merge upstream main

This commit is contained in:
song
2026-02-02 22:13:50 +08:00
parent 7ade9baa15
commit 0170d19fa7
319 changed files with 40485 additions and 8969 deletions

View File

@@ -82,70 +82,23 @@ func (s *GeminiMessagesCompatService) SelectAccountForModel(ctx context.Context,
}
func (s *GeminiMessagesCompatService) SelectAccountForModelWithExclusions(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}) (*Account, error) {
// 优先检查 context 中的强制平台(/antigravity 路由)
var platform string
forcePlatform, hasForcePlatform := ctx.Value(ctxkey.ForcePlatform).(string)
if hasForcePlatform && forcePlatform != "" {
platform = forcePlatform
} else if groupID != nil {
// 根据分组 platform 决定查询哪种账号
var group *Group
if ctxGroup, ok := ctx.Value(ctxkey.Group).(*Group); ok && IsGroupContextValid(ctxGroup) && ctxGroup.ID == *groupID {
group = ctxGroup
} else {
var err error
group, err = s.groupRepo.GetByIDLite(ctx, *groupID)
if err != nil {
return nil, fmt.Errorf("get group failed: %w", err)
}
}
platform = group.Platform
} else {
// 无分组时只使用原生 gemini 平台
platform = PlatformGemini
// 1. 确定目标平台和调度模式
// Determine target platform and scheduling mode
platform, useMixedScheduling, hasForcePlatform, err := s.resolvePlatformAndSchedulingMode(ctx, groupID)
if err != nil {
return nil, err
}
// gemini 分组支持混合调度(包含启用了 mixed_scheduling 的 antigravity 账户)
// 注意:强制平台模式不走混合调度
useMixedScheduling := platform == PlatformGemini && !hasForcePlatform
cacheKey := "gemini:" + sessionHash
if sessionHash != "" {
accountID, err := s.cache.GetSessionAccountID(ctx, derefGroupID(groupID), cacheKey)
if err == nil && accountID > 0 {
if _, excluded := excludedIDs[accountID]; !excluded {
account, err := s.getSchedulableAccount(ctx, accountID)
// 检查账号是否有效原生平台直接匹配antigravity 需要启用混合调度
if err == nil && account.IsSchedulableForModel(requestedModel) && (requestedModel == "" || s.isModelSupportedByAccount(account, requestedModel)) {
valid := false
if account.Platform == platform {
valid = true
} else if useMixedScheduling && account.Platform == PlatformAntigravity && account.IsMixedSchedulingEnabled() {
valid = true
}
if valid {
usable := true
if s.rateLimitService != nil && requestedModel != "" {
ok, err := s.rateLimitService.PreCheckUsage(ctx, account, requestedModel)
if err != nil {
log.Printf("[Gemini PreCheck] Account %d precheck error: %v", account.ID, err)
}
if !ok {
usable = false
}
}
if usable {
_ = s.cache.RefreshSessionTTL(ctx, derefGroupID(groupID), cacheKey, geminiStickySessionTTL)
return account, nil
}
}
}
}
}
// 2. 尝试粘性会话命中
// Try sticky session hit
if account := s.tryStickySessionHit(ctx, groupID, sessionHash, cacheKey, requestedModel, excludedIDs, platform, useMixedScheduling); account != nil {
return account, nil
}
// 查询可调度账户(强制平台模式:优先按分组查找,找不到再查全部)
// 3. 查询可调度账户(强制平台模式:优先按分组查找,找不到再查全部)
// Query schedulable accounts (force platform mode: try group first, fallback to all)
accounts, err := s.listSchedulableAccountsOnce(ctx, groupID, platform, hasForcePlatform)
if err != nil {
return nil, fmt.Errorf("query accounts failed: %w", err)
@@ -158,56 +111,9 @@ func (s *GeminiMessagesCompatService) SelectAccountForModelWithExclusions(ctx co
}
}
var selected *Account
for i := range accounts {
acc := &accounts[i]
if _, excluded := excludedIDs[acc.ID]; excluded {
continue
}
// 混合调度模式下原生平台直接通过antigravity 需要启用 mixed_scheduling
// 非混合调度模式antigravity 分组):不需要过滤
if useMixedScheduling && acc.Platform == PlatformAntigravity && !acc.IsMixedSchedulingEnabled() {
continue
}
if !acc.IsSchedulableForModel(requestedModel) {
continue
}
if requestedModel != "" && !s.isModelSupportedByAccount(acc, requestedModel) {
continue
}
if s.rateLimitService != nil && requestedModel != "" {
ok, err := s.rateLimitService.PreCheckUsage(ctx, acc, requestedModel)
if err != nil {
log.Printf("[Gemini PreCheck] Account %d precheck error: %v", acc.ID, err)
}
if !ok {
continue
}
}
if selected == nil {
selected = acc
continue
}
if acc.Priority < selected.Priority {
selected = acc
} else if acc.Priority == selected.Priority {
switch {
case acc.LastUsedAt == nil && selected.LastUsedAt != nil:
selected = acc
case acc.LastUsedAt != nil && selected.LastUsedAt == nil:
// keep selected (never used is preferred)
case acc.LastUsedAt == nil && selected.LastUsedAt == nil:
// Prefer OAuth accounts when both are unused (more compatible for Code Assist flows).
if acc.Type == AccountTypeOAuth && selected.Type != AccountTypeOAuth {
selected = acc
}
default:
if acc.LastUsedAt.Before(*selected.LastUsedAt) {
selected = acc
}
}
}
}
// 4. 按优先级 + LRU 选择最佳账号
// Select best account by priority + LRU
selected := s.selectBestGeminiAccount(ctx, accounts, requestedModel, excludedIDs, platform, useMixedScheduling)
if selected == nil {
if requestedModel != "" {
@@ -216,6 +122,8 @@ func (s *GeminiMessagesCompatService) SelectAccountForModelWithExclusions(ctx co
return nil, errors.New("no available Gemini accounts")
}
// 5. 设置粘性会话绑定
// Set sticky session binding
if sessionHash != "" {
_ = s.cache.SetSessionAccountID(ctx, derefGroupID(groupID), cacheKey, selected.ID, geminiStickySessionTTL)
}
@@ -223,6 +131,229 @@ func (s *GeminiMessagesCompatService) SelectAccountForModelWithExclusions(ctx co
return selected, nil
}
// resolvePlatformAndSchedulingMode 解析目标平台和调度模式。
// 返回:平台名称、是否使用混合调度、是否强制平台、错误。
//
// resolvePlatformAndSchedulingMode resolves target platform and scheduling mode.
// Returns: platform name, whether to use mixed scheduling, whether force platform, error.
func (s *GeminiMessagesCompatService) resolvePlatformAndSchedulingMode(ctx context.Context, groupID *int64) (platform string, useMixedScheduling bool, hasForcePlatform bool, err error) {
// 优先检查 context 中的强制平台(/antigravity 路由)
forcePlatform, hasForcePlatform := ctx.Value(ctxkey.ForcePlatform).(string)
if hasForcePlatform && forcePlatform != "" {
return forcePlatform, false, true, nil
}
if groupID != nil {
// 根据分组 platform 决定查询哪种账号
var group *Group
if ctxGroup, ok := ctx.Value(ctxkey.Group).(*Group); ok && IsGroupContextValid(ctxGroup) && ctxGroup.ID == *groupID {
group = ctxGroup
} else {
group, err = s.groupRepo.GetByIDLite(ctx, *groupID)
if err != nil {
return "", false, false, fmt.Errorf("get group failed: %w", err)
}
}
// gemini 分组支持混合调度(包含启用了 mixed_scheduling 的 antigravity 账户)
return group.Platform, group.Platform == PlatformGemini, false, nil
}
// 无分组时只使用原生 gemini 平台
return PlatformGemini, true, false, nil
}
// tryStickySessionHit 尝试从粘性会话获取账号。
// 如果命中且账号可用则返回账号;如果账号不可用则清理会话并返回 nil。
//
// tryStickySessionHit attempts to get account from sticky session.
// Returns account if hit and usable; clears session and returns nil if account unavailable.
func (s *GeminiMessagesCompatService) tryStickySessionHit(
ctx context.Context,
groupID *int64,
sessionHash, cacheKey, requestedModel string,
excludedIDs map[int64]struct{},
platform string,
useMixedScheduling bool,
) *Account {
if sessionHash == "" {
return nil
}
accountID, err := s.cache.GetSessionAccountID(ctx, derefGroupID(groupID), cacheKey)
if err != nil || accountID <= 0 {
return nil
}
if _, excluded := excludedIDs[accountID]; excluded {
return nil
}
account, err := s.getSchedulableAccount(ctx, accountID)
if err != nil {
return nil
}
// 检查账号是否需要清理粘性会话
// Check if sticky session should be cleared
if shouldClearStickySession(account) {
_ = s.cache.DeleteSessionAccountID(ctx, derefGroupID(groupID), cacheKey)
return nil
}
// 验证账号是否可用于当前请求
// Verify account is usable for current request
if !s.isAccountUsableForRequest(ctx, account, requestedModel, platform, useMixedScheduling) {
return nil
}
// 刷新会话 TTL 并返回账号
// Refresh session TTL and return account
_ = s.cache.RefreshSessionTTL(ctx, derefGroupID(groupID), cacheKey, geminiStickySessionTTL)
return account
}
// isAccountUsableForRequest 检查账号是否可用于当前请求。
// 验证:模型调度、模型支持、平台匹配、速率限制预检。
//
// isAccountUsableForRequest checks if account is usable for current request.
// Validates: model scheduling, model support, platform matching, rate limit precheck.
func (s *GeminiMessagesCompatService) isAccountUsableForRequest(
ctx context.Context,
account *Account,
requestedModel, platform string,
useMixedScheduling bool,
) bool {
// 检查模型调度能力
// Check model scheduling capability
if !account.IsSchedulableForModel(requestedModel) {
return false
}
// 检查模型支持
// Check model support
if requestedModel != "" && !s.isModelSupportedByAccount(account, requestedModel) {
return false
}
// 检查平台匹配
// Check platform matching
if !s.isAccountValidForPlatform(account, platform, useMixedScheduling) {
return false
}
// 速率限制预检
// Rate limit precheck
if !s.passesRateLimitPreCheck(ctx, account, requestedModel) {
return false
}
return true
}
// isAccountValidForPlatform 检查账号是否匹配目标平台。
// 原生平台直接匹配;混合调度模式下 antigravity 需要启用 mixed_scheduling。
//
// isAccountValidForPlatform checks if account matches target platform.
// Native platform matches directly; mixed scheduling mode requires antigravity to enable mixed_scheduling.
func (s *GeminiMessagesCompatService) isAccountValidForPlatform(account *Account, platform string, useMixedScheduling bool) bool {
if account.Platform == platform {
return true
}
if useMixedScheduling && account.Platform == PlatformAntigravity && account.IsMixedSchedulingEnabled() {
return true
}
return false
}
// passesRateLimitPreCheck 执行速率限制预检。
// 返回 true 表示通过预检或无需预检。
//
// passesRateLimitPreCheck performs rate limit precheck.
// Returns true if passed or precheck not required.
func (s *GeminiMessagesCompatService) passesRateLimitPreCheck(ctx context.Context, account *Account, requestedModel string) bool {
if s.rateLimitService == nil || requestedModel == "" {
return true
}
ok, err := s.rateLimitService.PreCheckUsage(ctx, account, requestedModel)
if err != nil {
log.Printf("[Gemini PreCheck] Account %d precheck error: %v", account.ID, err)
}
return ok
}
// selectBestGeminiAccount 从候选账号中选择最佳账号(优先级 + LRU + OAuth 优先)。
// 返回 nil 表示无可用账号。
//
// selectBestGeminiAccount selects best account from candidates (priority + LRU + OAuth preferred).
// Returns nil if no available account.
func (s *GeminiMessagesCompatService) selectBestGeminiAccount(
ctx context.Context,
accounts []Account,
requestedModel string,
excludedIDs map[int64]struct{},
platform string,
useMixedScheduling bool,
) *Account {
var selected *Account
for i := range accounts {
acc := &accounts[i]
// 跳过被排除的账号
if _, excluded := excludedIDs[acc.ID]; excluded {
continue
}
// 检查账号是否可用于当前请求
if !s.isAccountUsableForRequest(ctx, acc, requestedModel, platform, useMixedScheduling) {
continue
}
// 选择最佳账号
if selected == nil {
selected = acc
continue
}
if s.isBetterGeminiAccount(acc, selected) {
selected = acc
}
}
return selected
}
// isBetterGeminiAccount 判断 candidate 是否比 current 更优。
// 规则优先级更高数值更小优先同优先级时未使用过的优先OAuth > 非 OAuth其次是最久未使用的。
//
// isBetterGeminiAccount checks if candidate is better than current.
// Rules: higher priority (lower value) wins; same priority: never used (OAuth > non-OAuth) > least recently used.
func (s *GeminiMessagesCompatService) isBetterGeminiAccount(candidate, current *Account) bool {
// 优先级更高(数值更小)
if candidate.Priority < current.Priority {
return true
}
if candidate.Priority > current.Priority {
return false
}
// 同优先级,比较最后使用时间
switch {
case candidate.LastUsedAt == nil && current.LastUsedAt != nil:
// candidate 从未使用,优先
return true
case candidate.LastUsedAt != nil && current.LastUsedAt == nil:
// current 从未使用,保持
return false
case candidate.LastUsedAt == nil && current.LastUsedAt == nil:
// 都未使用,优先选择 OAuth 账号(更兼容 Code Assist 流程)
return candidate.Type == AccountTypeOAuth && current.Type != AccountTypeOAuth
default:
// 都使用过,选择最久未使用的
return candidate.LastUsedAt.Before(*current.LastUsedAt)
}
}
// isModelSupportedByAccount 根据账户平台检查模型支持
func (s *GeminiMessagesCompatService) isModelSupportedByAccount(account *Account, requestedModel string) bool {
if account.Platform == PlatformAntigravity {
@@ -1864,6 +1995,7 @@ func collectGeminiSSE(body io.Reader, isOAuth bool) (map[string]any, *ClaudeUsag
var last map[string]any
var lastWithParts map[string]any
var collectedTextParts []string // Collect all text parts for aggregation
usage := &ClaudeUsage{}
for {
@@ -1875,7 +2007,7 @@ func collectGeminiSSE(body io.Reader, isOAuth bool) (map[string]any, *ClaudeUsag
switch payload {
case "", "[DONE]":
if payload == "[DONE]" {
return pickGeminiCollectResult(last, lastWithParts), usage, nil
return mergeCollectedTextParts(pickGeminiCollectResult(last, lastWithParts), collectedTextParts), usage, nil
}
default:
var parsed map[string]any
@@ -1894,6 +2026,12 @@ func collectGeminiSSE(body io.Reader, isOAuth bool) (map[string]any, *ClaudeUsag
}
if parts := extractGeminiParts(parsed); len(parts) > 0 {
lastWithParts = parsed
// Collect text from each part for aggregation
for _, part := range parts {
if text, ok := part["text"].(string); ok && text != "" {
collectedTextParts = append(collectedTextParts, text)
}
}
}
}
}
@@ -1908,7 +2046,7 @@ func collectGeminiSSE(body io.Reader, isOAuth bool) (map[string]any, *ClaudeUsag
}
}
return pickGeminiCollectResult(last, lastWithParts), usage, nil
return mergeCollectedTextParts(pickGeminiCollectResult(last, lastWithParts), collectedTextParts), usage, nil
}
func pickGeminiCollectResult(last map[string]any, lastWithParts map[string]any) map[string]any {
@@ -1921,6 +2059,83 @@ func pickGeminiCollectResult(last map[string]any, lastWithParts map[string]any)
return map[string]any{}
}
// mergeCollectedTextParts merges all collected text chunks into the final response.
// This fixes the issue where non-streaming responses only returned the last chunk
// instead of the complete aggregated text.
func mergeCollectedTextParts(response map[string]any, textParts []string) map[string]any {
if len(textParts) == 0 {
return response
}
// Join all text parts
mergedText := strings.Join(textParts, "")
// Deep copy response
result := make(map[string]any)
for k, v := range response {
result[k] = v
}
// Get or create candidates
candidates, ok := result["candidates"].([]any)
if !ok || len(candidates) == 0 {
candidates = []any{map[string]any{}}
}
// Get first candidate
candidate, ok := candidates[0].(map[string]any)
if !ok {
candidate = make(map[string]any)
candidates[0] = candidate
}
// Get or create content
content, ok := candidate["content"].(map[string]any)
if !ok {
content = map[string]any{"role": "model"}
candidate["content"] = content
}
// Get existing parts
existingParts, ok := content["parts"].([]any)
if !ok {
existingParts = []any{}
}
// Find and update first text part, or create new one
newParts := make([]any, 0, len(existingParts)+1)
textUpdated := false
for _, p := range existingParts {
pm, ok := p.(map[string]any)
if !ok {
newParts = append(newParts, p)
continue
}
if _, hasText := pm["text"]; hasText && !textUpdated {
// Replace with merged text
newPart := make(map[string]any)
for k, v := range pm {
newPart[k] = v
}
newPart["text"] = mergedText
newParts = append(newParts, newPart)
textUpdated = true
} else {
newParts = append(newParts, pm)
}
}
if !textUpdated {
newParts = append([]any{map[string]any{"text": mergedText}}, newParts...)
}
content["parts"] = newParts
result["candidates"] = candidates
return result
}
type geminiNativeStreamResult struct {
usage *ClaudeUsage
firstTokenMs *int
@@ -2312,9 +2527,13 @@ func extractGeminiUsage(geminiResp map[string]any) *ClaudeUsage {
}
prompt, _ := asInt(usageMeta["promptTokenCount"])
cand, _ := asInt(usageMeta["candidatesTokenCount"])
cached, _ := asInt(usageMeta["cachedContentTokenCount"])
// 注意Gemini 的 promptTokenCount 包含 cachedContentTokenCount
// 但 Claude 的 input_tokens 不包含 cache_read_input_tokens需要减去
return &ClaudeUsage{
InputTokens: prompt,
OutputTokens: cand,
InputTokens: prompt - cached,
OutputTokens: cand,
CacheReadInputTokens: cached,
}
}