Swap AI coding models mid-project: keep context in AGENTS.md, not the model

The account is written from direct experience: the author recalls a model switch that reversed a design premise hammered out over two hours in the previous session, overwriting the code with a different approach. It frames the root cause structurally rather than as a model defect. What disappears at a switch is not the code but the reasoning around it — rejected alternatives, remaining tasks, naming and error-handling habits — and the new model, starting from a blank slate, fills that vacuum with its own defaults.

The piece also distinguishes how much is lost depending on what is swapped. Changing only the model inside the same agent preserves convention files, tools and permissions, though reasoning traces are often lost and prompt caching is invalidated, raising cost and latency right after the switch. Changing the agent itself loses agent-specific hooks, permissions and convention filenames, and requires migrating those settings by hand. The countermeasures, however, are identical in both cases, which is why they are presented as a single three-part discipline.

A recurring theme is that natural-language instructions in a conventions file are only requests, while tooling is enforcement. Models under pressure to clear errors may add noqa comments, skip tests or loosen config files, and having CI detect such changes mechanically is recommended. It also notes that agents readily inspect current state via git status but rarely dig into why changes were made with git log -p or git blame, so those steps are written into the session-start checklist.