Running a BGE-M3 local embedding setup on v0.10.3 using the pluggable provider layer from #172. ~/.gbrain/config.json has the embedding settings persisted:

{
  "engine": "pglite",
  "database_path": "/home/user/.gbrain/brain.pglite",
  "embedding": {
    "provider": "ollama",
    "model": "bge-m3",
    "dimensions": 1024,
    "base_url": "http://localhost:8085/v1"
  }
}

When PR #188 merged I wanted to apply the new schema migrations, so I ran gbrain init --pglite and got:

Cannot re-init: existing brain has a different embedding dimension.
  Existing: ollama / bge-m3 (1024d)
  Requested: openai / text-embedding-3-large (1536d)

The "Existing" line proves config.json was read somewhere in the flow. The "Requested" line is the baked default, not anything I passed on the CLI. So every re-init fails unless I repeat the flags I already wrote down in the config file.

Workaround:

gbrain init --pglite --provider ollama --model bge-m3 --dimensions 1024 --base-url http://localhost:8085/v1

Which works but feels like the config file is there for a reason.

Looking at src/commands/init.ts::resolveProviderWithGuard(), it calls resolveEmbeddingConfig(parseEmbeddingFlags(args)) using env vars and CLI flags, without threading the loaded config's embedding section in as defaults.

Rough shape of a fix:

const existing = loadConfig();
const defaults = existing?.embedding
  ? {
      provider: existing.embedding.provider,
      model: existing.embedding.model,
      dimensions: existing.embedding.dimensions,
      baseUrl: existing.embedding.base_url,
    }
  : {};
const resolved = resolveEmbeddingConfig({ ...defaults, ...parseEmbeddingFlags(args) });

Can take a crack at a PR if that shape looks right to you, or if you'd prefer a different fallback order let me know.