systematic-debugging

name	systematic-debugging
description	Use when encountering any bug, test failure, unexpected behavior, or broken build — before proposing any fix

Systematic Debugging

Overview

Random fixes waste time and introduce new bugs. Quick patches hide root causes.

Core principle: Find the root cause before touching anything. A fix without a root cause is a guess.

Announce at start: "I'm using the systematic-debugging skill for this issue."

---

The Iron Law

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST

If you haven't completed Phase 1, you cannot propose a fix. Not even a "small" one.

---

When to Use

Use for ANY technical issue:

• Test failures
• Bugs in production or development
• Unexpected behavior or wrong output
• Build or deploy failures
• Performance problems
• Integration issues

Use this ESPECIALLY when:

• You're under time pressure (emergencies make guessing tempting — don't)
• "Just one quick fix" seems obvious
• You've already tried a fix and it didn't work
• You don't fully understand why it's failing
• The same bug keeps coming back

Don't skip when:

• Issue seems simple (simple bugs have root causes too)
• You're in a hurry (systematic is faster than thrashing — always)
• Someone wants it fixed immediately (guessing takes longer than investigating)

---

The Four Phases

You MUST complete each phase before proceeding to the next.

---

Phase 1: Root Cause Investigation

Before attempting ANY fix:

1. Read error messages completely

   • Don't skim — read the full stack trace
   • Note exact error text, line numbers, file paths, error codes
   • Errors often contain the exact fix — read them before doing anything else

2. Reproduce it consistently

   • Can you trigger it reliably?
   • What are the exact steps?
   • Does it always fail, or only sometimes?
   • If not reproducible → gather more data first. Never guess on a flaky issue.

3. Check what changed

   • What was the last thing that changed before it broke?
   • git diff, recent commits, new dependencies, config changes
   • Environment differences (local vs CI vs prod)?
   • New env vars added or removed?

4. Instrument multi-component systems

   When the system has multiple layers (API → service → database, CI → build → deploy):

   Don't guess which layer — instrument each boundary:

   For each component boundary:
     - Log what enters the component
     - Log what exits the component
     - Verify config/env propagation at each layer
   
   Run once → gather evidence → identify which layer breaks
   Then investigate that specific layer
   

   Example for a failing deployment pipeline:

   # Layer 1: env vars present?
   echo "API_KEY: ${API_KEY:+SET}${API_KEY:-MISSING}"
   
   # Layer 2: passed to build script?
   env | grep API_KEY || echo "NOT in build environment"
   
   # Layer 3: available at runtime?
   curl /api/health | jq '.env.API_KEY'
   

   This tells you exactly where it breaks — don't fix before running this.

5. Trace the data flow

   • Where does the bad value originate?
   • What called this function with the bad input?
   • Keep tracing backward through the call stack until you find the source
   • Fix at the source — never at the symptom

Gate: You must be able to state: "The root cause is X because Y." If you can't, keep investigating.

---

Phase 2: Pattern Analysis

Find the pattern before forming a hypothesis:

1. Find working examples

   • Where does similar code work correctly in the same codebase?
   • What's different between the working version and the broken version?

2. Read references completely

   • If implementing a pattern (auth, pagination, webhooks), read the reference implementation in full
   • Don't skim and adapt — partial understanding guarantees bugs
   • Understand the pattern before applying it

3. List every difference

   • Working vs broken — list ALL differences, however small
   • Never assume "that can't matter" — it often does
   • Differences in: types, config keys, initialization order, async vs sync, env context

4. Map dependencies

   • What does this component assume about its environment?
   • What other components does it depend on?
   • What breaks if one of those assumptions is wrong?

Gate: You must be able to state: "The working version does X. The broken version does Y. The difference is Z."

---

Phase 3: Hypothesis and Testing

Scientific method — one variable at a time:

1. Form a single hypothesis

   • State it clearly: "I think X is the root cause because Y"
   • Write it down before testing
   • Be specific — "something in auth" is not a hypothesis

2. Test minimally

   • Make the SMALLEST possible change to test the hypothesis
   • One variable at a time — never bundle changes
   • If the test is destructive or irreversible, stage it first

3. Verify before continuing

   • Did the hypothesis hold? Yes → move to Phase 4
   • No → form a NEW hypothesis from the new evidence
   • Never stack fixes on top of a failed hypothesis

4. When you genuinely don't know

   • Say so: "I don't understand why X is happening"
   • Don't pretend — guessing confidently is worse than admitting uncertainty
   • Gather more evidence or ask for help

Gate: Hypothesis confirmed with minimal test. One hypothesis per test cycle.

---

Phase 4: Implementation

Fix the root cause, not the symptom:

1. Write a failing test case first

   • Simplest possible reproduction of the bug
   • Automated test if the project has a test framework
   • One-off script if not
   • This test must fail before your fix — that's how you know you're fixing the right thing

2. Implement one fix

   • Fix ONLY the root cause identified
   • One change at a time
   • No "while I'm here" improvements
   • No bundled refactoring

3. Verify the fix

   • Your test passes now?
   • No other tests broken?
   • Issue actually gone — not just hidden?

4. If the fix doesn't work

   • STOP
   • Count how many fixes you've tried
   • Under 3 → return to Phase 1 with the new evidence
   • 3 or more → STOP. The architecture is wrong. Don't attempt another fix.

5. If 3+ fixes have failed: question the architecture

   Signs you're fighting the wrong battle:

   • Each fix reveals a new problem in a different place
   • Fixing one thing breaks another
   • The fix requires "massive refactoring" to even attempt

   Stop fixing. Question fundamentals:

   • Is this pattern fundamentally sound?
   • Are we stuck with a broken approach out of inertia?
   • Should we refactor the architecture instead of patching symptoms?

   Discuss with the user before attempting anything else.

Gate: Test written → passes after fix → no regressions → can explain the fix in one sentence.

---

Red Flags — STOP and Return to Phase 1

If you catch yourself thinking any of these:

• "Quick fix for now, investigate later"
• "Just try changing X and see if it works"
• "It's probably X, let me fix that"
• "I'll add a few changes and run tests"
• "Skip the test, I'll verify manually"
• "I don't fully understand but this might work"
• "One more fix attempt" (after 2+ failures)
• Proposing solutions before tracing the data flow
• Listing fixes before completing Phase 1
• Each fix is revealing a new problem in a different place

ALL of these mean: STOP. Return to Phase 1.

3+ fixes failed? Don't fix again — question the architecture.

---

User Signals You're Doing It Wrong

Watch for these — they mean you're guessing instead of investigating:

User says	What it means
"Is that actually happening?"	You assumed without verifying
"Did you check the logs?"	You proposed a fix without gathering evidence
"Stop guessing"	You're cycling through fixes without root cause
"Why is it doing that?"	You explained the symptom but not the cause
"We're still stuck?"	Your approach isn't working — stop and restart Phase 1

When you see these: STOP. Return to Phase 1.

---

Common Rationalizations

Excuse	Reality
"Issue is simple, don't need process"	Simple bugs have root causes too. Process is fast for simple bugs.
"Emergency, no time for process"	Systematic debugging is FASTER than thrashing. Always.
"Just try this first, then investigate"	First fix sets the pattern. Do it right from the start.
"I'll write the test after the fix"	Untested fixes don't stick. Test first — it proves you fixed the right thing.
"Multiple changes at once saves time"	Can't isolate what worked. Guarantees new bugs.
"I skimmed the reference, I get the pattern"	Partial understanding guarantees bugs. Read it fully.
"I can see the problem"	Seeing the symptom ≠ understanding the root cause.
"One more fix attempt" (after 2+ failures)	3 failures = architectural problem. Stop fixing, question design.

---

Quick Reference

Phase	What you do	Gate
1. Root Cause	Read errors, reproduce, check changes, instrument layers, trace data flow	Can state: "Root cause is X because Y"
2. Pattern	Find working examples, read references, list differences	Can state: "Working does X. Broken does Y. Diff is Z."
3. Hypothesis	Form one hypothesis, test minimally, verify	Hypothesis confirmed with evidence
4. Implementation	Write failing test, single fix, verify, commit	Test passes, no regressions, one-sentence explanation

---

When Investigation Reveals "No Root Cause"

If systematic investigation shows the issue is truly environmental, timing-dependent, or external:

1. Document what you investigated and what you ruled out
2. Implement appropriate handling: retry logic, timeout, clear error message
3. Add logging/monitoring so the next occurrence has more data
4. Note it in learning-log.md for future reference

But: 95% of "no root cause" cases are incomplete investigation.

---

Supporting Techniques

Root cause tracing — when error is deep in a call stack:

• Start at the error site
• Ask: what called this? What value did it pass?
• Trace backward until you find where the bad value was introduced
• Fix there — not at the error site

Defense in depth — after finding root cause:

• Fix the source
• Add validation at the boundary where bad data entered
• Add a test that would have caught it earlier

Condition-based waiting — for timing/async issues:

• Replace arbitrary sleep() calls with polling for a specific condition
• while not ready: check_condition() beats sleep(2) every time

---

Related Skills

• continuous-improve — wraps this skill in the full Research → Plan → Execute → Verify → Reflect loop
• Any test-writing skill in your agent's library — use for Phase 4 test creation