feat(sdk): enumerate provider models of every modality in the model catalog

[heavy-d]

The SDK model catalog (/api/sdk/v1/models) built its list from getAllModels, which only enumerates language models from configured providers. Image, TTS, music, ASR, video, and embedding models appeared only when they happened to be in RECOMMENDED_MODELS — an SDK client asking for image_model compatibility got exactly one entry (GPT Image 2) while the web editor's picker, which goes through availableForKind, showed the full provider lists.

Add collectProviderCatalogModels: one pass per configured provider calling each getAvailable*Models once, no task filtering, so text_to_image and image_to_image capable models are both included. Per-list failures degrade to an empty list instead of dropping the provider.

The catalog service merges these into the existing gathering behind a 60s per-user TTL cache. Only the remote provider enumeration is cached — local download state (HF cache scan, download manager) stays fresh on every call, so a finished download still flips to ready_local immediately. Worker-scoped catalogs are unchanged and never enumerate providers.

getAllModels and availableForKind are untouched; existing consumers keep their exact behavior.

[Copilot]

Pull request overview

This PR expands the SDK model catalog endpoint (/api/sdk/v1/models) to include provider-enumerated models for non-LLM modalities (image, TTS, music, ASR, video, embeddings) so SDK consumers see the same breadth of models as the editor’s provider pickers, while keeping provider enumeration behind a short per-user TTL cache.

Changes:

• Add collectProviderCatalogModels() to enumerate all non-language-model modalities per configured provider (one call per getAvailable*Models per provider, no task filtering).
• Cache provider enumeration results for 60s per user in getSdkV1ModelCatalog, with an optional injection hook for tests/callers.
• Add Vitest coverage for inclusion, dedupe behavior, caching, and worker-scope behavior.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
packages/websocket/src/trpc/routers/models.ts	Adds collectProviderCatalogModels() to enumerate provider models across modalities for the SDK catalog.
packages/websocket/src/sdk/sdk-model-catalog-service.ts	Merges provider-enumerated models into the SDK catalog with a per-user TTL cache and test hook.
packages/websocket/tests/sdk-model-catalog-provider-models.test.ts	Adds tests validating provider-enumerated entries, dedupe, caching, injection, and worker-scope behavior.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (2)

packages/websocket/src/sdk/sdk-model-catalog-service.ts:182

• providerCatalogCache is a process-global Map keyed by userId but entries are never pruned. In a long-lived server handling many users, this can grow unbounded even though values are treated as TTL’d, since expired entries remain resident until overwritten for the same user.

      providerCatalogCache.delete(key);
    }
  }
}

packages/websocket/src/sdk/sdk-model-catalog-service.ts:227

• This code path ends up re-running provider configuration checks multiple times per request: getAllModels() calls getAvailableProviderIds() internally (models.ts:784), collectProviderCatalogModels() calls it again (models.ts:942), and getSdkV1ModelCatalog() also calls it here to build configuredProviderIds. On instances with many registered providers/secret backends, this adds avoidable latency and load.

Consider fetching providerIds once in getSdkV1ModelCatalog and threading it into provider enumeration (e.g., add an optional providerIds parameter to collectProviderCatalogModels / cache helper), so at least the catalog-specific enumeration doesn’t re-check configuration.

  let providerIds: readonly string[];
  if (args.query.scope === "worker") {
    if (!args.getWorkerModels) {
      throw new SdkModelCatalogServiceError(
        "Worker model catalogs are not available through this server."

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (1)

packages/websocket/src/sdk/sdk-model-catalog-service.ts:240

• getSdkV1ModelCatalog now calls getAvailableProviderIds directly, but collectProviderCatalogModels (invoked via getCachedProviderCatalogModels) also calls getAvailableProviderIds internally, and getAllModels does the same. That means each catalog request can re-run provider configuration checks multiple times and can (rarely) observe different provider-id snapshots inside one response.

Consider fetching providerIds once in getSdkV1ModelCatalog and threading that list into the provider-catalog enumeration (and ideally into getAllModels) so the request uses a consistent set and avoids repeated secret/config checks.

    [availableModels, providerCatalogModels, providerIds] = await Promise.all([
      getAllModels(args.userId),
      (args.getProviderCatalogModels ?? getCachedProviderCatalogModels)(
        args.userId
      ),
      getAvailableProviderIds(args.userId)
    ]);