Systematic debugging

Systematic debugging

Trigger. Something is broken and you do not yet know why: a crash, a wrong result, a hang, a flake, a test that fails only in CI, a regression after an upgrade. Also load it when you catch yourself about to change code in the hope that it helps.

Reference material (localization techniques per symptom class, intermittent-failure playbook, and the instrumentation recipes): see reference.md in this directory.

The rule

No edit before an explanation. You must be able to say this input reaches this line, which does this wrong thing, which produces this symptom — before you touch the code.

Guess-and-check feels faster and is not. Each speculative edit changes the system you are trying to understand, and a "fix" that removes a symptom without an explanation has usually moved the bug somewhere less visible.

Four phases

Do not skip forward. Each phase has an exit condition; if you cannot meet it, you are still in that phase.

Phase 1 — Reproduce

Exit condition: one command that fails, reliably, on demand.

1. Get the exact failure: full error, stack trace, exit code, wrong output. Not a paraphrase — the literal text.
2. Establish the environment: branch, commit, dependency versions, config, platform. Bugs that "only happen for them" are usually environment deltas.
3. Reduce the trigger to the smallest input and shortest command you can.
4. Run it three times. If it fails 3/3, continue. If it fails 1/3, you have an intermittent bug — see reference.md §intermittent before proceeding.

If you cannot reproduce it, say so and stop. Debugging a failure you cannot observe is writing fiction. Ask for the input, the logs, the version.

Phase 2 — Localize

Exit condition: the smallest region of code where correct input goes in and incorrect output comes out.

Bisection is the workhorse — over anything ordered:

Axis	Method
Time	git bisect between a known-good and known-bad commit
Call stack	Assert or log the invariant at the top, middle, and bottom
Input	Halve the input until the failure disappears
Config	Disable half the flags/plugins; repeat on the failing half
Data	Halve the dataset; find the minimal failing record

The discipline is the same every time: form a binary question, answer it by observation, halve the space. Ten halvings cover a thousand candidates.

Two rules that keep this honest:

• Verify the boundary. When you think you have found the transition point, confirm both sides: correct just before, incorrect just after.
• Instrument, don't infer. Print the actual value. A variable you believe holds 3 is the reason the bug survived this long.

Phase 3 — Explain

Exit condition: a written sentence that connects cause to symptom, which predicts something you have not yet tested.

Write it out:

When parse() receives a key containing :, the split on line 88 takes the first segment only, so the map is built with a truncated key, so the lookup on line 140 misses and returns the default.

Then test the prediction. A correct explanation predicts more than the one failure you have: it says which other inputs fail, and which superficially similar ones do not. Check one of each. If the prediction is wrong, your explanation is wrong — return to Phase 2.

This is the step people skip, and skipping it is how you get a fix that works on the reported input and nothing else.

Phase 4 — Fix and pin

Exit condition: a regression test observed failing before the fix and passing after.

1. Write the test first, from the explanation, and watch it fail with the real symptom — not ImportError, not a missing fixture.
2. Make the smallest change the explanation calls for. If the explanation does not call for it, it is not part of the fix.
3. Watch the test pass.
4. Run the full suite. A fix that breaks two other things is not a fix.
5. Check the explanation for siblings: if split(':') was wrong here, is the same pattern wrong in the three other places it appears?

Anti-patterns

• Shotgun editing. Changing several things at once, then testing. If it works you have learned nothing; if it does not, you have a bigger space.
• Fixing the symptom. Wrapping the crash in try/except, clamping the bad value, adding a retry. Legitimate only when the explanation says the condition is genuinely expected — and then it is not a bug fix.
• Blaming the platform. The compiler, the runtime, the standard library. Occasionally right; assume it last, and only with a minimal reproduction.
• Trusting the comment. Comments describe intent at writing time. Read the code.
• Stopping at "it stopped happening." Especially for intermittent bugs. Absence of a symptom over ten runs is weak evidence about a race.
• Rewriting to make the bug go away. Now you have an unexplained bug and a large diff.

When not to use

• The project has a symptom-indexed triage playbook for its own subsystems — use that first; it will localize faster than a generic bisection.
• You already know the cause and are just making the change.
• The task is adding new behaviour, not restoring intended behaviour — that is test-driven-development.