[FEATURE]: Add modalities config support for OpenAI-compatible providers — vision/image input broken by default #30071

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Problem
When using an OpenAI-compatible provider (e.g., Manifest, LocalAI, LM Studio, AnyScale) with a vision-capable model (Kimi 2.5, GPT-4o, Claude), image inputs are silently stripped before the API call is made. The model receives a text string: "ERROR: Cannot read image (this model does not support image input). Inform the user." — which is fabricated by the framework's unsupportedParts() check. The actual image_url content never reaches the provider.

Root Cause
@ai-sdk/openai-compatible hardcodes input: ["text"] for models from /v1/models responses since the OpenAI-compatible schema doesn't include modality metadata. When a user sends a message with image content, the framework checks model.capabilities.input.image — finds false — and strips the image.

Workaround (Confirmed Working in v1.15.12)
The modalities property on model definitions tells the framework the model accepts image inputs:

{
  "provider": {
    "manifest": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://localhost:2099/v1"
      },
      "models": {
        "static-image": {
          "name": "static-image",
          "headers": { "x-manifest-tier": "StaticImage" },
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        }
      }
    }
  }
}

After adding "modalities": { "input": ["text", "image"], ... }, the multimodal-looker subagent successfully reads and describes screenshots from Playwright.

Why This Needs Fixing

• Undocumented — The modalities property exists in packages/opencode/src/provider/models.ts but isn't in the config schema docs. Users have no way to discover it.
• Bad defaults — Auto-discovered OpenAI-compatible models default to input: ["text"], which silently breaks vision for any model that actually supports images.
• Others fixed this — KiloCode (PR Disable force dependency update while starting #8729) had the same issue and fixed it by defaulting input.image and attachment to true for auto-discovered models. OpenCode should do the same, or at minimum document the modalities escape hatch.

Suggested Fixes

• Default input: ["text", "image"] for auto-discovered OpenAI-compatible models — matches what KiloCode did. Most OpenAI-compatible providers expose at least one vision-capable model, and the worst case (model doesn't support images) is a graceful API error, not silent data loss.
• Document modalities in the provider config schema — add it to opencode.ai/config.json and the docs so users can opt in per-model.
• Allow /v1/models capabilities passthrough — if a provider's /v1/models response includes any capability fields, forward them to the AI SDK.