Debugging

Systematic Debugging

Stop guessing. Start investigating. Every bug follows a pattern - find the pattern, find the fix.

Announce: "Using systematic-debugging to investigate this issue."

Quick Start (3 Steps)

For straightforward bugs, this is often enough:

1. Read the error. Actually read it. File, line, message. Don't skip this.
2. Reproduce it. Run the failing command. See the exact output. No assumptions.
3. Trace backward. From the error line, follow the data. Where did the wrong value come from?

If these 3 steps solve it, you're done. If not, use the full system below.

The Full System: 4-Phase Cycle

Phase 1: Reproduce

Goal: See the bug happen with your own eyes.

• Run the exact command or action that triggers the bug
• Record the full error output (not a summary - the actual output)
• Note what you expected vs what you got
• Check: is the bug consistent or intermittent?

If you can't reproduce it, you can't fix it. Gather more context first.

Phase 2: Gather Evidence

Goal: Collect facts, not theories.

Read the relevant code. Don't guess what it does - read it.

Evidence Checklist:
- [ ] Full error message and stack trace
- [ ] The source code at the error location
- [ ] Recent changes to related files (git log --oneline -10 -- <file>)
- [ ] Input data that triggers the bug
- [ ] What the code should do vs what it actually does

Rule: Observe before editing. If you haven't read the code, you don't understand the bug. If you don't understand the bug, your fix will create the next bug.

Phase 3: Form Hypotheses

Goal: Generate 2-3 possible explanations. Then test them.

For each hypothesis:

1. State it clearly: "I think X fails because Y"
2. Define a test: "If this hypothesis is correct, then Z should be true"
3. Run the test
4. Record the result: confirmed, refuted, or inconclusive

Don't stop at the first hypothesis that feels right. The most obvious explanation is wrong about 40% of the time.

Phase 4: Fix and Verify

Goal: Make the smallest change that fixes the bug. Verify it works.

1. Make the fix
2. Run the reproduction case - does it pass now?
3. Run the full test suite - did you break anything else?
4. Explain why the fix works (if you can't explain it, you didn't fix it)

Done. If the fix doesn't work, go back to Phase 2 with new evidence.

ACH Method (Complex Bugs)

When a bug survives the 4-phase cycle, use Analysis of Competing Hypotheses:

Step 1: List All Hypotheses

Write down every possible explanation, including ones that seem unlikely:

H1: The database query returns stale cached data
H2: The auth token expires between request and response
H3: A race condition in the async handler
H4: The input validation silently drops a required field

Step 2: Build an Evidence Matrix

Evidence	H1 (Cache)	H2 (Token)	H3 (Race)	H4 (Validation)
Bug is intermittent	Consistent	Consistent	Supports	Inconsistent
Only happens under load	Neutral	Neutral	Supports	Inconsistent
Error log shows timeout	Inconsistent	Supports	Supports	Inconsistent
Fresh deploy doesn't fix	Supports	Neutral	Neutral	Supports

Step 3: Eliminate

Remove hypotheses that are inconsistent with the evidence. The survivor is your best lead.

In the example above, H4 is inconsistent with two pieces of evidence - eliminate it first. H1 is inconsistent with one - it's less likely. H2 and H3 both survive - investigate the one that's easier to test first.

When to Use ACH vs Quick Debug

Situation	Method
Clear error message, obvious location	Quick Start (3 steps)
Known area, unclear cause	4-Phase Cycle
Multiple possible causes, intermittent	ACH Method
Bug persists after 2+ fix attempts	ACH Method

When to Use This Skill

• Test fails and you don't immediately know why
• An error message appears in logs or console
• Code behaves differently than expected
• A previously working feature breaks
• Performance degrades without obvious cause

When NOT to Use

• You already know the exact fix (just fix it)
• It's a typo or syntax error (your linter already told you)
• The "bug" is actually a missing feature (use planning instead)

Delegation Pattern

Complex bugs benefit from delegation. Score: Exploration (+3) + 3+ files (+2) = 5+. Delegate the investigation, keep the fix.

## 1. TASK
Investigate why [SYMPTOM] occurs when [TRIGGER].

## 2. EXPECTED OUTCOME
Root cause analysis: error location (file:line), why it fails, suggested fix.

## 3. REQUIRED TOOLS
Read, Grep, Glob, Bash (read-only: test output, git log)

## 4. MUST DO
- Reproduce the error by running: [COMMAND]
- Read the source at the error location
- Check git log for recent changes
- Form at least 2 hypotheses and test each

## 5. MUST NOT DO
- Don't modify any files
- Don't run the application in production mode
- Don't access external services

## 6. CONTEXT
Error: [FULL_ERROR_MESSAGE]
Relevant files: [FILE_LIST]
Last working state: [COMMIT_OR_DATE]

Agent: debugger | Model: Sonnet | Effort: Medium (6 turns)

Anti-Patterns

Pattern	Problem	Fix
"Let me just try this..."	Random changes without understanding	Phase 2 first - gather evidence
Fixing the symptom	Suppressing the error instead of the cause	Trace to the root, not the surface
Changing multiple things at once	Can't tell which change fixed it	One change, one test, one commit
Ignoring the stack trace	The answer is literally in the output	Read the error. All of it.
"It works on my machine"	Environment assumptions	Reproduce in the failing environment
Debugging in production	Risk of data loss or downtime	Reproduce locally first

For the extended guide with recovery protocols, see references/anti-patterns.md.

With Starter System

If you have the Starter System installed:

• Use /remember failure: [description] to log bugs you've seen before
• Check /remember at the start of debugging - the bug might be a known pattern
• Use /handoff to pass debugging context when you run out of session time

With Kairn (Course)

When Kairn is configured, debugging gains:

• Recall before investigating: "Have I seen this error pattern before?" - automatic lookup across past sessions
• Learn after resolution: captures root cause + fix for future sessions automatically
• Pattern recognition across past bugs - "This looks like the auth timeout issue from 3 weeks ago"

Without Kairn: each debugging session starts fresh. The 4-phase cycle still works, you just can't learn from past sessions automatically.