perf: align with readme-magno optimizations (timestamp, cache logging)

- Change system prompt timestamp to daily-only format (was HH:MM,
  invalidating dynamic block every minute unnecessarily)
- Add CacheCreatedTokens and CacheReadTokens to UsageInfo
- Extract Anthropic cache_creation/cache_read tokens in parseResponse
- Log cache hit/miss stats per session for cost monitoring

Made-with: Cursor

pkg/agent/context.go

@@ -367,7 +367,7 @@ func (cb *ContextBuilder) LoadBootstrapFiles() string {
 // See: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
 // See: https://platform.openai.com/docs/guides/prompt-caching
 func (cb *ContextBuilder) buildDynamicContext(channel, chatID string) string {
-	now := time.Now().Format("2006-01-02 15:04 (Monday)")
+	now := time.Now().Format("2006-01-02 (Monday)")
 	rt := fmt.Sprintf("%s %s, Go %s", runtime.GOOS, runtime.GOARCH, runtime.Version())
 
 	var sb strings.Builder

pkg/agent/loop.go

@@ -661,7 +661,7 @@ func (al *AgentLoop) runLLMIteration(
 ) (string, int, error) {
 	iteration := 0
 	var finalContent string
-	var totalInputTokens, totalOutputTokens int
+	var totalInputTokens, totalOutputTokens, totalCacheCreated, totalCacheRead int
 
 	for iteration < agent.MaxIterations {
 		iteration++
@@ -809,6 +809,8 @@ func (al *AgentLoop) runLLMIteration(
 		if response.Usage != nil {
 			totalInputTokens += response.Usage.PromptTokens
 			totalOutputTokens += response.Usage.CompletionTokens
+			totalCacheCreated += response.Usage.CacheCreatedTokens
+			totalCacheRead += response.Usage.CacheReadTokens
 		}
 
 		logger.DebugCF("agent", "LLM response",
@@ -993,6 +995,8 @@ func (al *AgentLoop) runLLMIteration(
 				"input_tokens":   totalInputTokens,
 				"output_tokens":  totalOutputTokens,
 				"total_tokens":   totalInputTokens + totalOutputTokens,
+				"cache_created":  totalCacheCreated,
+				"cache_read":     totalCacheRead,
 			})
 	}
 

pkg/providers/anthropic/provider.go

@@ -306,6 +306,8 @@ func parseResponse(resp *anthropic.Message) *LLMResponse {
 			PromptTokens:       int(resp.Usage.InputTokens),
 			CompletionTokens:   int(resp.Usage.OutputTokens),
 			TotalTokens:        int(resp.Usage.InputTokens + resp.Usage.OutputTokens),
+			CacheCreatedTokens: int(resp.Usage.CacheCreationInputTokens),
+			CacheReadTokens:    int(resp.Usage.CacheReadInputTokens),
 		},
 	}
 }

pkg/providers/protocoltypes/types.go

@@ -45,6 +45,8 @@ type UsageInfo struct {
 	PromptTokens       int `json:"prompt_tokens"`
 	CompletionTokens   int `json:"completion_tokens"`
 	TotalTokens        int `json:"total_tokens"`
+	CacheCreatedTokens int `json:"cache_created_tokens,omitempty"`
+	CacheReadTokens    int `json:"cache_read_tokens,omitempty"`
 }
 
 // CacheControl marks a content block for LLM-side prefix caching.