feat(providers): add Ollama System One handler and decision-model discovery

[yourspraveen]

Draft: this depends on the shared POST /v1/systemone route, which is still being decided on #13987. Candidates: #14031, #14498, #14667, and now #15278 (a TypeSafe System One client, no public route yet). Once a base is chosen, I'll rebase onto it, connect this handler to the route, and mark the PR ready.

⚠️ base-red inherited: #15306

What

Ollama ships a local System One API (POST /v1/systemone, docs). It uses the same request/response format as TypeSafe Jev (choice / noul / score questions with calibrated probabilities). This PR adds the Ollama-specific pieces so ollama-local/<model> can serve it alongside the TypeSafe / OpenRouter backends proposed upstream. Models covered:

• Nimble and Tev (Ollama ≥ 0.35.0)
• Clef (27B) and Clef Flash (9B), tagged vision + decision (Ollama ≥ 0.35.1)

1. Discovery

• Ollama's decision capability maps to a systemone supported endpoint, spelled the same as feat(typesafe): add Jev System One API support #14498.
• A decision-only model no longer falls through to the images apiFormat.

2. Handler (open-sse/handlers/ollamaSystemOne.ts)

• Validation matches Ollama's own limits: 400 for invalid bodies, and 413 above 64 KiB, or above 32 MiB when images are present.
• Forwards only known fields, with the provider prefix stripped.
• images support for Clef / Clef Flash: base64 only. URLs and data URLs are rejected with a 400, as Ollama does.
• 60s default timeout to cover cold model loads; echoes the caller's model id; logs usage at $0 cost.
• Failure mapping:
  • 404 missing model → model lockout
  • 400 "not supported by System One" → returned as-is
  • 5xx / unreachable / timeout → connection cooldown
  • client abort → nothing marked
  • errors go through errorResponse() sanitization

3. Decision-only guard: keeping System One models out of the chat path

Ollama reports Clef as ["vision","decision"] with no completion. Before this guard, such a row looked like a chat model: /v1/models gave it no type, and a chat call went upstream only to come back as a 400.

• modelEndpointPolicy: a decision kind for systemone-only rows (chatSelectable: false). Nimble and Tev also advertise chat, so they stay chat.
• /v1/models tags systemone-only rows type: "decision", alongside embedding / rerank / image.
• decisionOnlyChatRejection runs in getModelInfoOrRetirementResponse, the resolver behind resolveModelOrError, which handles both direct chat calls and every combo target. It refuses such a model with a clear 400 before any credential is picked, so nothing is marked on the connection. The message reads as a model-scoped 400, so a combo advances to its next target (comboTargetDecision → advance). Only System One-capable providers (ollama-local) pay the stored-model lookup.

Related

• feat(providers): add TypeSafe System One client for Jev routing #15278 / feat(routing): add jev combo strategy driven by TypeSafe Jev #15279 (@miguelenes): a TypeSafe System One client, and a jev combo strategy that uses one System One choice call to order chat targets. That is the routing use described in feat(providers): add TypeSafe AI (Jev) as a provider + optional auto-routing classifier #13987 Part 2. Once the route lands, routing that judge call through /v1/systemone with a decision combo would let it use local Ollama models (Clef Flash / Nimble) with cloud fallback, not only TypeSafe. I proposed this on feat(providers): add TypeSafe AI (Jev) as a provider + optional auto-routing classifier #13987.
• feat(plugin): filter published catalog by capability presets #14554 (@maxmad64bis): capability presets for the OpenCode plugin catalog. The new type: "decision" is a natural thing for those presets to filter out of the chat picker.

Testing

• tests/unit/ollama-systemone.test.ts (23):
  • capability mapping, including Clef's vision + decision
  • validation, including images: base64 forwarded, URL / data URL / non-string rejected, the 32 MiB vs 64 KiB limits
  • URL building and failure classification
  • handler behaviour with injected fetch
• tests/unit/ollama-systemone-decision-guard.test.ts (5):
  • policy kind
  • /v1/models type
  • the 400 and that combos advance on it
  • chat / unknown / other-provider models untouched, with no lookup for other providers
  • resolveModelOrError refusing the model
• Regression: 102 test files touching the endpoint policy, the catalog classification and model resolution: 709 pass, 2 skipped, 0 fail.
• Live run against Ollama 0.35.0 on a Jetson:
  • tev1:4b → 200 (1.5s)
  • nimble → 200 (2.2s)
  • unpulled model → 404 + model lockout
  • gemma3:4b → 400, connection left untouched
• Clef / Clef Flash have not been run live yet; they need Ollama ≥ 0.35.1 and 11–18 GB.

Local quality gates

I ran the quality.yml fast-gate set locally against release/v3.8.52. The branch is rebased onto the current tip (23a1148486).

• Passing (36), including complexity-ratchets, file-size, mutation-test-coverage, typecheck:core, dashboard typecheck, cycles, and ESLint --max-warnings 0 on the changed files. The first run caught two problems:
  • handleOllamaSystemOne was at complexity 33. It's now split into small helpers with no behaviour change.
  • The handler test was missing from stryker.conf.json. It's added now.
• Not run locally: workflows, which needs actionlint and zizmor. This PR doesn't touch workflows.

Remaining before ready

• Rebase onto the chosen /v1/systemone base PR and connect the handler to the route
• Docs: docs/openapi.yaml, API_REFERENCE.md, LOCAL_OLLAMA_DOCKER.md
• changelog.d/features/15410-ollama-systemone.md
• Live test of Clef / Clef Flash (images path included) on Ollama ≥ 0.35.1

Refs #13987