debug

Diagnose bugs by reproducing symptoms and proving the causal mechanism before fixing.

1|Updated Sep 10, 2026
One-click install
npx skills add https://github.com/blauwtje/exo --skill debug-blauwtje
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: debug
Source: https://github.com/blauwtje/exo/tree/main/skills/debug
Command: npx skills add https://github.com/blauwtje/exo --skill debug-blauwtje

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? It prevents plausible patches that hide symptoms without explaining them, by requiring a reproduction and a proven causal line before any production edit is made. ## Core Features & Use Cases - Evidence-first debugging loop: Reproduce the failure as one bare command, instrument to separate two competing hypotheses, isolate the minimal trigger, then predict and fix only what the mechanism requires. - Delegated investigation and fixing: Dispatch an Opus investigator to prove the cause and a Sonnet fixer to apply the predicted change, handing off through a structured file so file dumps never reach the main session. - Profiling forensics: Diagnose slowdowns, memory growth, and captured traces from CPU profiles, heap snapshots, and trace artifacts, attributing hot paths to file, symbol, and line. - Use Case: A user reports a regression with a failing command. The skill reproduces it, narrows two hypotheses with one deciding observation, proves the causal line, applies the minimal fix, and re-runs the reproduction and test suite as proof. ## Quick Start Ask the AI to debug the reported symptom or failing command, for example: debug why the export endpoint returns a 500 error when the date range spans a month boundary.

Frequently Asked Questions about debug

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I debug a bug without guessing at the cause?▼

Reproduce the failure as one bare command, keep two hypotheses the reproduction leaves standing, and add one observation that tells them apart. Only after evidence names one causal line with a mechanism predicting the symptom do you make the predicted fix.

How to diagnose a slowdown or memory leak from a profile?▼

Capture a CPU profile, heap snapshot, or trace with the runtime's own profiler, narrow to the hot path or retainer chain, and map the frame to file, symbol, and line. A finding without a paired before/after capture or live-instrumentation proof stays a hypothesis.

When should I not use a reproduce-first debugging workflow?▼

Skip the investigation loop when a diagnostic's file, line, and symbol already match the source holding the invalid symbol, or when a user-stated cause checks out against the code and predicts the reproduction. Feature complaints also fall outside its scope.

What happens if the bug cannot be reproduced?▼

No production edit is made without a reproduction. The deliverable becomes reproduction evidence: the skill reports the attempted reproduction, both standing hypotheses, and every observation, then stops rather than guessing.

Why does debugging stop after two failed isolation rounds?▼

Two rounds that leave both hypotheses standing mean a third round is guessing. The skill reports the reproduction, both hypotheses, and all observations, and stops instead of patching blindly.