GitLab

Add post-sort shuffle for accounts with identical (priority, loadRate,
lastUsedAt) to break deterministic ordering when concurrent requests
read the same scheduler snapshot. Applies to both Antigravity and
OpenAI scheduling paths, plus the sortAccountsByPriorityAndLastUsed
helper.

Keeps upstream CallCount/ModelLoadInfo scheduling intact; shuffle is
additive and only randomises within equivalent-rank groups.

Without this routing guard, ForwardUpstream is never called because
Forward() always proceeds with the standard OAuth/cookie flow.

- In handleSmartRetry, use the actual upstream retryDelay to set model
  rate limit duration instead of always using the 30s default
- Return info.RetryDelay from shouldTriggerAntigravitySmartRetry when
  shouldRateLimitModel=true, so callers know the actual delay
- Extract getDefaultRateLimitDuration() and resolveResetTime() helpers
  to reduce duplication in handleUpstreamError 429 handling
- Improve debug logging with upstream_retry_delay and response body

ParseGatewayRequest only parsed Anthropic format (system/messages),
ignoring Gemini native format (systemInstruction/contents). This caused
GenerateSessionHash to produce identical hashes for all Gemini sessions.

Add protocol parameter to ParseGatewayRequest to branch between
Anthropic and Gemini parsing. Update GenerateSessionHash message
traversal to extract text from both formats.

Mix SessionContext (ClientIP, UserAgent, APIKeyID) into
GenerateSessionHash 3rd-level fallback to differentiate requests
from different users sending identical content.

Also switch hashContent from SHA256-truncated to XXHash64 for
better performance, and optimize Trie Lua script to match from
longest prefix first.

Previously, thoughtSignature cleanup only applied to Gemini CLI
requests (detected via x-gemini-api-privileged-user-id header or
tmp dir pattern). This caused 400 errors for non-CLI clients when
session cache expired and they sent stale signatures.

Remove the isGeminiCLIRequest guard so all clients benefit from
proactive thoughtSignature cleanup on session binding miss.

- Add `sort_order` field to groups table with migration
- Add `PUT /api/v1/admin/groups/sort-order` API for batch update
- Implement drag-and-drop UI using vue-draggable-plus
- All queries now order groups by sort_order
- Add i18n support (en/zh) for sort-related UI text
- Update test stubs to satisfy new interface methods

优化 /admin/users 页面的并发数列，显示「当前/最大」格式，
参考 AccountCapacityCell 的设计风格。

- 后端 UserHandler 注入 ConcurrencyService，批量查询用户当前并发数
- 新增 UserConcurrencyCell 组件，支持颜色状态（空闲灰/使用中黄/满载红）
- 前端 AdminUser 类型添加 current_concurrency 字段

Upstream accounts now use the standard APIKey type instead of a dedicated
upstream type. GetBaseURL() and new GetGeminiBaseURL() automatically append
/antigravity for Antigravity platform APIKey accounts, eliminating the need
for separate upstream forwarding methods.

- Remove ForwardUpstream, ForwardUpstreamGemini, testUpstreamConnection
- Remove upstream branch guards in Forward/ForwardGemini/TestConnection
- Add migration 052 to convert existing upstream accounts to apikey
- Update frontend CreateAccountModal to create apikey type
- Add unit tests for GetBaseURL and GetGeminiBaseURL

ForwardUpstream/ForwardUpstreamGemini should pipe the upstream response
directly to the client (headers + body), not parse it as SSE stream.

Replace manual header setting (Content-Type, anthropic-version, anthropic-beta)
with full client header passthrough in ForwardUpstream/ForwardUpstreamGemini.
Only authentication headers (Authorization, x-api-key) are overridden with
upstream account credentials. Hop-by-hop headers are excluded.

Add unit tests covering header passthrough, auth override, and hop-by-hop filtering.

v0.1.74 merged upstream accounts into the OAuth path, causing requests
to hit the wrong protocol and endpoint. Add three upstream-specific
methods (testUpstreamConnection, ForwardUpstream, ForwardUpstreamGemini)
that use base_url + apiKey auth and passthrough the original body, while
reusing the existing response handling and error/retry logic.

- Change antigravitySmartRetryMaxAttempts from 3 to 1 to prevent
  repeated rate limiting and long waits
- Clear sticky session binding (DeleteSessionAccountID) after smart
  retry exhaustion, so subsequent requests don't hit the same
  rate-limited account
- Add flow diagrams to Forward/ForwardGemini doc comments
- Add comprehensive unit tests covering:
  - Sticky session cleared on retry failure (429, 503, network error)
  - Sticky session NOT cleared on retry success
  - Sticky session NOT cleared for non-sticky requests (empty hash)
  - Sticky session NOT cleared on long delay path (handled by handler)
  - Nil cache safety (no panic)
  - MaxAttempts constant verification
  - End-to-end retryLoop → switchError propagation with session clear

The digest chain fallback is only needed for Gemini endpoints, not
for the Anthropic Messages API path. Remove the handler integration
while keeping the reusable service/repository layer for future use.

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.

- avoid panic by using safe UUID prefix truncation in Gemini digest fallback logs\n- remove unconditional Antigravity 429 full-body debug logs and honor log truncation config\n- align Antigravity quick preset mappings to opus 4.6-thinking targets only\n- restore scope rate-limit aggregation/output in ops availability stats

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.

Remove extra blank line that caused golangci-lint gofmt check to fail.

- Add GetAccessToken upstream branch tests (success/failure/empty/nil)
- Add mapAntigravityModel wildcard-target-equals-request edge case tests
- Add upstream account smart retry test case
- Add GeminiMessagesCompatService custom model_mapping and empty model tests

- 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

Key changes:
- Upgrade model mapping: Opus 4.5 → Opus 4.6-thinking with precise matching
- Unified rate limiting: scope-level → model-level with Redis snapshot sync
- Load-balanced scheduling by call count with smart retry mechanism
- Force cache billing support
- Model identity injection in prompts with leak prevention
- Thinking mode auto-handling (max_tokens/budget_tokens fix)
- Frontend: whitelist mode toggle, model mapping validation, status indicators
- Gemini session fallback with Redis Trie O(L) matching
- Ops: enhanced concurrency monitoring, account availability, retry logic
- Migration scripts: 049-051 for model mapping unification

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).

When extended thinking is enabled, Claude API requires max_tokens >
thinking.budget_tokens. If misconfigured, this auto-adjusts max_tokens
to budget_tokens + 1000 instead of returning a 400 error.

- Add ensureMaxTokensGreaterThanBudget helper function
- Extract Gemini25FlashThinkingBudgetLimit constant (24576)
- Log adjustment for debugging