providers: vLLM discovery silently drops every model whose owned_by is not "vllm" #57460

Summary

The built-in vllm provider discovers models from /health + /v1/models, then silently drops every entry whose owned_by is not exactly vllm or that carries no max_model_len. Any OpenAI-compatible gateway that reports its real owners (LiteLLM, Portkey, GoModel, a plain nginx-fronted upstream) is therefore invisible, with no error in the logs, and the operator has to write the model list by hand.

Environment

• opencode version: 2.0.32
• OS: Linux 6.18.51-brainykernel-zen3 (linux x64)
• Terminal: TERM_PROGRAM=ghostty, TERM=xterm-256color, COLORTERM=truecolor
• Shell: /usr/bin/zsh
• Install/channel: latest
• Active plugins: oh-my-openagent-v2, opencode-skill-creator (unrelated to the failure; it reproduces with an isolated project config)

Reproduction

1. Serve an OpenAI-compatible /v1/models on http://localhost:8090/v1 whose entries look like a normal gateway:

{"object":"list","data":[
  {"id":"my-model","object":"model","owned_by":"anthropic","created":1792503066,
   "metadata":{"context_window":1000000}}
]}

2. Serve GET /health returning {"status":"ok"} on the same host.
3. Point OpenCode at it in a project opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "providers": { "vllm": { "settings": { "baseURL": "http://localhost:8090/v1" } } }
}

4. Run opencode models.

Expected Behavior

The discovered model appears as vllm/my-model. The runtime already reached the endpoint successfully: with request logging on, OpenCode performs GET /health and GET /v1/models, both 200.

Actual Behavior

No vllm/* model is listed and nothing is logged as an error. The discovery result is discarded silently.

Bisecting the payload one field at a time, with everything else held constant:

So both conditions are required, and metadata, created and object are irrelevant.

Additional Context

• owned_by is informational in the OpenAI schema. Requiring the literal vllm makes the runtime unusable against every gateway that reports a real owner, which is the common case: LiteLLM reports the provider, Portkey reports the upstream, GoModel reports the provider instance. Serving owned_by: "vllm" is a lie only a gateway maintainer can opt into.
• Requiring max_model_len is reasonable for a real vLLM server, but a gateway that has no such field still serves the models; treating it as a hard filter hides them.
• Workaround used meanwhile: the gateway emits owned_by: "vllm" and fills max_model_len from its model catalog behind an opt-in flag. Once this filter is relaxed, that flag becomes unnecessary. Verified end to end: with the two fields present, OpenCode lists the model and an agent runs shell tools through it.
• Related: V2: Reconcile provider catalogs with /models endpoints #40524 (reconcile provider catalogs with /models), Feature: Dynamic context length detection for OpenAI-compatible providers #40908 (dynamic context length detection), Built-in vLLM discovery causes recurring 401 requests to oMLX on port 8000 #54347 (vLLM discovery probing port 8000).

A suggestion: when discovery receives entries but publishes none, log a single debug/warn line explaining why (e.g. skipped N models: owned_by is not "vllm"). Silent discard is what made this take a full payload bisect to diagnose.