LinkedIn Post (Text-Only, Teaser)

Pointing OpenCode V2 at a provider you run yourself isn’t hard. It’s just full of silent failures. I hit nine, and trap #1 cost me two days while the error blamed the wrong machine.

Trap #1: the provider id is a hyphen, not a slash.

openai-compatible/responses → wrong openai-compatible-responses → right

The wrong one still resolves. It just selects a different transport, so OpenCode sends a Responses-shaped body down the chat-completions path. Nothing errors at the lookup. You find out three layers later.

Field notes from the rest of the list:

• A V1-shaped config block is silently ignored by V2. Different key names, no warning.
• A credential in env only works in the shell that has it. Missing auth credential: apiKey is the symptom.
• baseURL with a leading space is a valid JSON string and an invalid URL. Use cat -et on macOS.
• Doubt the client before the server. A 30-second curl ruled out my server entirely.
• One error message is usually two failures. Read from where the request is issued, not where the error is printed.
• A UI reflects intent; the artifact reflects reality. Grep what was actually written.
• Preview data isn’t prod data. 410 ModelDeprecated in preview, 200 in prod, same code.
• Back up your config before any tool writes it. That one can cost data, not just time.

#OpenCode #DeveloperTools #Debugging #TechNotes #AI

---

Comment (First Comment)

Full story on Medium → https://medium.com/@vianhanif/[slug]

---

Notes

• Schedule via Fedica for D+1 (blog/Medium publish day + 1)
  • Post text-only (no links in main body)
  • CTA: list format performs well for saves/shares
  • Comment: “Full story on Medium → https://medium.com/@vianhanif/[slug]”
  • Schedule comment to post immediately after LinkedIn post goes live
  • Wait 10-15 minutes before engaging with comments