Why Agent-Written Code Patches Need a Clean-Environment Replay Before Merging

A software pairing methodology argues that code patches generated inside a single AI chat session should not be merged directly, as hidden conversational context — such as rejected approaches, private service names, and informal architecture decisions — never reaches the repository or future reviewers. The proposed protocol requires that any patch be fully reproducible by a second, memory-free environment working only from a written brief stored alongside the code, not from the original chat transcript. Two key roles are defined: a driver seeking a quick agent-assisted fix and a senior engineer who refuses to approve any merge based on chat history alone. The method identifies three common failure points — using a model-generated summary, exporting the full transcript, or comparing patches visually — all of which allow contaminated context to persist into the new environment. The core principle is that replayability from repository-level documentation is treated as mandatory, while speed on the first attempt is considered optional.