
Field Notes - Dec 09, '25
Executive Signals
- Adapters are the unit: ship by proven adapters, not calendar promises
- Submit is the blast radius: validate and dry-run until nothing surprises
- Throttle the lab: shared back-offices see bursts as abuse, pace yourself
- Targeted beats bulk: dealer-level control raises signal, limits collateral damage
- DOM over OCR: selectors for IDs, screenshots for audit trail
Engineering
Gate Go-Live by Tested Adapters
Calendar dates don’t move risk; test coverage does. Treat each integration adapter as its own release unit. Run the scheduler in no‑op to surface failures without touching external portals, then enable via per‑adapter flags. Require end‑to‑end manual submits with evidence and CRM writeback before flipping production.
- Require 2–3 clean manual submits per adapter, then 24‑hour soak
- Capture form validation, evidence screenshot, reference ID, and CRM writeback
- Flip only the adapters that pass; if 3 of 10 pass, ship 3
Submit Is the Blast Radius
Design the pipeline so defects die before any external POST. Push strict preflight checks and a dry‑run path that exercises everything except “Submit.” Use idempotency to prevent dupes, and make errors diagnostic by logging exact vendor output. Keep “Submit” behind feature flags per adapter and environment.
- Implement preflight validation plus a full dry‑run path
- Use idempotency keys; log and surface exact error payloads
- Feature‑flag submit per adapter and per environment
Test Without Spamming Shared Vendors
Many OEM portals ride the same back‑office provider, so bursty testing can look abusive. Coordinate test windows, throttle volume, and clearly tag traffic. Maintain an audit trail to defuse disputes and accelerate whitelisting.
- Pre‑notify provider with dealer IDs and test window; request auto‑cancel
- Cap to ≤5 submits per dealer/day during testing; stagger across adapters
- Tag payloads as TEST where allowed; keep an auditable log
Targeted Dealer Runs Beat Bulk OEM Runs
Early in rollout, scope traffic narrowly. If per‑dealer triggers don’t exist yet, constrain via CRM status as a stopgap, then build first‑class controls. Provide clear resubmission states and backoff, and index reference IDs to reconcile retries cleanly.
- Ship API/CLI for per‑dealer submits and “reset to new”
- Implement resubmission states with backoff and explicit transitions
- Index reference IDs to de‑dupe and reconcile across retries
Evidence: DOM for IDs, Screenshot for Audit
Extract the authoritative reference number from the DOM; keep a screenshot for audit. Treat OCR as a fallback path only. Harden selectors with tests and drift alerts, and write both the reference ID and touchpoint back to the CRM.
- Ship selector tests; alert on DOM/selector drift
- Store screenshot and parsed ID; write both back to the CRM
- If DOM parse fails, fall back to OCR and flag for hardening