Macrop

The same pure TypeScript engine runs on the server as the authoritative source and in the browser for optimistic previews: the UI updates instantly and is guaranteed to agree with what the server persists. The boundary is enforced, not intended: lint rules block any import across it, and CI runs a dedicated architecture gate on every push.

packages/core   pure macro engine: framework-free, no I/O
      │
      ├── apps/api   Hono · Drizzle · PostgreSQL    authoritative
      └── apps/web   React · TanStack Query         optimistic previews
                │
        hard HTTP boundary: cross-imports fail the lint

Meals can be drafted from a sentence or from a photo, and none of that machinery is allowed near the arithmetic. The model returns ingredients and quantities (no prompt’s output schema has a total in it), so every kcal, protein, carb and fat figure on screen is still the same pure engine multiplying macros by quantity.

The provenance rule that governs the staple catalog governs model output too: a row the model estimated stays private to the household that imported it, its micronutrients left null rather than invented, and only a hit against a real nutrition source is promoted into the global catalog. A guessed number never becomes everyone’s number.

Solo doesn’t mean unreviewed: I built the guardrails a team would have, then automated them.

lines of TypeScript

~148k

workspaces

3

test files

236

external nutrition APIs

3

releases

CI-gated

This repository’s first commit is not code; it is a 1,315-line specification: the domain model, the decision defaults, the version pins with their reasons, and a standing rule to ask before anything consequential or irreversible. The second commit is the application generated from it, 155 files in one pass. Everything after follows the same split, and the history records it precisely: as of August 2026, 96% of the non-merge commits carry a Claude co-author trailer, and not one is recorded as hand-written. I did not type this codebase. I specified it, constrained it, reviewed it and merged it.

That is what the ownership claim on this page means. The decisions live in written artifacts the agents had to obey: the genesis spec (deleted from the tree, still readable in the first commit), plan documents that precede their implementation commits in the log, eleven path-scoped rule files, and a test policy with teeth: an agent may add tests freely but cannot change an existing one without my explicit approval in that session, enforced by a hook that intercepts the edit. And the merges are the signature: every one of the repository’s 1,315 merges is mine, by pure coincidence exactly as many as the genesis spec has lines; both counts are real, and the history landed on the same number twice. The repository itself is private (deliberately: the code is the product), so the history behind these numbers is shown on request, the same way the demo account works.

The record also shows the discipline earning its keep on the app itself: English consent copy once shipped still in German, because the translation gate guarded key parity but not whether a value was still the source string: caught, fixed and pinned with a test so it cannot recur. What the record cannot do is reconstruct a per-line human/agent split, and I will not pretend it can. The claim is ownership of the specification, the merges and the verdicts; the audit machinery that patrolled all of it grew into the Heal Suite, whose page carries the other half of this disclosure.

A sample, not the full list: four decisions that were reasonable at the time, and what each one cost.