Reasoning efforts need an owner

1. Symptom

2. Why the partial fix is not enough

The patch only changes merge order when connection capabilities carry an empty list:

// current band-aid
...(caps.reasoningEfforts?.length
  ? { reasoningEfforts: caps.reasoningEfforts }
  : {}),

That unblocks catalog-derived values for the common “editor saved []” path. Three independent defects remain, any one of which still yields a wrong menu.

const efforts = model?.reasoningEfforts?.length
  ? model.reasoningEfforts
  : [...EFFORTS]; // none…ultra

3. Ownership model

Pick one source of truth per field, with a single merge function and no silent maximal fallbacks.

4. Catalog resolution policy

models.dev is useful and noisy. Treat it as a ranked lookup, not a bag union.

// target shape (illustrative)
function pickCatalogModel(catalog, modelId, hints?: { baseUrl?: string; providerHint?: string }) {
  const candidates = entriesEndingWith(catalog, modelId);
  if (candidates.length === 0) return null;
  const ranked = rankCandidates(candidates, hints); // vendor URL, provider id, option richness
  return ranked[0]; // ONE entry — do not fold options across the set
}

function effortsFromEntry(entry): ReasoningEffort[] | undefined {
  // toggle → include "none"; effort.values filtered to SUPPORTED_EFFORTS
  // no effort values but reasoning → ["none","high"]
  // reasoning false → undefined
}

5. Correct design (end-to-end)

5.1 Persistence semantics

5.2 Shared resolver module

Move pure resolution out of “frontend fetch helper” identity. Suggested home: packages/protocol or a small shared lib both Frontend and Host import.

• loadModelsDevCatalog() may stay environment-specific (fetch vs cache file), but resolveModelCapabilities must be pure and unit-tested without network.
• Frontend connectionsToModelProviders becomes a thin adapter over the shared resolver.
• Host, when accepting a run / setting reasoning, clamps the selected effort to resolved.reasoningEfforts. Unknown effort → nearest supported or reject.

5.3 Custom backend editor

5.4 Selector contract

// stop inventing a menu
if (model.capabilities.reasoning === false) return null;
const efforts = model.reasoningEfforts; // already resolved
if (!efforts?.length) return null; // or show Off-only — never full EFFORTS

The full EFFORTS constant remains the allow-list for parsing, not the default menu.

5.5 Base URL → vendor hints

Small static map is enough for v1; do not scrape HTML.

const VENDOR_HOST_HINTS: Record<string, string> = {
  "api.deepseek.com": "deepseek",
  "api.openai.com": "openai",
  "api.anthropic.com": "anthropic",
  "api.x.ai": "xai",
  // ...
};

Custom OpenAI-compatible proxies that re-host DeepSeek still get model-id matching; the hint only breaks ties toward the vendor when the URL is first-party.

6. Phased delivery

7. Acceptance criteria

8. Non-goals (this plan)

• Building a full models.dev mirror service
• Per-reseller effort accuracy when user points at OpenRouter etc. (later: provider hint from base URL of that reseller)
• Budget-token UIs (budget_tokens options in catalog)
• Changing DeepSeek wire format beyond effort allow-listing
• Requiring users to hand-edit effort lists for the default path

9. Likely touch points