What it is

Trapline watches a defined set of “patches” — sources where new postings appear — and maintains a clean ledger of what it finds. Each run reads the due patches from D1, searches with Parallel, extracts postings with Claude Haiku (Workers AI’s gpt-oss-120b as fallback), and dedupes on a strict identity: district + canonical employer + role, with an alias table so a venue, its operator and its management company count as one employer, not three phantom jobs. Pay is copied as printed, never estimated.

One runs row per day; one email, only if the ledger changed; no public URL, no fetch handler at all. Extractor candidates are evaluated, not adopted: a Workers-AI Llama model was tested twice (Oct 2026 and again Nov 2026) and rejected both times on the same measured defect — it promoted “[FOR HIRE]” self-advertisements into job rows and filled employer fields with respondent usernames.

Code

workers/trapline/src/index.ts — the Worker’s head (exact):

// Trapline circuit, Worker body — ported from an earlier no-code automation body that stopped when its
// hosted trial ended.
// proposals/N8N-TO-WORKERS-MIGRATION-2026-10-18.md §3. No fetch handler: the Worker answers no requests.

import { isRunHour, modelText } from "./logic.ts";
import { viaAnthropic, withFallback } from "./engine.ts";
import { webSearch } from "./parallel.ts";
import { runCircuit } from "./run.ts";

Achievements

  • Reliability, measured live from its D1 ledger on 2026-11-11: 25/25 daily runs from 2026-10-18 to 2026-11-11, 0 errors, 69 patch visits (max 12/day).
  • Survived a platform death: the n8n Cloud trial lapsed and deleted the original workspace; the circuit was re-implemented as a Worker from the exported workflow code and kept its ledger semantics intact.
  • Acceptance tests (A1, A3, A4, A5 plus delta, idempotence and seed tests) ship in the repo and run on node:sqlite.

Lessons

  • “Free tier” is a dependency like any other. The migration spec exists because a trial’s end date was treated as an architecture input from the start.
  • Test extractors on their failure modes, not their demos. The rejected model looked fine until its rows were audited field by field — twice, a month apart, with the same result.
  • Email only on change. A monitoring system that reports “nothing” every day trains its reader to ignore the day it matters.

Status: Active, in production · TypeScript, Cloudflare Workers cron, D1, Parallel, Claude Haiku / Workers AI