Debug Methodology

Systematic approach to debugging and problem-solving. Distilled from real production incidents and industry best practices.

⚠️ The Root Cause Imperative

Every fix MUST target the root cause. Workarounds are forbidden unless explicitly approved.

Before proposing ANY solution, pass the Root Cause Gate:

┌─────────────────────────────────────────────┐
│            ROOT CAUSE GATE                  │
│                                             │
│  1. What is the ACTUAL problem?             │
│  2. WHY does it happen? (not just WHAT)     │
│  3. Does my fix eliminate the WHY?           │
│     YES → proceed                           │
│     NO  → this is a workaround → STOP       │
│                                             │
│  Workaround test:                           │
│  "If I remove my fix, does the bug return?" │
│     YES → workaround (fix the cause instead)│
│     NO  → genuine fix ✅                    │
└─────────────────────────────────────────────┘

The 5 Whys — Mandatory for Non-Obvious Problems

Problem: API returns 524 timeout
  Why? → Cloudflare cuts connections >100s
  Why? → The API call takes >100s
  Why? → Using non-streaming request, server holds connection silent
  Why? → Code uses regular fetch, not streaming
  Fix: → Use streaming (server sends data continuously, Cloudflare won't cut)

  ❌ WRONG: Switch to faster model (workaround — avoids the timeout instead of fixing it)
  ✅ RIGHT: Use streaming API (root cause — Cloudflare needs ongoing data)

Common Workaround Traps

Problem	Workaround (❌)	Root Cause Fix (✅)
API timeout	Switch to faster model	Use streaming / fix the slow query
Data precision loss	Search by name instead of ID	Fix BigInt parsing
Search returns nothing	Try different search strategy	Fix the search implementation
Dependency conflict	Downgrade / pin version	Use correct environment (venv)
Feature doesn't work	Remove the feature	Debug why it fails

Self-check question: "Am I solving the problem, or avoiding it?"

Phase 1: STOP — Assess Before Acting

Before ANY fix attempt:

□ What is the EXACT symptom? (error message, behavior, screenshot)
□ When did it last work? What changed since then?
□ How is the service running? (process, env, startup command)

For running services:

ps -p <PID> -o command=        # How was it started?
ls .venv/ venv/ env/           # Virtual environment?
which python3 && python3 --version
which node && node --version

NEVER restart a service without first recording its original startup command.

Phase 2: Hypothesize — Form ONE Theory

Priority order:

1. Did I change something? → diff/revert first
2. Did the environment change? → versions, deps, configs
3. Did external inputs change? → API responses, data formats
4. Genuine new bug? → only after ruling out 1-3

Phase 3: Test — One Change at a Time

Change X → Test → Works? → Done
                → Fails? → REVERT X → new hypothesis

Do NOT stack changes.

Phase 4: Patch-Chain Detection

2 fix attempts failed → STOP. Revert ALL. Back to Phase 1.

You are likely:

• Fixing symptoms of a wrong fix
• In the wrong environment entirely
• Misunderstanding the architecture

Phase 5: Post-Fix Verification

After any fix, verify:

□ Does it solve the ORIGINAL problem? (not just silence the error)
□ Did I introduce new issues? (regression check)
□ Would removing my fix bring the bug back? (confirms causality)
□ Is the fix in the right layer? (not patching symptoms upstream)

Anti-Patterns

🚨 Workaround Addiction (NEW — Most Common!)

Bypassing the problem instead of fixing it. "It's slower but works" / "Use a different approach". → Ask: "Am I solving or avoiding?" If avoiding → find the real fix. → Workarounds are ONLY acceptable when: (1) explicitly approved by user, (2) clearly labeled as temporary, (3) a TODO is created for the real fix.

🚨 Drunk Man Anti-Pattern

Randomly changing things until the problem disappears. → Each change needs a hypothesis.

🚨 Streetlight Anti-Pattern

Looking where comfortable, not where the problem is. → "Is this where the bug IS, or where I KNOW HOW TO LOOK?"

🚨 Cargo Cult Fix

Copying a fix without understanding why it works. → Understand the mechanism first.

🚨 Ignoring the User

User says "it broke after you changed X" → immediately diff X. → User observations are the most valuable data.

Environment Checklist

□ Runtime: system or venv/nvm?
□ Dependencies: match expected versions?
□ Config: .env, config.json — recent changes?
□ Process manager: PM2/systemd — restart method?
□ Logs: tail -f before reproducing
□ Backup: snapshot before any change

Deployment Safety (Hardened SCP Flow)

Iron Rule: NEVER edit files directly on the server. NEVER overwrite server files without backup.

Standard deployment (every time, no exceptions):

1. PULL    scp server:/opt/apps/项目/ ./local-项目/
           (pull the files you need + related files)

2. EDIT    Make changes locally
           (complex multi-line → write full file, never sed)

3. VERIFY  node -c *.js                    # syntax check
           node -e "require('./file')"     # module load check
           (STOP if verification fails — do not proceed)

4. BACKUP  ssh server "cp file file.bak.$(date +%s)"

5. PUSH    scp ./local-file server:/opt/apps/项目/file

6. RESTART pm2 restart <app>
           (use SAME method as original — check ps/pm2 show first)

7. HEALTH  curl -s http://localhost:<port>/health
           pm2 logs <app> --lines 5 --nostream
           (if unhealthy → revert backup immediately)

Pull Scope Rules

Changing 1 file    → pull that file + its imports/importers
Changing routes    → also pull server.js (check mount points)
Changing frontend  → also pull index.html (check script tags)
Changing config    → also pull code that reads the config
Unsure what to pull → pull the whole project directory

What NOT to Do

❌ sed -i for multi-line code on server
❌ Skip node -c after editing .js
❌ pm2 restart before syntax verification
❌ Tell user to refresh before health check passes
❌ Push without backup

🚨 Server Code Modification Rules

Every code change on a server MUST be syntax-verified before restart/reload.

After editing .js files:
  □ node -c <file>                          # Syntax check
  □ node -e "require('./<file>')"           # Module load check (for route files)
  □ FAIL → DO NOT restart. DO NOT tell user to refresh. Fix first.

After editing .html files:
  □ Check critical tag closure (div/script/style)
  □ grep -c '<div' file && grep -c '</div' file   # Count match

Complex multi-line changes:
  □ Write complete file locally → scp upload
  □ NEVER use sed for multi-line code insertion (newlines get swallowed)
  □ If sed is unavoidable → verify with node -c immediately after

Restart sequence:
  □ node -c *.js passes → pm2 restart <app>
  □ Check pm2 logs --lines 5 for startup errors
  □ curl health endpoint to confirm service is up

Why:

sed -i

multi-line insertion silently corrupts JS (newlines become single line), causing syntax errors that break the entire page with no visible error to the user.

Decision Tree

Problem appears
  ├─ I just edited something? → DIFF → REVERT if suspect
  ├─ Service won't start? → CHECK startup command + env
  ├─ New error after fix? → STOP (patch chain!) → Revert → Phase 1
  ├─ User reports regression? → DIFF before/after
  ├─ Tempted to work around? → ROOT CAUSE GATE → fix the real issue
  └─ Intermittent? → CHECK logs + external deps + timing