Debugging Methodology

Systematic approach to finding and fixing bugs.

When to Use This Skill

Core Principles

1. Occam's Razor - Start with the simplest explanation
2. Binary Search - Isolate the problem area systematically
3. Preserve Evidence - Understand state before making changes
4. Document Hypotheses - Track what was tried and didn't work

Debugging Workflow

1. Understand → What is expected vs actual behavior?
2. Reproduce → Can you trigger the bug reliably?
3. Locate → Where in the code does it happen?
4. Diagnose → Why does it happen? (root cause)
5. Fix → Minimal change to resolve
6. Verify → Confirm fix works, no regressions

Common Bug Patterns

Diagnose at the Failure Point — Sentinel Values and the Resource Reading

Promoted from ~/.claude/rules/diagnose-at-the-failure-point.md, whose stub keeps the two gate lines. When a tool or library reports a failure, two surface features routinely misdirect the diagnosis: the named entity in the error, and the framing you inherited going in. Both are cheap to check against reality at the exact point of failure.

1. A named entity in an error may be a sentinel/default, not a real object

An error that names something by a low / zero / default identifier — Memory page 0 doesn't exist, id 0, an empty-string name, index -1, 0.0.0.0, a null UUID — is often reporting an uninitialized or sentinel default, not a real entity that misbehaved. Before theorizing on the entity, verify its identifier came from a real success (a completed allocation / lookup / registration), not a fallback path.

Canonical break (2026-07, cubecl CUDA): couldn't find resource for that handle: Memory page 0 doesn't exist was read — by a handoff issue, the upstream tracker, and me for many iterations — as a memory-pool reclaim race corrupting "page 0". It was MemoryLocation::uninit() = {pool:0,page:0}: a failed allocation left its handle unbound at the zero default, and every downstream lookup of an unbound handle reported "page 0." Zero pool reclaims had fired. The tell was mechanical: the count of distinct "missing" handles equalled the count of allocation failures exactly.

The tell is usually a count or ratio that matches something upstream (missing-entity count == failure count) — look for it before accepting the entity as real.

2. For any exhaustion symptom, read the actual resource at the failure point

OOM / out of memory / can't allocate / ENOSPC / too many open files / pool exhausted are symptoms with three different fixes — genuine exhaustion, fragmentation, or a manager/allocator bug — and one measurement at the failure point discriminates them:

Instrument the failing call site to print this; don't infer it from a sampler taken seconds earlier (peak may be between samples). In the same case, cuMemGetInfo at the failing malloc showed 0.58 GB free of 25.2 GB — genuinely full, not fragmented, not a pool bug. That one reading ended the allocator hunt: the working set simply exceeded the card. Every competing theory (reclaim race, cursor guard, sync-before-reclaim, a resolution "lever") had already died to a measurement, not an argument.

Evidence at the failure point > the framing you were handed > a plausible mechanism. A confidently-written issue, a maintainer's stated root cause, and an error's own wording are all inputs to verify, not conclusions. When a theory and a measurement disagree, the measurement wins — re-diagnose, don't patch the theory. It bites when an error names an entity by a suspicious round/zero/default value, when the fix for a "resource exhausted" failure depends on why, and when an inherited handoff or upstream tracker already asserts a root cause. Skipping the one cheap reading trades a five-second cuMemGetInfo/df/count for a multi-iteration hunt down a mechanism that never occurred. Siblings: agent-patterns-plugin:probe-input-integrity (an id you invent), git-plugin:git-issue-scoping and git-plugin:git-upstream-fix-check (don't trust the inherited framing), agent-patterns-plugin:tool-result-traps (a well-formed output that silently doesn't match reality).

Debugging Questions

When stuck, ask:

1. What changed recently that could cause this?
2. Does it happen in all environments or just one?
3. Is the bug in my code or a dependency?
4. What assumptions am I making that might be wrong?
5. Can I write a minimal reproduction?

Effective Debugging Practices

• Targeted changes: Form a hypothesis, change one thing at a time
• Use proper debuggers: Step through code with breakpoints when possible
• Find root causes: Trace issues to their origin, fix the source
• Reproduce first: Create a minimal reproduction before attempting a fix
• Verify the fix: Confirm the fix resolves the issue and passes tests

For system-level tools (memory analysis, profiling, strace/eBPF tracing, network debugging) and language-specific debugger commands, see REFERENCE.md.