Code decides; the model reads and drafts.

The assistant follows a script. If the caller wants a booking, it stays with them; if they ask about something else, it answers briefly and goes back to the booking.

Anything with consequences is decided by code: who is calling, which slots exist and what gets written to the calendar. Every request from the model goes through strict Zod checks (a library that checks each field has the right shape), and anything that does not fit is rejected before the calendar changes. Each client runs in its own container.

There were 63 calls and 34 bookings in the first 2 days, at about a minute and a half per call, with 0 technical errors. The production system has over 5,700 tests; the public template, on fake data, has 4,767 + 277, all passing.

It works like closing the till in a shop: you compare what the till says with the receipts, and if they do not match you see it straight away. Here the till is the exchange, and the receipts are the firm's books.

There is almost no model here, on purpose. Code downloads the data and matches it; it also does the maths. Automation reports, people decide.

Security mattered more than the technique. The keys can only read, so the platform cannot move money even if someone wanted it to. It is private; I can show it in a live demo if you ask, never with client data.

When I type a prompt, a hook (a small program that runs on every message) finds the note that fits best and adds a pointer of at most 1,200 characters, in 0.25 seconds. If the agent needs more, it asks the MCP server, which returns fragments.

Code decides what goes into the prompt: it scores by keywords and ranks by authority, and it never goes over 1,200 characters. The search is keyword based (BM25), and the tests run offline.

Every recording has a state in the queue, so I can always see which step it is on.

The model only summarises. Where the note goes and when to retry is decided by code.

It has 21 modules and 161 tests, including one that compares the output with a saved correct note (a golden-file test), plus a 1,300-line architecture doc. Memory use and speed are measured on a small VPS.

It works like a small office. The main session hands out the work and writes no code; junior searches and summarises, plumber does the routine work, senior takes what is hard or cannot be undone, and verifier checks at the end without touching anything.

Hooks are code, so they run every time, whatever the model decides. The per-prompt router has a --dry-run mode that shows what it would do without doing it.

The Stop hook came from a real failure: a session announced work, ended its turn without starting it, and my server sat idle for 3 hours. Since then, a turn that promises work it has not launched cannot end.

Nobody has to remember to run it. The timer starts it, and nobody presses a button.

Code decides when it runs and where the video goes.

It is private, so here I only describe how it works.

It is a public template, and all the data is synthetic: no real student appears anywhere.

The reports are deterministic, so the same data always gives the same report.

It has 11 unit tests, and the typecheck and lint are clean.