Find the smallest causal mechanism or sufficient causal set that explains the observed failure and downstream symptoms. Diagnosis stays separate from triage/review/correction delivery.

A useful model may be:

trigger + enabling conditions + propagation + missing containment/detection → observed failure

Record exact symptom, expected behavior, first known trigger, candidate/revision or event identity, environment/context, reproducibility, evidence, scope, and read/probe authority. A report, stack trace, correlation, changed artifact, or temporal order is evidence, not a cause.

Use when an engineering/software report still needs issue-validity classification. Use only when the diagnosis is settled and the correction is a software/build delivery job; otherwise return the diagnosis to the current correction owner.

Reproduce safely when possible; otherwise pin one equivalent direct observation. Separate primary failure from secondary errors, retries, compensating behavior, and recovery noise.

Write a small hypothesis table. For each hypothesis state trigger/mechanism, enabling conditions, propagation, evidence explained, one distinguishing observation, and the smallest safe probe.

Read probe commands when bounded source/history/Git evidence can discriminate hypotheses. Prefer existing observations, tests, logs, traces, configuration, history, measurements, and reversible diagnostics that fit the domain.

Choose each next observation or intervention for its ability to distinguish the live hypotheses, not because a preferred debugging ritual exists. Control material confounders where practical. Vary one factor at a time only when that probe can actually discriminate the mechanism; when interactions or coupled conditions are plausible, design the observation/probe to expose those interactions rather than pretending the factors are independent.

For a proposed causal explanation establish:

A factor lacking discriminating support is contextual/contributing/unresolved, not a confirmed root cause merely because it occurred before the failure.

Continue only while another safe observation can materially update the causal model. Stop when remaining hypotheses cannot be distinguished, no safe probe can change the diagnosis, required environment/observability/authority is unavailable, or the failure cannot be reproduced and no equivalent direct evidence exists.

Do not perform correction attempts inside Root Cause.

Return one:

Include failure identity, minimal mechanism/set, per-factor evidence and causal role, interactions/alternative sufficient paths when material, contributing/contextual/unresolved factors, propagation/containment, decisive evidence, falsified alternatives, affected boundary, confidence limits, and smallest useful next action.

When a durable diagnosis is needed, use the existing/user-selected destination; use only when that destination is a repository-scoped workspace. Use only when a substantial terminal diagnosis benefits from a visual view.