AI Debug

Debug

Purpose

Systematic debugging skill. Four phases, always in order. NEVER fix symptoms -- always find and fix the root cause. After 2 failed fix attempts, escalate to the user.

When to Use

• Test failures (expected vs actual mismatch)
• Runtime errors (exceptions, crashes, hangs)
• Regressions (worked before, broken now)
• Unexpected behavior (no error, but wrong result)

Process

Phase 1: Symptom Analysis (WHAT, WHEN, WHERE)

Gather facts before forming hypotheses:

1. WHAT: exact error message, stack trace, log output
2. WHEN: always? intermittent? after a specific change? under load?
3. WHERE: which file, function, line? which test? which environment?
4. SINCE WHEN: git log --oneline -20 -- what changed recently?

Output: symptom report with all facts classified as KNOWN or SUSPECTED.

Phase 2: Reproduction (MINIMAL REPRO)

Make the bug reproducible with the smallest possible case:

1. Run the failing test or reproduce the error
2. If not reproducible: document exact conditions and STOP (cannot debug what cannot be reproduced)
3. Strip to minimal repro: remove unrelated code, simplify inputs, isolate the component
4. Confirm: the minimal repro fails consistently

Output: exact command to reproduce the failure.

Phase 3: Root Cause (WHY)

Apply the 5 Whys to move from symptom to cause:

1. Why does it fail? -> [immediate cause]
2. Why does that happen? -> [deeper cause]
3. Why does that happen? -> [root cause] (Continue until you reach a cause you can fix directly)

Techniques (use as appropriate):

• Binary search: comment out code, add assertions to narrow the location
• Git bisect: git bisect start HEAD <known-good> to find the breaking commit
• Print tracing: add targeted print/log statements at decision points
• Diff analysis: git diff <known-good>..HEAD -- <file> to see what changed
• Assumption check: list every assumption the code makes, verify each one

Classification: identify the root cause category:

• Logic error (wrong condition, off-by-one, missing case)
• State corruption (mutation, shared state, race condition)
• Contract violation (caller sends wrong type, missing field)
• Environment (missing dependency, wrong version, config)
• Data (unexpected input, encoding, edge case)

Output: root cause statement (1-2 sentences, specific and testable).

Phase 4: Solution Design (FIX + REGRESSION TEST)

1. Design the fix: minimal change that addresses the root cause
   • Fix the ROOT CAUSE, not the symptom
   • One logical change only
   • If the fix is large, the root cause analysis may be wrong -- revisit Phase 3
2. Write regression test: a test that fails without the fix and passes with it
3. Apply the fix
4. Verify: regression test passes AND all existing tests pass
5. Check for siblings: does the same bug pattern exist elsewhere? (grep for similar code)

Escalation Protocol

Never retry the same approach. Never loop silently.

5 Whys Example

Symptom: test_parse_config_handles_empty fails with KeyError
Why 1: config["database"] raises KeyError
Why 2: parse_config returns empty dict when file is empty
Why 3: the YAML parser returns None for empty files, not empty dict
Root cause: missing None -> {} coercion after yaml.safe_load()
Fix: add `config = yaml.safe_load(f) or {}` instead of `config = yaml.safe_load(f)`

Common Mistakes

• Fixing the symptom (add a try/except) instead of the root cause
• Not writing a regression test for the fix
• Guessing without reproducing first
• Changing multiple things at once (change one thing, verify, repeat)
• Retrying the same approach that already failed
• Not checking for sibling bugs (same pattern elsewhere)

Integration

• Called by: /ai-dispatch (debug tasks), ai-build agent (when tests fail), user directly
• Calls: test runners (to reproduce), /ai-test (regression test)
• Transitions to: ai-build (fix implementation), /ai-commit (after verified fix)

$ARGUMENTS