It began as one person’s work across several companies: tasks in one tool, issues in another, deals in a third, and a routine of moving between them that nobody else could follow.
The first version was a personal work system for exactly that. Once it shipped, real work moved into it, and the brief grew: replace the task tool, the issue tracker and a light CRM with one place, on our own server, where every change leaves a mark and the system says what is slipping.
One action per real event
Every change goes through one named action, and every action writes one dotted event into the history: task.completed, cycle.closed, invoice.voided, deal.lost. Closing a project is never “update status”. An update records which fields changed, with the old and new values. Long text is stored as its length, and tokens and credentials never reach the log.
Serious actions ask for a reason: dismissing an insight, voiding an invoice, losing a deal, deleting a contact or a role, cancelling a meeting, deactivating a person. Repeats are harmless. A double click or a retry that finds the record already in its target state does nothing and logs nothing.
- 09:12webtask.completed · by Amina
- 09:40agenttask.cycle_changed · origin “planning-bot”
- 10:05webinsight.dismissed · reason “client paused the project”
- 16:58webinvoice.voided · reason “issued to the wrong company”
Capture first, organize later
A project, a task, a contact, a document or an expense can belong to no company at all. You never invent a dummy company to save a thought. A task inside a project takes the project’s company, with no override. Moving a record to another company shows what will change before it saves.
Unassigned is a real place, not a hole. Reports include it by default, so nothing captured in a hurry falls out of the numbers. Billing still names its company explicitly, even when ownership is empty.
Agents use the same doors as people
AI agents connect through MCP with a token owned by a person, and they act as that person: same permission checks, same actions, same rules as the screen. Every history row says how it came in, through the web, an agent, an automation or the API, and an agent’s row names the token it used.
Every agent write carries an idempotency key. The same key with the same arguments replays the first result. The same key with different arguments is refused. Agents never change work on their own: they suggest, a person clicks, and the normal action runs as that person.
Signals are arithmetic
The checks that say what is slipping are plain code. They always run, with or without an AI key. A model may reword an insight to read better; it never decides one.
| Signal | Fires when |
|---|---|
| Slipping forecast | Today plus open points × average cycle days ÷ average points done, over the last three cycles, lands after the target date. Critical past 14 days late. |
| Blocked chain | The longest chain of open blocking tasks is longer than 2. Critical past 4. |
| Scope creep | Points added after a cycle started exceed 20% of the plan. Critical past 50%. |
| Stale | Nothing has moved for 14 days. |
The machine watches. The decision stays human.
Told once, at the right hour
Email and WhatsApp wait for the end of each person’s quiet hours, in their own timezone. The in-app notice never waits. Each person is told once per event: the person who made the change is left out, and so is anyone who cannot see the task. A bulk edit sends one message per person, not one per task.
Reminders use a small ledger of what was sent, keyed on person, event, subject and local date. The row is written first, and the message goes only if the write succeeded, so nobody is reminded twice in a day, even when two workers race.
Hours become invoice lines
- Log the time
A timer or a manual entry, capped at one day.
- Draft the invoice
Including unbilled time locks those entries and writes one line per task, plus a line for other time, in hours to two decimals.
- Billed is billed
A billed entry can no longer be edited or deleted.
- Void to release
Voiding the invoice or deleting the draft releases the time. A credit note keeps it billed.
Failures that stay visible
Failed jobs were first written to files. The disk is wiped on every deploy, so failed agent runs, reminders, invites and webhooks disappeared, and nobody saw them. Failed jobs now live in a database table, counted in settings, with the newest fifty listed and a logged retry or forget on each. Everything else stays on Redis. The table grows only when something fails.
| Turned down | Why not |
|---|---|
| Failed jobs on disk | The disk is wiped on deploy. Failed reminders, invites and webhooks vanished unseen. |
English, on purpose
Our house rule is Somali and English from the first day. ZilaOS is English only, by decision, with no language files at all. Adding Somali later is a new decision and a full sweep, and we wrote that cost down before we chose.



