bsafronov Refactor Code
Change the requested structure with evidence that its callers still observe the same behavior.
- Define equivalence at the affected boundary: results and identity, side effects, ordering, errors, resource lifetime, and compatibility where the change touches them. Existing quirks remain behavior unless their correction is also in scope; an apparently cleaner result can still be a regression.
- Establish that boundary's behavior before changing its implementation. Reuse meaningful coverage; add characterization only where an exposed risk lacks proof. A pre-existing failure or unavailable check limits the baseline: record it, preserve the evidence, and use an independent observation where possible. A green run only after editing cannot establish before/after equivalence.
- Move the chosen responsibility while keeping callers' obligations intact. Preserve ordering and failure paths through extractions, including work that must not occur after failure. Keep useful behavior tests tied to the public boundary; replace an implementation-coupled assertion only with evidence that its behavior remains covered. Split genuinely independent moves when each can be verified; keep a coupled edit together when intermediate states cannot work.
- Repeat the same behavioral observations after the move, and verify the chosen structural boundary is now real rather than an unused wrapper. Treat a changed observation as a discrepancy to explain, not a reason to update expectations. Report the affected guarantees and any equivalence that remains unproved.
Stop when the requested structure exists and the relevant before/after behavior agrees. An unresolved contract choice can compose with interface design; an independently requested correction or feature remains a separate behavior change. Small cleanup inside an already green test-first slice does not require another Skill. This Skill owns preservation across structural change, not test-first sequencing, diagnosis, or mixed-version rollout.