The previous fallback (step 3) in GenerateSessionHash hashed system +
all messages together, producing a different hash each round as the
conversation grew ([a] -> [a,b] -> [a,b,c]). This made fallback sticky
sessions ineffective for multi-turn conversations.
Implement per-message Trie digest chain matching (reusing Gemini's Trie
infrastructure) so that the previous round's chain is always a prefix
of the current round's chain, enabling reliable session affinity.
Remove threshold-based waiting in both sticky session and antigravity
pre-check paths. When a model is rate-limited, immediately clear the
sticky session and switch accounts instead of waiting for short durations.
1. Frontend: replace hardcoded antigravityDefaultMappings with async
fetch from GET /admin/accounts/antigravity/default-model-mapping,
eliminating the duplicate data source that caused frontend/backend
mapping inconsistency.
2. Backend: convert handleSmartRetry and antigravityRetryLoop from
standalone functions to AntigravityGatewayService methods, enabling
Redis cache sync (updateAccountModelRateLimitInCache) after both
rate-limit write paths — long-delay branch and retry-exhausted branch.
- GetAccessToken: add upstream branch to read api_key from credentials
- shouldTriggerAntigravitySmartRetry: relax check from IsOAuth to Platform-based
- isModelSupportedByAccount/WithContext: replace IsAntigravityModelSupported
whitelist with mapAntigravityModel for unified scheduling/forwarding logic
- mapAntigravityModel: fix edge case where wildcard target equals request model
- Update tests for new behavior and add custom model_mapping test cases
The default fallback cooldown when rate limit reset time cannot be
parsed was 5 minutes, which is too aggressive and causes accounts
to be unnecessarily locked out. Reduce to 30 seconds for faster
recovery. Config override still works (unit remains minutes).
Kimi 等 Claude 兼容 API 返回缓存信息使用 OpenAI 风格的 cached_tokens 字段，
而非 Claude 标准的 cache_read_input_tokens，导致客户端收不到缓存命中信息且
内部计费缓存折扣为 0。
新增 reconcileCachedTokens 辅助函数，在 cache_read_input_tokens == 0 且
cached_tokens > 0 时自动填充，覆盖流式（message_start/message_delta）和
非流式两种响应路径。对 Claude 原生上游无影响。
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Kimi 等 Claude 兼容 API 返回缓存信息使用 OpenAI 风格的 cached_tokens 字段，
而非 Claude 标准的 cache_read_input_tokens，导致客户端收不到缓存命中信息且
内部计费缓存折扣为 0。
新增 reconcileCachedTokens 辅助函数，在 cache_read_input_tokens == 0 且
cached_tokens > 0 时自动填充，覆盖流式（message_start/message_delta）和
非流式两种响应路径。对 Claude 原生上游无影响。
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Previously the /v1/usage endpoint aggregated usage stats (today/total
tokens, cost, RPM/TPM) across all API Keys belonging to the user.
This made it impossible to distinguish usage from different API Keys
(e.g. balance vs subscription keys).
Now the usage stats are filtered by the current request's API Key ID,
so each key only sees its own usage data. The balance/remaining fields
are unaffected and still reflect the user-level wallet balance.
Changes:
- Add GetAPIKeyDashboardStats to repository interface and implementation
- Add getPerformanceStatsByAPIKey helper (also fixes TPM to include
cache_creation_tokens and cache_read_tokens)
- Add GetAPIKeyDashboardStats to UsageService
- Update Usage handler to call GetAPIKeyDashboardStats(apiKey.ID)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
