Exit Strategies

The loop engine supports four exit strategies that determine when autonomous iteration stops.

Strategy Overview

auto — Default

Uses the response analyzer's completion detection to identify when the task is done. Scans for completion keywords, checks test status, and validates output quality.

const strategy: ExitStrategy = {
  type: 'auto',
  shouldExit(analysis: AnalysisResult): boolean {
    return analysis.completionSignal.detected &&
           analysis.completionSignal.confidence >= 70
  }
}

Behavior:

• Checks for completion keywords (done, complete, finished)
• Requires confidence ≥ 70 to exit
• Falls back to circuit breaker if stagnation is detected
• Best for: General-purpose tasks where the agent determines completion

fix_plan — Structured

Follows a predefined fix plan — a list of steps that must be completed in order. Exits when all steps are done or the plan is exhausted.

interface FixPlan {
  steps: FixStep[]
  currentStep: number
}
 
interface FixStep {
  description: string
  completed: boolean
  attempts: number
  maxAttempts: number
}
 
const strategy: ExitStrategy = {
  type: 'fix_plan',
  shouldExit(analysis: AnalysisResult, plan: FixPlan): boolean {
    return plan.steps.every(step => step.completed) ||
           plan.steps.every(step => step.attempts >= step.maxAttempts)
  }
}

Behavior:

• Each iteration maps to a step in the plan
• Steps are marked complete when the response indicates success
• Steps that exceed maxAttempts are skipped with a warning
• Best for: Bug fix workflows where the remediation steps are known

done_signals — Confidence-Based

Exits when the cumulative confidence score reaches 40, regardless of specific completion keywords. More permissive than auto.

const strategy: ExitStrategy = {
  type: 'done_signals',
  shouldExit(analysis: AnalysisResult): boolean {
    return analysis.confidence >= 40
  }
}

Behavior:

• Lower threshold than auto (40 vs 70)
• Doesn't require explicit completion keywords
• May exit on partial completion — use with caution
• Best for: Open-ended exploration tasks where "good enough" is acceptable

circuit_breaker — Safety

Exits only when the circuit breaker opens. This strategy never exits on its own — it relies entirely on stagnation detection.

const strategy: ExitStrategy = {
  type: 'circuit_breaker',
  shouldExit(analysis: AnalysisResult, circuitState: string): boolean {
    return circuitState === 'OPEN'
  }
}

Behavior:

• Only exits when stagnation is confirmed
• Will run until maxIterations if no stagnation is detected
• Highest risk of wasted compute
• Best for: High-stakes tasks where premature exit is worse than extra iterations

Composite Strategies

Exit strategies can be combined — any trigger fires the exit:

const compositeStrategy: ExitStrategy = {
  type: 'composite',
  strategies: ['auto', 'circuit_breaker'],
  shouldExit(analysis: AnalysisResult, circuitState: string): boolean {
    return analysis.completionSignal.detected ||
           circuitState === 'OPEN'
  }
}