Agent Skills: Diagnose — Structured Debugging

Diagnose — Structured Debugging

Debug methodically. No guessing. No "let me try changing this and see." Follow the state machine: reproduce -> minimize -> hypothesize -> instrument -> fix -> regression-test.

The One Rule

If you have a fast, deterministic, agent-runnable pass/fail signal for the bug, you will find the cause. Bisection, hypothesis-testing, and instrumentation all consume that signal. If you don't have one, no amount of staring at code will save you. Spend disproportionate effort building the feedback loop.

Phase 1 — Build a Feedback Loop

Before anything else, establish a way to trigger the bug on command. Try these in order, from best to acceptable:

1. Failing test (unit, integration, or e2e) — ideal; fast, isolated, survives the fix
2. Curl/HTTP script against a running dev server
3. CLI invocation with fixture input, diff against known-good output
4. Headless browser script (Playwright/Puppeteer)
5. Replay a captured trace (HAR file, recorded request)
6. Throwaway harness — minimal script that exercises just the broken path
7. Property/fuzz loop — when the trigger is non-obvious
8. Bisection harness — git bisect run <script> to find the breaking commit
9. Differential loop — compare output of two versions (before/after, branch/main)
10. Manual HITL script — bash script that sets up state and tells you what to check (last resort)

Once you have a loop, optimize it: Can it run faster? Is the signal sharper? Is it deterministic?

Phase 2 — Reproduce

Run the feedback loop. Confirm:

• [ ] The bug appears
• [ ] It matches the user's description
• [ ] It's reproducible (runs 3+ times with same result)

If the bug is non-deterministic: Your goal is NOT to fix it yet. Your goal is to raise the reproduction rate until it's debuggable. Try: increase load, tighten timing, add contention, remove caching, pin to a single thread.

If reproduction fails entirely: Stop and say so explicitly. List what you tried. Ask the user for:

• Access to whatever environment reproduces it
• A captured artifact (HAR file, log dump, core dump, screen recording)
• Permission to add temporary instrumentation to the environment that reproduces it

Do not proceed to hypothesize without a working feedback loop.

Phase 3 — Minimize

Shrink the reproduction to the smallest possible case:

• Remove unrelated setup code
• Simplify inputs to the minimal trigger
• Isolate the failing path from the rest of the system

A minimal reproduction makes the cause obvious more often than you'd expect.

Phase 4 — Hypothesize

Generate 3-5 ranked hypotheses BEFORE changing any code. For each:

• State the hypothesis clearly: "The bug is caused by X"
• State a falsifiable prediction: "If X is the cause, then doing Y will make the bug disappear"
• Rank by likelihood based on the evidence so far

Do not skip to fixing. You may be right about hypothesis #1. You may be wrong. Prove it first.

Phase 5 — Instrument

Test your top hypothesis. One variable at a time.

Prefer in this order:

1. Debugger breakpoint — inspect state directly
2. Targeted log/print — one or two specific values at the suspected failure point
3. Assertion — add a runtime check that will fail loudly if your hypothesis is wrong
4. Broader logging — only if targeted approaches aren't conclusive

Tag all debug instrumentation with a unique prefix (e.g., [DEBUG-x7f2]) so you can find and remove every piece of it later.

If hypothesis #1 is disproven, move to #2. If all hypotheses are disproven, return to Phase 4 with new information from the instrumentation.

Phase 6 — Fix

Now — and only now — modify production code.

1. Write a regression test FIRST — turn the minimal reproduction from Phase 3 into a failing test
2. Run the test, confirm it fails
3. Apply the fix — the smallest change that addresses the root cause
4. Run the regression test, confirm it passes
5. Run the full test suite, confirm nothing else broke

The fix should address the root cause, not the symptom. If the bug is "null pointer in function X," the fix isn't a null check in X. It's figuring out why X received null and preventing that upstream.

Phase 7 — Cleanup

• Remove ALL debug instrumentation (search for your [DEBUG-...] tag)
• Remove any temporary test harnesses that aren't part of the permanent suite
• If the bug revealed an architectural problem, note it for the user — suggest running /review-architecture

When to Escalate

Escalate to the user (don't keep flailing) when:

• Reproduction fails after exhausting Phase 1 options
• All hypotheses are disproven and you have no new theories
• The fix requires changing a system you don't have context for
• The root cause is in a dependency or third-party service
• The bug is environment-specific and you can't reproduce the environment

Escalation is the right move, not a failure. State what you know, what you tried, and what you need.