Establish symptom, expected behavior, environment, revision and impact. Reproduce with a small input when possible. For intermittent failures, record frequency, logs, traces and conditions instead of waiting for an on-demand repro. Name the check that will prove the eventual fix (a failing regression test or a controlled signal), even before the cause is known.
State a falsifiable hypothesis per plausible cause, with predictions and evidence for and against. Keep observations separate from explanations. Inspect actual data/control flow and changes near onset.
Pick the experiment that discriminates among causes. A stale read alone cannot separate cache invalidation from replica lag. Use a controlled intervention, boundary trace or counterexample, accounting for confounders and observer effects.
Instrument the smallest useful boundaries: structured logs, no sensitive payloads, only the inputs, correlation and timing the question needs. For performance, profile the workload.
Update each hypothesis: refuted, supported, confirmed or unresolved, in a compact evidence table. Parallel read-only investigation is fine for independent questions when authorized; agents must not race to patch competing theories.
Once the cause is supported, make the smallest in-scope correction. Add a regression test that fails on the defect and passes on the fix where practical; otherwise keep a reproducible trace or controlled acceptance check and state its limits.
Check the fix, affected behavior and required repository gates. Remove temporary instrumentation unless intentionally kept. Report root-cause confidence and residual risk.

Stop conditions

Honor any investigation budget. If experiments stop yielding information, access is missing, or evidence cannot separate causes, report the smallest missing observation instead of guessing or repeating a patch. Urgent containment belongs in flow-triage under the user's operational authorization.

Red flags

"I know what this is": write the competing hypothesis down and test it anyway.
"Let me try one more change": no new patch without a new observation.
"It went away": an unexplained disappearance is not a fix.
"Three fixes failed": the model of the system is wrong; revisit assumptions, not the patch.

Anti-patterns

Treating correlation or one matching prediction as causation.
Blocking on deterministic reproduction before collecting better telemetry.
Stacking speculative fixes so nobody knows which cause mattered.
Accepting a test/import failure as the regression signal without inspecting it.

Done when

Cause and correction have discriminating evidence, or the investigation ends with a precise unresolved hypothesis and the next observation needed.
The final artifact and relevant checks support the claimed outcome.
Temporary instrumentation and uncertainty are accounted for.

Use flow-work for implementation or flow-triage when production impact needs containment.

flow-diagnose

Diagnose

Process