Model-Backed Code Review Prompts Need Version Pinning to Prevent Silent Drift

Teams increasingly place model-backed review calls in their pull request pipelines, but prompts and decoding settings are often edited without oversight. The article argues that such prompts, settings, and test fixtures should be treated as production dependencies rather than informal notes. It proposes a reviewable lockfile combined with a CI checker that fails when the prompt changes without an explicit lock update. To make the hashes reliable, the approach canonicalizes files by normalizing line endings, trimming whitespace, and sorting settings keys before hashing. The author distinguishes transport errors and nondeterminism, which call for retries and agreement checks, from prompt drift, which requires a reviewed lock bump.