Legacy code changes

Legacy Code Changes

Legacy code is code that makes money and scares you. It survived contact with reality — every weird if in it may be a bug fix for something that actually happened. The discipline: comprehension before modification, safety before improvement, incremental over heroic. This applies double to AI assistants, which have a strong rewrite bias: regenerating code they don't understand instead of minimally editing it.

Gate 1: Understand before touching (timeboxed, active)

• Trace one real flow end-to-end — entry point → decision points → side effects → output — for the exact behavior you must change. Reading the whole codebase is procrastination; tracing one path is comprehension.
• Use archaeology, not just reading: git log -p on the file (why does this weird line exist? — the commit message knows), git blame on the scary part, existing tests as executable documentation, and running the thing with a debugger/print on the path in question.
• Write down the 3–5 facts you learned that surprised you. If nothing surprised you, you haven't understood it yet — legacy code always surprises.
• Chesterton's Fence is the law: never delete or "fix" code you can't explain. That check for a null tenant on Tuesdays is either dead code or a ₹10-lakh lesson — find out which (git history, asking, logging it in prod) before removing.

Gate 2: Pin current behavior before changing it

• Where tests are missing, write characterization tests first: capture what the code actually does now (including behavior that looks wrong), so you can detect what your change breaks. You're not asserting correctness; you're building a tripwire. Feed the function its realistic inputs, snapshot the outputs, done — 30 minutes of pinning beats a week of "what else did I break".
• Can't test it because it's tangled in I/O and globals? Find or make the smallest seam: extract the decision logic from the side effects just enough to get it under test (code-quality's edges-and-core rule) — the minimal surgery, not a beautification pass.
• No time even for that? Then pin behavior operationally: run the golden path before and after, diff the outputs/DB state, and say honestly that this is the verification level (evidence rule).

Gate 3: The smallest change that works — then stop

• Match the existing style, patterns, and idiom even where you'd choose differently — a codebase with two conventions is worse than one with one mediocre convention. Improvements to the pattern are a separate, explicit proposal.
• Never mix the behavior change with cleanup in one commit (change-hygiene). The reviewer must be able to verify one claim at a time.
• Resist the pull to fix everything you see. Keep a written list of what you didn't fix (the "later ledger") instead of expanding the blast radius now. Touch radius ∝ risk in undertested code.

When "rewrite it" comes up — the honest calculus

Rewrites systematically lose because the old system's value is not its code but its accumulated edge-case knowledge, which the rewrite silently discards and then relearns from angry users. Default answer: no.

• The alternative that works: strangler pattern — put a boundary in front of the old code, route new functionality (or one migrated slice at a time) to new code behind the same interface, verify parity, expand slice by slice, delete the old path last. Every step shippable; reversal always possible.
• A rewrite is justified only when: the platform is dying under you (unsupported runtime, unbuildable), AND the scope is small enough to finish before requirements shift, AND you have the behavior pinned (tests/parity harness) so "done" is checkable. All three, in writing.
• Partial rewrites of the scary module follow the same rule in miniature: pin behavior → build replacement behind the seam → run both and diff (shadow mode) → cut over → delete.

Failure modes to hunt (in yourself and AI output)

• The "while I was in there" diff — 400 changed lines for a 5-line fix. Reject and re-slice.
• Regenerated-not-edited: an AI rewriting a function wholesale, silently dropping the weird branches (that Tuesday null check) it didn't understand. Diff old vs new branch by branch, not by vibes.
• "The tests pass" in a codebase whose tests don't cover the changed path — check coverage of the specific diff, not the suite's green.
• Confidence from cleanliness: new tidy code feels more correct than old ugly code. Feeling is not evidence; parity runs are.