debug

Debug — Systematic Debugging

Iron Law

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST. Symptom fixes are whack-a-mole debugging.

Process

0. Recall

• Query the knowledge graph: epic mem search "ERROR_MESSAGE_OR_FILE" for the error message or file name
• Check if a resolution or pattern node exists for this error category
• If a past resolution exists, apply it directly instead of re-debugging from scratch

Why: Past resolutions save 10-30 minutes of re-investigation. The knowledge graph exists so you don't repeat work.

1. Observe

• Read the full error message and stack trace
• Identify: What was expected? What actually happened?
• Note the exact file, line, and function

Why: Precise observation prevents misdiagnosis. Most debug loops start from skimming the error instead of reading it fully.

2. Hypothesize

Form 2-3 possible causes. Rank by likelihood.

Why: Ranked hypotheses prevent random code changes. Even a wrong hypothesis narrows the search space.

3. Isolate

• Narrow the scope: Binary search through recent changes
• Reproduce minimally: Smallest input that triggers the error
• Check assumptions: Print/log intermediate values
• Bisect: git log --oneline -10 — when did it last work?

Why: Isolation proves the cause before you touch any code. Without it, you're guessing.

4. Fix

• Fix the root cause, not the symptom
• If the fix is a workaround, add a TODO with explanation

Why: Root cause fixes are permanent. Symptom patches rot and hide the real bug.

5. Verify

• Run the failing test — must pass now
• Run the full test suite — no regressions
• If this was a bug, the regression test stays

Why: A passing fix without verification is wishful thinking. The full suite catches side effects.

6. Record

• If the fix involved 3+ debug steps, record a resolution node via epic mem add --title "..." --type resolution --body "..."
• Include: error category, root cause, solution approach
• This enables future sessions to skip re-debugging the same class of issues

Why: Recording turns one-time debugging into institutional knowledge. Future sessions benefit from your effort.

When to Trigger

• Any test failure during /go execution
• Runtime error in tool output
• "Unexpected" in any output

Anti-Rationalization

| Excuse | Rebuttal | What to do instead |

|--------|----------|-------------------|

| "Let me just try again" | Retrying without changing anything is not debugging. | Form a new hypothesis before each attempt. |

| "It works on my machine" | Environment differences are bugs too. | Reproduce in the actual target environment first. |

| "I'll just catch the exception" | That hides the bug, doesn't fix it. | Fix the root cause. Only catch what you can handle meaningfully. |

| "It's probably a flaky test" | Flaky tests hide real bugs. Prove it's flaky before dismissing. | Run it 3x in isolation. If it fails consistently, it's real. |

| "I'll add more logging" | Logging without a hypothesis is fishing. | Hypothesize first, then add targeted logging to confirm/deny. |

| "It's probably just a typo" | Assumptions are the enemy of debugging. Reproduce first, then diagnose. | Read the full error, reproduce it, and verify before assuming simplicity. |

| "Let me just try changing this one thing" | Whack-a-mole debugging creates more bugs. Follow the hypothesis → isolate → fix cycle. | Form a hypothesis, isolate the cause, then fix with evidence. |

| "The error message tells me everything" | Error messages describe symptoms, not root causes. Investigate the chain. | Trace the error to its origin. The message is the starting point, not the answer. |

Evidence Required

Before claiming the bug is fixed, show ALL of these:

• [ ] Root cause identified: one sentence explaining WHY it broke
• [ ] Fix applied: diff showing the change
• [ ] Previously failing test now passes (show output)
• [ ] Full test suite passes (no regressions)
• [ ] If bug came from a pattern, regression test added to prevent recurrence

"It works now" without root cause = symptom patch, not a fix.

Red Flags

• Changing random things hoping something works
• More than 3 retry attempts without forming a new hypothesis
• Suppressing errors instead of fixing them