Agent cost control: Fail fast on OpenRouter free models and add hard timeouts

- Disables tool capabilities for models with ':free' or '-free' suffix to prevent planning failures
- Limits conductor to 5 consecutive execution errors before aborting (prevents infinite retry loops)
- Adds a 25-minute context timeout to the executor to guarantee the process dies and the machine is auto-destroyed on Fly.io

Author: Test

cmd/gptcode/agent.go

@@ -6,6 +6,7 @@ import (
 	"os"
 	"os/exec"
 	"strings"
+	"time"
 
 	"github.com/spf13/cobra"
 
@@ -157,7 +158,8 @@ Please discover what failed by examining the code, tests, and standard CI config
 (such as npm test, mix test, go test, etc.) and write a fix. Once fixed, create a PR.`, branch, sha)
 	}
 
-	ctx := context.Background()
+	timeoutCtx, cancel := context.WithTimeout(context.Background(), 25*time.Minute)
+	defer cancel()
 
 	backendName := "openrouter"
 	baseURL := "https://openrouter.ai/api/v1"
@@ -195,7 +197,7 @@ Please discover what failed by examining the code, tests, and standard CI config
 	executor := modes.NewAutonomousExecutorWithLive(provider, ".", queryModel, language, liveClient, reportConfig, backendName)
 
 	fmt.Println("🤖 Starting AutoFix process...")
-	err = executor.Execute(ctx, prompt)
+	err = executor.Execute(timeoutCtx, prompt)
 	if err != nil {
 		fmt.Printf("❌ AutoFix failed: %v\n", err)
 		return err

internal/config/model_selector.go

@@ -167,6 +167,12 @@ func (ms *ModelSelector) loadCatalog() error {
 				}
 			}
 
+			// Special case: OpenRouter free models typically lack tool support
+			if strings.Contains(modelIDLower, ":free") || strings.Contains(modelIDLower, "-free") || strings.Contains(modelNameLower, " free") {
+				model.Capabilities.SupportsTools = false
+				model.Capabilities.SupportsFileOperations = false
+			}
+
 			if caps, ok := modelMap["capabilities"].(map[string]interface{}); ok {
 				if val, ok := caps["supports_tools"].(bool); ok {
 					model.Capabilities.SupportsTools = val

internal/maestro/conductor.go

@@ -288,6 +288,8 @@ func (c *Conductor) ExecuteTask(ctx context.Context, task string, complexity str
 		fmt.Fprintf(os.Stderr, "[MAESTRO] LoopDetector initialized with intent=%s\n", intent)
 	}
 
+	consecutiveErrors := 0
+
 	for {
 		// Check if we should continue (intent-aware limits + loop detection)
 		shouldContinue, stopReason := c.loopDetector.ShouldContinue()
@@ -335,8 +337,17 @@ func (c *Conductor) ExecuteTask(ctx context.Context, task string, complexity str
 		c.ReportProgress("editing", fmt.Sprintf("Completed - %d files changed", len(modifiedFiles)))
 		c.selector.RecordUsage(editBackend, editModel, err == nil, errorMsg(err))
 		if err != nil {
+			consecutiveErrors++
+			if consecutiveErrors >= 5 {
+				if os.Getenv("GPTCODE_DEBUG") == "1" {
+					fmt.Fprintf(os.Stderr, "[MAESTRO] Aborting after %d consecutive execution errors\n", consecutiveErrors)
+				}
+				c.ReportError("execution", fmt.Sprintf("Aborted: %d consecutive execution errors. Last error: %v", consecutiveErrors, err))
+				return fmt.Errorf("task aborted: %d consecutive execution errors: %w", consecutiveErrors, err)
+			}
+			
 			// LoopDetector will handle max iterations check on next iteration
-			fmt.Printf("[WARNING] Execution error: %v\n", err)
+			fmt.Printf("[WARNING] Execution error (%d/5): %v\n", consecutiveErrors, err)
 
 			// Use enhanced recovery system
 			recoveryCtx := &RecoveryContext{
@@ -374,6 +385,8 @@ func (c *Conductor) ExecuteTask(ctx context.Context, task string, complexity str
 			continue
 		}
 
+		consecutiveErrors = 0
+
 		// Record execution metrics
 		if c.Tracer != nil {
 			metrics := observability.Metrics{