soma-agentd model providers (OpenAI-compatible)

soma-agentd no longer runs local llama.cpp/GGUF inference inside the daemon process.

Instead, it proxies to an OpenAI-compatible HTTP endpoint (for example Ollama or LM Studio), while keeping the same local gRPC surface for Soma desktop.

What soma-agentd does now

• Binary: backend/bins/agentd (cargo run -p soma-agentd)
• IPC transport: Unix socket gRPC (proto/agent/v1/agent.proto)
• Upstream model provider: OpenAI-compatible HTTP APIs:
  • GET /models
  • POST /chat/completions
  • POST /embeddings

Configuration

soma-agentd CLI/env options:

• --socket-path / SOMA_AGENTD_SOCKET (default: /tmp/soma-agentd.sock)
• --provider-base-url / SOMA_AGENTD_PROVIDER_BASE_URL (default: http://127.0.0.1:11434/v1)
• --provider-api-key / SOMA_AGENTD_PROVIDER_API_KEY (optional bearer token)
• --default-chat-model / SOMA_AGENTD_DEFAULT_CHAT_MODEL (default: llama3.2)
• --default-embed-model / SOMA_AGENTD_DEFAULT_EMBED_MODEL (default: nomic-embed-text)
• --request-timeout-ms / SOMA_AGENTD_REQUEST_TIMEOUT_MS (default: 30000)

Example:

cargo run -p soma-agentd -- \
  --socket-path /tmp/soma-agentd-dev.sock \
  --provider-base-url http://127.0.0.1:11434/v1 \
  --default-chat-model llama3.2 \
  --default-embed-model nomic-embed-text

Model capabilities in Soma UI

Provider /models responses do not guarantee capability metadata (chat/embed/tool/image).

Soma therefore treats capabilities as local UI configuration:

• Global capability map lives in the Soma settings (agent.config in electron-store).
• Per-workspace overrides also live in agent.config.workspaces[space_id].
• These capability hints are local-only and are never synced to soma-daemon/network peers.

Notes

• ResolveDrift remains local and is still handled by soma-agentd (Yjs merge).
• Rerank is implemented via embeddings + cosine similarity in soma-agentd.
• If provider requests fail, check URL, API key, and that the upstream endpoint exposes OpenAI-compatible routes.