[rekram1-node]

I made model accept a "provider/model" string anywhere a text, image, or decision model is accepted, in both the Promise and Effect APIs. A model object still works for anything that needs settings.

await ai.llm.generate({ model: "openai/gpt-5.4", prompt: "Hi" })
await ai.llm.generate({ model: OpenAI.model("gpt-5.4", { baseURL }), prompt: "Hi" })

The text before the first / picks the provider and the rest is the model id, so openrouter/openai/gpt-5.4 sends openai/gpt-5.4 to OpenRouter. The string loads that provider's default model, imageModel or decisionModel with its env key and no settings. Each provider module is imported only when a string names it, inside the Effect, so failures are a typed AIError:

• unknown provider or no /: lists the known ids
• provider without that kind of model: Provider "anthropic" has no image models
• provider that needs settings, such as azure/...: the provider's own error (requires resourceName or baseURL)

Provider options for a string model are untyped. LLM.request still needs a model object because it builds the request synchronously, and so does compact, which takes a built request. This is a stopgap until strings resolve through a bundled catalog.

Accepted ids are each provider module's id, plus these models.dev ids where ours differ:

models.dev id	provider
fireworks-ai	fireworks
vercel	vercel-ai-gateway
google-vertex-anthropic	google-vertex (Messages API)

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" fill="none" stroke="currentColor" font-family="ui-monospace,monospace" font-size="11">
  <style>
    .s{animation:s 3s ease-in-out infinite}
    .m{animation:m 3s ease-in-out infinite}
    @keyframes s{0%,35%{opacity:1;transform:translateX(0)}55%,100%{opacity:.25;transform:translateX(6px)}}
    @keyframes m{0%,40%{opacity:.3}60%,100%{opacity:1}}
    @media (prefers-reduced-motion:reduce){.s,.m{animation:none}}
  </style>
  <g class="s"><rect x="8" y="22" width="112" height="28" rx="6" opacity=".5"/><text x="18" y="40" fill="currentColor" stroke="none">"openai/gpt-5.4"</text></g>
  <path d="M128 36h22m-6-5 6 5-6 5" opacity=".6"/>
  <g class="m"><rect x="158" y="14" width="74" height="44" rx="8"/><text x="168" y="33" fill="currentColor" stroke="none">OpenAI</text><text x="168" y="48" fill="currentColor" stroke="none" opacity=".6">.model()</text></g>
</svg>

---

Added by the opencode bot (AI-generated).

Public API: model accepts "provider/model" strings in the LLM, Image and Decision APIs and the ai client (backwards compatible)

[opencodejr]

The change works as described. I ran the new model-string tests and the full packages/ai suite, and all pass. The package's typecheck script needs bun check, which wasn't available here, so I ran its tsc fallback for both configs (including the new *.types.ts) instead; it passes. I also type-checked packages/core, which uses this package, and it passes. Strings resolve lazily inside the Effect, failures come back as a typed InvalidRequest, Object.hasOwn rejects names like constructor/x, and model objects and already-built requests are passed through unchanged.

Public API (@opencode/ai, backwards compatible):

• model now also accepts a "provider/model" string in:
  • LLM.generate, stream, loop and generateObject
  • Image.generate, stream, start and resume
  • Decision.generate
  • the matching ai.llm, ai.image and ai.decision methods
• The Model constraints of RequestInput, LoopInput, GenerateObjectOptions, GenerateObjectDynamicOptions, ImageRequestInput and DecisionRequestInput widen from X to X | string.
• LanguageModelProviderOptions, ImageModelOptions and DecisionModelOptions now map string to the untyped options instead of never.
• LLM.request and compact still need a model object.

Existing callers that pass model objects are unaffected.

One small inline note on the type test.

[opencodejr]

providerOptions is flat everywhere else (it goes to the route as-is, e.g. request.providerOptions?.store), so { openai: { store: false } } only type-checks because string models take an untyped Record<string, unknown>. At runtime it sends an openai key that the OpenAI route ignores. Could this use the flat shape ({ store: false }), so the test doesn't suggest a namespaced form that doesn't work?

[opencodejr]

The new commit fixes my earlier point: the string-model type test now uses flat providerOptions. The type-test config (tsconfig.types.json) still passes, and I have nothing new to flag. The public API impact is the same as before: model also accepts "provider/model" strings, and existing code that passes model objects still works.