MigrationGuard Harness — Development skill for Claude Code
An Evidence-Driven Agent Harness for Safe Dependency Migration, featuring Claude-powered code repair, differential verification, secure sandboxing, and checkpoint recovery.
How to install MigrationGuard Harness
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open LZH-1-G/MigrationGuard-Harness and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What MigrationGuard Harness does
An Evidence-Driven Agent Harness for Safe Dependency Migration, featuring Claude-powered code repair, differential verification, secure sandboxing, and checkpoint recovery.
Alternatives in Development
- Claudia — A powerful GUI app and Toolkit for Claude Code - Create custom agents, manage interactive Claude Code sessions 21k ★
- Zeroshot — Runs coding agents as a graph: one agent implements, independent agents review, failures go to repair, and not 1.9k ★
- Vibekit — A simple SDK for safely running Codex, Gemini CLI, and Claude Code in a secure sandbox 1.7k ★
README
MigrationGuard
**An evidence-driven agent harness for safe dependency migration.** MigrationGuard migrates Python projects from Pydantic v1 to v2 with deterministic codemods and Claude agents, and accepts the result only when sandboxed, differential evidence proves it is safe. It never trusts an agent's self-reported success.
$ migrationguard run ./billing --agent claude # real run, abridged
verify ✔ tests.pass: PASS — 7 passed
verify ✔ tests.differential: PASS — all 7 baseline-passing tests pass
verify ✘ behavior.probes: FAIL — 13 behavioural difference(s) ← codemods alone: tests green, behaviour changed
▶ verify -> repair
repair repair attempt 1: finished; 6 tool call(s), 0 denied; $0.2223, ...
verify ✔ tests.integrity: PASS — 0 protected file(s) changed, all re-derivable from the audited codemod
verify ✔ behavior.probes: PASS — 16 probes agree
review ✔ review.independent: PASS — The migration is faithful to v1 behaviour. ...
ACCEPTED: all blocking gates passed at iteration 2
Why
Coding agents are good at making a migration *look* done. The common shortcuts — a green test suite, an agent saying "all tests pass" — miss exactly the failures that matter in dependency upgrades:
- silent behaviour changes: Pydantic v2 stops coercing
123to"123", rejects42.7for anint, serialisesDecimalas a string and UTC asZ. Typical test suites stay green. - reward hacking: deleting, skipping or weakening failing tests, adding a
conftest.pythat deselects them, pinning back to v1, or importingpydantic.v1. - fragile execution: long agent runs die halfway (crash, rate limit, budget) and start over from scratch.
MigrationGuard combines four things to address that:
- Autonomous migration — deterministic libcst codemods for everything mechanical, a Claude Agent SDK repair agent for what needs judgement.
- Differential behavioural verification — the original project (pydantic 1.x) and t
Related Skills
03 Repair Batch Consistency
Repair Prompt: Batch Consistency and Mirror Recovery
Formux
Repair a form around completion, recovery, and trust
Claude Code Session Recovery
Diagnose and repair Claude Desktop session metadata bugs on Windows and Mac — missing conversation history, lo
ClaudeCraft
Claude Code Powered Multi-Agent Development Framework for High-Quality, Secure Software Development
Lazy Bird
Universal development automation for ANY project. Claude Code implements features, runs tests, creates PRs aut
Skill Evaluation Graph
Description Evidence-driven evaluation, benchmarking, and verified repair for Agent Skills across Codex, Claud
Related Agents
Hydra Task Steward
Create, checkpoint, hand off, or complete a Hydra task record. Use when work needs persisted state, a handoff,
AI Product Designer
The AI Product Designer designs LLM-, agent-, and ML-powered features inside the app: prompt UX, guardrails, l
Task Fixer
Repair agent that reads spec.md, verdict.json, and problems.md, applies the smallest safe fix, and refreshes e