Skip to content

refactor: deepen objective-evaluation choreography (PR 2) #41

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ElbertePlinio merged 2 commits into from
Jul 21, 2026
Merged

refactor: deepen objective-evaluation choreography (PR 2) #41

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
30f7e85
refactor: deepen objective-evaluation choreography (PR 2)
ElbertePlinio Jul 20, 2026
3c88138
test: add direct choreography contract test for objective evaluation
ElbertePlinio Jul 20, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
125 changes: 48 additions & 77 deletions app/lib/runner/agentic_run_orchestrator.dart
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,6 @@ import 'dart:io';
import 'package:crypto/crypto.dart';
import 'package:dart_arena/agent/agent_harness.dart';
import 'package:dart_arena/agent/agent_run_result.dart';
import 'package:dart_arena/analytics/result_primitives.dart';
import 'package:dart_arena/core/benchmark_task.dart';
import 'package:dart_arena/core/evaluation_context.dart';
import 'package:dart_arena/core/evaluation_result.dart';
Expand All @@ -19,6 +18,7 @@ import 'package:dart_arena/core/workspace_path.dart';
import 'package:dart_arena/evaluators/evaluator.dart';
import 'package:dart_arena/runner/evaluator_resource_limits.dart';
import 'package:dart_arena/runner/generated_code_sandbox.dart';
import 'package:dart_arena/runner/objective_evaluation.dart';
import 'package:dart_arena/runner/workdir_manager.dart';
import 'package:path/path.dart' as p;

Expand Down Expand Up @@ -333,35 +333,24 @@ class AgenticRunOrchestrator {
);
} else {
resultProvenance['gradingMode'] = 'clean_replay';
for (final evaluator in evaluators) {
final blocked = blockedEvaluationFor(
evaluatorId: evaluator.id,
previousResults: evaluations,
);
if (blocked != null) {
evaluations.add(blocked);
continue;
}
evaluations.add(
await evaluator.evaluate(
EvaluationContext(
workDir: gradingWorkspace,
response: response,
task: task,
previousResults: evaluations,
deniedEnvironmentKeys: workdirManager.deniedEnvironmentKeys,
generatedCodeSandbox: generatedCodeSandbox,
),
),
);
}
await runObjectiveEvaluators(
evaluators: evaluators,
evaluations: evaluations,
contextFor: (previousResults) => EvaluationContext(
workDir: gradingWorkspace,
response: response,
task: task,
previousResults: previousResults,
deniedEnvironmentKeys: workdirManager.deniedEnvironmentKeys,
generatedCodeSandbox: generatedCodeSandbox,
),
);
}

cancellationCheck?.call();
final aggregateScore = aggregate(evaluations, weights);
final primitives = determineResultPrimitives(
final outcome = finalizeObjectiveEvaluation(
evaluations: evaluations,
aggregateScore: aggregateScore,
weights: weights,
response: response,
);

Expand All @@ -371,15 +360,15 @@ class AgenticRunOrchestrator {
modelId: modelId,
taskId: task.id,
response: response,
evaluations: evaluations,
aggregateScore: aggregateScore,
evaluations: outcome.evaluations,
aggregateScore: outcome.aggregateScore,
completedAt: now(),
trialIndex: trialIndex,
taskVersion: task.version,
benchmarkTrack: task.track.name,
harnessId: harness.id,
primaryPass: primitives.primaryPass,
failureTag: primitives.failureTag,
primaryPass: outcome.primaryPass,
failureTag: outcome.failureTag,
patchText: patchText,
trajectoryLogPath: agentResult.trajectoryLogPath,
planId: planId,
Expand Down Expand Up @@ -434,10 +423,9 @@ class AgenticRunOrchestrator {
rationale: rationale,
details: {'error': error.toString()},
);
final aggregateScore = aggregate([evaluation], weights);
final primitives = determineResultPrimitives(
final outcome = finalizeObjectiveEvaluation(
evaluations: [evaluation],
aggregateScore: aggregateScore,
weights: weights,
response: response,
);
return TaskRunResult(
Expand All @@ -446,15 +434,15 @@ class AgenticRunOrchestrator {
modelId: modelId,
taskId: task.id,
response: response,
evaluations: [evaluation],
aggregateScore: aggregateScore,
evaluations: outcome.evaluations,
aggregateScore: outcome.aggregateScore,
completedAt: now(),
trialIndex: trialIndex,
taskVersion: task.version,
benchmarkTrack: task.track.name,
harnessId: harnessId,
primaryPass: primitives.primaryPass,
failureTag: primitives.failureTag,
primaryPass: outcome.primaryPass,
failureTag: outcome.failureTag,
planId: planId,
provenance: provenance,
);
Expand All @@ -481,29 +469,21 @@ class AgenticRunOrchestrator {
completionTokens: null,
latency: Duration.zero,
);
final evaluations = <EvaluationResult>[
environmentFailureEvaluation(
final evaluations = blockEvaluatorsForHardFailure(
evaluations: <EvaluationResult>[],
evaluators: applyTaskResourceLimitsToEvaluators(
task.evaluatorsFor(evaluatorConfig),
task,
),
failure: environmentFailureEvaluation(
rationale: rationale,
stderr: error.toString(),
phase: phase,
),
];
for (final evaluator in applyTaskResourceLimitsToEvaluators(
task.evaluatorsFor(evaluatorConfig),
task,
)) {
evaluations.add(
blockedEvaluationFor(
evaluatorId: evaluator.id,
previousResults: evaluations,
blockAllDownstream: true,
)!,
);
}
final aggregateScore = aggregate(evaluations, weights);
final primitives = determineResultPrimitives(
);
final outcome = finalizeObjectiveEvaluation(
evaluations: evaluations,
aggregateScore: aggregateScore,
weights: weights,
response: response,
);
return TaskRunResult(
Expand All @@ -512,15 +492,15 @@ class AgenticRunOrchestrator {
modelId: modelId,
taskId: task.id,
response: response,
evaluations: evaluations,
aggregateScore: aggregateScore,
evaluations: outcome.evaluations,
aggregateScore: outcome.aggregateScore,
completedAt: now(),
trialIndex: trialIndex,
taskVersion: task.version,
benchmarkTrack: task.track.name,
harnessId: harnessId,
primaryPass: primitives.primaryPass,
failureTag: primitives.failureTag,
primaryPass: outcome.primaryPass,
failureTag: outcome.failureTag,
planId: planId,
provenance: provenance,
);
Expand Down Expand Up @@ -614,24 +594,15 @@ class AgenticRunOrchestrator {
required PrepareFailed failure,
required String phase,
}) {
if (!hasHardDownstreamBlocker(evaluations)) {
evaluations.add(
environmentFailureEvaluation(
rationale: 'prepare failed',
stderr: failure.stderr,
phase: phase,
),
);
}
for (final evaluator in evaluators) {
evaluations.add(
blockedEvaluationFor(
evaluatorId: evaluator.id,
previousResults: evaluations,
blockAllDownstream: true,
)!,
);
}
blockEvaluatorsForHardFailure(
evaluations: evaluations,
evaluators: evaluators,
failure: environmentFailureEvaluation(
rationale: 'prepare failed',
stderr: failure.stderr,
phase: phase,
),
);
}

EvaluationResult _harnessEvaluation(AgentRunResult result) {
Expand Down
67 changes: 24 additions & 43 deletions app/lib/runner/codegen_task_executor.dart
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,19 +1,18 @@
import 'dart:async';

import 'package:dart_arena/analytics/result_primitives.dart';
import 'package:dart_arena/core/benchmark_task.dart';
import 'package:dart_arena/core/code_extractor.dart';
import 'package:dart_arena/core/evaluation_context.dart';
import 'package:dart_arena/core/evaluation_result.dart';
import 'package:dart_arena/core/evaluator_blocking.dart';
import 'package:dart_arena/core/evaluator_config.dart';
import 'package:dart_arena/core/model_response.dart';
import 'package:dart_arena/core/scoring.dart';
import 'package:dart_arena/core/task_run_result.dart';
import 'package:dart_arena/providers/model_provider.dart';
import 'package:dart_arena/providers/model_stream_event.dart';
import 'package:dart_arena/runner/evaluator_resource_limits.dart';
import 'package:dart_arena/runner/generated_code_sandbox.dart';
import 'package:dart_arena/runner/objective_evaluation.dart';
import 'package:dart_arena/runner/prompts/plan_aware_prompt.dart';
import 'package:dart_arena/runner/workdir_manager.dart';

Expand Down Expand Up @@ -155,52 +154,34 @@ class CodegenTaskExecutor {
final evaluations = <EvaluationResult>[];

if (prepResult is PrepareFailed) {
evaluations.add(
environmentFailureEvaluation(
blockEvaluatorsForHardFailure(
evaluations: evaluations,
evaluators: evaluators,
failure: environmentFailureEvaluation(
rationale: 'prepare failed',
stderr: prepResult.stderr,
),
);
for (final evaluator in evaluators) {
evaluations.add(
blockedEvaluationFor(
evaluatorId: evaluator.id,
previousResults: evaluations,
blockAllDownstream: true,
)!,
);
}
} else {
for (final evaluator in evaluators) {
cancellationCheck?.call();
final blocked = blockedEvaluationFor(
evaluatorId: evaluator.id,
previousResults: evaluations,
);
if (blocked != null) {
evaluations.add(blocked);
continue;
}
final result = await evaluator.evaluate(
EvaluationContext(
workDir: dir,
response: responseWithCode,
task: task,
previousResults: evaluations,
deniedEnvironmentKeys: workdirManager.deniedEnvironmentKeys,
generatedCodeSandbox: generatedCodeSandbox,
),
);
cancellationCheck?.call();
evaluations.add(result);
}
await runObjectiveEvaluators(
evaluators: evaluators,
evaluations: evaluations,
contextFor: (previousResults) => EvaluationContext(
workDir: dir,
response: responseWithCode,
task: task,
previousResults: previousResults,
deniedEnvironmentKeys: workdirManager.deniedEnvironmentKeys,
generatedCodeSandbox: generatedCodeSandbox,
),
cancellationCheck: cancellationCheck,
);
}

cancellationCheck?.call();
final aggregateScore = aggregate(evaluations, weights);
final primitives = determineResultPrimitives(
final outcome = finalizeObjectiveEvaluation(
evaluations: evaluations,
aggregateScore: aggregateScore,
weights: weights,
response: responseWithCode,
);

Expand All @@ -210,14 +191,14 @@ class CodegenTaskExecutor {
modelId: modelId,
taskId: task.id,
response: responseWithCode,
evaluations: evaluations,
aggregateScore: aggregateScore,
evaluations: outcome.evaluations,
aggregateScore: outcome.aggregateScore,
completedAt: now(),
trialIndex: trialIndex,
taskVersion: task.version,
benchmarkTrack: task.track.name,
primaryPass: primitives.primaryPass,
failureTag: primitives.failureTag,
primaryPass: outcome.primaryPass,
failureTag: outcome.failureTag,
planId: planId,
);
}
Expand Down
Loading
Loading