fix: bound uncached remainder by prompt-max(cached,write) and forward compact prompt_cache_key

relay/channel/openai/relay_image.go

@@ -80,6 +80,7 @@ func normalizeOpenAIUsage(usage *dto.Usage) {
 	if usage.InputTokensDetails != nil {
 		usage.PromptTokensDetails.CachedTokens = usage.InputTokensDetails.CachedTokens
 		usage.PromptTokensDetails.CachedCreationTokens = usage.InputTokensDetails.CachedCreationTokens
+		usage.PromptTokensDetails.CacheWriteTokens = usage.InputTokensDetails.CacheWriteTokens
 		usage.PromptTokensDetails.ImageTokens = usage.InputTokensDetails.ImageTokens
 		usage.PromptTokensDetails.TextTokens = usage.InputTokensDetails.TextTokens
 		usage.PromptTokensDetails.AudioTokens = usage.InputTokensDetails.AudioTokens

relay/responses_handler.go

@@ -40,6 +40,11 @@ func ResponsesHelper(c *gin.Context, info *relaycommon.RelayInfo) (newAPIError *
 	case *dto.OpenAIResponsesRequest:
 		responsesReq = req
 	case *dto.OpenAIResponsesCompactionRequest:
+		// Only fields documented for POST /v1/responses/compact are forwarded:
+		// model, input, instructions, previous_response_id, prompt_cache_key,
+		// prompt_cache_options, prompt_cache_retention, service_tier.
+		// Undocumented Codex-parity fields (tools, reasoning, text) are parsed
+		// for client compatibility but intentionally not sent upstream.
 		responsesReq = &dto.OpenAIResponsesRequest{
 			Model:                req.Model,
 			Input:                req.Input,
@@ -47,6 +52,7 @@ func ResponsesHelper(c *gin.Context, info *relaycommon.RelayInfo) (newAPIError *
 			PreviousResponseID:   req.PreviousResponseID,
 			ParallelToolCalls:    req.ParallelToolCalls,
 			ServiceTier:          req.ServiceTier,
+			PromptCacheKey:       req.PromptCacheKey,
 			PromptCacheOptions:   req.PromptCacheOptions,
 			PromptCacheRetention: req.PromptCacheRetention,
 		}