cognitive-bullwhip

What This Skill Does

The CognitiveBullwhip skill is a diagnostic powerhouse for complex AI agent architectures. Much like the supply chain phenomenon where minor consumer demand shifts lead to massive production volatility, AI agent chains often suffer from error amplification. A subtle nuance in an initial prompt can lead to a catastrophic hallucination or logic failure after several processing layers. This skill ingest your system's decision logs to identify, quantify, and map these amplification patterns. By pinpointing the specific layer where variance began to spiral, it moves your debugging process from symptom-chasing to surgical correction.

Installation

To integrate this diagnostic tool into your OpenClaw environment, execute the following command in your terminal: clawhub install openclaw/skills/skills/jkc3080/cognitive-bullwhip Once installed, ensure your agent's decision logs are formatted according to the skill's input schema to allow for accurate variance calculation.

Use Cases

• Post-Incident Forensics: When an agent produces an unexpected failure, use this skill to trace the ripple effect back to the originating input error.
• Stability Auditing: Run this as a CI/CD gate before moving agents from staging to production to ensure they aren't prone to cascading failures.
• Agent Optimization: Identify which specific processing layers or prompts are acting as 'amplifiers' for uncertainty, allowing you to tighten your constraints or improve prompt engineering in those specific zones.

Example Prompts

1. "Run a diagnostic on the last 24 hours of logs; I'm seeing erratic output behavior and need to find the root of the variance."
2. "Analyze the current decision log to see if a Cognitive Bullwhip effect is active, and provide a list of recommended intervention points."
3. "Compare our staging agent performance against the production log; identify which processing layer is causing the highest amplification of input noise."

Tips & Limitations

• Context is Key: The quality of the analysis is directly tied to the granularity of your decision_log. Ensure you are logging intermediate steps, not just final outputs.
• Not a Fix: This tool identifies the location and severity of the problem; it does not automatically rewrite your code or prompts. You must apply the recommended interventions manually.
• Window Selection: Use a shorter observation_window to detect high-frequency jitter, and a longer window to identify slow-burn 'drift' that may be occurring over several days.