merge: adopt main cache billing fixes, keep refactor/ui sidebar badge styling

dto/openai_response.go

@@ -270,12 +270,17 @@ type InputTokenDetails struct {
 // which field the upstream reported it in: Claude-derived conversions populate
 // CachedCreationTokens while OpenAI reports cache_write_tokens natively. Both
 // are billed at the cache-creation price; when both are present the larger
-// value wins so the same tokens are never double-counted.
+// value wins so the same tokens are never double-counted. Negative upstream
+// values are clamped to zero so they can never lower a charge.
 func (d InputTokenDetails) CacheCreationTokensTotal() int {
-	if d.CacheWriteTokens > d.CachedCreationTokens {
-		return d.CacheWriteTokens
+	total := d.CachedCreationTokens
+	if d.CacheWriteTokens > total {
+		total = d.CacheWriteTokens
 	}
-	return d.CachedCreationTokens
+	if total < 0 {
+		return 0
+	}
+	return total
 }
 
 type OutputTokenDetails struct {

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,
 		}

service/relayconvert/internal/oai_chat/to_claude_messages_resp.go

@@ -44,7 +44,8 @@ func buildClaudeUsageFromOpenAIUsage(oaiUsage *dto.Usage) *dto.ClaudeUsage {
 	if oaiUsage.PromptTokensDetails.CacheWriteTokens > 0 {
 		// OpenAI native cache-write usage counts cached and cache-write tokens
 		// inside prompt_tokens, while Claude semantics reports input_tokens
-		// excluding both; the uncached remainder clamps at zero.
+		// excluding both. Both counts are unadjusted prefixes and may overlap,
+		// so clamp a negative remainder at zero.
 		inputTokens = oaiUsage.PromptTokens - oaiUsage.PromptTokensDetails.CachedTokens - cacheCreationTokens
 		if inputTokens < 0 {
 			inputTokens = 0

service/relayconvert/internal/oai_chat/to_claude_messages_resp_test.go

@@ -92,7 +92,7 @@ func TestBuildClaudeUsageFromOpenAICacheWriteUsage(t *testing.T) {
 
 	require.NotNil(t, usage)
 	// Claude semantics reports input_tokens excluding cache read/write; the
-	// remainder 3619-2921-3616 clamps to 0.
+	// overlapping unadjusted prefixes drive the remainder negative, clamp to 0.
 	assert.Equal(t, 0, usage.InputTokens)
 	assert.Equal(t, 2921, usage.CacheReadInputTokens)
 	assert.Equal(t, 3616, usage.CacheCreationInputTokens)

service/text_quota.go

@@ -294,9 +294,10 @@ func calculateTextQuotaSummary(ctx *gin.Context, relayInfo *relaycommon.RelayInf
 			}
 		}
 
-		// OpenAI cache-write usage can report cached_tokens + cache_write_tokens
-		// exceeding prompt_tokens; the uncached remainder must clamp at zero so
-		// billing never subtracts more than the reported input.
+		// OpenAI cache-write usage reports unadjusted prefix counts, so
+		// cached_tokens + cache_write_tokens can exceed prompt_tokens and the
+		// remainder can go negative. Clamp at zero so overlap never turns into
+		// a negative base charge.
 		if baseTokens.IsNegative() {
 			baseTokens = decimal.Zero
 		}

service/text_quota_test.go

@@ -411,8 +411,8 @@ func TestCalculateTextQuotaSummaryBillsOpenAICacheWriteTokens(t *testing.T) {
 
 	t.Run("uncached remainder clamps to zero", func(t *testing.T) {
 		// Real OpenAI payload shape: cached_tokens + cache_write_tokens exceeds
-		// prompt_tokens, so the uncached remainder must clamp to 0 instead of
-		// producing a negative charge component.
+		// prompt_tokens because both are unadjusted prefix counts. The negative
+		// remainder must clamp to zero, never turn into a negative base charge.
 		usage := &dto.Usage{
 			PromptTokens:     3619,
 			CompletionTokens: 36,

service/tiered_settle.go

@@ -67,6 +67,8 @@ func BuildTieredTokenParams(usage *dto.Usage, isClaudeUsageSemantic bool, usedVa
 		}
 	}
 
+	// OpenAI cache-write usage reports unadjusted prefix counts, so cr + cc can
+	// exceed the prompt and drive the remainder negative. Clamp at zero.
 	if p < 0 {
 		p = 0
 	}