clone-drift-tracker

What this is for

The dangerous clones are not the identical ones - they are the ones that USED to be identical and silently diverged. One got a bugfix, the twin did not. This skill mines git history for code blocks that were once byte-identical at a past ref and now have different body hashes.

The static-only code-clone-detector cannot see drift through time. Mutation testing (Stryker, PIT) tests fail-ability dynamically but needs the full test run. This is the zero-runtime, deterministic temporal complement.

PROTECTION RULE - never ~/.claude/

Read-only skill. Guard required if write mode added later.

## What You Must Do When Invoked

Step 1

If -help is passed, print the ## Usage block below and stop.

Step 2

Confirm -ProjectDir is provided, the path exists, AND it is a git repo.

Step 3

Run: scripts/drift-scan.ps1 -ProjectDir "<path>" [-PastRef "v1.4"]

Step 4

If not given, defaults to HEAD~100 (compare against 100 commits ago).

Step 5

LLM reads the JSON output. For each drift pair:

Step 6

Read the currentSnippet and pastSnippet.

Step 7

What changed on the current side?

Step 8

Did the other side get the same fix? (Often the answer is no.)

Step 9

Is one side's change a missed migration, a missing bugfix, or an

Step 10

intentional specialization?

Step 11

Classify: missed-fix (high) / intentional-divergence (informational).

Step 12

Confidence: proven (clear functional gap), likely (likely gap),

Step 13

suspected (judgment call).

Step 14

Write clone-drift-report.md to the working directory.

Usage

/clone-drift                            # interactive
/clone-drift <dir>                      # against HEAD~100
/clone-drift <dir> -PastRef "v1.4.0"    # specific ref
/clone-drift -help                      # show usage

Returns JSON with pairs[]: {functionName, currentFile, pastFile, currentHash, pastHash, pastRef, pastSnippet, currentSnippet} plus summary.