Skip to content

[FEATURE]: Support decision models from local OpenAI-compatible providers (LM Studio /v1/systemone) #64306

Open

Description

@Mawericksracing

Description

LM Studio 0.4.28 ships native decision models and a Jev / System One compatible endpoint: POST /v1/systemone accepting { model, state, questions } and answering noul (predicate), choice and score questions. That is the same protocol OpenCode already speaks through @opencode/ai/providers/typesafe-ai, so a fully local decision backend is now technically available — for example rizzo-flow, a 4B Spark X2.5 LoRA with 1M context, ~2.6 GB at Q4_K_M.

OpenCode already exposes decisions (/api/experimental/generate/decisions) and supports them for 8 built-in cloud providers (baseten, cloudflare-ai-gateway, cloudflare-workers-ai, digitalocean, opencode, openrouter, typesafe-ai, vercel-ai-gateway). There is no way to point them at a local endpoint:

  • @opencode/ai/providers/openai-compatible has no decisions module.
  • Declaring "type": "decision" on a custom provider's model makes the model disappear from the catalog. A/B is unambiguous: without type the model appears in /api/model, with type: "decision" it is silently dropped (same result with a per-model package: "@opencode/ai/providers/typesafe-ai").
  • Overriding the native lmstudio provider is ignored: the log reports retained native value over legacy value.

Result: POST /api/experimental/generate/decisions for lmstudio/rizzo-flow returns

422 UnsupportedModelError
lmstudio/rizzo-flow is a chat model, not a decision model

Why it matters: local decisions would enable privacy-preserving, zero-cost, rate-limit-free classifiers for permission auto-approval, context pruning plugins (fast-jev-compaction style), routing and retry logic. No API key, no data leaving the machine.

Proposed change (any of these unblocks it)

  1. Add a decisions transport to openai-compatible that posts to {baseURL}/systemone (or {baseURL}/decisions) when a model is declared type: "decision". This also covers llama.cpp's System One API and any Jev-compatible server, not just LM Studio.
  2. Let a provider or a single model declare its decision endpoint explicitly (e.g. decision: { baseURL, model }), decoupling the capability from the 8 hardcoded provider packages.
  3. At minimum, stop discarding type: "decision" models from custom providers, and route them through the System One protocol.

Steps to reproduce

  1. LM Studio 0.4.28: download a decision model (rizzo-flow) and start the server on 127.0.0.1:1234.
  2. Declare a custom provider:
"providers": {
  "local-rizz": {
    "package": "@opencode/ai/providers/openai-compatible",
    "settings": { "baseURL": "http://127.0.0.1:1234/v1" },
    "models": {
      "rizzo-flow": { "name": "rizzo-flow 4B", "type": "decision" }
    }
  }
}
  1. opencode api get /api/model -> no local-rizz models.
  2. Remove "type": "decision" -> the model appears, but the decisions endpoint rejects it.
  3. POST /api/experimental/generate/decisions with {"model":{"providerID":"lmstudio","id":"rizzo-flow"},"input":[...],"questions":[...]} -> 422 UnsupportedModelError.

Environment

  • opencode 2.2.5
  • Windows 11 build 26300.9974, Windows PowerShell 5.1
  • LM Studio 0.4.28, POST http://127.0.0.1:1234/v1/systemone reachable and working

Related

Activity

  1. opencodejr commented on Nov 16, 2026

    @opencodejr
    Contributor

    Thanks for the detailed request and the clear A/B test.

    You're asking for decisions (/api/experimental/generate/decisions) to work with local decision models served over LM Studio's System One endpoint (/v1/systemone), from the native lmstudio provider or from a custom OpenAI-compatible provider.

    Most of this has already been worked on: #63200 ("feat(core): run LM Studio decision models") was merged on the main branch after 2.2.5. It adds:

    • a new @opencode/ai/providers/openai-compatible/decisions entrypoint that sends System One requests to the provider's base URL;
    • LM Studio discovery that now lists decision models as decision models with that package, instead of filtering them out;
    • support for running a local decision model without an API key (before, this failed with Missing auth credential: apiKey).

    From what I can tell, this change is not in a published release yet, so 2.2.5 behaves as you describe. Once a release includes it, lmstudio/rizzo-flow should be listed as a decision model.

    On a model disappearing when it has "type": "decision": the providers docs say type "Defaults to chat; only chat models are offered for sessions". That is probably why it drops out of the list you checked. On a release that includes #63200, setting the model's package to @opencode/ai/providers/openai-compatible/decisions (instead of the TypeSafe one) together with type: "decision" is the closest match to what the PR adds. I have not verified that this config works end to end.

    Your second proposal, an explicit per-provider or per-model decision endpoint setting, is not covered by that PR. Maintainers will consider it along with the rest of this request.

    Related:

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions