feat: add Ollama as an LLM provider (Ollama Cloud, default gpt-oss:120b-cloud)

[omjeem]

Summary

This PR adds Ollama as a third LLM provider alongside Anthropic and OpenAI, using Ollama Cloud's OpenAI-compatible endpoint. The goal is to make it easier to run Town with a free or lower-cost model, without requiring an Anthropic or OpenAI API key.

Changes

• Added apps/web/src/lib/ollama.ts using @ai-sdk/openai-compatible, so it integrates with the existing AI SDK call sites without requiring call-site changes.

• Added support for:

  • LLM_PROVIDER=ollama (chat)
  • CREATOR_PROVIDER=ollama (creator)

• Defaults to https://ollama.com/v1, with OLLAMA_BASE_URL available for local Ollama instances.

• Added the required configuration to .env.example.

• Existing deployments are unaffected—the provider priority remains anthropic → openai → ollama, so adding an Ollama configuration won't change existing behavior.

Notes for Review

• @ai-sdk/openai-compatible is intentionally pinned to 2.0.57. Version 3.x targets provider spec v4, while ai@6.0.205 still uses v3, which causes type-check failures.
• supportsStructuredOutputs: true is required. Without it, the SDK omits the JSON schema, causing gpt-oss to return invalid JSON for the moderator's generateObject. Enabling it correctly sends the schema as response_format.

The default model is gpt-oss:120b-cloud, which works with both Ollama Cloud and local Ollama. This can be overridden using OLLAMA_MODEL.

Verification

Tested against Ollama Cloud:

• ✅ Typecheck passes
• ✅ Build passes
• ✅ Structured outputs (test-moderator.ts)
• ✅ Tool calling
• ✅ Streaming
• ✅ Verified no regression with the default Anthropic flow

Out of Scope

lib/llm/router.ts (currently unused) and openai-image.ts (OpenAI-specific) were intentionally left unchanged.

[harshithmullapudi]

hey @omjeem where you able to run a town with the npc and having a chat?

[omjeem]

Yep! Ran a town locally with LLM_PROVIDER=ollama on gpt-oss:120b-cloud and had real NPC conversations end-to-end.

NPC chat (Founder) - streaming, in-context replies

Creator/town-building chat - works

Moderator routing (generateObject structured output) - smoke test passes all cases

All three call shapes (streaming, structured output, tool calling) verified against Ollama Cloud.