Five first-week blockers when wiring coding agents to an OpenAI-compatible backend
ObviousSpace8195 · reddit · 2026-11-03
A developer shares the most common first-week stalls teams hit when wiring coding agents (Cursor / Claude Code / Codex) to an OpenAI-compatible backend that bundles models, auth, DB and storage behind one project key — plus quick fixes:
- Stale env vars win: leftover OPENAI overrides beat the new baseURL/key. Print the resolved baseURL on the first request and clear overrides per shell/session.
- Assuming auth/db/storage bill like tokens: they're usually free at the project level; only model tokens cost money.
- Treating the gateway as a second product to install: same key, OpenAI-compatible calls — just change the model id.
- Agent losing context in a new chat: keep the skill + linked project in repo/session notes so clients don't re-login every thread.
- Five dashboards before one happy path: script one call (chat + one DB write or auth check) before expanding.
Ops notes, not a product pitch; the author invites specific errors for diagnosis.
More from coding & agent
- A one-night AI-built project nets 333 points and ~22k visitors — Aizkmusic · 2026-11-03
- Five manual checks before trusting an agent actually finished its job — Kodenstudio · 2026-11-03
- Open-source AI pit crew for iRacing runs 5 agents on Kimi, checks every number against telemetry — JeremyCMorgan · 2026-11-03
- Agent control plane benchmark: Smalltalk-80 cascades hit 100% compliance vs 2.5% for Gherkin — RandalSchwartz · 2026-11-03
- Karpathy: one prompt to Claude Code deploys a YouTube thumbnail-swap browser extension — amaarora · 2026-11-03
- Cherry Pi proposes layered harness architecture: agent loops compose with governed workspaces — sull · 2026-11-03