Repository navigation
Built-in vLLM discovery causes recurring 401 requests to oMLX on port 8000 #54347
Description
OpenCode v2.0.26 on macOS Apple Silicon automatically polls an authenticated oMLX server at http://127.0.0.1:8000/v1 every 30 seconds, producing GET /v1/models → 401: API key required.
oMLX is configured as a separate custom OpenAI-compatible provider with a valid API key. Chat completions work. I do not use vLLM.
Reproduction:
- Run oMLX with API-key authentication on port 8000.
- Configure the omlx provider with its API key and baseURL.
- Start the OpenCode background service.
- Observe recurring 401 messages in the oMLX logs.
Stopping the OpenCode service stops these requests. Adding an explicit Authorization header to the omlx provider does not resolve them.
Confirmed workaround:
Add "plugins": ["-opencode.provider.vllm"] at the configuration root, then restart the service. No new 401 messages appeared during approximately two minutes with the service running.
Source analysis at commit 9b4ec57:
- packages/core/src/plugin/provider/vllm.ts:14–26 defaults to port 8000 and requests /health and /v1/models.
- packages/core/src/plugin/provider/local.ts:34 sets a 30-second interval.
- local.ts:113–116 ignores discovery errors and continues polling.
- local.ts:142–193 uses the discovery provider API key.
- configured.ts:15–19 selects settings by provider ID, so omlx credentials do not authenticate the separate vllm probe.
Expected behavior: Avoid indefinitely polling an authenticated non-vLLM server merely because it occupies port 8000. Consider opt-in discovery, endpoint collision detection, or backoff after authentication failures.
Activity
opencode-agent commentedon Oct 10, 2026
Thanks for the detailed report and source analysis! I think this is a duplicate of #45172, which reports the same thing: OpenCode polls whatever is listening on
localhost:8000every 30 seconds and keeps going even when every request fails. A comment there traces it to the built-in vLLM discovery plugin and gives the same workaround you found ("plugins": ["-opencode.provider.vllm"]). The only difference is the status code. Your oMLX server answers/healthsuccessfully, so the probe goes on to/v1/modelsand gets a 401 there, while in #45172 it stopped at a 404 on/health.Possible duplicates and related issues, ordered by likelihood:
localhost:8000every 30s with no backoff after errors; same workaround127.0.0.1:8000The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different. Your suggestions (opt-in discovery, backoff after auth failures, collision detection) would be useful additions on #45172.
opencode-agent commentedon Oct 11, 2026
We haven't heard back, so I'm closing this as a duplicate of #45172. Both are caused by the built-in vLLM discovery plugin polling
localhost:8000every 30 seconds without backing off. Please follow #45172 for updates. Your suggestions (opt-in discovery, backoff after auth failures, endpoint collision detection) would be useful additions there. Until then, the workaround you found still applies: add"plugins": ["-opencode.provider.vllm"]to your config. If your problem turns out to be different, reply here and we can reopen. Thanks for the detailed report!