perf: reduce API costs with 6 optimization fixes

1. read_file: truncate output to 10K chars (was unlimited)
2. max_tool_iterations: reduce default from 50 to 15
3. web_fetch: reduce default from 50K to 20K chars
4. Anthropic: add cache_control on last tool definition for prompt caching
5. Token usage logging: track and log cumulative input/output tokens per session
6. max_tokens: reduce default from 32K to 8K, resolve real context window per model

Made-with: Cursor

pkg/agent/instance.go

@@ -148,6 +148,8 @@ func NewAgentInstance(
 
 	candidates := providers.ResolveCandidatesWithLookup(modelCfg, defaults.Provider, resolveFromModelList)
 
+	contextWindow := resolveContextWindow(model, maxTokens)
+
 	return &AgentInstance{
 		ID:             agentID,
 		Name:           agentName,
@@ -157,7 +159,7 @@ func NewAgentInstance(
 		MaxIterations:  maxIter,
 		MaxTokens:      maxTokens,
 		Temperature:    temperature,
-		ContextWindow:  maxTokens,
+		ContextWindow:  contextWindow,
 		Provider:       provider,
 		Sessions:       sessionsManager,
 		ContextBuilder: contextBuilder,
@@ -223,3 +225,29 @@ func expandHome(path string) string {
 	}
 	return path
 }
+
+// resolveContextWindow returns the actual context window size for a model.
+// This is used for summarization thresholds and should reflect the model's
+// real capacity rather than the output max_tokens setting.
+func resolveContextWindow(model string, maxTokens int) int {
+	m := strings.ToLower(model)
+	switch {
+	case strings.Contains(m, "claude"):
+		return 200000
+	case strings.Contains(m, "gpt-5"), strings.Contains(m, "gpt-4"):
+		return 128000
+	case strings.Contains(m, "gemini"):
+		return 1000000
+	case strings.Contains(m, "deepseek"):
+		return 64000
+	case strings.Contains(m, "llama"):
+		return 128000
+	case strings.Contains(m, "qwen"):
+		return 32000
+	default:
+		if maxTokens > 32000 {
+			return maxTokens
+		}
+		return 32000
+	}
+}

pkg/agent/loop.go

@@ -123,7 +123,7 @@ func registerSharedTools(
 		} else if searchTool != nil {
 			agent.Tools.Register(searchTool)
 		}
-		fetchTool, err := tools.NewWebFetchToolWithProxy(50000, cfg.Tools.Web.Proxy, cfg.Tools.Web.FetchLimitBytes)
+		fetchTool, err := tools.NewWebFetchToolWithProxy(20000, cfg.Tools.Web.Proxy, cfg.Tools.Web.FetchLimitBytes)
 		if err != nil {
 			logger.ErrorCF("agent", "Failed to create web fetch tool", map[string]any{"error": err.Error()})
 		} else {
@@ -661,6 +661,7 @@ func (al *AgentLoop) runLLMIteration(
 ) (string, int, error) {
 	iteration := 0
 	var finalContent string
+	var totalInputTokens, totalOutputTokens int
 
 	for iteration < agent.MaxIterations {
 		iteration++
@@ -805,6 +806,11 @@ func (al *AgentLoop) runLLMIteration(
 
 		go al.handleReasoning(ctx, response.Reasoning, opts.Channel, al.targetReasoningChannelID(opts.Channel))
 
+		if response.Usage != nil {
+			totalInputTokens += response.Usage.PromptTokens
+			totalOutputTokens += response.Usage.CompletionTokens
+		}
+
 		logger.DebugCF("agent", "LLM response",
 			map[string]any{
 				"agent_id":       agent.ID,
@@ -968,6 +974,17 @@ func (al *AgentLoop) runLLMIteration(
 		}
 	}
 
+	if totalInputTokens > 0 || totalOutputTokens > 0 {
+		logger.InfoCF("agent", "Session token usage",
+			map[string]any{
+				"agent_id":      agent.ID,
+				"iterations":    iteration,
+				"input_tokens":  totalInputTokens,
+				"output_tokens": totalOutputTokens,
+				"total_tokens":  totalInputTokens + totalOutputTokens,
+			})
+	}
+
 	return finalContent, iteration, nil
 }
 

pkg/config/defaults.go

@@ -30,9 +30,9 @@ func DefaultConfig() *Config {
 				RestrictToWorkspace: true,
 				Provider:            "",
 				Model:               "",
-				MaxTokens:           32768,
+			MaxTokens:           8192,
 			Temperature:         nil, // nil means use provider default
-				MaxToolIterations:   50,
+			MaxToolIterations:   15,
 			},
 		},
 		Bindings: []AgentBinding{},

pkg/providers/anthropic/provider.go

@@ -237,7 +237,7 @@ func buildParams(
 
 func translateTools(tools []ToolDefinition) []anthropic.ToolUnionParam {
 	result := make([]anthropic.ToolUnionParam, 0, len(tools))
-	for _, t := range tools {
+	for i, t := range tools {
 		tool := anthropic.ToolParam{
 			Name: t.Function.Name,
 			InputSchema: anthropic.ToolInputSchemaParam{
@@ -256,6 +256,9 @@ func translateTools(tools []ToolDefinition) []anthropic.ToolUnionParam {
 			}
 			tool.InputSchema.Required = required
 		}
+		if i == len(tools)-1 {
+			tool.CacheControl = anthropic.NewCacheControlEphemeralParam()
+		}
 		result = append(result, anthropic.ToolUnionParam{OfTool: &tool})
 	}
 	return result

pkg/tools/filesystem.go

@@ -117,6 +117,8 @@ func (t *ReadFileTool) Parameters() map[string]any {
 	}
 }
 
+const maxReadFileChars = 10000
+
 func (t *ReadFileTool) Execute(ctx context.Context, args map[string]any) *ToolResult {
 	path, ok := args["path"].(string)
 	if !ok {
@@ -127,7 +129,12 @@ func (t *ReadFileTool) Execute(ctx context.Context, args map[string]any) *ToolRe
 	if err != nil {
 		return ErrorResult(err.Error())
 	}
-	return NewToolResult(string(content))
+
+	text := string(content)
+	if len(text) > maxReadFileChars {
+		text = text[:maxReadFileChars] + fmt.Sprintf("\n\n... (truncated, showing %d of %d chars. Use exec with head/tail/sed for specific sections)", maxReadFileChars, len(content))
+	}
+	return NewToolResult(text)
 }
 
 type WriteFileTool struct {

pkg/tools/web.go

@@ -22,7 +22,7 @@ const (
 	perplexityTimeout = 30 * time.Second // Perplexity (LLM-based, slower)
 	fetchTimeout      = 60 * time.Second // WebFetchTool
 
-	defaultMaxChars = 50000
+	defaultMaxChars = 20000
 	maxRedirects    = 5
 )