Systematic debugging

Systematic Debugging

A discipline for hard bugs and performance regressions. Random fixes waste time and create new bugs; quick patches mask underlying issues. Skip phases only when you can explicitly justify it.

When exploring the codebase, read CONTEXT.md (if it exists) to build a clear mental model of the relevant modules, and check ADRs in the area you're touching.

Violating the letter of this process is violating the spirit of debugging.

The Two Iron Laws

These two gates are sequential, not contradictory — you must pass through both, in order:

1. No hypothesis without a feedback loop
2. No fix without root cause

• Law 1 (Phase 1): You may not start theorising about the cause until you have a tight, red-capable feedback loop — one command that catches this bug. Jumping straight to a hypothesis is the exact failure this skill prevents.
• Law 2 (Phases 3–5): You may not propose or apply a fix until you have traced the root cause. Symptom fixes are failure.

If you catch yourself reading code to build a theory before a feedback loop exists, stop — you have violated Law 1. If you catch yourself proposing a fix before you understand why the bug happens, stop — you have violated Law 2.

Phase 0 — Orient

Get your bearings before Phase 1 (read errors, check recent changes, read CONTEXT.md) — see examples.md for the mechanics. Orientation does not entitle you to a fix.

Phase 1 — Build a feedback loop

This is the skill. Everything else is mechanical. If you have a tight pass/fail signal for the bug — one that goes red on this bug — you will find the cause; bisection, hypothesis-testing, and instrumentation all just consume it. If you don't have one, no amount of staring at code will save you.

Spend disproportionate effort here. Be aggressive. Be creative. Refuse to give up.

The ten ways to construct a loop (failing test, curl, CLI, headless browser, replay, throwaway harness, fuzz loop, bisection harness, differential loop, HITL script), the sub-technique for locating where a multi-component bug breaks, and tightening/non-deterministic-bug guidance: see examples.md.

Build the right feedback loop, and the bug is 90% fixed.

When you genuinely cannot build a loop

Stop and say so explicitly. List what you tried. Ask the user for: (a) access to whatever environment reproduces it, (b) a captured artifact (HAR file, log dump, core dump, screen recording with timestamps), or (c) permission to add temporary production instrumentation. Do not proceed to hypothesise without a loop.

Completion criterion — a tight loop that goes red

Phase 1 is done when the loop is tight and red-capable: you can name one command — a script path, a test invocation, a curl — that you have already run at least once (paste the invocation and its output), and that is:

• Red-capable — it drives the actual bug code path and asserts the user's exact symptom, so it can go red on this bug and green once fixed. Not "runs without erroring" — it must be able to catch this specific bug.
• Deterministic — same verdict every run (flaky bugs: a pinned, high reproduction rate, per above).
• Fast — seconds, not minutes.
• Agent-runnable — you can run it unattended; a human in the loop only via scripts/hitl-loop.template.sh.

If you catch yourself reading code to build a theory before this command exists, stop — jumping straight to a hypothesis is the exact failure this skill prevents (Iron Law 1). No red-capable command, no Phase 2.

Phase 2 — Reproduce + minimise

Run the loop. Watch it go red — the bug appears.

Confirm:

• The loop produces the failure mode the user described — not a different failure that happens to be nearby. Wrong bug = wrong fix.
• The failure is reproducible across multiple runs (or, for non-deterministic bugs, reproducible at a high enough rate to debug against).
• You have captured the exact symptom (error message, wrong output, slow timing) so later phases can verify the fix actually addresses it.

Minimise

Once it's red, shrink the repro to the smallest scenario that still goes red. Cut inputs, callers, config, data, and steps one at a time, re-running the loop after each cut — keep only what's load-bearing for the failure.

Why bother: a minimal repro shrinks the hypothesis space in Phase 3 (fewer moving parts left to suspect) and becomes the clean regression test in Phase 5.

Done when every remaining element is load-bearing — removing any one of them makes the loop go green.

Do not proceed until you have reproduced and minimised.

Phase 3 — Hypothesise

Generate 3–5 ranked hypotheses before testing any of them. Single-hypothesis generation anchors on the first plausible idea.

Each hypothesis must be falsifiable: state the prediction it makes.

Format: "If <X> is the cause, then <changing Y> will make the bug disappear / <changing Z> will make it worse."

If you cannot state the prediction, the hypothesis is a vibe — discard or sharpen it.

Census discipline — never truncate evidence you intend to count

A command whose output feeds a completeness claim must not be truncated. | head, | tail, -m, and a query limit silently turn all into some, and the rows they drop are invisible in the result you paste.

• Print the count first (grep -c, count(*)), then the rows, and state the count in the finding.
• If you truncate for readability, you forfeit the words every, all, only, none, the N writers until you re-run unbounded.

This bites hardest when enumerating the writers of a value, the callers of a function, or the consumers of a field — exactly the lists a fix's completeness rests on. A fix that routed six of nine writers is not a partial fix; it is a bug report you wrote about yourself and then closed.

Pattern-analysis techniques for generating and sharpening hypotheses, and presenting the ranked list to the user: see patterns.md. For bugs deep in the call stack, trace the bad value backward to its origin — see root-cause-tracing.md in this directory for the complete backward-tracing technique. Fix at the source, not at the symptom.

Phase 4 — Instrument

Each probe must map to a specific prediction from Phase 3. Change one variable at a time.

Instrumentation tool preference (debugger over logs, never "log everything") and the perf-regression measure-first branch: see examples.md.

Tag every debug log with a unique prefix, e.g. [DEBUG-a4f2]. Cleanup at the end becomes a single grep. Untagged logs survive; tagged logs die.

When the bug involves waiting on a condition (timeouts, races, "it works if I add a sleep"), replace arbitrary timeouts with condition polling — see condition-based-waiting.md in this directory.

Phase 5 — Fix + regression test

Verify the claims that justify the fix shape

A fix's shape is usually decided by a claim that some other code depends on the current arrangement: "keeping these two values apart is what lets X compare them", "removing this would blind the Y check", "Z reads this field". That claim is the difference between re-pointing one line and adding a field, a second writer, and a paired change in another service.

Every such claim names a consumer. Go read the consumer, and state the file and line where it reads the value. If you cannot name one, the claim is void — and so is the fix it was protecting.

Do this even when the claim arrives as settled. These claims are where inherited analysis fails, for three reasons:

• They are negative claims. "Nothing else reads this" and "this is what X compares" cannot be settled by looking at the value being written. Reading the writer proves nothing; only the reader decides.
• The consumer lives outside what you are changing — another service, another repo, the schema. Checking a fix's scope keeps you in code you already have open; checking its justification does not, so it quietly gets skipped.
• They come pre-stamped as verified by a ticket, a prior PR, a commit message, a passing test's name, or your own earlier summary. A test proving a value is written or sent proves nothing reads it.

Cheapest sufficient check: grep the whole system for the field name, read every hit, and confirm the schema has the column at all. A justification that survives that is real; one that does not was holding up the wrong design.

The same rule applies to a claim someone else makes. A review comment, ticket, or inherited summary asserting "this flow works today", "an admin can do X", or "nothing else reads this" is a claim about behaviour, and it arrives pre-stamped as authoritative — which is exactly why it skips the check. Go make the behaviour happen before you act on the finding; the test you would write to lock it down is the check. A finding whose facts are all correct and whose premise is wrong will send you to fix something that was never broken, and the fix will look reasonable to everyone.

Changing a guard: measure what it gates

Before altering a conditional that decides whether a write happens, count the records that pass it now and the records that would pass after. A guard's behaviour is a property of the data flowing through it, not of the expression.

An expression that reads as obviously wrong can be load-bearing for most of the fleet, and an obviously-better replacement can silently stop work for a large population. Report both numbers; if you cannot obtain them, say the change is unmeasured rather than letting the cleaner expression imply it is safer.

Write the regression test before the fix — but only if there is a correct seam for it.

A correct seam is one where the test exercises the real bug pattern as it occurs at the call site. If the only available seam is too shallow (single-caller test when the bug needs multiple callers, unit test that can't replicate the chain that triggered the bug), a regression test there gives false confidence.

If no correct seam exists, that itself is the finding. Note it. The codebase architecture is preventing the bug from being locked down. Flag this for Phase 6.

Reporting the confirmed root cause to the user with dmi-superpowers:say (the full four-beat — this is a finding the user must understand and decide on, not status narration): see patterns.md.

If a correct seam exists:

1. Turn the minimised repro into a failing test at that seam.
2. Watch it fail.
3. Apply the fix — address the root cause, ONE change at a time. No "while I'm here" improvements, no bundled refactoring.
4. Watch it pass.
5. Re-run the Phase 1 feedback loop against the original (un-minimised) scenario.

After finding the root cause, consider whether validation belongs at more than one layer — see defense-in-depth.md in this directory.

If the fix doesn't work — count your attempts

• STOP. Count: how many fixes have you tried?
• If < 3: return to Phase 3, re-rank hypotheses with the new information. Don't pile fixes on top.
• If ≥ 3: STOP and question the architecture. Do not attempt fix #4 without an architectural discussion.

Pattern indicating an architectural problem:

• Each fix reveals new shared state / coupling / problem in a different place.
• Fixes require "massive refactoring" to implement.
• Each fix creates new symptoms elsewhere.

Then question fundamentals: Is this pattern fundamentally sound? Are we sticking with it through sheer inertia? Should we refactor the architecture instead of continuing to fix symptoms? Discuss with your human partner before attempting more fixes. This is NOT a failed hypothesis — this is a wrong architecture.

Phase 6 — Cleanup + post-mortem

Required before declaring done:

• Original repro no longer reproduces (re-run the Phase 1 loop)
• Regression test passes (or absence of a correct seam is documented)
• All [DEBUG-...] instrumentation removed (grep the prefix)
• Throwaway prototypes deleted (or moved to a clearly-marked debug location)
• The hypothesis that turned out correct is stated in the commit / PR message — so the next debugger learns

Then ask: what would have prevented this bug? If the answer involves architectural change (no good test seam, tangled callers, hidden coupling), hand off to the dmi-superpowers:improve-codebase-architecture skill with the specifics. Make the recommendation after the fix is in, not before — you have more information now than when you started.

Guardrails (always-on)

These apply at every phase, not just at the end.

Red Flags — STOP and follow the process

If you catch yourself thinking:

• "Quick fix for now, investigate later"
• "Just try changing X and see if it works"
• "Add multiple changes, run tests"
• "Skip the test, I'll manually verify"
• "It's probably X, let me fix that"
• "I don't fully understand but this might work"
• "Pattern says X but I'll adapt it differently"
• "This separation is what lets X compare Y" (naming a consumer you have not opened)
• "Every writer now goes through…" (from a search you piped through head)
• "The reviewer says this path works, so I'll protect it" (never ran the path)
• "Removing this would blind Z" (never read Z)
• "Here are the main problems: [lists fixes without investigation]"
• Proposing solutions before building a feedback loop or tracing data flow
• "One more fix attempt" (when already tried 2+)
• Each fix reveals new problem in a different place

ALL of these mean: STOP. Return to Phase 1 (no loop yet) or Phase 3 (loop exists, no confirmed root cause).

If 3+ fixes failed: question the architecture (see Phase 5).

Your human partner's signals you're doing it wrong

Watch for these redirections:

• "Is that not happening?" — you assumed without verifying
• "Are we sure it reads that?" — you took a consumer claim on trust
• "Will it show us...?" — you should have added evidence gathering
• "Stop guessing" — you're proposing fixes without understanding
• "Ultra-think this" — question fundamentals, not just symptoms
• "We're stuck?" (frustrated) — your approach isn't working

When you see these: STOP. Return to Phase 1.

The Common Rationalizations catalog and the "no root cause" edge case (95% of those are incomplete investigation): see patterns.md.

Supporting files

In this directory:

• examples.md — orientation mechanics, ways to construct a feedback loop, locating where a multi-component bug breaks, tightening the loop, and instrumentation tool preference.
• patterns.md — pattern-analysis hypothesis generation, reporting practices, the common-rationalizations catalog, and the "no root cause" edge case.
• root-cause-tracing.md — trace bugs backward through the call stack to the original trigger (Phase 3).
• condition-based-waiting.md (+ condition-based-waiting-example.ts) — replace arbitrary timeouts with condition polling (Phase 4).
• defense-in-depth.md — add validation at multiple layers after finding the root cause (Phase 5).
• scripts/hitl-loop.template.sh — structured human-in-the-loop reproduction harness (Phase 1, last resort).
• find-polluter.sh — locate a test that pollutes shared state for a later test.

Related skills:

• dmi-superpowers:test-driven-development — for creating the failing regression test (Phase 5).
• dmi-superpowers:verification-before-completion — verify the fix worked before claiming success.
• dmi-superpowers:improve-codebase-architecture — hand-off target when the post-mortem reveals an architectural cause (Phase 6).