The harness drives; the model is the passenger

Every step a program can do exactly and repeatably is a step a language model can get wrong, and it costs the model attention it should spend on the actual work. So whatever can be done deterministically is done by the harness, and the model is left to create.

The user, 2026-10-10: “everything that can mechanically/deterministically be done by the harness we should handle it, so llm can focus on creating.”

• Git is an automated module: opening a folder that is not a repository runs git init, and EVERY agent action on a repository is committed or tagged automatically.
• Every design asks “can the harness do this deterministically?”; if yes, the model never has to. Retroactive.
• Project-specific harness data lives in the project’s .bridlewise/ folder, git-ignored (checkpoints included).
• Two consequences have their own records: PARTIAL-FAILURE and TESTS-TEST-THEMSELVES.

What is missing: not built yet. engine/src/git/workspace-lifecycle.ts can initialize an empty folder as a repository, but nothing calls it when a project is opened, and no agent action is committed or checkpointed automatically. The git-auto design carries it out.