# `defaultChatModel` required for embedding-only openai-compatible providers — blocks startup

[gokcanakalin]

Summary

ProviderRegistry.fromConfig rejects openai-compatible providers that omit
defaultChatModel, even when the provider is embedding-only. This forces
users to keep a dummy/placeholder chat model on embedding providers, and any
attempt to "clean up" the redundant key makes every startup mode
(tui, run, serve) fail.

Environment

• atomic-agent 0.6.0 (Linux, x86-64, self-contained binary)

Steps to reproduce

1. Define an embedding-only provider in ~/.atomic-agent/config.json:

{
  "llm": {
    "activeTextProvider": "openrouter",
    "activeEmbeddingProvider": "local-embed",
    "providers": [
      { "id": "openrouter", "kind": "openai-compatible", "baseUrl": "https://openrouter.ai/api/v1", "...": "..." },
      {
        "id": "local-embed",
        "kind": "openai-compatible",
        "apiKey": "not-needed",
        "baseUrl": "http://127.0.0.1:11435",
        "defaultEmbeddingModel": "nomic-embed-text-v2-moe-Q8_0.gguf",
        "userModels": [
          { "id": "nomic-embed-text-v2-moe-Q8_0.gguf", "kind": "embedding", "dim": 768 }
        ]
      }
    ]
  }
}

2. Run atomic-agent tui (or run / serve).

Actual behavior

Immediate fatal error; the TUI never renders:

Error: openai-compatible provider "local-embed" requires baseUrl and defaultChatModel
    at ProviderRegistry.fromConfig
    at async createAgentRuntime
    at async tuiCommand

Expected behavior

An embedding-only provider (all userModels[].kind === "embedding", and/or
selected only via llm.activeEmbeddingProvider + defaultEmbeddingModel)
should not require defaultChatModel.

Why this matters

• The config schema already distinguishes embeddings explicitly
  (defaultEmbeddingModel, userModels[].kind: "embedding",
  llm.activeEmbeddingProvider), so the chat-model requirement contradicts
  the schema's own design.
• Users must keep a placeholder chat model that doesn't exist on the
  embedding server (e.g. defaultChatModel: "z-ai/glm-5.3-flash" on a local
  llama.cpp endpoint), which is confusing and looks like a misconfiguration.
• The validation only fires at runtime startup, not at config-write time, so
  a "harmless cleanup" edit breaks the agent with no warning until the next
  launch.

Suggested fixes (any of)

1. Skip the defaultChatModel requirement when all of a provider's
   userModels are kind: "embedding" (or when defaultEmbeddingModel is
   set and the provider is not the active text provider).
2. Default defaultChatModel to the first kind: "chat" user model (error
   only if none exists AND the provider is used for chat).
3. At minimum, validate the config at write time (atomic-agent config set)
   and improve the error message to explain embedding-only providers still
   need the key.

Workaround

Keep any defaultChatModel value on embedding-only providers, e.g.:

"defaultChatModel": "placeholder"