Skip to content

feat(gemini,antigravity): add Gemini 4 Hydrogen/Carbon support and dynamic effort variant routing - #8778

Closed
universonic wants to merge 7 commits into
Wei-Shaw:mainfrom
universonic:feat/gemini-4-hydrogen-claude55-effort
Closed

universonic wants to merge 7 commits into
Wei-Shaw:mainfrom
universonic:feat/gemini-4-hydrogen-claude55-effort

Conversation

@universonic

Copy link
Copy Markdown

背景说明

本 PR 基于 #8661(feat(gemini): add Gemini 4 Carbon support,保留原提交记录)构建(计费部分与 #8746 存在交集),在此基础上新增了对 gemini-4-hydrogen 的完整支持,并统一优化了 Antigravity 渠道下的 effort 动态变体路由、混合分组冷却对齐与计费档位归因。


改了什么 & 为什么要这样改

1. 新增 gemini-4-hydrogen 全链路支持(含 xhigh 档位)

  • 改动内容:
    • 在 gemini / geminicli 默认目录与前端 gemini 白名单中新增 gemini-4-hydrogen;在 Antigravity 目录、默认映射与前端预设中新增 gemini-4-hydrogen-{low,medium,high,xhigh} 及动态自映射 gemini-4-hydrogen -> gemini-4-hydrogen。
    • 在 request_transformer.go、gemini_compat_thinking.go、antigravity_gemini_thinking_variant.go 与 gemini_reasoning_effort.go 中为 gemini-4-hydrogen 支持 low / medium / high / xhigh 四个 thinkingLevel(max 映射为 xhigh;其余不支持 xhigh 的模型自动钳制/降级为 high),禁用 thinkingBudget。
    • 在价格快照、BillingService 兜底计费及 PricingService 档位别名归一中新增 gemini-4-hydrogen 定价(标准 $1.50 / $7.50 / $0.075 每百万 token,Priority 1.8x,1M 输入上下文 / 64K 同步输出,仅隐式缓存无存储费)。
  • 原因:gemini-4-hydrogen 与 Carbon 同样只接受 thinkingLevel(传 thinkingBudget 或 minimal 会触发上游 HTTP 400),但额外开放了 xhigh 深度思考档位,且在 Antigravity 上游目录中拆分为四个带档位后缀的模型 ID;若沿用旧预算路径或缺少 -xhigh 变体映射会导致请求 400、xhigh 档位丢失或上游 404。

2. 修复并统一 Antigravity 前置变体选择与兼容层的 effort 解析

  • 改动内容:
    • 提取统一的 ResolveClaudeThinkingLevel(优先级:thinking.type == "disabled" > output_config.effort > thinking.budget_tokens),接入 AntigravityGatewayService.Forward 与 prepareAntigravityCompatCall。
    • 在 buildAntigravityCompatGeminiBody 中补齐 applyGeminiCompatThinkingConfig,并统一 /v1/chat/completions 对嵌套 reasoning.effort 与平铺 reasoning_effort 的解析与请求转换。
  • 原因:feat(gemini): add Gemini 4 Carbon support #8661 为 ClaudeRequest 引入了 OutputConfig,但网关前置的变体挑选函数此前仅读取 thinking.budget_tokens(未传 thinking 时恒回退 high),而下游转换器见到已带档位后缀的模型又会跳过 effort 写入;同时 Antigravity 的 OpenAI 兼容层对裸模型(如 gemini-4-carbon)遗漏了 thinkingConfig 注入,会导致客户端传入的 output_config.effort / reasoning_effort 被静默忽略或覆盖为 high。

3. 支持 Antigravity 下 Claude 5.5(claude-opus-5-5 / claude-sonnet-5-5)随 effort 动态映射变体

  • 改动内容:
    • 将 DefaultAntigravityModelMapping 中的 claude-opus-5-5 与 claude-sonnet-5-5 由固定 -high 调整为动态自映射,纳入 isAntigravityThinkingVariantFamily 按请求 effort 自动路由至 -low / -medium / -high(缺省仍落到 -high;若管理员显式配置 base -> *-high 则仍优先尊重固定映射;单独添加 base -> base 预设时自动回退默认分档目录避免发裸名 404)。
    • 前端 useModelWhitelist.ts 同步补齐 Opus 5.5 / Sonnet 5.5 动态预设与 -low 预设;后端 configuredUpstreamModelsForCapabilitySync 与 GetUpstreamModelMetadata 支持动态分档裸键的能力同步与多档位元数据聚合。
  • 原因:Antigravity 上游将 Claude 5.5 拆分为 -low / -medium / -high 三个独立模型,此前默认写死 -high 会使 Claude Code / OpenCode 等客户端调节 effort 时无法切到对应档位。该分档逻辑严格限制在 PlatformAntigravity 内,原生 PlatformAnthropic(OAuth / SetupToken / APIKey / Bedrock)保持裸模型 ID 与既有行为完全不变。

4. 调度准入、模型级冷却归属与实际生效计费档位对齐

  • 改动内容:
    • 调度准入(gateway_scheduling.go、gemini_messages_compat_service.go)支持仅同步了后缀变体(含单 -xhigh 目录)的 Antigravity 账号接纳裸模型请求。
    • 新增 WithAntigravityThinkingLevelFromClaudeBody(读取分组策略改写后的最终请求体)、WithAntigravityThinkingLevelFromGeminiBody,配合 resolveFinalAntigravityModelKey 与 antigravityRateLimitModelKey(含 TestConnection)使调度预检与转发限流统一按实际目标变体读写冷却状态。
    • 新增 effectiveAntigravityGeminiReasoningEffort,区分入站请求的 RequestedReasoningEffort 与上游实际生效的 ReasoningEffort(覆盖档位钳制、账号缺档降级、签名/预算整流重试及 web_search 降级至 gemini-2.5-flash 等场景)。
  • 原因:消除“转发阶段支持裸模型分档但调度阶段因无裸键提前拒单”、“-low 限流误冷却默认 -high 或在混合分组(Antigravity + Anthropic 订阅)中因调度与转发档位不一致白白消耗故障切换次数”、以及“上游实际档位发生钳制/降级时仍按未生效高档倍率计费”等边界问题。

本地验证

  • go test -tags=unit ./internal/domain ./internal/pkg/antigravity/... ./internal/pkg/gemini/... ./internal/pkg/geminicli/... ./internal/service ./internal/handler/...
  • go vet -tags=unit ./internal/domain ./internal/pkg/antigravity/... ./internal/pkg/gemini/... ./internal/pkg/geminicli/... ./internal/service ./internal/handler/...
  • golangci-lint run ./internal/domain ./internal/pkg/antigravity ./internal/pkg/gemini ./internal/pkg/geminicli ./internal/service ./internal/handler
  • vitest run src/composables/__tests__/useModelWhitelist.spec.ts
  • vue-tsc --noEmit & eslint src/composables/useModelWhitelist.ts src/composables/__tests__/useModelWhitelist.spec.ts

@github-actions

github-actions Bot commented Nov 14, 2026 •

Copy link
Copy Markdown
Contributor

All contributors have signed the CLA. ✅
Posted by the CLA Assistant Lite bot.

@universonic

Copy link
Copy Markdown
Author

I have read the CLA Document and I hereby sign the CLA

github-actions Bot added a commit that referenced this pull request Nov 14, 2026
doublepi123 added a commit to doublepi123/sub2api that referenced this pull request Nov 14, 2026
Port upstream Wei-Shaw#8778 (which also carries the Wei-Shaw#8746 Hydrogen
billing) on top of the fork's Carbon adaptation.

Hydrogen:
- Antigravity lists Hydrogen only as gemini-4-hydrogen-low/-medium/-high/
  -xhigh (production account 8). Default mapping adds the four suffixes and
  a bare self-mapped alias that resolves to a real tier; static catalog,
  Gemini API / CLI catalogs, frontend whitelist and presets follow.
- xhigh/max select -xhigh where the account has it and fall back to -high
  for families without it; native thinkingBudget is normalized to
  thinkingLevel for Hydrogen as for Carbon.
- Pricing: $1.50 / $7.50 / $0.075 per MTok, priority 1.8x.

Claude 5.5 on Antigravity:
- claude-opus-5-5 / claude-sonnet-5-5 default mappings become dynamic
  self-maps routed by effort (default stays -high); explicit fixed admin
  mappings still win; native Anthropic accounts are untouched.
- Scheduling admits bare names on tier-only accounts; model cooldowns and
  billing tiers follow the variant actually forwarded.

Fork semantics kept over the PR:
- Carbon keeps the bare self-map plus measured -low/-medium/-high suffixes
  and defaults to medium when the client pins no level.
- Variant selection keeps the fork's body-based parser where an explicit
  thinking budget beats effort; PR tests were adapted to that contract.
@universonic

Copy link
Copy Markdown
Author

更新说明:同步 upstream/main 并补充交叉适配

已合并最新 upstream/main(d70676756 -> 7af2c717f)解决与 #8673 的文件冲突,并追加提交 b6fd4fac9 完成与上游新合入机制(#8673、#8739)的语义对齐及边界修复:

  1. Antigravity 兼容层思考配置合并对齐(fix(antigravity): stream Gemini compat thoughts and anchor pre-content timeout after headers #8673)

    • 变体前置解析保留支持 xhigh 且在未显式传 effort 时屏蔽合成默认值的 geminiThinkingLevelFromClaudeRequest。
    • applyGeminiCompatThinkingConfig 的 includeThoughts 同时覆盖流式请求(stream=true,防止长思考阶段空闲超时并流式输出推理摘要)与显式启用思考且非 none/minimal 的非流式请求(恢复非流式 reasoning_content / reasoning 项输出,同时保持未传 effort 或 effort=none 的非流式请求不强制请求摘要)。
  2. 兼容入口分组 Effort 策略与模型冷却预检对齐(fix(gateway): enforce effort policies on Anthropic compatibility routes #8739)

    • 新增 WithAntigravityThinkingLevelFromOpenAIBody,在 /v1/chat/completions 与 /v1/responses 完成分组 effort 策略改写后,将最终有效档位写入 Context,使混合分组调度时的模型级冷却预检与实际转发变体一致。
  3. 策略改写前的 RequestedReasoningEffort 原始归因保留(fix(gateway): enforce effort policies on Anthropic compatibility routes #8739)

    • AntigravityGatewayService 兼容入口优先读取策略改写前写入 Context 的原始请求档位(无 Context 时回退提取原始请求体 canonical 值,保留 max 等原值),避免被分组策略改写或模型档位归一化覆盖。
  4. 纯 web_search 降级模型的有效档位清零

    • 当请求仅携带 web_search 工具被降级至仅支持预算的 gemini-2.5-flash 时,effectiveAntigravityGeminiReasoningEffort 明确裁决有效分档为 nil,阻止回退逻辑恢复降级前的分档 ReasoningEffort。
  5. 回归测试

    • 新增 TestGatewayCompatibilityHandlersAntigravityEffortPolicySchedulingAndAttribution 覆盖 /v1/chat/completions 与 /v1/responses 在策略改写下的变体冷却调度与原始/生效档位归因,并补充非流式显式思考摘要与 web_search 降级单测。

@universonic

Copy link
Copy Markdown
Author

@Wei-Shaw 这些改动对提升 Gemini 用户使用体验有关键帮助,请求合并,谢谢。

doublepi123 added a commit to doublepi123/sub2api that referenced this pull request Nov 17, 2026
Sync 185 upstream commits (106 non-merge), including the Gemini 4
Hydrogen billing PR Wei-Shaw#8746, Vertex multi-region endpoints, OpenAI Ultra
Fast group gating, scheduler bucket-merged snapshot loading and the
Gemini native content session hash.

Conflict resolutions that keep fork behavior:
- VERSION and the fork's single-job release workflow are kept.
- Hydrogen billing keeps the fork's tier-alias match; Wei-Shaw#8746's card and
  snapshot entry were already present from the Wei-Shaw#8778 port, so nothing is
  duplicated.
- Antigravity compat variant selection keeps the fork's body-based level
  (explicit thinking budget beats effort); upstream's now-unused
  geminiThinkingLevelFromClaudeBody helper is dropped. includeThoughts
  follows upstream (enabled on streams) together with its new tests.
- Chat Completions / Responses to Anthropic keep the fork's Kiro branch and
  gain upstream's one-shot retry after a concurrent token rotation.
- Scheduler credential projection keeps the Kiro catalog and Gemini
  explicit-cache keys and adds openai_capabilities.
- Gemini native sticky sessions take upstream's content hash, which keeps
  the fork's API-key isolation and stable first-turn anchor.

Fork follow-ups in this merge:
- 252_fork_group_usage_rollup_published_timezone.sql restores the
  published-timezone invalidation from fork migration 228, which upstream
  251 replaced with the session timezone; verified on Postgres 18 with a
  UTC session against Asia/Shanghai buckets.
- The Kiro snapshot test stub implements ListModelAvailabilityCandidates,
  now used by the database fallback loader.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants