lbjlaq/Antigravity-Manager v5.0.1-beta.6

Antigravity Tools v5.0.1-beta.6

on GitHub

• [Inbound Model Gate & Punishment-Chain Exemptions; Fix 429/503 Collateral Damage to the Account Pool] (Ref #3764, Ref #3769, Thanks to @jeikl):
  • Four-Stage Model Gate: Missing model → 404 model not exist; custom mappings take precedence (exact / case / normalized / wildcard) and the raw ID is preserved otherwise; the mapped ID must exist in the union of downstream-exposed models, the agent tool list, the official catalog and mapping values; if no account holds positive quota → 502 no quota accounts. Fails open when the pool has no model metadata yet, so a cold start cannot wipe out all traffic. (Ref #3764)
  • All Entry Points Covered: Beyond POST generate / messages / chat (including Responses), the gate now also guards :countTokens colon syntax, /v1beta/models/:model/countTokens, and all four protocol GET probe endpoints: /v1/models/:model, /v1/model/:model, /v1/responses/models/:model, /responses/models/:model, /v1/models/claude/:model, /v1beta/models/:model. (Ref #3764)
  • Punishment Chain Fully Exempt: is_gateway_self_answer is the single source of truth and is wired into policy::classify (highest priority → InternalGatewayMessage), rate-limit parsing, mark_rate_limited(_fast/_async), classify_rate_limit_reason and mark_account_transient_failure. Gateway-generated 404/502 no longer lock accounts, unbind sticky sessions, or write zero-quota markers, eliminating collateral damage to innocent accounts. (Ref #3764)
  • Legacy Model Families Retired: Claude 3.x / < 4.6 and Gemini 2.x are hard-blocked ahead of custom mappings; the gemini-2.5-* → 3.1-flash-lite internal alias, legacy entries in the OpenCode static catalog and VariantType::Gemini25Thinking were removed. Image-generation families are exempt and accept -preview / -1k / -2k / -4k / aspect-ratio variants plus version-synonym normalization. (Ref #3769)
  • Hardcoded Fallback Models Removed: All three INTERNAL_BACKGROUND_TASK constants were deleted; background summary models are now derived from the live pool (newest flash-low → bare flash → skip). Scheduler warm-up fallbacks are pool-driven and skip when no flash model exists instead of falling back to a hardcoded ID. (Ref #3769)
  • Image Model Visibility & Quota Consistency: The model list only registers a normalized base image model when an account holds positive real quota (covering /v1/models, /v1beta/models and the agent one-click config), and the quota gate counts same-family variant quota so a listed model never answers with 502. (Ref #3769)
  • Frontend Observability: The traffic-log account column highlights model not exist (rose) and no quota accounts (amber) with high contrast in both light and dark themes, while the forwarded-payload panel shows an English remediation guide (detail + hint). (Ref #3764)
  • QoS Ladder Probe Fixed: The probe now sends a complete official v1internal envelope (project / requestId / request / model / userAgent), picks a Flash-tier model from the account's real metadata, and falls back Daily → Sandbox. Measured 200 upstream (TTFT ≈ 1.0–1.9s), so jailed accounts can self-heal back to Healthy. (Ref #3769)

• [新增入站模型门禁与惩罚链豁免，修复 429/503 误伤号池] (Ref #3764, Ref #3769, Thanks to @jeikl):
  • 模型门禁四道判定: 缺失 model → 404 model not exist；自定义映射优先（精确 / 大小写 / 规范化 / 通配符），无映射保留原值；映射后必须存在于「下游开放列表 ∪ agent 工具列表 ∪ 官方目录 ∪ 映射值」；全池无正配额 → 502 no quota accounts。冷启动无模型元数据时 fail-open，避免启动窗口误杀全部流量。(Ref #3764)
  • 全入口覆盖: 在 POST generate / messages / chat（含 Responses）之外，补齐 :countTokens 冒号语法与 /v1beta/models/:model/countTokens，并为四大协议 GET 探测端点接入门禁：/v1/models/:model、/v1/model/:model、/v1/responses/models/:model、/responses/models/:model、/v1/models/claude/:model、/v1beta/models/:model。(Ref #3764)
  • 惩罚链 100% 豁免: is_gateway_self_answer 作为单一真理源，接入 policy::classify（最高优先级归为 InternalGatewayMessage）、限流解析、mark_rate_limited(_fast/_async)、classify_rate_limit_reason 与 mark_account_transient_failure；网关代答的 404/502 不再锁号、不再解绑粘性会话、不再写入零配额标记，彻底杜绝无辜账号连坐。(Ref #3764)
  • 旧版模型全面下线: Claude 3.x / < 4.6 与 Gemini 2.x 硬门禁（优先于自定义映射），删除 gemini-2.5-* → 3.1-flash-lite 内部别名、opencode 静态目录 legacy 条目与 VariantType::Gemini25Thinking；生图模型族独立豁免，并支持 -preview / -1k / -2k / -4k / 比例 变体与版本同义名归一。(Ref #3769)
  • 去除兜底模型硬编码: 删除三处 INTERNAL_BACKGROUND_TASK 常量，后台摘要模型改为号池动态推导（最新 flash-low → 无后缀 flash → 跳过）；调度器预热兜底同步改为号池驱动，无可用 flash 时跳过而非硬编码回退。(Ref #3769)
  • 生图可见性与配额一致性: 列表侧仅在账号真实配额 > 0 时，才把尺寸变体归一后的基底生图模型登记进 /v1/models、/v1beta/models 与 agent 一键配置；配额闸门同步折算同族变体配额，避免「列表有、请求 502」的不一致。(Ref #3769)
  • 前端可观测性: 流量日志账号列高亮展示 model not exist（玫红）/ no quota accounts（琥珀），亮色暗色均保持高对比；转出报文面板展示英文排查指引（detail + hint）。(Ref #3764)
  • QoS 阶梯探针修复: 补齐官方 v1internal 信封（project / requestId / request / model / userAgent），探针模型改为从账号真实元数据挑选 Flash 档，端点走 Daily → Sandbox 灾备；实测上游返回 200（TTFT ≈ 1.0–1.9s），小黑屋账号可正常自愈回到 Healthy。(Ref #3769)

Full Changelog: v5.0.1-beta.5...v5.0.1-beta.6