refactor, Let the caller create the assessment of a "symflower fix" run

evaluate/task/symflower-fix.go

@@ -5,7 +5,6 @@ import (
 	"time"
 
 	pkgerrors "github.com/pkg/errors"
-	"github.com/symflower/eval-dev-quality/evaluate/metrics"
 	"github.com/symflower/eval-dev-quality/language"
 	"github.com/symflower/eval-dev-quality/log"
 	evaltask "github.com/symflower/eval-dev-quality/task"
@@ -33,27 +32,20 @@ func symflowerFix(logger *log.Logger, repositoryPath string, language language.L
 }
 
 // ExecuteWithSymflowerFix runs the "symflower fix" command and calculates the new assessments.
-func ExecuteWithSymflowerFix(ctx evaltask.Context, logger *log.Logger, packagePath string) (assessments metrics.Assessments, problems []error, err error) {
+func ExecuteWithSymflowerFix(ctx evaltask.Context, logger *log.Logger, packagePath string) (testResult *language.TestResult, processingTime uint64, problems []error, err error) {
 	// Run "symflower fix"  if the model response fails to execute.
 	logger.Print("model response alone failed execution, attempting to fix with \"symflower fix \"")
 
 	duration, err := symflowerFix(logger, packagePath, ctx.Language)
 	if err != nil {
-		return nil, problems, pkgerrors.WithStack(err)
+		return nil, 0, nil, pkgerrors.WithStack(err)
 	}
 
 	testResult, ps, err := ctx.Language.ExecuteTests(logger, packagePath)
 	problems = append(problems, ps...)
 	if err != nil {
-		return nil, problems, pkgerrors.WithMessage(err, "symflower fix")
+		return testResult, duration, problems, pkgerrors.WithMessage(err, "symflower fix")
 	}
-	logger.Printf("with symflower repair: Executes tests with %d coverage objects", testResult.Coverage)
 
-	// Symflower was able to fix a failure so now update the assessment with the improved results.
-	assessments = metrics.NewAssessments()
-	assessments[metrics.AssessmentKeyProcessingTime] = duration
-	assessments.Award(metrics.AssessmentKeyFilesExecuted)
-	assessments.AwardPoints(metrics.AssessmentKeyCoverage, testResult.Coverage)
-
-	return assessments, problems, nil
+	return testResult, duration, problems, nil
 }

evaluate/task/task-write-test.go

@@ -96,7 +96,7 @@ func (t *TaskWriteTests) Run(ctx evaltask.Context) (repositoryAssessment map[eva
 
 			// Run "symflower fix" if the model response fails to execute.
 			if ctx.Language.ID() == "golang" { // Currently we only support Go for "symflower fix".
-				withSymflowerFixAssessments, ps, err := ExecuteWithSymflowerFix(ctx, taskLogger.Logger, ctx.Repository.DataPath())
+				withSymflowerFixTestResult, processingTime, ps, err := ExecuteWithSymflowerFix(ctx, taskLogger.Logger, ctx.Repository.DataPath())
 				problems = append(problems, ps...)
 				if err != nil {
 					problems = append(problems, err)
@@ -106,6 +106,14 @@ func (t *TaskWriteTests) Run(ctx evaltask.Context) (repositoryAssessment map[eva
 
 					continue
 				} else {
+					ctx.Logger.Printf("with symflower repair: Executes tests with %d coverage objects", withSymflowerFixTestResult.Coverage)
+
+					// Symflower was able to fix a failure so now update the assessment with the improved results.
+					withSymflowerFixAssessments := metrics.NewAssessments()
+					withSymflowerFixAssessments[metrics.AssessmentKeyProcessingTime] = processingTime
+					withSymflowerFixAssessments.Award(metrics.AssessmentKeyFilesExecuted)
+					withSymflowerFixAssessments.AwardPoints(metrics.AssessmentKeyCoverage, withSymflowerFixTestResult.Coverage)
+
 					withSymflowerAssessmentForFile = metrics.CombineWithSymflowerFixAssessments(modelAssessmentForFile, withSymflowerFixAssessments)
 				}
 			}