soma-agentd model-provider status

soma-agentd is no longer a standalone binary and no longer an OpenAI-compatible model provider or proxy.

It ships as a library crate (backend/crates/agentd, embedded by the Tauri host via the desktop-agent crate) that the src-tauri host loads in-process. Model/provider access for chat/embed/rerank goes through desktop-agent's OpenAI-compatible client; the provider configuration source of truth is the Tauri store, surfaced to the renderer through @soma/sdk.

What soma-agentd does now

• Library crate: backend/crates/agentd — exposes an in-process AgentHandle owned by the Tauri host's desktop-agent crate.
• Local helpers retained:
  • agentStatus
  • listModels (returns an empty model list; the desktop side talks to Ollama / a remote endpoint instead)
  • resolveDrift (Yjs update merge)
• The Unix-socket gRPC surface, the --socket-path / SOMA_AGENTD_SOCKET flags, and the persisted background-task store are gone (the desktop side keeps an in-memory background-task store in JS).

Configuration

The agent runtime shares the single embedded SQLite database with the daemon (its tables fold in from the former agentd.db). There are no SOMA_AGENTD_* environment variables.

The previous SOMA_AGENTD_PROVIDER_BASE_URL, SOMA_AGENTD_PROVIDER_API_KEY, SOMA_AGENTD_DEFAULT_CHAT_MODEL, SOMA_AGENTD_DEFAULT_EMBED_MODEL, and SOMA_AGENTD_REQUEST_TIMEOUT_MS knobs are no longer agentd options — model-provider configuration lives in the desktop config layer instead.