Agent Skills: Systematic Debugging

Systematic Debugging

When to Use

• A bug, error, exception, or crash needs investigation
• Something is "not working" and the cause is unclear
• A test is failing and the reason isn't obvious
• Unexpected behavior needs troubleshooting in any language or framework

Core Workflow

Follow these five phases sequentially. Do not skip ahead to fixing before completing isolation and tracing.

Phase 1: Reproduce

Establish a reliable way to trigger the bug before doing anything else.

1. Read the full error message, stack trace, and logs — note exact text, line numbers, and error codes
2. Create a minimal reproduction case that triggers the issue consistently
3. Record the exact steps, inputs, and environment that cause the failure

Checkpoint: Can you trigger the bug on demand? If intermittent, gather more data before proceeding.

Phase 2: Isolate

Narrow down where the failure originates.

1. Use binary search to find the failing component — disable or stub out halves of the system
2. Check recent changes with git log --oneline -20 and git diff against the last known good state
3. Add targeted logging or use a debugger to observe state at key boundaries

# Find which commit introduced the bug
git bisect start
git bisect bad HEAD
git bisect good <last-known-good-commit>
# Git will checkout midpoints — test each one and mark good/bad

Checkpoint: The bug is traced to a specific function, module, or data flow.

Phase 3: Trace to Root Cause

Understand why the failure happens — not just where.

1. Read the code path completely from entry point through the failure site
2. Check assumptions: what does each function expect vs. what it actually receives?
3. Trace data flow backward — where does the bad value originate?
4. Verify with evidence: add assertions or print statements to confirm your hypothesis

# Example: verify assumptions about incoming data
def process_order(order):
    assert order.status == "pending", f"Expected pending, got {order.status}"
    assert order.items, "Order has no items"
    # ... rest of processing

Checkpoint: The chain of causation from trigger to symptom is explained, with supporting evidence (logs, assertions, debugger output).

Phase 4: Fix at Root Cause

Apply a targeted fix that addresses the actual cause, not just the symptom.

1. Fix the root cause, not a downstream effect
2. Keep the fix minimal — change only what's necessary
3. Avoid "band-aid" fixes that mask the underlying problem (e.g., adding a try/except around a crash without fixing why it crashes)

Phase 5: Verify

Confirm the fix works and doesn't introduce regressions.

1. Run the reproduction case from Phase 1 — confirm the bug is gone
2. Run the full test suite to check for regressions
3. Test edge cases related to the fix
4. If the bug was missing a test, add one that would have caught it

Checkpoint: Reproduction case passes, test suite is green, and you have a new test covering this bug.

Key Anti-Patterns to Avoid

• Shotgun debugging: making random changes hoping something works
• Fix-and-pray: applying a fix without understanding the cause
• Skipping reproduction: jumping to code changes without confirming you can trigger the bug
• Fixing symptoms: wrapping errors in try/catch instead of fixing what produces them

See anti-patterns.md for the full catalog.

Related Skills

• root-cause-tracing: Deep call-stack tracing techniques — use after Phase 2 when the bug is deep in execution chains
• verification-before-completion: Mandatory verification gates — reinforces Phase 5 before claiming a fix is complete

Deep-Dive References

• workflow.md: Detailed phase-by-phase instructions with decision trees
• examples.md: Worked debugging examples across languages
• troubleshooting.md: Common debugging scenarios and solutions
• anti-patterns.md: Patterns to avoid and how to recognize them