chore: set cache control markers on tools

intercept_anthropic_messages_base.go

@@ -44,6 +44,16 @@ func (i *AnthropicMessagesInterceptionBase) injectTools() {
 		return
 	}
 
+	// Any existing tool definitions.
+	for _, tool := range i.req.Tools {
+		if tool.OfTool == nil {
+			continue
+		}
+
+		// Explicitly unset all cache control settings, we'll set one at the end.
+		tool.OfTool.CacheControl = anthropic.CacheControlEphemeralParam{}
+	}
+
 	// Inject tools.
 	for _, tool := range i.mcpProxy.ListTools() {
 		i.req.Tools = append(i.req.Tools, anthropic.ToolUnionParam{
@@ -55,10 +65,18 @@ func (i *AnthropicMessagesInterceptionBase) injectTools() {
 				Name:        tool.ID,
 				Description: anthropic.String(tool.Description),
 				Type:        anthropic.ToolTypeCustom,
+				// Explicitly unset all cache control settings, we'll set one at the end.
+				CacheControl: anthropic.CacheControlEphemeralParam{},
 			},
 		})
 	}
 
+	// See https://docs.claude.com/en/docs/build-with-claude/prompt-caching.
+	// "The cache_control parameter on the last tool definition caches all tool definitions."
+	if count := len(i.req.Tools); count > 0 {
+		i.req.Tools[count-1].OfTool.CacheControl = anthropic.NewCacheControlEphemeralParam()
+	}
+
 	// Note: Parallel tool calls are disabled to avoid tool_use/tool_result block mismatches.
 	i.req.ToolChoice = anthropic.ToolChoiceUnionParam{
 		OfAny: &anthropic.ToolChoiceAnyParam{
@@ -68,6 +86,28 @@ func (i *AnthropicMessagesInterceptionBase) injectTools() {
 	}
 }
 
+// removeUnnecessaryCacheMarkers removes any cache control settings which are unnecessarily set by the client
+// and/or may interfere with the cache control we need to implement ourselves.
+func (i *AnthropicMessagesInterceptionBase) removeUnnecessaryCacheMarkers() {
+	if i.req == nil {
+		return
+	}
+
+	// Explicitly unset any cache control markers on "assistant" messages; these should never be set
+	// since it's more beneficial for us to cache tool definitions, and Anthropic only allows for 4
+	// cache markers...
+	// https://docs.claude.com/en/docs/build-with-claude/prompt-caching#when-to-use-multiple-breakpoints
+	for _, msg := range i.req.Messages {
+		if msg.Role == anthropic.MessageParamRoleAssistant {
+			for _, c := range msg.Content {
+				if c.OfText != nil {
+					c.OfText.CacheControl = anthropic.CacheControlEphemeralParam{}
+				}
+			}
+		}
+	}
+}
+
 // isSmallFastModel checks if the model is a small/fast model (Haiku 3.5).
 // These models are optimized for tasks like code autocomplete and other small, quick operations.
 // See `ANTHROPIC_SMALL_FAST_MODEL`: https://docs.anthropic.com/en/docs/claude-code/settings#environment-variables

intercept_anthropic_messages_blocking.go

@@ -41,6 +41,7 @@ func (i *AnthropicMessagesBlockingInterception) ProcessRequest(w http.ResponseWr
 
 	ctx := r.Context()
 
+	i.removeUnnecessaryCacheMarkers()
 	i.injectTools()
 
 	var (

intercept_anthropic_messages_streaming.go

@@ -80,6 +80,7 @@ func (i *AnthropicMessagesStreamingInterception) ProcessRequest(w http.ResponseW
 			logger.Warn(ctx, "failed to determine last user prompt", slog.Error(err))
 		}
 
+		i.removeUnnecessaryCacheMarkers()
 		// Only inject tools into "actual" request.
 		i.injectTools()
 	}