Reproduce. Run the failing thing yourself and capture the exact error, command, and environment. If you cannot reproduce, stop and gather the reporter's exact conditions; do not fix blind. Side observations found on the way get noted and set aside, not chased.
Trace the intended path. Before reading logs, read the code/docs to state what SHOULD happen, step by step. You cannot recognize the broken step without knowing the intended sequence.
Check live state. What is actually on disk, in the database, running as a process, deployed? A grep that finds no trace of a migration is not proof it never ran; completed one-off scripts are routinely deleted. Absence in code is not absence in reality.
Find the smoking gun. Logs, CI runs, git history. Cheapest checks first: installed version vs latest, the artifact's own declared type or config, the library's supported list. Only conclude "unsupported" or "impossible" after checking the declared capability directly.
Prove the mechanism. Confirm with a live probe, not log trust (example: the blob that returned 403 in the failing run returns 200 now, so it was a permission-timing issue, not missing data). A differential test beats an assumption: when two configurations should differ, run both and compare.
Only now read source for the WHY, and state the cause in one sentence with the evidence that proves it, plus what changes AND what deliberately does not change because it was not the cause.
Fix, minimally. Fix the upstream cause, not the symptom (a service that dies when the machine sleeps needs the sleep addressed, not a restart loop).
Verify with the original failing check, unchanged. Then run the wider gates (tests, lint, typecheck). Add a regression test that would have caught this bug. If recovery of live state was part of the fix, its success doubles as diagnosis confirmation.

Isolation techniques (when the cause is entangled)

Single-variable toggles: flip one factor at a time (feature flag, config bit, environment) and record the result per state. Never attempt a compound fix while two candidate causes are still entangled.
Necessary-but-not-sufficient check: after any partial fix, re-run the original failing check. The first improving number is not proof of a complete fix; say "necessary but not sufficient" and keep going.
Ranked compounding causes: many incidents have several contributing causes (stale binary + lost race + shared timeout). Enumerate and rank them; do not stop at the first plausible one.
Confound check: before concluding "the fix did not work" or "the test is wrong", look for an independent confound (a second process racing the test, a hook rewriting the file). Re-run clean.
Instrumentation-first for weird results: a flat or extreme result that fails to discriminate across conditions that should matter is an instrumentation bug until a targeted, isolated reproduction says otherwise. Do not report it as a finding, and do not let it block unrelated work while you wait to reproduce it.
Doc-vs-code conflicts: when documentation (or an Architecture Decision Record) and code disagree, determine the actual source of truth empirically (test the code), then fix whichever artifact is wrong, making the fix more precise about where the invariant is enforced. Never silently pick a side.

Anti-shortcut gates (hard rules)

Never delete, skip, or loosen a failing test to make a suite pass. If a test is genuinely wrong, prove it (git history: did the tested thing move or disappear?) and say so explicitly.
Never add a broad catch, silent fallback, or retry loop to make a symptom disappear. That converts a loud bug into a silent one.
Never weaken a designated safety mechanism (rate limit, halt, permission gate) because it fired. When a guardrail fires during a bad outcome, the default hypothesis is that the new code is incompatible with a correct guardrail. Changing the guardrail requires separate, explicit user sign-off.
Never read or print secrets (.env contents, tokens) while debugging credential or environment issues; test credentials differentially instead (call the API with each and compare behavior).
Retry a flaky operation at most once under changed conditions. If it fails again under conditions you believe clean, stop retrying and surface a structural hypothesis instead.

Verification is layered

An automated check passing covers only what that check was built to detect. Before declaring a generated or visual artifact done, do one perceptual/manual pass for defect classes the validator cannot see (label collisions, off-canvas content, semantic errors). If the ideal verification is structurally unavailable, run the nearest safe non-mutating substitute and state exactly what it does and does not cover; never report "verified" against a silently downgraded check.

Numerical and scientific code

For parity work against a reference implementation, tolerance judgment, and trajectory comparison, read references/numerical-debugging.md before starting. Summary of its hard rules: compare from identical initialization, know the float noise floor before calling a residual a bug, reserve "bit-exact" for residuals at machine precision, and test whether the metric itself is well-posed before chasing an unmet target.

When to stop

Root cause stated in one sentence, with cited evidence.
Fix verified by the original failing check plus the standard gates.
Regression test added (or an explicit reason why not).
If the investigation dead-ends: document it (goal, issue, root cause if known, lesson, alternative) in the project's scratch/history notes and the issue tracker; a documented dead end is a valid outcome, a silent one is not.
If you are deep in a long session and the remaining fix is risky, prefer a validated, independently re-runnable handoff artifact (script plus ordered notes, verified to run) over a rushed multi-file edit.

debugging

Debugging

The loop