CodeClone 2.1: why your AI agent's diff isn't enough

Tasking an AI agent with a change typically produces a clean diff and passing tests, but that surface tells you nothing about whether the agent stayed in scope, touched forbidden paths, or introduced duplicate logic along the way. CodeClone 2.1.0a1 positions itself as a Structural Change Controller that moves governance earlier than the diff. Agents declare their intent over MCP before editing, and CodeClone responds with the structural blast radius, the allowed file scope, explicit do-not-touch boundaries, and a regression budget; once editing finishes, the actual change is reconciled against that declared contract.

A core design principle is that no LLM ever judges whether a change is safe — every boundary and verification step relies on deterministic facts derived from the repository, avoiding the correlated failure mode where one model's rationalization convinces another. The alpha also ships a read-only Live Implementation Context feature to curb in-session context loss, and an Engineering Memory layer that requires human approval before agent-generated knowledge becomes trusted across sessions. Additional capabilities include agent trajectory tracking, multi-agent coordination, and an opt-in observability layer.

The tool is available as a prerelease on PyPI and runs entirely locally with no cloud dependency. It reflects a broader argument that code review in agent-assisted development needs to govern process and intent, not just judge the resulting diff.