cc-debugging

Most debugging time goes to finding and understanding the defect; the fix is usually obvious once you understand it. Every action tests a hypothesis — no guessing, no random changes.

First action on any bug: run the failing test or repro and read its actual output. Do this before reading the source in depth and before editing anything — the observed failure, not the code you infer it from, is what you debug. If the usual runner is unavailable (e.g. no pytest), fall back to a runnable form (python3 -c, a direct script) and capture that output now, not after a fix.

Shared numeric thresholds (20:1 debugger variation, ~50% of fixes wrong first time): Read(${CLAUDEPLUGINROOT}/references/cc-foundations.md).

Scientific Debugging Method

STABILIZE → LOCATE → HYPOTHESIZE → EXPERIMENT → FIX → TEST → SEARCH

Step 1: STABILIZE

Get a reliable reproduction — you cannot debug what you cannot reproduce.

Precondition on editing: the first tool action of the session is running the failing test or repro, and its output is captured, BEFORE any Edit to implementation code. The order is run-test → then edit, never edit-then-test. If the standard runner is missing, run the repro another way (python3 -c, a direct script) and capture that — the missing runner does not excuse skipping the observed failure.

• Reduce to the smallest case that still fails.
• Intermittent failures are usually initialization errors, timing issues, or dangling pointers.
• Record the exact conditions: inputs, environment, order of operations.

Step 2: LOCATE

Narrow the suspicious region before forming a hypothesis.

• Binary search: disable code sections until the failure disappears → the bug is in what you removed.
• Check recently changed code first (defects cluster around changes).
• Check modules with prior defect history (defects cluster by module).
• Look for patterns: specific data? specific user? specific environment?

Step 3: HYPOTHESIZE

Form a specific, testable hypothesis — not "the bug is somewhere in module X."

• Good: "The counter isn't reset between requests because X shares state with Y."
• Use all available data: logs, stack traces, variable values, test outputs.
• One hypothesis at a time; rank competing candidates; brainstorm alternatives before committing.

Step 4: EXPERIMENT

Design a test that will disprove the hypothesis, not confirm it.

• Add targeted logging or assertions at the suspected site, or write a failing test that would pass if the hypothesis holds.
• Observe before changing production code.
• Record results; update or discard the hypothesis based on what you see.

Step 5: FIX

Fix the root cause, not the symptom.

• Understand the program vicinity (hundreds of lines, not just the bug line).
• Rule out competing hypotheses before committing to the diagnosis.
• Make one change at a time; keep the original source.

Step 6: TEST

Verify the fix actually works.

• Triangulate: multiple different test cases, not just the original repro.
• Add a regression test that would have caught this bug.
• Run the full test suite and report the result.

Step 7: SEARCH

Defects cluster — if this bug existed, similar ones likely exist nearby.

Precondition on completing: before reporting the fix complete, a search for the same defect pattern (grep/Glob) has been run and its result recorded.

• Search for the same pattern elsewhere in the codebase.
• Check the module's other methods for similar logic.
• Check other code from the same author or era.

Common Defects Quick Check

Rule these out before deep investigation:

• Off-by-one: loop bounds (< vs <=), array index vs length
• Null / undefined dereference before checking
• Race condition (intermittent, timing-dependent)
• Uninitialized variable
• Incorrect operator precedence (add explicit parentheses)
• Floating-point equality (== instead of epsilon comparison)
• Resource leak: file handle, connection, or lock not released on an error path
• Logic inversion: wrong branch taken

When quick checks fail and the systematic method stalls, the brute-force techniques (full code review, isolate in a harness, rewrite the section) and the full defect catalog are in Read(${CLAUDESKILLDIR}/checklists.md).

Confessional debugging

Explain the problem out loud, to a person or a rubber duck. Articulation frequently reveals the bug before the listener responds.

Chain