fix(gateway): 修复 Qwen3.8 推理参数并增加上游自愈

统一 Qwen3.8 固定思考、推理档位、预算互斥与温度下限规则,并覆盖 Chat Completions 和原生 Responses 请求。

新增安全参数纠正、有限重试和并发安全的进程内 LRU 学习;补充共享行为向量、httptest 回归与真实模型验收用例。

验证:go test ./... -count=1 通过,gofmt 检查通过;真实 Qwen3.8 流式、非流式及缓存重学习验收通过。
This commit is contained in:
2026-08-01 22:50:45 +08:00
parent 50a13de70b
commit 9c093974a1
9 changed files with 1547 additions and 33 deletions
@@ -32,7 +32,7 @@ var openAIResponsesRequestParameters = stringSet(
var gatewayOpenAIRequestExtensions = stringSet(
"runMode", "run_mode", "conversationId", "conversation_id", "sessionId", "session_id",
"requestId", "request_id", "signal", "userMessage", "user_message", "platformId",
"platform_id", "options", "enable_thinking", "thinking_budget_tokens", "enable_web_search",
"platform_id", "options", "enable_thinking", "thinking_budget", "thinking_budget_tokens", "enable_web_search",
"modelType", "model_type", "capability", "capabilityType", "mode", "simulation", "testMode",
"cacheAffinityKey", "cache_affinity_key", "simulationDurationMs", "simulationDurationSeconds",
"simulationMinDurationMs", "simulationMaxDurationMs", "simulationMinDurationSeconds",