Leads Master

Selling websites by phone runs on a spreadsheet of prospects, three PDFs of prices, a notebook of call-backs and one WhatsApp thread, and the thread is the only place the last conversation exists. Leads Master replaces all four with one lead record that a playbook, a call, a follow-up, a quotation and a report all key off. The screens are the smallest part. Most of the work is in the rules: what a quote may total, which status may follow which, and which offline write is safe to replay.

Cold-calling small businesses for website work is a bookkeeping problem wearing a sales costume. The script lives in one person's head and leaves with them. Prices come from a PDF, so two prospects get two quotes for the same work. Follow-ups go on paper, so the interested lead goes cold because nobody called back on Thursday. Call notes end up in four places, so nobody can answer how many conversations happened last month. None of that is a screen problem, which is why the rules went into SQL and the screens became a view of them.

• Designed the schema across eight D1 migrations: guides, playbooks, calls, leads, follow-ups, quotations, automation and offline receipts.
• Wrote 39 route handlers with zod on every write path, behind one typed repository per domain so no page builds its own SQL.
• Built all 26 pages, including the playbook graph editor, Call Mode, the quote builder and the analytics view.
• Wrote the four places a record could be lost or doubled: playbook publish, call-path folding, the idempotency receipt and the offline queue.
• Wrote the 257 tests, each database test building a throwaway SQLite file from the real migrations.

1. 01

   A playbook is a graph, and publishing is a snapshot

   validateGraph() · published_snapshot_json · WHERE published_version = ?

   A message step owns its responses and each response points at the next step, so a loop is a warning rather than an error. Validation is one pure function called twice, live in the builder and again on the server before publishing, so the two cannot disagree. Publishing stores an immutable snapshot behind a version guard, and every call starts from a copy of it, which means editing a playbook cannot change a call that is already open.

2. 02

   Position is folded, never stored as a pointer

   call_session_steps, append-only

   Going back in a call appends an event instead of rewinding a cursor. The current step is re-derived by folding the log, so a caller who changes their mind afterwards replaces the abandoned route without rewriting what happened.

3. 03

   Money is an integer, and the server recalculates it

   quotes · quote_line_items · triggers in 0006_quotes.sql

   Totals are integer minor units with the currency exponent, discounts are basis points, and no float reaches the database. The browser never submits a total: both write paths re-read the price rows and run the same totals function the preview and the print view use. A trigger rejects any line whose total is not quantity times unit price, and a second one rejects a line in a different currency from its quote.

4. 04

   An offline write lands once, or not at all

   src/lib/offline/policy.ts · mutation_receipts

   The service worker never touches the API. Writes the application knows are safe to replay queue in IndexedDB behind a five-pattern allowlist, each carrying an Idempotency-Key. The receipt is a reservation rather than a lookup: the first request claims it with a private nonce and re-reads it, so two simultaneous replays cannot both run the handler, and a refused write releases its key so a corrected retry still works.

5. 05

   One follow-up rule, five screens

   src/lib/leads/follow-up-time.ts · follow-up-consistency.test.ts

   A follow-up is stored both as a calendar day and as a UTC instant, and comparing those directly is how a task ends up due on one screen and overdue on another. One module compares days in the reporting zone, and the dashboard counters, the lead badge, the list filters, analytics and the automation rules all call it. A cross-feature test fails if any screen drifts.

6. 06

   Automation reads records and never writes to them

   automation_rules · automation_alerts · automation_runs

   Five rules, no cron, evaluated when you press run. Reminders upsert by dedupe key, so running twice opens nothing new and a reminder whose condition clears resolves itself. There is no analytics table either: the reporting page reads the same rows the list pages read, so a figure cannot disagree with the list it came from.

• 01

  There is no authentication

  No owner_id, no login, no per-user isolation. Anyone who can reach a deployment can read and edit every lead, call, quotation and playbook. Cloning the repository is safe; exposing it is not, until Cloudflare Access or real authentication sits in front. The README carries that warning above the fold rather than in a footnote.

• 02

  No CSV in or out

  Bringing a spreadsheet in, or handing data to a CRM, is manual or a script you write. The duplicate check runs on the way in only if the write comes through the app.

• 03

  A fourth market is a migration

  The three seeded markets are editable in the UI. A new one needs SQL, because the market codes sit in constraints rather than in a form.

• 04

  Thin coverage in two places

  The analytics composition has one cross-feature test, the follow-up rule. The automation open, re-open and resolve cycle has none, which is the next thing to pin.

• 05

  No background scheduler

  Nothing runs while the app is closed and nothing is sent anywhere. No cron trigger is configured, so automation is a button rather than a queue.

• 06

  A quotation is printed, not downloaded

  It leaves the building through the browser print pipeline and a print stylesheet. There is no generated PDF and no read-only share link.

The loop closes: plan a session, run a published playbook, log the outcome, and send a quotation priced from the catalog the session was planned against. What the repository deliberately does not publish is business impact. It keeps no analytics on its own use, so the number of leads added, calls logged and quotes sent stays blank rather than being estimated. The code is the evidence, and it is the only thing claimed here.