Mitate Hits 2,990 Commits in 6 Weeks on Spare Time

The development record starts from a deliberate decision to split the product into four physically separate repositories rather than one codebase. Mitate holds the architecture policy, API contracts and conformance fixtures; mitate-agent handles the design layer; mitate-application handles safe execution; and data-service-api holds the shared Data Service API spec and two database implementations. Keeping AI from reading into the neighboring layer's internals appears to be the reason each layer could be worked on at the same time.

The contract-first flow then turned that separation into concrete work orders. For capability DM-0065 on Sept 28, the contract revision added 36 fixtures, then Application changed 41 files, then Agent changed 26 files, followed by 71 mutation tests and an evidence audit. The audit did not pass cleanly: it recorded 62 fulfilled, 5 unfulfilled and 4 lacking evidence, and the supervising AI accepted the work conditionally while leaving unverified items on the record. That habit of writing down what was not checked is what allows the project to move without pretending the work is finished.

A second pattern is the use of parallel sessions. The Sept 8 contract v0.24.0 record shows four Opus instances drafting in parallel, followed by three Sonnet instances implementing in parallel, while a separate session landed a health-check CLI in mitate-application first. Parallelism here is a byproduct of separation rather than a scheduling trick. The author also notes that the hardest constraint was AI resource exhaustion—the weekly quota did not last a week—which is why model assignment by role mattered.