diff --git a/pkg/agent/pipeline_llm.go b/pkg/agent/pipeline_llm.go
index b4c87634ed..a56318a3d5 100644
--- a/pkg/agent/pipeline_llm.go
+++ b/pkg/agent/pipeline_llm.go
@@ -564,6 +564,15 @@ func (p *Pipeline) CallLLM(
 		llmResponseFields["prompt_tokens"] = exec.response.Usage.PromptTokens
 		llmResponseFields["completion_tokens"] = exec.response.Usage.CompletionTokens
 		llmResponseFields["total_tokens"] = exec.response.Usage.TotalTokens
+		if exec.response.Usage.PromptCacheHitTokens != nil {
+			llmResponseFields["prompt_cache_hit_tokens"] = *exec.response.Usage.PromptCacheHitTokens
+		}
+		if exec.response.Usage.PromptCacheMissTokens != nil {
+			llmResponseFields["prompt_cache_miss_tokens"] = *exec.response.Usage.PromptCacheMissTokens
+		}
+		if exec.response.Usage.PromptTokensDetails != nil {
+			llmResponseFields["cached_tokens"] = exec.response.Usage.PromptTokensDetails.CachedTokens
+		}
 	}
 	logger.DebugCF("agent", "LLM response", llmResponseFields)
 
diff --git a/pkg/providers/protocoltypes/types.go b/pkg/providers/protocoltypes/types.go
index b60c467014..03501cc40e 100644
--- a/pkg/providers/protocoltypes/types.go
+++ b/pkg/providers/protocoltypes/types.go
@@ -53,6 +53,19 @@ type UsageInfo struct {
 	PromptTokens     int `json:"prompt_tokens"`
 	CompletionTokens int `json:"completion_tokens"`
 	TotalTokens      int `json:"total_tokens"`
+
+	// Cache metadata (optional). DeepSeek reports prompt_cache_hit_tokens /
+	// prompt_cache_miss_tokens; OpenAI-compatible endpoints may report
+	// prompt_tokens_details.cached_tokens. Pointers distinguish "not reported"
+	// (nil) from a real zero.
+	PromptCacheHitTokens  *int                 `json:"prompt_cache_hit_tokens"`
+	PromptCacheMissTokens *int                 `json:"prompt_cache_miss_tokens"`
+	PromptTokensDetails   *PromptTokensDetails `json:"prompt_tokens_details"`
+}
+
+// PromptTokensDetails carries OpenAI-style prompt token breakdowns.
+type PromptTokensDetails struct {
+	CachedTokens int `json:"cached_tokens"`
 }
 
 // CacheControl marks a content block for LLM-side prefix caching.