providers: custom and plugin-added providers never become available #53912

Summary

On OpenCode 2.0.24 a provider that is not in the built-in catalog can never be made available — neither via the providers block in opencode.json, nor from a plugin via ctx.provider.transform -> editor.add({ info, models }). The provider/models do appear in the plugin registry (ctx.provider.list() / ctx.model.list()), but they are absent from GET /api/provider and GET /api/model, and opencode run --model <provider>/<model> fails with Model unavailable. This blocks using any custom/local OpenAI-compatible server (e.g. llama.cpp / llama-swap).

Related observations:

• Config providers is applied to catalog providers (adding a model to opencode works), but a new provider id is parsed (visible in GET /api/config) and never materialized.
• integrationID pointing at a real integration with an active key credential, sourceConnection, and canonical do not help.
• Built-in local discovery does issue requests (LM Studio GET /api/v1/models on 127.0.0.1:1234; vLLM GET /health + GET /v1/models on 127.0.0.1:8000) but no active provider appears either.
• Plugin console.log does not reach opencode.log.

Environment

• opencode version: 2.0.24 (@opencode/cli, npm latest)
• OS: Microsoft Windows [Version 10.0.19045.6332] (Windows 10 22H2, x64)
• Terminal: Unavailable: TERM=xterm-256color, TERM_PROGRAM/COLORTERM unset
• Shell: C:\WINDOWS\system32\cmd.exe
• Install/channel: release
• Active plugins: localai.discovery (global), opencode.models-disabler (global); no plugins entry in config

Reproduction

Custom provider via config:

1. In the global opencode.json (Windows: %USERPROFILE%\.config\opencode\opencode.json) add:

   {
     "providers": {
       "acme": {
         "name": "Acme",
         "env": ["ACME_API_KEY"],
         "package": "@opencode/ai/providers/openai-compatible",
         "settings": { "baseURL": "https://llm.acme.example/v1" },
         "models": { "qwen3-coder": { "name": "Qwen 3 Coder" } }
       }
     }
   }

2. opencode service restart

3. opencode api get /api/provider -> acme is absent

4. opencode run --model acme/qwen3-coder "hi" -> Error: Model unavailable: acme/qwen3-coder

Same provider via a plugin:

1. ~/.config/opencode/plugins/foo/index.js:

   export default {
     id: "foo",
     async setup(ctx) {
       await ctx.provider.transform((editor) => {
         editor.add({
           info: {
             id: "acme",
             name: "Acme",
             activation: "enabled",
             package: "@opencode/ai/providers/openai-compatible",
             settings: { baseURL: "https://llm.acme.example/v1" },
           },
           models: [/* one valid Model.Info object */],
         })
       })
     },
   }

2. Inside the plugin, ctx.provider.list() shows acme and ctx.model.list() shows its model.

3. opencode api get /api/provider still lacks acme; opencode run --model acme/<model> -> Model unavailable.

Expected Behavior

Per the V2 docs (Providers -> Custom, and build/plugins -> ctx.provider.transform / editor.add), a non-catalog provider with a valid package, settings.baseURL and at least one model should become available and selectable, the same way catalog providers do.

Actual Behavior

The provider stays inactive and its models are unselectable.

SessionRunnerModel.ModelUnavailableError: Model unavailable: acme/qwen3-coder

GET /api/provider returns only opencode and opencode-go (both have integrationID: "opencode" and an active OAuth credential).

Additional Context

Things that DID work:

• Adding/overriding a model on a catalog provider via providers (e.g. providers.opencode.models.config-test-model -> opencode/config-test-model resolves).
• Adding a model under an available provider from a plugin: ctx.model.transform(editor => editor.update("opencode", alias, draft => ...)).

Things that did NOT work:

• New provider id via providers (not even materialized).
• ctx.provider.transform -> editor.add (registered but never active), including with an existing integrationID + active credential, with sourceConnection, and with canonical.
• Built-in LM Studio / vLLM discovery endpoints answering correctly at their default ports.

Workaround in use (confirms the catalog-provider path works): attach local models to the active opencode provider as alias models with modelID set to a catalog model, model-level settings.baseURL pointing at the local server, and body.model overriding the outgoing model name. This makes opencode/local-<model> selectable and functional.