oadmin/sub2api
1
0
Fork 0
Code Issues Pull Requests 1 Actions 2 Packages Projects Releases Wiki Activity

Merge branch 'Wei-Shaw:main' into main

Browse Source
This commit is contained in:
程序猿MT
2026-02-12 23:12:41 +08:00
committed by GitHub
parent 8da5fac69e c7b42148a5
commit 174d7c774d
30 changed files with 583 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
backend/internal/pkg/antigravity/claude_types.go
View File

@@ -27,7 +27,7 @@ type ClaudeMessage struct {
// ThinkingConfig Thinking 配置
type ThinkingConfig struct {
Type string `json:"type"` // "enabled" or "disabled"
Type string `json:"type"` // "enabled" / "adaptive" / "disabled"
BudgetTokens int `json:"budget_tokens,omitempty"` // thinking budget
}

1
backend/internal/pkg/antigravity/gemini_types.go
View File

@@ -155,6 +155,7 @@ type GeminiUsageMetadata struct {
CandidatesTokenCount int `json:"candidatesTokenCount,omitempty"`
CachedContentTokenCount int `json:"cachedContentTokenCount,omitempty"`
TotalTokenCount int `json:"totalTokenCount,omitempty"`
ThoughtsTokenCount int `json:"thoughtsTokenCount,omitempty"` // thinking tokens（按输出价格计费）
}
// GeminiGroundingMetadata Gemini grounding 元数据（Web Search）

32
backend/internal/pkg/antigravity/request_transformer.go
View File

@@ -64,6 +64,10 @@ const MaxTokensBudgetPadding = 1000
// Gemini 2.5 Flash thinking budget 上限
const Gemini25FlashThinkingBudgetLimit = 24576
// 对于 Antigravity 的 Claude（budget-only）模型，该语义最终等价为 thinkingBudget=24576。
// 这里复用相同数值以保持行为一致。
const ClaudeAdaptiveHighThinkingBudgetTokens = Gemini25FlashThinkingBudgetLimit
// ensureMaxTokensGreaterThanBudget 确保 max_tokens > budget_tokens
// Claude API 要求启用 thinking 时，max_tokens 必须大于 thinking.budget_tokens
// 返回调整后的 maxTokens 和是否进行了调整
@@ -96,7 +100,7 @@ func TransformClaudeToGeminiWithOptions(claudeReq *ClaudeRequest, projectID, map
}
// 检测是否启用 thinking
isThinkingEnabled := claudeReq.Thinking != nil && claudeReq.Thinking.Type == "enabled"
isThinkingEnabled := claudeReq.Thinking != nil && (claudeReq.Thinking.Type == "enabled" || claudeReq.Thinking.Type == "adaptive")
// 只有 Gemini 模型支持 dummy thought workaround
// Claude 模型通过 Vertex/Google API 需要有效的 thought signatures
@@ -198,8 +202,7 @@ type modelInfo struct {
// modelInfoMap 模型前缀 → 模型信息映射
// 只有在此映射表中的模型才会注入身份提示词
// 注意：当前 claude-opus-4-6 会被映射到 claude-opus-4-5-thinking，
// 但保留此条目以便后续 Antigravity 上游支持 4.6 时快速切换
// 注意：模型映射逻辑在网关层完成；这里仅用于按模型前缀判断是否注入身份提示词。
var modelInfoMap = map[string]modelInfo{
"claude-opus-4-5": {DisplayName: "Claude Opus 4.5", CanonicalID: "claude-opus-4-5-20250929"},
"claude-opus-4-6": {DisplayName: "Claude Opus 4.6", CanonicalID: "claude-opus-4-6"},
@@ -593,6 +596,10 @@ func maxOutputTokensLimit(model string) int {
return maxOutputTokensUpperBound
}
func isAntigravityOpus46Model(model string) bool {
return strings.HasPrefix(strings.ToLower(model), "claude-opus-4-6")
}
func buildGenerationConfig(req *ClaudeRequest) *GeminiGenerationConfig {
maxLimit := maxOutputTokensLimit(req.Model)
config := &GeminiGenerationConfig{
@@ -606,25 +613,36 @@ func buildGenerationConfig(req *ClaudeRequest) *GeminiGenerationConfig {
}
// Thinking 配置
if req.Thinking != nil && req.Thinking.Type == "enabled" {
if req.Thinking != nil && (req.Thinking.Type == "enabled" || req.Thinking.Type == "adaptive") {
config.ThinkingConfig = &GeminiThinkingConfig{
IncludeThoughts: true,
}
// - thinking.type=enabled：budget_tokens>0 用显式预算
// - thinking.type=adaptive：仅在 Antigravity 的 Opus 4.6 上覆写为 （24576）
budget := -1
if req.Thinking.BudgetTokens > 0 {
budget := req.Thinking.BudgetTokens
budget = req.Thinking.BudgetTokens
}
if req.Thinking.Type == "adaptive" && isAntigravityOpus46Model(req.Model) {
budget = ClaudeAdaptiveHighThinkingBudgetTokens
}
// 正预算需要做上限与 max_tokens 约束；动态预算（-1）直接透传给上游。
if budget > 0 {
// gemini-2.5-flash 上限
if strings.Contains(req.Model, "gemini-2.5-flash") && budget > Gemini25FlashThinkingBudgetLimit {
budget = Gemini25FlashThinkingBudgetLimit
}
config.ThinkingConfig.ThinkingBudget = budget
// 自动修正：max_tokens 必须大于 budget_tokens
// 自动修正：max_tokens 必须大于 budget_tokens（Claude 上游要求）
if adjusted, ok := ensureMaxTokensGreaterThanBudget(config.MaxOutputTokens, budget); ok {
log.Printf("[Antigravity] Auto-adjusted max_tokens from %d to %d (must be > budget_tokens=%d)",
config.MaxOutputTokens, adjusted, budget)
config.MaxOutputTokens = adjusted
}
}
config.ThinkingConfig.ThinkingBudget = budget
}
if config.MaxOutputTokens > maxLimit {

90
backend/internal/pkg/antigravity/request_transformer_test.go
View File

@@ -259,3 +259,93 @@ func TestBuildTools_CustomTypeTools(t *testing.T) {
})
}
}
func TestBuildGenerationConfig_ThinkingDynamicBudget(t *testing.T) {
tests := []struct {
name string
model string
thinking *ThinkingConfig
wantBudget int
wantPresent bool
}{
{
name: "enabled without budget defaults to dynamic (-1)",
model: "claude-opus-4-6-thinking",
thinking: &ThinkingConfig{Type: "enabled"},
wantBudget: -1,
wantPresent: true,
},
{
name: "enabled with budget uses the provided value",
model: "claude-opus-4-6-thinking",
thinking: &ThinkingConfig{Type: "enabled", BudgetTokens: 1024},
wantBudget: 1024,
wantPresent: true,
},
{
name: "enabled with -1 budget uses dynamic (-1)",
model: "claude-opus-4-6-thinking",
thinking: &ThinkingConfig{Type: "enabled", BudgetTokens: -1},
wantBudget: -1,
wantPresent: true,
},
{
name: "adaptive on opus4.6 maps to high budget (24576)",
model: "claude-opus-4-6-thinking",
thinking: &ThinkingConfig{Type: "adaptive", BudgetTokens: 20000},
wantBudget: ClaudeAdaptiveHighThinkingBudgetTokens,
wantPresent: true,
},
{
name: "adaptive on non-opus model keeps default dynamic (-1)",
model: "claude-sonnet-4-5-thinking",
thinking: &ThinkingConfig{Type: "adaptive"},
wantBudget: -1,
wantPresent: true,
},
{
name: "disabled does not emit thinkingConfig",
model: "claude-opus-4-6-thinking",
thinking: &ThinkingConfig{Type: "disabled", BudgetTokens: 1024},
wantBudget: 0,
wantPresent: false,
},
{
name: "nil thinking does not emit thinkingConfig",
model: "claude-opus-4-6-thinking",
thinking: nil,
wantBudget: 0,
wantPresent: false,
},
}
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
req := &ClaudeRequest{
Model: tt.model,
Thinking: tt.thinking,
}
cfg := buildGenerationConfig(req)
if cfg == nil {
t.Fatalf("expected non-nil generationConfig")
}
if tt.wantPresent {
if cfg.ThinkingConfig == nil {
t.Fatalf("expected thinkingConfig to be present")
}
if !cfg.ThinkingConfig.IncludeThoughts {
t.Fatalf("expected includeThoughts=true")
}
if cfg.ThinkingConfig.ThinkingBudget != tt.wantBudget {
t.Fatalf("expected thinkingBudget=%d, got %d", tt.wantBudget, cfg.ThinkingConfig.ThinkingBudget)
}
return
}
if cfg.ThinkingConfig != nil {
t.Fatalf("expected thinkingConfig to be nil, got %+v", cfg.ThinkingConfig)
}
})
}
}

2
backend/internal/pkg/antigravity/response_transformer.go
View File

@@ -280,7 +280,7 @@ func (p *NonStreamingProcessor) buildResponse(geminiResp *GeminiResponse, respon
if geminiResp.UsageMetadata != nil {
cached := geminiResp.UsageMetadata.CachedContentTokenCount
usage.InputTokens = geminiResp.UsageMetadata.PromptTokenCount - cached
usage.OutputTokens = geminiResp.UsageMetadata.CandidatesTokenCount
usage.OutputTokens = geminiResp.UsageMetadata.CandidatesTokenCount + geminiResp.UsageMetadata.ThoughtsTokenCount
usage.CacheReadInputTokens = cached
}

4
backend/internal/pkg/antigravity/stream_transformer.go
View File

@@ -85,7 +85,7 @@ func (p *StreamingProcessor) ProcessLine(line string) []byte {
if geminiResp.UsageMetadata != nil {
cached := geminiResp.UsageMetadata.CachedContentTokenCount
p.inputTokens = geminiResp.UsageMetadata.PromptTokenCount - cached
p.outputTokens = geminiResp.UsageMetadata.CandidatesTokenCount
p.outputTokens = geminiResp.UsageMetadata.CandidatesTokenCount + geminiResp.UsageMetadata.ThoughtsTokenCount
p.cacheReadTokens = cached
}
@@ -146,7 +146,7 @@ func (p *StreamingProcessor) emitMessageStart(v1Resp *V1InternalResponse) []byte
if v1Resp.Response.UsageMetadata != nil {
cached := v1Resp.Response.UsageMetadata.CachedContentTokenCount
usage.InputTokens = v1Resp.Response.UsageMetadata.PromptTokenCount - cached
usage.OutputTokens = v1Resp.Response.UsageMetadata.CandidatesTokenCount
usage.OutputTokens = v1Resp.Response.UsageMetadata.CandidatesTokenCount + v1Resp.Response.UsageMetadata.ThoughtsTokenCount
usage.CacheReadInputTokens = cached
}