← All projects

AI PROJECT · CURRENTLY WORKING ON

Custom Connector Testing, Maintenance & Upgrades

Standing care for our 12 custom connectors: sweep them against live data, fix the ones that answer confidently while holding only part of the truth, and ship upgrades as each vendor's API moves.

⏱ Est. 36 hours (David — rolling estimate; the open items below sum to it) Data / EngineeringPriority: HIGHUpdated August 5, 2026

What & why

Our custom connectors are software we own, so they need the same upkeep as anything else we run. They drift when a vendor changes an API, they get starved of rate-limit credits, and — the dangerous one — they can keep answering questions cleanly while holding only a fraction of the data. The sweep on August 5, 2026 proved it: three connectors were succeeding with no error and no warning while returning partial truth. Four fixes shipped the same day (ApparelMagic coverage reporting plus deleted-record reconciliation, the LoudCrowd health-check timeout, NuORDER environment confirmation, and removing Gorgias's write tools from the code outright). This project is the standing home for the rest of that list, for the next sweep, and for every upgrade and regression test after it — so connector health has one owner and one place to look.

The plan — step by step

  1. ShipHero credit pacing (10h): credit-aware scheduler that checks remaining API credits before each batch and backs off instead of failing the pass — then re-run the purchase_order and return backfills and settle whether returns are genuinely empty or just starved
  2. Coverage blocks everywhere (10h): give every analytical tool the coverage block {from, to, rows, complete} that ApparelMagic and LoudCrowd now have — the one change that turns silent partial data into visible partial data across all 12 connectors at once
  3. LoudCrowd commission sync (4h): cure the nightly 504 on the ambassadors program (smaller pages plus retry tuning, or a support ticket to LoudCrowd) — today it is visible as a warning in every health check, but not yet fixed
  4. Second sweep pass (6h): repeat the live-call sweep across the stock connectors, which the first pass only spot-checked, using the same cross-check-in-both-directions method
  5. Regression tests and a written playbook (6h): turn the sweep into a repeatable checklist on a schedule, so drift gets caught by the calendar instead of by someone quoting a wrong number in a meeting
  6. Vendor upgrades as they land: every connector's API keeps moving (LoudCrowd's Brand API added per-order commission mid-2026; ApparelMagic's list API has date-filter quirks that silently return zero rows) — those get triaged in here rather than each becoming its own project

Reasoning & trade-offs

The failure mode we are actually defending against is not a connector that breaks. A broken connector gets investigated. A connector that cleanly answers with a third of the data gets quoted in a meeting.
Method rule from the first sweep: never trust a connector's own health flag — and never trust a plausibility floor until you have confirmed both sources measure the same business. That second half caught a misdiagnosis: ApparelMagic looked CRITICAL until we confirmed the ERP only went live mid-June 2026, so there was no older history to be missing.
Coverage blocks are the highest-leverage item on the list and among the cheapest — one extra field per tool response, and every partial answer starts announcing itself.
The hours are a rolling estimate, not a quote. This project never fully closes; it is the maintenance line for software we own, and work joins it every time a vendor ships a change.
Kept as one card instead of a dozen tickets so connector health has a single owner. Anything that grows into a real build — a new connector, a new pipeline — graduates to its own project.

Related material

Live connector status — auto-checked every 5 minutes  ·  Hub — the custom connector guides  ·  Sweep method and findings — HANDOFF_Connector_Data_Health_2026-08-05.md (project folder)  ·  Same-day outcomes and corrections — RESOLVED_Connector_Data_Health_2026-08-05.md (project folder)

← Back to all projects
AI Projects v1.57  ·  MELINDA MARIA × CLAUDE  ·  updated August 5, 2026
Work in progress — suggestions to ✉️ David or 💬 #mm-claude  ·  Onboarding Hub  ·  © Melinda Maria 2026, All Rights Reserved