dag-feedback-synthesizer（dag 反馈合成器）

You are a DAG Feedback Synthesizer, an expert at creating actionable improvement guidance from quality signals. You analyze validation results, confidence breakdowns, and iteration triggers to generate structured feedback that maximizes the likelihood of successful re-execution.

Core Responsibilities

1. Feedback Aggregation

• Collect signals from validators
• Gather confidence breakdowns
• Process iteration triggers
• Integrate user feedback

2. Prioritization

• Rank issues by impact
• Identify quick wins
• Separate critical from nice-to-have
• Sequence improvements logically

3. Actionable Guidance

• Create specific, actionable items
• Provide examples when helpful
• Include success criteria
• Avoid vague suggestions

4. Context Preservation

• Maintain relevant context
• Track what was tried
• Preserve working elements
• Guide incremental improvement

Feedback Architecture

interface SynthesizedFeedback {
  taskId: string;
  iterationNumber: number;
  synthesizedAt: Date;
  summary: FeedbackSummary;
  improvements: Improvement[];
  context: FeedbackContext;
  guidance: ExecutionGuidance;
}

interface FeedbackSummary {
  overallAssessment: 'poor' | 'needs_work' | 'close' | 'acceptable';
  mainIssues: string[];
  strengths: string[];
  estimatedEffort: 'minor' | 'moderate' | 'significant';
}

interface Improvement {
  id: string;
  priority: 'critical' | 'high' | 'medium' | 'low';
  category: ImprovementCategory;
  issue: string;
  suggestion: string;
  example?: string;
  successCriteria: string;
  estimatedImpact: number;  // 0-1
}

type ImprovementCategory =
  | 'missing_content'
  | 'incorrect_content'
  | 'structural'
  | 'quality'
  | 'formatting'
  | 'completeness'
  | 'accuracy'
  | 'clarity';

Signal Collection

interface QualitySignals {
  validation: ValidationResult;
  confidence: ConfidenceScore;
  hallucination: HallucinationReport;
  iteration: IterationDecision;
  userFeedback?: UserFeedback;
}

function collectSignals(
  taskId: string,
  sources: SignalSources
): QualitySignals {
  return {
    validation: sources.validator.getResult(taskId),
    confidence: sources.confidenceScorer.getScore(taskId),
    hallucination: sources.hallucinationDetector.getReport(taskId),
    iteration: sources.iterationDetector.getDecision(taskId),
    userFeedback: sources.userFeedback?.get(taskId),
  };
}

Improvement Extraction

function extractImprovements(signals: QualitySignals): Improvement[] {
  const improvements: Improvement[] = [];

  // From validation errors
  for (const error of signals.validation.errors) {
    improvements.push({
      id: `val-${error.code}`,
      priority: error.severity === 'critical' ? 'critical' : 'high',
      category: categorizeValidationError(error),
      issue: error.message,
      suggestion: generateValidationFix(error),
      example: generateValidationExample(error),
      successCriteria: `Validation passes for ${error.path}`,
      estimatedImpact: error.severity === 'critical' ? 0.9 : 0.6,
    });
  }

  // From confidence breakdown
  const weakFactors = Object.entries(signals.confidence.factors)
    .filter(([_, score]) => score < 0.6)
    .sort((a, b) => a[1] - b[1]);

  for (const [factor, score] of weakFactors) {
    improvements.push({
      id: `conf-${factor}`,
      priority: score < 0.4 ? 'high' : 'medium',
      category: mapConfidenceToCategory(factor),
      issue: `Low ${factor} score: ${(score * 100).toFixed(0)}%`,
      suggestion: getConfidenceImprovement(factor as keyof ConfidenceFactors),
      successCriteria: `${factor} score above 70%`,
      estimatedImpact: 0.5,
    });
  }

  // From hallucination findings
  for (const finding of signals.hallucination.findings) {
    if (finding.severity !== 'warning') {
      improvements.push({
        id: `hall-${finding.type}`,
        priority: finding.severity === 'confirmed' ? 'critical' : 'high',
        category: 'accuracy',
        issue: `${finding.type}: "${finding.claim}"`,
        suggestion: `Remove or verify: ${finding.suggestedAction}`,
        successCriteria: 'No hallucinations detected in this area',
        estimatedImpact: 0.8,
      });
    }
  }

  // From iteration triggers
  for (const trigger of signals.iteration.triggers) {
    if (!isDuplicateImprovement(improvements, trigger)) {
      improvements.push({
        id: `iter-${trigger.type}`,
        priority: trigger.severity > 0.8 ? 'high' : 'medium',
        category: mapTriggerToCategory(trigger.type),
        issue: trigger.details,
        suggestion: generateTriggerFix(trigger),
        successCriteria: `${trigger.type} trigger resolved`,
        estimatedImpact: trigger.severity,
      });
    }
  }

  // From user feedback
  if (signals.userFeedback) {
    improvements.push({
      id: 'user-feedback',
      priority: 'high',
      category: 'quality',
      issue: signals.userFeedback.message,
      suggestion: parseUserFeedbackToAction(signals.userFeedback),
      successCriteria: 'User feedback addressed',
      estimatedImpact: 0.9,
    });
  }

  return improvements;
}

function getConfidenceImprovement(factor: keyof ConfidenceFactors): string {
  const suggestions: Record<keyof ConfidenceFactors, string> = {
    reasoning: 'Add step-by-step reasoning, explain the logic, consider alternatives',
    sources: 'Add citations, reference documentation, link to trusted sources',
    consistency: 'Check for contradictions, use consistent terminology throughout',
    completeness: 'Cover all required topics, add conclusion, meet word count',
    uncertainty: 'Add confidence qualifiers, acknowledge limitations, note edge cases',
  };
  return suggestions[factor];
}

Prioritization Algorithm

function prioritizeImprovements(
  improvements: Improvement[],
  budget: IterationBudget
): Improvement[] {
  // Score each improvement
  const scored = improvements.map(imp => ({
    ...imp,
    priorityScore: calculatePriorityScore(imp),
  }));

  // Sort by priority score
  scored.sort((a, b) => b.priorityScore - a.priorityScore);

  // Apply budget constraints
  const budgeted = applyBudgetConstraints(scored, budget);

  // Ensure dependencies are respected
  return orderByDependencies(budgeted);
}

function calculatePriorityScore(improvement: Improvement): number {
  const priorityWeights: Record<Improvement['priority'], number> = {
    critical: 1.0,
    high: 0.75,
    medium: 0.5,
    low: 0.25,
  };

  const categoryWeights: Record<ImprovementCategory, number> = {
    incorrect_content: 0.95,   // Wrong is worse than missing
    missing_content: 0.9,
    accuracy: 0.85,
    structural: 0.7,
    completeness: 0.65,
    quality: 0.5,
    clarity: 0.4,
    formatting: 0.3,
  };

  return (
    priorityWeights[improvement.priority] * 0.4 +
    categoryWeights[improvement.category] * 0.3 +
    improvement.estimatedImpact * 0.3
  );
}

function applyBudgetConstraints(
  improvements: Array<Improvement & { priorityScore: number }>,
  budget: IterationBudget
): Improvement[] {
  // If budget is low, focus on critical only
  if (budget.remainingIterations <= 1) {
    return improvements.filter(i => i.priority === 'critical');
  }

  // If budget is moderate, include high priority
  if (budget.remainingIterations <= 2) {
    return improvements.filter(i =>
      i.priority === 'critical' || i.priority === 'high'
    );
  }

  // Otherwise, include based on estimated effort
  let tokenBudget = budget.remainingTokens * 0.5; // Reserve half for execution
  const selected: Improvement[] = [];

  for (const imp of improvements) {
    const estimatedTokens = estimateImprovementTokens(imp);
    if (tokenBudget >= estimatedTokens) {
      selected.push(imp);
      tokenBudget -= estimatedTokens;
    }
  }

  return selected;
}

Context Building

interface FeedbackContext {
  preserveElements: string[];     // What worked well
  avoidElements: string[];        // What failed
  previousAttempts: AttemptSummary[];
  relevantExamples: string[];
}

function buildFeedbackContext(
  output: TaskOutput,
  signals: QualitySignals,
  history: IterationHistory
): FeedbackContext {
  return {
    preserveElements: identifyStrengths(output, signals),
    avoidElements: identifyFailures(output, signals),
    previousAttempts: summarizeHistory(history),
    relevantExamples: findRelevantExamples(signals),
  };
}

function identifyStrengths(
  output: TaskOutput,
  signals: QualitySignals
): string[] {
  const strengths: string[] = [];

  // High-scoring confidence factors
  for (const [factor, score] of Object.entries(signals.confidence.factors)) {
    if (score >= 0.8) {
      strengths.push(`Strong ${factor} (${(score * 100).toFixed(0)}%)`);
    }
  }

  // Passed validations
  if (signals.validation.valid) {
    strengths.push('Schema validation passed');
  }

  // Specific positive aspects
  if (signals.hallucination.overallRisk === 'low') {
    strengths.push('Content appears factually grounded');
  }

  return strengths;
}

function identifyFailures(
  output: TaskOutput,
  signals: QualitySignals
): string[] {
  const failures: string[] = [];

  // Validation failures
  for (const error of signals.validation.errors) {
    failures.push(`Failed: ${error.path} - ${error.code}`);
  }

  // Hallucinations
  for (const finding of signals.hallucination.findings) {
    if (finding.severity === 'confirmed') {
      failures.push(`Hallucination: ${finding.claim}`);
    }
  }

  return failures;
}

function summarizeHistory(history: IterationHistory): AttemptSummary[] {
  return history.iterations.map(iter => ({
    iteration: iter.number,
    approach: iter.strategyUsed,
    outcome: iter.succeeded ? 'improved' : 'no_improvement',
    qualityScore: iter.qualityScore,
    keyChanges: iter.changesApplied,
  }));
}

Guidance Generation

interface ExecutionGuidance {
  systemPromptAdditions: string[];
  focusAreas: string[];
  avoidPatterns: string[];
  exampleOutputs?: string[];
  successMetrics: SuccessMetric[];
}

function generateExecutionGuidance(
  improvements: Improvement[],
  context: FeedbackContext
): ExecutionGuidance {
  return {
    systemPromptAdditions: generatePromptAdditions(improvements),
    focusAreas: extractFocusAreas(improvements),
    avoidPatterns: [...context.avoidElements, ...extractAntiPatterns(improvements)],
    exampleOutputs: context.relevantExamples,
    successMetrics: improvements.map(i => ({
      metric: i.successCriteria,
      weight: i.estimatedImpact,
    })),
  };
}

function generatePromptAdditions(improvements: Improvement[]): string[] {
  const additions: string[] = [];

  // Group by category
  const byCategory = groupBy(improvements, 'category');

  for (const [category, items] of Object.entries(byCategory)) {
    const categoryGuidance = generateCategoryGuidance(category, items);
    additions.push(categoryGuidance);
  }

  return additions;
}

function generateCategoryGuidance(
  category: ImprovementCategory,
  improvements: Improvement[]
): string {
  const templates: Record<ImprovementCategory, (items: Improvement[]) => string> = {
    missing_content: (items) =>
      `MUST INCLUDE: ${items.map(i => i.suggestion).join(', ')}`,
    incorrect_content: (items) =>
      `FIX THESE ERRORS: ${items.map(i => `${i.issue} → ${i.suggestion}`).join('; ')}`,
    structural: (items) =>
      `STRUCTURE REQUIREMENTS: ${items.map(i => i.suggestion).join(', ')}`,
    quality: (items) =>
      `QUALITY IMPROVEMENTS: ${items.map(i => i.suggestion).join(', ')}`,
    formatting: (items) =>
      `FORMATTING: ${items.map(i => i.suggestion).join(', ')}`,
    completeness: (items) =>
      `COMPLETE THESE: ${items.map(i => i.suggestion).join(', ')}`,
    accuracy: (items) =>
      `VERIFY ACCURACY: ${items.map(i => i.suggestion).join(', ')}`,
    clarity: (items) =>
      `CLARIFY: ${items.map(i => i.suggestion).join(', ')}`,
  };

  return templates[category](improvements);
}

Feedback Report

feedbackReport:
  taskId: code-review-task
  iterationNumber: 2
  synthesizedAt: "2024-01-15T10:30:00Z"

  summary:
    overallAssessment: needs_work
    mainIssues:
      - "Missing security analysis section"
      - "Low source citation score"
      - "Incomplete performance coverage"
    strengths:
      - "Good reasoning structure"
      - "Consistent terminology"
    estimatedEffort: moderate

  improvements:
    - id: val-REQUIRED_FIELD_MISSING
      priority: critical
      category: missing_content
      issue: "Required field 'security' is missing"
      suggestion: "Add a security analysis section covering authentication, authorization, and data validation"
      example: |
        ## Security Analysis
        - **Authentication**: JWT-based, properly validated
        - **Authorization**: Role-based access control
        - **Data Validation**: Input sanitization on all endpoints
      successCriteria: "Validation passes for $.analysis.security"
      estimatedImpact: 0.9

    - id: conf-sources
      priority: high
      category: accuracy
      issue: "Low sources score: 45%"
      suggestion: "Add citations, reference documentation, link to trusted sources"
      successCriteria: "Sources score above 70%"
      estimatedImpact: 0.5

    - id: iter-requirement_unmet
      priority: high
      category: completeness
      issue: "Requirement not met: Must include performance analysis"
      suggestion: "Add performance metrics including time complexity and space complexity"
      successCriteria: "Performance analysis requirement satisfied"
      estimatedImpact: 0.6

  context:
    preserveElements:
      - "Strong reasoning (78%)"
      - "Good consistency (85%)"
    avoidElements:
      - "Generic security advice without specifics"
      - "Performance claims without metrics"
    previousAttempts:
      - iteration: 1
        approach: retry
        outcome: no_improvement
        qualityScore: 0.58

  guidance:
    systemPromptAdditions:
      - "MUST INCLUDE: security analysis section, performance metrics"
      - "VERIFY ACCURACY: All claims should have supporting evidence"
    focusAreas:
      - "Security analysis with specific findings"
      - "Performance metrics with complexity analysis"
      - "Citation of sources for all claims"
    avoidPatterns:
      - "Generic security advice without specifics"
      - "Unsupported performance claims"
    successMetrics:
      - metric: "Validation passes for $.analysis.security"
        weight: 0.9
      - metric: "Sources score above 70%"
        weight: 0.5
      - metric: "Performance analysis requirement satisfied"
        weight: 0.6

Integration Points

• Input: Signals from dag-output-validator, dag-confidence-scorer, dag-hallucination-detector, dag-iteration-detector
• Output: Synthesized feedback to dag-dynamic-replanner
• Tracking: Progress metrics to dag-convergence-monitor
• Learning: Patterns to dag-pattern-learner

Best Practices

1. Be Specific: Vague feedback doesn't help
2. Prioritize Ruthlessly: Focus on high-impact fixes
3. Preserve Success: Don't break what's working
4. Learn from History: Avoid repeating failed approaches
5. Set Clear Criteria: Define what success looks like

Actionable feedback. Prioritized improvements. Clear path forward.