| name | dev-debug |
| description | Debug and resolve problems. Use when the user has a bug, an error, an unexpected behavior, or wants to understand why something is not working. |
| allowed-tools | ["Read","Glob","Grep","Bash"] |
| context | fork |
| model | opus |
| argument-hint | [error-description] |
Debug a Problem
Goal
Identify the root cause of a bug methodically via a systematic 4-phase approach.
Systematic Methodology (4 Phases)
┌──────────────────────────────────────────────────────────────────┐
│ SYSTEMATIC DEBUGGING │
├──────────────────────────────────────────────────────────────────┤
│ │
│ PHASE 1: OBSERVATION Collect without interpreting │
│ ═══════════════════ │
│ - Reproduce the exact symptom │
│ - Document the environment │
│ - Capture logs, stack traces, states │
│ - DO NOT jump to conclusions │
│ │
│ PHASE 2: HYPOTHESES Reason systematically │
│ ════════════════════ │
│ - List ALL possible causes │
│ - Rank by probability (high/medium/low) │
│ - Define a validation test for each hypothesis │
│ - Use the 5 Whys technique │
│ │
│ PHASE 3: INVESTIGATION Prove, do not assume │
│ ══════════════════════ │
│ - Test ONE hypothesis at a time │
│ - Use strategic tracing and logging │
│ - Isolate with binary search (code or git bisect) │
│ - Document each tested hypothesis (confirmed/refuted) │
│ │
│ PHASE 4: VERIFICATION Confirm the fix is real │
│ ════════════════════ │
│ - Reproduce the original bug (must fail without the fix) │
│ - Apply the minimal fix │
│ - Prove the bug is fixed │
│ - Verify the absence of side effects │
│ - Add a non-regression test │
│ - Defense in depth: assertions on invariants │
│ │
└──────────────────────────────────────────────────────────────────┘
Phase 1: Observation
Key questions:
- What is happening exactly?
- What should happen?
- When did it start?
- Is it 100% reproducible?
- What are the aggravating/mitigating factors?
tail -100 logs/app.log 2>/dev/null
git log --oneline -10
git log --oneline -5 -- src/path/suspect/
Phase 2: Hypotheses
Hypothesis matrix
| # | Hypothesis | Probability | Validation test |
|---|
| 1 | [Most likely] | High | [How to verify] |
| 2 | [Secondary] | Medium | [How to verify] |
| 3 | [Less likely] | Low | [How to verify] |
5 Whys technique
Problem: The application crashes at login
1. Why? -> The JWT token is invalid
2. Why? -> The token has expired
3. Why? -> The refresh token was not called
4. Why? -> The interceptor did not detect the expiration
5. Why? -> Timezone bug in the comparison
Root cause: Timezone bug in the refresh logic
Common causes by type
| Bug type | Frequent causes |
|---|
| Null/Undefined | Missing data, race condition, API changed |
| Type error | Wrong type, JSON parsing, implicit conversion |
| Off-by-one | Array index, loop, < vs <= comparison |
| Race condition | Async without await, shared state, timing |
| Memory leak | Event listeners, closures, circular references |
| Regression | Recent change, side effect, dependency update |
Phase 3: Investigation
Tracing techniques
function trace(label: string, data: unknown) {
console.log(`[TRACE:${label}]`, JSON.stringify(data, null, 2));
}
trace('INPUT', { args });
trace('STATE', { variables });
trace('OUTPUT', { result });
trace('BRANCH', { condition });
Binary search in the code
1. Comment out half of the suspect code
2. Does the bug persist?
- Yes -> The bug is in the remaining half
- No -> The bug is in the commented half
3. Repeat until isolating the exact line
Git bisect (find the faulty commit)
git bisect start
git bisect bad
git bisect good <commit>
git bisect reset
Phase 4: Verification (MANDATORY)
Prove the fix works
1. WITHOUT the fix: reproduce the bug -> failure confirmed
2. WITH the fix: same scenario -> success confirmed
3. Existing tests -> all pass
4. Non-regression test -> written and passes
5. Side effects -> verified absent
Defense in depth
function processPayment(amount: number, userId: string) {
assert(amount > 0, 'Payment amount must be positive');
assert(userId, 'User ID is required');
}
Completion checklist
[ ] Bug reproduced reliably
[ ] Root cause identified (not just the symptom)
[ ] Minimal fix applied (no opportunistic refactoring)
[ ] Non-regression test added
[ ] Existing tests pass
[ ] No side effects
[ ] Fix documentation (descriptive commit message)
Expected output
## Diagnosis: [Bug description]
### Phase 1 - Observation
**Symptom:** [What happens]
**Expected behavior:** [What should happen]
**Reproduction:** [Steps 1, 2, 3...]
### Phase 2 - Hypotheses
| # | Hypothesis | Probability | Result |
|---|------------|-------------|--------|
| 1 | [...] | High | Confirmed/Refuted |
### Phase 3 - Investigation
**Root cause:** `src/xxx.ts:42` - [Technical explanation]
**5 Whys:** [Causal chain]
### Phase 4 - Verification
- [x] Bug reproduced
- [x] Fix applied
- [x] Non-regression test added
- [x] All tests pass
- [x] No side effects
Rules
- Do not assume - verify (Phase 4 mandatory)
- One bug at a time
- Understand BEFORE fixing
- Always add a non-regression test
- Document every tested hypothesis, even refuted ones
- The fix must be MINIMAL - no opportunistic refactoring
Iron Law: No fix without root cause
IMPORTANT: NEVER propose a fix before identifying the root cause. Symptom fixes mask the real problem and create new bugs.
Red flags (rationalizations to avoid)
| Phrase | Problem |
|---|
| "Quick fix for now" | Avoids the root cause |
| "Let's just try changing X" | Guess-and-check, not systematic |
| "I don't really understand but it should work" | Blind fix |
| "It's urgent, no time to investigate" | Systematic investigation is FASTER |
Rule of 3 failures
After 3 failed fix attempts: STOP. Do not attempt a 4th fix.
- Question the basic assumptions
- Broaden the search perimeter
- Check whether the problem is architectural (not just a local bug)
- Consider
git bisect to find the faulty commit
Metrics
| Approach | Average time | 1st-attempt fix rate |
|---|
| Systematic (4 phases) | 15-30 min | ~95% |
| Random fixing (guess-and-check) | 2-3h | ~40% |