Loop Breaker

LLMs often get stuck in repetitive cycles, burning tokens on identical tool calls or redundant apologies that never change the state. When a model repeats an action, it usually expects a different result from a static environment, leading to a "doom loop" that only ends when the user manually aborts the session.

What it does

• Identifies repeats by canonicalizing tool inputs and lowercasing replies to detect duplicates regardless of whitespace or key order.
• Tracks streaks using a consecutive-only counter to distinguish between legitimate re-reads and stuck cycles.
• Applies a ladder that triggers a forced strategy shift at four repeats and a hard stop at eight.
• Enforces logic checks to differentiate between bounded polling (legitimate) and blind retries of failing commands (doom loop).

How it works

1. Log actions using the loop-watch.sh script or a manual tally to record every tool call and assistant reply.
2. Evaluate the ladder after every turn to see if the current consecutive streak has reached the nudge (4) or stop (8) threshold.
3. Pivot or Report by either switching to a different tool/input when nudged or providing a structured "Unresolved" report when stopped.

Frameworks & tools

This skill is framework-agnostic and relies on a Bash script (loop-watch.sh) for state tracking. It works within any environment supporting tool calls or long-context reasoning, including Claude Code, Cursor, and custom agentic loops.

Why this beats prompting it yourself

Generic "don't repeat yourself" instructions are often ignored during high-pressure debugging. This skill provides a mathematical threshold and a specific "ladder" protocol that forces the agent to stop and report instead of hallucinating progress through repetition.

Use cases

• Fixing stuck searches when an agent queries the same term repeatedly despite zero results.
• Breaking apology loops where a model keeps restating its intention without executing code.
• Managing cheap models that lack the reasoning depth to realize they are stuck in a retry cycle.

Known limitations

Does not automatically detect A/B/A/B alternating loops; these require manual intervention as the consecutive counter resets. Requires consistent logging of tool calls to function accurately.