Safe Change

Know exactly what breaks before you touch it.

Safe Change maps the blast radius of any code change — importers, API routes, test coverage, ENV vars, database migrations — then gives you a risk score and waits for your go/no-go before running the verify gate.

It is the proactive companion to deep-debugging: catch impact before the bug ships, not after.

Quick Reference

When to Use

Activate this skill whenever:

• You are about to rename or extract a function/class used across multiple files
• You are changing the signature of a service method
• You are modifying a NestJS controller or Next.js API route
• You are touching a file that is imported by more than 3 other files
• You are changing code that reads from process.env
• There is an active DB migration in the migrations folder
• A teammate asks "is it safe to change X?"

Do not skip this step for "small" changes. Most production incidents start as changes that felt small.

How It Works — 6 Phases

Phase 1 — Detect Stack

The agent reads package.json at the project root to determine:

• Is this NestJS, Next.js, or generic TypeScript?
• Are both frameworks present (monorepo)?

Phase 2 — Build Impact Map

Run scripts/scan-impact.mjs from the project root.

The script uses regex-based static analysis (no AST compiler, zero install friction) to collect:

Phase 3 — Risk Score

Heuristic scoring (see table below). The agent reads the JSON output and renders it as a human-readable report.

Phase 4 — Render Report

The agent formats the JSON from scan-impact.mjs into a Markdown report (see Output Format below). It presents the report and explicitly states the risk score at the top.

Phase 5 — Checkpoint (mandatory)

After presenting the report the agent must pause and ask:

> "Risk score is [Low/Medium/High]. Do you want to proceed with this change? (yes / no / adjust scope)"

Do not proceed until the user confirms. This is the core safety gate.

Phase 6 — Verify Gate

After the change is made, run scripts/verify-gate.sh from the project root.

The gate runs in order:

1. tsc --noEmit — type-check
2. npm run lint (if script exists)
3. npm test (if script exists)
4. npm run build (if script exists)

Stops on first failure. Color-coded output. Non-zero exit on failure.

Output Format

The agent renders the JSON from scan-impact.mjs as:

## Safe Change Report — src/notifications/notifications.service.ts

**Risk Score: MEDIUM**
Risk factors: 5 importers, 1 API route, test coverage exists

### Importers (5)
- src/users/users.service.ts
- src/appointments/appointments.service.ts
- src/billing/billing.service.ts
- src/reports/reports.service.ts
- src/audit/audit.service.ts

### API Routes Affected
| Controller | Endpoints |
|------------|-----------|
| NotificationsController | POST /notifications, GET /notifications |

### Test Coverage
- src/notifications/notifications.service.spec.ts ✓
- Gap: no

### ENV Variables Referenced
- SMTP_HOST
- SMTP_USER
- SMTP_PASS

### Recent Migrations (last 7 days)
- None

---
**Checkpoint:** Do you want to proceed with this change? (yes / no / adjust scope)

Limitations

See references/limitations.md for the full list. Key constraints:

• Regex-based: dynamic imports (import(path)) are not detected
• Re-exports through barrel files (index.ts) may undercount importers
• Decorator aliases (custom @Route() wrapping @Controller) are not detected
• TypeScript only — no Python, Go, Rust adapters in v0.1

Companion Skills

• deep-debugging — use after a bug ships; safe-change is what you run before
• self-improving-agent — log the impact report as a learning when a change causes an incident

Scripts

File Structure

safe-change/
├── SKILL.md                          # This file
├── README.md                         # Marketing overview
├── package.json                      # ClawHub metadata
├── scripts/
│   ├── scan-impact.mjs               # Impact analyzer (Node ESM, no deps)
│   └── verify-gate.sh                # Verify gate (bash)
├── references/
│   ├── example-impact-report.md      # Full SMTP swap example
│   ├── usage.md                      # How agent invokes the scripts
│   └── limitations.md                # Known limitations + trade-offs
└── assets/
    └── SKILL-TEMPLATE.md             # Template for creating similar skills