troubleshoot

Preserve evidence first, build a failure contract, then run hypothesis-driven experiments with bounded commands. The workflow localizes the earliest divergence, proves the causal chain, and only then applies and verifies a repair at the narrowest boundary.

A restart, rollback, failover, or cache clear can mitigate impact but does not prove why the failure occurred. Root cause requires a trigger-to-invariant-to-symptom causal chain, a counterfactual, and elimination of alternatives.

Production and unconfirmed environments remain read-only until the user explicitly authorizes the exact live action. Confirmed non-production targets may receive bounded, reversible changes with an identified target and rollback path.

Skip it for known mechanical fixes, routine feature work, syntax errors, formatting, lint, and straightforward dependency installation. Use dedicated skills for code review, security assessment, or domain-specific Terraform, Helm, and cloud implementation.

A remediation budget limits each blocker to five attempts and 120 active minutes by default. Every retry requires newly acquired evidence and a genuinely new falsifiable hypothesis; when the budget is exhausted the skill reports instead of patching speculatively.