Automake

Recipe

1. Inspect the full conversation and repository, resolving every Optimizer, Evaluator, and run input as supplied, safely inferable, defaultable, or missing.
2. Treat only a clear request to start the Automake ratchet now as run-now intent, resolving every missing input safely and skipping all questions and approvals.
3. Otherwise ask one missing Optimizer input per turn until its Goal is concrete, without repeating known facts or asking about run limits.
4. Present the complete Optimizer approval block in this order: **Approve this Optimizer instruction?**, the Goal, smallest candidate scope, constraints, cheap checks, and repository context without evaluation criteria, then exactly 1. Approved (Recommended) and 2. No — changes needed.
5. Ask one missing Evaluator input when needed.
6. Present the complete Evaluator approval block in the same order: **Approve this Evaluator instruction?**, its independent Goal, BETTER rule, task-shaped evidence, exact verdict format, complexity cost, and candidate-gate policy, then the same two choices.
7. Resolve iterations, consecutive-failure limit, and observable success condition from supplied values or defaults.
8. Present the complete run gate in this order: **Ready to run Automake?**, the resolved summary, then exactly 1. Run (Recommended) and 2. No — changes needed.
9. On run approval, establish a clean committed Git baseline automatically, excluding likely local secrets without reading or committing them, and create one run-local Automake state directory.
10. Run one fresh write-bounded Optimizer and one fresh read-only Evaluator per candidate without crossing their instructions, accepting only a parseable evidence-backed BETTER verdict.
11. Ratchet accepted commits forward, restore every rejection exactly, and record every launch, attempt, result, evaluation or rejection, learning, state transition, final kept commit, and stop reason.

Details