# Debugging > Finding the cause of a fault instead of a change that makes the symptom go away: reproducing before diagnosing, shrinking the reproduction until nothing is removable, stating a hypothesis that predicts an observation, changing one variable at a time, bisecting, and choosing which evidence to collect from a running production system before it is destroyed. Use when a fix is being guessed at, when a change "seems to work", when the same bug keeps coming back, when a fault cannot be reproduced, when a production incident needs a cause rather than a restart, when print statements are being added everywhere, or when several changes were made at once and it now works. Does not cover JVM performance triage (java-performance), GC (jvm-gc-tuning), live thread diagnosis (concurrency-diagnostics), heap dump mechanics (heap-dump-analysis), or deliberately injecting failures (distributed-systems-testing). - Skill: `robsonkades/debugging` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add robsonkades/debugging` - Raw SKILL.md: https://api.skillmd.com/api/skills/robsonkades/debugging/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: robsonkades (https://skillmd.com/u/robsonkades) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/robsonkades/debugging --- # Debugging ## Purpose Turn a report into a cause. The alternative — changing things until the symptom disappears — produces code nobody understands, a bug that returns under a slightly different input, and no way to tell whether the change helped or moved the failure somewhere quieter. The discipline is cheap and it is nearly always skipped under pressure, which is exactly when guessing is most expensive. ## Workflow 1. **Restate the fault as an observation.** "Customer 88123 saw a negative balance at 14:02" is an observation. "The refund logic is broken" is a hypothesis wearing a report's clothes, and adopting it early is how the wrong subsystem gets investigated for a day. 2. **Reproduce it.** Deterministically if possible, intermittently if not — but know which, because "I cannot reproduce it" and "it reproduces one time in ten" lead to different work. If it only happens in production, collect evidence within the mitigation budget (`references/production-evidence.md`). 3. **Shrink.** Remove inputs, steps, data and configuration until removing anything more makes the fault disappear, while preserving the original failure signature. Stop when the next reduction costs more than it helps; a minimal reproduction need not identify the full cause. 4. **State a hypothesis that predicts something you have not yet looked at.** "If the cause is the missing time zone, then the row written at 23:30 local will carry yesterday's date." Also name an observation that would refute it. Explaining existing evidence is useful; a discriminating prediction makes the next experiment useful. 5. **Test it by changing exactly one thing**, and record the result whether it confirms or refutes. Two changes at once means a confirmed hypothesis is still ambiguous. 6. **Check the explanation against the distribution** — frequency, timing, affected and unaffected users. Record remaining contradictions and uncertainty instead of forcing one cause to explain unrelated failures. 7. **Write the failing test, then fix, then verify** (tdd). The reproduction from step 3 is a candidate test. Use the narrowest level preserving the failure; a race may need controlled scheduling or integration coverage. Record when reproduction remains intermittent. Inspect deployed versus source JDK/toolchain, dependencies, JVM flags, configuration and data versions before version-sensitive diagnostics or changes. This workflow has no universal Java baseline. Preserve the target environment; upgrading it is a separate decision. Return the fault, evidence, tested/refuted hypotheses, fix or mitigation, validation and unresolved gaps. Missing access or reproduction permits an evidence plan, not an invented root cause. ## Rules - Prefer a causal explanation before a permanent fix. During an incident, a reversible mitigation may precede diagnosis; label it as mitigation, preserve the evidence the response budget allows, and do not present symptom disappearance as root-cause proof. - Prefer one controlled variable per experiment. If a batch helps, it implicates the batch but does not isolate a member or interaction; reduce/revert the batch in a controlled environment. - Read the whole stack trace, including cause and suppressed chains. The deepest application frame is a useful boundary candidate, not a verdict: framework callbacks, generated code, reflection, native frames, and library defects can move or hide the causal frame. - Balance evidence preservation against customer impact and the incident commander's authority. Capture cheap, non-disruptive evidence first when the error budget permits; mitigate immediately when delay is unsafe, and record which volatile evidence the action destroyed (`references/production-evidence.md`). - Symptom disappearance is evidence of recovery, not proof of cause or durable correction. - Question the assumption that the fault is where the symptom is. Corrupted state is written in one place and observed in another, often much later; the write is the bug. - Consider `git bisect` when verified good/bad revisions and a reliable classifier exist. It can narrow a monotone change in roughly log₂(n) classifications; build cost, skipped commits and intermittent or nonmonotone outcomes affect that benefit. See the reference's exit protocol. - Delete the debugging output before the change ships, and if a log line was genuinely useful, promote it deliberately with a level and structure (structured-logging) rather than leaving a `System.out.println`. - Timebox according to incident severity and experiment cost. Repeated refutations are a signal to revisit observations, harness assumptions and the system model or seek another perspective. ## References - **The method in detail** — `references/method.md`. Shrinking a reproduction, differential diagnosis (what changed — code, data, config, dependency, traffic, time), bisection over commits and over data, reading exception chains, and the specific traps of intermittent and heisenbug faults. Read when the fault resists the workflow above. - **Evidence from a running system** — `references/production-evidence.md`. What each source can and cannot answer — logs, metrics, traces, thread dumps, heap dumps, JFR, database state, deployment history — with volatility, cost and collection order during an incident. Read before touching a production system that is currently faulty.