Behavioral Loop Detection

Behavioral Loop Detection

Overview

Detects when an agent is stuck in a repetitive action loop by maintaining a rolling 20-action history window. Compares recent actions using normalized similarity scoring and applies escalating interventions before the agent wastes tokens or blocks progress.

Designed for any agentic context — browser automation, file editing, API calls, or multi-step orchestration — not tied to any specific tool type.

When to Use

Invoke this skill when:

• An agent task has been running for more than 10 steps without clear progress
• The same tool is being called repeatedly with similar arguments
• A task appears stalled with no TaskUpdate(completed) after a long sequence
• You want proactive loop-prevention in a custom orchestrator

Skill({ skill: 'behavioral-loop-detection' });

The Iron Law

SAME ACTION × 3 = REPLAN
SAME ACTION × 5 = EXPLORE
SAME ACTION × 8 = FORCE-DONE

Never let an agent silently loop forever. Each repetition consumes tokens and produces no value.

Action History Window

Maintenance Rules

• Keep a rolling buffer of the last 20 actions (FIFO — drop oldest when full)
• Each entry in the buffer: { toolName, normalizedArgs, timestamp, stepIndex }
• Reset the buffer only when a fundamentally different tool/action type appears
• Never reset on argument variation alone (e.g., same tool, different file path = still similar)

Normalized Action Comparison

Two actions are similar when:

1. The toolName is identical, AND
2. The normalized argument string similarity score ≥ 0.75 (Jaccard over word tokens)

Normalization rules (apply in order):

1. Convert all values to lowercase
2. Strip file paths down to basename only (e.g., /a/b/file.ts → file.ts)
3. Remove timestamps, UUIDs (8-4-4-4-12 hex pattern), and numeric IDs
4. Sort object keys alphabetically before stringifying
5. Truncate the result to 200 characters

Similarity scoring:

function jaccardSimilarity(a, b) {
  const setA = new Set(a.split(/\s+/));
  const setB = new Set(b.split(/\s+/));
  const intersection = new Set([...setA].filter(x => setB.has(x)));
  const union = new Set([...setA, ...setB]);
  return intersection.size / union.size;
}

Escalation Ladder

• Escalation levels are cumulative: Level 2 fires at step 5 regardless of whether Level 1 fired
• After FORCE-DONE signal, call TaskUpdate({ status: 'completed', metadata: { partial: true, loopDetected: true } })

Workflow

Step 1: Initialize the Action Buffer

At the start of task execution or when this skill is loaded, create the buffer:

Command:

const actionBuffer = {
  window: [], // Array of last 20 action entries
  maxSize: 20,
  similarRunLength: 0, // Current streak of similar actions
  lastNormalized: null, // Normalized args of last action
};

Expected output: In-memory buffer ready; no file I/O required unless persisting across sessions.

Verify: actionBuffer.window.length === 0 on initialization.

Step 2: Record Each Action

Before executing any tool call, append to the buffer:

Command:

function recordAction(buffer, toolName, rawArgs) {
  const normalized = normalizeArgs(toolName, rawArgs);
  const entry = {
    toolName,
    normalizedArgs: normalized,
    timestamp: Date.now(),
    stepIndex: buffer.window.length,
  };
  if (buffer.window.length >= buffer.maxSize) {
    buffer.window.shift(); // drop oldest
  }
  buffer.window.push(entry);
  return entry;
}

Expected output: Buffer contains the new entry; length never exceeds 20.

Verify: buffer.window.length <= 20 after each call.

Step 3: Check Similarity Against Last Action

After recording, compute similarity and update the run-length counter:

Command:

function checkSimilarity(buffer, currentEntry) {
  if (!buffer.lastNormalized) {
    buffer.lastNormalized = currentEntry.normalizedArgs;
    buffer.similarRunLength = 1;
    return { similar: false, runLength: 1 };
  }
  const score = jaccardSimilarity(buffer.lastNormalized, currentEntry.normalizedArgs);
  if (score >= 0.75) {
    buffer.similarRunLength += 1;
  } else {
    buffer.similarRunLength = 1;
    buffer.lastNormalized = currentEntry.normalizedArgs;
  }
  return { similar: score >= 0.75, runLength: buffer.similarRunLength, score };
}

Expected output: { similar: boolean, runLength: number, score: number }

Verify: runLength increments only when score >= 0.75.

Step 4: Apply Escalation Rule

Evaluate the run-length and emit the appropriate intervention:

Command:

function applyEscalation(runLength, taskId) {
  if (runLength >= 8) {
    console.error(`[loop-detection] FORCE-DONE: ${runLength} similar actions. Task ${taskId}`);
    return {
      level: 3,
      action: 'force-done',
      message:
        'Loop limit reached (8 repetitions). Mark this task complete with partial results and explain what was not accomplished.',
    };
  }
  if (runLength >= 5) {
    console.error(`[loop-detection] EXPLORE: ${runLength} similar actions. Task ${taskId}`);
    return {
      level: 2,
      action: 'explore',
      message:
        'You have repeated a similar action 5 times. The current approach is failing. Try a completely different tool or method.',
    };
  }
  if (runLength >= 3) {
    console.error(`[loop-detection] REPLAN: ${runLength} similar actions. Task ${taskId}`);
    return {
      level: 1,
      action: 'replan',
      message:
        'You have repeated a similar action 3 times. Stop and produce a revised plan before continuing.',
    };
  }
  return { level: 0, action: 'continue', message: null };
}

Expected output: { level: 0|1|2|3, action: string, message: string|null }

Verify: level === 3 triggers TaskUpdate({ status: 'completed', metadata: { partial: true, loopDetected: true } }).

Step 5: Handle FORCE-DONE

When level === 3, the agent MUST stop and complete the task:

Command:

// In the agent's task loop, when applyEscalation returns level 3:
TaskUpdate({
  taskId: context.taskId,
  status: 'completed',
  metadata: {
    summary:
      'Task partially completed. Loop detected after 8 similar actions. ' + partialResultsSummary,
    partial: true,
    loopDetected: true,
    loopDetails: {
      runLength: runLength,
      lastAction: buffer.lastNormalized,
      bufferSnapshot: buffer.window.slice(-5),
    },
  },
});

Expected output: Task marked completed with partial: true metadata.

Verify: TaskList() shows the task as completed, not in_progress.

Enforcement Hooks

Input validated against schemas/input.schema.json before execution. Output contract defined in schemas/output.schema.json.

Pre-execute hook at hooks/pre-execute.cjs validates that taskId is provided. Post-execute hook at hooks/post-execute.cjs emits a loop-detection event to tool-events.jsonl.

Integration Points

• master-orchestrator: Integrate after each wave of agent spawns; check for stalled tasks
• developer: Active during file-editing loops (same file edited 3+ times without test progress)
• qa: Active during test-fix loops (same test file modified 3+ times without green)
• planner: Active during re-planning loops (same plan section rewritten 3+ times)

Anti-Patterns

• Never reset the buffer on every tool call — that defeats detection
• Never apply force-done for level < 3 — nudges are sufficient at levels 1 and 2
• Never compare raw unormalized args — path differences will mask real loops
• Never skip the similarity check when toolName differs — tool variety alone ≠ not looping

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md for previously detected loop patterns.

After completing: If a loop was detected and force-done triggered, append to .claude/context/memory/learnings.md:

## Behavioral Loop Detection — [date]
- Task [taskId]: [toolName] looped [N] times. Final normalized args: [args].
- Root cause hypothesis: [explain why the agent got stuck]
- Recommendation: [what to change in the task prompt or approach]

After issues: Append to .claude/context/memory/issues.md if force-done fires more than twice in a session.

Related Skills

• verification-before-completion — Pre-completion gates that prevent false success
• judge-verification — Independent LLM judge that verifies task completion
• error-recovery-escalation — 5-level error recovery before force-done
• context-compressor — Compress context when loop detection fires repeatedly