Regression Guard

The problem

AI coding agents break previously working code 3 out of 4 times (SWE-CI benchmark, March 2026). Developers spend 38% of their week fixing bugs they didn't write. 88% of teams need 2-3 deployment cycles just to verify a single AI fix.

The issue isn't that AI writes bad code. It's that it silently breaks things in other files while you're not looking. Tests pass. The build passes. Then something breaks in production because the agent changed a function signature and didn't update a caller three directories away.

How Regression Guard fixes this

It runs automatically after every code change — a checkpoint before your code goes anywhere near a test suite. The agent classifies the risk, traces dependencies, and checks for scope creep. You only hear about it when something's wrong.

• Silent on trivial changes — no noise, no friction. Most changes pass invisibly.
• Escalates automatically — a 2-line edit that changes a function signature? That's a Tier 3 check. No manual trigger needed.
• Catches what tests miss — behavioral contract checks caught a breaking signature change that both `tsc` and `vite build` silently passed in our tests.
• Scope creep detection — the agent checks "did I actually do what was asked?" and flags when it went off-piste.

Three tiers, right-sized checking

Not every change needs a full audit. Regression Guard classifies risk and scales verification accordingly:

• 🟢 Tier 1 — Trivial edits (1 file, <10 lines). Signature scan + import check. ~200 tokens.
• 🟡 Tier 2 — Moderate changes. Caller trace + alignment check + test suite. ~1,500 tokens.
• 🔴 Tier 3 — Risky changes. Transitive dependency trace + behavioral contract + full alignment audit. ~4,000 tokens.

Works everywhere

Claude Code, OpenClaw, Cursor, Codex CLI, Gemini CLI — no configuration required. Drop it in your skills folder and it works immediately.

Tested and validated

Tested by Claude Opus 4.7 on a real TypeScript codebase. The behavioral contract check and caller trace caught regressions that TypeScript's type checker and Vite's build process both missed. The checkpoint philosophy — reveal problems, let the agent decide — was specifically validated as the right design approach.