg *config.Config + onUpdate func() // Callback when settings are updated (for cache invalidation) + version string // Application version + webSearchManagerBuilder WebSearchManagerBuilder + antigravityUAVersionCache atomic.Value // *cachedAntigravityUserAgentVersion + antigravityUAVersionSF singleflight.Group + openAICodexUACache atomic.Value // *cachedOpenAICodexUserAgent + openAICodexUASF singleflight.Group + openAIAllowCodexPluginCache atomic.Value // *cachedOpenAIAllowCodexPlugin + openAIAllowCodexPluginSF singleflight.Group } // DefaultPlatformQuotaSetting 单 platform 三档限额（nil = 沿用上层；0 = 显式禁用；>0 = 上限） @@ -1015,6 +1028,54 @@ func (s *SettingService) GetOpenAICodexUserAgent(ctx context.Context) string { return fallback } +// IsOpenAIAllowClaudeCodeCodexPluginEnabled 全局开关：是否额外放行 Claude Code 的 Codex 插件（默认关闭）。 +// 仅在调用方已确认账号 codex_cli_only 开启时读取，避免对非受限账号产生无谓查询。 +// 使用进程内 atomic.Value 缓存（60s TTL），避免在每个网关请求热路径上访问 DB。 +func (s *SettingService) IsOpenAIAllowClaudeCodeCodexPluginEnabled(ctx context.Context) bool { + if cached, ok := s.openAIAllowCodexPluginCache.Load().(*cachedOpenAIAllowCodexPlugin); ok && cached != nil { + if time.Now().UnixNano() < cached.expiresAt { + return cached.value + } + } + result, _, _ := s.openAIAllowCodexPluginSF.Do("openai_allow_codex_plugin_enabled", func() (any, error) { + if cached, ok := s.openAIAllowCodexPluginCache.Load().(*cachedOpenAIAllowCodexPlugin); ok && cached != nil { + if time.Now().UnixNano() < cached.expiresAt { + return cached.value, nil + } + } + dbCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), openAIAllowCodexPluginDBTimeout) + defer cancel() + value, err := s.settingRepo.GetValue(dbCtx, SettingKeyOpenAIAllowClaudeCodeCodexPlugin) + if err != nil { + if errors.Is(err, ErrSettingNotFound) { + // 设置不存在 → 默认关闭，正常 TTL 缓存 + s.openAIAllowCodexPluginCache.Store(&cachedOpenAIAllowCodexPlugin{ + value: false, + expiresAt: time.Now().Add(openAIAllowCodexPluginCacheTTL).UnixNano(), + }) + return false, nil + } + slog.Warn("failed to get openai_allow_claude_code_codex_plugin setting", "error", err) + // DB 错误 → 安全默认关闭，短 TTL 快速重试 + s.openAIAllowCodexPluginCache.Store(&cachedOpenAIAllowCodexPlugin{ + value: false, + expiresAt: time.Now().Add(openAIAllowCodexPluginErrorTTL).UnixNano(), + }) + return false, nil + } + enabled := value == "true" + s.openAIAllowCodexPluginCache.Store(&cachedOpenAIAllowCodexPlugin{ + value: enabled, + expiresAt: time.Now().Add(openAIAllowCodexPluginCacheTTL).UnixNano(), + }) + return enabled, nil + }) + if val, ok := result.(bool); ok { + return val + } + return false +} + // SetOnUpdateCallback sets a callback function to be called when settings are updated // This is used for cache invalidation (e.g., HTML cache in frontend server) func (s *SettingService) SetOnUpdateCallback(callback func()) { @@ -1816,6 +1877,7 @@ func (s *SettingService) buildSystemSettingsUpdates(ctx context.Context, setting updates[SettingKeyRewriteMessageCacheControl] = strconv.FormatBool(settings.RewriteMessageCacheControl) updates[SettingKeyAntigravityUserAgentVersion] = antigravity.NormalizeUserAgentVersion(settings.AntigravityUserAgentVersion) updates[SettingKeyOpenAICodexUserAgent] = strings.TrimSpace(settings.OpenAICodexUserAgent) + updates[SettingKeyOpenAIAllowClaudeCodeCodexPlugin] = strconv.FormatBool(settings.OpenAIAllowClaudeCodeCodexPlugin) updates[SettingPaymentVisibleMethodAlipaySource] = settings.PaymentVisibleMethodAlipaySource updates[SettingPaymentVisibleMethodWxpaySource] = settings.PaymentVisibleMethodWxpaySource updates[SettingPaymentVisibleMethodAlipayEnabled] = strconv.FormatBool(settings.PaymentVisibleMethodAlipayEnabled) @@ -1968,6 +2030,11 @@ func (s *SettingService) refreshCachedSettings(settings *SystemSettings) { if s.cfg != nil { s.cfg.SetTrustForwardedIPForAPIKeyACL(settings.APIKeyACLTrustForwardedIP) } + s.openAIAllowCodexPluginSF.Forget("openai_allow_codex_plugin_enabled") + s.openAIAllowCodexPluginCache.Store(&cachedOpenAIAllowCodexPlugin{ + value: settings.OpenAIAllowClaudeCodeCodexPlugin, + expiresAt: time.Now().Add(openAIAllowCodexPluginCacheTTL).UnixNano(), + }) if s.onUpdate != nil { s.onUpdate() // Invalidate cache after settings update } @@ -3233,6 +3300,7 @@ func (s *SettingService) parseSettings(settings map[string]string) *SystemSettin } result.AntigravityUserAgentVersion = antigravity.NormalizeUserAgentVersion(settings[SettingKeyAntigravityUserAgentVersion]) result.OpenAICodexUserAgent = strings.TrimSpace(settings[SettingKeyOpenAICodexUserAgent]) + result.OpenAIAllowClaudeCodeCodexPlugin = settings[SettingKeyOpenAIAllowClaudeCodeCodexPlugin] == "true" // Web search emulation: quick enabled check from the JSON config if raw := settings[SettingKeyWebSearchEmulationConfig]; raw != "" { diff --git a/backend/internal/service/setting_service_openai_allow_claude_code_test.go b/backend/internal/service/setting_service_openai_allow_claude_code_test.go new file mode 100644 index 00000000..22059f07 --- /dev/null +++ b/backend/internal/service/setting_service_openai_allow_claude_code_test.go @@ -0,0 +1,55 @@ +package service + +import ( + "context" + "testing" + + "github.com/Wei-Shaw/sub2api/internal/config" + "github.com/stretchr/testify/require" +) + +type allowClaudeCodeSettingRepoStub struct{ values map[string]string } + +func (s *allowClaudeCodeSettingRepoStub) Get(ctx context.Context, key string) (*Setting, error) { + panic("unused") +} +func (s *allowClaudeCodeSettingRepoStub) GetValue(ctx context.Context, key string) (string, error) { + if v, ok := s.values[key]; ok { + return v, nil + } + return "", ErrSettingNotFound +} +func (s *allowClaudeCodeSettingRepoStub) Set(ctx context.Context, key, value string) error { + panic("unused") +} +func (s *allowClaudeCodeSettingRepoStub) GetMultiple(ctx context.Context, keys []string) (map[string]string, error) { + panic("unused") +} +func (s *allowClaudeCodeSettingRepoStub) SetMultiple(ctx context.Context, settings map[string]string) error { + panic("unused") +} +func (s *allowClaudeCodeSettingRepoStub) GetAll(ctx context.Context) (map[string]string, error) { + panic("unused") +} +func (s *allowClaudeCodeSettingRepoStub) Delete(ctx context.Context, key string) error { + panic("unused") +} + +func TestSettingService_IsOpenAIAllowClaudeCodeCodexPluginEnabled(t *testing.T) { + t.Run("默认关闭（设置缺失）", func(t *testing.T) { + svc := NewSettingService(&allowClaudeCodeSettingRepoStub{values: map[string]string{}}, &config.Config{}) + require.False(t, svc.IsOpenAIAllowClaudeCodeCodexPluginEnabled(context.Background())) + }) + t.Run("值为 true 时开启", func(t *testing.T) { + svc := NewSettingService(&allowClaudeCodeSettingRepoStub{values: map[string]string{ + SettingKeyOpenAIAllowClaudeCodeCodexPlugin: "true", + }}, &config.Config{}) + require.True(t, svc.IsOpenAIAllowClaudeCodeCodexPluginEnabled(context.Background())) + }) + t.Run("值非 true 时关闭", func(t *testing.T) { + svc := NewSettingService(&allowClaudeCodeSettingRepoStub{values: map[string]string{ + SettingKeyOpenAIAllowClaudeCodeCodexPlugin: "false", + }}, &config.Config{}) + require.False(t, svc.IsOpenAIAllowClaudeCodeCodexPluginEnabled(context.Background())) + }) +} diff --git a/backend/internal/service/settings_view.go b/backend/internal/service/settings_view.go index 3f961ab2..7b45ef1a 100644 --- a/backend/internal/service/settings_view.go +++ b/backend/internal/service/settings_view.go @@ -195,6 +195,7 @@ type SystemSettings struct { RewriteMessageCacheControl bool // 是否改写 messages[*].content[*].cache_control（默认 false） AntigravityUserAgentVersion string // Antigravity 上游 User-Agent 版本号；空值使用配置/默认值 OpenAICodexUserAgent string // OpenAI Codex 上游完整 User-Agent；空值使用内置默认 + OpenAIAllowClaudeCodeCodexPlugin bool // 全局开关：是否额外放行 Claude Code 的 Codex 插件（默认 false） // Web Search Emulation WebSearchEmulationEnabled bool // 是否启用 web search 模拟 diff --git a/frontend/src/api/admin/settings.ts b/frontend/src/api/admin/settings.ts index d2b878cc..6d8e6cee 100644 --- a/frontend/src/api/admin/settings.ts +++ b/frontend/src/api/admin/settings.ts @@ -560,6 +560,7 @@ export interface SystemSettings { rewrite_message_cache_control: boolean; antigravity_user_agent_version: string; openai_codex_user_agent: string; + openai_allow_claude_code_codex_plugin: boolean; web_search_emulation_enabled?: boolean; // Payment configuration @@ -792,6 +793,7 @@ export interface UpdateSettingsRequest { rewrite_message_cache_control?: boolean; antigravity_user_agent_version?: string; openai_codex_user_agent?: string; + openai_allow_claude_code_codex_plugin?: boolean; // Payment configuration payment_enabled?: boolean; risk_control_enabled?: boolean; diff --git a/frontend/src/components/account/BulkEditAccountModal.vue b/frontend/src/components/account/BulkEditAccountModal.vue index c8d53220..6e71fe4b 100644 --- a/frontend/src/components/account/BulkEditAccountModal.vue +++ b/frontend/src/components/account/BulkEditAccountModal.vue @@ -742,6 +742,50 @@ + +
+
+ + +
+
+

+ {{ t('admin.accounts.openai.codexCLIOnlyAllowClaudeCodeDesc') }} +

+ +
+
+
@@ -1219,6 +1263,7 @@ const enableOpenAIPassthrough = ref(false) const enableOpenAIWSMode = ref(false) const enableOpenAIAPIKeyWSMode = ref(false) const enableCodexCLIOnly = ref(false) +const enableCodexCLIOnlyAllowClaudeCode = ref(false) const enableOpenAICompactMode = ref(false) const enableOpenAICompactModelMapping = ref(false) const enableRpmLimit = ref(false) @@ -1246,6 +1291,7 @@ const openaiPassthroughEnabled = ref(false) const openaiOAuthResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const openaiAPIKeyResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const codexCLIOnlyEnabled = ref(false) +const codexCLIOnlyAllowClaudeCodeEnabled = ref(false) const openAICompactMode = ref('auto') const openAICompactModelMappings = ref([]) const rpmLimitEnabled = ref(false) @@ -1496,6 +1542,11 @@ const buildUpdatePayload = (): Record | null => { extra.codex_cli_only = codexCLIOnlyEnabled.value } + if (enableCodexCLIOnlyAllowClaudeCode.value) { + const extra = ensureExtra() + extra.codex_cli_only_allowed_clients = codexCLIOnlyAllowClaudeCodeEnabled.value ? ['claude_code'] : [] + } + if (enableOpenAICompactMode.value) { const extra = ensureExtra() extra.openai_compact_mode = openAICompactMode.value @@ -1602,6 +1653,7 @@ const handleSubmit = async () => { enableOpenAIWSMode.value || enableOpenAIAPIKeyWSMode.value || enableCodexCLIOnly.value || + enableCodexCLIOnlyAllowClaudeCode.value || enableOpenAICompactMode.value || enableOpenAICompactModelMapping.value || enableRpmLimit.value || @@ -1704,6 +1756,7 @@ watch( enableOpenAIWSMode.value = false enableOpenAIAPIKeyWSMode.value = false enableCodexCLIOnly.value = false + enableCodexCLIOnlyAllowClaudeCode.value = false enableOpenAICompactMode.value = false enableOpenAICompactModelMapping.value = false enableRpmLimit.value = false @@ -1727,6 +1780,7 @@ watch( openaiOAuthResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF codexCLIOnlyEnabled.value = false + codexCLIOnlyAllowClaudeCodeEnabled.value = false openAICompactMode.value = 'auto' openAICompactModelMappings.value = [] rpmLimitEnabled.value = false diff --git a/frontend/src/components/account/CreateAccountModal.vue b/frontend/src/components/account/CreateAccountModal.vue index 331295f7..c699df37 100644 --- a/frontend/src/components/account/CreateAccountModal.vue +++ b/frontend/src/components/account/CreateAccountModal.vue @@ -2635,6 +2635,32 @@ />
+
+
+ +

+ {{ t('admin.accounts.openai.codexCLIOnlyAllowClaudeCodeDesc') }} +

+
+ +
@@ -3353,6 +3379,7 @@ const openAIResponsesMode = ref('auto') const openaiOAuthResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const openaiAPIKeyResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const codexCLIOnlyEnabled = ref(false) +const codexCLIOnlyAllowClaudeCodeEnabled = ref(false) const anthropicPassthroughEnabled = ref(false) const webSearchEmulationMode = ref('default') const webSearchGlobalEnabled = ref(false) @@ -3724,6 +3751,7 @@ watch( openaiOAuthResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF codexCLIOnlyEnabled.value = false + codexCLIOnlyAllowClaudeCodeEnabled.value = false } if (newPlatform !== 'anthropic') { anthropicPassthroughEnabled.value = false @@ -3744,6 +3772,7 @@ watch( ([category, platform]) => { if (platform === 'openai' && category !== 'oauth-based') { codexCLIOnlyEnabled.value = false + codexCLIOnlyAllowClaudeCodeEnabled.value = false } if (platform !== 'anthropic' || category !== 'apikey') { anthropicPassthroughEnabled.value = false @@ -4123,6 +4152,7 @@ const resetForm = () => { openaiOAuthResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF codexCLIOnlyEnabled.value = false + codexCLIOnlyAllowClaudeCodeEnabled.value = false anthropicPassthroughEnabled.value = false webSearchEmulationMode.value = 'default' // Reset quota control state @@ -4201,6 +4231,15 @@ const buildOpenAIExtra = (base?: Record): Record +
+
+ +

+ {{ t('admin.accounts.openai.codexCLIOnlyAllowClaudeCodeDesc') }} +

+
+ +
('auto') const openaiOAuthResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const openaiAPIKeyResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const codexCLIOnlyEnabled = ref(false) +const codexCLIOnlyAllowClaudeCodeEnabled = ref(false) type CodexImageGenerationBridgeMode = 'inherit' | 'enabled' | 'disabled' const codexImageGenerationBridgeMode = ref('inherit') const anthropicPassthroughEnabled = ref(false) @@ -2728,6 +2755,7 @@ const syncFormFromAccount = (newAccount: Account | null) => { openaiOAuthResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF codexCLIOnlyEnabled.value = false + codexCLIOnlyAllowClaudeCodeEnabled.value = false codexImageGenerationBridgeMode.value = 'inherit' anthropicPassthroughEnabled.value = false webSearchEmulationMode.value = 'default' @@ -2759,6 +2787,9 @@ const syncFormFromAccount = (newAccount: Account | null) => { }) if (newAccount.type === 'oauth') { codexCLIOnlyEnabled.value = extra?.codex_cli_only === true + codexCLIOnlyAllowClaudeCodeEnabled.value = + Array.isArray(extra?.codex_cli_only_allowed_clients) && + (extra.codex_cli_only_allowed_clients as unknown[]).includes('claude_code') } const credentials = newAccount.credentials as Record | undefined const compactMappings = credentials?.compact_model_mapping as Record | undefined @@ -3877,6 +3908,12 @@ const handleSubmit = async () => { } else { delete newExtra.codex_cli_only } + // 仅当 codex_cli_only 开启且子开关开启时写入 Claude Code 插件白名单，否则清除避免孤立字段 + if (codexCLIOnlyEnabled.value && codexCLIOnlyAllowClaudeCodeEnabled.value) { + newExtra.codex_cli_only_allowed_clients = ['claude_code'] + } else { + delete newExtra.codex_cli_only_allowed_clients + } } updatePayload.extra = newExtra diff --git a/frontend/src/components/account/__tests__/BulkEditAccountModal.spec.ts b/frontend/src/components/account/__tests__/BulkEditAccountModal.spec.ts index caa307fc..3ae75ee9 100644 --- a/frontend/src/components/account/__tests__/BulkEditAccountModal.spec.ts +++ b/frontend/src/components/account/__tests__/BulkEditAccountModal.spec.ts @@ -197,6 +197,25 @@ describe('BulkEditAccountModal', () => { }) }) + it('OpenAI OAuth 批量编辑应提交 codex_cli_only_allowed_clients 字段', async () => { + const wrapper = mountModal({ + selectedPlatforms: ['openai'], + selectedTypes: ['oauth'] + }) + + await wrapper.get('#bulk-edit-openai-codex-allow-claude-code-enabled').setValue(true) + await wrapper.get('#bulk-edit-openai-codex-allow-claude-code-toggle').trigger('click') + await wrapper.get('#bulk-edit-account-form').trigger('submit.prevent') + await flushPromises() + + expect(adminAPI.accounts.bulkUpdate).toHaveBeenCalledTimes(1) + expect(adminAPI.accounts.bulkUpdate).toHaveBeenCalledWith([1, 2], { + extra: { + codex_cli_only_allowed_clients: ['claude_code'] + } + }) + }) + it('OpenAI API Key 批量编辑应提交 API Key 专属 WS mode 字段', async () => { const wrapper = mountModal({ selectedPlatforms: ['openai'], diff --git a/frontend/src/i18n/locales/en.ts b/frontend/src/i18n/locales/en.ts index ff5ea651..956b5e7a 100644 --- a/frontend/src/i18n/locales/en.ts +++ b/frontend/src/i18n/locales/en.ts @@ -3338,6 +3338,9 @@ export default { codexCLIOnly: 'Codex official clients only', codexCLIOnlyDesc: 'Only applies to OpenAI OAuth. When enabled, only Codex official client families are allowed; when disabled, the gateway bypasses this restriction and keeps existing behavior.', + codexCLIOnlyAllowClaudeCode: "Also allow Claude Code's Codex plugin", + codexCLIOnlyAllowClaudeCodeDesc: + 'Only takes effect when the switch above is on. Additionally allows requests from the Claude Code Codex plugin (exact match on originator=Claude Code) without weakening blocking of other non-official clients.', codexImageGenerationBridge: 'Codex image-generation bridge', codexImageGenerationBridgeDesc: 'Account policy takes precedence over channel and global settings. Only controls whether Codex requests through the /responses text endpoint receive the image_generation tool; standalone image-generation endpoints are unaffected.', @@ -5577,6 +5580,9 @@ export default { openaiCodexUserAgent: 'OpenAI Codex UA', openaiCodexUserAgentPlaceholder: 'codex-tui/0.125.0 (Ubuntu 22.4.0; x86_64) xterm-256color (codex-tui; 0.125.0)', openaiCodexUserAgentHint: 'Used to bypass Cloudflare browser-UA challenges on the OpenAI upstream. Only applies when the client User-Agent is detected as a browser (Mozilla/...). Leave empty to use the built-in default.', + openaiAllowClaudeCodeCodexPlugin: "Allow using the Codex plugin in Claude Code", + openaiAllowClaudeCodeCodexPluginDesc: + "Global switch; only affects OpenAI OAuth accounts that have 'Codex official clients only' enabled. When on, all such accounts additionally allow requests from the Claude Code Codex plugin (exact match on originator=Claude Code) without per-account config; upstream requests remain pass-through.", }, webSearchEmulation: { title: 'Web Search Emulation', diff --git a/frontend/src/i18n/locales/zh.ts b/frontend/src/i18n/locales/zh.ts index b8ac7d2c..2bdebf06 100644 --- a/frontend/src/i18n/locales/zh.ts +++ b/frontend/src/i18n/locales/zh.ts @@ -3483,6 +3483,8 @@ export default { responsesStatusForcedChatCompletions: '已强制 Chat Completions', codexCLIOnly: '仅允许 Codex 官方客户端', codexCLIOnlyDesc: '仅对 OpenAI OAuth 生效。开启后仅允许 Codex 官方客户端家族访问；关闭后完全绕过并保持原逻辑。', + codexCLIOnlyAllowClaudeCode: '额外放行 Claude Code 的 Codex 插件', + codexCLIOnlyAllowClaudeCodeDesc: '仅在上方开关开启时生效。额外放行通过 Claude Code 的 Codex 插件发起的请求（精确匹配 originator=Claude Code），不影响对其他非官方客户端的拦截。', codexImageGenerationBridge: 'Codex 图片生成桥接', codexImageGenerationBridgeDesc: '账号级策略优先于渠道和全局配置。仅控制 Codex 走 /responses 文本端点时是否注入 image_generation 工具；不影响独立图片生成接口。', @@ -5733,6 +5735,9 @@ export default { openaiCodexUserAgent: 'OpenAI Codex UA', openaiCodexUserAgentPlaceholder: 'codex-tui/0.125.0 (Ubuntu 22.4.0; x86_64) xterm-256color (codex-tui; 0.125.0)', openaiCodexUserAgentHint: '用于规避 OpenAI 上游 Cloudflare 对浏览器 UA 的访问质询。仅在检测到客户端 User-Agent 为浏览器（Mozilla/...）时生效，其他客户端原样透传。留空使用内置默认值。', + openaiAllowClaudeCodeCodexPlugin: '允许在 Claude Code 中使用 Codex 插件', + openaiAllowClaudeCodeCodexPluginDesc: + '全局开关，仅对已开启「仅允许 Codex 官方客户端」的 OpenAI OAuth 账号生效。开启后，所有此类账号都额外放行通过 Claude Code 的 Codex 插件发起的请求（精确匹配 originator=Claude Code），无需逐账号配置；上游请求仍保持透传。', }, webSearchEmulation: { title: 'Web Search 模拟', diff --git a/frontend/src/views/admin/SettingsView.vue b/frontend/src/views/admin/SettingsView.vue index 68eb4849..239ce2d7 100644 --- a/frontend/src/views/admin/SettingsView.vue +++ b/frontend/src/views/admin/SettingsView.vue @@ -3948,6 +3948,19 @@ }}

+ + +
+
+ +

+ {{ t("admin.settings.gatewayForwarding.openaiAllowClaudeCodeCodexPluginDesc") }} +

+
+ +
@@ -7162,6 +7175,7 @@ const form = reactive({ rewrite_message_cache_control: false, antigravity_user_agent_version: "", openai_codex_user_agent: "", + openai_allow_claude_code_codex_plugin: false, // 余额、订阅到期与账号限额通知 balance_low_notify_enabled: false, balance_low_notify_threshold: 0, @@ -8267,6 +8281,7 @@ async function saveSettings() { form.antigravity_user_agent_version?.trim() || "", openai_codex_user_agent: form.openai_codex_user_agent?.trim() || "", + openai_allow_claude_code_codex_plugin: form.openai_allow_claude_code_codex_plugin, // Payment configuration payment_enabled: form.payment_enabled, risk_control_enabled: form.risk_control_enabled, From ddf91e9a7f7f7c72160d5097ffe2211ee6c20930 Mon Sep 17 00:00:00 2001 From: alfadb Date: Wed, 27 May 2026 15:46:05 +0800 Subject: [PATCH 03/20] =?UTF-8?q?fix(gateway):=20=E6=8C=89=E6=9C=80?= =?UTF-8?q?=E7=BB=88=20anthropic-beta=20header=20=E5=AF=B9=20body.context?= =?UTF-8?q?=5Fmanagement=20=E5=81=9A=E8=83=BD=E5=8A=9B=E7=BB=B4=E5=BA=A6?= =?UTF-8?q?=20sanitize?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 上游 Anthropic 在 body 含 `context_management` 但最终发出去的 `anthropic-beta` header 不含 `context-management-2025-06-27` 时会拒收： { "type": "invalid_request_error", "message": "context_management: Extra inputs are not permitted" } (HTTP 400, request_id 形如 req_011C...) 该 400 在 haiku 路径上触发,因为三个 beta header 构造器有意排除了 context-management beta: - HaikuBetaHeader (messages, OAuth / mimic CC) - APIKeyHaikuBetaHeader (messages, API-key) - CountTokensBetaHeader (count_tokens, 所有认证类型) 但 body 中仍然带着 `context_management` 字段,原因有二: 1. normalizeClaudeOAuthRequestBody 在 thinking_enabled / thinking_adaptive 打开时为 `clear_thinking_20251015` 主动注入; 2. 客户端 (Claude Code CLI >= 2.1.87) 原样发送, 网关透传时一并转发。 修复方案: 能力维度对称约束 ========================== 对齐已有的 Bedrock 模式 (`backend/internal/service/bedrock_request.go` 中的 `sanitizeBedrockFieldsForBetaTokens`): 根据 **最终** 发出的 `anthropic-beta` header 决定是否保留 `body.context_management`, 而不是按 model 名或路由分类来决定。 新增纯函数: sanitizeAnthropicBodyForBetaTokens(body, betaHeader) (body, changed) 如果 `betaHeader` 不含 `context-management-2025-06-27`, 用 sjson 把 body 字段 strip 掉; 否则原样返回。 在所有 Anthropic / Anthropic-兼容 上游出口都接入: | 路径 | sanitize 接入点 | |--------------------------------------------|-------------------------------------------------------| | /v1/messages OAuth mimic CC | buildUpstreamRequest | | /v1/messages OAuth 真 CC 透传 | buildUpstreamRequest | | /v1/messages API-key | buildUpstreamRequest | | /v1/messages API-key passthrough | buildUpstreamRequestAnthropicAPIKeyPassthrough | | /v1/messages Vertex / service-account | buildUpstreamRequestAnthropicVertex | | /v1/messages/count_tokens (全部 4 条路径) | buildCountTokensRequest, | | | buildCountTokensRequestAnthropicAPIKeyPassthrough | | Antigravity Anthropic-兼容 上游 | AntigravityGatewayService.ForwardUpstream | | Bedrock | (已由 sanitizeBedrockFieldsForBetaTokens 处理) | 为什么要重排 (而不是加一行调用) ================================ sanitize 必须 **在** `signBillingHeaderCCH` 之前运行。CCH 对整个 body 取 xxHash64 摘要后写入 billing header 里 5 位十六进制的 `cch` 字段; 如果先签名再 strip, 上游对发出去的 body 重算 hash 会和 `cch` 不一致, 请求被判为 third-party。这就要求在 `http.NewRequest` 之前算出最终的 `anthropic-beta` header, 所以把原本内联在 builder 里的 beta 计算逻辑 抽成了两个纯函数: - computeFinalAnthropicBeta (messages 路径: mimic 不透传 客户端 beta) - computeFinalCountTokensAnthropicBeta (count_tokens 路径: mimic 不 跳过白名单透传) 两者逐位保留原行为: - mimic 路径在 messages 上跳过客户端 beta, 在 count_tokens 上合并 - API-key 路径尊重 `InjectBetaForAPIKey` 开关 - dropSet (`defaultDroppedBetasSet` + BetaPolicy filter) 应用在主路径, passthrough / Vertex 路径有意不应用 —— 这条原有的不对称行为本 PR 不动。 一条语义测试 (`TestSanitizeMustBeBeforeCCHSigning_HashConsistency`) 把 顺序约束文档化并强制守住: 它证明 `sanitize -> signBillingHeaderCCH` 产生的 `cch` 与最终 body 一致, 而 `signBillingHeaderCCH -> sanitize` 产生的 `cch` 会被上游 hash 重算判失败。 为什么是能力维度 (而不是 haiku 模型名匹配) ========================================== 最朴素的"按 model 名 strip"方案 (`strings.Contains(modelID, "haiku") -> DeleteBytes "context_management"`) 有四个真实失败模式: 1. 过度删除。CLI >= 2.1.87 的真 Claude Code 客户端在 haiku 上同时 发送 body 字段 **和** `anthropic-beta: context-management-2025-06-27`。 一律 strip 会让该用户的 `clear_thinking_20251015` 静默失效。 2. 别名漂移。未来的 haiku 别名 (`claude-3-haiku-...`, `claude-haiku-...` 等) 改变匹配面; 任何新别名都会悄悄绕过 strip。 3. count_tokens 漏覆盖。count_tokens 有自己的 builder 和不同的 beta header 集合; 在一个地方做 model 名检查会漏掉这条路径。 4. API-key passthrough 早退。passthrough builder 在 model 名 strip 之前就 return 了, strip 根本不执行。 能力维度沿着 header 端到端走, 上述 4 个 case 都由构造方式保证正确, 不依赖任何 modelID 匹配。 防御项 ====== - 当 `sjson.DeleteBytes` 在 `gjson` 刚验证过字段存在的 body 上失败时, `sanitizeAnthropicBodyForBetaTokens` 会记 warning 日志 —— 这种情况 现实中仅在请求中途被破坏时发生, 日志把此前会静默发生的 body / header 不一致暴露出来。 - `header_util.go` 新增 `deleteHeaderAllForms`: 在白名单透传已经写入 canonical 大小写的 `Anthropic-Beta` 之后再覆盖, 否则会同时留下两条。 测试 ==== `backend/internal/service` 下新增 44 个测试: - 纯函数: anthropicBetaTokensContains x 5, sanitize keep/strip x 6, computeFinal{Anthropic,CountTokens}AnthropicBeta x 12 - normalize 回归 x 5 - buildUpstreamRequest 端到端 x 4 (OAuth mimic haiku strip / mimic sonnet preserve / 真 CC haiku 带客户端 beta preserve / API-key haiku strip) - buildCountTokensRequest 端到端 x 2 - buildUpstreamRequestAnthropicAPIKeyPassthrough x 2 (strip / preserve) - buildCountTokensRequestAnthropicAPIKeyPassthrough x 2 (strip / preserve) - buildUpstreamRequestAnthropicVertex x 2 (strip / preserve, 含 outgoing `anthropic-beta` header 对称断言) - CCH 顺序语义测试 x 1 unit 套件全过 (本机 88s), `golangci-lint` 0 issues。 已知局限 (本 PR 范围外) ======================== - Vertex 路径用透传过来的客户端 `anthropic-beta` header 作为 sanitize 依据, 而不是 Vertex 侧的能力矩阵。最坏情况是过度 strip (= 当前 main 的行为, 主路径本来什么都不 strip); 不是 regression。完整的 Vertex 能力模型属于单独的 PR。 - Vertex builder 仍然不应用 BetaPolicy filter / dropSet。这是该 builder 早 return 的既有架构决策, 本 PR 不动。 - count_tokens mimic 在 haiku 上仍然注入 `context-management-2025-06-27` (因为原 count_tokens mimic 逻辑并不像 messages mimic 那样排除它)。 本 PR 逐位保留 main 的行为; 是否要让它与 messages mimic 的排除策略 统一是另一个问题。 - `sanitizeAnthropicBodyForBetaTokens` 目前只处理 `context_management <-> context-management-2025-06-27` 这一对。如果 Anthropic 后续推出更多 beta-gated body 字段, 可以在后续 PR 重构为 `{body 路径 -> required beta token}` 注册表的形式。 --- .../service/antigravity_gateway_service.go | 10 +- ...y_anthropic_vertex_service_account_test.go | 64 ++ .../gateway_context_management_test.go | 667 ++++++++++++++++++ backend/internal/service/gateway_request.go | 64 ++ backend/internal/service/gateway_service.go | 279 ++++++-- backend/internal/service/header_util.go | 14 + 6 files changed, 1024 insertions(+), 74 deletions(-) create mode 100644 backend/internal/service/gateway_context_management_test.go diff --git a/backend/internal/service/antigravity_gateway_service.go b/backend/internal/service/antigravity_gateway_service.go index 9882b010..2b849bdd 100644 --- a/backend/internal/service/antigravity_gateway_service.go +++ b/backend/internal/service/antigravity_gateway_service.go @@ -4209,6 +4209,14 @@ func (s *AntigravityGatewayService) ForwardUpstream(ctx context.Context, c *gin. // 构建上游请求 URL upstreamURL := baseURL + "/v1/messages" + // 能力维度 sanitize：Anthropic-compatible 上游透传路径也需要保证 body↔beta header + // 对称。客户端 anthropic-beta header 不含 context-management-2025-06-27 但 body 带 + // context_management 时 strip，与 Anthropic 直连 / Bedrock / Vertex 路径保持一致。 + clientBeta := c.GetHeader("anthropic-beta") + if sanitized, changed := sanitizeAnthropicBodyForBetaTokens(body, clientBeta); changed { + body = sanitized + } + // 创建请求 req, err := http.NewRequestWithContext(ctx, http.MethodPost, upstreamURL, bytes.NewReader(body)) if err != nil { @@ -4224,7 +4232,7 @@ func (s *AntigravityGatewayService) ForwardUpstream(ctx context.Context, c *gin. if v := c.GetHeader("anthropic-version"); v != "" { req.Header.Set("anthropic-version", v) } - if v := c.GetHeader("anthropic-beta"); v != "" { + if v := clientBeta; v != "" { req.Header.Set("anthropic-beta", v) } diff --git a/backend/internal/service/gateway_anthropic_vertex_service_account_test.go b/backend/internal/service/gateway_anthropic_vertex_service_account_test.go index aa779805..2f42b0ab 100644 --- a/backend/internal/service/gateway_anthropic_vertex_service_account_test.go +++ b/backend/internal/service/gateway_anthropic_vertex_service_account_test.go @@ -66,3 +66,67 @@ func readRequestBodyForTest(t *testing.T, req *http.Request) []byte { require.NoError(t, err) return body } + +// Vertex 路径回归保护：同样需要 +// body↔beta header 能力维度对称。客户端 header 不带 context-management beta +// 但 body 带 context_management 字段 → Vertex builder 必须 strip 字段，与 Anthropic +// 直连 / Bedrock 路径保持一致。 +func TestGatewayService_BuildAnthropicVertexServiceAccount_StripsContextManagementWhenBetaMissing(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + // 客户端 header 只带 interleaved-thinking，不带 context-management-2025-06-27 + c.Request.Header.Set("Anthropic-Beta", "interleaved-thinking-2025-05-14") + + account := &Account{ + ID: 302, Platform: PlatformAnthropic, Type: AccountTypeServiceAccount, + Credentials: map[string]any{"project_id": "vertex-proj", "location": "us-east5"}, + } + // body 带了 context_management 字段（客户端透传 / normalize 补齐 / mimicry 注入等场景都可能导致） + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015","keep":"all"}]},"messages":[{"role":"user","content":"hi"}]}`) + + svc := &GatewayService{} + req, err := svc.buildUpstreamRequest( + context.Background(), c, account, body, + "vertex-token", "service_account", "claude-haiku-4-5@20251001", false, false, + ) + require.NoError(t, err) + + got := readRequestBodyForTest(t, req) + require.False(t, gjson.GetBytes(got, "context_management").Exists(), + "Vertex 路径下客户端 header 缺 context-management beta 时，必须 strip body 同名字段") + // header 对称断言：覆盖未来某人在 Vertex builder 里加入与 sanitize 不一致的 header 处理。 + outBeta := getHeaderRaw(req.Header, "anthropic-beta") + require.False(t, anthropicBetaTokensContains(outBeta, "context-management-2025-06-27"), + "与 body 对称：outgoing anthropic-beta header 也不含 context-management beta") +} + +// Vertex 路径反面：客户端 header 含 context-management beta 时保留字段。 +func TestGatewayService_BuildAnthropicVertexServiceAccount_PreservesContextManagementWhenBetaPresent(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + c.Request.Header.Set("Anthropic-Beta", "interleaved-thinking-2025-05-14,context-management-2025-06-27") + + account := &Account{ + ID: 303, Platform: PlatformAnthropic, Type: AccountTypeServiceAccount, + Credentials: map[string]any{"project_id": "vertex-proj", "location": "us-east5"}, + } + body := []byte(`{"model":"claude-sonnet-4-6","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + + svc := &GatewayService{} + req, err := svc.buildUpstreamRequest( + context.Background(), c, account, body, + "vertex-token", "service_account", "claude-sonnet-4-6@20260218", false, false, + ) + require.NoError(t, err) + + got := readRequestBodyForTest(t, req) + require.True(t, gjson.GetBytes(got, "context_management").Exists(), + "Vertex + 客户端 header 包含 context-management beta 时字段必须保留") + outBeta := getHeaderRaw(req.Header, "anthropic-beta") + require.True(t, anthropicBetaTokensContains(outBeta, "context-management-2025-06-27"), + "与 body 对称：outgoing anthropic-beta header 同步含 context-management beta") +} diff --git a/backend/internal/service/gateway_context_management_test.go b/backend/internal/service/gateway_context_management_test.go new file mode 100644 index 00000000..c2263bdc --- /dev/null +++ b/backend/internal/service/gateway_context_management_test.go @@ -0,0 +1,667 @@ +//go:build unit + +package service + +import ( + "context" + "io" + "net/http" + "net/http/httptest" + "regexp" + "strings" + "testing" + + "github.com/Wei-Shaw/sub2api/internal/config" + "github.com/Wei-Shaw/sub2api/internal/pkg/claude" + "github.com/gin-gonic/gin" + "github.com/stretchr/testify/require" + "github.com/tidwall/gjson" +) + +// ============================================================================ +// 背景 +// ============================================================================ +// +// Anthropic 上游对 body.context_management 字段实施 Pydantic schema 校验： +// 当且仅当 anthropic-beta header 含 context-management-2025-06-27 时接受。 +// 否则报： +// "context_management: Extra inputs are not permitted" +// +// 本仓采用能力维度对称约束（与 Bedrock 路径的 sanitizeBedrockFieldsForBetaTokens +// 对称）：在所有 Anthropic 直连出口，按最终 anthropic-beta header 是否含上述 token +// 决定 body 是否保留同名字段。 +// +// 本文件覆盖： +// 1) sanitizeAnthropicBodyForBetaTokens 纯函数 +// 2) anthropicBetaTokensContains 解析辅助函数 +// 3) computeFinalAnthropicBeta / computeFinalCountTokensAnthropicBeta 各路径 +// 4) normalizeClaudeOAuthRequestBody 的 context_management 补齐行为（不再按 model 短路） + +// ============================================================================ +// anthropicBetaTokensContains +// ============================================================================ + +func TestAnthropicBetaTokensContains_EmptyInputs(t *testing.T) { + require.False(t, anthropicBetaTokensContains("", "context-management-2025-06-27")) + require.False(t, anthropicBetaTokensContains("oauth-2025-04-20", "")) +} + +func TestAnthropicBetaTokensContains_SingleToken(t *testing.T) { + require.True(t, anthropicBetaTokensContains("context-management-2025-06-27", "context-management-2025-06-27")) +} + +func TestAnthropicBetaTokensContains_MultiTokenComma(t *testing.T) { + header := "oauth-2025-04-20,context-management-2025-06-27,interleaved-thinking-2025-05-14" + require.True(t, anthropicBetaTokensContains(header, "context-management-2025-06-27")) + require.True(t, anthropicBetaTokensContains(header, "oauth-2025-04-20")) + require.False(t, anthropicBetaTokensContains(header, "fast-mode-2026-02-01")) +} + +func TestAnthropicBetaTokensContains_ToleratesWhitespace(t *testing.T) { + header := "oauth-2025-04-20 , context-management-2025-06-27 , interleaved-thinking-2025-05-14" + require.True(t, anthropicBetaTokensContains(header, "context-management-2025-06-27")) +} + +func TestAnthropicBetaTokensContains_SubstringNotMatched(t *testing.T) { + // 严格 token 比较，不应被子串误匹配 + require.False(t, anthropicBetaTokensContains("context-management-2025-06-27-rev2", "context-management-2025-06-27"), + "必须按 token 边界匹配，不允许 prefix 子串误命中") +} + +// ============================================================================ +// sanitizeAnthropicBodyForBetaTokens +// ============================================================================ + +func TestSanitizeAnthropicBodyForBetaTokens_NoFieldNoChange(t *testing.T) { + body := []byte(`{"model":"claude-haiku-4-5","messages":[]}`) + out, changed := sanitizeAnthropicBodyForBetaTokens(body, "oauth-2025-04-20") + require.False(t, changed) + require.Equal(t, string(body), string(out)) +} + +func TestSanitizeAnthropicBodyForBetaTokens_FieldKeptWhenBetaPresent(t *testing.T) { + body := []byte(`{"model":"claude-opus-4-7","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + out, changed := sanitizeAnthropicBodyForBetaTokens(body, + "oauth-2025-04-20,context-management-2025-06-27,interleaved-thinking-2025-05-14") + require.False(t, changed) + require.True(t, gjson.GetBytes(out, "context_management").Exists()) + require.Equal(t, "clear_thinking_20251015", + gjson.GetBytes(out, "context_management.edits.0.type").String()) +} + +func TestSanitizeAnthropicBodyForBetaTokens_FieldStrippedWhenBetaMissing(t *testing.T) { + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + out, changed := sanitizeAnthropicBodyForBetaTokens(body, "oauth-2025-04-20,interleaved-thinking-2025-05-14") + require.True(t, changed) + require.False(t, gjson.GetBytes(out, "context_management").Exists(), + "header 不含 context-management beta 时必须 strip 同名字段") +} + +func TestSanitizeAnthropicBodyForBetaTokens_FieldStrippedWhenBetaEmpty(t *testing.T) { + body := []byte(`{"context_management":{"edits":[]},"messages":[]}`) + out, changed := sanitizeAnthropicBodyForBetaTokens(body, "") + require.True(t, changed) + require.False(t, gjson.GetBytes(out, "context_management").Exists()) +} + +func TestSanitizeAnthropicBodyForBetaTokens_EmptyBody(t *testing.T) { + out, changed := sanitizeAnthropicBodyForBetaTokens([]byte{}, "") + require.False(t, changed) + require.Empty(t, out) + + out, changed = sanitizeAnthropicBodyForBetaTokens(nil, "") + require.False(t, changed) + require.Empty(t, out) +} + +// ★ 关键回归断言：能力维度 sanitize 解决了 "真 CC + haiku" 路径的过度删除问题。 +// 真实 Claude Code CLI 2.1.87+ 客户端 header 含 context-management beta； +// 即使 model 是 haiku，sanitize 也不应剥离功能字段。 +func TestSanitizeAnthropicBodyForBetaTokens_HaikuRealCCClientPreservesField(t *testing.T) { + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015","keep":"all"}]},"messages":[]}`) + // 真 Claude Code CLI 2.1.87+ 客户端 header 含 context-management beta + clientBeta := "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,context-management-2025-06-27" + out, changed := sanitizeAnthropicBodyForBetaTokens(body, clientBeta) + require.False(t, changed, + "真 CC 客户端 header 含 context-management beta 时，haiku body 字段必须保留（功能不丢）") + require.True(t, gjson.GetBytes(out, "context_management").Exists()) +} + +// ============================================================================ +// computeFinalAnthropicBeta — 关键路径 +// ============================================================================ + +func newTestGatewayServiceForBeta(injectBetaForAPIKey bool) *GatewayService { + cfg := &config.Config{} + cfg.Gateway.InjectBetaForAPIKey = injectBetaForAPIKey + return &GatewayService{cfg: cfg} +} + +func TestComputeFinalAnthropicBeta_OAuthMimic_NonHaiku_IncludesContextManagement(t *testing.T) { + s := newTestGatewayServiceForBeta(false) + final, ok := s.computeFinalAnthropicBeta("oauth", true, "claude-sonnet-4-6", http.Header{}, []byte(`{}`), nil) + require.True(t, ok) + require.True(t, anthropicBetaTokensContains(final, claude.BetaContextManagement), + "OAuth mimic non-haiku 必须注入完整 CC mimicry beta，含 context-management-2025-06-27") + require.True(t, anthropicBetaTokensContains(final, claude.BetaOAuth)) + require.True(t, anthropicBetaTokensContains(final, claude.BetaClaudeCode)) +} + +func TestComputeFinalAnthropicBeta_OAuthMimic_Haiku_ExcludesContextManagement(t *testing.T) { + s := newTestGatewayServiceForBeta(false) + final, ok := s.computeFinalAnthropicBeta("oauth", true, "claude-haiku-4-5", http.Header{}, []byte(`{}`), nil) + require.True(t, ok) + require.False(t, anthropicBetaTokensContains(final, claude.BetaContextManagement), + "OAuth mimic haiku 仅注入 oauth + interleaved-thinking，不含 context-management") + require.True(t, anthropicBetaTokensContains(final, claude.BetaOAuth)) + require.True(t, anthropicBetaTokensContains(final, claude.BetaInterleavedThinking)) +} + +func TestComputeFinalAnthropicBeta_OAuthMimic_IgnoresClientBeta(t *testing.T) { + // mimic 路径下原代码白名单透传被跳过，client beta 应被忽略 + s := newTestGatewayServiceForBeta(false) + hdr := http.Header{} + hdr.Set("anthropic-beta", "custom-experimental-beta") + final, ok := s.computeFinalAnthropicBeta("oauth", true, "claude-sonnet-4-6", hdr, []byte(`{}`), nil) + require.True(t, ok) + require.False(t, strings.Contains(final, "custom-experimental-beta"), + "mimic 路径必须忽略客户端 anthropic-beta header") +} + +func TestComputeFinalAnthropicBeta_OAuthTransparent_NonHaiku_PreservesClientContextManagement(t *testing.T) { + // 真 CC 客户端透传：客户端 header 中的 context-management beta 必须保留 + s := newTestGatewayServiceForBeta(false) + hdr := http.Header{} + hdr.Set("anthropic-beta", "claude-code-20250219,oauth-2025-04-20,context-management-2025-06-27") + final, ok := s.computeFinalAnthropicBeta("oauth", false, "claude-sonnet-4-6", hdr, []byte(`{}`), nil) + require.True(t, ok) + require.True(t, anthropicBetaTokensContains(final, claude.BetaContextManagement)) +} + +func TestComputeFinalAnthropicBeta_OAuthTransparent_Haiku_RealCCPreservesContextManagement(t *testing.T) { + // haiku 透传 + 客户端带 context-management beta → 必须保留 + // （能力维度核心场景：避免 model-name 误删客户端透传的功能 beta） + s := newTestGatewayServiceForBeta(false) + hdr := http.Header{} + hdr.Set("anthropic-beta", "claude-code-20250219,oauth-2025-04-20,context-management-2025-06-27,interleaved-thinking-2025-05-14") + final, ok := s.computeFinalAnthropicBeta("oauth", false, "claude-haiku-4-5", hdr, []byte(`{}`), nil) + require.True(t, ok) + require.True(t, anthropicBetaTokensContains(final, claude.BetaContextManagement), + "真 CC + haiku + 客户端带 context-management beta → 透传必须保留") +} + +func TestComputeFinalAnthropicBeta_APIKey_PassesClientBetaThroughDropSet(t *testing.T) { + s := newTestGatewayServiceForBeta(false) + hdr := http.Header{} + hdr.Set("anthropic-beta", "oauth-2025-04-20,custom-beta") + final, ok := s.computeFinalAnthropicBeta("apikey", false, "claude-sonnet-4-6", hdr, []byte(`{}`), nil) + require.True(t, ok) + require.True(t, anthropicBetaTokensContains(final, "oauth-2025-04-20")) + require.True(t, anthropicBetaTokensContains(final, "custom-beta")) +} + +func TestComputeFinalAnthropicBeta_APIKey_NoClientBetaInjectOff_ShouldNotSet(t *testing.T) { + s := newTestGatewayServiceForBeta(false) + final, ok := s.computeFinalAnthropicBeta("apikey", false, "claude-sonnet-4-6", http.Header{}, []byte(`{}`), nil) + require.False(t, ok, "API-key + 客户端未传 + InjectBetaForAPIKey 关 → 不应主动设置 anthropic-beta") + require.Equal(t, "", final) +} + +// ============================================================================ +// computeFinalCountTokensAnthropicBeta +// ============================================================================ + +func TestComputeFinalCountTokensAnthropicBeta_OAuthMimic_AlwaysIncludesContextManagement(t *testing.T) { + // count_tokens 路径下 mimic 不按 haiku 排除：始终注入完整 mimicry beta + s := newTestGatewayServiceForBeta(false) + final, ok := s.computeFinalCountTokensAnthropicBeta("oauth", true, "claude-haiku-4-5", http.Header{}, []byte(`{}`), nil) + require.True(t, ok) + require.True(t, anthropicBetaTokensContains(final, claude.BetaContextManagement), + "count_tokens + mimic 即使 haiku 也注入 context-management beta（与 messages 不同）") + require.True(t, anthropicBetaTokensContains(final, claude.BetaTokenCounting), + "count_tokens 路径必须含 token-counting beta") +} + +// 重构等价性回归： +// 原 main buildCountTokensRequest 在 count_tokens mimic 分支上不跳过白名单透传 +// （与 messages mimic 不同），incomingBeta 取自客户端透传。重构后必须从 clientHeaders +// 拿同一个值并 merge，否则会丢失客户端 beta。 +func TestComputeFinalCountTokensAnthropicBeta_OAuthMimic_PreservesClientBeta(t *testing.T) { + s := newTestGatewayServiceForBeta(false) + hdr := http.Header{} + hdr.Set("anthropic-beta", "custom-experimental-beta,context-1m-2025-08-07") + final, ok := s.computeFinalCountTokensAnthropicBeta("oauth", true, "claude-haiku-4-5", hdr, []byte(`{}`), nil) + require.True(t, ok) + require.True(t, anthropicBetaTokensContains(final, "custom-experimental-beta"), + "count_tokens mimic 不同于 messages mimic：原代码会保留客户端透传的 beta") + require.True(t, anthropicBetaTokensContains(final, "context-1m-2025-08-07"), + "客户端透传的其他 beta token 同样需要保留") + require.True(t, anthropicBetaTokensContains(final, claude.BetaContextManagement), + "同时 FullClaudeCodeMimicryBetas 不打折扣") + require.True(t, anthropicBetaTokensContains(final, claude.BetaTokenCounting), + "同时补齐 token-counting beta") +} + +// messages mimic 路径反向验证：原代码会跳过白名单透传， +// 客户端 beta 不会进入 mimic 计算。重构后 messages computeFinalAnthropicBeta +// mimic 分支依然不该使用 clientBeta。 +func TestComputeFinalAnthropicBeta_OAuthMimic_IgnoresClientBetaExplicit(t *testing.T) { + s := newTestGatewayServiceForBeta(false) + hdr := http.Header{} + hdr.Set("anthropic-beta", "custom-experimental-beta") + final, ok := s.computeFinalAnthropicBeta("oauth", true, "claude-sonnet-4-6", hdr, []byte(`{}`), nil) + require.True(t, ok) + require.False(t, anthropicBetaTokensContains(final, "custom-experimental-beta"), + "messages mimic 原代码跳过白名单透传 → 客户端 beta 不进入计算。"+ + "与 count_tokens mimic 是不同的设计，不能合并为同一函数。") +} + +func TestComputeFinalCountTokensAnthropicBeta_OAuthTransparent_NoClientBetaInjectsDefault(t *testing.T) { + // 真 CC 客户端透传 + 客户端未传 anthropic-beta → 用 CountTokensBetaHeader 兜底 + s := newTestGatewayServiceForBeta(false) + final, ok := s.computeFinalCountTokensAnthropicBeta("oauth", false, "claude-haiku-4-5", http.Header{}, []byte(`{}`), nil) + require.True(t, ok) + require.Equal(t, claude.CountTokensBetaHeader, final) + // CountTokensBetaHeader 不含 context-management beta + require.False(t, anthropicBetaTokensContains(final, claude.BetaContextManagement)) +} + +func TestComputeFinalCountTokensAnthropicBeta_OAuthTransparent_AppendsBetaTokenCounting(t *testing.T) { + s := newTestGatewayServiceForBeta(false) + hdr := http.Header{} + hdr.Set("anthropic-beta", "oauth-2025-04-20,context-management-2025-06-27") + final, ok := s.computeFinalCountTokensAnthropicBeta("oauth", false, "claude-sonnet-4-6", hdr, []byte(`{}`), nil) + require.True(t, ok) + require.True(t, anthropicBetaTokensContains(final, claude.BetaTokenCounting), + "客户端未带 token-counting beta 时必须补齐") + require.True(t, anthropicBetaTokensContains(final, claude.BetaContextManagement), + "客户端带的 context-management beta 必须保留") +} + +// ============================================================================ +// normalizeClaudeOAuthRequestBody — 回归：context_management 补齐恢复原行为 +// ============================================================================ +// +// 重构后该函数不再按 model 名短路：thinking=enabled/adaptive 时补齐 context_management， +// 与 model 无关。strip 责任移交 sanitizeAnthropicBodyForBetaTokens（在 +// buildUpstreamRequest 层按最终 beta header 执行）。 + +func TestNormalizeClaudeOAuthRequestBody_InjectsContextManagement_ThinkingEnabled(t *testing.T) { + body := []byte(`{"model":"claude-sonnet-4-6","thinking":{"type":"enabled","budget_tokens":1000},"messages":[]}`) + out, _ := normalizeClaudeOAuthRequestBody(body, "claude-sonnet-4-6", claudeOAuthNormalizeOptions{}) + require.True(t, gjson.GetBytes(out, "context_management").Exists()) + require.Equal(t, "clear_thinking_20251015", + gjson.GetBytes(out, "context_management.edits.0.type").String()) +} + +func TestNormalizeClaudeOAuthRequestBody_InjectsContextManagement_ThinkingAdaptive(t *testing.T) { + body := []byte(`{"model":"claude-opus-4-7","thinking":{"type":"adaptive"},"messages":[]}`) + out, _ := normalizeClaudeOAuthRequestBody(body, "claude-opus-4-7", claudeOAuthNormalizeOptions{}) + require.True(t, gjson.GetBytes(out, "context_management").Exists()) +} + +func TestNormalizeClaudeOAuthRequestBody_HaikuStillInjects_StripDeferredToSanitize(t *testing.T) { + // haiku + thinking=enabled：normalize 阶段仍按 CLI mimicry 行为补齐字段； + // strip 由 buildUpstreamRequest 层的 sanitize 兜底（如果 final beta 不含 token）。 + body := []byte(`{"model":"claude-haiku-4-5","thinking":{"type":"enabled","budget_tokens":1000},"messages":[]}`) + out, _ := normalizeClaudeOAuthRequestBody(body, "claude-haiku-4-5", claudeOAuthNormalizeOptions{}) + require.True(t, gjson.GetBytes(out, "context_management").Exists(), + "normalize 不再按 model 名短路；strip 责任移交 sanitize 层") +} + +func TestNormalizeClaudeOAuthRequestBody_PreservesClientContextManagement(t *testing.T) { + body := []byte(`{"model":"claude-opus-4-7","context_management":{"edits":[{"type":"custom_strategy"}]},"thinking":{"type":"enabled","budget_tokens":1000},"messages":[]}`) + out, _ := normalizeClaudeOAuthRequestBody(body, "claude-opus-4-7", claudeOAuthNormalizeOptions{}) + require.Equal(t, "custom_strategy", + gjson.GetBytes(out, "context_management.edits.0.type").String(), + "客户端透传的 context_management 内容必须原样保留") +} + +func TestNormalizeClaudeOAuthRequestBody_NoThinking_NoInject(t *testing.T) { + body := []byte(`{"model":"claude-sonnet-4-6","messages":[]}`) + out, _ := normalizeClaudeOAuthRequestBody(body, "claude-sonnet-4-6", claudeOAuthNormalizeOptions{}) + require.False(t, gjson.GetBytes(out, "context_management").Exists()) +} + +// ============================================================================ +// passthrough 集成测试：buildUpstreamRequest- +// AnthropicAPIKeyPassthrough 与 buildCountTokensRequestAnthropicAPIKeyPassthrough +// 路径上 sanitize 是否生效。 +// ============================================================================ + +// passthrough 集成测试不设 base_url，避开 validateUpstreamBaseURL 对 cfg.Security 的依赖。 +// targetURL 会走默认 claudeAPIURL，sanitize 逻辑与 baseURL 是否存在无关。 +func newAnthropicAPIKeyPassthroughAccountForBetaTest() *Account { + return &Account{ + ID: 501, + Name: "anthropic-apikey-passthrough-ctxmgmt-test", + Platform: PlatformAnthropic, + Type: AccountTypeAPIKey, + Credentials: map[string]any{ + "api_key": "upstream-key", + }, + Extra: map[string]any{"anthropic_passthrough": true}, + Status: StatusActive, + Schedulable: true, + } +} + +func readUpstreamBodyForTest(t *testing.T, req *http.Request) []byte { + t.Helper() + require.NotNil(t, req.Body) + b, err := io.ReadAll(req.Body) + require.NoError(t, err) + return b +} + +func TestBuildUpstreamRequestAnthropicAPIKeyPassthrough_StripsContextManagementWhenClientHeaderMissingBeta(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + // 客户端仅带 oauth beta，不带 context-management-2025-06-27 + c.Request.Header.Set("Anthropic-Beta", "oauth-2025-04-20") + + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildUpstreamRequestAnthropicAPIKeyPassthrough( + context.Background(), c, newAnthropicAPIKeyPassthroughAccountForBetaTest(), body, "token", + ) + require.NoError(t, err) + require.False(t, gjson.GetBytes(readUpstreamBodyForTest(t, req), "context_management").Exists(), + "API-key passthrough + 客户端未带 context-management beta → strip body 字段") +} + +func TestBuildUpstreamRequestAnthropicAPIKeyPassthrough_PreservesContextManagementWhenClientHeaderHasBeta(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + c.Request.Header.Set("Anthropic-Beta", "oauth-2025-04-20,context-management-2025-06-27") + + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildUpstreamRequestAnthropicAPIKeyPassthrough( + context.Background(), c, newAnthropicAPIKeyPassthroughAccountForBetaTest(), body, "token", + ) + require.NoError(t, err) + require.True(t, gjson.GetBytes(readUpstreamBodyForTest(t, req), "context_management").Exists(), + "API-key passthrough + 客户端带 context-management beta → 字段保留（不过度删除）") +} + +func TestBuildCountTokensRequestAnthropicAPIKeyPassthrough_StripsContextManagementWhenClientHeaderMissingBeta(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages/count_tokens", nil) + c.Request.Header.Set("Anthropic-Beta", "oauth-2025-04-20,token-counting-2024-11-01") + + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildCountTokensRequestAnthropicAPIKeyPassthrough( + context.Background(), c, newAnthropicAPIKeyPassthroughAccountForBetaTest(), body, "token", + ) + require.NoError(t, err) + require.False(t, gjson.GetBytes(readUpstreamBodyForTest(t, req), "context_management").Exists(), + "count_tokens passthrough + 客户端未带 context-management beta → strip") +} + +// ============================================================================ +// 集成测试：buildUpstreamRequest +// 全路径验证上游 outgoing body 与 anthropic-beta header 严格对称。 +// 这个测试能挡住未来某人忘调 sanitize / 将 sanitize 挪到 CCH 之后 等 regression。 +// ============================================================================ + +func TestBuildUpstreamRequest_OAuthMimicHaiku_StripsContextManagementEndToEnd(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + + account := &Account{ID: 401, Platform: PlatformAnthropic, Type: AccountTypeOAuth, + Credentials: map[string]any{"access_token": "oauth-tok"}, + Status: StatusActive, + Schedulable: true, + } + // haiku + mimic CC → final beta = HaikuBetaHeader（不含 context-management）→ + // body 必须 strip。 + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildUpstreamRequest( + context.Background(), c, account, body, + "oauth-tok", "oauth", "claude-haiku-4-5", false, true, // mimicClaudeCode=true + ) + require.NoError(t, err) + + outBody := readUpstreamBodyForTest(t, req) + outBeta := getHeaderRaw(req.Header, "anthropic-beta") + + require.False(t, gjson.GetBytes(outBody, "context_management").Exists(), + "OAuth mimic + haiku 端到端：outgoing body 不应含 context_management") + require.False(t, anthropicBetaTokensContains(outBeta, claude.BetaContextManagement), + "对称约束：outgoing anthropic-beta header 也不带 context-management beta") +} + +func TestBuildUpstreamRequest_OAuthMimicNonHaiku_PreservesContextManagementEndToEnd(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + + account := &Account{ID: 402, Platform: PlatformAnthropic, Type: AccountTypeOAuth, + Credentials: map[string]any{"access_token": "oauth-tok"}, + Status: StatusActive, + Schedulable: true, + } + // sonnet + mimic CC → final beta = FullClaudeCodeMimicryBetas（含 context-management）→ + // body 保留。 + body := []byte(`{"model":"claude-sonnet-4-6","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildUpstreamRequest( + context.Background(), c, account, body, + "oauth-tok", "oauth", "claude-sonnet-4-6", false, true, + ) + require.NoError(t, err) + + outBody := readUpstreamBodyForTest(t, req) + outBeta := getHeaderRaw(req.Header, "anthropic-beta") + + require.True(t, gjson.GetBytes(outBody, "context_management").Exists(), + "OAuth mimic + non-haiku：outgoing body 必须保留 context_management。") + require.True(t, anthropicBetaTokensContains(outBeta, claude.BetaContextManagement), + "对称约束：outgoing anthropic-beta header 同时含 context-management beta") +} + +func TestBuildUpstreamRequest_OAuthTransparentHaikuWithRealCCBeta_PreservesField(t *testing.T) { + // 端到端验证：真 CC 客户端 + haiku + 客户端 header 带 context-management beta + // → final beta 透传 → 不应该过度删除 body 字段 + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + c.Request.Header.Set("Anthropic-Beta", + "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,context-management-2025-06-27") + + account := &Account{ID: 403, Platform: PlatformAnthropic, Type: AccountTypeOAuth, + Credentials: map[string]any{"access_token": "oauth-tok"}, + Status: StatusActive, Schedulable: true, + } + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015","keep":"all"}]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildUpstreamRequest( + context.Background(), c, account, body, + "oauth-tok", "oauth", "claude-haiku-4-5", false, false, // mimicClaudeCode=false（真 CC） + ) + require.NoError(t, err) + + outBody := readUpstreamBodyForTest(t, req) + outBeta := getHeaderRaw(req.Header, "anthropic-beta") + + require.True(t, anthropicBetaTokensContains(outBeta, claude.BetaContextManagement), + "真 CC 透传路径：客户端 header 中的 context-management beta 必须保留") + require.True(t, gjson.GetBytes(outBody, "context_management").Exists(), + "回归保护：真 CC + haiku + 客户端带 beta token 时，clear_thinking_20251015 功能不能静默失效") +} + +// CCH 顺序语义测试：sanitize 必须在 signBillingHeaderCCH 之前， +// 否则签名的 hash 与最终发送的 body 不一致，被 Anthropic 判 third-party。 +// +// 该测试不走 buildUpstreamRequest 完整路径（需要 mock SettingService 成本高）， +// 而是直接验证两个顺序产生的 cch 不同，证明二者不可交换。 +// 测试名本身是语义约束的文档化 marker。 +func TestSanitizeMustBeBeforeCCHSigning_HashConsistency(t *testing.T) { + // 构造 body：含 context_management + cch=00000 占位符 + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"system":[{"type":"text","text":"x-anthropic-billing-header: cc_version=2.1.92; cch=00000;"}],"messages":[]}`) + + // 最终发送场景：final beta 不含 context-management beta → sanitize 会 strip + finalBeta := "oauth-2025-04-20,interleaved-thinking-2025-05-14" + + extractCCH := func(t *testing.T, b []byte) string { + t.Helper() + m := regexp.MustCompile(`\bcch=([0-9a-fA-F]{5})\b`).FindSubmatch(b) + require.NotNil(t, m, "body 里找不到 cch=<5hex> ：%s", string(b)) + return string(m[1]) + } + + // === 正确顺序：sanitize → signBillingHeaderCCH === + // 1. strip context_management + sanitizedFirst, changed := sanitizeAnthropicBodyForBetaTokens(body, finalBeta) + require.True(t, changed) + require.False(t, gjson.GetBytes(sanitizedFirst, "context_management").Exists()) + // 2. 基于“strip 后的 body”算 hash + correctFinal := signBillingHeaderCCH(sanitizedFirst) + correctCCH := extractCCH(t, correctFinal) + require.NotEqual(t, "00000", correctCCH, "placeholder 应被替换") + + // === 错误顺序：signBillingHeaderCCH → sanitize（未来 regression 场景）=== + // 1. 先基于“含 context_management 的 body”算 hash → cch=H_with + signedFirst := signBillingHeaderCCH(body) + wrongCCH := extractCCH(t, signedFirst) + require.NotEqual(t, "00000", wrongCCH) + // 2. 后 strip context_management → body 变化但 cch 仍是 H_with + wrongFinal, _ := sanitizeAnthropicBodyForBetaTokens(signedFirst, finalBeta) + wrongFinalCCH := extractCCH(t, wrongFinal) + + // === 关键断言 === + // 上游验证逻辑：将 outgoing body 的 cch 还原为 00000、重算 hash、与 cch 字段比较。 + // 模拟上游验证：用发送 body 算出“期望的 cch”，与发送 body 里的 cch 字段比。 + recomputeExpected := func(b []byte, currentCCH string) string { + t.Helper() + // 把 cch= 还原为 cch=00000 + re := regexp.MustCompile(`(\bcch=)` + currentCCH + `(\b)`) + restored := re.ReplaceAll(b, []byte("${1}00000${2}")) + return extractCCH(t, signBillingHeaderCCH(restored)) + } + + // 正确顺序：发送 body 的 cch == 重算 hash → 上游验证过 + require.Equal(t, correctCCH, recomputeExpected(correctFinal, correctCCH), + "正确顺序：final body 里的 cch 与重算 hash 一致 → 上游验证通过") + + // 错误顺序：发送 body 的 cch 是“含 ctx 算的”，但最终 body 不含 ctx → 重算 hash 不同 + require.NotEqual(t, wrongFinalCCH, recomputeExpected(wrongFinal, wrongFinalCCH), + "错误顺序：final body 里的 cch 是基于含 ctx 的 body 算的，"+ + "但发送 body 已 strip ctx → 上游重算 hash 与 cch 不一致 → 被判 third-party。"+ + "这是 buildUpstreamRequest / buildCountTokensRequest 里 sanitize 必须在 "+ + "signBillingHeaderCCH 之前的原因。") +} + +// count_tokens 主路径 E2E 集成测试 +func TestBuildCountTokensRequest_OAuthMimicHaiku_PreservesContextManagementEndToEnd(t *testing.T) { + // count_tokens 路径下 mimic 不按 haiku 排除，始终注入 BetaContextManagement + // → sanitize 看到最终 beta header 含 context-management beta → 字段保留。 + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages/count_tokens", nil) + + account := &Account{ID: 411, Platform: PlatformAnthropic, Type: AccountTypeOAuth, + Credentials: map[string]any{"access_token": "oauth-tok"}, + Status: StatusActive, Schedulable: true, + } + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildCountTokensRequest( + context.Background(), c, account, body, + "oauth-tok", "oauth", "claude-haiku-4-5", true, // mimicClaudeCode=true + ) + require.NoError(t, err) + + outBody := readUpstreamBodyForTest(t, req) + outBeta := getHeaderRaw(req.Header, "anthropic-beta") + + require.True(t, anthropicBetaTokensContains(outBeta, claude.BetaContextManagement), + "count_tokens mimic 始终注入 context-management beta") + require.True(t, gjson.GetBytes(outBody, "context_management").Exists(), + "对称约束：final beta 含 token 时 body 字段保留") + require.True(t, anthropicBetaTokensContains(outBeta, claude.BetaTokenCounting), + "count_tokens 路径必须含 token-counting beta") +} + +func TestBuildCountTokensRequest_APIKeyHaiku_StripsContextManagementEndToEnd(t *testing.T) { + // API-key + haiku + 客户端 header 不带 context-management beta → final beta 不含 → strip + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages/count_tokens", nil) + c.Request.Header.Set("Anthropic-Beta", "interleaved-thinking-2025-05-14") + + account := &Account{ID: 412, Platform: PlatformAnthropic, Type: AccountTypeAPIKey, + Credentials: map[string]any{"api_key": "sk-ant-xxx"}, + Status: StatusActive, Schedulable: true, + } + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildCountTokensRequest( + context.Background(), c, account, body, + "sk-ant-xxx", "apikey", "claude-haiku-4-5", false, + ) + require.NoError(t, err) + + outBody := readUpstreamBodyForTest(t, req) + require.False(t, gjson.GetBytes(outBody, "context_management").Exists(), + "count_tokens API-key + 客户端未带 beta token → body strip") +} + +// count_tokens passthrough preserve 测试 +func TestBuildCountTokensRequestAnthropicAPIKeyPassthrough_PreservesContextManagementWhenClientHeaderHasBeta(t *testing.T) { + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages/count_tokens", nil) + c.Request.Header.Set("Anthropic-Beta", "oauth-2025-04-20,context-management-2025-06-27,token-counting-2024-11-01") + + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[{"type":"clear_thinking_20251015"}]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildCountTokensRequestAnthropicAPIKeyPassthrough( + context.Background(), c, newAnthropicAPIKeyPassthroughAccountForBetaTest(), body, "token", + ) + require.NoError(t, err) + require.True(t, gjson.GetBytes(readUpstreamBodyForTest(t, req), "context_management").Exists(), + "count_tokens passthrough + 客户端带 context-management beta → 字段保留") +} + +func TestBuildUpstreamRequest_APIKeyHaikuWithContextManagement_StripsField(t *testing.T) { + // API-key + haiku + body 带 context_management + 客户端 header 未带 context-management beta + // → final beta 不含 → body 字段被 strip + gin.SetMode(gin.TestMode) + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages", nil) + c.Request.Header.Set("Anthropic-Beta", "interleaved-thinking-2025-05-14") + + account := &Account{ID: 404, Platform: PlatformAnthropic, Type: AccountTypeAPIKey, + Credentials: map[string]any{"api_key": "sk-ant-xxx"}, + Status: StatusActive, Schedulable: true, + } + body := []byte(`{"model":"claude-haiku-4-5","context_management":{"edits":[]},"messages":[]}`) + svc := &GatewayService{cfg: &config.Config{}} + req, err := svc.buildUpstreamRequest( + context.Background(), c, account, body, + "sk-ant-xxx", "apikey", "claude-haiku-4-5", false, false, + ) + require.NoError(t, err) + + outBody := readUpstreamBodyForTest(t, req) + require.False(t, gjson.GetBytes(outBody, "context_management").Exists(), + "API-key + haiku + 客户端未带 beta token → body 字段必须被 strip") +} diff --git a/backend/internal/service/gateway_request.go b/backend/internal/service/gateway_request.go index 498336a4..91f7601c 100644 --- a/backend/internal/service/gateway_request.go +++ b/backend/internal/service/gateway_request.go @@ -12,6 +12,7 @@ import ( "github.com/Wei-Shaw/sub2api/internal/domain" "github.com/Wei-Shaw/sub2api/internal/pkg/antigravity" + "github.com/Wei-Shaw/sub2api/internal/pkg/logger" "github.com/tidwall/gjson" "github.com/tidwall/sjson" ) @@ -665,6 +666,69 @@ func removeThinkingDependentContextStrategies(body []byte) []byte { return body } +// anthropicBetaContextManagementToken 是 context_management 字段受的 beta token。 +// 与 claude.BetaContextManagement 保持一致；在本文件本地定义以避免震荡 +// claude package 的该常量含义。 +const anthropicBetaContextManagementToken = "context-management-2025-06-27" + +// sanitizeAnthropicBodyForBetaTokens 是对 Anthropic 直连路径上 body↔beta header +// **能力维度**对称约束的统一实现，与 Bedrock 路径的 +// `sanitizeBedrockFieldsForBetaTokens` 对称。 +// +// 问题场景： +// - context_management 是 Claude Code CLI 2.1.87+ 默认携带的 beta 字段 +// （含 clear_thinking_20251015 等清理策略） +// - 其被 Anthropic 上游接受的前提是 anthropic-beta header 含 +// `context-management-2025-06-27` +// - 若两侧不一致上游 Pydantic schema 拒收： +// "context_management: Extra inputs are not permitted" +// +// 本函数按最终发送的 anthropic-beta header 决定是否保留 body 中的 +// context_management 字段：缺 beta token → strip。这将限制完全建立在 +// "能力维度" 上，与 model 名 / token type / mimicry 子路径无关。 +// +// 调用约束：必须在 CCH 签名之前调用，否则签名 hash 与最终 body +// 不一致，上游会以 third-party 拒收。 +// +// 返回 (sanitized, changed)：changed 表示是否发生实际删除，供调用方决定 +// 是否重用原 body 引用。 +func sanitizeAnthropicBodyForBetaTokens(body []byte, anthropicBetaHeader string) ([]byte, bool) { + if len(body) == 0 { + return body, false + } + if !gjson.GetBytes(body, "context_management").Exists() { + return body, false + } + if anthropicBetaTokensContains(anthropicBetaHeader, anthropicBetaContextManagementToken) { + return body, false + } + if b, err := sjson.DeleteBytes(body, "context_management"); err == nil { + return b, true + } else { + // 不应发生：gjson 刚验证过字段存在 + body 是合法 JSON。如果 sjson 仍报错， + // 调用方会拿到 (body, false)，但此前 computeFinalAnthropicBeta 已按“strip 后” + // 计算了 finalBeta——两侧会不一致。记录 warning 最小限度提醒运维。 + logger.LegacyPrintf("service.gateway", + "[CtxMgmtSanitize] sjson.DeleteBytes failed unexpectedly: %v (body len=%d). "+ + "body and final anthropic-beta header may be out of sync.", err, len(body)) + } + return body, false +} + +// anthropicBetaTokensContains 检测逗号分隔的 anthropic-beta header 是否含指定 token。 +// 宋体空格宽容；区分大小写（Anthropic beta token 始终是小写）。 +func anthropicBetaTokensContains(header, token string) bool { + if header == "" || token == "" { + return false + } + for _, part := range strings.Split(header, ",") { + if strings.TrimSpace(part) == token { + return true + } + } + return false +} + // FilterSignatureSensitiveBlocksForRetry is a stronger retry filter for cases where upstream errors indicate // signature/thought_signature validation issues involving tool blocks. // diff --git a/backend/internal/service/gateway_service.go b/backend/internal/service/gateway_service.go index 4a8175a4..7c48f243 100644 --- a/backend/internal/service/gateway_service.go +++ b/backend/internal/service/gateway_service.go @@ -1155,6 +1155,12 @@ func normalizeClaudeOAuthRequestBody(body []byte, modelID string, opts claudeOAu // context_management：thinking.type 为 enabled/adaptive 时，真实 CLI 会自动 // 附带 {"edits":[{"type":"clear_thinking_20251015","keep":"all"}]}。 // 客户端显式传了就透传；否则按 CLI 行为补齐。 + // + // 注：本函数不按 model 名决定是否保留 context_management。“最终 beta + // header 不含 context-management-2025-06-27 时 strip 字段”的能力维度 + // 对称约束由 sanitizeAnthropicBodyForBetaTokens 在 buildUpstreamRequest / + // buildCountTokensRequest 层统一执行，与 Bedrock 路径的 + // sanitizeBedrockFieldsForBetaTokens 对称。 if !gjson.GetBytes(out, "context_management").Exists() { thinkingType := gjson.GetBytes(out, "thinking.type").String() if thinkingType == "enabled" || thinkingType == "adaptive" { @@ -5248,6 +5254,17 @@ func (s *GatewayService) buildUpstreamRequestAnthropicAPIKeyPassthrough( targetURL = validatedURL + "/v1/messages?beta=true" } + // 能力维度 body sanitize：透传路径上 anthropic-beta header 原样透传客户端值， + // 依此决定是否保留 body 中的 context_management。避免“客户端 body 带字段但 + // header 忘记带 beta token”的客户端 bug 在透传场景下让上游 400。 + clientBeta := "" + if c != nil && c.Request != nil { + clientBeta = getHeaderRaw(c.Request.Header, "anthropic-beta") + } + if sanitized, changed := sanitizeAnthropicBodyForBetaTokens(body, clientBeta); changed { + body = sanitized + } + req, err := http.NewRequestWithContext(ctx, http.MethodPost, targetURL, bytes.NewReader(body)) if err != nil { return nil, err @@ -6106,6 +6123,29 @@ func (s *GatewayService) buildUpstreamRequest(ctx context.Context, c *gin.Contex if fingerprint != nil { body = syncBillingHeaderVersion(body, fingerprint.UserAgent) } + + // === 计算最终 anthropic-beta header（先于 body sanitize 与 CCH 签名）=== + // + // 顺序约束： + // 1) 算 finalBeta（纯函数，不依赖 req.Header；mimicry 路径会忽略客户端 beta， + // 与原“OAuth + mimicClaudeCode 跳过白名单透传”行为对齐） + // 2) 按 finalBeta 做能力维度 body sanitize（如 context-management beta 缺失 → + // strip body.context_management，与 Bedrock 路径对称） + // 3) CCH 签名（必须使用 strip 后的 body，否则 hash 与最终 body 不一致 → + // 被 Anthropic 判 third-party） + // 4) NewRequest（body 至此最终敲定） + // 5) 透传白名单 / fingerprint / mimic header / 写入 finalBeta + policyFilterSet := s.getBetaPolicyFilterSet(ctx, c, account, modelID) + effectiveDropSet := mergeDropSets(policyFilterSet) + finalBetaHeader, finalBetaShouldSet := s.computeFinalAnthropicBeta( + tokenType, mimicClaudeCode, modelID, clientHeaders, body, effectiveDropSet, + ) + + // 能力维度 body sanitize：与最终 anthropic-beta header 对称 + if sanitized, changed := sanitizeAnthropicBodyForBetaTokens(body, finalBetaHeader); changed { + body = sanitized + } + // CCH 签名：将 cch=00000 占位符替换为 xxHash64 签名（需在所有 body 修改之后） if enableCCH { body = signBillingHeaderCCH(body) @@ -6156,46 +6196,18 @@ func (s *GatewayService) buildUpstreamRequest(ctx context.Context, c *gin.Contex applyClaudeOAuthHeaderDefaults(req) } - // Build effective drop set: merge static defaults with dynamic beta policy filter rules - policyFilterSet := s.getBetaPolicyFilterSet(ctx, c, account, modelID) - effectiveDropSet := mergeDropSets(policyFilterSet) + // OAuth + mimic Claude Code：强制注入 CLI 指纹相关 header + // （user-agent/x-stainless-*/x-app/Accept/x-stainless-helper-method/x-client-request-id） + if tokenType == "oauth" && mimicClaudeCode { + applyClaudeCodeMimicHeaders(req, reqStream) + } - // 处理 anthropic-beta header（OAuth 账号需要包含 oauth beta） - if tokenType == "oauth" { - if mimicClaudeCode { - // 非 Claude Code 客户端：按 opencode 的策略处理： - // - 强制 Claude Code 指纹相关请求头（尤其是 user-agent/x-stainless/x-app） - // - 保留 incoming beta 的同时，确保 OAuth 所需 beta 存在 - applyClaudeCodeMimicHeaders(req, reqStream) - - incomingBeta := getHeaderRaw(req.Header, "anthropic-beta") - // Claude Code OAuth credentials are scoped to Claude Code. - // Non-haiku models MUST include claude-code beta for Anthropic to recognize - // this as a legitimate Claude Code request; without it, the request is - // rejected as third-party ("out of extra usage"). - // Haiku models are exempt from third-party detection and don't need it. - requiredBetas := []string{claude.BetaOAuth, claude.BetaInterleavedThinking} - if !strings.Contains(strings.ToLower(modelID), "haiku") { - requiredBetas = claude.FullClaudeCodeMimicryBetas() - } - setHeaderRaw(req.Header, "anthropic-beta", mergeAnthropicBetaDropping(requiredBetas, incomingBeta, effectiveDropSet)) - } else { - // Claude Code 客户端：尽量透传原始 header，仅补齐 oauth beta - clientBetaHeader := getHeaderRaw(req.Header, "anthropic-beta") - setHeaderRaw(req.Header, "anthropic-beta", stripBetaTokensWithSet(s.getBetaHeader(modelID, clientBetaHeader), effectiveDropSet)) - } - } else { - // API-key accounts: apply beta policy filter to strip controlled tokens - if existingBeta := getHeaderRaw(req.Header, "anthropic-beta"); existingBeta != "" { - setHeaderRaw(req.Header, "anthropic-beta", stripBetaTokensWithSet(existingBeta, effectiveDropSet)) - } else if s.cfg != nil && s.cfg.Gateway.InjectBetaForAPIKey { - // API-key：仅在请求显式使用 beta 特性且客户端未提供时，按需补齐（默认关闭） - if requestNeedsBetaFeatures(body) { - if beta := defaultAPIKeyBetaHeader(body); beta != "" { - setHeaderRaw(req.Header, "anthropic-beta", beta) - } - } - } + // 写入最终 anthropic-beta header + // 注：透传分支白名单可能写入了客户端 anthropic-beta，无条件 Del 一次再按 finalBeta + // 决定是否 set，确保 dropSet 过滤后的结果一定覆盖客户端原始值。 + deleteHeaderAllForms(req.Header, "anthropic-beta") + if finalBetaShouldSet { + setHeaderRaw(req.Header, "anthropic-beta", finalBetaHeader) } // 同步 X-Claude-Code-Session-Id 头：取 body 中已处理的 metadata.user_id 的 session_id 覆盖 @@ -6242,6 +6254,16 @@ func (s *GatewayService) buildUpstreamRequestAnthropicVertex( if err != nil { return nil, err } + + // 能力维度 sanitize：Vertex 路径上 anthropic-beta header 原样透传客户端值 + // （下面白名单跳过 anthropic-version 但保留 anthropic-beta），依此决定是否 + // 保留 body 中的 context_management，与 Anthropic 直连 / Bedrock 路径对称。 + if c != nil && c.Request != nil { + clientBeta := getHeaderRaw(c.Request.Header, "anthropic-beta") + if sanitized, changed := sanitizeAnthropicBodyForBetaTokens(vertexBody, clientBeta); changed { + vertexBody = sanitized + } + } fullURL, err := buildVertexAnthropicURL(account.VertexProjectID(), account.VertexLocation(modelID), modelID, reqStream) if err != nil { return nil, err @@ -6410,6 +6432,121 @@ func mergeAnthropicBetaDropping(required []string, incoming string, drop map[str return strings.Join(out, ",") } +// computeFinalAnthropicBeta 计算发往上游的最终 anthropic-beta header 值。 +// +// 设计动机：将原本在 buildUpstreamRequest 内联在一起、依赖 req.Header 的 +// anthropic-beta 计算逻辑抽成纯函数。这样调用方可以在 NewRequest 之前 +// 就提前拿到最终 beta header，进而能按它对 body 做能力维度 sanitize 后再做 +// CCH 签名——一举修复了以下之前由顺序依赖导致的能力维度 sanitize +// 无法部署的问题（签名与最终 body 不一致可以被判 third-party）。 +// +// 返回 (value, shouldSet)： +// - shouldSet=false 意为“不主动设置 anthropic-beta header”，与原代码“ +// API-key 账号 + 客户端未传 anthropic-beta + InjectBetaForAPIKey 未开启或 +// requestNeedsBetaFeatures=false”的行为对齐。 +// - shouldSet=true 时 value 可能为空字符串（例如客户端透传的 beta 被 dropSet +// 全部过滤掉），这与原代码中 setHeaderRaw 的结果一致。 +// +// clientHeaders 是客户端原始 HTTP header（通常为 c.Request.Header）；nil 时按“客户端 +// 未传”处理。body 是已经 metadata 重写 / billing version sync 之后但未 sanitize 上游 +// 不兼容字段之前的版本。 +func (s *GatewayService) computeFinalAnthropicBeta( + tokenType string, + mimicClaudeCode bool, + modelID string, + clientHeaders http.Header, + body []byte, + effectiveDropSet map[string]struct{}, +) (string, bool) { + clientBeta := "" + if clientHeaders != nil { + clientBeta = getHeaderRaw(clientHeaders, "anthropic-beta") + } + + if tokenType == "oauth" { + if mimicClaudeCode { + // mimic 路径：原代码跳过白名单透传，incomingBeta 总是空字符串。 + // 这里传空 string 以严格对齐原行为。 + requiredBetas := []string{claude.BetaOAuth, claude.BetaInterleavedThinking} + if !strings.Contains(strings.ToLower(modelID), "haiku") { + requiredBetas = claude.FullClaudeCodeMimicryBetas() + } + return mergeAnthropicBetaDropping(requiredBetas, "", effectiveDropSet), true + } + // 真 Claude Code 客户端透传路径 + return stripBetaTokensWithSet(s.getBetaHeader(modelID, clientBeta), effectiveDropSet), true + } + + // API-key accounts + if clientBeta != "" { + return stripBetaTokensWithSet(clientBeta, effectiveDropSet), true + } + if s.cfg != nil && s.cfg.Gateway.InjectBetaForAPIKey { + if requestNeedsBetaFeatures(body) { + if beta := defaultAPIKeyBetaHeader(body); beta != "" { + return beta, true + } + } + } + return "", false +} + +// computeFinalCountTokensAnthropicBeta 是 count_tokens 路径上 anthropic-beta header 的 +// 计算纯函数。语义与 computeFinalAnthropicBeta 对齐，但备份了 count_tokens 独有的 +// 两条特殊规则： +// +// - OAuth mimic：requiredBetas 为 FullClaudeCodeMimicryBetas + BetaTokenCounting +// （与 messages 不同的是：不按 haiku 排除；count_tokens 始终携带 token-counting beta） +// - OAuth 透传 + 客户端未传 anthropic-beta：补齐 CountTokensBetaHeader +// - OAuth 透传 + 客户端传了：补齐 BetaTokenCounting（如果未含） +// +// 返回语义同 computeFinalAnthropicBeta。 +func (s *GatewayService) computeFinalCountTokensAnthropicBeta( + tokenType string, + mimicClaudeCode bool, + modelID string, + clientHeaders http.Header, + body []byte, + effectiveDropSet map[string]struct{}, +) (string, bool) { + clientBeta := "" + if clientHeaders != nil { + clientBeta = getHeaderRaw(clientHeaders, "anthropic-beta") + } + + if tokenType == "oauth" { + if mimicClaudeCode { + // 与原代码严格等价：original buildCountTokensRequest 在 count_tokens mimic + // 分支上**不**会跳过白名单透传（与 messages mimic 路径不同），所以 + // incomingBeta = req.Header[anthropic-beta] = 客户端透传过来的 client beta。 + // 重构后直接从 clientHeaders 拿同一个值，保持行为一致。 + requiredBetas := append(claude.FullClaudeCodeMimicryBetas(), claude.BetaTokenCounting) + return mergeAnthropicBetaDropping(requiredBetas, clientBeta, effectiveDropSet), true + } + if clientBeta == "" { + return claude.CountTokensBetaHeader, true + } + beta := s.getBetaHeader(modelID, clientBeta) + if !strings.Contains(beta, claude.BetaTokenCounting) { + beta = beta + "," + claude.BetaTokenCounting + } + return stripBetaTokensWithSet(beta, effectiveDropSet), true + } + + // API-key accounts + if clientBeta != "" { + return stripBetaTokensWithSet(clientBeta, effectiveDropSet), true + } + if s.cfg != nil && s.cfg.Gateway.InjectBetaForAPIKey { + if requestNeedsBetaFeatures(body) { + if beta := defaultAPIKeyBetaHeader(body); beta != "" { + return beta, true + } + } + } + return "", false +} + // stripBetaTokens removes the given beta tokens from a comma-separated header value. func stripBetaTokens(header string, tokens []string) string { if header == "" || len(tokens) == 0 { @@ -9312,6 +9449,15 @@ func (s *GatewayService) buildCountTokensRequestAnthropicAPIKeyPassthrough( targetURL = validatedURL + "/v1/messages/count_tokens?beta=true" } + // 同 buildUpstreamRequestAnthropicAPIKeyPassthrough：能力维度 sanitize。 + clientBeta := "" + if c != nil && c.Request != nil { + clientBeta = getHeaderRaw(c.Request.Header, "anthropic-beta") + } + if sanitized, changed := sanitizeAnthropicBodyForBetaTokens(body, clientBeta); changed { + body = sanitized + } + req, err := http.NewRequestWithContext(ctx, http.MethodPost, targetURL, bytes.NewReader(body)) if err != nil { return nil, err @@ -9402,6 +9548,19 @@ func (s *GatewayService) buildCountTokensRequest(ctx context.Context, c *gin.Con if ctFingerprint != nil && ctEnableFP { body = syncBillingHeaderVersion(body, ctFingerprint.UserAgent) } + + // === 计算最终 anthropic-beta header（先于 body sanitize 与 CCH 签名）=== + // 顺序约束同 buildUpstreamRequest。 + ctEffectiveDropSet := mergeDropSets(s.getBetaPolicyFilterSet(ctx, c, account, modelID)) + finalBetaHeader, finalBetaShouldSet := s.computeFinalCountTokensAnthropicBeta( + tokenType, mimicClaudeCode, modelID, clientHeaders, body, ctEffectiveDropSet, + ) + + // 能力维度 body sanitize：与最终 anthropic-beta header 对称 + if sanitized, changed := sanitizeAnthropicBodyForBetaTokens(body, finalBetaHeader); changed { + body = sanitized + } + if ctEnableCCH { body = signBillingHeaderCCH(body) } @@ -9445,41 +9604,15 @@ func (s *GatewayService) buildCountTokensRequest(ctx context.Context, c *gin.Con applyClaudeOAuthHeaderDefaults(req) } - // Build effective drop set for count_tokens: merge static defaults with dynamic beta policy filter rules - ctEffectiveDropSet := mergeDropSets(s.getBetaPolicyFilterSet(ctx, c, account, modelID)) + // OAuth + mimic Claude Code：强制注入 CLI 指纹 header + if tokenType == "oauth" && mimicClaudeCode { + applyClaudeCodeMimicHeaders(req, false) + } - // OAuth 账号：处理 anthropic-beta header - if tokenType == "oauth" { - if mimicClaudeCode { - applyClaudeCodeMimicHeaders(req, false) - - incomingBeta := getHeaderRaw(req.Header, "anthropic-beta") - requiredBetas := append(claude.FullClaudeCodeMimicryBetas(), claude.BetaTokenCounting) - setHeaderRaw(req.Header, "anthropic-beta", mergeAnthropicBetaDropping(requiredBetas, incomingBeta, ctEffectiveDropSet)) - } else { - clientBetaHeader := getHeaderRaw(req.Header, "anthropic-beta") - if clientBetaHeader == "" { - setHeaderRaw(req.Header, "anthropic-beta", claude.CountTokensBetaHeader) - } else { - beta := s.getBetaHeader(modelID, clientBetaHeader) - if !strings.Contains(beta, claude.BetaTokenCounting) { - beta = beta + "," + claude.BetaTokenCounting - } - setHeaderRaw(req.Header, "anthropic-beta", stripBetaTokensWithSet(beta, ctEffectiveDropSet)) - } - } - } else { - // API-key accounts: apply beta policy filter to strip controlled tokens - if existingBeta := getHeaderRaw(req.Header, "anthropic-beta"); existingBeta != "" { - setHeaderRaw(req.Header, "anthropic-beta", stripBetaTokensWithSet(existingBeta, ctEffectiveDropSet)) - } else if s.cfg != nil && s.cfg.Gateway.InjectBetaForAPIKey { - // API-key：与 messages 同步的按需 beta 注入（默认关闭） - if requestNeedsBetaFeatures(body) { - if beta := defaultAPIKeyBetaHeader(body); beta != "" { - setHeaderRaw(req.Header, "anthropic-beta", beta) - } - } - } + // 写入最终 anthropic-beta header（Del 一次避免白名单透传值残留） + deleteHeaderAllForms(req.Header, "anthropic-beta") + if finalBetaShouldSet { + setHeaderRaw(req.Header, "anthropic-beta", finalBetaHeader) } // 同步 X-Claude-Code-Session-Id 头：取 body 中已处理的 metadata.user_id 的 session_id 覆盖 diff --git a/backend/internal/service/header_util.go b/backend/internal/service/header_util.go index 1091070d..f8da068d 100644 --- a/backend/internal/service/header_util.go +++ b/backend/internal/service/header_util.go @@ -109,6 +109,20 @@ func addHeaderRaw(h http.Header, key, value string) { h[key] = append(h[key], value) } +// deleteHeaderAllForms removes a header in all common key forms (raw, wire casing, +// canonical) so subsequent setHeaderRaw will not coexist with a passthrough value +// written under a different casing. +func deleteHeaderAllForms(h http.Header, key string) { + if h == nil || key == "" { + return + } + h.Del(key) // canonical + delete(h, key) + if wk := resolveWireCasing(key); wk != key { + delete(h, wk) + } +} + // getHeaderRaw reads a header value, trying multiple key forms to handle the mismatch // between Go canonical keys, wire casing keys, and raw keys: // 1. exact key as provided From 20f5340784484d8be82c2139bb40ff34f1a7b715 Mon Sep 17 00:00:00 2001 From: JIA-ss <627723154@qq.com> Date: Thu, 28 May 2026 00:38:25 +0800 Subject: [PATCH 04/20] =?UTF-8?q?fix(apicompat):=20Responses=E2=86=92Chat?= =?UTF-8?q?=20=E8=BD=AC=E6=8D=A2=E8=A1=A5=E9=BD=90=20completion=5Ftokens?= =?UTF-8?q?=5Fdetails=20=E9=80=8F=E4=BC=A0?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit OpenAI Responses API 在 gpt-5.x 等 reasoning 模型上会返回 output_tokens_details.reasoning_tokens, 但 ResponsesToChatCompletions 只映射了 input_tokens_details.cached_tokens, 导致客户端拿到的 chat.completion.usage 中 completion_tokens 出现无法解释的波动 (短 prompt 也可能 30+ token), 且缺失 reasoning_tokens 细分字段, 难以与 OpenAI 原生 Chat Completions 响应对账。 按 OpenAI 官方 CompletionUsage schema (openai/openai-go SDK completion.go) 补齐所有 token-details 字段, 全部 omitempty: prompt_tokens_details: - cached_tokens (原已支持) - audio_tokens (新增) completion_tokens_details: - reasoning_tokens (新增) - audio_tokens (新增) - accepted_prediction_tokens (新增) - rejected_prediction_tokens (新增) 实现细节: - 抽出 promptDetailsFromResponses / completionDetailsFromResponses 两个 helper, 全零字段返回 nil - 非流路径 ResponsesToChatCompletions 复用已存在的 chatUsageFromResponsesUsage helper, 消除两条路径间的重复 - 非 reasoning / 非 audio 上游 (Anthropic, Gemini, gpt-4o) 不填这些 字段, helper 返回 nil → CompletionTokensDetails 不输出, 对现有响应 字节级兼容 新增单测: - TestResponsesToChatCompletions_ReasoningTokens - TestResponsesToChatCompletions_AllTokenDetailsPassThrough - TestResponsesToChatCompletions_NoReasoningTokensWhenZero - TestResponsesEventToChatChunks_CompletedWithReasoningTokens --- .../chatcompletions_responses_test.go | 135 ++++++++++++++++++ .../apicompat/responses_to_chatcompletions.go | 58 +++++--- backend/internal/pkg/apicompat/types.go | 30 +++- 3 files changed, 198 insertions(+), 25 deletions(-) diff --git a/backend/internal/pkg/apicompat/chatcompletions_responses_test.go b/backend/internal/pkg/apicompat/chatcompletions_responses_test.go index 016c2415..b03b012f 100644 --- a/backend/internal/pkg/apicompat/chatcompletions_responses_test.go +++ b/backend/internal/pkg/apicompat/chatcompletions_responses_test.go @@ -663,6 +663,115 @@ func TestResponsesToChatCompletions_CachedTokens(t *testing.T) { assert.Equal(t, 80, chat.Usage.PromptTokensDetails.CachedTokens) } +func TestResponsesToChatCompletions_ReasoningTokens(t *testing.T) { + resp := &ResponsesResponse{ + ID: "resp_reasoning", + Status: "completed", + Output: []ResponsesOutput{ + { + Type: "message", + Content: []ResponsesContentPart{{Type: "output_text", Text: "ping"}}, + }, + }, + Usage: &ResponsesUsage{ + InputTokens: 24, + OutputTokens: 33, + TotalTokens: 57, + OutputTokensDetails: &ResponsesOutputTokensDetails{ + ReasoningTokens: 32, + }, + }, + } + + chat := ResponsesToChatCompletions(resp, "gpt-5.5") + require.NotNil(t, chat.Usage) + assert.Equal(t, 33, chat.Usage.CompletionTokens) + require.NotNil(t, chat.Usage.CompletionTokensDetails) + assert.Equal(t, 32, chat.Usage.CompletionTokensDetails.ReasoningTokens) +} + +func TestResponsesToChatCompletions_AllTokenDetailsPassThrough(t *testing.T) { + // Covers the full OpenAI CompletionUsage detail field set so future audio + // and prediction-outputs responses propagate without further changes. + resp := &ResponsesResponse{ + ID: "resp_full_details", + Status: "completed", + Output: []ResponsesOutput{ + { + Type: "message", + Content: []ResponsesContentPart{{Type: "output_text", Text: "x"}}, + }, + }, + Usage: &ResponsesUsage{ + InputTokens: 100, + OutputTokens: 50, + TotalTokens: 150, + InputTokensDetails: &ResponsesInputTokensDetails{ + CachedTokens: 60, + AudioTokens: 4, + }, + OutputTokensDetails: &ResponsesOutputTokensDetails{ + ReasoningTokens: 30, + AudioTokens: 2, + AcceptedPredictionTokens: 10, + RejectedPredictionTokens: 3, + }, + }, + } + + chat := ResponsesToChatCompletions(resp, "gpt-5.5") + require.NotNil(t, chat.Usage) + require.NotNil(t, chat.Usage.PromptTokensDetails) + assert.Equal(t, 60, chat.Usage.PromptTokensDetails.CachedTokens) + assert.Equal(t, 4, chat.Usage.PromptTokensDetails.AudioTokens) + + require.NotNil(t, chat.Usage.CompletionTokensDetails) + assert.Equal(t, 30, chat.Usage.CompletionTokensDetails.ReasoningTokens) + assert.Equal(t, 2, chat.Usage.CompletionTokensDetails.AudioTokens) + assert.Equal(t, 10, chat.Usage.CompletionTokensDetails.AcceptedPredictionTokens) + assert.Equal(t, 3, chat.Usage.CompletionTokensDetails.RejectedPredictionTokens) + + raw, err := json.Marshal(chat.Usage) + require.NoError(t, err) + assert.Contains(t, string(raw), `"prompt_tokens_details"`) + assert.Contains(t, string(raw), `"completion_tokens_details"`) + assert.Contains(t, string(raw), `"reasoning_tokens":30`) + assert.Contains(t, string(raw), `"accepted_prediction_tokens":10`) +} + +func TestResponsesToChatCompletions_NoReasoningTokensWhenZero(t *testing.T) { + // Non-reasoning models do not return reasoning_tokens. The mapping must + // omit completion_tokens_details entirely rather than emitting a zero-valued + // field, so non-reasoning responses stay clean. + resp := &ResponsesResponse{ + ID: "resp_no_reasoning", + Status: "completed", + Output: []ResponsesOutput{ + { + Type: "message", + Content: []ResponsesContentPart{{Type: "output_text", Text: "hi"}}, + }, + }, + Usage: &ResponsesUsage{ + InputTokens: 10, + OutputTokens: 5, + TotalTokens: 15, + OutputTokensDetails: &ResponsesOutputTokensDetails{ + ReasoningTokens: 0, + }, + }, + } + + chat := ResponsesToChatCompletions(resp, "gpt-4o") + require.NotNil(t, chat.Usage) + assert.Nil(t, chat.Usage.CompletionTokensDetails) + + raw, err := json.Marshal(chat.Usage) + require.NoError(t, err) + assert.NotContains(t, string(raw), "completion_tokens_details") + assert.NotContains(t, string(raw), "reasoning_tokens") +} + func TestResponsesToChatCompletions_WebSearch(t *testing.T) { resp := &ResponsesResponse{ ID: "resp_ws", @@ -825,6 +934,32 @@ func TestResponsesEventToChatChunks_Completed(t *testing.T) { assert.Equal(t, 30, chunks[1].Usage.PromptTokensDetails.CachedTokens) } +func TestResponsesEventToChatChunks_CompletedWithReasoningTokens(t *testing.T) { + state := NewResponsesEventToChatState() + state.Model = "gpt-5.5" + state.IncludeUsage = true + + chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{ + Type: "response.completed", + Response: &ResponsesResponse{ + Status: "completed", + Usage: &ResponsesUsage{ + InputTokens: 24, + OutputTokens: 33, + TotalTokens: 57, + OutputTokensDetails: &ResponsesOutputTokensDetails{ + ReasoningTokens: 32, + }, + }, + }, + }, state) + require.Len(t, chunks, 2) + + require.NotNil(t, chunks[1].Usage) + require.NotNil(t, chunks[1].Usage.CompletionTokensDetails) + assert.Equal(t, 32, chunks[1].Usage.CompletionTokensDetails.ReasoningTokens) +} + func TestResponsesEventToChatChunks_ResponseDone(t *testing.T) { state := NewResponsesEventToChatState() state.Model = "gpt-4o" diff --git a/backend/internal/pkg/apicompat/responses_to_chatcompletions.go b/backend/internal/pkg/apicompat/responses_to_chatcompletions.go index 7e8354ee..8809b4fc 100644 --- a/backend/internal/pkg/apicompat/responses_to_chatcompletions.go +++ b/backend/internal/pkg/apicompat/responses_to_chatcompletions.go @@ -81,19 +81,7 @@ func ResponsesToChatCompletions(resp *ResponsesResponse, model string) *ChatComp FinishReason: finishReason, }} - if resp.Usage != nil { - usage := &ChatUsage{ - PromptTokens: resp.Usage.InputTokens, - CompletionTokens: resp.Usage.OutputTokens, - TotalTokens: resp.Usage.InputTokens + resp.Usage.OutputTokens, - } - if resp.Usage.InputTokensDetails != nil && resp.Usage.InputTokensDetails.CachedTokens > 0 { - usage.PromptTokensDetails = &ChatTokenDetails{ - CachedTokens: resp.Usage.InputTokensDetails.CachedTokens, - } - } - out.Usage = usage - } + out.Usage = chatUsageFromResponsesUsage(resp.Usage) return out } @@ -341,14 +329,48 @@ func chatUsageFromResponsesUsage(u *ResponsesUsage) *ChatUsage { CompletionTokens: u.OutputTokens, TotalTokens: u.InputTokens + u.OutputTokens, } - if u.InputTokensDetails != nil && u.InputTokensDetails.CachedTokens > 0 { - usage.PromptTokensDetails = &ChatTokenDetails{ - CachedTokens: u.InputTokensDetails.CachedTokens, - } - } + usage.PromptTokensDetails = promptDetailsFromResponses(u.InputTokensDetails) + usage.CompletionTokensDetails = completionDetailsFromResponses(u.OutputTokensDetails) return usage } +// promptDetailsFromResponses maps Responses-API input_tokens_details into a +// Chat-Completions prompt_tokens_details. Returns nil when nothing would be +// emitted, so upstreams that do not break down prompt usage stay clean. +func promptDetailsFromResponses(src *ResponsesInputTokensDetails) *ChatTokenDetails { + if src == nil { + return nil + } + if src.CachedTokens == 0 && src.AudioTokens == 0 { + return nil + } + return &ChatTokenDetails{ + CachedTokens: src.CachedTokens, + AudioTokens: src.AudioTokens, + } +} + +// completionDetailsFromResponses maps Responses-API output_tokens_details +// into a Chat-Completions completion_tokens_details. Mirrors the OpenAI +// official CompletionUsage schema: reasoning_tokens, audio_tokens, and +// the predicted-outputs accepted/rejected counts. Returns nil when nothing +// would be emitted so non-reasoning, non-audio responses stay clean. +func completionDetailsFromResponses(src *ResponsesOutputTokensDetails) *ChatTokenDetails { + if src == nil { + return nil + } + if src.ReasoningTokens == 0 && src.AudioTokens == 0 && + src.AcceptedPredictionTokens == 0 && src.RejectedPredictionTokens == 0 { + return nil + } + return &ChatTokenDetails{ + ReasoningTokens: src.ReasoningTokens, + AudioTokens: src.AudioTokens, + AcceptedPredictionTokens: src.AcceptedPredictionTokens, + RejectedPredictionTokens: src.RejectedPredictionTokens, + } +} + func makeChatDeltaChunk(state *ResponsesEventToChatState, delta ChatDelta) ChatCompletionsChunk { return ChatCompletionsChunk{ ID: state.ID, diff --git a/backend/internal/pkg/apicompat/types.go b/backend/internal/pkg/apicompat/types.go index 8b576647..b4451f23 100644 --- a/backend/internal/pkg/apicompat/types.go +++ b/backend/internal/pkg/apicompat/types.go @@ -362,11 +362,15 @@ func (u *ResponsesUsage) UnmarshalJSON(data []byte) error { // ResponsesInputTokensDetails breaks down input token usage. type ResponsesInputTokensDetails struct { CachedTokens int `json:"cached_tokens,omitempty"` + AudioTokens int `json:"audio_tokens,omitempty"` } // ResponsesOutputTokensDetails breaks down output token usage. type ResponsesOutputTokensDetails struct { - ReasoningTokens int `json:"reasoning_tokens,omitempty"` + ReasoningTokens int `json:"reasoning_tokens,omitempty"` + AudioTokens int `json:"audio_tokens,omitempty"` + AcceptedPredictionTokens int `json:"accepted_prediction_tokens,omitempty"` + RejectedPredictionTokens int `json:"rejected_prediction_tokens,omitempty"` } // --------------------------------------------------------------------------- @@ -517,15 +521,27 @@ type ChatChoice struct { // ChatUsage holds token counts in Chat Completions format. type ChatUsage struct { - PromptTokens int `json:"prompt_tokens"` - CompletionTokens int `json:"completion_tokens"` - TotalTokens int `json:"total_tokens"` - PromptTokensDetails *ChatTokenDetails `json:"prompt_tokens_details,omitempty"` + PromptTokens int `json:"prompt_tokens"` + CompletionTokens int `json:"completion_tokens"` + TotalTokens int `json:"total_tokens"` + PromptTokensDetails *ChatTokenDetails `json:"prompt_tokens_details,omitempty"` + CompletionTokensDetails *ChatTokenDetails `json:"completion_tokens_details,omitempty"` } -// ChatTokenDetails provides a breakdown of token usage. +// ChatTokenDetails provides a breakdown of token usage. The same type is +// reused for both prompt_tokens_details and completion_tokens_details; +// unset fields are omitted so each side only emits the fields that apply. +// +// Field set mirrors OpenAI's official CompletionUsage schema: +// - prompt_tokens_details: cached_tokens, audio_tokens +// - completion_tokens_details: reasoning_tokens, audio_tokens, +// accepted_prediction_tokens, rejected_prediction_tokens type ChatTokenDetails struct { - CachedTokens int `json:"cached_tokens,omitempty"` + CachedTokens int `json:"cached_tokens,omitempty"` + AudioTokens int `json:"audio_tokens,omitempty"` + ReasoningTokens int `json:"reasoning_tokens,omitempty"` + AcceptedPredictionTokens int `json:"accepted_prediction_tokens,omitempty"` + RejectedPredictionTokens int `json:"rejected_prediction_tokens,omitempty"` } // ChatCompletionsChunk is a single streaming chunk from POST /v1/chat/completions. From d7bed40dda4a5bcf7cb969d73c221829de02a91d Mon Sep 17 00:00:00 2001 From: siyuan <740665504@qq.com> Date: Thu, 28 May 2026 01:27:11 +0800 Subject: [PATCH 05/20] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=20OpenAI=20WS=20?= =?UTF-8?q?=E5=85=BC=E5=AE=B9=E6=80=A7=E4=B8=8E=20usage=20=E7=BB=9F?= =?UTF-8?q?=E8=AE=A1?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 对齐 WS 与流式终态 usage 解析，补齐 failed/done/incomplete/cancelled 等事件 - 兼容后续 WS response.create 省略 model，保持模型映射与权限判断一致 - 补齐 passthrough header 透传和图片 usage 字段映射 --- .../service/openai_gateway_service.go | 2 +- .../service/openai_gateway_service_test.go | 6 + .../internal/service/openai_ws_forwarder.go | 39 ++- ..._ws_forwarder_hotpath_optimization_test.go | 18 ++ ...penai_ws_forwarder_ingress_session_test.go | 252 ++++++++++++++++++ .../openai_ws_forwarder_success_test.go | 64 +++++ .../service/openai_ws_v2/passthrough_relay.go | 26 +- .../passthrough_relay_internal_test.go | 40 ++- .../openai_ws_v2_passthrough_adapter.go | 11 +- 9 files changed, 439 insertions(+), 19 deletions(-) diff --git a/backend/internal/service/openai_gateway_service.go b/backend/internal/service/openai_gateway_service.go index f93cc221..7a04f78e 100644 --- a/backend/internal/service/openai_gateway_service.go +++ b/backend/internal/service/openai_gateway_service.go @@ -4861,7 +4861,7 @@ func (s *OpenAIGatewayService) parseSSEUsageBytes(data []byte, usage *OpenAIUsag return } eventType := gjson.GetBytes(data, "type").String() - if eventType != "response.completed" && eventType != "response.done" && + if eventType != "response.completed" && eventType != "response.done" && eventType != "response.failed" && eventType != "response.incomplete" && eventType != "response.cancelled" && eventType != "response.canceled" { return } diff --git a/backend/internal/service/openai_gateway_service_test.go b/backend/internal/service/openai_gateway_service_test.go index 8bed920d..8aad2fa6 100644 --- a/backend/internal/service/openai_gateway_service_test.go +++ b/backend/internal/service/openai_gateway_service_test.go @@ -2218,6 +2218,12 @@ func TestParseSSEUsage_SelectiveParsing(t *testing.T) { require.Equal(t, 15, usage.OutputTokens) require.Equal(t, 4, usage.CacheReadInputTokens) + // failed 事件在部分上游路径也会携带已消耗 usage，应与 WS/passthrough 保持一致 + svc.parseSSEUsage(`{"type":"response.failed","response":{"usage":{"input_tokens":17,"output_tokens":19,"input_tokens_details":{"cached_tokens":6}}}}`, usage) + require.Equal(t, 17, usage.InputTokens) + require.Equal(t, 19, usage.OutputTokens) + require.Equal(t, 6, usage.CacheReadInputTokens) + svc.parseSSEUsage(`{"type":"response.completed","response":{"usage":{"prompt_tokens":21,"completion_tokens":8,"prompt_tokens_details":{"cached_tokens":6}}}}`, usage) require.Equal(t, 21, usage.InputTokens) require.Equal(t, 8, usage.OutputTokens) diff --git a/backend/internal/service/openai_ws_forwarder.go b/backend/internal/service/openai_ws_forwarder.go index b8e558ae..75f6559b 100644 --- a/backend/internal/service/openai_ws_forwarder.go +++ b/backend/internal/service/openai_ws_forwarder.go @@ -369,7 +369,12 @@ func openAIWSEventMayContainToolCalls(eventType string) bool { } func openAIWSEventShouldParseUsage(eventType string) bool { - return eventType == "response.completed" || strings.TrimSpace(eventType) == "response.completed" + switch strings.TrimSpace(eventType) { + case "response.completed", "response.done", "response.failed", "response.incomplete", "response.cancelled", "response.canceled": + return true + default: + return false + } } func parseOpenAIWSEventEnvelope(message []byte) (eventType string, responseID string, response gjson.Result) { @@ -2484,6 +2489,7 @@ func (s *OpenAIGatewayService) ProxyResponsesWebSocketFromClient( imageInputSize string payloadBytes int } + ingressSessionOriginalModel := "" applyPayloadMutation := func(current []byte, path string, value any) ([]byte, error) { next, err := sjson.SetBytes(current, path, value) @@ -2547,12 +2553,21 @@ func (s *OpenAIGatewayService) ProxyResponsesWebSocketFromClient( } originalModel := strings.TrimSpace(values[1].String()) + modelMissing := originalModel == "" if originalModel == "" { - return openAIWSClientPayload{}, NewOpenAIWSClientCloseError( - coderws.StatusPolicyViolation, - "model is required in response.create payload", - nil, - ) + // 入站 WS 长会话里，部分客户端只在第一轮 response.create 上声明 + // model，后续 turn 复用同一 session-level model。为避免因省略 + // model 直接断开用户连接，这里回落到上一轮已通过校验的客户端模型， + // 并在下方写回上游 payload，保证账号模型映射/fast policy/图片权限 + // 仍按同一模型执行。 + originalModel = ingressSessionOriginalModel + if originalModel == "" { + return openAIWSClientPayload{}, NewOpenAIWSClientCloseError( + coderws.StatusPolicyViolation, + "model is required in response.create payload", + nil, + ) + } } promptCacheKey := strings.TrimSpace(values[2].String()) previousResponseID := strings.TrimSpace(values[3].String()) @@ -2572,7 +2587,7 @@ func (s *OpenAIGatewayService) ProxyResponsesWebSocketFromClient( normalized = next } upstreamModel := normalizeOpenAIModelForUpstream(account, account.GetMappedModel(originalModel)) - if upstreamModel != originalModel { + if modelMissing || upstreamModel != originalModel { next, setErr := applyPayloadMutation(normalized, "model", upstreamModel) if setErr != nil { return openAIWSClientPayload{}, NewOpenAIWSClientCloseError(coderws.StatusPolicyViolation, "invalid websocket request payload", setErr) @@ -2602,11 +2617,10 @@ func (s *OpenAIGatewayService) ProxyResponsesWebSocketFromClient( // single integration point for all WS ingress turns (first + follow-up // frames flow through here). // - // Model fallback: parseClientPayload above rejects any frame whose - // "model" field is missing (line ~2493-2500), so by the time we - // reach this point upstreamModel is always derived from a non-empty - // per-frame model. The capturedSessionModel fallback used in the - // passthrough adapter is therefore not needed in this path. + // Model fallback: first turn still requires model at the handler layer； + // follow-up response.create frames may omit it and then reuse + // ingressSessionOriginalModel. We always write a concrete upstream model + // before evaluating policy, so whitelist / filter behavior remains stable. policyApplied, blocked, policyErr := s.applyOpenAIFastPolicyToWSResponseCreate(ctx, account, upstreamModel, normalized) if policyErr != nil { return openAIWSClientPayload{}, NewOpenAIWSClientCloseError(coderws.StatusPolicyViolation, "invalid websocket request payload", policyErr) @@ -2635,6 +2649,7 @@ func (s *OpenAIGatewayService) ProxyResponsesWebSocketFromClient( ) } normalized = policyApplied + ingressSessionOriginalModel = originalModel return openAIWSClientPayload{ payloadRaw: normalized, diff --git a/backend/internal/service/openai_ws_forwarder_hotpath_optimization_test.go b/backend/internal/service/openai_ws_forwarder_hotpath_optimization_test.go index 0350bde9..2622f7f2 100644 --- a/backend/internal/service/openai_ws_forwarder_hotpath_optimization_test.go +++ b/backend/internal/service/openai_ws_forwarder_hotpath_optimization_test.go @@ -39,6 +39,24 @@ func TestParseOpenAIWSResponseUsageFromCompletedEvent(t *testing.T) { require.Equal(t, 4, usage.CacheReadInputTokens) } +func TestOpenAIWSEventShouldParseUsageTerminalEvents(t *testing.T) { + t.Parallel() + + for _, eventType := range []string{ + "response.completed", + "response.done", + "response.failed", + "response.incomplete", + "response.cancelled", + "response.canceled", + } { + require.True(t, openAIWSEventShouldParseUsage(eventType), eventType) + require.True(t, openAIWSEventShouldParseUsage(" "+eventType+" "), eventType) + } + require.False(t, openAIWSEventShouldParseUsage("response.output_text.delta")) + require.False(t, openAIWSEventShouldParseUsage("")) +} + func TestOpenAIWSErrorEventHelpers_ConsistentWithWrapper(t *testing.T) { message := []byte(`{"type":"error","error":{"type":"invalid_request_error","code":"invalid_request","message":"invalid input"}}`) codeRaw, errTypeRaw, errMsgRaw := parseOpenAIWSErrorEventFields(message) diff --git a/backend/internal/service/openai_ws_forwarder_ingress_session_test.go b/backend/internal/service/openai_ws_forwarder_ingress_session_test.go index edb6fbcd..b7f1bc4f 100644 --- a/backend/internal/service/openai_ws_forwarder_ingress_session_test.go +++ b/backend/internal/service/openai_ws_forwarder_ingress_session_test.go @@ -164,6 +164,140 @@ func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_KeepLeaseAcrossT require.Len(t, captureConn.writes, 2, "应向同一上游连接发送两轮 response.create") } +func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_FollowupCreateCanOmitModel(t *testing.T) { + gin.SetMode(gin.TestMode) + + cfg := &config.Config{} + cfg.Security.URLAllowlist.Enabled = false + cfg.Security.URLAllowlist.AllowInsecureHTTP = true + cfg.Gateway.OpenAIWS.Enabled = true + cfg.Gateway.OpenAIWS.OAuthEnabled = true + cfg.Gateway.OpenAIWS.APIKeyEnabled = true + cfg.Gateway.OpenAIWS.ResponsesWebsocketsV2 = true + cfg.Gateway.OpenAIWS.MaxConnsPerAccount = 1 + cfg.Gateway.OpenAIWS.MinIdlePerAccount = 0 + cfg.Gateway.OpenAIWS.MaxIdlePerAccount = 1 + cfg.Gateway.OpenAIWS.QueueLimitPerConn = 8 + cfg.Gateway.OpenAIWS.DialTimeoutSeconds = 3 + cfg.Gateway.OpenAIWS.ReadTimeoutSeconds = 3 + cfg.Gateway.OpenAIWS.WriteTimeoutSeconds = 3 + + captureConn := &openAIWSCaptureConn{ + events: [][]byte{ + []byte(`{"type":"response.completed","response":{"id":"resp_omit_model_1","model":"gpt-5.1","usage":{"input_tokens":1,"output_tokens":1}}}`), + []byte(`{"type":"response.completed","response":{"id":"resp_omit_model_2","model":"gpt-5.1","usage":{"input_tokens":1,"output_tokens":1}}}`), + }, + } + captureDialer := &openAIWSCaptureDialer{conn: captureConn} + pool := newOpenAIWSConnPool(cfg) + pool.setClientDialerForTest(captureDialer) + svc := &OpenAIGatewayService{ + cfg: cfg, + httpUpstream: &httpUpstreamRecorder{}, + cache: &stubGatewayCache{}, + openaiWSResolver: NewOpenAIWSProtocolResolver(cfg), + toolCorrector: NewCodexToolCorrector(), + openaiWSPool: pool, + } + account := &Account{ + ID: 115, + Name: "openai-ingress-omit-model", + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Credentials: map[string]any{ + "api_key": "sk-test", + "model_mapping": map[string]any{ + "client-model": "gpt-5.1", + }, + }, + Extra: map[string]any{ + "responses_websockets_v2_enabled": true, + }, + } + + serverErrCh := make(chan error, 1) + wsServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { + conn, err := coderws.Accept(w, r, &coderws.AcceptOptions{ + CompressionMode: coderws.CompressionContextTakeover, + }) + if err != nil { + serverErrCh <- err + return + } + defer func() { + _ = conn.CloseNow() + }() + + rec := httptest.NewRecorder() + ginCtx, _ := gin.CreateTestContext(rec) + req := r.Clone(r.Context()) + req.Header = req.Header.Clone() + req.Header.Set("User-Agent", "unit-test-agent/1.0") + ginCtx.Request = req + + readCtx, cancel := context.WithTimeout(r.Context(), 3*time.Second) + msgType, firstMessage, readErr := conn.Read(readCtx) + cancel() + if readErr != nil { + serverErrCh <- readErr + return + } + if msgType != coderws.MessageText && msgType != coderws.MessageBinary { + serverErrCh <- errors.New("unsupported websocket client message type") + return + } + + serverErrCh <- svc.ProxyResponsesWebSocketFromClient(r.Context(), ginCtx, conn, account, "sk-test", firstMessage, nil) + })) + defer wsServer.Close() + + dialCtx, cancelDial := context.WithTimeout(context.Background(), 3*time.Second) + clientConn, _, err := coderws.Dial(dialCtx, "ws"+strings.TrimPrefix(wsServer.URL, "http"), nil) + cancelDial() + require.NoError(t, err) + defer func() { + _ = clientConn.CloseNow() + }() + + writeCtx, cancelWrite := context.WithTimeout(context.Background(), 3*time.Second) + err = clientConn.Write(writeCtx, coderws.MessageText, []byte(`{"type":"response.create","model":"client-model","stream":false}`)) + cancelWrite() + require.NoError(t, err) + + readCtx, cancelRead := context.WithTimeout(context.Background(), 3*time.Second) + _, firstEvent, readErr := clientConn.Read(readCtx) + cancelRead() + require.NoError(t, readErr) + require.Equal(t, "resp_omit_model_1", gjson.GetBytes(firstEvent, "response.id").String()) + + writeCtx, cancelWrite = context.WithTimeout(context.Background(), 3*time.Second) + err = clientConn.Write(writeCtx, coderws.MessageText, []byte(`{"type":"response.create","stream":false,"previous_response_id":"resp_omit_model_1"}`)) + cancelWrite() + require.NoError(t, err) + + readCtx, cancelRead = context.WithTimeout(context.Background(), 3*time.Second) + _, secondEvent, readErr := clientConn.Read(readCtx) + cancelRead() + require.NoError(t, readErr) + require.Equal(t, "resp_omit_model_2", gjson.GetBytes(secondEvent, "response.id").String()) + _ = clientConn.Close(coderws.StatusNormalClosure, "done") + + select { + case serverErr := <-serverErrCh: + require.NoError(t, serverErr) + case <-time.After(5 * time.Second): + t.Fatal("等待 ingress websocket 结束超时") + } + + require.Len(t, captureConn.writes, 2) + require.Equal(t, "gpt-5.1", gjson.Get(requestToJSONString(captureConn.writes[0]), "model").String()) + require.Equal(t, "gpt-5.1", gjson.Get(requestToJSONString(captureConn.writes[1]), "model").String()) + require.Equal(t, "resp_omit_model_1", gjson.Get(requestToJSONString(captureConn.writes[1]), "previous_response_id").String()) +} + func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_DedicatedModeDoesNotReuseConnAcrossSessions(t *testing.T) { gin.SetMode(gin.TestMode) @@ -441,6 +575,124 @@ func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_PassthroughModeR require.Len(t, upstreamConn.writes, 1, "passthrough 模式应透传首条 response.create") } +func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_PassthroughHeadersUsePromptCacheAndTurnState(t *testing.T) { + gin.SetMode(gin.TestMode) + + cfg := &config.Config{} + cfg.Security.URLAllowlist.Enabled = false + cfg.Security.URLAllowlist.AllowInsecureHTTP = true + cfg.Gateway.OpenAIWS.Enabled = true + cfg.Gateway.OpenAIWS.OAuthEnabled = true + cfg.Gateway.OpenAIWS.APIKeyEnabled = true + cfg.Gateway.OpenAIWS.ResponsesWebsocketsV2 = true + cfg.Gateway.OpenAIWS.ModeRouterV2Enabled = true + cfg.Gateway.OpenAIWS.IngressModeDefault = OpenAIWSIngressModeCtxPool + cfg.Gateway.OpenAIWS.DialTimeoutSeconds = 3 + cfg.Gateway.OpenAIWS.ReadTimeoutSeconds = 3 + cfg.Gateway.OpenAIWS.WriteTimeoutSeconds = 3 + + upstreamConn := &openAIWSCaptureConn{ + events: [][]byte{ + []byte(`{"type":"response.completed","response":{"id":"resp_passthrough_headers","model":"gpt-5.1","usage":{"input_tokens":1,"output_tokens":1}}}`), + }, + } + captureDialer := &openAIWSCaptureDialer{conn: upstreamConn} + svc := &OpenAIGatewayService{ + cfg: cfg, + httpUpstream: &httpUpstreamRecorder{}, + cache: &stubGatewayCache{}, + openaiWSResolver: NewOpenAIWSProtocolResolver(cfg), + toolCorrector: NewCodexToolCorrector(), + openaiWSPassthroughDialer: captureDialer, + } + account := &Account{ + ID: 453, + Name: "openai-ingress-passthrough-headers", + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Credentials: map[string]any{ + "access_token": "oauth-token", + }, + Extra: map[string]any{ + "openai_oauth_responses_websockets_v2_mode": OpenAIWSIngressModePassthrough, + }, + } + + serverErrCh := make(chan error, 1) + wsServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { + conn, err := coderws.Accept(w, r, &coderws.AcceptOptions{ + CompressionMode: coderws.CompressionContextTakeover, + }) + if err != nil { + serverErrCh <- err + return + } + defer func() { + _ = conn.CloseNow() + }() + + rec := httptest.NewRecorder() + ginCtx, _ := gin.CreateTestContext(rec) + req := r.Clone(r.Context()) + req.Header = req.Header.Clone() + req.Header.Set("User-Agent", "codex_cli_rs/0.98.0") + req.Header.Set(openAIWSTurnStateHeader, "turn-state-1") + req.Header.Set(openAIWSTurnMetadataHeader, "turn-meta-1") + ginCtx.Request = req + + readCtx, cancel := context.WithTimeout(r.Context(), 3*time.Second) + msgType, firstMessage, readErr := conn.Read(readCtx) + cancel() + if readErr != nil { + serverErrCh <- readErr + return + } + if msgType != coderws.MessageText && msgType != coderws.MessageBinary { + serverErrCh <- errors.New("unsupported websocket client message type") + return + } + + serverErrCh <- svc.ProxyResponsesWebSocketFromClient(r.Context(), ginCtx, conn, account, "oauth-token", firstMessage, nil) + })) + defer wsServer.Close() + + dialCtx, cancelDial := context.WithTimeout(context.Background(), 3*time.Second) + clientConn, _, err := coderws.Dial(dialCtx, "ws"+strings.TrimPrefix(wsServer.URL, "http"), nil) + cancelDial() + require.NoError(t, err) + defer func() { + _ = clientConn.CloseNow() + }() + + writeCtx, cancelWrite := context.WithTimeout(context.Background(), 3*time.Second) + err = clientConn.Write(writeCtx, coderws.MessageText, []byte(`{"type":"response.create","model":"gpt-5.1","stream":false,"prompt_cache_key":"pcache_passthrough"}`)) + cancelWrite() + require.NoError(t, err) + + readCtx, cancelRead := context.WithTimeout(context.Background(), 3*time.Second) + _, event, readErr := clientConn.Read(readCtx) + cancelRead() + require.NoError(t, readErr) + require.Equal(t, "resp_passthrough_headers", gjson.GetBytes(event, "response.id").String()) + _ = clientConn.Close(coderws.StatusNormalClosure, "done") + + select { + case serverErr := <-serverErrCh: + if serverErr != nil { + require.Contains(t, serverErr.Error(), "StatusNormalClosure") + } + case <-time.After(5 * time.Second): + t.Fatal("等待 passthrough websocket 结束超时") + } + + require.Equal(t, isolateOpenAISessionID(0, "pcache_passthrough"), captureDialer.lastHeaders.Get("session_id")) + require.Equal(t, "turn-state-1", captureDialer.lastHeaders.Get(openAIWSTurnStateHeader)) + require.Equal(t, "turn-meta-1", captureDialer.lastHeaders.Get(openAIWSTurnMetadataHeader)) +} + func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_ModeOffReturnsPolicyViolation(t *testing.T) { gin.SetMode(gin.TestMode) diff --git a/backend/internal/service/openai_ws_forwarder_success_test.go b/backend/internal/service/openai_ws_forwarder_success_test.go index cd816533..e949560f 100644 --- a/backend/internal/service/openai_ws_forwarder_success_test.go +++ b/backend/internal/service/openai_ws_forwarder_success_test.go @@ -727,6 +727,70 @@ func TestOpenAIGatewayService_Forward_WSv2_HeaderSessionFallbackFromPromptCacheK require.True(t, gjson.Get(requestToJSONString(captureConn.lastWrite), "stream").Exists()) } +func TestOpenAIGatewayService_Forward_WSv2_ResponseDoneUsageParsed(t *testing.T) { + gin.SetMode(gin.TestMode) + + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/openai/v1/responses", nil) + c.Request.Header.Set("User-Agent", "unit-test-agent/1.0") + + cfg := &config.Config{} + cfg.Security.URLAllowlist.Enabled = false + cfg.Security.URLAllowlist.AllowInsecureHTTP = true + cfg.Gateway.OpenAIWS.Enabled = true + cfg.Gateway.OpenAIWS.OAuthEnabled = true + cfg.Gateway.OpenAIWS.APIKeyEnabled = true + cfg.Gateway.OpenAIWS.ResponsesWebsocketsV2 = true + cfg.Gateway.OpenAIWS.MaxConnsPerAccount = 1 + cfg.Gateway.OpenAIWS.MinIdlePerAccount = 0 + cfg.Gateway.OpenAIWS.MaxIdlePerAccount = 1 + + captureConn := &openAIWSCaptureConn{ + events: [][]byte{ + []byte(`{"type":"response.done","response":{"id":"resp_done_usage","model":"gpt-5.1","usage":{"input_tokens":13,"output_tokens":8,"input_tokens_details":{"cached_tokens":5},"cache_creation_input_tokens":2,"output_tokens_details":{"image_tokens":4}}}}`), + }, + } + captureDialer := &openAIWSCaptureDialer{conn: captureConn} + pool := newOpenAIWSConnPool(cfg) + pool.setClientDialerForTest(captureDialer) + + svc := &OpenAIGatewayService{ + cfg: cfg, + httpUpstream: &httpUpstreamRecorder{}, + cache: &stubGatewayCache{}, + openaiWSResolver: NewOpenAIWSProtocolResolver(cfg), + toolCorrector: NewCodexToolCorrector(), + openaiWSPool: pool, + } + account := &Account{ + ID: 32, + Name: "openai-ws-done", + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Credentials: map[string]any{ + "api_key": "sk-test", + }, + Extra: map[string]any{ + "responses_websockets_v2_enabled": true, + }, + } + + body := []byte(`{"model":"gpt-5.1","stream":false,"input":[{"type":"input_text","text":"hi"}]}`) + result, err := svc.Forward(context.Background(), c, account, body) + require.NoError(t, err) + require.NotNil(t, result) + require.Equal(t, "resp_done_usage", result.RequestID) + require.Equal(t, 13, result.Usage.InputTokens) + require.Equal(t, 8, result.Usage.OutputTokens) + require.Equal(t, 5, result.Usage.CacheReadInputTokens) + require.Equal(t, 2, result.Usage.CacheCreationInputTokens) + require.Equal(t, 4, result.Usage.ImageOutputTokens) +} + func TestOpenAIGatewayService_Forward_WSv1_Unsupported(t *testing.T) { gin.SetMode(gin.TestMode) diff --git a/backend/internal/service/openai_ws_v2/passthrough_relay.go b/backend/internal/service/openai_ws_v2/passthrough_relay.go index 35c7569d..6aba3b7d 100644 --- a/backend/internal/service/openai_ws_v2/passthrough_relay.go +++ b/backend/internal/service/openai_ws_v2/passthrough_relay.go @@ -25,6 +25,7 @@ type Usage struct { OutputTokens int CacheCreationInputTokens int CacheReadInputTokens int + ImageOutputTokens int } type RelayResult struct { @@ -756,8 +757,21 @@ func parseUsageAndAccumulate( } inputResult := gjson.GetBytes(message, "response.usage.input_tokens") + if !inputResult.Exists() { + inputResult = gjson.GetBytes(message, "response.usage.prompt_tokens") + } outputResult := gjson.GetBytes(message, "response.usage.output_tokens") + if !outputResult.Exists() { + outputResult = gjson.GetBytes(message, "response.usage.completion_tokens") + } cachedResult := gjson.GetBytes(message, "response.usage.input_tokens_details.cached_tokens") + if !cachedResult.Exists() { + cachedResult = gjson.GetBytes(message, "response.usage.prompt_tokens_details.cached_tokens") + } + imageTokens := usageResult.Get("output_tokens_details.image_tokens").Int() + if imageTokens == 0 { + imageTokens = usageResult.Get("completion_tokens_details.image_tokens").Int() + } inputTokens, inputOK := parseUsageIntField(inputResult, true) outputTokens, outputOK := parseUsageIntField(outputResult, true) @@ -771,14 +785,18 @@ func parseUsageAndAccumulate( return Usage{} } parsedUsage := Usage{ - InputTokens: inputTokens, - OutputTokens: outputTokens, - CacheReadInputTokens: cachedTokens, + InputTokens: inputTokens, + OutputTokens: outputTokens, + CacheCreationInputTokens: int(usageResult.Get("cache_creation_input_tokens").Int()), + CacheReadInputTokens: cachedTokens, + ImageOutputTokens: int(imageTokens), } state.usage.InputTokens += parsedUsage.InputTokens state.usage.OutputTokens += parsedUsage.OutputTokens + state.usage.CacheCreationInputTokens += parsedUsage.CacheCreationInputTokens state.usage.CacheReadInputTokens += parsedUsage.CacheReadInputTokens + state.usage.ImageOutputTokens += parsedUsage.ImageOutputTokens return parsedUsage } @@ -840,7 +858,7 @@ func isTerminalEvent(eventType string) bool { func shouldParseUsage(eventType string) bool { switch eventType { - case "response.completed", "response.done", "response.failed": + case "response.completed", "response.done", "response.failed", "response.incomplete", "response.cancelled", "response.canceled": return true default: return false diff --git a/backend/internal/service/openai_ws_v2/passthrough_relay_internal_test.go b/backend/internal/service/openai_ws_v2/passthrough_relay_internal_test.go index 52104482..13c51f66 100644 --- a/backend/internal/service/openai_ws_v2/passthrough_relay_internal_test.go +++ b/backend/internal/service/openai_ws_v2/passthrough_relay_internal_test.go @@ -300,20 +300,41 @@ func TestParseUsageAndEnrichCoverage(t *testing.T) { require.Equal(t, 0, state.usage.OutputTokens) require.Equal(t, 0, state.usage.CacheReadInputTokens) - parseUsageAndAccumulate(state, []byte(`{"type":"response.completed","response":{"usage":{"input_tokens":2,"output_tokens":1,"input_tokens_details":{"cached_tokens":1}}}}`), "response.completed", nil) + parseUsageAndAccumulate(state, []byte(`{"type":"response.completed","response":{"usage":{"input_tokens":2,"output_tokens":1,"input_tokens_details":{"cached_tokens":1},"cache_creation_input_tokens":4,"output_tokens_details":{"image_tokens":3}}}}`), "response.completed", nil) require.Equal(t, 2, state.usage.InputTokens) require.Equal(t, 1, state.usage.OutputTokens) require.Equal(t, 1, state.usage.CacheReadInputTokens) + require.Equal(t, 4, state.usage.CacheCreationInputTokens) + require.Equal(t, 3, state.usage.ImageOutputTokens) result := &RelayResult{} enrichResult(result, state, 5*time.Millisecond) require.Equal(t, state.usage.InputTokens, result.Usage.InputTokens) + require.Equal(t, state.usage.CacheCreationInputTokens, result.Usage.CacheCreationInputTokens) + require.Equal(t, state.usage.ImageOutputTokens, result.Usage.ImageOutputTokens) require.Equal(t, 5*time.Millisecond, result.Duration) parseUsageAndAccumulate(state, []byte(`{"type":"response.in_progress","response":{"usage":{"input_tokens":9}}}`), "response.in_progress", nil) require.Equal(t, 2, state.usage.InputTokens) enrichResult(nil, state, 0) } +func TestParseUsageAndAccumulateAcceptsChatUsageAliases(t *testing.T) { + t.Parallel() + + state := &relayState{} + got := parseUsageAndAccumulate( + state, + []byte(`{"type":"response.done","response":{"usage":{"prompt_tokens":12,"completion_tokens":6,"prompt_tokens_details":{"cached_tokens":4},"completion_tokens_details":{"image_tokens":2}}}}`), + "response.done", + nil, + ) + require.Equal(t, 12, got.InputTokens) + require.Equal(t, 6, got.OutputTokens) + require.Equal(t, 4, got.CacheReadInputTokens) + require.Equal(t, 2, got.ImageOutputTokens) + require.Equal(t, got, state.usage) +} + func TestEmitTurnCompleteCoverage(t *testing.T) { t.Parallel() @@ -377,6 +398,23 @@ func TestIsTokenEventCoverageBranches(t *testing.T) { require.True(t, isTokenEvent("response.done")) } +func TestShouldParseUsageTerminalEvents(t *testing.T) { + t.Parallel() + + for _, eventType := range []string{ + "response.completed", + "response.done", + "response.failed", + "response.incomplete", + "response.cancelled", + "response.canceled", + } { + require.True(t, shouldParseUsage(eventType), eventType) + } + require.False(t, shouldParseUsage("response.output_text.delta")) + require.False(t, shouldParseUsage("")) +} + func TestRelayTurnTimingHelpersCoverage(t *testing.T) { t.Parallel() diff --git a/backend/internal/service/openai_ws_v2_passthrough_adapter.go b/backend/internal/service/openai_ws_v2_passthrough_adapter.go index 17543dc0..c93d0981 100644 --- a/backend/internal/service/openai_ws_v2_passthrough_adapter.go +++ b/backend/internal/service/openai_ws_v2_passthrough_adapter.go @@ -312,6 +312,7 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough( // goroutine）和 OnTurnComplete / final result（runUpstreamToClient // goroutine）之间同步当前 turn 的 usage metadata。 usageMeta.initFromFirstFrame(firstClientMessage) + promptCacheKey := strings.TrimSpace(gjson.GetBytes(firstClientMessage, "prompt_cache_key").String()) wsURL, err := s.buildOpenAIResponsesWSURL(account) if err != nil { @@ -338,7 +339,13 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough( if s.cfg != nil && s.cfg.Gateway.ForceCodexCLI { isCodexCLI = true } - headers, _ := s.buildOpenAIWSHeaders(c, account, token, wsDecision, isCodexCLI, "", "", "") + turnState := "" + turnMetadata := "" + if c != nil { + turnState = strings.TrimSpace(c.GetHeader(openAIWSTurnStateHeader)) + turnMetadata = strings.TrimSpace(c.GetHeader(openAIWSTurnMetadataHeader)) + } + headers, _ := s.buildOpenAIWSHeaders(c, account, token, wsDecision, isCodexCLI, turnState, turnMetadata, promptCacheKey) proxyURL := "" if account.ProxyID != nil && account.Proxy != nil { proxyURL = account.Proxy.URL() @@ -519,6 +526,7 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough( OutputTokens: turn.Usage.OutputTokens, CacheCreationInputTokens: turn.Usage.CacheCreationInputTokens, CacheReadInputTokens: turn.Usage.CacheReadInputTokens, + ImageOutputTokens: turn.Usage.ImageOutputTokens, }, Model: turn.RequestModel, ServiceTier: usageMeta.serviceTier.Load(), @@ -593,6 +601,7 @@ func (s *OpenAIGatewayService) proxyResponsesWebSocketV2Passthrough( OutputTokens: relayResult.Usage.OutputTokens, CacheCreationInputTokens: relayResult.Usage.CacheCreationInputTokens, CacheReadInputTokens: relayResult.Usage.CacheReadInputTokens, + ImageOutputTokens: relayResult.Usage.ImageOutputTokens, }, Model: relayResult.RequestModel, ServiceTier: usageMeta.serviceTier.Load(), From 27600b1d2c9579f83abbb1ee469bfc45b908f840 Mon Sep 17 00:00:00 2001 From: Pluviobyte Date: Thu, 28 May 2026 05:40:50 +0000 Subject: [PATCH 06/20] fix(gateway): filter count_tokens generation fields Anthropic count_tokens rejects generation-only fields such as temperature, top_p, top_k, stream, and stop sequences. Passing the original messages payload through unchanged can turn otherwise valid requests into upstream 400 errors. Sanitize only the count_tokens upstream payload after the gateway's existing request normalization, preserving fields that existing compatibility paths rely on while removing parameters the count_tokens endpoint does not accept. Fixes #2764 Co-authored-by: Cursor --- ...teway_anthropic_apikey_passthrough_test.go | 60 +++++++++++++++++++ backend/internal/service/gateway_service.go | 21 +++++++ 2 files changed, 81 insertions(+) diff --git a/backend/internal/service/gateway_anthropic_apikey_passthrough_test.go b/backend/internal/service/gateway_anthropic_apikey_passthrough_test.go index 5cb03f30..9062c517 100644 --- a/backend/internal/service/gateway_anthropic_apikey_passthrough_test.go +++ b/backend/internal/service/gateway_anthropic_apikey_passthrough_test.go @@ -476,6 +476,66 @@ func TestGatewayService_AnthropicAPIKeyPassthrough_ModelMappingPreservesOtherFie require.Equal(t, int64(1024), gjson.GetBytes(sentBody, "max_tokens").Int(), "max_tokens 不应被修改") } +func TestGatewayService_AnthropicAPIKeyPassthrough_CountTokensFiltersGenerationFields(t *testing.T) { + gin.SetMode(gin.TestMode) + + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + c.Request = httptest.NewRequest(http.MethodPost, "/v1/messages/count_tokens", nil) + + body := []byte(`{"model":"claude-sonnet-4-20250514","system":[{"type":"text","text":"sys"}],"messages":[{"role":"user","content":"hello"}],"tools":[{"name":"tool","input_schema":{"type":"object"}}],"temperature":0.7,"top_p":0.9,"top_k":40,"stream":true,"stop_sequences":["END"],"max_tokens":1024,"thinking":{"type":"enabled","budget_tokens":5000}}`) + parsed := &ParsedRequest{ + Body: body, + Model: "claude-sonnet-4-20250514", + } + + upstreamRespBody := `{"input_tokens":42}` + upstream := &anthropicHTTPUpstreamRecorder{ + resp: &http.Response{ + StatusCode: http.StatusOK, + Header: http.Header{"Content-Type": []string{"application/json"}}, + Body: io.NopCloser(strings.NewReader(upstreamRespBody)), + }, + } + + svc := &GatewayService{ + cfg: &config.Config{Gateway: config.GatewayConfig{MaxLineSize: defaultMaxLineSize}}, + httpUpstream: upstream, + rateLimitService: &RateLimitService{}, + } + + account := &Account{ + ID: 302, + Name: "count-token-filter-test", + Platform: PlatformAnthropic, + Type: AccountTypeAPIKey, + Concurrency: 1, + Credentials: map[string]any{ + "api_key": "upstream-key", + "base_url": "https://api.anthropic.com", + }, + Extra: map[string]any{"anthropic_passthrough": true}, + Status: StatusActive, + Schedulable: true, + } + + err := svc.ForwardCountTokens(context.Background(), c, account, parsed) + require.NoError(t, err) + + sentBody := upstream.lastBody + require.False(t, gjson.GetBytes(sentBody, "temperature").Exists()) + require.False(t, gjson.GetBytes(sentBody, "top_p").Exists()) + require.False(t, gjson.GetBytes(sentBody, "top_k").Exists()) + require.False(t, gjson.GetBytes(sentBody, "stream").Exists()) + require.False(t, gjson.GetBytes(sentBody, "stop_sequences").Exists()) + require.Equal(t, "claude-sonnet-4-20250514", gjson.GetBytes(sentBody, "model").String()) + require.Equal(t, "sys", gjson.GetBytes(sentBody, "system.0.text").String()) + require.Equal(t, "hello", gjson.GetBytes(sentBody, "messages.0.content").String()) + require.Equal(t, "tool", gjson.GetBytes(sentBody, "tools.0.name").String()) + require.Equal(t, int64(1024), gjson.GetBytes(sentBody, "max_tokens").Int()) + require.Equal(t, "enabled", gjson.GetBytes(sentBody, "thinking.type").String()) +} + // TestGatewayService_AnthropicAPIKeyPassthrough_EmptyModelSkipsMapping // 确保空模型名不会触发映射逻辑 func TestGatewayService_AnthropicAPIKeyPassthrough_EmptyModelSkipsMapping(t *testing.T) { diff --git a/backend/internal/service/gateway_service.go b/backend/internal/service/gateway_service.go index 4a8175a4..a787e3eb 100644 --- a/backend/internal/service/gateway_service.go +++ b/backend/internal/service/gateway_service.go @@ -9311,6 +9311,7 @@ func (s *GatewayService) buildCountTokensRequestAnthropicAPIKeyPassthrough( } targetURL = validatedURL + "/v1/messages/count_tokens?beta=true" } + body = sanitizeCountTokensRequestBody(body) req, err := http.NewRequestWithContext(ctx, http.MethodPost, targetURL, bytes.NewReader(body)) if err != nil { @@ -9405,6 +9406,7 @@ func (s *GatewayService) buildCountTokensRequest(ctx context.Context, c *gin.Con if ctEnableCCH { body = signBillingHeaderCCH(body) } + body = sanitizeCountTokensRequestBody(body) req, err := http.NewRequestWithContext(ctx, "POST", targetURL, bytes.NewReader(body)) if err != nil { @@ -9501,6 +9503,25 @@ func (s *GatewayService) buildCountTokensRequest(ctx context.Context, c *gin.Con return req, nil } +func sanitizeCountTokensRequestBody(body []byte) []byte { + out := body + for _, path := range []string{ + "temperature", + "top_p", + "top_k", + "stream", + "stop_sequences", + "stop", + } { + if gjson.GetBytes(out, path).Exists() { + if next, ok := deleteJSONPathBytes(out, path); ok { + out = next + } + } + } + return out +} + // countTokensError 返回 count_tokens 错误响应 func (s *GatewayService) countTokensError(c *gin.Context, status int, errType, message string) { c.JSON(status, gin.H{ From b15375dfb4b3fa40aaec53012385106c8c29d3ae Mon Sep 17 00:00:00 2001 From: wucm667 Date: Thu, 28 May 2026 17:27:01 +0800 Subject: [PATCH 07/20] fix(admin): handle already up-to-date updates --- .../internal/handler/admin/system_handler.go | 26 +++- .../handler/admin/system_handler_test.go | 144 ++++++++++++++++++ backend/internal/service/update_service.go | 8 +- .../internal/service/update_service_test.go | 64 ++++++++ 4 files changed, 239 insertions(+), 3 deletions(-) create mode 100644 backend/internal/handler/admin/system_handler_test.go create mode 100644 backend/internal/service/update_service_test.go diff --git a/backend/internal/handler/admin/system_handler.go b/backend/internal/handler/admin/system_handler.go index 3e2022c7..fb6c0ef7 100644 --- a/backend/internal/handler/admin/system_handler.go +++ b/backend/internal/handler/admin/system_handler.go @@ -2,6 +2,7 @@ package admin import ( "context" + "errors" "net/http" "strconv" "strings" @@ -17,12 +18,18 @@ import ( // SystemHandler handles system-related operations type SystemHandler struct { - updateSvc *service.UpdateService + updateSvc systemUpdateService lockSvc *service.SystemOperationLockService } +type systemUpdateService interface { + CheckUpdate(ctx context.Context, force bool) (*service.UpdateInfo, error) + PerformUpdate(ctx context.Context) error + Rollback() error +} + // NewSystemHandler creates a new SystemHandler -func NewSystemHandler(updateSvc *service.UpdateService, lockSvc *service.SystemOperationLockService) *SystemHandler { +func NewSystemHandler(updateSvc systemUpdateService, lockSvc *service.SystemOperationLockService) *SystemHandler { return &SystemHandler{ updateSvc: updateSvc, lockSvc: lockSvc, @@ -67,6 +74,21 @@ func (h *SystemHandler) PerformUpdate(c *gin.Context) { }() if err := h.updateSvc.PerformUpdate(ctx); err != nil { + if errors.Is(err, service.ErrNoUpdateAvailable) { + info, checkErr := h.updateSvc.CheckUpdate(ctx, false) + if checkErr != nil { + releaseReason = "SYSTEM_UPDATE_FAILED" + return nil, checkErr + } + succeeded = true + return gin.H{ + "message": "Already up to date", + "already_up_to_date": true, + "current_version": info.CurrentVersion, + "latest_version": info.LatestVersion, + "operation_id": lock.OperationID(), + }, nil + } releaseReason = "SYSTEM_UPDATE_FAILED" return nil, err } diff --git a/backend/internal/handler/admin/system_handler_test.go b/backend/internal/handler/admin/system_handler_test.go new file mode 100644 index 00000000..0f33a452 --- /dev/null +++ b/backend/internal/handler/admin/system_handler_test.go @@ -0,0 +1,144 @@ +//go:build unit + +package admin + +import ( + "context" + "encoding/json" + "errors" + "net/http" + "net/http/httptest" + "testing" + "time" + + "github.com/Wei-Shaw/sub2api/internal/service" + "github.com/gin-gonic/gin" + "github.com/stretchr/testify/require" +) + +type systemHandlerUpdateServiceStub struct { + performErr error + updateInfo *service.UpdateInfo + checkErr error + checkForces []bool + performCall int +} + +func (s *systemHandlerUpdateServiceStub) CheckUpdate(_ context.Context, force bool) (*service.UpdateInfo, error) { + s.checkForces = append(s.checkForces, force) + return s.updateInfo, s.checkErr +} + +func (s *systemHandlerUpdateServiceStub) PerformUpdate(context.Context) error { + s.performCall++ + return s.performErr +} + +func (s *systemHandlerUpdateServiceStub) Rollback() error { + return nil +} + +type systemUpdateResponseEnvelope struct { + Code int `json:"code"` + Message string `json:"message"` + Data struct { + Message string `json:"message"` + AlreadyUpToDate bool `json:"already_up_to_date"` + CurrentVersion string `json:"current_version"` + LatestVersion string `json:"latest_version"` + OperationID string `json:"operation_id"` + } `json:"data"` +} + +type systemUpdateErrorEnvelope struct { + Code int `json:"code"` + Message string `json:"message"` +} + +func newSystemHandlerTestRouter(t *testing.T, updateSvc *systemHandlerUpdateServiceStub, repo *memoryIdempotencyRepoStub) *gin.Engine { + t.Helper() + gin.SetMode(gin.TestMode) + service.SetDefaultIdempotencyCoordinator(nil) + t.Cleanup(func() { + service.SetDefaultIdempotencyCoordinator(nil) + }) + + lockSvc := service.NewSystemOperationLockService(repo, service.IdempotencyConfig{ + ProcessingTimeout: time.Second, + SystemOperationTTL: time.Minute, + }) + handler := NewSystemHandler(updateSvc, lockSvc) + + router := gin.New() + router.POST("/api/v1/admin/system/update", handler.PerformUpdate) + return router +} + +func requireSystemLockStatus(t *testing.T, repo *memoryIdempotencyRepoStub, wantStatus string) { + t.Helper() + repo.mu.Lock() + defer repo.mu.Unlock() + + for _, record := range repo.data { + if record.Status == wantStatus { + return + } + } + t.Fatalf("system lock status %q not found in records: %#v", wantStatus, repo.data) +} + +func TestSystemHandlerPerformUpdateAlreadyUpToDateReturnsOK(t *testing.T) { + updateSvc := &systemHandlerUpdateServiceStub{ + performErr: service.ErrNoUpdateAvailable, + updateInfo: &service.UpdateInfo{ + CurrentVersion: "0.1.132", + LatestVersion: "0.1.132", + HasUpdate: false, + }, + } + repo := newMemoryIdempotencyRepoStub() + router := newSystemHandlerTestRouter(t, updateSvc, repo) + + rec := httptest.NewRecorder() + req := httptest.NewRequest(http.MethodPost, "/api/v1/admin/system/update", nil) + req.Header.Set("Idempotency-Key", "already-up-to-date") + router.ServeHTTP(rec, req) + + require.Equal(t, http.StatusOK, rec.Code) + require.Equal(t, 1, updateSvc.performCall) + require.Equal(t, []bool{false}, updateSvc.checkForces) + requireSystemLockStatus(t, repo, service.IdempotencyStatusSucceeded) + + var body systemUpdateResponseEnvelope + require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &body)) + require.Equal(t, 0, body.Code) + require.Equal(t, "success", body.Message) + require.Equal(t, "Already up to date", body.Data.Message) + require.True(t, body.Data.AlreadyUpToDate) + require.Equal(t, "0.1.132", body.Data.CurrentVersion) + require.Equal(t, "0.1.132", body.Data.LatestVersion) + require.NotEmpty(t, body.Data.OperationID) +} + +func TestSystemHandlerPerformUpdateFailureStillReturnsInternalError(t *testing.T) { + updateSvc := &systemHandlerUpdateServiceStub{ + performErr: errors.New("download failed"), + } + repo := newMemoryIdempotencyRepoStub() + router := newSystemHandlerTestRouter(t, updateSvc, repo) + + rec := httptest.NewRecorder() + req := httptest.NewRequest(http.MethodPost, "/api/v1/admin/system/update", nil) + req.Header.Set("Idempotency-Key", "real-failure") + router.ServeHTTP(rec, req) + + require.Equal(t, http.StatusInternalServerError, rec.Code) + require.Equal(t, 1, updateSvc.performCall) + require.Empty(t, updateSvc.checkForces) + requireSystemLockStatus(t, repo, service.IdempotencyStatusFailedRetryable) + + var body systemUpdateErrorEnvelope + require.NoError(t, json.Unmarshal(rec.Body.Bytes(), &body)) + require.Equal(t, http.StatusInternalServerError, body.Code) + require.Equal(t, "internal error", body.Message) +} diff --git a/backend/internal/service/update_service.go b/backend/internal/service/update_service.go index 34ad4610..de8c5e16 100644 --- a/backend/internal/service/update_service.go +++ b/backend/internal/service/update_service.go @@ -17,6 +17,12 @@ import ( "strconv" "strings" "time" + + infraerrors "github.com/Wei-Shaw/sub2api/internal/pkg/errors" +) + +var ( + ErrNoUpdateAvailable = infraerrors.Conflict("ALREADY_UP_TO_DATE", "no update available; current version is latest") ) const ( @@ -146,7 +152,7 @@ func (s *UpdateService) PerformUpdate(ctx context.Context) error { } if !info.HasUpdate { - return fmt.Errorf("no update available") + return ErrNoUpdateAvailable } // Find matching archive and checksum for current platform diff --git a/backend/internal/service/update_service_test.go b/backend/internal/service/update_service_test.go new file mode 100644 index 00000000..8d8310d4 --- /dev/null +++ b/backend/internal/service/update_service_test.go @@ -0,0 +1,64 @@ +//go:build unit + +package service + +import ( + "context" + "errors" + "testing" + "time" + + "github.com/stretchr/testify/require" +) + +type updateServiceCacheStub struct { + data string +} + +func (s *updateServiceCacheStub) GetUpdateInfo(context.Context) (string, error) { + if s.data == "" { + return "", errors.New("cache miss") + } + return s.data, nil +} + +func (s *updateServiceCacheStub) SetUpdateInfo(_ context.Context, data string, _ time.Duration) error { + s.data = data + return nil +} + +type updateServiceGitHubClientStub struct { + release *GitHubRelease +} + +func (s *updateServiceGitHubClientStub) FetchLatestRelease(context.Context, string) (*GitHubRelease, error) { + return s.release, nil +} + +func (s *updateServiceGitHubClientStub) DownloadFile(context.Context, string, string, int64) error { + panic("DownloadFile should not be called when no update is available") +} + +func (s *updateServiceGitHubClientStub) FetchChecksumFile(context.Context, string) ([]byte, error) { + panic("FetchChecksumFile should not be called when no update is available") +} + +func TestUpdateServicePerformUpdateNoUpdateReturnsSentinel(t *testing.T) { + svc := NewUpdateService( + &updateServiceCacheStub{}, + &updateServiceGitHubClientStub{ + release: &GitHubRelease{ + TagName: "v0.1.132", + Name: "v0.1.132", + }, + }, + "0.1.132", + "release", + ) + + err := svc.PerformUpdate(context.Background()) + + require.Error(t, err) + require.True(t, errors.Is(err, ErrNoUpdateAvailable)) + require.ErrorIs(t, err, ErrNoUpdateAvailable) +} From 56e96fdd8c0fbb42c833c15808d953a08d3fa40e Mon Sep 17 00:00:00 2001 From: gaoren002 Date: Thu, 28 May 2026 10:03:41 +0000 Subject: [PATCH 08/20] fix: classify concurrency acquire failures --- .../handler/concurrency_error_response.go | 27 ++++++++ .../concurrency_error_response_test.go | 63 +++++++++++++++++++ backend/internal/handler/gateway_handler.go | 6 +- backend/internal/handler/gateway_helper.go | 3 + .../handler/gateway_helper_hotpath_test.go | 19 ++++++ .../handler/openai_gateway_handler.go | 6 +- 6 files changed, 118 insertions(+), 6 deletions(-) create mode 100644 backend/internal/handler/concurrency_error_response.go create mode 100644 backend/internal/handler/concurrency_error_response_test.go diff --git a/backend/internal/handler/concurrency_error_response.go b/backend/internal/handler/concurrency_error_response.go new file mode 100644 index 00000000..52abf735 --- /dev/null +++ b/backend/internal/handler/concurrency_error_response.go @@ -0,0 +1,27 @@ +package handler + +import ( + "context" + "errors" + "fmt" + "net/http" +) + +const statusClientClosedRequest = 499 + +func concurrencyErrorResponse(err error, slotType string) (int, string, string) { + var concurrencyErr *ConcurrencyError + if errors.As(err, &concurrencyErr) { + if concurrencyErr.SlotType != "" { + slotType = concurrencyErr.SlotType + } + return http.StatusTooManyRequests, "rate_limit_error", + fmt.Sprintf("Concurrency limit exceeded for %s, please retry later", slotType) + } + + if errors.Is(err, context.Canceled) { + return statusClientClosedRequest, "api_error", "context canceled" + } + + return http.StatusServiceUnavailable, "api_error", "Service temporarily unavailable, please retry later" +} diff --git a/backend/internal/handler/concurrency_error_response_test.go b/backend/internal/handler/concurrency_error_response_test.go new file mode 100644 index 00000000..a2e6b9ab --- /dev/null +++ b/backend/internal/handler/concurrency_error_response_test.go @@ -0,0 +1,63 @@ +package handler + +import ( + "context" + "errors" + "net/http" + "testing" + + "github.com/stretchr/testify/require" +) + +func TestConcurrencyErrorResponse(t *testing.T) { + tests := []struct { + name string + err error + slotType string + wantStatus int + wantType string + wantMessage string + }{ + { + name: "true concurrency timeout remains rate limit", + err: &ConcurrencyError{SlotType: "account", IsTimeout: true}, + slotType: "user", + wantStatus: http.StatusTooManyRequests, + wantType: "rate_limit_error", + wantMessage: "Concurrency limit exceeded for account, please retry later", + }, + { + name: "client cancellation is not classified as concurrency limit", + err: context.Canceled, + slotType: "user", + wantStatus: statusClientClosedRequest, + wantType: "api_error", + wantMessage: "context canceled", + }, + { + name: "deadline exceeded is service unavailable", + err: context.DeadlineExceeded, + slotType: "user", + wantStatus: http.StatusServiceUnavailable, + wantType: "api_error", + wantMessage: "Service temporarily unavailable, please retry later", + }, + { + name: "redis acquire error is service unavailable", + err: errors.New("redis unavailable"), + slotType: "user", + wantStatus: http.StatusServiceUnavailable, + wantType: "api_error", + wantMessage: "Service temporarily unavailable, please retry later", + }, + } + + for _, tt := range tests { + t.Run(tt.name, func(t *testing.T) { + status, errType, message := concurrencyErrorResponse(tt.err, tt.slotType) + require.Equal(t, tt.wantStatus, status) + require.Equal(t, tt.wantType, errType) + require.Equal(t, tt.wantMessage, message) + }) + } +} diff --git a/backend/internal/handler/gateway_handler.go b/backend/internal/handler/gateway_handler.go index 4695a791..a24611f9 100644 --- a/backend/internal/handler/gateway_handler.go +++ b/backend/internal/handler/gateway_handler.go @@ -1471,10 +1471,10 @@ func (h *GatewayHandler) calculateSubscriptionRemaining(group *service.Group, su return min } -// handleConcurrencyError handles concurrency-related errors with proper 429 response +// handleConcurrencyError handles concurrency-related acquire errors. func (h *GatewayHandler) handleConcurrencyError(c *gin.Context, err error, slotType string, streamStarted bool) { - h.handleStreamingAwareError(c, http.StatusTooManyRequests, "rate_limit_error", - fmt.Sprintf("Concurrency limit exceeded for %s, please retry later", slotType), streamStarted) + status, errType, message := concurrencyErrorResponse(err, slotType) + h.handleStreamingAwareError(c, status, errType, message, streamStarted) } func (h *GatewayHandler) handleFailoverExhausted(c *gin.Context, failoverErr *service.UpstreamFailoverError, platform string, streamStarted bool) { diff --git a/backend/internal/handler/gateway_helper.go b/backend/internal/handler/gateway_helper.go index 09e6c09b..e4897502 100644 --- a/backend/internal/handler/gateway_helper.go +++ b/backend/internal/handler/gateway_helper.go @@ -336,6 +336,9 @@ func (h *ConcurrencyHelper) waitForSlotWithPingTimeout(c *gin.Context, slotType for { select { case <-ctx.Done(): + if parentErr := c.Request.Context().Err(); parentErr != nil { + return nil, parentErr + } return nil, &ConcurrencyError{ SlotType: slotType, IsTimeout: true, diff --git a/backend/internal/handler/gateway_helper_hotpath_test.go b/backend/internal/handler/gateway_helper_hotpath_test.go index 4a677199..d57c396c 100644 --- a/backend/internal/handler/gateway_helper_hotpath_test.go +++ b/backend/internal/handler/gateway_helper_hotpath_test.go @@ -280,6 +280,25 @@ func TestWaitForSlotWithPingTimeout_TimeoutAndStreamPing(t *testing.T) { }) } +func TestWaitForSlotWithPingTimeout_ParentContextCanceled(t *testing.T) { + cache := &helperConcurrencyCacheStub{ + accountSeq: []bool{false}, + } + concurrency := service.NewConcurrencyService(cache) + helper := NewConcurrencyHelper(concurrency, SSEPingFormatNone, 5*time.Millisecond) + c, _ := newHelperTestContext(http.MethodPost, "/v1/messages") + reqCtx, cancel := context.WithCancel(c.Request.Context()) + c.Request = c.Request.WithContext(reqCtx) + cancel() + + streamStarted := false + release, err := helper.waitForSlotWithPingTimeout(c, "account", 101, 2, time.Second, false, &streamStarted, true) + require.Nil(t, release) + require.ErrorIs(t, err, context.Canceled) + var cErr *ConcurrencyError + require.False(t, errors.As(err, &cErr)) +} + func TestWaitForSlotWithPingTimeout_AcquireError(t *testing.T) { errCache := &helperConcurrencyCacheStubWithError{ err: errors.New("redis unavailable"), diff --git a/backend/internal/handler/openai_gateway_handler.go b/backend/internal/handler/openai_gateway_handler.go index a51eee86..2d0524cd 100644 --- a/backend/internal/handler/openai_gateway_handler.go +++ b/backend/internal/handler/openai_gateway_handler.go @@ -1685,10 +1685,10 @@ func (h *OpenAIGatewayHandler) acquireImageGenerationSlot(c *gin.Context, stream return nil, false } -// handleConcurrencyError handles concurrency-related errors with proper 429 response +// handleConcurrencyError handles concurrency-related acquire errors. func (h *OpenAIGatewayHandler) handleConcurrencyError(c *gin.Context, err error, slotType string, streamStarted bool) { - h.handleStreamingAwareError(c, http.StatusTooManyRequests, "rate_limit_error", - fmt.Sprintf("Concurrency limit exceeded for %s, please retry later", slotType), streamStarted) + status, errType, message := concurrencyErrorResponse(err, slotType) + h.handleStreamingAwareError(c, status, errType, message, streamStarted) } func (h *OpenAIGatewayHandler) handleFailoverExhausted(c *gin.Context, failoverErr *service.UpstreamFailoverError, streamStarted bool) { From 2bd3125d0fe23515fd42ee3c6651bbe38e47593c Mon Sep 17 00:00:00 2001 From: Wey Gu Date: Thu, 28 May 2026 22:08:02 +0800 Subject: [PATCH 09/20] Preserve usage request context --- backend/internal/handler/gateway_handler.go | 7 +-- .../gateway_handler_chat_completions.go | 2 +- .../handler/gateway_handler_responses.go | 2 +- .../internal/handler/gemini_v1beta_handler.go | 2 +- .../handler/openai_chat_completions.go | 2 +- backend/internal/handler/openai_embeddings.go | 2 +- .../handler/openai_gateway_handler.go | 44 +++++++++++++--- .../openai_gateway_usage_context_test.go | 41 +++++++++++++++ backend/internal/handler/openai_images.go | 2 +- .../handler/usage_record_submit_task_test.go | 24 ++++----- .../server/middleware/client_request_id.go | 6 ++- .../middleware/client_request_id_test.go | 50 +++++++++++++++++++ 12 files changed, 154 insertions(+), 30 deletions(-) create mode 100644 backend/internal/handler/openai_gateway_usage_context_test.go create mode 100644 backend/internal/server/middleware/client_request_id_test.go diff --git a/backend/internal/handler/gateway_handler.go b/backend/internal/handler/gateway_handler.go index 4695a791..a6749191 100644 --- a/backend/internal/handler/gateway_handler.go +++ b/backend/internal/handler/gateway_handler.go @@ -510,7 +510,7 @@ func (h *GatewayHandler) Messages(c *gin.Context) { // 使用量记录通过有界 worker 池提交，避免请求热路径创建无界 goroutine。 quotaPlatform := service.QuotaPlatform(c.Request.Context(), apiKey) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(c.Request.Context(), func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.RecordUsageInput{ Result: result, ParsedRequest: parsedReq, @@ -905,7 +905,7 @@ func (h *GatewayHandler) Messages(c *gin.Context) { // 使用量记录通过有界 worker 池提交，避免请求热路径创建无界 goroutine。 quotaPlatform := service.QuotaPlatform(c.Request.Context(), currentAPIKey) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(c.Request.Context(), func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.RecordUsageInput{ Result: result, ParsedRequest: parsedReq, @@ -2056,10 +2056,11 @@ func (h *GatewayHandler) maybeLogCompatibilityFallbackMetrics(reqLog *zap.Logger ) } -func (h *GatewayHandler) submitUsageRecordTask(task service.UsageRecordTask) { +func (h *GatewayHandler) submitUsageRecordTask(parent context.Context, task service.UsageRecordTask) { if task == nil { return } + task = wrapUsageRecordTaskContext(parent, task) if h.usageRecordWorkerPool != nil { h.usageRecordWorkerPool.Submit(task) return diff --git a/backend/internal/handler/gateway_handler_chat_completions.go b/backend/internal/handler/gateway_handler_chat_completions.go index acbdc261..daf6e6ea 100644 --- a/backend/internal/handler/gateway_handler_chat_completions.go +++ b/backend/internal/handler/gateway_handler_chat_completions.go @@ -292,7 +292,7 @@ func (h *GatewayHandler) ChatCompletions(c *gin.Context) { upstreamEndpoint := GetUpstreamEndpoint(c, account.Platform) quotaPlatform := service.QuotaPlatform(c.Request.Context(), apiKey) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(c.Request.Context(), func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.RecordUsageInput{ Result: result, QuotaPlatform: quotaPlatform, diff --git a/backend/internal/handler/gateway_handler_responses.go b/backend/internal/handler/gateway_handler_responses.go index 6a083f31..f57b9989 100644 --- a/backend/internal/handler/gateway_handler_responses.go +++ b/backend/internal/handler/gateway_handler_responses.go @@ -267,7 +267,7 @@ func (h *GatewayHandler) Responses(c *gin.Context) { upstreamEndpoint := GetUpstreamEndpoint(c, account.Platform) quotaPlatform := service.QuotaPlatform(c.Request.Context(), apiKey) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(c.Request.Context(), func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.RecordUsageInput{ Result: result, QuotaPlatform: quotaPlatform, diff --git a/backend/internal/handler/gemini_v1beta_handler.go b/backend/internal/handler/gemini_v1beta_handler.go index 27ea4404..0b33ca3e 100644 --- a/backend/internal/handler/gemini_v1beta_handler.go +++ b/backend/internal/handler/gemini_v1beta_handler.go @@ -528,7 +528,7 @@ func (h *GatewayHandler) GeminiV1BetaModels(c *gin.Context) { inboundEndpoint := GetInboundEndpoint(c) upstreamEndpoint := GetUpstreamEndpoint(c, account.Platform) quotaPlatform := service.QuotaPlatform(c.Request.Context(), apiKey) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(c.Request.Context(), func(ctx context.Context) { if err := h.gatewayService.RecordUsageWithLongContext(ctx, &service.RecordUsageLongContextInput{ Result: result, QuotaPlatform: quotaPlatform, diff --git a/backend/internal/handler/openai_chat_completions.go b/backend/internal/handler/openai_chat_completions.go index 17f0d47e..9f63ef1f 100644 --- a/backend/internal/handler/openai_chat_completions.go +++ b/backend/internal/handler/openai_chat_completions.go @@ -273,7 +273,7 @@ func (h *OpenAIGatewayHandler) ChatCompletions(c *gin.Context) { inboundEndpoint := GetInboundEndpoint(c) upstreamEndpoint := resolveRawCCUpstreamEndpoint(c, account) - h.submitOpenAIUsageRecordTask(result, func(ctx context.Context) { + h.submitOpenAIUsageRecordTask(c.Request.Context(), result, func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.OpenAIRecordUsageInput{ Result: result, APIKey: apiKey, diff --git a/backend/internal/handler/openai_embeddings.go b/backend/internal/handler/openai_embeddings.go index bbb67044..b64ac41d 100644 --- a/backend/internal/handler/openai_embeddings.go +++ b/backend/internal/handler/openai_embeddings.go @@ -220,7 +220,7 @@ func (h *OpenAIGatewayHandler) Embeddings(c *gin.Context) { inboundEndpoint := GetInboundEndpoint(c) upstreamEndpoint := GetUpstreamEndpoint(c, account.Platform) - h.submitOpenAIUsageRecordTask(result, func(ctx context.Context) { + h.submitOpenAIUsageRecordTask(c.Request.Context(), result, func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.OpenAIRecordUsageInput{ Result: result, APIKey: apiKey, diff --git a/backend/internal/handler/openai_gateway_handler.go b/backend/internal/handler/openai_gateway_handler.go index a51eee86..86503f30 100644 --- a/backend/internal/handler/openai_gateway_handler.go +++ b/backend/internal/handler/openai_gateway_handler.go @@ -12,6 +12,7 @@ import ( "time" "github.com/Wei-Shaw/sub2api/internal/config" + "github.com/Wei-Shaw/sub2api/internal/pkg/ctxkey" pkghttputil "github.com/Wei-Shaw/sub2api/internal/pkg/httputil" "github.com/Wei-Shaw/sub2api/internal/pkg/ip" "github.com/Wei-Shaw/sub2api/internal/pkg/logger" @@ -46,6 +47,31 @@ func resolveOpenAIMessagesDispatchMappedModel(apiKey *service.APIKey, requestedM return strings.TrimSpace(apiKey.Group.ResolveMessagesDispatchModel(requestedModel)) } +func usageRecordContext(parent context.Context, base context.Context) context.Context { + if base == nil { + base = context.Background() + } + if parent == nil { + return base + } + if clientRequestID, _ := parent.Value(ctxkey.ClientRequestID).(string); strings.TrimSpace(clientRequestID) != "" { + base = context.WithValue(base, ctxkey.ClientRequestID, strings.TrimSpace(clientRequestID)) + } + if requestID, _ := parent.Value(ctxkey.RequestID).(string); strings.TrimSpace(requestID) != "" { + base = context.WithValue(base, ctxkey.RequestID, strings.TrimSpace(requestID)) + } + return base +} + +func wrapUsageRecordTaskContext(parent context.Context, task service.UsageRecordTask) service.UsageRecordTask { + if task == nil { + return nil + } + return func(ctx context.Context) { + task(usageRecordContext(parent, ctx)) + } +} + // NewOpenAIGatewayHandler creates a new OpenAIGatewayHandler func NewOpenAIGatewayHandler( gatewayService *service.OpenAIGatewayService, @@ -437,7 +463,7 @@ func (h *OpenAIGatewayHandler) Responses(c *gin.Context) { upstreamEndpoint := GetUpstreamEndpoint(c, account.Platform) // 使用量记录通过有界 worker 池提交，避免请求热路径创建无界 goroutine。 - h.submitOpenAIUsageRecordTask(result, func(ctx context.Context) { + h.submitOpenAIUsageRecordTask(c.Request.Context(), result, func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.OpenAIRecordUsageInput{ Result: result, APIKey: apiKey, @@ -821,7 +847,7 @@ func (h *OpenAIGatewayHandler) Messages(c *gin.Context) { inboundEndpoint := GetInboundEndpoint(c) upstreamEndpoint := GetUpstreamEndpoint(c, account.Platform) - h.submitOpenAIUsageRecordTask(result, func(ctx context.Context) { + h.submitOpenAIUsageRecordTask(c.Request.Context(), result, func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.OpenAIRecordUsageInput{ Result: result, APIKey: apiKey, @@ -1424,7 +1450,7 @@ func (h *OpenAIGatewayHandler) ResponsesWebSocket(c *gin.Context) { h.gatewayService.ReportOpenAIAccountScheduleResult(account.ID, true, result.FirstTokenMs) inboundEndpoint := GetInboundEndpoint(c) upstreamEndpoint := GetUpstreamEndpoint(c, account.Platform) - h.submitOpenAIUsageRecordTask(result, func(taskCtx context.Context) { + h.submitOpenAIUsageRecordTask(ctx, result, func(taskCtx context.Context) { if err := h.gatewayService.RecordUsage(taskCtx, &service.OpenAIRecordUsageInput{ Result: result, APIKey: apiKey, @@ -1609,10 +1635,11 @@ func getContextInt64(c *gin.Context, key string) (int64, bool) { } } -func (h *OpenAIGatewayHandler) submitUsageRecordTask(task service.UsageRecordTask) { +func (h *OpenAIGatewayHandler) submitUsageRecordTask(parent context.Context, task service.UsageRecordTask) { if task == nil { return } + task = wrapUsageRecordTaskContext(parent, task) if h.usageRecordWorkerPool != nil { h.usageRecordWorkerPool.Submit(task) return @@ -1631,18 +1658,19 @@ func (h *OpenAIGatewayHandler) submitUsageRecordTask(task service.UsageRecordTas task(ctx) } -func (h *OpenAIGatewayHandler) submitOpenAIUsageRecordTask(result *service.OpenAIForwardResult, task service.UsageRecordTask) { +func (h *OpenAIGatewayHandler) submitOpenAIUsageRecordTask(parent context.Context, result *service.OpenAIForwardResult, task service.UsageRecordTask) { if result != nil && result.ImageCount > 0 { - h.submitMandatoryUsageRecordTask(task) + h.submitMandatoryUsageRecordTask(parent, task) return } - h.submitUsageRecordTask(task) + h.submitUsageRecordTask(parent, task) } -func (h *OpenAIGatewayHandler) submitMandatoryUsageRecordTask(task service.UsageRecordTask) { +func (h *OpenAIGatewayHandler) submitMandatoryUsageRecordTask(parent context.Context, task service.UsageRecordTask) { if task == nil { return } + task = wrapUsageRecordTaskContext(parent, task) if h.usageRecordWorkerPool != nil { if mode := h.usageRecordWorkerPool.Submit(task); mode != service.UsageRecordSubmitModeDropped { return diff --git a/backend/internal/handler/openai_gateway_usage_context_test.go b/backend/internal/handler/openai_gateway_usage_context_test.go new file mode 100644 index 00000000..7091c9c0 --- /dev/null +++ b/backend/internal/handler/openai_gateway_usage_context_test.go @@ -0,0 +1,41 @@ +package handler + +import ( + "context" + "testing" + + "github.com/Wei-Shaw/sub2api/internal/pkg/ctxkey" + "github.com/stretchr/testify/require" +) + +func TestSubmitUsageRecordTaskCopiesRequestContext(t *testing.T) { + parent := context.WithValue(context.Background(), ctxkey.ClientRequestID, "client-request-123") + parent = context.WithValue(parent, ctxkey.RequestID, "request-456") + + var gotClientRequestID string + var gotRequestID string + h := &GatewayHandler{} + h.submitUsageRecordTask(parent, func(ctx context.Context) { + gotClientRequestID, _ = ctx.Value(ctxkey.ClientRequestID).(string) + gotRequestID, _ = ctx.Value(ctxkey.RequestID).(string) + }) + + require.Equal(t, "client-request-123", gotClientRequestID) + require.Equal(t, "request-456", gotRequestID) +} + +func TestOpenAISubmitUsageRecordTaskCopiesRequestContext(t *testing.T) { + parent := context.WithValue(context.Background(), ctxkey.ClientRequestID, "openai-client-request-123") + parent = context.WithValue(parent, ctxkey.RequestID, "openai-request-456") + + var gotClientRequestID string + var gotRequestID string + h := &OpenAIGatewayHandler{} + h.submitUsageRecordTask(parent, func(ctx context.Context) { + gotClientRequestID, _ = ctx.Value(ctxkey.ClientRequestID).(string) + gotRequestID, _ = ctx.Value(ctxkey.RequestID).(string) + }) + + require.Equal(t, "openai-client-request-123", gotClientRequestID) + require.Equal(t, "openai-request-456", gotRequestID) +} diff --git a/backend/internal/handler/openai_images.go b/backend/internal/handler/openai_images.go index bbb08014..36339d4b 100644 --- a/backend/internal/handler/openai_images.go +++ b/backend/internal/handler/openai_images.go @@ -311,7 +311,7 @@ func (h *OpenAIGatewayHandler) Images(c *gin.Context) { if result != nil { upstreamModel = result.UpstreamModel } - h.submitMandatoryUsageRecordTask(func(ctx context.Context) { + h.submitMandatoryUsageRecordTask(c.Request.Context(), func(ctx context.Context) { if err := h.gatewayService.RecordUsage(ctx, &service.OpenAIRecordUsageInput{ Result: result, APIKey: apiKey, diff --git a/backend/internal/handler/usage_record_submit_task_test.go b/backend/internal/handler/usage_record_submit_task_test.go index e4c2837a..ebe5c3df 100644 --- a/backend/internal/handler/usage_record_submit_task_test.go +++ b/backend/internal/handler/usage_record_submit_task_test.go @@ -29,7 +29,7 @@ func TestGatewayHandlerSubmitUsageRecordTask_WithPool(t *testing.T) { h := &GatewayHandler{usageRecordWorkerPool: pool} done := make(chan struct{}) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { close(done) }) @@ -44,7 +44,7 @@ func TestGatewayHandlerSubmitUsageRecordTask_WithoutPoolSyncFallback(t *testing. h := &GatewayHandler{} var called atomic.Bool - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { if _, ok := ctx.Deadline(); !ok { t.Fatal("expected deadline in fallback context") } @@ -57,7 +57,7 @@ func TestGatewayHandlerSubmitUsageRecordTask_WithoutPoolSyncFallback(t *testing. func TestGatewayHandlerSubmitUsageRecordTask_NilTask(t *testing.T) { h := &GatewayHandler{} require.NotPanics(t, func() { - h.submitUsageRecordTask(nil) + h.submitUsageRecordTask(context.Background(), nil) }) } @@ -66,12 +66,12 @@ func TestGatewayHandlerSubmitUsageRecordTask_WithoutPool_TaskPanicRecovered(t *t var called atomic.Bool require.NotPanics(t, func() { - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { panic("usage task panic") }) }) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { called.Store(true) }) require.True(t, called.Load(), "panic 后后续任务应仍可执行") @@ -82,7 +82,7 @@ func TestOpenAIGatewayHandlerSubmitUsageRecordTask_WithPool(t *testing.T) { h := &OpenAIGatewayHandler{usageRecordWorkerPool: pool} done := make(chan struct{}) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { close(done) }) @@ -97,7 +97,7 @@ func TestOpenAIGatewayHandlerSubmitUsageRecordTask_WithoutPoolSyncFallback(t *te h := &OpenAIGatewayHandler{} var called atomic.Bool - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { if _, ok := ctx.Deadline(); !ok { t.Fatal("expected deadline in fallback context") } @@ -110,7 +110,7 @@ func TestOpenAIGatewayHandlerSubmitUsageRecordTask_WithoutPoolSyncFallback(t *te func TestOpenAIGatewayHandlerSubmitUsageRecordTask_NilTask(t *testing.T) { h := &OpenAIGatewayHandler{} require.NotPanics(t, func() { - h.submitUsageRecordTask(nil) + h.submitUsageRecordTask(context.Background(), nil) }) } @@ -119,12 +119,12 @@ func TestOpenAIGatewayHandlerSubmitUsageRecordTask_WithoutPool_TaskPanicRecovere var called atomic.Bool require.NotPanics(t, func() { - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { panic("usage task panic") }) }) - h.submitUsageRecordTask(func(ctx context.Context) { + h.submitUsageRecordTask(context.Background(), func(ctx context.Context) { called.Store(true) }) require.True(t, called.Load(), "panic 后后续任务应仍可执行") @@ -152,7 +152,7 @@ func TestOpenAIGatewayHandlerSubmitMandatoryUsageRecordTask_DroppedTaskSyncFallb pool.Submit(func(ctx context.Context) {}) var called atomic.Bool - h.submitMandatoryUsageRecordTask(func(ctx context.Context) { + h.submitMandatoryUsageRecordTask(context.Background(), func(ctx context.Context) { called.Store(true) }) close(release) @@ -182,7 +182,7 @@ func TestOpenAIGatewayHandlerSubmitOpenAIUsageRecordTask_ImageResultUsesMandator pool.Submit(func(ctx context.Context) {}) var called atomic.Bool - h.submitOpenAIUsageRecordTask(&service.OpenAIForwardResult{ImageCount: 1}, func(ctx context.Context) { + h.submitOpenAIUsageRecordTask(context.Background(), &service.OpenAIForwardResult{ImageCount: 1}, func(ctx context.Context) { called.Store(true) }) close(release) diff --git a/backend/internal/server/middleware/client_request_id.go b/backend/internal/server/middleware/client_request_id.go index 6838d6af..5f886646 100644 --- a/backend/internal/server/middleware/client_request_id.go +++ b/backend/internal/server/middleware/client_request_id.go @@ -11,6 +11,8 @@ import ( "go.uber.org/zap" ) +const clientRequestIDHeader = "X-Client-Request-ID" + // ClientRequestID ensures every request has a unique client_request_id in request.Context(). // // This is used by the Ops monitoring module for end-to-end request correlation. @@ -21,12 +23,14 @@ func ClientRequestID() gin.HandlerFunc { return } - if v := c.Request.Context().Value(ctxkey.ClientRequestID); v != nil { + if v, _ := c.Request.Context().Value(ctxkey.ClientRequestID).(string); strings.TrimSpace(v) != "" { + c.Header(clientRequestIDHeader, strings.TrimSpace(v)) c.Next() return } id := uuid.New().String() + c.Header(clientRequestIDHeader, id) ctx := context.WithValue(c.Request.Context(), ctxkey.ClientRequestID, id) requestLogger := logger.FromContext(ctx).With(zap.String("client_request_id", strings.TrimSpace(id))) ctx = logger.IntoContext(ctx, requestLogger) diff --git a/backend/internal/server/middleware/client_request_id_test.go b/backend/internal/server/middleware/client_request_id_test.go new file mode 100644 index 00000000..394c1612 --- /dev/null +++ b/backend/internal/server/middleware/client_request_id_test.go @@ -0,0 +1,50 @@ +package middleware + +import ( + "context" + "net/http" + "net/http/httptest" + "testing" + + "github.com/Wei-Shaw/sub2api/internal/pkg/ctxkey" + "github.com/gin-gonic/gin" + "github.com/stretchr/testify/require" +) + +func TestClientRequestIDGeneratesAndExposesID(t *testing.T) { + gin.SetMode(gin.TestMode) + router := gin.New() + router.Use(ClientRequestID()) + router.GET("/", func(c *gin.Context) { + value, _ := c.Request.Context().Value(ctxkey.ClientRequestID).(string) + c.String(http.StatusOK, value) + }) + + req := httptest.NewRequest(http.MethodGet, "/", nil) + w := httptest.NewRecorder() + + router.ServeHTTP(w, req) + + require.Equal(t, http.StatusOK, w.Code) + require.NotEmpty(t, w.Body.String()) + require.Equal(t, w.Body.String(), w.Header().Get(clientRequestIDHeader)) +} + +func TestClientRequestIDPreservesExistingContextID(t *testing.T) { + gin.SetMode(gin.TestMode) + router := gin.New() + router.Use(ClientRequestID()) + router.GET("/", func(c *gin.Context) { + value, _ := c.Request.Context().Value(ctxkey.ClientRequestID).(string) + c.String(http.StatusOK, value) + }) + + w := httptest.NewRecorder() + req := httptest.NewRequest(http.MethodGet, "/", nil) + req = req.WithContext(context.WithValue(req.Context(), ctxkey.ClientRequestID, "existing-client-request-id")) + router.ServeHTTP(w, req) + + require.Equal(t, http.StatusOK, w.Code) + require.Equal(t, "existing-client-request-id", w.Body.String()) + require.Equal(t, "existing-client-request-id", w.Header().Get(clientRequestIDHeader)) +} From ed1b57c5975578f62a88e95c3eaad3fa21e0efc0 Mon Sep 17 00:00:00 2001 From: shaw Date: Fri, 29 May 2026 08:58:10 +0800 Subject: [PATCH 10/20] fix(openai): gate routing by endpoint capability --- .../handler/openai_chat_completions.go | 3 +- backend/internal/handler/openai_embeddings.go | 10 +- .../handler/openai_gateway_handler.go | 9 +- backend/internal/service/account.go | 83 ++++++++ .../service/openai_account_scheduler.go | 55 +++-- .../service/openai_account_scheduler_test.go | 193 ++++++++++++++++++ .../service/openai_gateway_service.go | 62 +++--- .../internal/service/openai_images_test.go | 73 +++++++ .../service/openai_ws_account_sticky_test.go | 46 +++++ .../internal/service/openai_ws_forwarder.go | 39 +++- .../components/account/CreateAccountModal.vue | 69 ++++++- .../components/account/EditAccountModal.vue | 95 ++++++++- .../__tests__/EditAccountModal.spec.ts | 57 ++++++ frontend/src/i18n/locales/en.ts | 5 + frontend/src/i18n/locales/zh.ts | 5 + frontend/src/types/index.ts | 1 + 16 files changed, 740 insertions(+), 65 deletions(-) diff --git a/backend/internal/handler/openai_chat_completions.go b/backend/internal/handler/openai_chat_completions.go index 17f0d47e..9805bf8a 100644 --- a/backend/internal/handler/openai_chat_completions.go +++ b/backend/internal/handler/openai_chat_completions.go @@ -127,7 +127,7 @@ func (h *OpenAIGatewayHandler) ChatCompletions(c *gin.Context) { for { reqLog.Debug("openai_chat_completions.account_selecting", zap.Int("excluded_account_count", len(failedAccountIDs))) - selection, scheduleDecision, err := h.gatewayService.SelectAccountWithScheduler( + selection, scheduleDecision, err := h.gatewayService.SelectAccountWithSchedulerForCapability( c.Request.Context(), apiKey.GroupID, "", @@ -135,6 +135,7 @@ func (h *OpenAIGatewayHandler) ChatCompletions(c *gin.Context) { reqModel, failedAccountIDs, service.OpenAIUpstreamTransportAny, + service.OpenAIEndpointCapabilityChatCompletions, false, ) if err != nil { diff --git a/backend/internal/handler/openai_embeddings.go b/backend/internal/handler/openai_embeddings.go index bbb67044..81713f7f 100644 --- a/backend/internal/handler/openai_embeddings.go +++ b/backend/internal/handler/openai_embeddings.go @@ -107,7 +107,7 @@ func (h *OpenAIGatewayHandler) Embeddings(c *gin.Context) { routingStart := time.Now() for { - selection, _, err := h.gatewayService.SelectAccountWithScheduler( + selection, _, err := h.gatewayService.SelectAccountWithSchedulerForCapability( c.Request.Context(), apiKey.GroupID, "", @@ -115,6 +115,7 @@ func (h *OpenAIGatewayHandler) Embeddings(c *gin.Context) { reqModel, failedAccountIDs, service.OpenAIUpstreamTransportHTTPSSE, + service.OpenAIEndpointCapabilityEmbeddings, false, ) if err != nil { @@ -140,13 +141,6 @@ func (h *OpenAIGatewayHandler) Embeddings(c *gin.Context) { return } account := selection.Account - if account.Type != service.AccountTypeAPIKey { - if selection.ReleaseFunc != nil { - selection.ReleaseFunc() - } - failedAccountIDs[account.ID] = struct{}{} - continue - } setOpsSelectedAccount(c, account.ID, account.Platform) accountReleaseFunc, accountAcquired := h.acquireResponsesAccountSlot(c, apiKey.GroupID, "", selection, false, &streamStarted, reqLog) diff --git a/backend/internal/handler/openai_gateway_handler.go b/backend/internal/handler/openai_gateway_handler.go index a51eee86..1d661748 100644 --- a/backend/internal/handler/openai_gateway_handler.go +++ b/backend/internal/handler/openai_gateway_handler.go @@ -266,7 +266,7 @@ func (h *OpenAIGatewayHandler) Responses(c *gin.Context) { for { // Select account supporting the requested model reqLog.Debug("openai.account_selecting", zap.Int("excluded_account_count", len(failedAccountIDs))) - selection, scheduleDecision, err := h.gatewayService.SelectAccountWithScheduler( + selection, scheduleDecision, err := h.gatewayService.SelectAccountWithSchedulerForCapability( c.Request.Context(), apiKey.GroupID, previousResponseID, @@ -274,6 +274,7 @@ func (h *OpenAIGatewayHandler) Responses(c *gin.Context) { reqModel, failedAccountIDs, service.OpenAIUpstreamTransportAny, + service.OpenAIEndpointCapabilityChatCompletions, requireCompact, ) if err != nil { @@ -675,7 +676,7 @@ func (h *OpenAIGatewayHandler) Messages(c *gin.Context) { currentRoutingModel = effectiveMappedModel } reqLog.Debug("openai_messages.account_selecting", zap.Int("excluded_account_count", len(failedAccountIDs))) - selection, scheduleDecision, err := h.gatewayService.SelectAccountWithScheduler( + selection, scheduleDecision, err := h.gatewayService.SelectAccountWithSchedulerForCapability( c.Request.Context(), apiKey.GroupID, "", // no previous_response_id @@ -683,6 +684,7 @@ func (h *OpenAIGatewayHandler) Messages(c *gin.Context) { currentRoutingModel, failedAccountIDs, service.OpenAIUpstreamTransportAny, + service.OpenAIEndpointCapabilityChatCompletions, false, ) if err != nil { @@ -1273,7 +1275,7 @@ func (h *OpenAIGatewayHandler) ResponsesWebSocket(c *gin.Context) { for { reqLog.Debug("openai.websocket_account_selecting", zap.Int("excluded_account_count", len(failedAccountIDs))) - selection, scheduleDecision, err := h.gatewayService.SelectAccountWithScheduler( + selection, scheduleDecision, err := h.gatewayService.SelectAccountWithSchedulerForCapability( ctx, apiKey.GroupID, previousResponseID, @@ -1281,6 +1283,7 @@ func (h *OpenAIGatewayHandler) ResponsesWebSocket(c *gin.Context) { reqModel, failedAccountIDs, service.OpenAIUpstreamTransportResponsesWebsocketV2, + service.OpenAIEndpointCapabilityChatCompletions, false, ) if err != nil { diff --git a/backend/internal/service/account.go b/backend/internal/service/account.go index d488aa75..e3ca9c5d 100644 --- a/backend/internal/service/account.go +++ b/backend/internal/service/account.go @@ -66,6 +66,15 @@ type Account struct { modelMappingCacheRawSig uint64 } +type OpenAIEndpointCapability string + +const ( + OpenAIEndpointCapabilityChatCompletions OpenAIEndpointCapability = "chat_completions" + OpenAIEndpointCapabilityEmbeddings OpenAIEndpointCapability = "embeddings" +) + +const openAIEndpointCapabilitiesCredentialKey = "openai_capabilities" + type TempUnschedulableRule struct { ErrorCode int `json:"error_code"` Keywords []string `json:"keywords"` @@ -1122,6 +1131,80 @@ func (a *Account) GetOpenAISessionID() string { return strings.TrimSpace(a.GetExtraString("openai_session_id")) } +func (a *Account) SupportsOpenAIEndpointCapability(capability OpenAIEndpointCapability) bool { + if a == nil { + return false + } + if capability == "" { + return true + } + if !a.IsOpenAI() { + return false + } + switch capability { + case OpenAIEndpointCapabilityChatCompletions: + case OpenAIEndpointCapabilityEmbeddings: + if a.Type != AccountTypeAPIKey { + return false + } + default: + return false + } + + configured, found := a.openAIEndpointCapabilitySet() + if !found { + return true + } + return configured[string(capability)] +} + +func (a *Account) openAIEndpointCapabilitySet() (map[string]bool, bool) { + if a == nil || a.Credentials == nil { + return nil, false + } + raw, found := a.Credentials[openAIEndpointCapabilitiesCredentialKey] + if !found || raw == nil { + return nil, false + } + + result := make(map[string]bool) + add := func(value string) { + value = strings.ToLower(strings.TrimSpace(value)) + if value == "" { + return + } + result[value] = true + } + + switch capabilities := raw.(type) { + case []any: + for _, item := range capabilities { + if value, ok := item.(string); ok { + add(value) + } + } + case []string: + for _, value := range capabilities { + add(value) + } + case map[string]any: + for key, value := range capabilities { + enabled, ok := value.(bool) + if ok && enabled { + add(key) + } + } + case map[string]bool: + for key, enabled := range capabilities { + if enabled { + add(key) + } + } + } + + return result, true +} + func (a *Account) SupportsOpenAIImageCapability(capability OpenAIImagesCapability) bool { if !a.IsOpenAI() { return false diff --git a/backend/internal/service/openai_account_scheduler.go b/backend/internal/service/openai_account_scheduler.go index a8ac391a..1eca08b1 100644 --- a/backend/internal/service/openai_account_scheduler.go +++ b/backend/internal/service/openai_account_scheduler.go @@ -44,6 +44,7 @@ type OpenAIAccountScheduleRequest struct { PreviousResponseID string RequestedModel string RequiredTransport OpenAIUpstreamTransport + RequiredCapability OpenAIEndpointCapability RequiredImageCapability OpenAIImagesCapability RequireCompact bool ExcludedIDs map[int64]struct{} @@ -263,12 +264,13 @@ func (s *defaultOpenAIAccountScheduler) Select( previousResponseID := strings.TrimSpace(req.PreviousResponseID) if previousResponseID != "" { - selection, err := s.service.SelectAccountByPreviousResponseID( + selection, err := s.service.selectAccountByPreviousResponseIDForCapability( ctx, req.GroupID, previousResponseID, req.RequestedModel, req.ExcludedIDs, + req.RequiredCapability, req.RequireCompact, ) if err != nil { @@ -363,7 +365,7 @@ func (s *defaultOpenAIAccountScheduler) selectBySessionHash( _ = s.service.deleteStickySessionAccountID(ctx, req.GroupID, sessionHash) return nil, nil } - account = s.service.recheckSelectedOpenAIAccountFromDB(ctx, account, req.RequestedModel, req.RequireCompact) + account = s.service.recheckSelectedOpenAIAccountFromDB(ctx, account, req.RequestedModel, req.RequireCompact, req.RequiredCapability) if account == nil || !s.isAccountTransportCompatible(account, req.RequiredTransport) { _ = s.service.deleteStickySessionAccountID(ctx, req.GroupID, sessionHash) return nil, nil @@ -791,11 +793,11 @@ func (s *defaultOpenAIAccountScheduler) tryAcquireOpenAISelectionOrder( compactBlocked := false for i := 0; i < len(selectionOrder); i++ { candidate := selectionOrder[i] - fresh := s.service.resolveFreshSchedulableOpenAIAccount(ctx, candidate.account, req.RequestedModel, false) + fresh := s.service.resolveFreshSchedulableOpenAIAccount(ctx, candidate.account, req.RequestedModel, false, req.RequiredCapability) if fresh == nil || !s.isAccountTransportCompatible(fresh, req.RequiredTransport) || !s.isAccountRequestCompatible(ctx, fresh, req) { continue } - fresh = s.service.recheckSelectedOpenAIAccountFromDB(ctx, fresh, req.RequestedModel, false) + fresh = s.service.recheckSelectedOpenAIAccountFromDB(ctx, fresh, req.RequestedModel, false, req.RequiredCapability) if fresh == nil || !s.isAccountTransportCompatible(fresh, req.RequiredTransport) || !s.isAccountRequestCompatible(ctx, fresh, req) { continue } @@ -930,11 +932,11 @@ func (s *defaultOpenAIAccountScheduler) selectByLoadBalance( cfg := s.service.schedulingConfig() // WaitPlan.MaxConcurrency 使用 Concurrency（非 EffectiveLoadFactor），因为 WaitPlan 控制的是 Redis 实际并发槽位等待。 for _, candidate := range selectionOrder { - fresh := s.service.resolveFreshSchedulableOpenAIAccount(ctx, candidate.account, req.RequestedModel, false) + fresh := s.service.resolveFreshSchedulableOpenAIAccount(ctx, candidate.account, req.RequestedModel, false, req.RequiredCapability) if fresh == nil || !s.isAccountTransportCompatible(fresh, req.RequiredTransport) || !s.isAccountRequestCompatible(ctx, fresh, req) { continue } - fresh = s.service.recheckSelectedOpenAIAccountFromDB(ctx, fresh, req.RequestedModel, false) + fresh = s.service.recheckSelectedOpenAIAccountFromDB(ctx, fresh, req.RequestedModel, false, req.RequiredCapability) if fresh == nil || !s.isAccountTransportCompatible(fresh, req.RequiredTransport) || !s.isAccountRequestCompatible(ctx, fresh, req) { continue } @@ -981,7 +983,7 @@ func (s *defaultOpenAIAccountScheduler) isAccountRequestCompatible(ctx context.C s.service.isUpstreamModelRestrictedByChannel(ctx, *req.GroupID, account, req.RequestedModel, req.RequireCompact) { return false } - return account.SupportsOpenAIImageCapability(req.RequiredImageCapability) + return accountSupportsOpenAICapabilities(account, req.RequiredCapability, req.RequiredImageCapability) } func (s *defaultOpenAIAccountScheduler) ReportResult(accountID int64, success bool, firstTokenMs *int) { @@ -1104,7 +1106,21 @@ func (s *OpenAIGatewayService) SelectAccountWithScheduler( requiredTransport OpenAIUpstreamTransport, requireCompact bool, ) (*AccountSelectionResult, OpenAIAccountScheduleDecision, error) { - return s.selectAccountWithScheduler(ctx, groupID, previousResponseID, sessionHash, requestedModel, excludedIDs, requiredTransport, "", requireCompact) + return s.selectAccountWithScheduler(ctx, groupID, previousResponseID, sessionHash, requestedModel, excludedIDs, requiredTransport, "", "", requireCompact) +} + +func (s *OpenAIGatewayService) SelectAccountWithSchedulerForCapability( + ctx context.Context, + groupID *int64, + previousResponseID string, + sessionHash string, + requestedModel string, + excludedIDs map[int64]struct{}, + requiredTransport OpenAIUpstreamTransport, + requiredCapability OpenAIEndpointCapability, + requireCompact bool, +) (*AccountSelectionResult, OpenAIAccountScheduleDecision, error) { + return s.selectAccountWithScheduler(ctx, groupID, previousResponseID, sessionHash, requestedModel, excludedIDs, requiredTransport, requiredCapability, "", requireCompact) } func (s *OpenAIGatewayService) SelectAccountWithSchedulerForImages( @@ -1115,13 +1131,13 @@ func (s *OpenAIGatewayService) SelectAccountWithSchedulerForImages( excludedIDs map[int64]struct{}, requiredCapability OpenAIImagesCapability, ) (*AccountSelectionResult, OpenAIAccountScheduleDecision, error) { - selection, decision, err := s.selectAccountWithScheduler(ctx, groupID, "", sessionHash, requestedModel, excludedIDs, OpenAIUpstreamTransportHTTPSSE, requiredCapability, false) + selection, decision, err := s.selectAccountWithScheduler(ctx, groupID, "", sessionHash, requestedModel, excludedIDs, OpenAIUpstreamTransportHTTPSSE, "", requiredCapability, false) if err == nil && selection != nil && selection.Account != nil { return selection, decision, nil } // 如果要求 native 能力（如指定了模型）但没有可用的 APIKey 账号，回退到 basic（OAuth 账号） if requiredCapability == OpenAIImagesCapabilityNative { - return s.selectAccountWithScheduler(ctx, groupID, "", sessionHash, requestedModel, excludedIDs, OpenAIUpstreamTransportHTTPSSE, OpenAIImagesCapabilityBasic, false) + return s.selectAccountWithScheduler(ctx, groupID, "", sessionHash, requestedModel, excludedIDs, OpenAIUpstreamTransportHTTPSSE, "", OpenAIImagesCapabilityBasic, false) } return selection, decision, err } @@ -1134,6 +1150,7 @@ func (s *OpenAIGatewayService) selectAccountWithScheduler( requestedModel string, excludedIDs map[int64]struct{}, requiredTransport OpenAIUpstreamTransport, + requiredCapability OpenAIEndpointCapability, requiredImageCapability OpenAIImagesCapability, requireCompact bool, ) (*AccountSelectionResult, OpenAIAccountScheduleDecision, error) { @@ -1144,14 +1161,14 @@ func (s *OpenAIGatewayService) selectAccountWithScheduler( if requiredTransport == OpenAIUpstreamTransportAny || requiredTransport == OpenAIUpstreamTransportHTTPSSE { effectiveExcludedIDs := cloneExcludedAccountIDs(excludedIDs) for { - selection, err := s.selectAccountWithLoadAwareness(ctx, groupID, sessionHash, requestedModel, effectiveExcludedIDs, requireCompact) + selection, err := s.selectAccountWithLoadAwareness(ctx, groupID, sessionHash, requestedModel, effectiveExcludedIDs, requireCompact, requiredCapability) if err != nil { return nil, decision, err } if selection == nil || selection.Account == nil { return selection, decision, nil } - if selection.Account.SupportsOpenAIImageCapability(requiredImageCapability) { + if accountSupportsOpenAICapabilities(selection.Account, requiredCapability, requiredImageCapability) { return selection, decision, nil } if selection.ReleaseFunc != nil { @@ -1169,14 +1186,15 @@ func (s *OpenAIGatewayService) selectAccountWithScheduler( effectiveExcludedIDs := cloneExcludedAccountIDs(excludedIDs) for { - selection, err := s.selectAccountWithLoadAwareness(ctx, groupID, sessionHash, requestedModel, effectiveExcludedIDs, requireCompact) + selection, err := s.selectAccountWithLoadAwareness(ctx, groupID, sessionHash, requestedModel, effectiveExcludedIDs, requireCompact, requiredCapability) if err != nil { return nil, decision, err } if selection == nil || selection.Account == nil { return selection, decision, nil } - if s.isOpenAIAccountTransportCompatible(selection.Account, requiredTransport) { + if s.isOpenAIAccountTransportCompatible(selection.Account, requiredTransport) && + accountSupportsOpenAICapabilities(selection.Account, requiredCapability, requiredImageCapability) { return selection, decision, nil } if selection.ReleaseFunc != nil { @@ -1213,12 +1231,21 @@ func (s *OpenAIGatewayService) selectAccountWithScheduler( PreviousResponseID: previousResponseID, RequestedModel: requestedModel, RequiredTransport: requiredTransport, + RequiredCapability: requiredCapability, RequiredImageCapability: requiredImageCapability, RequireCompact: requireCompact, ExcludedIDs: excludedIDs, }) } +func accountSupportsOpenAICapabilities(account *Account, requiredCapability OpenAIEndpointCapability, requiredImageCapability OpenAIImagesCapability) bool { + if account == nil { + return false + } + return account.SupportsOpenAIEndpointCapability(requiredCapability) && + account.SupportsOpenAIImageCapability(requiredImageCapability) +} + func cloneExcludedAccountIDs(excludedIDs map[int64]struct{}) map[int64]struct{} { if len(excludedIDs) == 0 { return nil diff --git a/backend/internal/service/openai_account_scheduler_test.go b/backend/internal/service/openai_account_scheduler_test.go index ba20ee5f..fedf7e9c 100644 --- a/backend/internal/service/openai_account_scheduler_test.go +++ b/backend/internal/service/openai_account_scheduler_test.go @@ -393,6 +393,64 @@ func TestOpenAIGatewayService_SelectAccountWithScheduler_DefaultDisabled_Require require.Equal(t, openAIAccountScheduleLayerLoadBalance, decision.Layer) } +func TestOpenAIGatewayService_SelectAccountWithScheduler_DefaultDisabled_EmbeddingsSkipsChatOnlyAccount(t *testing.T) { + resetOpenAIAdvancedSchedulerSettingCacheForTest() + + ctx := context.Background() + groupID := int64(10110) + accounts := []Account{ + { + ID: 36031, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions"}, + }, + }, + { + ID: 36032, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 5, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions", "embeddings"}, + }, + }, + } + cfg := &config.Config{} + cfg.Gateway.Scheduling.LoadBatchEnabled = false + svc := &OpenAIGatewayService{ + accountRepo: schedulerTestOpenAIAccountRepo{accounts: accounts}, + cache: &schedulerTestGatewayCache{}, + cfg: cfg, + concurrencyService: NewConcurrencyService(schedulerTestConcurrencyCache{}), + } + + selection, decision, err := svc.SelectAccountWithSchedulerForCapability( + ctx, + &groupID, + "", + "", + "text-embedding-3-small", + nil, + OpenAIUpstreamTransportHTTPSSE, + OpenAIEndpointCapabilityEmbeddings, + false, + ) + require.NoError(t, err) + require.NotNil(t, selection) + require.NotNil(t, selection.Account) + require.Equal(t, int64(36032), selection.Account.ID) + require.Equal(t, openAIAccountScheduleLayerLoadBalance, decision.Layer) +} + func TestOpenAIGatewayService_SelectAccountWithScheduler_EnabledUsesAdvancedPreviousResponseRouting(t *testing.T) { resetOpenAIAdvancedSchedulerSettingCacheForTest() @@ -458,6 +516,141 @@ func TestOpenAIGatewayService_SelectAccountWithScheduler_EnabledUsesAdvancedPrev require.True(t, decision.StickyPreviousHit) } +func TestOpenAIGatewayService_SelectAccountWithScheduler_Enabled_EmbeddingsSkipsChatOnlyAccount(t *testing.T) { + resetOpenAIAdvancedSchedulerSettingCacheForTest() + + ctx := context.Background() + groupID := int64(10111) + accounts := []Account{ + { + ID: 37011, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions"}, + }, + }, + { + ID: 37012, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 5, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions", "embeddings"}, + }, + }, + } + cfg := &config.Config{} + cfg.Gateway.Scheduling.LoadBatchEnabled = false + svc := &OpenAIGatewayService{ + accountRepo: schedulerTestOpenAIAccountRepo{accounts: accounts}, + cache: &schedulerTestGatewayCache{}, + cfg: cfg, + rateLimitService: newOpenAIAdvancedSchedulerRateLimitService("true"), + concurrencyService: NewConcurrencyService(schedulerTestConcurrencyCache{}), + } + + selection, decision, err := svc.SelectAccountWithSchedulerForCapability( + ctx, + &groupID, + "", + "", + "text-embedding-3-small", + nil, + OpenAIUpstreamTransportHTTPSSE, + OpenAIEndpointCapabilityEmbeddings, + false, + ) + require.NoError(t, err) + require.NotNil(t, selection) + require.NotNil(t, selection.Account) + require.Equal(t, int64(37012), selection.Account.ID) + require.Equal(t, openAIAccountScheduleLayerLoadBalance, decision.Layer) + require.Equal(t, 1, decision.CandidateCount) +} + +func TestOpenAIGatewayService_SelectAccountWithScheduler_Enabled_EmbeddingsSkipsChatOnlyStickyBindings(t *testing.T) { + resetOpenAIAdvancedSchedulerSettingCacheForTest() + + ctx := context.Background() + groupID := int64(10112) + accounts := []Account{ + { + ID: 37021, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions"}, + }, + Extra: map[string]any{ + "openai_apikey_responses_websockets_v2_enabled": true, + }, + }, + { + ID: 37022, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 5, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions", "embeddings"}, + }, + Extra: map[string]any{ + "openai_apikey_responses_websockets_v2_enabled": true, + }, + }, + } + cfg := newSchedulerTestOpenAIWSV2Config() + cfg.Gateway.Scheduling.LoadBatchEnabled = false + cache := &schedulerTestGatewayCache{ + sessionBindings: map[string]int64{ + "openai:session_hash_embeddings": 37021, + }, + } + svc := &OpenAIGatewayService{ + accountRepo: schedulerTestOpenAIAccountRepo{accounts: accounts}, + cache: cache, + cfg: cfg, + rateLimitService: newOpenAIAdvancedSchedulerRateLimitService("true"), + concurrencyService: NewConcurrencyService(schedulerTestConcurrencyCache{}), + } + store := svc.getOpenAIWSStateStore() + require.NoError(t, store.BindResponseAccount(ctx, groupID, "resp_embeddings_chat_only", 37021, time.Hour)) + + selection, decision, err := svc.SelectAccountWithSchedulerForCapability( + ctx, + &groupID, + "resp_embeddings_chat_only", + "session_hash_embeddings", + "text-embedding-3-small", + nil, + OpenAIUpstreamTransportHTTPSSE, + OpenAIEndpointCapabilityEmbeddings, + false, + ) + require.NoError(t, err) + require.NotNil(t, selection) + require.NotNil(t, selection.Account) + require.Equal(t, int64(37022), selection.Account.ID) + require.Equal(t, openAIAccountScheduleLayerLoadBalance, decision.Layer) + require.False(t, decision.StickyPreviousHit) + require.False(t, decision.StickySessionHit) + require.Equal(t, int64(37022), cache.sessionBindings["openai:session_hash_embeddings"]) +} + func TestOpenAIGatewayService_OpenAIAccountSchedulerMetrics_DisabledNoOp(t *testing.T) { resetOpenAIAdvancedSchedulerSettingCacheForTest() diff --git a/backend/internal/service/openai_gateway_service.go b/backend/internal/service/openai_gateway_service.go index f93cc221..77587f69 100644 --- a/backend/internal/service/openai_gateway_service.go +++ b/backend/internal/service/openai_gateway_service.go @@ -1279,7 +1279,7 @@ func (s *OpenAIGatewayService) SelectAccountForModel(ctx context.Context, groupI // SelectAccountForModelWithExclusions selects an account supporting the requested model while excluding specified accounts. // SelectAccountForModelWithExclusions 选择支持指定模型的账号，同时排除指定的账号。 func (s *OpenAIGatewayService) SelectAccountForModelWithExclusions(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}) (*Account, error) { - return s.selectAccountForModelWithExclusions(ctx, groupID, sessionHash, requestedModel, excludedIDs, false, 0) + return s.selectAccountForModelWithExclusions(ctx, groupID, sessionHash, requestedModel, excludedIDs, false, 0, "") } // noAvailableOpenAISelectionError builds the standard "no account available" error @@ -1312,13 +1312,16 @@ func openAICompactSupportTier(account *Account) int { // isOpenAIAccountEligibleForRequest centralises the schedulable / OpenAI / model / // compact-support checks used during account selection. -func isOpenAIAccountEligibleForRequest(ctx context.Context, account *Account, requestedModel string, requireCompact bool) bool { +func isOpenAIAccountEligibleForRequest(ctx context.Context, account *Account, requestedModel string, requireCompact bool, requiredCapability OpenAIEndpointCapability) bool { if account == nil || !account.IsOpenAI() || !account.IsSchedulableForModelWithContext(ctx, requestedModel) { return false } if requestedModel != "" && !account.IsModelSupported(requestedModel) { return false } + if !account.SupportsOpenAIEndpointCapability(requiredCapability) { + return false + } if requireCompact && openAICompactSupportTier(account) == 0 { return false } @@ -1366,7 +1369,7 @@ func resolveOpenAIAccountUpstreamModelForRequest(account *Account, requestedMode return upstreamModel } -func (s *OpenAIGatewayService) selectAccountForModelWithExclusions(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool, stickyAccountID int64) (*Account, error) { +func (s *OpenAIGatewayService) selectAccountForModelWithExclusions(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool, stickyAccountID int64, requiredCapability OpenAIEndpointCapability) (*Account, error) { if s.checkChannelPricingRestriction(ctx, groupID, requestedModel) { slog.Warn("channel pricing restriction blocked request", "group_id", derefGroupID(groupID), @@ -1376,7 +1379,7 @@ func (s *OpenAIGatewayService) selectAccountForModelWithExclusions(ctx context.C // 1. 尝试粘性会话命中 // Try sticky session hit - if account := s.tryStickySessionHit(ctx, groupID, sessionHash, requestedModel, excludedIDs, requireCompact, stickyAccountID); account != nil { + if account := s.tryStickySessionHit(ctx, groupID, sessionHash, requestedModel, excludedIDs, requireCompact, stickyAccountID, requiredCapability); account != nil { return account, nil } @@ -1389,7 +1392,7 @@ func (s *OpenAIGatewayService) selectAccountForModelWithExclusions(ctx context.C // 3. 按优先级 + LRU 选择最佳账号 // Select by priority + LRU - selected, compactBlocked := s.selectBestAccount(ctx, groupID, accounts, requestedModel, excludedIDs, requireCompact) + selected, compactBlocked := s.selectBestAccount(ctx, groupID, accounts, requestedModel, excludedIDs, requireCompact, requiredCapability) if selected == nil { return nil, noAvailableOpenAISelectionError(requestedModel, compactBlocked) @@ -1414,7 +1417,7 @@ func (s *OpenAIGatewayService) selectAccountForModelWithExclusions(ctx context.C // // tryStickySessionHit attempts to get account from sticky session. // Returns account if hit and usable; clears session and returns nil if account is unavailable. -func (s *OpenAIGatewayService) tryStickySessionHit(ctx context.Context, groupID *int64, sessionHash, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool, stickyAccountID int64) *Account { +func (s *OpenAIGatewayService) tryStickySessionHit(ctx context.Context, groupID *int64, sessionHash, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool, stickyAccountID int64, requiredCapability OpenAIEndpointCapability) *Account { if sessionHash == "" { return nil } @@ -1446,14 +1449,14 @@ func (s *OpenAIGatewayService) tryStickySessionHit(ctx context.Context, groupID // 验证账号是否可用于当前请求 // Verify account is usable for current request - if !isOpenAIAccountEligibleForRequest(ctx, account, requestedModel, false) { + if !isOpenAIAccountEligibleForRequest(ctx, account, requestedModel, false, requiredCapability) { return nil } if s.isOpenAIAccountRuntimeBlocked(account) { _ = s.deleteStickySessionAccountID(ctx, groupID, sessionHash) return nil } - account = s.recheckSelectedOpenAIAccountFromDB(ctx, account, requestedModel, requireCompact) + account = s.recheckSelectedOpenAIAccountFromDB(ctx, account, requestedModel, requireCompact, requiredCapability) if account == nil { _ = s.deleteStickySessionAccountID(ctx, groupID, sessionHash) return nil @@ -1477,7 +1480,7 @@ func (s *OpenAIGatewayService) tryStickySessionHit(ctx context.Context, groupID // Returns nil if no available account. The second return reports whether at // least one candidate was filtered out solely because it lacks compact support // (only meaningful when requireCompact=true). -func (s *OpenAIGatewayService) selectBestAccount(ctx context.Context, groupID *int64, accounts []Account, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool) (*Account, bool) { +func (s *OpenAIGatewayService) selectBestAccount(ctx context.Context, groupID *int64, accounts []Account, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool, requiredCapability OpenAIEndpointCapability) (*Account, bool) { var selected *Account selectedCompactTier := -1 compactBlocked := false @@ -1492,11 +1495,11 @@ func (s *OpenAIGatewayService) selectBestAccount(ctx context.Context, groupID *i continue } - fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, acc, requestedModel, false) + fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, acc, requestedModel, false, requiredCapability) if fresh == nil { continue } - fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, false) + fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, false, requiredCapability) if fresh == nil { continue } @@ -1573,10 +1576,10 @@ func (s *OpenAIGatewayService) isBetterAccount(candidate, current *Account) bool // SelectAccountWithLoadAwareness selects an account with load-awareness and wait plan. func (s *OpenAIGatewayService) SelectAccountWithLoadAwareness(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}) (*AccountSelectionResult, error) { - return s.selectAccountWithLoadAwareness(ctx, groupID, sessionHash, requestedModel, excludedIDs, false) + return s.selectAccountWithLoadAwareness(ctx, groupID, sessionHash, requestedModel, excludedIDs, false, "") } -func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool) (*AccountSelectionResult, error) { +func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool, requiredCapability OpenAIEndpointCapability) (*AccountSelectionResult, error) { if s.checkChannelPricingRestriction(ctx, groupID, requestedModel) { slog.Warn("channel pricing restriction blocked request", "group_id", derefGroupID(groupID), @@ -1593,7 +1596,7 @@ func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Contex } } if s.concurrencyService == nil || !cfg.LoadBatchEnabled { - account, err := s.selectAccountForModelWithExclusions(ctx, groupID, sessionHash, requestedModel, excludedIDs, requireCompact, stickyAccountID) + account, err := s.selectAccountForModelWithExclusions(ctx, groupID, sessionHash, requestedModel, excludedIDs, requireCompact, stickyAccountID, requiredCapability) if err != nil { return nil, err } @@ -1646,8 +1649,8 @@ func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Contex if clearSticky { _ = s.deleteStickySessionAccountID(ctx, groupID, sessionHash) } - if !clearSticky && isOpenAIAccountEligibleForRequest(ctx, account, requestedModel, false) { - account = s.recheckSelectedOpenAIAccountFromDB(ctx, account, requestedModel, requireCompact) + if !clearSticky && isOpenAIAccountEligibleForRequest(ctx, account, requestedModel, false, requiredCapability) { + account = s.recheckSelectedOpenAIAccountFromDB(ctx, account, requestedModel, requireCompact, requiredCapability) if account == nil { _ = s.deleteStickySessionAccountID(ctx, groupID, sessionHash) } else if s.isOpenAIAccountRuntimeBlocked(account) { @@ -1691,15 +1694,12 @@ func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Contex // Scheduler snapshots can be temporarily stale (bucket rebuild is throttled); // re-check schedulability here so recently rate-limited/overloaded accounts // are not selected again before the bucket is rebuilt. - if !acc.IsSchedulable() { + if !isOpenAIAccountEligibleForRequest(ctx, acc, requestedModel, false, requiredCapability) { continue } if s.isOpenAIAccountRuntimeBlocked(acc) { continue } - if requestedModel != "" && !acc.IsModelSupported(requestedModel) { - continue - } if needsUpstreamCheck && s.isUpstreamModelRestrictedByChannel(ctx, *groupID, acc, requestedModel, requireCompact) { continue } @@ -1779,11 +1779,11 @@ func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Contex } for _, item := range selectionOrder { - fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, item.account, requestedModel, false) + fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, item.account, requestedModel, false, requiredCapability) if fresh == nil { continue } - fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, requireCompact) + fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, requireCompact, requiredCapability) if fresh == nil { continue } @@ -1813,11 +1813,11 @@ func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Contex ordered = prioritizeOpenAICompactAccounts(ordered) } for _, acc := range ordered { - fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, acc, requestedModel, false) + fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, acc, requestedModel, false, requiredCapability) if fresh == nil { continue } - fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, requireCompact) + fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, requireCompact, requiredCapability) if fresh == nil { continue } @@ -1858,11 +1858,11 @@ func (s *OpenAIGatewayService) selectAccountWithLoadAwareness(ctx context.Contex candidates = prioritizeOpenAICompactAccounts(candidates) } for _, acc := range candidates { - fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, acc, requestedModel, false) + fresh := s.resolveFreshSchedulableOpenAIAccount(ctx, acc, requestedModel, false, requiredCapability) if fresh == nil { continue } - fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, requireCompact) + fresh = s.recheckSelectedOpenAIAccountFromDB(ctx, fresh, requestedModel, requireCompact, requiredCapability) if fresh == nil { continue } @@ -1910,7 +1910,7 @@ func (s *OpenAIGatewayService) tryAcquireAccountSlot(ctx context.Context, accoun return s.concurrencyService.AcquireAccountSlot(ctx, accountID, maxConcurrency) } -func (s *OpenAIGatewayService) resolveFreshSchedulableOpenAIAccount(ctx context.Context, account *Account, requestedModel string, requireCompact bool) *Account { +func (s *OpenAIGatewayService) resolveFreshSchedulableOpenAIAccount(ctx context.Context, account *Account, requestedModel string, requireCompact bool, requiredCapability OpenAIEndpointCapability) *Account { if account == nil { return nil } @@ -1924,7 +1924,7 @@ func (s *OpenAIGatewayService) resolveFreshSchedulableOpenAIAccount(ctx context. fresh = current } - if !isOpenAIAccountEligibleForRequest(ctx, fresh, requestedModel, requireCompact) { + if !isOpenAIAccountEligibleForRequest(ctx, fresh, requestedModel, requireCompact, requiredCapability) { return nil } if s.isOpenAIAccountRuntimeBlocked(fresh) { @@ -1933,12 +1933,12 @@ func (s *OpenAIGatewayService) resolveFreshSchedulableOpenAIAccount(ctx context. return fresh } -func (s *OpenAIGatewayService) recheckSelectedOpenAIAccountFromDB(ctx context.Context, account *Account, requestedModel string, requireCompact bool) *Account { +func (s *OpenAIGatewayService) recheckSelectedOpenAIAccountFromDB(ctx context.Context, account *Account, requestedModel string, requireCompact bool, requiredCapability OpenAIEndpointCapability) *Account { if account == nil { return nil } if s.schedulerSnapshot == nil || s.accountRepo == nil { - if !isOpenAIAccountEligibleForRequest(ctx, account, requestedModel, requireCompact) { + if !isOpenAIAccountEligibleForRequest(ctx, account, requestedModel, requireCompact, requiredCapability) { return nil } return account @@ -1948,7 +1948,7 @@ func (s *OpenAIGatewayService) recheckSelectedOpenAIAccountFromDB(ctx context.Co if err != nil || latest == nil { return nil } - if !isOpenAIAccountEligibleForRequest(ctx, latest, requestedModel, requireCompact) { + if !isOpenAIAccountEligibleForRequest(ctx, latest, requestedModel, requireCompact, requiredCapability) { return nil } if s.isOpenAIAccountRuntimeBlocked(latest) { diff --git a/backend/internal/service/openai_images_test.go b/backend/internal/service/openai_images_test.go index 854e9f6d..a87e96c1 100644 --- a/backend/internal/service/openai_images_test.go +++ b/backend/internal/service/openai_images_test.go @@ -413,6 +413,79 @@ func TestAccountSupportsOpenAIImageCapability_OAuthSupportsNative(t *testing.T) require.True(t, account.SupportsOpenAIImageCapability(OpenAIImagesCapabilityNative)) } +func TestAccountSupportsOpenAIEndpointCapability(t *testing.T) { + t.Run("OpenAI APIKey 默认兼容 chat 和 embeddings", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + } + + require.True(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityChatCompletions)) + require.True(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityEmbeddings)) + }) + + t.Run("OpenAI OAuth 默认仅兼容 chat", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeOAuth, + } + + require.True(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityChatCompletions)) + require.False(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityEmbeddings)) + }) + + t.Run("显式列表支持同时声明 chat 和 embeddings", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions", "embeddings"}, + }, + } + + require.True(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityChatCompletions)) + require.True(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityEmbeddings)) + }) + + t.Run("显式列表只声明 chat 时不支持 embeddings", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions"}, + }, + } + + require.True(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityChatCompletions)) + require.False(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityEmbeddings)) + }) + + t.Run("显式 map 支持单独关闭 chat 并开启 embeddings", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Credentials: map[string]any{ + "openai_capabilities": map[string]any{ + "chat_completions": false, + "embeddings": true, + }, + }, + } + + require.False(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityChatCompletions)) + require.True(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapabilityEmbeddings)) + }) + + t.Run("未知能力不应默认放行", func(t *testing.T) { + account := &Account{ + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + } + + require.False(t, account.SupportsOpenAIEndpointCapability(OpenAIEndpointCapability("unknown"))) + }) +} + func TestBuildOpenAIImagesURL_HandlesVersionedBaseURL(t *testing.T) { require.Equal(t, "https://image-upstream.example/v1/images/generations", diff --git a/backend/internal/service/openai_ws_account_sticky_test.go b/backend/internal/service/openai_ws_account_sticky_test.go index 4005a921..c8b28a46 100644 --- a/backend/internal/service/openai_ws_account_sticky_test.go +++ b/backend/internal/service/openai_ws_account_sticky_test.go @@ -268,6 +268,52 @@ func TestOpenAIGatewayService_SelectAccountByPreviousResponseID_BusyKeepsSticky( require.Equal(t, int64(21), selection.WaitPlan.AccountID) } +func TestOpenAIGatewayService_SelectAccountByPreviousResponseID_CapabilityMismatchKeepsSticky(t *testing.T) { + ctx := context.Background() + groupID := int64(25) + account := Account{ + ID: 31, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Credentials: map[string]any{ + "openai_capabilities": []any{"chat_completions"}, + }, + Extra: map[string]any{ + "openai_apikey_responses_websockets_v2_enabled": true, + }, + } + cache := &stubGatewayCache{} + store := NewOpenAIWSStateStore(cache) + cfg := newOpenAIWSV2TestConfig() + svc := &OpenAIGatewayService{ + accountRepo: stubOpenAIAccountRepo{accounts: []Account{account}}, + cache: cache, + cfg: cfg, + concurrencyService: NewConcurrencyService(stubConcurrencyCache{}), + openaiWSStateStore: store, + } + + require.NoError(t, store.BindResponseAccount(ctx, groupID, "resp_prev_capability", account.ID, time.Hour)) + + selection, err := svc.selectAccountByPreviousResponseIDForCapability( + ctx, + &groupID, + "resp_prev_capability", + "text-embedding-3-small", + nil, + OpenAIEndpointCapabilityEmbeddings, + false, + ) + require.NoError(t, err) + require.Nil(t, selection) + boundAccountID, getErr := store.GetResponseAccount(ctx, groupID, "resp_prev_capability") + require.NoError(t, getErr) + require.Equal(t, account.ID, boundAccountID) +} + func newOpenAIWSV2TestConfig() *config.Config { cfg := &config.Config{} cfg.Gateway.OpenAIWS.Enabled = true diff --git a/backend/internal/service/openai_ws_forwarder.go b/backend/internal/service/openai_ws_forwarder.go index b8e558ae..5fd5cffc 100644 --- a/backend/internal/service/openai_ws_forwarder.go +++ b/backend/internal/service/openai_ws_forwarder.go @@ -3987,6 +3987,18 @@ func (s *OpenAIGatewayService) SelectAccountByPreviousResponseID( requestedModel string, excludedIDs map[int64]struct{}, requireCompact bool, +) (*AccountSelectionResult, error) { + return s.selectAccountByPreviousResponseIDForCapability(ctx, groupID, previousResponseID, requestedModel, excludedIDs, "", requireCompact) +} + +func (s *OpenAIGatewayService) selectAccountByPreviousResponseIDForCapability( + ctx context.Context, + groupID *int64, + previousResponseID string, + requestedModel string, + excludedIDs map[int64]struct{}, + requiredCapability OpenAIEndpointCapability, + requireCompact bool, ) (*AccountSelectionResult, error) { if s == nil { return nil, nil @@ -4027,12 +4039,31 @@ func (s *OpenAIGatewayService) SelectAccountByPreviousResponseID( if requestedModel != "" && !account.IsModelSupported(requestedModel) { return nil, nil } - account = s.recheckSelectedOpenAIAccountFromDB(ctx, account, requestedModel, requireCompact) - if account == nil { - _ = store.DeleteResponseAccount(ctx, derefGroupID(groupID), responseID) + if !account.SupportsOpenAIEndpointCapability(requiredCapability) { return nil, nil } - // 兜底：若上游 compact 能力刚被探测为不支持，但 sticky 还在，需要主动放弃。 + if s.schedulerSnapshot != nil && s.accountRepo != nil { + latest, latestErr := s.accountRepo.GetByID(ctx, account.ID) + if latestErr != nil || latest == nil { + _ = store.DeleteResponseAccount(ctx, derefGroupID(groupID), responseID) + return nil, nil + } + if shouldClearStickySession(latest, requestedModel) || !latest.IsOpenAI() || !latest.IsSchedulable() { + _ = store.DeleteResponseAccount(ctx, derefGroupID(groupID), responseID) + return nil, nil + } + if requestedModel != "" && !latest.IsModelSupported(requestedModel) { + return nil, nil + } + if !latest.SupportsOpenAIEndpointCapability(requiredCapability) { + return nil, nil + } + if s.isOpenAIAccountRuntimeBlocked(latest) { + _ = store.DeleteResponseAccount(ctx, derefGroupID(groupID), responseID) + return nil, nil + } + account = latest + } if requireCompact && openAICompactSupportTier(account) == 0 { _ = store.DeleteResponseAccount(ctx, derefGroupID(groupID), responseID) return nil, nil diff --git a/frontend/src/components/account/CreateAccountModal.vue b/frontend/src/components/account/CreateAccountModal.vue index 331295f7..665c4695 100644 --- a/frontend/src/components/account/CreateAccountModal.vue +++ b/frontend/src/components/account/CreateAccountModal.vue @@ -2679,7 +2679,7 @@
@@ -2696,6 +2696,26 @@ />
+
+ +
+ +
+

{{ t('admin.accounts.openai.endpointCapabilitiesDesc') }}

+
@@ -3172,7 +3192,8 @@ import type { CreateAccountRequest, CodexSessionImportMessage, OpenAICompactMode, - OpenAIResponsesMode + OpenAIResponsesMode, + OpenAIEndpointCapability } from '@/types' import BaseDialog from '@/components/common/BaseDialog.vue' import ConfirmDialog from '@/components/common/ConfirmDialog.vue' @@ -3350,6 +3371,7 @@ const autoPauseOnExpired = ref(true) const openaiPassthroughEnabled = ref(false) const openAICompactMode = ref('auto') const openAIResponsesMode = ref('auto') +const openAIEndpointCapabilities = ref(['chat_completions', 'embeddings']) const openaiOAuthResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const openaiAPIKeyResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const codexCLIOnlyEnabled = ref(false) @@ -3412,6 +3434,43 @@ const openAIResponsesModeOptions = computed(() => [ { value: 'force_responses', label: t('admin.accounts.openai.responsesModeForceResponses') }, { value: 'force_chat_completions', label: t('admin.accounts.openai.responsesModeForceChatCompletions') } ]) +const openAIEndpointCapabilityOptions = computed<{ value: OpenAIEndpointCapability; label: string }[]>(() => [ + { value: 'chat_completions', label: t('admin.accounts.openai.capabilityChatCompletions') }, + { value: 'embeddings', label: t('admin.accounts.openai.capabilityEmbeddings') } +]) + +const normalizeOpenAIEndpointCapabilities = (values: OpenAIEndpointCapability[]) => { + const allowed: OpenAIEndpointCapability[] = ['chat_completions', 'embeddings'] + const selected = allowed.filter((value) => values.includes(value)) + return selected.length > 0 ? selected : allowed +} + +const toggleOpenAIEndpointCapability = (capability: OpenAIEndpointCapability, event?: Event) => { + if (openAIEndpointCapabilities.value.includes(capability)) { + if (openAIEndpointCapabilities.value.length <= 1) { + const input = event?.target as HTMLInputElement | null + if (input) input.checked = true + return + } + openAIEndpointCapabilities.value = openAIEndpointCapabilities.value.filter( + (value) => value !== capability + ) + return + } + openAIEndpointCapabilities.value = normalizeOpenAIEndpointCapabilities([ + ...openAIEndpointCapabilities.value, + capability + ]) +} + +const applyOpenAIEndpointCapabilities = (credentials: Record) => { + const capabilities = normalizeOpenAIEndpointCapabilities(openAIEndpointCapabilities.value) + if (capabilities.length === 2) { + delete credentials.openai_capabilities + return + } + credentials.openai_capabilities = capabilities +} function buildAntigravityExtra(): Record | undefined { const extra: Record = {} @@ -3721,6 +3780,7 @@ watch( } if (newPlatform !== 'openai') { openaiPassthroughEnabled.value = false + openAIEndpointCapabilities.value = ['chat_completions', 'embeddings'] openaiOAuthResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF codexCLIOnlyEnabled.value = false @@ -4120,6 +4180,7 @@ const resetForm = () => { openaiPassthroughEnabled.value = false openAICompactMode.value = 'auto' openAIResponsesMode.value = 'auto' + openAIEndpointCapabilities.value = ['chat_completions', 'embeddings'] openaiOAuthResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF codexCLIOnlyEnabled.value = false @@ -4498,6 +4559,7 @@ const handleSubmit = async () => { } } if (form.platform === 'openai') { + applyOpenAIEndpointCapabilities(credentials) const compactModelMapping = buildOpenAICompactModelMapping() if (compactModelMapping) { credentials.compact_model_mapping = compactModelMapping @@ -4620,6 +4682,9 @@ const createAccountAndFinish = async ( } } if (platform === 'openai') { + if (type === 'apikey') { + applyOpenAIEndpointCapabilities(credentials) + } const compactModelMapping = buildOpenAICompactModelMapping() if (compactModelMapping) { credentials.compact_model_mapping = compactModelMapping diff --git a/frontend/src/components/account/EditAccountModal.vue b/frontend/src/components/account/EditAccountModal.vue index f44b5d38..3cb10591 100644 --- a/frontend/src/components/account/EditAccountModal.vue +++ b/frontend/src/components/account/EditAccountModal.vue @@ -1439,7 +1439,7 @@
@@ -1459,6 +1459,26 @@
{{ t(openAIResponsesStatusKey) }}
+
+ +
+ +
+

{{ t('admin.accounts.openai.endpointCapabilitiesDesc') }}

+
@@ -2245,7 +2265,15 @@ import { useAppStore } from '@/stores/app' import { useAuthStore } from '@/stores/auth' import { adminAPI } from '@/api/admin' import { useQuotaNotifyState } from '@/composables/useQuotaNotifyState' -import type { Account, Proxy, AdminGroup, CheckMixedChannelResponse, OpenAICompactMode, OpenAIResponsesMode } from '@/types' +import type { + Account, + Proxy, + AdminGroup, + CheckMixedChannelResponse, + OpenAICompactMode, + OpenAIResponsesMode, + OpenAIEndpointCapability +} from '@/types' import BaseDialog from '@/components/common/BaseDialog.vue' import ConfirmDialog from '@/components/common/ConfirmDialog.vue' import Select from '@/components/common/Select.vue' @@ -2433,6 +2461,7 @@ const customBaseUrl = ref('') const openaiPassthroughEnabled = ref(false) const openAICompactMode = ref('auto') const openAIResponsesMode = ref('auto') +const openAIEndpointCapabilities = ref(['chat_completions', 'embeddings']) const openaiOAuthResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const openaiAPIKeyResponsesWebSocketV2Mode = ref(OPENAI_WS_MODE_OFF) const codexCLIOnlyEnabled = ref(false) @@ -2539,6 +2568,63 @@ const openAIResponsesModeOptions = computed(() => [ { value: 'force_responses', label: t('admin.accounts.openai.responsesModeForceResponses') }, { value: 'force_chat_completions', label: t('admin.accounts.openai.responsesModeForceChatCompletions') } ]) +const openAIEndpointCapabilityOptions = computed<{ value: OpenAIEndpointCapability; label: string }[]>(() => [ + { value: 'chat_completions', label: t('admin.accounts.openai.capabilityChatCompletions') }, + { value: 'embeddings', label: t('admin.accounts.openai.capabilityEmbeddings') } +]) + +const normalizeOpenAIEndpointCapabilities = (values: OpenAIEndpointCapability[]) => { + const allowed: OpenAIEndpointCapability[] = ['chat_completions', 'embeddings'] + const selected = allowed.filter((value) => values.includes(value)) + return selected.length > 0 ? selected : allowed +} + +const readOpenAIEndpointCapabilities = (credentials?: Record): OpenAIEndpointCapability[] => { + const raw = credentials?.openai_capabilities + if (Array.isArray(raw)) { + return normalizeOpenAIEndpointCapabilities( + raw.filter((value): value is OpenAIEndpointCapability => + value === 'chat_completions' || value === 'embeddings' + ) + ) + } + if (raw !== null && typeof raw === 'object') { + const capabilityMap = raw as Record + return normalizeOpenAIEndpointCapabilities( + openAIEndpointCapabilityOptions.value + .map((option) => option.value) + .filter((value) => capabilityMap[value] === true) + ) + } + return ['chat_completions', 'embeddings'] +} + +const toggleOpenAIEndpointCapability = (capability: OpenAIEndpointCapability, event?: Event) => { + if (openAIEndpointCapabilities.value.includes(capability)) { + if (openAIEndpointCapabilities.value.length <= 1) { + const input = event?.target as HTMLInputElement | null + if (input) input.checked = true + return + } + openAIEndpointCapabilities.value = openAIEndpointCapabilities.value.filter( + (value) => value !== capability + ) + return + } + openAIEndpointCapabilities.value = normalizeOpenAIEndpointCapabilities([ + ...openAIEndpointCapabilities.value, + capability + ]) +} + +const applyOpenAIEndpointCapabilities = (credentials: Record) => { + const capabilities = normalizeOpenAIEndpointCapabilities(openAIEndpointCapabilities.value) + if (capabilities.length === 2) { + delete credentials.openai_capabilities + return + } + credentials.openai_capabilities = capabilities +} const normalizeOpenAIResponsesMode = (mode: unknown): OpenAIResponsesMode => { if (mode === 'force_responses' || mode === 'force_chat_completions') { return mode @@ -2724,6 +2810,7 @@ const syncFormFromAccount = (newAccount: Account | null) => { openaiPassthroughEnabled.value = false openAICompactMode.value = 'auto' openAIResponsesMode.value = 'auto' + openAIEndpointCapabilities.value = ['chat_completions', 'embeddings'] openAICompactModelMappings.value = [] openaiOAuthResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF openaiAPIKeyResponsesWebSocketV2Mode.value = OPENAI_WS_MODE_OFF @@ -2736,6 +2823,9 @@ const syncFormFromAccount = (newAccount: Account | null) => { openAICompactMode.value = (extra?.openai_compact_mode as OpenAICompactMode) || 'auto' if (newAccount.type === 'apikey') { openAIResponsesMode.value = normalizeOpenAIResponsesMode(extra?.openai_responses_mode) + openAIEndpointCapabilities.value = readOpenAIEndpointCapabilities( + newAccount.credentials as Record | undefined + ) } const codexImageGenerationBridgeValue = typeof extra?.codex_image_generation_bridge === 'boolean' ? extra.codex_image_generation_bridge @@ -3476,6 +3566,7 @@ const handleSubmit = async () => { newCredentials.model_mapping = currentCredentials.model_mapping } if (props.account.platform === 'openai') { + applyOpenAIEndpointCapabilities(newCredentials) const compactModelMapping = buildModelMappingObject('mapping', [], openAICompactModelMappings.value) if (compactModelMapping) { newCredentials.compact_model_mapping = compactModelMapping diff --git a/frontend/src/components/account/__tests__/EditAccountModal.spec.ts b/frontend/src/components/account/__tests__/EditAccountModal.spec.ts index 0b8e939c..db012a30 100644 --- a/frontend/src/components/account/__tests__/EditAccountModal.spec.ts +++ b/frontend/src/components/account/__tests__/EditAccountModal.spec.ts @@ -310,6 +310,63 @@ describe('EditAccountModal', () => { expect(updateAccountMock.mock.calls[0]?.[1]?.extra?.openai_responses_supported).toBe(true) }) + it('submits OpenAI APIKey endpoint capabilities from credentials', async () => { + const account = buildAccount() + account.credentials.openai_capabilities = ['chat_completions'] + updateAccountMock.mockReset() + checkMixedChannelRiskMock.mockReset() + checkMixedChannelRiskMock.mockResolvedValue({ has_risk: false }) + updateAccountMock.mockResolvedValue(account) + + const wrapper = mountModal(account) + + expect(wrapper.findAll('input[type="checkbox"]').some((input) => (input.element as HTMLInputElement).checked)).toBe(true) + + await wrapper.get('form#edit-account-form').trigger('submit.prevent') + + expect(updateAccountMock).toHaveBeenCalledTimes(1) + expect(updateAccountMock.mock.calls[0]?.[1]?.credentials?.openai_capabilities).toEqual([ + 'chat_completions' + ]) + }) + + it('keeps at least one OpenAI APIKey endpoint capability selected', async () => { + const account = buildAccount() + updateAccountMock.mockReset() + checkMixedChannelRiskMock.mockReset() + checkMixedChannelRiskMock.mockResolvedValue({ has_risk: false }) + updateAccountMock.mockResolvedValue(account) + + const wrapper = mountModal(account) + + const chatCheckbox = wrapper.get( + '[data-testid="openai-endpoint-capability-chat_completions"]' + ) + const embeddingsCheckbox = wrapper.get( + '[data-testid="openai-endpoint-capability-embeddings"]' + ) + + expect(chatCheckbox.element.checked).toBe(true) + expect(embeddingsCheckbox.element.checked).toBe(true) + + await embeddingsCheckbox.setValue(false) + + expect(chatCheckbox.element.checked).toBe(true) + expect(embeddingsCheckbox.element.checked).toBe(false) + + await chatCheckbox.setValue(false) + + expect(chatCheckbox.element.checked).toBe(true) + expect(embeddingsCheckbox.element.checked).toBe(false) + + await wrapper.get('form#edit-account-form').trigger('submit.prevent') + + expect(updateAccountMock).toHaveBeenCalledTimes(1) + expect(updateAccountMock.mock.calls[0]?.[1]?.credentials?.openai_capabilities).toEqual([ + 'chat_completions' + ]) + }) + it('submits account-level Codex image generation bridge override', async () => { const account = buildAccount() account.extra = { diff --git a/frontend/src/i18n/locales/en.ts b/frontend/src/i18n/locales/en.ts index 41c3c495..ec659dd4 100644 --- a/frontend/src/i18n/locales/en.ts +++ b/frontend/src/i18n/locales/en.ts @@ -3353,6 +3353,11 @@ export default { responsesModeAuto: 'Auto', responsesModeForceResponses: 'Force Responses', responsesModeForceChatCompletions: 'Force Chat Completions', + endpointCapabilities: 'Endpoint capabilities', + endpointCapabilitiesDesc: + 'Used by account routing. Both endpoints are allowed by default; if the upstream only supports one, select only the supported endpoint.', + capabilityChatCompletions: 'Chat Completions', + capabilityEmbeddings: 'Embeddings', responsesStatusAutoSupported: 'Auto probe: Responses', responsesStatusAutoUnsupported: 'Auto probe: Chat Completions', responsesStatusAutoUnknown: 'Auto probe: unknown', diff --git a/frontend/src/i18n/locales/zh.ts b/frontend/src/i18n/locales/zh.ts index 8ff8ea80..36b0d8c3 100644 --- a/frontend/src/i18n/locales/zh.ts +++ b/frontend/src/i18n/locales/zh.ts @@ -3499,6 +3499,11 @@ export default { responsesModeAuto: '自动', responsesModeForceResponses: '强制 Responses', responsesModeForceChatCompletions: '强制 Chat Completions', + endpointCapabilities: '端点能力', + endpointCapabilitiesDesc: + '用于调度筛选。默认两个端点都可用；如果上游只支持其中一个，请只勾选实际支持的端点。', + capabilityChatCompletions: 'Chat Completions', + capabilityEmbeddings: 'Embeddings', responsesStatusAutoSupported: '自动探测：Responses', responsesStatusAutoUnsupported: '自动探测：Chat Completions', responsesStatusAutoUnknown: '自动探测：未探测', diff --git a/frontend/src/types/index.ts b/frontend/src/types/index.ts index eae5e455..c2136169 100644 --- a/frontend/src/types/index.ts +++ b/frontend/src/types/index.ts @@ -997,6 +997,7 @@ export interface CodexUsageSnapshot { export type OpenAICompactMode = 'auto' | 'force_on' | 'force_off' export type OpenAIResponsesMode = 'auto' | 'force_responses' | 'force_chat_completions' +export type OpenAIEndpointCapability = 'chat_completions' | 'embeddings' export interface OpenAICompactState { openai_compact_mode?: OpenAICompactMode From 37044b83eb67ed8680c3edc7249c9c2a75dd56fa Mon Sep 17 00:00:00 2001 From: shaw Date: Fri, 29 May 2026 09:23:06 +0800 Subject: [PATCH 11/20] fix(openai): clarify endpoint capability UI --- .../components/account/CreateAccountModal.vue | 31 +++++++++++++- .../components/account/EditAccountModal.vue | 42 +++++++++++++++++-- .../__tests__/EditAccountModal.spec.ts | 31 ++++++++++++++ frontend/src/i18n/locales/en.ts | 10 ++++- frontend/src/i18n/locales/zh.ts | 9 +++- 5 files changed, 114 insertions(+), 9 deletions(-) diff --git a/frontend/src/components/account/CreateAccountModal.vue b/frontend/src/components/account/CreateAccountModal.vue index 665c4695..5f5a11d7 100644 --- a/frontend/src/components/account/CreateAccountModal.vue +++ b/frontend/src/components/account/CreateAccountModal.vue @@ -2692,10 +2692,18 @@ +

{{ t('admin.accounts.autoPauseThresholdHint') }}

+
+
+ + +

{{ t('admin.accounts.autoPauseThresholdHint') }}

+
+
+
([]) const customErrorCodeInput = ref(null) const interceptWarmupRequests = ref(false) const autoPauseOnExpired = ref(false) +const autoPause5hThreshold = ref(null) +const autoPause7dThreshold = ref(null) const mixedScheduling = ref(false) // For antigravity accounts: enable mixed scheduling const allowOverages = ref(false) // For antigravity accounts: enable AI Credits overages const antigravityModelRestrictionMode = ref<'whitelist' | 'mapping'>('whitelist') @@ -2862,9 +2896,11 @@ const syncFormFromAccount = (newAccount: Account | null) => { // Load mixed scheduling setting (only for antigravity accounts) mixedScheduling.value = false allowOverages.value = false - const extra = newAccount.extra as Record | undefined - mixedScheduling.value = extra?.mixed_scheduling === true - allowOverages.value = extra?.allow_overages === true + const extra = newAccount.extra as Record | undefined + mixedScheduling.value = extra?.mixed_scheduling === true + allowOverages.value = extra?.allow_overages === true + autoPause5hThreshold.value = typeof extra?.auto_pause_5h_threshold === 'number' ? extra.auto_pause_5h_threshold * 100 : null + autoPause7dThreshold.value = typeof extra?.auto_pause_7d_threshold === 'number' ? extra.auto_pause_7d_threshold * 100 : null // Load OpenAI passthrough toggle (OpenAI OAuth/API Key) openaiPassthroughEnabled.value = false @@ -3987,9 +4023,9 @@ const handleSubmit = async () => { } // For OpenAI OAuth/API Key accounts, handle passthrough mode in extra - if (props.account.platform === 'openai' && (props.account.type === 'oauth' || props.account.type === 'apikey')) { - const currentExtra = (props.account.extra as Record) || {} - const newExtra: Record = { ...currentExtra } + if (props.account.platform === 'openai' && (props.account.type === 'oauth' || props.account.type === 'apikey')) { + const currentExtra = (props.account.extra as Record) || {} + const newExtra: Record = { ...currentExtra } const hadCodexCLIOnlyEnabled = currentExtra.codex_cli_only === true if (props.account.type === 'oauth') { newExtra.openai_oauth_responses_websockets_v2_mode = openaiOAuthResponsesWebSocketV2Mode.value @@ -4011,15 +4047,25 @@ const handleSubmit = async () => { } else { newExtra.openai_compact_mode = openAICompactMode.value } - if (props.account.type === 'apikey') { + if (props.account.type === 'apikey') { if (!openAITextGenerationCapabilityEnabled.value || openAIResponsesMode.value === 'auto') { delete newExtra.openai_responses_mode } else { newExtra.openai_responses_mode = openAIResponsesMode.value } - } + } + if (autoPause5hThreshold.value != null && autoPause5hThreshold.value > 0) { + newExtra.auto_pause_5h_threshold = autoPause5hThreshold.value / 100 + } else { + delete newExtra.auto_pause_5h_threshold + } + if (autoPause7dThreshold.value != null && autoPause7dThreshold.value > 0) { + newExtra.auto_pause_7d_threshold = autoPause7dThreshold.value / 100 + } else { + delete newExtra.auto_pause_7d_threshold + } - delete newExtra.codex_image_generation_bridge_enabled + delete newExtra.codex_image_generation_bridge_enabled if (codexImageGenerationBridgeMode.value === 'inherit') { delete newExtra.codex_image_generation_bridge } else { diff --git a/frontend/src/components/account/__tests__/EditAccountModal.spec.ts b/frontend/src/components/account/__tests__/EditAccountModal.spec.ts index 4561924f..6db63831 100644 --- a/frontend/src/components/account/__tests__/EditAccountModal.spec.ts +++ b/frontend/src/components/account/__tests__/EditAccountModal.spec.ts @@ -330,6 +330,28 @@ describe('EditAccountModal', () => { ]) }) + it('submits OpenAI quota auto-pause thresholds in extra', async () => { + const account = buildAccount() + account.extra = { + auto_pause_5h_threshold: 0.9, + auto_pause_7d_threshold: 0.8 + } + updateAccountMock.mockReset() + checkMixedChannelRiskMock.mockReset() + checkMixedChannelRiskMock.mockResolvedValue({ has_risk: false }) + updateAccountMock.mockResolvedValue(account) + + const wrapper = mountModal(account) + + await wrapper.get('[data-testid="auto-pause-5h-threshold"]').setValue('95') + await wrapper.get('[data-testid="auto-pause-7d-threshold"]').setValue('96') + await wrapper.get('form#edit-account-form').trigger('submit.prevent') + + expect(updateAccountMock).toHaveBeenCalledTimes(1) + expect(updateAccountMock.mock.calls[0]?.[1]?.extra?.auto_pause_5h_threshold).toBe(0.95) + expect(updateAccountMock.mock.calls[0]?.[1]?.extra?.auto_pause_7d_threshold).toBe(0.96) + }) + it('keeps at least one OpenAI APIKey endpoint capability selected', async () => { const account = buildAccount() updateAccountMock.mockReset() diff --git a/frontend/src/i18n/locales/en.ts b/frontend/src/i18n/locales/en.ts index b2aeb2f8..fa2e5a92 100644 --- a/frontend/src/i18n/locales/en.ts +++ b/frontend/src/i18n/locales/en.ts @@ -3475,6 +3475,9 @@ export default { 'When enabled, warmup requests like title generation will return mock responses without consuming upstream tokens', autoPauseOnExpired: 'Auto Pause On Expired', autoPauseOnExpiredDesc: 'When enabled, the account will auto pause scheduling after it expires', + autoPause5hThreshold: '5h Usage Threshold (%)', + autoPause7dThreshold: '7d Usage Threshold (%)', + autoPauseThresholdHint: 'Leave empty or set 0 to disable. Reaching the threshold only skips the account during scheduling and does not modify schedulable.', // Quota control (Anthropic OAuth/SetupToken only) quotaControl: { title: 'Quota Control', diff --git a/frontend/src/i18n/locales/zh.ts b/frontend/src/i18n/locales/zh.ts index 85d1feee..2364f9c4 100644 --- a/frontend/src/i18n/locales/zh.ts +++ b/frontend/src/i18n/locales/zh.ts @@ -3613,6 +3613,9 @@ export default { interceptWarmupRequestsDesc: '启用后，标题生成等预热请求将返回 mock 响应，不消耗上游 token', autoPauseOnExpired: '过期自动暂停调度', autoPauseOnExpiredDesc: '启用后，账号过期将自动暂停调度', + autoPause5hThreshold: '5h 用量阈值(%)', + autoPause7dThreshold: '7d 用量阈值(%)', + autoPauseThresholdHint: '填 0 或留空表示不启用；达到阈值后仅在调度时跳过账号，不修改 schedulable。', // Quota control (Anthropic OAuth/SetupToken only) quotaControl: { title: '配额控制', From 8b7a8227060e3eac20b5c0e331b50b2dc99f5e13 Mon Sep 17 00:00:00 2001 From: wucm667 Date: Fri, 29 May 2026 12:20:30 +0800 Subject: [PATCH 16/20] fix(account): address review on OpenAI quota auto-pause - gate previous_response_id sticky path with quota auto-pause check at both the snapshot and DB-recheck stages (previously bypassed, #1) - skip pausing when the usage window already reset to avoid a stale stuck-pause; carry codex_*_reset_at / reset_after_seconds / codex_usage_updated_at through the scheduler snapshot whitelist (#2) - remove the incomplete limit mode; percentage threshold only (#3) - add global default 5h/7d threshold inputs to the Ops settings dialog with validation and en/zh i18n (#4) - downgrade account_auto_paused_by_quota log from Info to Debug; it fires per-candidate on the scheduling hot path (#5) Co-Authored-By: Claude Opus 4.8 --- .../internal/repository/scheduler_cache.go | 7 +- .../repository/scheduler_cache_unit_test.go | 18 +++-- .../service/openai_account_scheduler_test.go | 58 +++++++++++++-- .../service/openai_gateway_service.go | 72 +++++++++++++------ .../service/openai_ws_account_sticky_test.go | 40 +++++++++++ .../internal/service/openai_ws_forwarder.go | 10 +++ frontend/src/api/admin/ops.ts | 6 ++ frontend/src/i18n/locales/en.ts | 8 ++- frontend/src/i18n/locales/zh.ts | 8 ++- .../ops/components/OpsSettingsDialog.vue | 65 +++++++++++++++++ 10 files changed, 257 insertions(+), 35 deletions(-) diff --git a/backend/internal/repository/scheduler_cache.go b/backend/internal/repository/scheduler_cache.go index ec8c72dc..ff3c4301 100644 --- a/backend/internal/repository/scheduler_cache.go +++ b/backend/internal/repository/scheduler_cache.go @@ -550,10 +550,13 @@ func filterSchedulerExtra(extra map[string]any) map[string]any { "openai_responses_supported", "codex_5h_used_percent", "codex_7d_used_percent", + "codex_5h_reset_at", + "codex_7d_reset_at", + "codex_5h_reset_after_seconds", + "codex_7d_reset_after_seconds", + "codex_usage_updated_at", "auto_pause_5h_threshold", "auto_pause_7d_threshold", - "auto_pause_5h_limit", - "auto_pause_7d_limit", } filtered := make(map[string]any) for _, key := range keys { diff --git a/backend/internal/repository/scheduler_cache_unit_test.go b/backend/internal/repository/scheduler_cache_unit_test.go index fabc6bad..9e4ec23e 100644 --- a/backend/internal/repository/scheduler_cache_unit_test.go +++ b/backend/internal/repository/scheduler_cache_unit_test.go @@ -80,10 +80,15 @@ func TestBuildSchedulerMetadataAccount_KeepsQuotaAutoPauseFields(t *testing.T) { account := service.Account{ ID: 88, Extra: map[string]any{ - "codex_5h_used_percent": 12.34, - "codex_7d_used_percent": 56.78, - "auto_pause_5h_threshold": 0.95, - "auto_pause_7d_threshold": 0.96, + "codex_5h_used_percent": 12.34, + "codex_7d_used_percent": 56.78, + "codex_5h_reset_at": "2026-05-29T10:00:00Z", + "codex_7d_reset_at": "2026-06-01T10:00:00Z", + "codex_5h_reset_after_seconds": 300, + "codex_7d_reset_after_seconds": 600, + "codex_usage_updated_at": "2026-05-29T09:00:00Z", + "auto_pause_5h_threshold": 0.95, + "auto_pause_7d_threshold": 0.96, }, } @@ -91,6 +96,11 @@ func TestBuildSchedulerMetadataAccount_KeepsQuotaAutoPauseFields(t *testing.T) { require.Equal(t, 12.34, got.Extra["codex_5h_used_percent"]) require.Equal(t, 56.78, got.Extra["codex_7d_used_percent"]) + require.Equal(t, "2026-05-29T10:00:00Z", got.Extra["codex_5h_reset_at"]) + require.Equal(t, "2026-06-01T10:00:00Z", got.Extra["codex_7d_reset_at"]) + require.Equal(t, 300, got.Extra["codex_5h_reset_after_seconds"]) + require.Equal(t, 600, got.Extra["codex_7d_reset_after_seconds"]) + require.Equal(t, "2026-05-29T09:00:00Z", got.Extra["codex_usage_updated_at"]) require.Equal(t, 0.95, got.Extra["auto_pause_5h_threshold"]) require.Equal(t, 0.96, got.Extra["auto_pause_7d_threshold"]) } diff --git a/backend/internal/service/openai_account_scheduler_test.go b/backend/internal/service/openai_account_scheduler_test.go index 37810870..531769a7 100644 --- a/backend/internal/service/openai_account_scheduler_test.go +++ b/backend/internal/service/openai_account_scheduler_test.go @@ -704,7 +704,6 @@ func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_AutoPauseBy5hT Extra: map[string]any{ "codex_5h_used_percent": 95.0, "auto_pause_5h_threshold": 0.95, - "auto_pause_5h_limit": 100, }, } secondary := Account{ID: 35002, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} @@ -729,7 +728,6 @@ func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_AllowsBelow5hT Extra: map[string]any{ "codex_5h_used_percent": 80.0, "auto_pause_5h_threshold": 0.95, - "auto_pause_5h_limit": 100, }, } secondary := Account{ID: 35102, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} @@ -754,7 +752,6 @@ func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_AutoPauseBy7dT Extra: map[string]any{ "codex_7d_used_percent": 95.0, "auto_pause_7d_threshold": 0.95, - "auto_pause_7d_limit": 200, }, } secondary := Account{ID: 35202, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} @@ -790,7 +787,6 @@ func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_UsesGlobalDefa Priority: 0, Extra: map[string]any{ "codex_5h_used_percent": 95.0, - "auto_pause_5h_limit": 100, }, } secondary := Account{ID: 35402, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} @@ -802,6 +798,60 @@ func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_UsesGlobalDefa require.Equal(t, int64(35402), account.ID) } +func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_StaleUsageWindowResetSkipsPause(t *testing.T) { + ctx := context.Background() + // Usage is over threshold but the window's reset time has already passed, so the + // cached percentage is stale (the real window rolled over) and the account must NOT + // stay paused — otherwise it could be skipped forever with no traffic to refresh it. + primary := Account{ + ID: 35501, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Extra: map[string]any{ + "codex_5h_used_percent": 99.0, + "auto_pause_5h_threshold": 0.95, + "codex_5h_reset_at": time.Now().Add(-time.Minute).Format(time.RFC3339), + }, + } + secondary := Account{ID: 35502, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} + svc := &OpenAIGatewayService{accountRepo: schedulerTestOpenAIAccountRepo{accounts: []Account{primary, secondary}}, cfg: &config.Config{}} + + account, err := svc.SelectAccountForModelWithExclusions(ctx, nil, "", "gpt-5.1", nil) + require.NoError(t, err) + require.NotNil(t, account) + require.Equal(t, int64(35501), account.ID) +} + +func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_FreshUsageWindowStillPauses(t *testing.T) { + ctx := context.Background() + // Same as above but the window has not reset yet, so the account stays paused. + primary := Account{ + ID: 35601, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Extra: map[string]any{ + "codex_5h_used_percent": 99.0, + "auto_pause_5h_threshold": 0.95, + "codex_5h_reset_at": time.Now().Add(time.Hour).Format(time.RFC3339), + }, + } + secondary := Account{ID: 35602, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} + svc := &OpenAIGatewayService{accountRepo: schedulerTestOpenAIAccountRepo{accounts: []Account{primary, secondary}}, cfg: &config.Config{}} + + account, err := svc.SelectAccountForModelWithExclusions(ctx, nil, "", "gpt-5.1", nil) + require.NoError(t, err) + require.NotNil(t, account) + require.Equal(t, int64(35602), account.ID) +} + func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_SkipsFreshlyRateLimitedSnapshotCandidate(t *testing.T) { ctx := context.Background() groupID := int64(10102) diff --git a/backend/internal/service/openai_gateway_service.go b/backend/internal/service/openai_gateway_service.go index 268f985c..e2534cc2 100644 --- a/backend/internal/service/openai_gateway_service.go +++ b/backend/internal/service/openai_gateway_service.go @@ -1328,13 +1328,12 @@ func isOpenAIAccountEligibleForRequest(ctx context.Context, account *Account, re return false } if paused, reason := shouldAutoPauseOpenAIAccountByQuota(ctx, account); paused { - slog.Info("account_auto_paused_by_quota", + // Debug level: this fires per-candidate on the scheduling hot path, so Info + // would amplify into log spam once several accounts cross the threshold. + slog.Debug("account_auto_paused_by_quota", "account_id", account.ID, - "usage_5h_percent", readOpenAIQuotaUsedPercent(account.Extra, "5h"), - "usage_7d_percent", readOpenAIQuotaUsedPercent(account.Extra, "7d"), - "threshold_type", reason.window, + "window", reason.window, "threshold", reason.threshold, - "limit", reason.limit, "utilization", reason.utilization, ) return false @@ -1354,7 +1353,6 @@ func isOpenAIAccountEligibleForRequest(ctx context.Context, account *Account, re type openAIQuotaAutoPauseDecision struct { window string threshold float64 - limit float64 utilization float64 } @@ -1363,18 +1361,15 @@ func shouldAutoPauseOpenAIAccountByQuota(ctx context.Context, account *Account) return false, openAIQuotaAutoPauseDecision{} } threshold5h, threshold7d := resolveOpenAIQuotaAutoPauseThresholds(ctx, account) + now := time.Now() if threshold5h > 0 { - if utilization, limit, ok := resolveOpenAIQuotaUtilization(account.Extra, "5h"); ok { - if utilization >= threshold5h { - return true, openAIQuotaAutoPauseDecision{window: "5h", threshold: threshold5h, limit: limit, utilization: utilization} - } + if utilization, ok := resolveOpenAIQuotaUtilization(account.Extra, "5h", now); ok && utilization >= threshold5h { + return true, openAIQuotaAutoPauseDecision{window: "5h", threshold: threshold5h, utilization: utilization} } } if threshold7d > 0 { - if utilization, limit, ok := resolveOpenAIQuotaUtilization(account.Extra, "7d"); ok { - if utilization >= threshold7d { - return true, openAIQuotaAutoPauseDecision{window: "7d", threshold: threshold7d, limit: limit, utilization: utilization} - } + if utilization, ok := resolveOpenAIQuotaUtilization(account.Extra, "7d", now); ok && utilization >= threshold7d { + return true, openAIQuotaAutoPauseDecision{window: "7d", threshold: threshold7d, utilization: utilization} } } return false, openAIQuotaAutoPauseDecision{} @@ -1431,18 +1426,49 @@ func resolveAccountExtraNumber(extra map[string]any, keys ...string) (float64, b return 0, false } -func resolveOpenAIQuotaUtilization(extra map[string]any, window string) (float64, float64, bool) { - limitKeys := []string{"auto_pause_" + window + "_limit", "quota_" + window + "_limit", window + "_limit"} - if limit, ok := resolveAccountExtraNumber(extra, limitKeys...); ok && limit > 0 { - if usage, ok := resolveAccountExtraNumber(extra, "usage_"+window); ok && usage >= 0 { - return usage / limit, limit, true - } - } +// resolveOpenAIQuotaUtilization returns the current utilization ratio (0..1) for the +// given Codex usage window. ok=false means there is no usable signal to pause on: +// either no snapshot exists, or the window has already rolled over so the cached +// percentage is stale. The stale guard matters because a paused account stops +// receiving requests, so its snapshot is never refreshed from upstream headers — +// without this check an old used_percent would keep the account paused forever even +// after the real window reset. +func resolveOpenAIQuotaUtilization(extra map[string]any, window string, now time.Time) (float64, bool) { usedPercent := readOpenAIQuotaUsedPercent(extra, window) if usedPercent <= 0 { - return 0, 0, false + return 0, false } - return usedPercent / 100, 100, true + if openAIQuotaWindowReset(extra, window, now) { + return 0, false + } + return usedPercent / 100, true +} + +// openAIQuotaWindowReset reports whether the Codex usage window's reset time has +// already passed relative to now. It prefers the absolute codex__reset_at +// timestamp and falls back to codex__reset_after_seconds anchored at +// codex_usage_updated_at, mirroring AccountUsageService's window-progress logic. +func openAIQuotaWindowReset(extra map[string]any, window string, now time.Time) bool { + if len(extra) == 0 { + return false + } + if resetAtRaw, ok := extra["codex_"+window+"_reset_at"]; ok { + if resetAt, err := parseTime(fmt.Sprint(resetAtRaw)); err == nil { + return !now.Before(resetAt) + } + } + resetAfter := parseExtraInt(extra["codex_"+window+"_reset_after_seconds"]) + if resetAfter <= 0 { + return false + } + base := now + if updatedRaw, ok := extra["codex_usage_updated_at"]; ok { + if updatedAt, err := parseTime(fmt.Sprint(updatedRaw)); err == nil { + base = updatedAt + } + } + resetAt := base.Add(time.Duration(resetAfter) * time.Second) + return !now.Before(resetAt) } func readOpenAIQuotaUsedPercent(extra map[string]any, window string) float64 { diff --git a/backend/internal/service/openai_ws_account_sticky_test.go b/backend/internal/service/openai_ws_account_sticky_test.go index c8b28a46..6fc44298 100644 --- a/backend/internal/service/openai_ws_account_sticky_test.go +++ b/backend/internal/service/openai_ws_account_sticky_test.go @@ -48,6 +48,46 @@ func TestOpenAIGatewayService_SelectAccountByPreviousResponseID_Hit(t *testing.T } } +func TestOpenAIGatewayService_SelectAccountByPreviousResponseID_QuotaAutoPausedMiss(t *testing.T) { + ctx := context.Background() + groupID := int64(23) + account := Account{ + ID: 77, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 2, + Extra: map[string]any{ + "openai_apikey_responses_websockets_v2_enabled": true, + "codex_5h_used_percent": 96.0, + "auto_pause_5h_threshold": 0.95, + }, + } + cache := &stubGatewayCache{} + store := NewOpenAIWSStateStore(cache) + cfg := newOpenAIWSV2TestConfig() + svc := &OpenAIGatewayService{ + accountRepo: stubOpenAIAccountRepo{accounts: []Account{account}}, + cache: cache, + cfg: cfg, + concurrencyService: NewConcurrencyService(stubConcurrencyCache{}), + openaiWSStateStore: store, + } + + require.NoError(t, store.BindResponseAccount(ctx, groupID, "resp_prev_quota", account.ID, time.Hour)) + + selection, err := svc.SelectAccountByPreviousResponseID(ctx, &groupID, "resp_prev_quota", "gpt-5.1", nil, false) + require.NoError(t, err) + require.Nil(t, selection, "超过 5h 配额阈值的账号不应继续命中 previous_response_id 粘连") + + // Auto-pause is transient, so the binding is preserved: the chain can resume on the + // same account once the quota window resets. + boundAccountID, getErr := store.GetResponseAccount(ctx, groupID, "resp_prev_quota") + require.NoError(t, getErr) + require.Equal(t, account.ID, boundAccountID) +} + func TestOpenAIGatewayService_SelectAccountByPreviousResponseID_RateLimitedMiss(t *testing.T) { ctx := context.Background() groupID := int64(23) diff --git a/backend/internal/service/openai_ws_forwarder.go b/backend/internal/service/openai_ws_forwarder.go index 6eea0191..878ff486 100644 --- a/backend/internal/service/openai_ws_forwarder.go +++ b/backend/internal/service/openai_ws_forwarder.go @@ -4045,6 +4045,13 @@ func (s *OpenAIGatewayService) selectAccountByPreviousResponseIDForCapability( if !account.SupportsOpenAIEndpointCapability(requiredCapability) { return nil, nil } + // Quota auto-pause must also gate the previous_response_id sticky path; otherwise an + // account over its 5h/7d threshold keeps serving the same response chain even though + // normal scheduling skips it. Pause is transient, so fall through to normal scheduling + // without deleting the binding (the window may reset before the next turn). + if paused, _ := shouldAutoPauseOpenAIAccountByQuota(ctx, account); paused { + return nil, nil + } if s.schedulerSnapshot != nil && s.accountRepo != nil { latest, latestErr := s.accountRepo.GetByID(ctx, account.ID) if latestErr != nil || latest == nil { @@ -4061,6 +4068,9 @@ func (s *OpenAIGatewayService) selectAccountByPreviousResponseIDForCapability( if !latest.SupportsOpenAIEndpointCapability(requiredCapability) { return nil, nil } + if paused, _ := shouldAutoPauseOpenAIAccountByQuota(ctx, latest); paused { + return nil, nil + } if s.isOpenAIAccountRuntimeBlocked(latest) { _ = store.DeleteResponseAccount(ctx, derefGroupID(groupID), responseID) return nil, nil diff --git a/frontend/src/api/admin/ops.ts b/frontend/src/api/admin/ops.ts index 69235668..847fc8c9 100644 --- a/frontend/src/api/admin/ops.ts +++ b/frontend/src/api/admin/ops.ts @@ -778,9 +778,15 @@ export interface OpsAlertRuntimeSettings { thresholds: OpsMetricThresholds // 指标阈值配置 } +export interface OpsOpenAIAccountQuotaAutoPauseSettings { + default_threshold_5h: number // 0~1，0 表示不启用全局默认 5h 阈值 + default_threshold_7d: number // 0~1，0 表示不启用全局默认 7d 阈值 +} + export interface OpsAdvancedSettings { data_retention: OpsDataRetentionSettings aggregation: OpsAggregationSettings + openai_account_quota_auto_pause: OpsOpenAIAccountQuotaAutoPauseSettings ignore_count_tokens_errors: boolean ignore_context_canceled: boolean ignore_no_available_accounts: boolean diff --git a/frontend/src/i18n/locales/en.ts b/frontend/src/i18n/locales/en.ts index fa2e5a92..8ab90961 100644 --- a/frontend/src/i18n/locales/en.ts +++ b/frontend/src/i18n/locales/en.ts @@ -5193,6 +5193,11 @@ export default { aggregation: 'Pre-aggregation Tasks', enableAggregation: 'Enable Pre-aggregation', aggregationHint: 'Pre-aggregation improves query performance for long time windows', + openaiQuotaAutoPause: 'OpenAI Account Quota Auto-pause', + openaiQuotaAutoPauseHint: 'When an OpenAI account reaches its 5h / 7d usage threshold, the scheduler skips it automatically and resumes once the window rolls over. Per-account thresholds take precedence over this global default.', + openaiQuotaAutoPauseDefault5h: 'Default 5h usage threshold (%)', + openaiQuotaAutoPauseDefault7d: 'Default 7d usage threshold (%)', + openaiQuotaAutoPauseThresholdHint: 'Value 0-100; leave blank or 0 to disable the global default threshold.', errorFiltering: 'Error Filtering', ignoreCountTokensErrors: 'Ignore count_tokens errors', ignoreCountTokensErrorsHint: 'When enabled, errors from count_tokens requests will not be written to the error log.', @@ -5223,7 +5228,8 @@ export default { slaMinPercentRange: 'SLA minimum percentage must be between 0 and 100', ttftP99MaxRange: 'TTFT P99 maximum must be a number ≥ 0', requestErrorRateMaxRange: 'Request error rate maximum must be between 0 and 100', - upstreamErrorRateMaxRange: 'Upstream error rate maximum must be between 0 and 100' + upstreamErrorRateMaxRange: 'Upstream error rate maximum must be between 0 and 100', + openaiQuotaAutoPauseRange: 'OpenAI quota auto-pause threshold must be between 0 and 100' } }, concurrency: { diff --git a/frontend/src/i18n/locales/zh.ts b/frontend/src/i18n/locales/zh.ts index 2364f9c4..4f1d1f13 100644 --- a/frontend/src/i18n/locales/zh.ts +++ b/frontend/src/i18n/locales/zh.ts @@ -5352,6 +5352,11 @@ export default { aggregation: '预聚合任务', enableAggregation: '启用预聚合任务', aggregationHint: '预聚合可提升长时间窗口查询性能', + openaiQuotaAutoPause: 'OpenAI 账号配额自动暂停', + openaiQuotaAutoPauseHint: '当 OpenAI 账号 5h / 7d 用量达到阈值时，调度会自动跳过该账号；窗口滚动后自动恢复。账号级阈值优先于此全局默认值。', + openaiQuotaAutoPauseDefault5h: '默认 5h 用量阈值 (%)', + openaiQuotaAutoPauseDefault7d: '默认 7d 用量阈值 (%)', + openaiQuotaAutoPauseThresholdHint: '取值 0-100，留空或 0 表示不启用全局默认阈值。', errorFiltering: '错误过滤', ignoreCountTokensErrors: '忽略 count_tokens 错误', ignoreCountTokensErrorsHint: '启用后，count_tokens 请求的错误将不会写入错误日志。', @@ -5383,7 +5388,8 @@ export default { slaMinPercentRange: 'SLA最低百分比必须在0-100之间', ttftP99MaxRange: 'TTFT P99最大值必须大于等于0', requestErrorRateMaxRange: '请求错误率最大值必须在0-100之间', - upstreamErrorRateMaxRange: '上游错误率最大值必须在0-100之间' + upstreamErrorRateMaxRange: '上游错误率最大值必须在0-100之间', + openaiQuotaAutoPauseRange: 'OpenAI 配额自动暂停阈值必须在 0-100 之间' } }, concurrency: { diff --git a/frontend/src/views/admin/ops/components/OpsSettingsDialog.vue b/frontend/src/views/admin/ops/components/OpsSettingsDialog.vue index 5dba5b1d..bfb7a65f 100644 --- a/frontend/src/views/admin/ops/components/OpsSettingsDialog.vue +++ b/frontend/src/views/admin/ops/components/OpsSettingsDialog.vue @@ -50,6 +50,10 @@ async function loadAllSettings() { runtimeSettings.value = runtime emailConfig.value = email advancedSettings.value = advanced + // 兼容旧 payload：后端未返回该字段时补默认值，保证表单可绑定 + if (advancedSettings.value && !advancedSettings.value.openai_account_quota_auto_pause) { + advancedSettings.value.openai_account_quota_auto_pause = { default_threshold_5h: 0, default_threshold_7d: 0 } + } // 如果后端返回了阈值，使用后端的值；否则保持默认值 if (thresholds && Object.keys(thresholds).length > 0) { metricThresholds.value = { @@ -119,6 +123,28 @@ function removeRecipient(target: 'alert' | 'report', email: string) { if (idx >= 0) list.splice(idx, 1) } +// OpenAI 账号配额自动暂停：后端按 0~1 分数存储，UI 按百分比(0~100)展示 +const quotaAutoPause5hPercent = computed({ + get() { + const v = advancedSettings.value?.openai_account_quota_auto_pause?.default_threshold_5h + return v && v > 0 ? Math.round(v * 1000) / 10 : null + }, + set(val) { + if (!advancedSettings.value?.openai_account_quota_auto_pause) return + advancedSettings.value.openai_account_quota_auto_pause.default_threshold_5h = val != null && val > 0 ? val / 100 : 0 + } +}) +const quotaAutoPause7dPercent = computed({ + get() { + const v = advancedSettings.value?.openai_account_quota_auto_pause?.default_threshold_7d + return v && v > 0 ? Math.round(v * 1000) / 10 : null + }, + set(val) { + if (!advancedSettings.value?.openai_account_quota_auto_pause) return + advancedSettings.value.openai_account_quota_auto_pause.default_threshold_7d = val != null && val > 0 ? val / 100 : 0 + } +}) + // 验证 const validation = computed(() => { const errors: string[] = [] @@ -145,6 +171,11 @@ const validation = computed(() => { if (hourly_metrics_retention_days < 0 || hourly_metrics_retention_days > 365) { errors.push(t('admin.ops.settings.validation.retentionDaysRange')) } + + const { default_threshold_5h, default_threshold_7d } = advancedSettings.value.openai_account_quota_auto_pause + if (default_threshold_5h < 0 || default_threshold_5h > 1 || default_threshold_7d < 0 || default_threshold_7d > 1) { + errors.push(t('admin.ops.settings.validation.openaiQuotaAutoPauseRange')) + } } // 验证指标阈值 @@ -473,6 +504,40 @@ async function saveAllSettings() {
+ +
+
{{ t('admin.ops.settings.openaiQuotaAutoPause') }}
+

{{ t('admin.ops.settings.openaiQuotaAutoPauseHint') }}

+ +
+
+ + +
+
+ + +
+
+

{{ t('admin.ops.settings.openaiQuotaAutoPauseThresholdHint') }}

+
+
{{ t('admin.ops.settings.errorFiltering') }}
From c9caadb3782a4e658ce46789df092d0b15d58172 Mon Sep 17 00:00:00 2001 From: wucm667 Date: Fri, 29 May 2026 14:32:45 +0800 Subject: [PATCH 17/20] fix(account): address second-round review on quota auto-pause - TopK initial filter now drops quota-paused accounts: fold the quota check into isAccountRequestCompatible so session-hash, TopK pool, and per-candidate rechecks all skip paused accounts. Previously the candidate pool was built without the quota check, so paused accounts could fill TopK and leave the scheduler returning "no available accounts" even with healthy ones available. - Add per-account explicit disable flags auto_pause_5h_disabled / auto_pause_7d_disabled with toggles in EditAccountModal. Without these, leaving the account threshold blank silently falls back to the global default, so admins could not exempt a single account once a global default existed. Disable is per-window: an account can opt out of 5h auto-pause while still honoring 7d. Schedule snapshot whitelist includes the new fields, i18n EN/ZH updated, threshold-hint text revised to explain "blank = global default". - Move quota auto-pause settings off the request hot path: replace the per-repo TTL+singleflight sync DB read with a per-SettingService stale-while-revalidate in-memory snapshot. Get is non-blocking (atomic.Pointer load + async refresh on staleness); writes via UpdateOpsAdvancedSettings push directly into the cache through an injected sink; wire warms the cache at startup. Adds Warm (sync) for tests/init and SetOpenAIQuotaAutoPauseSettings (sink target). Co-Authored-By: Claude Opus 4.7 --- backend/cmd/server/wire_gen.go | 2 +- .../internal/repository/scheduler_cache.go | 2 + .../repository/scheduler_cache_unit_test.go | 4 + .../service/openai_account_scheduler.go | 7 + .../service/openai_account_scheduler_test.go | 136 ++++++++++++++ .../service/openai_gateway_service.go | 44 ++++- backend/internal/service/ops_service.go | 15 ++ backend/internal/service/ops_settings.go | 12 +- .../service/ops_settings_advanced_test.go | 50 ++++- backend/internal/service/setting_service.go | 176 +++++++++++------- backend/internal/service/wire.go | 42 ++++- .../components/account/EditAccountModal.vue | 60 ++++++ .../__tests__/EditAccountModal.spec.ts | 21 +++ frontend/src/i18n/locales/en.ts | 5 +- frontend/src/i18n/locales/zh.ts | 5 +- 15 files changed, 505 insertions(+), 76 deletions(-) diff --git a/backend/cmd/server/wire_gen.go b/backend/cmd/server/wire_gen.go index 465f5e25..6e8be8fc 100644 --- a/backend/cmd/server/wire_gen.go +++ b/backend/cmd/server/wire_gen.go @@ -195,7 +195,7 @@ func initializeApplication(buildInfo handler.BuildInfo) (*Application, error) { gatewayService := service.NewGatewayService(accountRepository, groupRepository, usageLogRepository, usageBillingRepository, userRepository, userSubscriptionRepository, userGroupRateRepository, gatewayCache, configConfig, schedulerSnapshotService, concurrencyService, billingService, rateLimitService, billingCacheService, identityService, httpUpstream, deferredService, claudeTokenProvider, sessionLimitCache, rpmCache, digestSessionStore, settingService, tlsFingerprintProfileService, channelService, modelPricingResolver, balanceNotifyService, serviceUserPlatformQuotaRepository) geminiMessagesCompatService := service.NewGeminiMessagesCompatService(accountRepository, groupRepository, gatewayCache, schedulerSnapshotService, geminiTokenProvider, rateLimitService, httpUpstream, antigravityGatewayService, configConfig) opsSystemLogSink := service.ProvideOpsSystemLogSink(opsRepository) - opsService := service.NewOpsService(opsRepository, settingRepository, configConfig, accountRepository, userRepository, concurrencyService, gatewayService, openAIGatewayService, geminiMessagesCompatService, antigravityGatewayService, opsSystemLogSink) + opsService := service.ProvideOpsService(opsRepository, settingRepository, configConfig, accountRepository, userRepository, concurrencyService, gatewayService, openAIGatewayService, geminiMessagesCompatService, antigravityGatewayService, opsSystemLogSink, settingService) encryptionKey, err := payment.ProvideEncryptionKey(configConfig) if err != nil { return nil, err diff --git a/backend/internal/repository/scheduler_cache.go b/backend/internal/repository/scheduler_cache.go index ff3c4301..cf19deda 100644 --- a/backend/internal/repository/scheduler_cache.go +++ b/backend/internal/repository/scheduler_cache.go @@ -557,6 +557,8 @@ func filterSchedulerExtra(extra map[string]any) map[string]any { "codex_usage_updated_at", "auto_pause_5h_threshold", "auto_pause_7d_threshold", + "auto_pause_5h_disabled", + "auto_pause_7d_disabled", } filtered := make(map[string]any) for _, key := range keys { diff --git a/backend/internal/repository/scheduler_cache_unit_test.go b/backend/internal/repository/scheduler_cache_unit_test.go index 9e4ec23e..a4667591 100644 --- a/backend/internal/repository/scheduler_cache_unit_test.go +++ b/backend/internal/repository/scheduler_cache_unit_test.go @@ -89,6 +89,8 @@ func TestBuildSchedulerMetadataAccount_KeepsQuotaAutoPauseFields(t *testing.T) { "codex_usage_updated_at": "2026-05-29T09:00:00Z", "auto_pause_5h_threshold": 0.95, "auto_pause_7d_threshold": 0.96, + "auto_pause_5h_disabled": true, + "auto_pause_7d_disabled": false, }, } @@ -103,4 +105,6 @@ func TestBuildSchedulerMetadataAccount_KeepsQuotaAutoPauseFields(t *testing.T) { require.Equal(t, "2026-05-29T09:00:00Z", got.Extra["codex_usage_updated_at"]) require.Equal(t, 0.95, got.Extra["auto_pause_5h_threshold"]) require.Equal(t, 0.96, got.Extra["auto_pause_7d_threshold"]) + require.Equal(t, true, got.Extra["auto_pause_5h_disabled"]) + require.Equal(t, false, got.Extra["auto_pause_7d_disabled"]) } diff --git a/backend/internal/service/openai_account_scheduler.go b/backend/internal/service/openai_account_scheduler.go index fd28fa86..47a8142a 100644 --- a/backend/internal/service/openai_account_scheduler.go +++ b/backend/internal/service/openai_account_scheduler.go @@ -974,6 +974,13 @@ func (s *defaultOpenAIAccountScheduler) isAccountRequestCompatible(ctx context.C if s != nil && s.service != nil && s.service.isOpenAIAccountRuntimeBlocked(account) { return false } + // Quota auto-pause must be evaluated during the initial filter too. Without it the + // TopK candidate pool can be filled with paused accounts and the later fresh/DB + // rechecks won't reach healthy accounts that fell outside TopK — manifesting as + // "no available accounts" even though healthy ones exist. + if paused, _ := shouldAutoPauseOpenAIAccountByQuota(ctx, account); paused { + return false + } if req.RequestedModel != "" && !account.IsModelSupported(req.RequestedModel) { return false } diff --git a/backend/internal/service/openai_account_scheduler_test.go b/backend/internal/service/openai_account_scheduler_test.go index 531769a7..da5f0a66 100644 --- a/backend/internal/service/openai_account_scheduler_test.go +++ b/backend/internal/service/openai_account_scheduler_test.go @@ -798,6 +798,63 @@ func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_UsesGlobalDefa require.Equal(t, int64(35402), account.ID) } +// Regression: a per-account explicit-disable flag exempts the account from auto-pause +// even when a global default threshold is set. Without this, "leave threshold blank" +// silently falls back to global default and admins have no way to whitelist a single +// account. +func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_PerAccountDisableOverridesGlobalDefault(t *testing.T) { + ctx := withOpenAIQuotaAutoPauseSettings(context.Background(), OpsOpenAIAccountQuotaAutoPauseSettings{DefaultThreshold5h: 0.95}) + // Account has high usage AND no per-account threshold (would normally fall back to + // the global default and get paused), but the explicit disable flag is set. + primary := Account{ + ID: 35701, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Extra: map[string]any{ + "codex_5h_used_percent": 99.0, + "auto_pause_5h_disabled": true, + }, + } + secondary := Account{ID: 35702, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} + svc := &OpenAIGatewayService{accountRepo: schedulerTestOpenAIAccountRepo{accounts: []Account{primary, secondary}}, cfg: &config.Config{}} + + account, err := svc.SelectAccountForModelWithExclusions(ctx, nil, "", "gpt-5.1", nil) + require.NoError(t, err) + require.NotNil(t, account) + require.Equal(t, int64(35701), account.ID) +} + +// Disable is per-window: disabling only 5h must still allow 7d auto-pause to fire. +func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_PerWindowDisableScoped(t *testing.T) { + ctx := context.Background() + primary := Account{ + ID: 35801, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Extra: map[string]any{ + "codex_5h_used_percent": 99.0, + "codex_7d_used_percent": 99.0, + "auto_pause_5h_disabled": true, + "auto_pause_7d_threshold": 0.95, + }, + } + secondary := Account{ID: 35802, Platform: PlatformOpenAI, Type: AccountTypeAPIKey, Status: StatusActive, Schedulable: true, Concurrency: 1, Priority: 5} + svc := &OpenAIGatewayService{accountRepo: schedulerTestOpenAIAccountRepo{accounts: []Account{primary, secondary}}, cfg: &config.Config{}} + + account, err := svc.SelectAccountForModelWithExclusions(ctx, nil, "", "gpt-5.1", nil) + require.NoError(t, err) + require.NotNil(t, account) + require.Equal(t, int64(35802), account.ID, "7d auto-pause must still fire even though 5h is disabled") +} + func TestOpenAIGatewayService_SelectAccountForModelWithExclusions_StaleUsageWindowResetSkipsPause(t *testing.T) { ctx := context.Background() // Usage is over threshold but the window's reset time has already passed, so the @@ -1399,6 +1456,85 @@ func TestOpenAIGatewayService_SelectAccountWithScheduler_LoadBalanceTopKFallback } } +// Regression: TopK initial filter must drop quota-auto-paused accounts. Otherwise +// the candidate pool is filled with paused accounts, healthy accounts fall outside +// TopK, and the scheduler returns "no available accounts" even though healthy ones +// exist. +func TestOpenAIGatewayService_SelectAccountWithScheduler_LoadBalanceTopKExcludesQuotaPaused(t *testing.T) { + ctx := context.Background() + groupID := int64(110) + accounts := []Account{ + { + ID: 37001, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 0, + Extra: map[string]any{ + "codex_5h_used_percent": 96.0, + "auto_pause_5h_threshold": 0.95, + }, + }, + { + ID: 37002, + Platform: PlatformOpenAI, + Type: AccountTypeAPIKey, + Status: StatusActive, + Schedulable: true, + Concurrency: 1, + Priority: 5, + }, + } + + cfg := &config.Config{} + cfg.Gateway.OpenAIWS.LBTopK = 1 // TopK=1 makes the bug fatal: paused account would crowd out the healthy one entirely + cfg.Gateway.OpenAIWS.SchedulerScoreWeights.Priority = 0.4 + cfg.Gateway.OpenAIWS.SchedulerScoreWeights.Load = 1.0 + cfg.Gateway.OpenAIWS.SchedulerScoreWeights.Queue = 1.0 + + concurrencyCache := schedulerTestConcurrencyCache{ + loadMap: map[int64]*AccountLoadInfo{ + 37001: {AccountID: 37001, LoadRate: 5, WaitingCount: 0}, + 37002: {AccountID: 37002, LoadRate: 5, WaitingCount: 0}, + }, + acquireResults: map[int64]bool{ + 37002: true, + }, + } + + svc := &OpenAIGatewayService{ + accountRepo: schedulerTestOpenAIAccountRepo{accounts: accounts}, + cache: &schedulerTestGatewayCache{}, + cfg: cfg, + rateLimitService: newOpenAIAdvancedSchedulerRateLimitService("true"), + concurrencyService: NewConcurrencyService(concurrencyCache), + } + + selection, decision, err := svc.SelectAccountWithScheduler( + ctx, + &groupID, + "", + "", + "gpt-5.1", + nil, + OpenAIUpstreamTransportAny, + false, + ) + require.NoError(t, err) + require.NotNil(t, selection) + require.NotNil(t, selection.Account) + require.Equal(t, int64(37002), selection.Account.ID) + require.Equal(t, openAIAccountScheduleLayerLoadBalance, decision.Layer) + // Only the healthy account should ever enter the candidate pool; the paused one + // must be filtered out at the initial-filter stage. + require.Equal(t, 1, decision.CandidateCount) + if selection.ReleaseFunc != nil { + selection.ReleaseFunc() + } +} + func TestOpenAIGatewayService_OpenAIAccountSchedulerMetrics(t *testing.T) { ctx := context.Background() groupID := int64(12) diff --git a/backend/internal/service/openai_gateway_service.go b/backend/internal/service/openai_gateway_service.go index e2534cc2..b1ae6a9a 100644 --- a/backend/internal/service/openai_gateway_service.go +++ b/backend/internal/service/openai_gateway_service.go @@ -1360,14 +1360,21 @@ func shouldAutoPauseOpenAIAccountByQuota(ctx context.Context, account *Account) if account == nil || !account.IsOpenAI() { return false, openAIQuotaAutoPauseDecision{} } + // Per-account explicit-disable flags must take precedence over the global default. + // Without these, leaving the account threshold blank means "use global default", + // so an admin has no way to exempt a single account from auto-pause once a global + // default exists. The disable flag is per-window so an account can opt out of + // only 5h or only 7d auto-pause. + disabled5h := resolveAccountExtraBool(account.Extra, "auto_pause_5h_disabled") + disabled7d := resolveAccountExtraBool(account.Extra, "auto_pause_7d_disabled") threshold5h, threshold7d := resolveOpenAIQuotaAutoPauseThresholds(ctx, account) now := time.Now() - if threshold5h > 0 { + if !disabled5h && threshold5h > 0 { if utilization, ok := resolveOpenAIQuotaUtilization(account.Extra, "5h", now); ok && utilization >= threshold5h { return true, openAIQuotaAutoPauseDecision{window: "5h", threshold: threshold5h, utilization: utilization} } } - if threshold7d > 0 { + if !disabled7d && threshold7d > 0 { if utilization, ok := resolveOpenAIQuotaUtilization(account.Extra, "7d", now); ok && utilization >= threshold7d { return true, openAIQuotaAutoPauseDecision{window: "7d", threshold: threshold7d, utilization: utilization} } @@ -1375,6 +1382,39 @@ func shouldAutoPauseOpenAIAccountByQuota(ctx context.Context, account *Account) return false, openAIQuotaAutoPauseDecision{} } +// resolveAccountExtraBool reads a bool-like value from account extra, tolerating +// the few shapes JSON unmarshalling may produce (real bool, "true"/"false" +// strings, 0/1 numbers). +func resolveAccountExtraBool(extra map[string]any, key string) bool { + if len(extra) == 0 { + return false + } + value, ok := extra[key] + if !ok || value == nil { + return false + } + switch v := value.(type) { + case bool: + return v + case string: + parsed, err := strconv.ParseBool(strings.TrimSpace(v)) + return err == nil && parsed + case float64: + return v != 0 + case float32: + return v != 0 + case int: + return v != 0 + case int64: + return v != 0 + case json.Number: + if i, err := v.Int64(); err == nil { + return i != 0 + } + } + return false +} + func resolveOpenAIQuotaAutoPauseThresholds(ctx context.Context, account *Account) (float64, float64) { threshold5h, _ := resolveAccountExtraNumber(account.Extra, "auto_pause_5h_threshold") threshold7d, _ := resolveAccountExtraNumber(account.Extra, "auto_pause_7d_threshold") diff --git a/backend/internal/service/ops_service.go b/backend/internal/service/ops_service.go index 1cea72fa..2d7c5bd4 100644 --- a/backend/internal/service/ops_service.go +++ b/backend/internal/service/ops_service.go @@ -41,6 +41,11 @@ type OpsService struct { // cleanupReloader 由 wire 在 OpsCleanupService 构造完成后通过 SetCleanupReloader 注入。 // 解耦避免 OpsService -> OpsCleanupService 的硬依赖（cleanup 也读 settings，会循环）。 cleanupReloader CleanupReloader + + // quotaAutoPauseSink 由 wire 注入（通常是 SettingService.SetOpenAIQuotaAutoPauseSettings）。 + // UpdateOpsAdvancedSettings 写入新配置后调用，把最新的 quota auto-pause 全局默认阈值 + // 立即同步到调度热路径读取的内存缓存，避免下次请求才能感知新值。 + quotaAutoPauseSink func(OpsOpenAIAccountQuotaAutoPauseSettings) } // CleanupReloader 由 OpsCleanupService 实现。 @@ -57,6 +62,16 @@ func (s *OpsService) SetCleanupReloader(r CleanupReloader) { s.cleanupReloader = r } +// SetOpenAIQuotaAutoPauseSettingsSink 由 wire 注入，把最新的 quota auto-pause 全局默认 +// 阈值 push 到调度热路径读取的内存缓存。同 SetCleanupReloader 的解耦目的：避免 OpsService +// 持有 *SettingService 引入循环依赖。 +func (s *OpsService) SetOpenAIQuotaAutoPauseSettingsSink(sink func(OpsOpenAIAccountQuotaAutoPauseSettings)) { + if s == nil { + return + } + s.quotaAutoPauseSink = sink +} + func NewOpsService( opsRepo OpsRepository, settingRepo SettingRepository, diff --git a/backend/internal/service/ops_settings.go b/backend/internal/service/ops_settings.go index 23e92e5a..472f4e32 100644 --- a/backend/internal/service/ops_settings.go +++ b/backend/internal/service/ops_settings.go @@ -490,12 +490,12 @@ func (s *OpsService) UpdateOpsAdvancedSettings(ctx context.Context, cfg *OpsAdva if err := s.settingRepo.Set(ctx, SettingKeyOpsAdvancedSettings, string(raw)); err != nil { return nil, err } - cacheKey := openAIQuotaAutoPauseSettingsCacheKey(s.settingRepo) - openAIQuotaAutoPauseSettingsSF.Forget(cacheKey) - storeOpenAIQuotaAutoPauseSettingsCache(s.settingRepo, &cachedOpenAIQuotaAutoPauseSettings{ - settings: cfg.OpenAIAccountQuotaAutoPause, - expiresAt: time.Now().Add(openAIQuotaAutoPauseSettingsCacheTTL).UnixNano(), - }) + // Push the new quota auto-pause settings straight into the in-memory cache that + // the OpenAI scheduling hot path reads, so the next request observes the new value + // without waiting for the background refresher's TTL. + if s.quotaAutoPauseSink != nil { + s.quotaAutoPauseSink(cfg.OpenAIAccountQuotaAutoPause) + } // notify cleanup service to reload schedule/enabled. if s.cleanupReloader != nil { diff --git a/backend/internal/service/ops_settings_advanced_test.go b/backend/internal/service/ops_settings_advanced_test.go index d8598fe0..62803f94 100644 --- a/backend/internal/service/ops_settings_advanced_test.go +++ b/backend/internal/service/ops_settings_advanced_test.go @@ -4,6 +4,7 @@ import ( "context" "encoding/json" "testing" + "time" "github.com/Wei-Shaw/sub2api/internal/config" ) @@ -103,11 +104,58 @@ func TestGetOpenAIQuotaAutoPauseSettings_ReadsDefaultsFromOpsAdvancedSettings(t repo.values[SettingKeyOpsAdvancedSettings] = `{"openai_account_quota_auto_pause":{"default_threshold_5h":0.95,"default_threshold_7d":0.9}}` svc := NewSettingService(repo, &config.Config{}) - settings := svc.GetOpenAIQuotaAutoPauseSettings(context.Background()) + // Warm the in-memory cache synchronously so the assertion below is deterministic. + // GetOpenAIQuotaAutoPauseSettings is non-blocking on the hot path (returns the + // cached value, refreshes asynchronously); for tests and startup, Warm is the + // synchronous entry point that guarantees a populated cache. + settings := svc.WarmOpenAIQuotaAutoPauseSettings(context.Background()) if settings.DefaultThreshold5h != 0.95 { t.Fatalf("DefaultThreshold5h = %v, want 0.95", settings.DefaultThreshold5h) } if settings.DefaultThreshold7d != 0.9 { t.Fatalf("DefaultThreshold7d = %v, want 0.9", settings.DefaultThreshold7d) } + + // Subsequent Get must hit the warm cache and return the same value without any DB + // access — that's the hot-path invariant. + cached := svc.GetOpenAIQuotaAutoPauseSettings(context.Background()) + if cached.DefaultThreshold5h != 0.95 || cached.DefaultThreshold7d != 0.9 { + t.Fatalf("cached read = %+v, want {0.95, 0.9}", cached) + } +} + +// Hot-path invariant: a Get with cold cache must return immediately (zero defaults) +// rather than blocking on the DB. The async refresher will populate the cache for +// subsequent calls. +func TestGetOpenAIQuotaAutoPauseSettings_ColdCacheNonBlocking(t *testing.T) { + repo := newRuntimeSettingRepoStub() + repo.values[SettingKeyOpsAdvancedSettings] = `{"openai_account_quota_auto_pause":{"default_threshold_5h":0.7}}` + svc := NewSettingService(repo, &config.Config{}) + + start := time.Now() + settings := svc.GetOpenAIQuotaAutoPauseSettings(context.Background()) + elapsed := time.Since(start) + if elapsed > 50*time.Millisecond { + t.Fatalf("cold-cache Get must be non-blocking, took %v", elapsed) + } + // Cold cache means we get zero defaults (the async refresh hasn't completed yet). + if settings.DefaultThreshold5h != 0 || settings.DefaultThreshold7d != 0 { + t.Fatalf("cold-cache Get = %+v, want zeroes", settings) + } +} + +// Explicit cache write (e.g. from UpdateOpsAdvancedSettings) must be visible on the +// very next read without any DB roundtrip. +func TestSetOpenAIQuotaAutoPauseSettings_VisibleImmediately(t *testing.T) { + svc := NewSettingService(newRuntimeSettingRepoStub(), &config.Config{}) + + svc.SetOpenAIQuotaAutoPauseSettings(OpsOpenAIAccountQuotaAutoPauseSettings{ + DefaultThreshold5h: 0.88, + DefaultThreshold7d: 0.77, + }) + + got := svc.GetOpenAIQuotaAutoPauseSettings(context.Background()) + if got.DefaultThreshold5h != 0.88 || got.DefaultThreshold7d != 0.77 { + t.Fatalf("after Set, Get = %+v, want {0.88, 0.77}", got) + } } diff --git a/backend/internal/service/setting_service.go b/backend/internal/service/setting_service.go index 40e98b88..98acdb80 100644 --- a/backend/internal/service/setting_service.go +++ b/backend/internal/service/setting_service.go @@ -14,7 +14,6 @@ import ( "sort" "strconv" "strings" - "sync" "sync/atomic" "time" @@ -162,28 +161,7 @@ const openAIQuotaAutoPauseSettingsCacheTTL = 60 * time.Second const openAIQuotaAutoPauseSettingsErrorTTL = 5 * time.Second const openAIQuotaAutoPauseSettingsDBTimeout = 5 * time.Second -var openAIQuotaAutoPauseSettingsCache sync.Map // map[string]*cachedOpenAIQuotaAutoPauseSettings -var openAIQuotaAutoPauseSettingsSF singleflight.Group - -func openAIQuotaAutoPauseSettingsCacheKey(repo SettingRepository) string { - if repo == nil { - return "nil" - } - return fmt.Sprintf("%T:%p", repo, repo) -} - -func loadOpenAIQuotaAutoPauseSettingsCache(repo SettingRepository) (*cachedOpenAIQuotaAutoPauseSettings, bool) { - value, ok := openAIQuotaAutoPauseSettingsCache.Load(openAIQuotaAutoPauseSettingsCacheKey(repo)) - if !ok || value == nil { - return nil, false - } - cached, ok := value.(*cachedOpenAIQuotaAutoPauseSettings) - return cached, ok && cached != nil -} - -func storeOpenAIQuotaAutoPauseSettingsCache(repo SettingRepository, cached *cachedOpenAIQuotaAutoPauseSettings) { - openAIQuotaAutoPauseSettingsCache.Store(openAIQuotaAutoPauseSettingsCacheKey(repo), cached) -} +const openAIQuotaAutoPauseSettingsRefreshKey = "openai_quota_auto_pause_settings" // DefaultSubscriptionGroupReader validates group references used by default subscriptions. type DefaultSubscriptionGroupReader interface { @@ -209,6 +187,15 @@ type SettingService struct { openAICodexUASF singleflight.Group openAIAllowCodexPluginCache atomic.Value // *cachedOpenAIAllowCodexPlugin openAIAllowCodexPluginSF singleflight.Group + + // openAIQuotaAutoPauseSettingsCache holds the most recently observed quota auto-pause + // settings. GetOpenAIQuotaAutoPauseSettings reads this atomic.Value on the request hot + // path without ever blocking on the DB; when the cached entry expires, a background + // goroutine refreshes it via openAIQuotaAutoPauseSettingsSF (stale-while-revalidate). + // This per-service field also gives tests natural isolation — each SettingService + // instance owns its own cache, no shared package-level state. + openAIQuotaAutoPauseSettingsCache atomic.Value // *cachedOpenAIQuotaAutoPauseSettings + openAIQuotaAutoPauseSettingsSF singleflight.Group } // DefaultPlatformQuotaSetting 单 platform 三档限额（nil = 沿用上层；0 = 显式禁用；>0 = 上限） @@ -2060,9 +2047,17 @@ func (s *SettingService) refreshCachedSettings(settings *SystemSettings) { enabled: settings.OpenAIAdvancedSchedulerEnabled, expiresAt: time.Now().Add(openAIAdvancedSchedulerSettingCacheTTL).UnixNano(), }) - cacheKey := openAIQuotaAutoPauseSettingsCacheKey(s.settingRepo) - openAIQuotaAutoPauseSettingsSF.Forget(cacheKey) - openAIQuotaAutoPauseSettingsCache.Delete(cacheKey) + // Invalidate the quota auto-pause cache and let the next read trigger a fresh load. + // We can't know from here whether ops_advanced_settings was also touched, so be + // defensive: store an expired entry — GetOpenAIQuotaAutoPauseSettings will serve + // stale and kick off an async refresh, never blocking the request that follows. + s.openAIQuotaAutoPauseSettingsSF.Forget(openAIQuotaAutoPauseSettingsRefreshKey) + if cached, _ := s.openAIQuotaAutoPauseSettingsCache.Load().(*cachedOpenAIQuotaAutoPauseSettings); cached != nil { + s.openAIQuotaAutoPauseSettingsCache.Store(&cachedOpenAIQuotaAutoPauseSettings{ + settings: cached.settings, + expiresAt: 0, + }) + } if s.cfg != nil { s.cfg.SetTrustForwardedIPForAPIKeyACL(settings.APIKeyACLTrustForwardedIP) } @@ -4484,49 +4479,104 @@ func (s *SettingService) GetClaudeCodeVersionBounds(ctx context.Context) (min, m return b.min, b.max } +// GetOpenAIQuotaAutoPauseSettings returns the current global default quota auto-pause +// settings. It is invoked on the OpenAI scheduling hot path (once per request) and is +// therefore designed to never block on the DB: +// +// - Fresh cached value → returned immediately. +// - Stale or empty cache → the last known value is returned, and a background +// goroutine refreshes the cache via singleflight (stale-while-revalidate). +// - First call with no cache yet → zero defaults are returned and the same async +// refresh is kicked off; the next call gets the freshly populated value. +// +// Callers that need the freshly persisted value synchronously (tests, post-update +// confirmation, optional startup warm-up) should call WarmOpenAIQuotaAutoPauseSettings. func (s *SettingService) GetOpenAIQuotaAutoPauseSettings(ctx context.Context) OpsOpenAIAccountQuotaAutoPauseSettings { - if cached, ok := loadOpenAIQuotaAutoPauseSettingsCache(s.settingRepo); ok { - if time.Now().UnixNano() < cached.expiresAt { - return cached.settings + if s == nil { + return OpsOpenAIAccountQuotaAutoPauseSettings{} + } + cached, _ := s.openAIQuotaAutoPauseSettingsCache.Load().(*cachedOpenAIQuotaAutoPauseSettings) + now := time.Now().UnixNano() + if cached != nil && now < cached.expiresAt { + return cached.settings + } + // Stale or unset: trigger background refresh without blocking this request. + // singleflight.DoChan dedupes concurrent refreshes; we deliberately ignore the + // returned channel — the result is observable via the atomic cache. + s.openAIQuotaAutoPauseSettingsSF.DoChan(openAIQuotaAutoPauseSettingsRefreshKey, func() (any, error) { + s.refreshOpenAIQuotaAutoPauseSettings(context.Background()) + return nil, nil + }) + if cached != nil { + return cached.settings // serve stale value while revalidating + } + return OpsOpenAIAccountQuotaAutoPauseSettings{} +} + +// WarmOpenAIQuotaAutoPauseSettings synchronously loads the quota auto-pause settings +// into the in-memory cache. Useful for application startup (so the first request hits +// a warm cache) and for tests that need deterministic reads immediately after +// constructing the service. +func (s *SettingService) WarmOpenAIQuotaAutoPauseSettings(ctx context.Context) OpsOpenAIAccountQuotaAutoPauseSettings { + if s == nil { + return OpsOpenAIAccountQuotaAutoPauseSettings{} + } + s.refreshOpenAIQuotaAutoPauseSettings(ctx) + cached, _ := s.openAIQuotaAutoPauseSettingsCache.Load().(*cachedOpenAIQuotaAutoPauseSettings) + if cached == nil { + return OpsOpenAIAccountQuotaAutoPauseSettings{} + } + return cached.settings +} + +// refreshOpenAIQuotaAutoPauseSettings reads the latest settings from the DB and stores +// them into the in-memory cache. On error it stores the prior value (or zero defaults +// if nothing is cached yet) with the shorter error TTL so the next refresh comes +// sooner. Always uses its own timeout-bounded context to keep refresh latency +// predictable regardless of the caller. +func (s *SettingService) refreshOpenAIQuotaAutoPauseSettings(ctx context.Context) { + if s == nil || s.settingRepo == nil { + return + } + dbCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), openAIQuotaAutoPauseSettingsDBTimeout) + defer cancel() + + settings := OpsOpenAIAccountQuotaAutoPauseSettings{} + ttl := openAIQuotaAutoPauseSettingsCacheTTL + raw, err := s.settingRepo.GetValue(dbCtx, SettingKeyOpsAdvancedSettings) + if err == nil { + cfg := defaultOpsAdvancedSettings() + if strings.TrimSpace(raw) != "" { + if jsonErr := json.Unmarshal([]byte(raw), cfg); jsonErr == nil { + normalizeOpsAdvancedSettings(cfg) + } } + settings = cfg.OpenAIAccountQuotaAutoPause + } else if !errors.Is(err, ErrSettingNotFound) { + // Real error: keep serving prior value but refresh sooner. + if prior, _ := s.openAIQuotaAutoPauseSettingsCache.Load().(*cachedOpenAIQuotaAutoPauseSettings); prior != nil { + settings = prior.settings + } + ttl = openAIQuotaAutoPauseSettingsErrorTTL } - cacheKey := openAIQuotaAutoPauseSettingsCacheKey(s.settingRepo) - result, _, _ := openAIQuotaAutoPauseSettingsSF.Do(cacheKey, func() (any, error) { - if cached, ok := loadOpenAIQuotaAutoPauseSettingsCache(s.settingRepo); ok { - if time.Now().UnixNano() < cached.expiresAt { - return cached.settings, nil - } - } - - settings := OpsOpenAIAccountQuotaAutoPauseSettings{} - ttl := openAIQuotaAutoPauseSettingsCacheTTL - if s != nil && s.settingRepo != nil { - dbCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), openAIQuotaAutoPauseSettingsDBTimeout) - defer cancel() - raw, err := s.settingRepo.GetValue(dbCtx, SettingKeyOpsAdvancedSettings) - if err == nil { - cfg := defaultOpsAdvancedSettings() - if strings.TrimSpace(raw) != "" { - if jsonErr := json.Unmarshal([]byte(raw), cfg); jsonErr == nil { - normalizeOpsAdvancedSettings(cfg) - } - } - settings = cfg.OpenAIAccountQuotaAutoPause - } else { - ttl = openAIQuotaAutoPauseSettingsErrorTTL - } - } - - storeOpenAIQuotaAutoPauseSettingsCache(s.settingRepo, &cachedOpenAIQuotaAutoPauseSettings{ - settings: settings, - expiresAt: time.Now().Add(ttl).UnixNano(), - }) - return settings, nil + s.openAIQuotaAutoPauseSettingsCache.Store(&cachedOpenAIQuotaAutoPauseSettings{ + settings: settings, + expiresAt: time.Now().Add(ttl).UnixNano(), }) +} - settings, _ := result.(OpsOpenAIAccountQuotaAutoPauseSettings) - return settings +// SetOpenAIQuotaAutoPauseSettings writes the given settings directly into the in-memory +// cache. Called from settings-write code paths so that the next read reflects the new +// value immediately, without waiting for the background refresh. +func (s *SettingService) SetOpenAIQuotaAutoPauseSettings(settings OpsOpenAIAccountQuotaAutoPauseSettings) { + if s == nil { + return + } + s.openAIQuotaAutoPauseSettingsCache.Store(&cachedOpenAIQuotaAutoPauseSettings{ + settings: settings, + expiresAt: time.Now().Add(openAIQuotaAutoPauseSettingsCacheTTL).UnixNano(), + }) } // GetRectifierSettings 获取请求整流器配置 diff --git a/backend/internal/service/wire.go b/backend/internal/service/wire.go index b22e10ae..d3e4ce51 100644 --- a/backend/internal/service/wire.go +++ b/backend/internal/service/wire.go @@ -396,6 +396,46 @@ func ProvideBackupService( return svc } +// ProvideOpsService constructs OpsService and wires the SettingService-backed quota +// auto-pause cache sink. Mirrors the SetCleanupReloader pattern: OpsService doesn't +// hold a *SettingService reference, but wire injects a tiny callback so writes to +// ops_advanced_settings immediately propagate into the scheduler hot-path cache. +func ProvideOpsService( + opsRepo OpsRepository, + settingRepo SettingRepository, + cfg *config.Config, + accountRepo AccountRepository, + userRepo UserRepository, + concurrencyService *ConcurrencyService, + gatewayService *GatewayService, + openAIGatewayService *OpenAIGatewayService, + geminiCompatService *GeminiMessagesCompatService, + antigravityGatewayService *AntigravityGatewayService, + systemLogSink *OpsSystemLogSink, + settingService *SettingService, +) *OpsService { + svc := NewOpsService( + opsRepo, + settingRepo, + cfg, + accountRepo, + userRepo, + concurrencyService, + gatewayService, + openAIGatewayService, + geminiCompatService, + antigravityGatewayService, + systemLogSink, + ) + if settingService != nil { + svc.SetOpenAIQuotaAutoPauseSettingsSink(settingService.SetOpenAIQuotaAutoPauseSettings) + // Optional warm-up so the first scheduled request after process start observes + // a populated cache rather than zero defaults. Best-effort, sync-bounded. + settingService.WarmOpenAIQuotaAutoPauseSettings(context.Background()) + } + return svc +} + // ProvideSettingService wires SettingService with group reader and proxy repo. func ProvideSettingService(settingRepo SettingRepository, groupRepo GroupRepository, proxyRepo ProxyRepository, cfg *config.Config) *SettingService { svc := NewSettingService(settingRepo, cfg) @@ -481,7 +521,7 @@ var ProviderSet = wire.NewSet( NewDataManagementService, ProvideBackupService, ProvideOpsSystemLogSink, - NewOpsService, + ProvideOpsService, ProvideOpsMetricsCollector, ProvideOpsAggregationService, ProvideOpsAlertEvaluatorService, diff --git a/frontend/src/components/account/EditAccountModal.vue b/frontend/src/components/account/EditAccountModal.vue index 470c0bfb..8dc85d0e 100644 --- a/frontend/src/components/account/EditAccountModal.vue +++ b/frontend/src/components/account/EditAccountModal.vue @@ -1791,6 +1791,28 @@ v-if="account?.platform === 'openai'" class="border-t border-gray-200 pt-4 dark:border-dark-600 space-y-4" > +
+
+ + +
+

{{ t('admin.accounts.autoPauseDisabledHint') }}

+

{{ t('admin.accounts.autoPauseThresholdHint') }}

+
+
+ + +
+

{{ t('admin.accounts.autoPauseDisabledHint') }}

+

{{ t('admin.accounts.autoPauseThresholdHint') }}

@@ -2481,6 +2527,8 @@ const interceptWarmupRequests = ref(false) const autoPauseOnExpired = ref(false) const autoPause5hThreshold = ref(null) const autoPause7dThreshold = ref(null) +const autoPause5hDisabled = ref(false) +const autoPause7dDisabled = ref(false) const mixedScheduling = ref(false) // For antigravity accounts: enable mixed scheduling const allowOverages = ref(false) // For antigravity accounts: enable AI Credits overages const antigravityModelRestrictionMode = ref<'whitelist' | 'mapping'>('whitelist') @@ -2901,6 +2949,8 @@ const syncFormFromAccount = (newAccount: Account | null) => { allowOverages.value = extra?.allow_overages === true autoPause5hThreshold.value = typeof extra?.auto_pause_5h_threshold === 'number' ? extra.auto_pause_5h_threshold * 100 : null autoPause7dThreshold.value = typeof extra?.auto_pause_7d_threshold === 'number' ? extra.auto_pause_7d_threshold * 100 : null + autoPause5hDisabled.value = extra?.auto_pause_5h_disabled === true + autoPause7dDisabled.value = extra?.auto_pause_7d_disabled === true // Load OpenAI passthrough toggle (OpenAI OAuth/API Key) openaiPassthroughEnabled.value = false @@ -4064,6 +4114,16 @@ const handleSubmit = async () => { } else { delete newExtra.auto_pause_7d_threshold } + if (autoPause5hDisabled.value) { + newExtra.auto_pause_5h_disabled = true + } else { + delete newExtra.auto_pause_5h_disabled + } + if (autoPause7dDisabled.value) { + newExtra.auto_pause_7d_disabled = true + } else { + delete newExtra.auto_pause_7d_disabled + } delete newExtra.codex_image_generation_bridge_enabled if (codexImageGenerationBridgeMode.value === 'inherit') { diff --git a/frontend/src/components/account/__tests__/EditAccountModal.spec.ts b/frontend/src/components/account/__tests__/EditAccountModal.spec.ts index 6db63831..f4865de9 100644 --- a/frontend/src/components/account/__tests__/EditAccountModal.spec.ts +++ b/frontend/src/components/account/__tests__/EditAccountModal.spec.ts @@ -352,6 +352,27 @@ describe('EditAccountModal', () => { expect(updateAccountMock.mock.calls[0]?.[1]?.extra?.auto_pause_7d_threshold).toBe(0.96) }) + it('submits OpenAI quota auto-pause disable flag in extra', async () => { + // Toggling the per-account disable flag must persist as auto_pause_5h_disabled + // so an admin can exempt one account from auto-pause even when a global default + // threshold is configured (otherwise leaving the threshold blank would silently + // fall back to the global default). + const account = buildAccount() + updateAccountMock.mockReset() + checkMixedChannelRiskMock.mockReset() + checkMixedChannelRiskMock.mockResolvedValue({ has_risk: false }) + updateAccountMock.mockResolvedValue(account) + + const wrapper = mountModal(account) + + await wrapper.get('[data-testid="auto-pause-5h-disabled"]').trigger('click') + await wrapper.get('form#edit-account-form').trigger('submit.prevent') + + expect(updateAccountMock).toHaveBeenCalledTimes(1) + expect(updateAccountMock.mock.calls[0]?.[1]?.extra?.auto_pause_5h_disabled).toBe(true) + expect(updateAccountMock.mock.calls[0]?.[1]?.extra?.auto_pause_7d_disabled).toBeUndefined() + }) + it('keeps at least one OpenAI APIKey endpoint capability selected', async () => { const account = buildAccount() updateAccountMock.mockReset() diff --git a/frontend/src/i18n/locales/en.ts b/frontend/src/i18n/locales/en.ts index 8ab90961..6735029c 100644 --- a/frontend/src/i18n/locales/en.ts +++ b/frontend/src/i18n/locales/en.ts @@ -3477,7 +3477,10 @@ export default { autoPauseOnExpiredDesc: 'When enabled, the account will auto pause scheduling after it expires', autoPause5hThreshold: '5h Usage Threshold (%)', autoPause7dThreshold: '7d Usage Threshold (%)', - autoPauseThresholdHint: 'Leave empty or set 0 to disable. Reaching the threshold only skips the account during scheduling and does not modify schedulable.', + autoPauseThresholdHint: 'Leave empty or set 0 to use the global default threshold (configured in Ops settings); set a value to override the global default. Reaching the threshold only skips the account during scheduling and does not modify schedulable.', + autoPause5hDisabled: 'Disable 5h auto-pause', + autoPause7dDisabled: 'Disable 7d auto-pause', + autoPauseDisabledHint: 'When enabled, this account is never auto-paused (even if a global default threshold is configured).', // Quota control (Anthropic OAuth/SetupToken only) quotaControl: { title: 'Quota Control', diff --git a/frontend/src/i18n/locales/zh.ts b/frontend/src/i18n/locales/zh.ts index 4f1d1f13..abb8dff7 100644 --- a/frontend/src/i18n/locales/zh.ts +++ b/frontend/src/i18n/locales/zh.ts @@ -3615,7 +3615,10 @@ export default { autoPauseOnExpiredDesc: '启用后，账号过期将自动暂停调度', autoPause5hThreshold: '5h 用量阈值(%)', autoPause7dThreshold: '7d 用量阈值(%)', - autoPauseThresholdHint: '填 0 或留空表示不启用；达到阈值后仅在调度时跳过账号，不修改 schedulable。', + autoPauseThresholdHint: '留空或填 0 表示使用全局默认阈值（在运维设置中配置）；填具体值则覆盖全局默认。达到阈值后仅在调度时跳过账号，不修改 schedulable。', + autoPause5hDisabled: '禁用 5h 自动暂停', + autoPause7dDisabled: '禁用 7d 自动暂停', + autoPauseDisabledHint: '开启后该账号永不进入自动暂停（即使全局默认阈值已配置）。', // Quota control (Anthropic OAuth/SetupToken only) quotaControl: { title: '配额控制', From 0a521f09fbc481ccdd2a0d6bccfaa8f80df220fc Mon Sep 17 00:00:00 2001 From: Pluviobyte Date: Fri, 29 May 2026 06:46:49 +0000 Subject: [PATCH 18/20] fix(gemini): close tool_use block before text in messages streaming When the Gemini->Anthropic streaming bridge for the /v1/messages endpoint receives a functionCall part followed by a text part, the text branch in handleStreamingResponse opened a new text content block without closing the already-open tool_use block. The tool block's content_block_stop was only emitted at end-of-stream, after the text block's content_block_start, so the Anthropic SSE stream contained overlapping/unterminated content blocks. Clients that assemble messages by block index (e.g. Claude Code) can drop the tool input or mis-parse the response. The functionCall branch already closes an open text block before opening a tool block, and the chat-completions sibling closes the tool block in its text branch via closeOpenTool(). This applies the same symmetric handling to the messages variant: close any open tool_use block (resetting openToolIndex/openToolName/ seenToolJSON) before starting text. Adds a regression test that replays a tool->text Gemini stream and asserts the Anthropic content-block lifecycle never overlaps. --- .../service/gemini_messages_compat_service.go | 16 +++ .../gemini_messages_compat_service_test.go | 105 ++++++++++++++++++ 2 files changed, 121 insertions(+) diff --git a/backend/internal/service/gemini_messages_compat_service.go b/backend/internal/service/gemini_messages_compat_service.go index 516556ca..64f19b2e 100644 --- a/backend/internal/service/gemini_messages_compat_service.go +++ b/backend/internal/service/gemini_messages_compat_service.go @@ -2031,6 +2031,22 @@ func (s *GeminiMessagesCompatService) handleStreamingResponse(c *gin.Context, re parts := extractGeminiParts(geminiResp) for _, part := range parts { if text, ok := part["text"].(string); ok && text != "" { + // Close an open tool_use block before starting text, mirroring + // the functionCall branch (which closes open text blocks) and + // the chat-completions sibling's closeOpenTool(). Otherwise a + // tool→text sequence keeps the tool_use block open while the + // text block starts, emitting overlapping Anthropic content + // blocks that violate the SSE contract. + if openToolIndex >= 0 { + writeSSE(c.Writer, "content_block_stop", map[string]any{ + "type": "content_block_stop", + "index": openToolIndex, + }) + openToolIndex = -1 + openToolName = "" + seenToolJSON = "" + } + delta, newSeen := computeGeminiTextDelta(seenText, text) seenText = newSeen if delta == "" { diff --git a/backend/internal/service/gemini_messages_compat_service_test.go b/backend/internal/service/gemini_messages_compat_service_test.go index d0560344..79db633a 100644 --- a/backend/internal/service/gemini_messages_compat_service_test.go +++ b/backend/internal/service/gemini_messages_compat_service_test.go @@ -832,3 +832,108 @@ func TestParseGeminiRateLimitResetTime(t *testing.T) { }) } } + +// TestGeminiMessagesHandleStreamingResponse_ClosesToolBlockBeforeText guards the +// tool→text ordering in the Gemini→Anthropic (messages) streaming bridge. When +// Gemini emits a functionCall part followed by a text part, the tool_use content +// block must be closed before the text block opens; otherwise the Anthropic SSE +// stream contains overlapping content blocks. The chat-completions sibling +// already enforces this via closeOpenTool(). +func TestGeminiMessagesHandleStreamingResponse_ClosesToolBlockBeforeText(t *testing.T) { + gin.SetMode(gin.TestMode) + + upstreamBody := `data: {"candidates":[{"content":{"parts":[{"functionCall":{"name":"get_weather","args":{"city":"SF"}}}]}}]}` + "\n\n" + + `data: {"candidates":[{"content":{"parts":[{"text":"All done."}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":5,"candidatesTokenCount":3}}` + "\n\n" + + "data: [DONE]\n\n" + + resp := &http.Response{ + StatusCode: http.StatusOK, + Header: http.Header{"Content-Type": []string{"text/event-stream"}}, + Body: io.NopCloser(strings.NewReader(upstreamBody)), + } + + rec := httptest.NewRecorder() + c, _ := gin.CreateTestContext(rec) + + svc := &GeminiMessagesCompatService{} + result, err := svc.handleStreamingResponse(c, resp, time.Now(), "claude-3-5-sonnet") + require.NoError(t, err) + require.NotNil(t, result) + + events := parseAnthropicContentBlockEvents(t, rec.Body.String()) + + // Anthropic allows at most one content block open at a time: every + // content_block_start must be matched by a content_block_stop before the + // next start. Replay the lifecycle and assert there is no overlap. + open := -1 + blockTypes := map[int]string{} + textStarted := false + toolClosed := false + toolClosedBeforeText := false + for _, ev := range events { + switch ev.event { + case "content_block_start": + require.Equalf(t, -1, open, + "content block %d opened while block %d was still open (overlapping blocks)", ev.index, open) + open = ev.index + blockTypes[ev.index] = ev.blockType + if ev.blockType == "text" { + textStarted = true + if toolClosed { + toolClosedBeforeText = true + } + } + case "content_block_stop": + require.Equalf(t, open, ev.index, + "content_block_stop index %d does not match the open block %d", ev.index, open) + if blockTypes[ev.index] == "tool_use" { + toolClosed = true + } + open = -1 + } + } + + require.True(t, textStarted, "expected a text content block to be emitted after the tool call") + require.True(t, toolClosedBeforeText, "tool_use block must be closed before the text block starts") + require.Equal(t, -1, open, "stream ended with a content block still open") +} + +type anthropicContentBlockEvent struct { + event string + index int + blockType string +} + +// parseAnthropicContentBlockEvents extracts content_block_start/stop events (with +// their index and, for starts, the content block type) from an Anthropic SSE body. +func parseAnthropicContentBlockEvents(t *testing.T, raw string) []anthropicContentBlockEvent { + t.Helper() + var events []anthropicContentBlockEvent + for _, chunk := range strings.Split(raw, "\n\n") { + var eventName, dataLine string + for _, line := range strings.Split(chunk, "\n") { + switch { + case strings.HasPrefix(line, "event:"): + eventName = strings.TrimSpace(strings.TrimPrefix(line, "event:")) + case strings.HasPrefix(line, "data:"): + dataLine = strings.TrimSpace(strings.TrimPrefix(line, "data:")) + } + } + if eventName != "content_block_start" && eventName != "content_block_stop" { + continue + } + var payload struct { + Index int `json:"index"` + ContentBlock struct { + Type string `json:"type"` + } `json:"content_block"` + } + require.NoError(t, json.Unmarshal([]byte(dataLine), &payload)) + events = append(events, anthropicContentBlockEvent{ + event: eventName, + index: payload.Index, + blockType: payload.ContentBlock.Type, + }) + } + return events +} From 68901cbfff783af794d96028be8dad3e532c0fe7 Mon Sep 17 00:00:00 2001 From: shaw Date: Fri, 29 May 2026 16:29:29 +0800 Subject: [PATCH 19/20] chore(pricing): update model pricing metadata --- .../model_prices_and_context_window.json | 3307 ++++++++--------- 1 file changed, 1578 insertions(+), 1729 deletions(-) diff --git a/backend/resources/model-pricing/model_prices_and_context_window.json b/backend/resources/model-pricing/model_prices_and_context_window.json index 3cae8c8b..e88ed2da 100644 --- a/backend/resources/model-pricing/model_prices_and_context_window.json +++ b/backend/resources/model-pricing/model_prices_and_context_window.json @@ -1,135 +1,4 @@ { - "claude-3-5-haiku-20241022": { - "cache_creation_input_token_cost": 1e-06, - "cache_creation_input_token_cost_above_1hr": 6e-06, - "cache_read_input_token_cost": 8e-08, - "deprecation_date": "2025-10-01", - "input_cost_per_token": 8e-07, - "litellm_provider": "anthropic", - "max_input_tokens": 200000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_token": 4e-06, - "search_context_cost_per_query": { - "search_context_size_high": 0.01, - "search_context_size_low": 0.01, - "search_context_size_medium": 0.01 - }, - "supports_assistant_prefill": true, - "supports_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true, - "tool_use_system_prompt_tokens": 264 - }, - "claude-3-5-haiku-latest": { - "cache_creation_input_token_cost": 1.25e-06, - "cache_creation_input_token_cost_above_1hr": 6e-06, - "cache_read_input_token_cost": 1e-07, - "deprecation_date": "2025-10-01", - "input_cost_per_token": 1e-06, - "litellm_provider": "anthropic", - "max_input_tokens": 200000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_token": 5e-06, - "search_context_cost_per_query": { - "search_context_size_high": 0.01, - "search_context_size_low": 0.01, - "search_context_size_medium": 0.01 - }, - "supports_assistant_prefill": true, - "supports_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true, - "tool_use_system_prompt_tokens": 264 - }, - "claude-3-5-sonnet-20240620": { - "cache_creation_input_token_cost": 3.75e-06, - "cache_creation_input_token_cost_above_1hr": 6e-06, - "cache_read_input_token_cost": 3e-07, - "deprecation_date": "2025-06-01", - "input_cost_per_token": 3e-06, - "litellm_provider": "anthropic", - "max_input_tokens": 200000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "supports_assistant_prefill": true, - "supports_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_tool_choice": true, - "supports_vision": true, - "tool_use_system_prompt_tokens": 159 - }, - "claude-3-5-sonnet-20241022": { - "cache_creation_input_token_cost": 3.75e-06, - "cache_creation_input_token_cost_above_1hr": 6e-06, - "cache_read_input_token_cost": 3e-07, - "deprecation_date": "2025-10-01", - "input_cost_per_token": 3e-06, - "litellm_provider": "anthropic", - "max_input_tokens": 200000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "search_context_cost_per_query": { - "search_context_size_high": 0.01, - "search_context_size_low": 0.01, - "search_context_size_medium": 0.01 - }, - "supports_assistant_prefill": true, - "supports_computer_use": true, - "supports_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true, - "tool_use_system_prompt_tokens": 159 - }, - "claude-3-5-sonnet-latest": { - "cache_creation_input_token_cost": 3.75e-06, - "cache_creation_input_token_cost_above_1hr": 6e-06, - "cache_read_input_token_cost": 3e-07, - "deprecation_date": "2025-06-01", - "input_cost_per_token": 3e-06, - "litellm_provider": "anthropic", - "max_input_tokens": 200000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "search_context_cost_per_query": { - "search_context_size_high": 0.01, - "search_context_size_low": 0.01, - "search_context_size_medium": 0.01 - }, - "supports_assistant_prefill": true, - "supports_computer_use": true, - "supports_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true, - "tool_use_system_prompt_tokens": 159 - }, "claude-3-7-sonnet-20250219": { "cache_creation_input_token_cost": 3.75e-06, "cache_creation_input_token_cost_above_1hr": 6e-06, @@ -159,34 +28,6 @@ "supports_web_search": true, "tool_use_system_prompt_tokens": 159 }, - "claude-3-7-sonnet-latest": { - "cache_creation_input_token_cost": 3.75e-06, - "cache_creation_input_token_cost_above_1hr": 6e-06, - "cache_read_input_token_cost": 3e-07, - "deprecation_date": "2025-06-01", - "input_cost_per_token": 3e-06, - "litellm_provider": "anthropic", - "max_input_tokens": 200000, - "max_output_tokens": 64000, - "max_tokens": 64000, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "search_context_cost_per_query": { - "search_context_size_high": 0.01, - "search_context_size_low": 0.01, - "search_context_size_medium": 0.01 - }, - "supports_assistant_prefill": true, - "supports_computer_use": true, - "supports_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_tool_choice": true, - "supports_vision": true, - "tool_use_system_prompt_tokens": 159 - }, "claude-3-haiku-20240307": { "cache_creation_input_token_cost": 3e-07, "cache_creation_input_token_cost_above_1hr": 6e-06, @@ -226,28 +67,9 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 395 }, - "claude-3-opus-latest": { - "cache_creation_input_token_cost": 1.875e-05, - "cache_creation_input_token_cost_above_1hr": 6e-06, - "cache_read_input_token_cost": 1.5e-06, - "deprecation_date": "2025-03-01", - "input_cost_per_token": 1.5e-05, - "litellm_provider": "anthropic", - "max_input_tokens": 200000, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 7.5e-05, - "supports_assistant_prefill": true, - "supports_function_calling": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_tool_choice": true, - "supports_vision": true, - "tool_use_system_prompt_tokens": 395 - }, "claude-4-opus-20250514": { "cache_creation_input_token_cost": 1.875e-05, + "cache_creation_input_token_cost_above_1hr": 3e-05, "cache_read_input_token_cost": 1.5e-06, "input_cost_per_token": 1.5e-05, "litellm_provider": "anthropic", @@ -274,6 +96,7 @@ }, "claude-4-sonnet-20250514": { "cache_creation_input_token_cost": 3.75e-06, + "cache_creation_input_token_cost_above_1hr": 6e-06, "cache_creation_input_token_cost_above_200k_tokens": 7.5e-06, "cache_read_input_token_cost": 3e-07, "cache_read_input_token_cost_above_200k_tokens": 6e-07, @@ -447,6 +270,7 @@ "supports_assistant_prefill": true, "supports_computer_use": true, "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -474,6 +298,7 @@ "supports_assistant_prefill": true, "supports_computer_use": true, "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -485,18 +310,14 @@ "claude-opus-4-6": { "cache_creation_input_token_cost": 6.25e-06, "cache_creation_input_token_cost_above_1hr": 1e-05, - "cache_creation_input_token_cost_above_200k_tokens": 1.25e-05, "cache_read_input_token_cost": 5e-07, - "cache_read_input_token_cost_above_200k_tokens": 1e-06, "input_cost_per_token": 5e-06, - "input_cost_per_token_above_200k_tokens": 1e-05, "litellm_provider": "anthropic", "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 2.5e-05, - "output_cost_per_token_above_200k_tokens": 3.75e-05, "provider_specific_entry": { "fast": 6.0, "us": 1.1 @@ -506,9 +327,13 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "supports_adaptive_thinking": true, "supports_assistant_prefill": false, "supports_computer_use": true, "supports_function_calling": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true, + "supports_output_config": true, "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -520,18 +345,14 @@ "claude-opus-4-6-20260205": { "cache_creation_input_token_cost": 6.25e-06, "cache_creation_input_token_cost_above_1hr": 1e-05, - "cache_creation_input_token_cost_above_200k_tokens": 1.25e-05, "cache_read_input_token_cost": 5e-07, - "cache_read_input_token_cost_above_200k_tokens": 1e-06, "input_cost_per_token": 5e-06, - "input_cost_per_token_above_200k_tokens": 1e-05, "litellm_provider": "anthropic", "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 2.5e-05, - "output_cost_per_token_above_200k_tokens": 3.75e-05, "provider_specific_entry": { "fast": 6.0, "us": 1.1 @@ -541,9 +362,13 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "supports_adaptive_thinking": true, "supports_assistant_prefill": false, "supports_computer_use": true, "supports_function_calling": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true, + "supports_output_config": true, "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -555,18 +380,14 @@ "claude-opus-4-6-thinking": { "cache_creation_input_token_cost": 6.25e-06, "cache_creation_input_token_cost_above_1hr": 1e-05, - "cache_creation_input_token_cost_above_200k_tokens": 1.25e-05, "cache_read_input_token_cost": 5e-07, - "cache_read_input_token_cost_above_200k_tokens": 1e-06, "input_cost_per_token": 5e-06, - "input_cost_per_token_above_200k_tokens": 1e-05, "litellm_provider": "anthropic", "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, "mode": "chat", "output_cost_per_token": 2.5e-05, - "output_cost_per_token_above_200k_tokens": 3.75e-05, "provider_specific_entry": { "fast": 6.0, "us": 1.1 @@ -576,9 +397,13 @@ "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "supports_adaptive_thinking": true, "supports_assistant_prefill": false, "supports_computer_use": true, "supports_function_calling": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true, + "supports_output_config": true, "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -587,6 +412,114 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346 }, + "claude-opus-4-7": { + "cache_creation_input_token_cost": 6.25e-06, + "cache_creation_input_token_cost_above_1hr": 1e-05, + "cache_read_input_token_cost": 5e-07, + "input_cost_per_token": 5e-06, + "litellm_provider": "anthropic", + "max_input_tokens": 1000000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 2.5e-05, + "provider_specific_entry": { + "fast": 6.0, + "us": 1.1 + }, + "search_context_cost_per_query": { + "search_context_size_high": 0.01, + "search_context_size_low": 0.01, + "search_context_size_medium": 0.01 + }, + "supports_adaptive_thinking": true, + "supports_assistant_prefill": false, + "supports_computer_use": true, + "supports_function_calling": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true, + "supports_output_config": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_xhigh_reasoning_effort": true, + "tool_use_system_prompt_tokens": 346 + }, + "claude-opus-4-7-20260416": { + "cache_creation_input_token_cost": 6.25e-06, + "cache_creation_input_token_cost_above_1hr": 1e-05, + "cache_read_input_token_cost": 5e-07, + "input_cost_per_token": 5e-06, + "litellm_provider": "anthropic", + "max_input_tokens": 1000000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 2.5e-05, + "provider_specific_entry": { + "fast": 6.0, + "us": 1.1 + }, + "search_context_cost_per_query": { + "search_context_size_high": 0.01, + "search_context_size_low": 0.01, + "search_context_size_medium": 0.01 + }, + "supports_adaptive_thinking": true, + "supports_assistant_prefill": false, + "supports_computer_use": true, + "supports_function_calling": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true, + "supports_output_config": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_xhigh_reasoning_effort": true, + "tool_use_system_prompt_tokens": 346 + }, + "claude-opus-4-8": { + "cache_creation_input_token_cost": 6.25e-06, + "cache_creation_input_token_cost_above_1hr": 1e-05, + "cache_read_input_token_cost": 5e-07, + "input_cost_per_token": 5e-06, + "litellm_provider": "anthropic", + "max_input_tokens": 1000000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 2.5e-05, + "provider_specific_entry": { + "fast": 6.0, + "us": 1.1 + }, + "search_context_cost_per_query": { + "search_context_size_high": 0.01, + "search_context_size_low": 0.01, + "search_context_size_medium": 0.01 + }, + "supports_adaptive_thinking": true, + "supports_assistant_prefill": false, + "supports_computer_use": true, + "supports_function_calling": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true, + "supports_output_config": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_xhigh_reasoning_effort": true, + "tool_use_system_prompt_tokens": 346 + }, "claude-sonnet-4-20250514": { "cache_creation_input_token_cost": 3.75e-06, "cache_creation_input_token_cost_above_1hr": 6e-06, @@ -621,6 +554,7 @@ }, "claude-sonnet-4-5": { "cache_creation_input_token_cost": 3.75e-06, + "cache_creation_input_token_cost_above_1hr": 6e-06, "cache_creation_input_token_cost_above_200k_tokens": 7.5e-06, "cache_read_input_token_cost": 3e-07, "cache_read_input_token_cost_above_200k_tokens": 6e-07, @@ -651,6 +585,7 @@ }, "claude-sonnet-4-5-20250929": { "cache_creation_input_token_cost": 3.75e-06, + "cache_creation_input_token_cost_above_1hr": 6e-06, "cache_creation_input_token_cost_above_200k_tokens": 7.5e-06, "cache_read_input_token_cost": 3e-07, "cache_read_input_token_cost_above_200k_tokens": 6e-07, @@ -682,6 +617,7 @@ }, "claude-sonnet-4-5-20250929-v1:0": { "cache_creation_input_token_cost": 3.75e-06, + "cache_creation_input_token_cost_above_1hr": 6e-06, "cache_creation_input_token_cost_above_200k_tokens": 7.5e-06, "cache_read_input_token_cost": 3e-07, "cache_read_input_token_cost_above_200k_tokens": 6e-07, @@ -707,26 +643,27 @@ }, "claude-sonnet-4-6": { "cache_creation_input_token_cost": 3.75e-06, - "cache_creation_input_token_cost_above_200k_tokens": 7.5e-06, + "cache_creation_input_token_cost_above_1hr": 6e-06, "cache_read_input_token_cost": 3e-07, - "cache_read_input_token_cost_above_200k_tokens": 6e-07, "input_cost_per_token": 3e-06, - "input_cost_per_token_above_200k_tokens": 6e-06, "litellm_provider": "anthropic", - "max_input_tokens": 200000, + "max_input_tokens": 1000000, "max_output_tokens": 64000, "max_tokens": 64000, "mode": "chat", "output_cost_per_token": 1.5e-05, - "output_cost_per_token_above_200k_tokens": 2.25e-05, "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, "search_context_size_medium": 0.01 }, + "supports_adaptive_thinking": true, "supports_assistant_prefill": true, "supports_computer_use": true, "supports_function_calling": true, + "supports_max_reasoning_effort": true, + "supports_minimal_reasoning_effort": true, + "supports_output_config": true, "supports_pdf_input": true, "supports_prompt_caching": true, "supports_reasoning": true, @@ -735,6 +672,54 @@ "supports_vision": true, "tool_use_system_prompt_tokens": 346 }, + "codex-auto-review": { + "cache_read_input_token_cost": 5e-07, + "cache_read_input_token_cost_above_272k_tokens": 1e-06, + "cache_read_input_token_cost_flex": 2.5e-07, + "cache_read_input_token_cost_priority": 1e-06, + "input_cost_per_token": 5e-06, + "input_cost_per_token_above_272k_tokens": 1e-05, + "input_cost_per_token_batches": 2.5e-06, + "input_cost_per_token_flex": 2.5e-06, + "input_cost_per_token_priority": 1e-05, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 3e-05, + "output_cost_per_token_above_272k_tokens": 4.5e-05, + "output_cost_per_token_batches": 1.5e-05, + "output_cost_per_token_flex": 1.5e-05, + "output_cost_per_token_priority": 6e-05, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, "deepseek-chat": { "cache_read_input_token_cost": 2.8e-08, "input_cost_per_token": 2.8e-07, @@ -792,478 +777,9 @@ "supports_reasoning": true, "supports_tool_choice": true }, - "gemini-1.0-pro": { - "input_cost_per_character": 1.25e-07, - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "input_cost_per_video_per_second": 0.002, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 32760, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 3.75e-07, - "output_cost_per_token": 1.5e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#google_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true - }, - "gemini-1.0-pro-001": { - "deprecation_date": "2025-04-09", - "input_cost_per_character": 1.25e-07, - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "input_cost_per_video_per_second": 0.002, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 32760, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 3.75e-07, - "output_cost_per_token": 1.5e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true - }, - "gemini-1.0-pro-002": { - "deprecation_date": "2025-04-09", - "input_cost_per_character": 1.25e-07, - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "input_cost_per_video_per_second": 0.002, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 32760, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 3.75e-07, - "output_cost_per_token": 1.5e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true - }, - "gemini-1.0-pro-vision": { - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "litellm_provider": "vertex_ai-vision-models", - "max_images_per_prompt": 16, - "max_input_tokens": 16384, - "max_output_tokens": 2048, - "max_tokens": 2048, - "max_video_length": 2, - "max_videos_per_prompt": 1, - "mode": "chat", - "output_cost_per_token": 1.5e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.0-pro-vision-001": { - "deprecation_date": "2025-04-09", - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "litellm_provider": "vertex_ai-vision-models", - "max_images_per_prompt": 16, - "max_input_tokens": 16384, - "max_output_tokens": 2048, - "max_tokens": 2048, - "max_video_length": 2, - "max_videos_per_prompt": 1, - "mode": "chat", - "output_cost_per_token": 1.5e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.0-ultra": { - "input_cost_per_character": 1.25e-07, - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "input_cost_per_video_per_second": 0.002, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 8192, - "max_output_tokens": 2048, - "max_tokens": 2048, - "mode": "chat", - "output_cost_per_character": 3.75e-07, - "output_cost_per_token": 1.5e-06, - "source": "As of Jun, 2024. There is no available doc on vertex ai pricing gemini-1.0-ultra-001. Using gemini-1.0-pro pricing. Got max_tokens info here: https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true - }, - "gemini-1.0-ultra-001": { - "input_cost_per_character": 1.25e-07, - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "input_cost_per_video_per_second": 0.002, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 8192, - "max_output_tokens": 2048, - "max_tokens": 2048, - "mode": "chat", - "output_cost_per_character": 3.75e-07, - "output_cost_per_token": 1.5e-06, - "source": "As of Jun, 2024. There is no available doc on vertex ai pricing gemini-1.0-ultra-001. Using gemini-1.0-pro pricing. Got max_tokens info here: https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true - }, - "gemini-1.5-flash": { - "deprecation_date": "2025-09-29", - "input_cost_per_audio_per_second": 2e-06, - "input_cost_per_audio_per_second_above_128k_tokens": 4e-06, - "input_cost_per_character": 1.875e-08, - "input_cost_per_character_above_128k_tokens": 2.5e-07, - "input_cost_per_image": 2e-05, - "input_cost_per_image_above_128k_tokens": 4e-05, - "input_cost_per_token": 7.5e-08, - "input_cost_per_token_above_128k_tokens": 1e-06, - "input_cost_per_video_per_second": 2e-05, - "input_cost_per_video_per_second_above_128k_tokens": 4e-05, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 7.5e-08, - "output_cost_per_character_above_128k_tokens": 1.5e-07, - "output_cost_per_token": 3e-07, - "output_cost_per_token_above_128k_tokens": 6e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-flash-001": { - "deprecation_date": "2025-05-24", - "input_cost_per_audio_per_second": 2e-06, - "input_cost_per_audio_per_second_above_128k_tokens": 4e-06, - "input_cost_per_character": 1.875e-08, - "input_cost_per_character_above_128k_tokens": 2.5e-07, - "input_cost_per_image": 2e-05, - "input_cost_per_image_above_128k_tokens": 4e-05, - "input_cost_per_token": 7.5e-08, - "input_cost_per_token_above_128k_tokens": 1e-06, - "input_cost_per_video_per_second": 2e-05, - "input_cost_per_video_per_second_above_128k_tokens": 4e-05, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 7.5e-08, - "output_cost_per_character_above_128k_tokens": 1.5e-07, - "output_cost_per_token": 3e-07, - "output_cost_per_token_above_128k_tokens": 6e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-flash-002": { - "deprecation_date": "2025-09-24", - "input_cost_per_audio_per_second": 2e-06, - "input_cost_per_audio_per_second_above_128k_tokens": 4e-06, - "input_cost_per_character": 1.875e-08, - "input_cost_per_character_above_128k_tokens": 2.5e-07, - "input_cost_per_image": 2e-05, - "input_cost_per_image_above_128k_tokens": 4e-05, - "input_cost_per_token": 7.5e-08, - "input_cost_per_token_above_128k_tokens": 1e-06, - "input_cost_per_video_per_second": 2e-05, - "input_cost_per_video_per_second_above_128k_tokens": 4e-05, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 7.5e-08, - "output_cost_per_character_above_128k_tokens": 1.5e-07, - "output_cost_per_token": 3e-07, - "output_cost_per_token_above_128k_tokens": 6e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-1.5-flash", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-flash-exp-0827": { - "deprecation_date": "2025-09-29", - "input_cost_per_audio_per_second": 2e-06, - "input_cost_per_audio_per_second_above_128k_tokens": 4e-06, - "input_cost_per_character": 1.875e-08, - "input_cost_per_character_above_128k_tokens": 2.5e-07, - "input_cost_per_image": 2e-05, - "input_cost_per_image_above_128k_tokens": 4e-05, - "input_cost_per_token": 4.688e-09, - "input_cost_per_token_above_128k_tokens": 1e-06, - "input_cost_per_video_per_second": 2e-05, - "input_cost_per_video_per_second_above_128k_tokens": 4e-05, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 1.875e-08, - "output_cost_per_character_above_128k_tokens": 3.75e-08, - "output_cost_per_token": 4.6875e-09, - "output_cost_per_token_above_128k_tokens": 9.375e-09, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-flash-preview-0514": { - "deprecation_date": "2025-09-29", - "input_cost_per_audio_per_second": 2e-06, - "input_cost_per_audio_per_second_above_128k_tokens": 4e-06, - "input_cost_per_character": 1.875e-08, - "input_cost_per_character_above_128k_tokens": 2.5e-07, - "input_cost_per_image": 2e-05, - "input_cost_per_image_above_128k_tokens": 4e-05, - "input_cost_per_token": 7.5e-08, - "input_cost_per_token_above_128k_tokens": 1e-06, - "input_cost_per_video_per_second": 2e-05, - "input_cost_per_video_per_second_above_128k_tokens": 4e-05, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 1.875e-08, - "output_cost_per_character_above_128k_tokens": 3.75e-08, - "output_cost_per_token": 4.6875e-09, - "output_cost_per_token_above_128k_tokens": 9.375e-09, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-pro": { - "deprecation_date": "2025-09-29", - "input_cost_per_audio_per_second": 3.125e-05, - "input_cost_per_audio_per_second_above_128k_tokens": 6.25e-05, - "input_cost_per_character": 3.125e-07, - "input_cost_per_character_above_128k_tokens": 6.25e-07, - "input_cost_per_image": 0.00032875, - "input_cost_per_image_above_128k_tokens": 0.0006575, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_128k_tokens": 2.5e-06, - "input_cost_per_video_per_second": 0.00032875, - "input_cost_per_video_per_second_above_128k_tokens": 0.0006575, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 2097152, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 1.25e-06, - "output_cost_per_character_above_128k_tokens": 2.5e-06, - "output_cost_per_token": 5e-06, - "output_cost_per_token_above_128k_tokens": 1e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-pro-001": { - "deprecation_date": "2025-05-24", - "input_cost_per_audio_per_second": 3.125e-05, - "input_cost_per_audio_per_second_above_128k_tokens": 6.25e-05, - "input_cost_per_character": 3.125e-07, - "input_cost_per_character_above_128k_tokens": 6.25e-07, - "input_cost_per_image": 0.00032875, - "input_cost_per_image_above_128k_tokens": 0.0006575, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_128k_tokens": 2.5e-06, - "input_cost_per_video_per_second": 0.00032875, - "input_cost_per_video_per_second_above_128k_tokens": 0.0006575, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 1.25e-06, - "output_cost_per_character_above_128k_tokens": 2.5e-06, - "output_cost_per_token": 5e-06, - "output_cost_per_token_above_128k_tokens": 1e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-pro-002": { - "deprecation_date": "2025-09-24", - "input_cost_per_audio_per_second": 3.125e-05, - "input_cost_per_audio_per_second_above_128k_tokens": 6.25e-05, - "input_cost_per_character": 3.125e-07, - "input_cost_per_character_above_128k_tokens": 6.25e-07, - "input_cost_per_image": 0.00032875, - "input_cost_per_image_above_128k_tokens": 0.0006575, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_128k_tokens": 2.5e-06, - "input_cost_per_video_per_second": 0.00032875, - "input_cost_per_video_per_second_above_128k_tokens": 0.0006575, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 2097152, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 1.25e-06, - "output_cost_per_character_above_128k_tokens": 2.5e-06, - "output_cost_per_token": 5e-06, - "output_cost_per_token_above_128k_tokens": 1e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-1.5-pro", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gemini-1.5-pro-preview-0215": { - "deprecation_date": "2025-09-29", - "input_cost_per_audio_per_second": 3.125e-05, - "input_cost_per_audio_per_second_above_128k_tokens": 6.25e-05, - "input_cost_per_character": 3.125e-07, - "input_cost_per_character_above_128k_tokens": 6.25e-07, - "input_cost_per_image": 0.00032875, - "input_cost_per_image_above_128k_tokens": 0.0006575, - "input_cost_per_token": 7.8125e-08, - "input_cost_per_token_above_128k_tokens": 1.5625e-07, - "input_cost_per_video_per_second": 0.00032875, - "input_cost_per_video_per_second_above_128k_tokens": 0.0006575, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 1.25e-06, - "output_cost_per_character_above_128k_tokens": 2.5e-06, - "output_cost_per_token": 3.125e-07, - "output_cost_per_token_above_128k_tokens": 6.25e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, - "gemini-1.5-pro-preview-0409": { - "deprecation_date": "2025-09-29", - "input_cost_per_audio_per_second": 3.125e-05, - "input_cost_per_audio_per_second_above_128k_tokens": 6.25e-05, - "input_cost_per_character": 3.125e-07, - "input_cost_per_character_above_128k_tokens": 6.25e-07, - "input_cost_per_image": 0.00032875, - "input_cost_per_image_above_128k_tokens": 0.0006575, - "input_cost_per_token": 7.8125e-08, - "input_cost_per_token_above_128k_tokens": 1.5625e-07, - "input_cost_per_video_per_second": 0.00032875, - "input_cost_per_video_per_second_above_128k_tokens": 0.0006575, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 1.25e-06, - "output_cost_per_character_above_128k_tokens": 2.5e-06, - "output_cost_per_token": 3.125e-07, - "output_cost_per_token_above_128k_tokens": 6.25e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_tool_choice": true - }, - "gemini-1.5-pro-preview-0514": { - "deprecation_date": "2025-09-29", - "input_cost_per_audio_per_second": 3.125e-05, - "input_cost_per_audio_per_second_above_128k_tokens": 6.25e-05, - "input_cost_per_character": 3.125e-07, - "input_cost_per_character_above_128k_tokens": 6.25e-07, - "input_cost_per_image": 0.00032875, - "input_cost_per_image_above_128k_tokens": 0.0006575, - "input_cost_per_token": 7.8125e-08, - "input_cost_per_token_above_128k_tokens": 1.5625e-07, - "input_cost_per_video_per_second": 0.00032875, - "input_cost_per_video_per_second_above_128k_tokens": 0.0006575, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 1.25e-06, - "output_cost_per_character_above_128k_tokens": 2.5e-06, - "output_cost_per_token": 3.125e-07, - "output_cost_per_token_above_128k_tokens": 6.25e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, "gemini-2.0-flash": { "cache_read_input_token_cost": 2.5e-08, - "deprecation_date": "2026-03-31", + "deprecation_date": "2026-06-01", "input_cost_per_audio_token": 7e-07, "input_cost_per_token": 1e-07, "litellm_provider": "vertex_ai-language-models", @@ -1278,6 +794,11 @@ "max_videos_per_prompt": 10, "mode": "chat", "output_cost_per_token": 4e-07, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/pricing#2_0flash", "supported_modalities": [ "text", @@ -1303,7 +824,7 @@ }, "gemini-2.0-flash-001": { "cache_read_input_token_cost": 3.75e-08, - "deprecation_date": "2026-03-31", + "deprecation_date": "2026-06-01", "input_cost_per_audio_token": 1e-06, "input_cost_per_token": 1.5e-07, "litellm_provider": "vertex_ai-language-models", @@ -1318,54 +839,11 @@ "max_videos_per_prompt": 10, "mode": "chat", "output_cost_per_token": 6e-07, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text", - "image" - ], - "supports_audio_output": true, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.0-flash-exp": { - "cache_read_input_token_cost": 3.75e-08, - "input_cost_per_audio_per_second": 0, - "input_cost_per_audio_per_second_above_128k_tokens": 0, - "input_cost_per_character": 0, - "input_cost_per_character_above_128k_tokens": 0, - "input_cost_per_image": 0, - "input_cost_per_image_above_128k_tokens": 0, - "input_cost_per_token": 1.5e-07, - "input_cost_per_token_above_128k_tokens": 0, - "input_cost_per_video_per_second": 0, - "input_cost_per_video_per_second_above_128k_tokens": 0, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 0, - "output_cost_per_character_above_128k_tokens": 0, - "output_cost_per_token": 6e-07, - "output_cost_per_token_above_128k_tokens": 0, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_modalities": [ "text", @@ -1410,7 +888,7 @@ }, "gemini-2.0-flash-lite": { "cache_read_input_token_cost": 1.875e-08, - "deprecation_date": "2026-03-31", + "deprecation_date": "2026-06-01", "input_cost_per_audio_token": 7.5e-08, "input_cost_per_token": 7.5e-08, "litellm_provider": "vertex_ai-language-models", @@ -1424,6 +902,11 @@ "max_videos_per_prompt": 10, "mode": "chat", "output_cost_per_token": 3e-07, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-2.0-flash", "supported_modalities": [ "text", @@ -1446,7 +929,7 @@ }, "gemini-2.0-flash-lite-001": { "cache_read_input_token_cost": 1.875e-08, - "deprecation_date": "2026-03-31", + "deprecation_date": "2026-06-01", "input_cost_per_audio_token": 7.5e-08, "input_cost_per_token": 7.5e-08, "litellm_provider": "vertex_ai-language-models", @@ -1460,6 +943,11 @@ "max_videos_per_prompt": 10, "mode": "chat", "output_cost_per_token": 3e-07, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-2.0-flash", "supported_modalities": [ "text", @@ -1480,235 +968,6 @@ "supports_vision": true, "supports_web_search": true }, - "gemini-2.0-flash-live-preview-04-09": { - "cache_read_input_token_cost": 7.5e-08, - "input_cost_per_audio_token": 3e-06, - "input_cost_per_image": 3e-06, - "input_cost_per_token": 5e-07, - "input_cost_per_video_per_second": 3e-06, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_audio_token": 1.2e-05, - "output_cost_per_token": 2e-06, - "rpm": 10, - "source": "https://cloud.google.com/vertex-ai/docs/generative-ai/model-reference/gemini#gemini-2-0-flash-live-preview-04-09", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text", - "audio" - ], - "supports_audio_output": true, - "supports_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_url_context": true, - "supports_vision": true, - "supports_web_search": true, - "tpm": 250000 - }, - "gemini-2.0-flash-preview-image-generation": { - "cache_read_input_token_cost": 2.5e-08, - "deprecation_date": "2025-11-14", - "input_cost_per_audio_token": 7e-07, - "input_cost_per_token": 1e-07, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_token": 4e-07, - "source": "https://ai.google.dev/pricing#2_0flash", - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text", - "image" - ], - "supports_audio_input": true, - "supports_audio_output": true, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.0-flash-thinking-exp": { - "cache_read_input_token_cost": 0.0, - "deprecation_date": "2025-12-02", - "input_cost_per_audio_per_second": 0, - "input_cost_per_audio_per_second_above_128k_tokens": 0, - "input_cost_per_character": 0, - "input_cost_per_character_above_128k_tokens": 0, - "input_cost_per_image": 0, - "input_cost_per_image_above_128k_tokens": 0, - "input_cost_per_token": 0, - "input_cost_per_token_above_128k_tokens": 0, - "input_cost_per_video_per_second": 0, - "input_cost_per_video_per_second_above_128k_tokens": 0, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 0, - "output_cost_per_character_above_128k_tokens": 0, - "output_cost_per_token": 0, - "output_cost_per_token_above_128k_tokens": 0, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-2.0-flash", - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text", - "image" - ], - "supports_audio_output": true, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.0-flash-thinking-exp-01-21": { - "cache_read_input_token_cost": 0.0, - "deprecation_date": "2025-12-02", - "input_cost_per_audio_per_second": 0, - "input_cost_per_audio_per_second_above_128k_tokens": 0, - "input_cost_per_character": 0, - "input_cost_per_character_above_128k_tokens": 0, - "input_cost_per_image": 0, - "input_cost_per_image_above_128k_tokens": 0, - "input_cost_per_token": 0, - "input_cost_per_token_above_128k_tokens": 0, - "input_cost_per_video_per_second": 0, - "input_cost_per_video_per_second_above_128k_tokens": 0, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65536, - "max_pdf_size_mb": 30, - "max_tokens": 65536, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_character": 0, - "output_cost_per_character_above_128k_tokens": 0, - "output_cost_per_token": 0, - "output_cost_per_token_above_128k_tokens": 0, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#gemini-2.0-flash", - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text", - "image" - ], - "supports_audio_output": false, - "supports_function_calling": false, - "supports_parallel_function_calling": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": false, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.0-pro-exp-02-05": { - "cache_read_input_token_cost": 3.125e-07, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 2097152, - "max_output_tokens": 8192, - "max_pdf_size_mb": 30, - "max_tokens": 8192, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text" - ], - "supports_audio_input": true, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_video_input": true, - "supports_vision": true, - "supports_web_search": true - }, "gemini-2.5-computer-use-preview-10-2025": { "input_cost_per_token": 1.25e-06, "input_cost_per_token_above_200k_tokens": 2.5e-06, @@ -1751,6 +1010,11 @@ "mode": "chat", "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", "supported_endpoints": [ "/v1/chat/completions", @@ -1773,6 +1037,7 @@ "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, + "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, "supports_url_context": true, @@ -1821,6 +1086,7 @@ "supports_pdf_input": true, "supports_prompt_caching": true, "supports_response_schema": true, + "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, "supports_url_context": true, @@ -1828,57 +1094,6 @@ "supports_web_search": false, "tpm": 8000000 }, - "gemini-2.5-flash-image-preview": { - "cache_read_input_token_cost": 7.5e-08, - "deprecation_date": "2026-01-15", - "input_cost_per_audio_token": 1e-06, - "input_cost_per_image_token": 3e-07, - "input_cost_per_token": 3e-07, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "image_generation", - "output_cost_per_image": 0.039, - "output_cost_per_image_token": 3e-05, - "output_cost_per_reasoning_token": 3e-05, - "output_cost_per_token": 3e-05, - "rpm": 100000, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions", - "/v1/batch" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text", - "image" - ], - "supports_audio_output": false, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_url_context": true, - "supports_vision": true, - "supports_web_search": true, - "tpm": 8000000 - }, "gemini-2.5-flash-lite": { "cache_read_input_token_cost": 1e-08, "input_cost_per_audio_token": 3e-07, @@ -1896,6 +1111,11 @@ "mode": "chat", "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", "supported_endpoints": [ "/v1/chat/completions", @@ -1918,6 +1138,7 @@ "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, + "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, "supports_url_context": true, @@ -1942,6 +1163,11 @@ "mode": "chat", "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", "supported_endpoints": [ "/v1/chat/completions", @@ -1987,6 +1213,11 @@ "mode": "chat", "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", "supported_endpoints": [ "/v1/chat/completions", @@ -2087,96 +1318,6 @@ "supports_audio_input": true, "supports_audio_output": true }, - "gemini-2.5-flash-preview-04-17": { - "cache_read_input_token_cost": 3.75e-08, - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 1.5e-07, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_reasoning_token": 3.5e-06, - "output_cost_per_token": 6e-07, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions", - "/v1/batch" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text" - ], - "supports_audio_output": false, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.5-flash-preview-05-20": { - "cache_read_input_token_cost": 7.5e-08, - "deprecation_date": "2025-11-18", - "input_cost_per_audio_token": 1e-06, - "input_cost_per_token": 3e-07, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_reasoning_token": 2.5e-06, - "output_cost_per_token": 2.5e-06, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions", - "/v1/batch" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text" - ], - "supports_audio_output": false, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_url_context": true, - "supports_vision": true, - "supports_web_search": true - }, "gemini-2.5-flash-preview-09-2025": { "cache_read_input_token_cost": 7.5e-08, "input_cost_per_audio_token": 1e-06, @@ -2194,6 +1335,11 @@ "mode": "chat", "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://developers.googleblog.com/en/continuing-to-bring-you-our-latest-models-with-an-improved-gemini-2-5-flash-and-flash-lite-release/", "supported_endpoints": [ "/v1/chat/completions", @@ -2251,6 +1397,11 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_above_200k_tokens": 1.5e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -2271,199 +1422,13 @@ "supports_prompt_caching": true, "supports_reasoning": true, "supports_response_schema": true, + "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, "supports_video_input": true, "supports_vision": true, "supports_web_search": true }, - "gemini-2.5-pro-exp-03-25": { - "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text" - ], - "supports_audio_input": true, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_video_input": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.5-pro-preview-03-25": { - "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, - "deprecation_date": "2025-12-02", - "input_cost_per_audio_token": 1.25e-06, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions", - "/v1/batch" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text" - ], - "supports_audio_output": false, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.5-pro-preview-05-06": { - "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, - "deprecation_date": "2025-12-02", - "input_cost_per_audio_token": 1.25e-06, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions", - "/v1/batch" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text" - ], - "supported_regions": [ - "global" - ], - "supports_audio_output": false, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, - "gemini-2.5-pro-preview-06-05": { - "cache_read_input_token_cost": 1.25e-07, - "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, - "input_cost_per_audio_token": 1.25e-06, - "input_cost_per_token": 1.25e-06, - "input_cost_per_token_above_200k_tokens": 2.5e-06, - "litellm_provider": "vertex_ai-language-models", - "max_audio_length_hours": 8.4, - "max_audio_per_prompt": 1, - "max_images_per_prompt": 3000, - "max_input_tokens": 1048576, - "max_output_tokens": 65535, - "max_pdf_size_mb": 30, - "max_tokens": 65535, - "max_video_length": 1, - "max_videos_per_prompt": 10, - "mode": "chat", - "output_cost_per_token": 1e-05, - "output_cost_per_token_above_200k_tokens": 1.5e-05, - "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions", - "/v1/batch" - ], - "supported_modalities": [ - "text", - "image", - "audio", - "video" - ], - "supported_output_modalities": [ - "text" - ], - "supports_audio_output": false, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true, - "supports_web_search": true - }, "gemini-2.5-pro-preview-tts": { "cache_read_input_token_cost": 1.25e-07, "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, @@ -2483,6 +1448,11 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_above_200k_tokens": 1.5e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-pro-preview", "supported_modalities": [ "text" @@ -2500,7 +1470,7 @@ "supports_vision": true, "supports_web_search": true }, - "gemini-3-flash-preview": { + "gemini-3-flash": { "cache_read_input_token_cost": 5e-08, "cache_read_input_token_cost_priority": 9e-08, "input_cost_per_audio_token": 1e-06, @@ -2521,6 +1491,11 @@ "output_cost_per_reasoning_token": 3e-06, "output_cost_per_token": 3e-06, "output_cost_per_token_priority": 5.4e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, "source": "https://ai.google.dev/pricing/gemini-3", "supported_endpoints": [ "/v1/chat/completions", @@ -2549,7 +1524,65 @@ "supports_tool_choice": true, "supports_url_context": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "web_search_billing_unit": "per_query" + }, + "gemini-3-flash-preview": { + "cache_read_input_token_cost": 5e-08, + "cache_read_input_token_cost_priority": 9e-08, + "input_cost_per_audio_token": 1e-06, + "input_cost_per_audio_token_priority": 1.8e-06, + "input_cost_per_token": 5e-07, + "input_cost_per_token_priority": 9e-07, + "litellm_provider": "vertex_ai-language-models", + "max_audio_length_hours": 8.4, + "max_audio_per_prompt": 1, + "max_images_per_prompt": 3000, + "max_input_tokens": 1048576, + "max_output_tokens": 65535, + "max_pdf_size_mb": 30, + "max_tokens": 65535, + "max_video_length": 1, + "max_videos_per_prompt": 10, + "mode": "chat", + "output_cost_per_reasoning_token": 3e-06, + "output_cost_per_token": 3e-06, + "output_cost_per_token_priority": 5.4e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, + "source": "https://ai.google.dev/pricing/gemini-3", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/completions", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image", + "audio", + "video" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_output": false, + "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_url_context": true, + "supports_vision": true, + "supports_web_search": true, + "web_search_billing_unit": "per_query" }, "gemini-3-pro-image-preview": { "input_cost_per_image": 0.0011, @@ -2564,6 +1597,11 @@ "output_cost_per_image_token": 0.00012, "output_cost_per_token": 1.2e-05, "output_cost_per_token_batches": 6e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -2581,9 +1619,11 @@ "supports_function_calling": false, "supports_prompt_caching": true, "supports_response_schema": true, + "supports_service_tier": true, "supports_system_messages": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "web_search_billing_unit": "per_query" }, "gemini-3-pro-preview": { "cache_creation_input_token_cost_above_200k_tokens": 2.5e-07, @@ -2591,6 +1631,7 @@ "cache_read_input_token_cost_above_200k_tokens": 4e-07, "cache_read_input_token_cost_above_200k_tokens_priority": 7.2e-07, "cache_read_input_token_cost_priority": 3.6e-07, + "deprecation_date": "2026-03-26", "input_cost_per_token": 2e-06, "input_cost_per_token_above_200k_tokens": 4e-06, "input_cost_per_token_above_200k_tokens_priority": 7.2e-06, @@ -2612,6 +1653,11 @@ "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "output_cost_per_token_batches": 6e-06, "output_cost_per_token_priority": 2.16e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -2639,6 +1685,240 @@ "supports_tool_choice": true, "supports_video_input": true, "supports_vision": true, + "supports_web_search": true, + "web_search_billing_unit": "per_query" + }, + "gemini-3.1-flash-image": { + "input_cost_per_image": 0.00056, + "input_cost_per_token": 5e-07, + "litellm_provider": "vertex_ai-language-models", + "max_input_tokens": 65536, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "image_generation", + "output_cost_per_image": 0.0672, + "output_cost_per_image_token": 6e-05, + "output_cost_per_token": 3e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/completions", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text", + "image" + ], + "supports_function_calling": false, + "supports_prompt_caching": true, + "supports_response_schema": true, + "supports_system_messages": true, + "supports_vision": true, + "supports_web_search": true, + "web_search_billing_unit": "per_query" + }, + "gemini-3.1-flash-image-preview": { + "input_cost_per_image": 0.00056, + "input_cost_per_token": 5e-07, + "litellm_provider": "vertex_ai-language-models", + "max_input_tokens": 65536, + "max_output_tokens": 32768, + "max_tokens": 32768, + "mode": "image_generation", + "output_cost_per_image": 0.0672, + "output_cost_per_image_token": 6e-05, + "output_cost_per_token": 3e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/completions", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text", + "image" + ], + "supports_function_calling": false, + "supports_prompt_caching": true, + "supports_response_schema": true, + "supports_system_messages": true, + "supports_vision": true, + "supports_web_search": true, + "web_search_billing_unit": "per_query" + }, + "gemini-3.1-flash-lite": { + "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost_batches": 1.25e-08, + "cache_read_input_token_cost_flex": 1.25e-08, + "cache_read_input_token_cost_per_audio_token": 5e-08, + "cache_read_input_token_cost_priority": 4.5e-08, + "input_cost_per_audio_token": 5e-07, + "input_cost_per_token": 2.5e-07, + "input_cost_per_token_batches": 1.25e-07, + "input_cost_per_token_flex": 1.25e-07, + "input_cost_per_token_priority": 4.5e-07, + "litellm_provider": "vertex_ai-language-models", + "max_audio_length_hours": 8.4, + "max_audio_per_prompt": 1, + "max_images_per_prompt": 3000, + "max_input_tokens": 1048576, + "max_output_tokens": 65536, + "max_pdf_size_mb": 30, + "max_tokens": 65536, + "max_video_length": 1, + "max_videos_per_prompt": 10, + "mode": "chat", + "output_cost_per_reasoning_token": 1.5e-06, + "output_cost_per_token": 1.5e-06, + "output_cost_per_token_batches": 7.5e-07, + "output_cost_per_token_flex": 7.5e-07, + "output_cost_per_token_priority": 2.7e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, + "source": "https://ai.google.dev/gemini-api/docs/pricing#gemini-3.1-flash-lite", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/completions", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image", + "audio", + "video" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true, + "supports_audio_output": false, + "supports_code_execution": true, + "supports_file_search": true, + "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_url_context": true, + "supports_video_input": true, + "supports_vision": true, + "supports_web_search": true, + "web_search_billing_unit": "per_query" + }, + "gemini-3.1-flash-lite-preview": { + "cache_read_input_token_cost": 2.5e-08, + "cache_read_input_token_cost_per_audio_token": 5e-08, + "input_cost_per_audio_token": 5e-07, + "input_cost_per_token": 2.5e-07, + "litellm_provider": "vertex_ai-language-models", + "max_audio_length_hours": 8.4, + "max_audio_per_prompt": 1, + "max_images_per_prompt": 3000, + "max_input_tokens": 1048576, + "max_output_tokens": 65536, + "max_pdf_size_mb": 30, + "max_tokens": 65536, + "max_video_length": 1, + "max_videos_per_prompt": 10, + "mode": "chat", + "output_cost_per_reasoning_token": 1.5e-06, + "output_cost_per_token": 1.5e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, + "source": "https://ai.google.dev/gemini-api/docs/models", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/completions", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image", + "audio", + "video" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true, + "supports_audio_output": false, + "supports_code_execution": true, + "supports_file_search": true, + "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_url_context": true, + "supports_video_input": true, + "supports_vision": true, + "supports_web_search": true, + "web_search_billing_unit": "per_query" + }, + "gemini-3.1-flash-live-preview": { + "input_cost_per_audio_token": 3e-06, + "input_cost_per_image_token": 1e-06, + "input_cost_per_token": 7.5e-07, + "input_cost_per_video_per_second": 3.3333333333333335e-05, + "litellm_provider": "gemini", + "max_input_tokens": 131072, + "max_output_tokens": 65536, + "max_tokens": 65536, + "mode": "chat", + "output_cost_per_audio_token": 1.2e-05, + "output_cost_per_token": 4.5e-06, + "source": "https://ai.google.dev/gemini-api/docs/pricing", + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "text", + "image", + "audio", + "video" + ], + "supported_output_modalities": [ + "text", + "audio" + ], + "supports_audio_input": true, + "supports_audio_output": true, + "supports_function_calling": true, + "supports_vision": true, "supports_web_search": true }, "gemini-3.1-pro-high": { @@ -2669,6 +1949,11 @@ "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "output_cost_per_token_batches": 6e-06, "output_cost_per_token_priority": 2.16e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", "supported_endpoints": [ "/v1/chat/completions", @@ -2697,7 +1982,8 @@ "supports_url_context": true, "supports_video_input": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "web_search_billing_unit": "per_query" }, "gemini-3.1-pro-low": { "cache_creation_input_token_cost_above_200k_tokens": 2.5e-07, @@ -2727,6 +2013,11 @@ "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "output_cost_per_token_batches": 6e-06, "output_cost_per_token_priority": 2.16e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", "supported_endpoints": [ "/v1/chat/completions", @@ -2755,7 +2046,8 @@ "supports_url_context": true, "supports_video_input": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "web_search_billing_unit": "per_query" }, "gemini-3.1-pro-preview": { "cache_creation_input_token_cost_above_200k_tokens": 2.5e-07, @@ -2785,6 +2077,11 @@ "output_cost_per_token_above_200k_tokens_priority": 3.24e-05, "output_cost_per_token_batches": 6e-06, "output_cost_per_token_priority": 2.16e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", "supported_endpoints": [ "/v1/chat/completions", @@ -2813,7 +2110,8 @@ "supports_url_context": true, "supports_video_input": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "web_search_billing_unit": "per_query" }, "gemini-3.1-pro-preview-customtools": { "cache_creation_input_token_cost_above_200k_tokens": 2.5e-07, @@ -2837,6 +2135,11 @@ "output_cost_per_token": 1.2e-05, "output_cost_per_token_above_200k_tokens": 1.8e-05, "output_cost_per_token_batches": 6e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models", "supported_endpoints": [ "/v1/chat/completions", @@ -2864,7 +2167,67 @@ "supports_url_context": true, "supports_video_input": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "web_search_billing_unit": "per_query" + }, + "gemini-3.5-flash": { + "cache_read_input_token_cost": 1.5e-07, + "cache_read_input_token_cost_priority": 2.7e-07, + "input_cost_per_audio_token": 1e-06, + "input_cost_per_audio_token_priority": 1.8e-06, + "input_cost_per_token": 1.5e-06, + "input_cost_per_token_priority": 2.7e-06, + "litellm_provider": "vertex_ai-language-models", + "max_audio_length_hours": 8.4, + "max_audio_per_prompt": 1, + "max_images_per_prompt": 3000, + "max_input_tokens": 1048576, + "max_output_tokens": 65535, + "max_pdf_size_mb": 30, + "max_tokens": 65535, + "max_video_length": 1, + "max_videos_per_prompt": 10, + "mode": "chat", + "output_cost_per_reasoning_token": 9e-06, + "output_cost_per_token": 9e-06, + "output_cost_per_token_priority": 1.62e-05, + "search_context_cost_per_query": { + "search_context_size_high": 0.014, + "search_context_size_low": 0.014, + "search_context_size_medium": 0.014 + }, + "source": "https://ai.google.dev/pricing/gemini-3", + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/completions", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image", + "audio", + "video" + ], + "supported_output_modalities": [ + "text" + ], + "supports_audio_input": true, + "supports_audio_output": false, + "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_url_context": true, + "supports_video_input": true, + "supports_vision": true, + "supports_web_search": true, + "web_search_billing_unit": "per_query" }, "gemini-embedding-001": { "input_cost_per_token": 1.5e-07, @@ -2876,6 +2239,35 @@ "output_vector_size": 3072, "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models" }, + "gemini-embedding-2": { + "input_cost_per_audio_per_second": 0.00016, + "input_cost_per_image": 0.00012, + "input_cost_per_token": 2e-07, + "input_cost_per_video_per_second": 0.00079, + "litellm_provider": "vertex_ai-embedding-models", + "max_input_tokens": 8192, + "max_tokens": 8192, + "mode": "embedding", + "output_cost_per_token": 0, + "output_vector_size": 3072, + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", + "supports_multimodal": true, + "uses_embed_content": true + }, + "gemini-embedding-2-preview": { + "input_cost_per_audio_per_second": 0.00016, + "input_cost_per_image": 0.00012, + "input_cost_per_token": 2e-07, + "input_cost_per_video_per_second": 0.00079, + "litellm_provider": "vertex_ai-embedding-models", + "max_input_tokens": 8192, + "max_tokens": 8192, + "mode": "embedding", + "output_cost_per_token": 0, + "output_vector_size": 3072, + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", + "uses_embed_content": true + }, "gemini-exp-1206": { "cache_read_input_token_cost": 3e-08, "input_cost_per_audio_token": 1e-06, @@ -2894,6 +2286,11 @@ "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, "rpm": 100000, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", "supported_endpoints": [ "/v1/chat/completions", @@ -2930,13 +2327,11 @@ "max_input_tokens": 1000000, "max_output_tokens": 8192, "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 0, + "mode": "embedding", "output_cost_per_token": 0, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/gemini-experimental", - "supports_function_calling": false, - "supports_parallel_function_calling": true, - "supports_tool_choice": true + "output_vector_size": 3072, + "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", + "uses_embed_content": true }, "gemini-flash-latest": { "cache_read_input_token_cost": 3e-08, @@ -2956,6 +2351,11 @@ "output_cost_per_reasoning_token": 2.5e-06, "output_cost_per_token": 2.5e-06, "rpm": 100000, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview", "supported_endpoints": [ "/v1/chat/completions", @@ -3003,6 +2403,11 @@ "output_cost_per_reasoning_token": 4e-07, "output_cost_per_token": 4e-07, "rpm": 15, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-lite", "supported_endpoints": [ "/v1/chat/completions", @@ -3046,13 +2451,17 @@ "max_tokens": 65535, "max_video_length": 1, "max_videos_per_prompt": 10, - "mode": "chat", + "mode": "realtime", "output_cost_per_audio_token": 1.2e-05, "output_cost_per_token": 2e-06, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://ai.google.dev/gemini-api/docs/pricing", "supported_endpoints": [ - "/v1/chat/completions", - "/v1/completions" + "/vertex_ai/live" ], "supported_modalities": [ "text", @@ -3077,38 +2486,6 @@ "supports_vision": true, "supports_web_search": true }, - "gemini-pro": { - "input_cost_per_character": 1.25e-07, - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "input_cost_per_video_per_second": 0.002, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 32760, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 3.75e-07, - "output_cost_per_token": 1.5e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true - }, - "gemini-pro-experimental": { - "input_cost_per_character": 0, - "input_cost_per_token": 0, - "litellm_provider": "vertex_ai-language-models", - "max_input_tokens": 1000000, - "max_output_tokens": 8192, - "max_tokens": 8192, - "mode": "chat", - "output_cost_per_character": 0, - "output_cost_per_token": 0, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/gemini-experimental", - "supports_function_calling": false, - "supports_parallel_function_calling": true, - "supports_tool_choice": true - }, "gemini-pro-latest": { "cache_read_input_token_cost": 1.25e-07, "cache_read_input_token_cost_above_200k_tokens": 2.5e-07, @@ -3128,6 +2505,11 @@ "output_cost_per_token": 1e-05, "output_cost_per_token_above_200k_tokens": 1.5e-05, "rpm": 2000, + "search_context_cost_per_query": { + "search_context_size_high": 0.035, + "search_context_size_low": 0.035, + "search_context_size_medium": 0.035 + }, "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -3155,24 +2537,6 @@ "supports_web_search": true, "tpm": 800000 }, - "gemini-pro-vision": { - "input_cost_per_image": 0.0025, - "input_cost_per_token": 5e-07, - "litellm_provider": "vertex_ai-vision-models", - "max_images_per_prompt": 16, - "max_input_tokens": 16384, - "max_output_tokens": 2048, - "max_tokens": 2048, - "max_video_length": 2, - "max_videos_per_prompt": 1, - "mode": "chat", - "output_cost_per_token": 1.5e-06, - "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/learn/models#foundation_models", - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_tool_choice": true, - "supports_vision": true - }, "gemini-robotics-er-1.5-preview": { "cache_read_input_token_cost": 0, "input_cost_per_audio_token": 1e-06, @@ -3236,31 +2600,6 @@ "supports_system_messages": true, "supports_tool_choice": true }, - "gpt-3.5-turbo-0301": { - "input_cost_per_token": 1.5e-06, - "litellm_provider": "openai", - "max_input_tokens": 4097, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 2e-06, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, - "gpt-3.5-turbo-0613": { - "input_cost_per_token": 1.5e-06, - "litellm_provider": "openai", - "max_input_tokens": 4097, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 2e-06, - "supports_function_calling": true, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, "gpt-3.5-turbo-1106": { "deprecation_date": "2026-09-28", "input_cost_per_token": 1e-06, @@ -3288,18 +2627,6 @@ "supports_system_messages": true, "supports_tool_choice": true }, - "gpt-3.5-turbo-16k-0613": { - "input_cost_per_token": 3e-06, - "litellm_provider": "openai", - "max_input_tokens": 16385, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 4e-06, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, "gpt-3.5-turbo-instruct": { "input_cost_per_token": 1.5e-06, "litellm_provider": "text-completion-openai", @@ -3347,6 +2674,7 @@ "supports_tool_choice": true }, "gpt-4-0314": { + "deprecation_date": "2026-03-26", "input_cost_per_token": 3e-05, "litellm_provider": "openai", "max_input_tokens": 8192, @@ -3354,7 +2682,6 @@ "max_tokens": 4096, "mode": "chat", "output_cost_per_token": 6e-05, - "supports_prompt_caching": true, "supports_system_messages": true, "supports_tool_choice": true }, @@ -3387,57 +2714,6 @@ "supports_system_messages": true, "supports_tool_choice": true }, - "gpt-4-1106-vision-preview": { - "deprecation_date": "2024-12-06", - "input_cost_per_token": 1e-05, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 3e-05, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gpt-4-32k": { - "input_cost_per_token": 6e-05, - "litellm_provider": "openai", - "max_input_tokens": 32768, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 0.00012, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, - "gpt-4-32k-0314": { - "input_cost_per_token": 6e-05, - "litellm_provider": "openai", - "max_input_tokens": 32768, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 0.00012, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, - "gpt-4-32k-0613": { - "input_cost_per_token": 6e-05, - "litellm_provider": "openai", - "max_input_tokens": 32768, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 0.00012, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, "gpt-4-turbo": { "input_cost_per_token": 1e-05, "litellm_provider": "openai", @@ -3485,21 +2761,6 @@ "supports_system_messages": true, "supports_tool_choice": true }, - "gpt-4-vision-preview": { - "deprecation_date": "2024-12-06", - "input_cost_per_token": 1e-05, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_token": 3e-05, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, "gpt-4.1": { "cache_read_input_token_cost": 5e-07, "cache_read_input_token_cost_priority": 8.75e-07, @@ -3535,7 +2796,8 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "gpt-4.1-2025-04-14": { "cache_read_input_token_cost": 5e-07, @@ -3569,7 +2831,8 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "gpt-4.1-mini": { "cache_read_input_token_cost": 1e-07, @@ -3606,7 +2869,8 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "gpt-4.1-mini-2025-04-14": { "cache_read_input_token_cost": 1e-07, @@ -3640,7 +2904,8 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "gpt-4.1-nano": { "cache_read_input_token_cost": 2.5e-08, @@ -3713,47 +2978,6 @@ "supports_tool_choice": true, "supports_vision": true }, - "gpt-4.5-preview": { - "cache_read_input_token_cost": 3.75e-05, - "input_cost_per_token": 7.5e-05, - "input_cost_per_token_batches": 3.75e-05, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 16384, - "max_tokens": 16384, - "mode": "chat", - "output_cost_per_token": 0.00015, - "output_cost_per_token_batches": 7.5e-05, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gpt-4.5-preview-2025-02-27": { - "cache_read_input_token_cost": 3.75e-05, - "deprecation_date": "2025-07-14", - "input_cost_per_token": 7.5e-05, - "input_cost_per_token_batches": 3.75e-05, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 16384, - "max_tokens": 16384, - "mode": "chat", - "output_cost_per_token": 0.00015, - "output_cost_per_token_batches": 7.5e-05, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_response_schema": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, "gpt-4o": { "cache_read_input_token_cost": 1.25e-06, "cache_read_input_token_cost_priority": 2.125e-06, @@ -3857,23 +3081,6 @@ "supports_system_messages": true, "supports_tool_choice": true }, - "gpt-4o-audio-preview-2024-10-01": { - "input_cost_per_audio_token": 4e-05, - "input_cost_per_token": 2.5e-06, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 16384, - "max_tokens": 16384, - "mode": "chat", - "output_cost_per_audio_token": 8e-05, - "output_cost_per_token": 1e-05, - "supports_audio_input": true, - "supports_audio_output": true, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, "gpt-4o-audio-preview-2024-12-17": { "input_cost_per_audio_token": 4e-05, "input_cost_per_token": 2.5e-06, @@ -4077,7 +3284,7 @@ "supports_vision": true }, "gpt-4o-mini-transcribe": { - "input_cost_per_audio_token": 3e-06, + "input_cost_per_audio_token": 1.25e-06, "input_cost_per_token": 1.25e-06, "litellm_provider": "openai", "max_input_tokens": 16000, @@ -4089,7 +3296,7 @@ ] }, "gpt-4o-mini-transcribe-2025-03-20": { - "input_cost_per_audio_token": 3e-06, + "input_cost_per_audio_token": 1.25e-06, "input_cost_per_token": 1.25e-06, "litellm_provider": "openai", "max_input_tokens": 16000, @@ -4101,7 +3308,7 @@ ] }, "gpt-4o-mini-transcribe-2025-12-15": { - "input_cost_per_audio_token": 3e-06, + "input_cost_per_audio_token": 1.25e-06, "input_cost_per_token": 1.25e-06, "litellm_provider": "openai", "max_input_tokens": 16000, @@ -4184,25 +3391,6 @@ "supports_system_messages": true, "supports_tool_choice": true }, - "gpt-4o-realtime-preview-2024-10-01": { - "cache_creation_input_audio_token_cost": 2e-05, - "cache_read_input_token_cost": 2.5e-06, - "input_cost_per_audio_token": 0.0001, - "input_cost_per_token": 5e-06, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 4096, - "max_tokens": 4096, - "mode": "chat", - "output_cost_per_audio_token": 0.0002, - "output_cost_per_token": 2e-05, - "supports_audio_input": true, - "supports_audio_output": true, - "supports_function_calling": true, - "supports_parallel_function_calling": true, - "supports_system_messages": true, - "supports_tool_choice": true - }, "gpt-4o-realtime-preview-2024-12-17": { "cache_read_input_token_cost": 2.5e-06, "input_cost_per_audio_token": 4e-05, @@ -4286,7 +3474,7 @@ "supports_vision": true }, "gpt-4o-transcribe": { - "input_cost_per_audio_token": 6e-06, + "input_cost_per_audio_token": 2.5e-06, "input_cost_per_token": 2.5e-06, "litellm_provider": "openai", "max_input_tokens": 16000, @@ -4298,7 +3486,7 @@ ] }, "gpt-4o-transcribe-diarize": { - "input_cost_per_audio_token": 6e-06, + "input_cost_per_audio_token": 2.5e-06, "input_cost_per_token": 2.5e-06, "litellm_provider": "openai", "max_input_tokens": 16000, @@ -4337,7 +3525,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4346,7 +3536,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-2025-08-07": { "cache_read_input_token_cost": 1.25e-07, @@ -4376,7 +3568,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4385,7 +3579,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-chat": { "cache_read_input_token_cost": 1.25e-07, @@ -4409,7 +3605,9 @@ "text" ], "supports_function_calling": false, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": false, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4417,7 +3615,8 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": false, - "supports_vision": true + "supports_vision": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-chat-latest": { "cache_read_input_token_cost": 1.25e-07, @@ -4441,7 +3640,9 @@ "text" ], "supports_function_calling": false, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": false, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4449,7 +3650,8 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": false, - "supports_vision": true + "supports_vision": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-codex": { "cache_read_input_token_cost": 1.25e-07, @@ -4471,7 +3673,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4479,7 +3683,9 @@ "supports_response_schema": true, "supports_system_messages": false, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-mini": { "cache_read_input_token_cost": 2.5e-08, @@ -4509,7 +3715,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4518,7 +3726,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-mini-2025-08-07": { "cache_read_input_token_cost": 2.5e-08, @@ -4548,7 +3758,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4557,7 +3769,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-nano": { "cache_read_input_token_cost": 5e-09, @@ -4585,7 +3799,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4593,7 +3809,9 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-nano-2025-08-07": { "cache_read_input_token_cost": 5e-09, @@ -4620,7 +3838,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4628,7 +3848,9 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-pro": { "input_cost_per_token": 1.5e-05, @@ -4652,7 +3874,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": false, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4661,7 +3885,8 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-pro-2025-10-06": { "input_cost_per_token": 1.5e-05, @@ -4685,7 +3910,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": false, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4694,7 +3921,8 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-search-api": { "cache_read_input_token_cost": 1.25e-07, @@ -4706,6 +3934,8 @@ "mode": "chat", "output_cost_per_token": 1e-05, "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4713,7 +3943,8 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5-search-api-2025-10-14": { "cache_read_input_token_cost": 1.25e-07, @@ -4725,6 +3956,7 @@ "mode": "chat", "output_cost_per_token": 1e-05, "supports_function_calling": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4732,7 +3964,8 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.1": { "cache_read_input_token_cost": 1.25e-07, @@ -4759,7 +3992,9 @@ "image" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4768,7 +4003,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.1-2025-11-13": { "cache_read_input_token_cost": 1.25e-07, @@ -4795,7 +4032,9 @@ "image" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4804,7 +4043,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.1-chat-latest": { "cache_read_input_token_cost": 1.25e-07, @@ -4831,7 +4072,9 @@ "image" ], "supports_function_calling": false, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": true, "supports_parallel_function_calling": false, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4839,7 +4082,9 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": false, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.1-codex": { "cache_read_input_token_cost": 1.25e-07, @@ -4864,7 +4109,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4872,7 +4119,9 @@ "supports_response_schema": true, "supports_system_messages": false, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.1-codex-max": { "cache_read_input_token_cost": 1.25e-07, @@ -4894,7 +4143,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4902,7 +4153,9 @@ "supports_response_schema": true, "supports_system_messages": false, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true }, "gpt-5.1-codex-mini": { "cache_read_input_token_cost": 2.5e-08, @@ -4927,7 +4180,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4935,7 +4190,9 @@ "supports_response_schema": true, "supports_system_messages": false, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.2": { "cache_read_input_token_cost": 1.75e-07, @@ -4963,7 +4220,9 @@ "image" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -4972,7 +4231,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true }, "gpt-5.2-2025-12-11": { "cache_read_input_token_cost": 1.75e-07, @@ -5000,7 +4261,9 @@ "image" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": true, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -5009,7 +4272,9 @@ "supports_service_tier": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true }, "gpt-5.2-chat-latest": { "cache_read_input_token_cost": 1.75e-07, @@ -5035,7 +4300,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -5043,7 +4310,9 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.2-codex": { "cache_read_input_token_cost": 1.75e-07, @@ -5068,7 +4337,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -5076,7 +4347,9 @@ "supports_response_schema": true, "supports_system_messages": false, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true }, "gpt-5.2-pro": { "input_cost_per_token": 2.1e-05, @@ -5098,7 +4371,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -5107,7 +4382,8 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true }, "gpt-5.2-pro-2025-12-11": { "input_cost_per_token": 2.1e-05, @@ -5129,7 +4405,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -5138,17 +4416,21 @@ "supports_system_messages": true, "supports_tool_choice": true, "supports_vision": true, - "supports_web_search": true + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true }, - "gpt-5.4": { - "cache_read_input_token_cost": 2.5e-07, - "input_cost_per_token": 2.5e-06, + "gpt-5.3-chat-latest": { + "cache_read_input_token_cost": 1.75e-07, + "cache_read_input_token_cost_priority": 3.5e-07, + "input_cost_per_token": 1.75e-06, + "input_cost_per_token_priority": 3.5e-06, "litellm_provider": "openai", - "max_input_tokens": 1050000, - "max_output_tokens": 128000, - "max_tokens": 128000, + "max_input_tokens": 128000, + "max_output_tokens": 16384, + "max_tokens": 16384, "mode": "chat", - "output_cost_per_token": 1.5e-05, + "output_cost_per_token": 1.4e-05, + "output_cost_per_token_priority": 2.8e-05, "supported_endpoints": [ "/v1/chat/completions", "/v1/responses" @@ -5157,111 +4439,13 @@ "text", "image" ], - "supported_output_modalities": [ - "text", - "image" - ], - "supports_function_calling": true, - "supports_native_streaming": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_service_tier": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "codex-auto-review": { - "cache_read_input_token_cost": 2.5e-07, - "input_cost_per_token": 2.5e-06, - "litellm_provider": "openai", - "max_input_tokens": 1050000, - "max_output_tokens": 128000, - "max_tokens": 128000, - "mode": "chat", - "output_cost_per_token": 1.5e-05, - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/responses" - ], - "supported_modalities": [ - "text", - "image" - ], - "supported_output_modalities": [ - "text", - "image" - ], - "supports_function_calling": true, - "supports_native_streaming": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_service_tier": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gpt-5.4-mini": { - "cache_read_input_token_cost": 7.5e-08, - "input_cost_per_token": 7.5e-07, - "litellm_provider": "openai", - "max_input_tokens": 400000, - "max_output_tokens": 128000, - "max_tokens": 128000, - "mode": "chat", - "output_cost_per_token": 4.5e-06, - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/batch", - "/v1/responses" - ], - "supported_modalities": [ - "text", - "image" - ], - "supported_output_modalities": [ - "text" - ], - "supports_function_calling": true, - "supports_native_streaming": true, - "supports_parallel_function_calling": true, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_response_schema": true, - "supports_service_tier": true, - "supports_system_messages": true, - "supports_tool_choice": true, - "supports_vision": true - }, - "gpt-5.4-nano": { - "cache_read_input_token_cost": 2e-08, - "input_cost_per_token": 2e-07, - "litellm_provider": "openai", - "max_input_tokens": 400000, - "max_output_tokens": 128000, - "max_tokens": 128000, - "mode": "chat", - "output_cost_per_token": 1.25e-06, - "supported_endpoints": [ - "/v1/chat/completions", - "/v1/batch", - "/v1/responses" - ], - "supported_modalities": [ - "text", - "image" - ], "supported_output_modalities": [ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -5269,7 +4453,9 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false }, "gpt-5.3-codex": { "cache_read_input_token_cost": 1.75e-07, @@ -5294,7 +4480,9 @@ "text" ], "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, "supports_native_streaming": true, + "supports_none_reasoning_effort": false, "supports_parallel_function_calling": true, "supports_pdf_input": true, "supports_prompt_caching": true, @@ -5302,8 +4490,586 @@ "supports_response_schema": true, "supports_system_messages": false, "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false + }, + "gpt-5.3-codex-spark": { + "cache_read_input_token_cost": 1.75e-07, + "cache_read_input_token_cost_priority": 3.5e-07, + "input_cost_per_token": 1.75e-06, + "input_cost_per_token_priority": 3.5e-06, + "litellm_provider": "openai", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "responses", + "output_cost_per_token": 1.4e-05, + "output_cost_per_token_priority": 2.8e-05, + "supported_endpoints": [ + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, + "supports_native_streaming": true, + "supports_none_reasoning_effort": false, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_system_messages": false, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": false + }, + "gpt-5.4": { + "cache_read_input_token_cost": 2.5e-07, + "cache_read_input_token_cost_above_272k_tokens": 5e-07, + "cache_read_input_token_cost_flex": 1.3e-07, + "cache_read_input_token_cost_priority": 5e-07, + "input_cost_per_token": 2.5e-06, + "input_cost_per_token_above_272k_tokens": 5e-06, + "input_cost_per_token_batches": 1.25e-06, + "input_cost_per_token_flex": 1.25e-06, + "input_cost_per_token_priority": 5e-06, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.5e-05, + "output_cost_per_token_above_272k_tokens": 2.25e-05, + "output_cost_per_token_batches": 7.5e-06, + "output_cost_per_token_flex": 7.5e-06, + "output_cost_per_token_priority": 3e-05, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.4-2026-03-05": { + "cache_read_input_token_cost": 2.5e-07, + "cache_read_input_token_cost_above_272k_tokens": 5e-07, + "cache_read_input_token_cost_flex": 1.3e-07, + "cache_read_input_token_cost_priority": 5e-07, + "input_cost_per_token": 2.5e-06, + "input_cost_per_token_above_272k_tokens": 5e-06, + "input_cost_per_token_batches": 1.25e-06, + "input_cost_per_token_flex": 1.25e-06, + "input_cost_per_token_priority": 5e-06, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.5e-05, + "output_cost_per_token_above_272k_tokens": 2.25e-05, + "output_cost_per_token_batches": 7.5e-06, + "output_cost_per_token_flex": 7.5e-06, + "output_cost_per_token_priority": 3e-05, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, "supports_vision": true }, + "gpt-5.4-mini": { + "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost_batches": 3.75e-08, + "cache_read_input_token_cost_flex": 3.75e-08, + "cache_read_input_token_cost_priority": 1.5e-07, + "input_cost_per_token": 7.5e-07, + "input_cost_per_token_batches": 3.75e-07, + "input_cost_per_token_flex": 3.75e-07, + "input_cost_per_token_priority": 1.5e-06, + "litellm_provider": "openai", + "max_input_tokens": 400000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.5e-06, + "output_cost_per_token_batches": 2.25e-06, + "output_cost_per_token_flex": 2.25e-06, + "output_cost_per_token_priority": 9e-06, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.4-mini-2026-03-17": { + "cache_read_input_token_cost": 7.5e-08, + "cache_read_input_token_cost_batches": 3.75e-08, + "cache_read_input_token_cost_flex": 3.75e-08, + "cache_read_input_token_cost_priority": 1.5e-07, + "input_cost_per_token": 7.5e-07, + "input_cost_per_token_batches": 3.75e-07, + "input_cost_per_token_flex": 3.75e-07, + "input_cost_per_token_priority": 1.5e-06, + "litellm_provider": "openai", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 4.5e-06, + "output_cost_per_token_batches": 2.25e-06, + "output_cost_per_token_flex": 2.25e-06, + "output_cost_per_token_priority": 9e-06, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.4-nano": { + "cache_read_input_token_cost": 2e-08, + "cache_read_input_token_cost_batches": 1e-08, + "cache_read_input_token_cost_flex": 1e-08, + "input_cost_per_token": 2e-07, + "input_cost_per_token_batches": 1e-07, + "input_cost_per_token_flex": 1e-07, + "litellm_provider": "openai", + "max_input_tokens": 400000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.25e-06, + "output_cost_per_token_batches": 6.25e-07, + "output_cost_per_token_flex": 6.25e-07, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.4-nano-2026-03-17": { + "cache_read_input_token_cost": 2e-08, + "cache_read_input_token_cost_batches": 1e-08, + "cache_read_input_token_cost_flex": 1e-08, + "input_cost_per_token": 2e-07, + "input_cost_per_token_batches": 1e-07, + "input_cost_per_token_flex": 1e-07, + "litellm_provider": "openai", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.25e-06, + "output_cost_per_token_batches": 6.25e-07, + "output_cost_per_token_flex": 6.25e-07, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.4-pro": { + "cache_read_input_token_cost": 3e-06, + "cache_read_input_token_cost_above_272k_tokens": 6e-06, + "input_cost_per_token": 3e-05, + "input_cost_per_token_above_272k_tokens": 6e-05, + "input_cost_per_token_batches": 1.5e-05, + "input_cost_per_token_flex": 1.5e-05, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "responses", + "output_cost_per_token": 0.00018, + "output_cost_per_token_above_272k_tokens": 0.00027, + "output_cost_per_token_batches": 9e-05, + "output_cost_per_token_flex": 9e-05, + "supported_endpoints": [ + "/v1/responses", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, + "supports_native_streaming": true, + "supports_none_reasoning_effort": false, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": false, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.4-pro-2026-03-05": { + "cache_read_input_token_cost": 3e-06, + "cache_read_input_token_cost_above_272k_tokens": 6e-06, + "input_cost_per_token": 3e-05, + "input_cost_per_token_above_272k_tokens": 6e-05, + "input_cost_per_token_batches": 1.5e-05, + "input_cost_per_token_flex": 1.5e-05, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "responses", + "output_cost_per_token": 0.00018, + "output_cost_per_token_above_272k_tokens": 0.00027, + "output_cost_per_token_batches": 9e-05, + "output_cost_per_token_flex": 9e-05, + "supported_endpoints": [ + "/v1/responses", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": true, + "supports_native_streaming": true, + "supports_none_reasoning_effort": false, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": false, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.5": { + "cache_read_input_token_cost": 5e-07, + "cache_read_input_token_cost_above_272k_tokens": 1e-06, + "cache_read_input_token_cost_flex": 2.5e-07, + "cache_read_input_token_cost_priority": 1e-06, + "input_cost_per_token": 5e-06, + "input_cost_per_token_above_272k_tokens": 1e-05, + "input_cost_per_token_batches": 2.5e-06, + "input_cost_per_token_flex": 2.5e-06, + "input_cost_per_token_priority": 1e-05, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 3e-05, + "output_cost_per_token_above_272k_tokens": 4.5e-05, + "output_cost_per_token_batches": 1.5e-05, + "output_cost_per_token_flex": 1.5e-05, + "output_cost_per_token_priority": 6e-05, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.5-2026-04-23": { + "cache_read_input_token_cost": 5e-07, + "cache_read_input_token_cost_above_272k_tokens": 1e-06, + "cache_read_input_token_cost_flex": 2.5e-07, + "cache_read_input_token_cost_priority": 1e-06, + "input_cost_per_token": 5e-06, + "input_cost_per_token_above_272k_tokens": 1e-05, + "input_cost_per_token_batches": 2.5e-06, + "input_cost_per_token_flex": 2.5e-06, + "input_cost_per_token_priority": 1e-05, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 3e-05, + "output_cost_per_token_above_272k_tokens": 4.5e-05, + "output_cost_per_token_batches": 1.5e-05, + "output_cost_per_token_flex": 1.5e-05, + "output_cost_per_token_priority": 6e-05, + "supported_endpoints": [ + "/v1/chat/completions", + "/v1/batch", + "/v1/responses" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": true, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.5-pro": { + "cache_read_input_token_cost": 3e-06, + "cache_read_input_token_cost_above_272k_tokens": 6e-06, + "input_cost_per_token": 3e-05, + "input_cost_per_token_above_272k_tokens": 6e-05, + "input_cost_per_token_batches": 1.5e-05, + "input_cost_per_token_flex": 1.5e-05, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "responses", + "output_cost_per_token": 0.00018, + "output_cost_per_token_above_272k_tokens": 0.00027, + "output_cost_per_token_batches": 9e-05, + "output_cost_per_token_flex": 9e-05, + "supported_endpoints": [ + "/v1/responses", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_low_reasoning_effort": false, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": false, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": false, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, + "gpt-5.5-pro-2026-04-23": { + "cache_read_input_token_cost": 3e-06, + "cache_read_input_token_cost_above_272k_tokens": 6e-06, + "input_cost_per_token": 3e-05, + "input_cost_per_token_above_272k_tokens": 6e-05, + "input_cost_per_token_batches": 1.5e-05, + "input_cost_per_token_flex": 1.5e-05, + "litellm_provider": "openai", + "max_input_tokens": 1050000, + "max_output_tokens": 128000, + "max_tokens": 128000, + "mode": "responses", + "output_cost_per_token": 0.00018, + "output_cost_per_token_above_272k_tokens": 0.00027, + "output_cost_per_token_batches": 9e-05, + "output_cost_per_token_flex": 9e-05, + "supported_endpoints": [ + "/v1/responses", + "/v1/batch" + ], + "supported_modalities": [ + "text", + "image" + ], + "supported_output_modalities": [ + "text" + ], + "supports_function_calling": true, + "supports_low_reasoning_effort": false, + "supports_minimal_reasoning_effort": false, + "supports_native_streaming": true, + "supports_none_reasoning_effort": false, + "supports_parallel_function_calling": true, + "supports_pdf_input": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": false, + "supports_service_tier": true, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": true, + "supports_web_search": true, + "supports_xhigh_reasoning_effort": true + }, "gpt-audio": { "input_cost_per_audio_token": 3.2e-05, "input_cost_per_token": 2.5e-06, @@ -5340,6 +5106,39 @@ "supports_tool_choice": true, "supports_vision": false }, + "gpt-audio-1.5": { + "input_cost_per_audio_token": 3.2e-05, + "input_cost_per_token": 2.5e-06, + "litellm_provider": "openai", + "max_input_tokens": 128000, + "max_output_tokens": 16384, + "max_tokens": 16384, + "mode": "chat", + "output_cost_per_audio_token": 6.4e-05, + "output_cost_per_token": 1e-05, + "supported_endpoints": [ + "/v1/chat/completions" + ], + "supported_modalities": [ + "text", + "audio" + ], + "supported_output_modalities": [ + "text", + "audio" + ], + "supports_audio_input": true, + "supports_audio_output": true, + "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, + "supports_prompt_caching": false, + "supports_reasoning": false, + "supports_response_schema": false, + "supports_system_messages": true, + "supports_tool_choice": true, + "supports_vision": false + }, "gpt-audio-2025-08-28": { "input_cost_per_audio_token": 3.2e-05, "input_cost_per_token": 2.5e-06, @@ -5540,6 +5339,38 @@ "supports_pdf_input": true, "supports_vision": true }, + "gpt-image-2": { + "cache_read_input_image_token_cost": 2e-06, + "cache_read_input_token_cost": 1.25e-06, + "input_cost_per_image_token": 8e-06, + "input_cost_per_token": 5e-06, + "litellm_provider": "openai", + "mode": "image_generation", + "output_cost_per_image_token": 3e-05, + "output_cost_per_token": 1e-05, + "supported_endpoints": [ + "/v1/images/generations", + "/v1/images/edits" + ], + "supports_pdf_input": true, + "supports_vision": true + }, + "gpt-image-2-2026-04-21": { + "cache_read_input_image_token_cost": 2e-06, + "cache_read_input_token_cost": 1.25e-06, + "input_cost_per_image_token": 8e-06, + "input_cost_per_token": 5e-06, + "litellm_provider": "openai", + "mode": "image_generation", + "output_cost_per_image_token": 3e-05, + "output_cost_per_token": 1e-05, + "supported_endpoints": [ + "/v1/images/generations", + "/v1/images/edits" + ], + "supports_pdf_input": true, + "supports_vision": true + }, "gpt-realtime": { "cache_creation_input_audio_token_cost": 4e-07, "cache_read_input_token_cost": 4e-07, @@ -5572,6 +5403,70 @@ "supports_system_messages": true, "supports_tool_choice": true }, + "gpt-realtime-1.5": { + "cache_creation_input_audio_token_cost": 4e-07, + "cache_read_input_token_cost": 4e-07, + "input_cost_per_audio_token": 3.2e-05, + "input_cost_per_image": 5e-06, + "input_cost_per_token": 4e-06, + "litellm_provider": "openai", + "max_input_tokens": 32000, + "max_output_tokens": 4096, + "max_tokens": 4096, + "mode": "chat", + "output_cost_per_audio_token": 6.4e-05, + "output_cost_per_token": 1.6e-05, + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "text", + "image", + "audio" + ], + "supported_output_modalities": [ + "text", + "audio" + ], + "supports_audio_input": true, + "supports_audio_output": true, + "supports_function_calling": true, + "supports_parallel_function_calling": true, + "supports_system_messages": true, + "supports_tool_choice": true + }, + "gpt-realtime-2": { + "cache_creation_input_audio_token_cost": 4e-07, + "cache_read_input_token_cost": 4e-07, + "input_cost_per_audio_token": 3.2e-05, + "input_cost_per_image": 5e-06, + "input_cost_per_token": 4e-06, + "litellm_provider": "openai", + "max_input_tokens": 32000, + "max_output_tokens": 4096, + "max_tokens": 4096, + "mode": "chat", + "output_cost_per_audio_token": 6.4e-05, + "output_cost_per_token": 1.6e-05, + "supported_endpoints": [ + "/v1/realtime" + ], + "supported_modalities": [ + "text", + "image", + "audio" + ], + "supported_output_modalities": [ + "text", + "audio" + ], + "supports_audio_input": true, + "supports_audio_output": true, + "supports_function_calling": true, + "supports_parallel_function_calling": true, + "supports_system_messages": true, + "supports_tool_choice": true + }, "gpt-realtime-2025-08-28": { "cache_creation_input_audio_token_cost": 4e-07, "cache_read_input_token_cost": 4e-07, @@ -5720,62 +5615,6 @@ "supports_tool_choice": true, "supports_vision": true }, - "o1-mini": { - "cache_read_input_token_cost": 5.5e-07, - "input_cost_per_token": 1.1e-06, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 65536, - "max_tokens": 65536, - "mode": "chat", - "output_cost_per_token": 4.4e-06, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_vision": true - }, - "o1-mini-2024-09-12": { - "cache_read_input_token_cost": 1.5e-06, - "deprecation_date": "2025-10-27", - "input_cost_per_token": 3e-06, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 65536, - "max_tokens": 65536, - "mode": "chat", - "output_cost_per_token": 1.2e-05, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_vision": true - }, - "o1-preview": { - "cache_read_input_token_cost": 7.5e-06, - "input_cost_per_token": 1.5e-05, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 6e-05, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_vision": true - }, - "o1-preview-2024-09-12": { - "cache_read_input_token_cost": 7.5e-06, - "input_cost_per_token": 1.5e-05, - "litellm_provider": "openai", - "max_input_tokens": 128000, - "max_output_tokens": 32768, - "max_tokens": 32768, - "mode": "chat", - "output_cost_per_token": 6e-05, - "supports_pdf_input": true, - "supports_prompt_caching": true, - "supports_reasoning": true, - "supports_vision": true - }, "o1-pro": { "input_cost_per_token": 0.00015, "input_cost_per_token_batches": 7.5e-05, @@ -5876,7 +5715,8 @@ "supports_response_schema": true, "supports_service_tier": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o3-2025-04-16": { "cache_read_input_token_cost": 5e-07, @@ -5908,7 +5748,8 @@ "supports_response_schema": true, "supports_service_tier": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o3-deep-research": { "cache_read_input_token_cost": 2.5e-06, @@ -5941,7 +5782,8 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o3-deep-research-2025-06-26": { "cache_read_input_token_cost": 2.5e-06, @@ -5974,7 +5816,8 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o3-mini": { "cache_read_input_token_cost": 5.5e-07, @@ -6038,7 +5881,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o3-pro-2025-06-10": { "input_cost_per_token": 2e-05, @@ -6068,7 +5912,8 @@ "supports_reasoning": true, "supports_response_schema": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o4-mini": { "cache_read_input_token_cost": 2.75e-07, @@ -6093,7 +5938,8 @@ "supports_response_schema": true, "supports_service_tier": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o4-mini-2025-04-16": { "cache_read_input_token_cost": 2.75e-07, @@ -6112,7 +5958,8 @@ "supports_response_schema": true, "supports_service_tier": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o4-mini-deep-research": { "cache_read_input_token_cost": 5e-07, @@ -6145,7 +5992,8 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true }, "o4-mini-deep-research-2025-06-26": { "cache_read_input_token_cost": 5e-07, @@ -6178,6 +6026,7 @@ "supports_response_schema": true, "supports_system_messages": true, "supports_tool_choice": true, - "supports_vision": true + "supports_vision": true, + "supports_web_search": true } } From 7321e4dea807651dd6fc309eee3d057ba75e7e70 Mon Sep 17 00:00:00 2001 From: "github-actions[bot]" <41898282+github-actions[bot]@users.noreply.github.com> Date: Fri, 29 May 2026 08:50:18 +0000 Subject: [PATCH 20/20] chore: sync VERSION to 0.1.133 [skip ci] --- backend/cmd/server/VERSION | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/backend/cmd/server/VERSION b/backend/cmd/server/VERSION index 7b9dfc4d..56ebc9e5 100644 --- a/backend/cmd/server/VERSION +++ b/backend/cmd/server/VERSION @@ -1 +1 @@ -0.1.132 +0.1.133