First-run setup

A fresh install opens first-run setup until it is completed once in that desktop profile; completion is recorded in onboarding.json in the desktop profile folder. An install that already had Synoec data (organization.json in Synoec’s home) before the launch skips it. Launch with --desktop-onboarding (or SYNOEC_DESKTOP_ONBOARDING=1) to see it on an existing install, and add --desktop-test-profile=<folder> to try it without touching your own data.

It walks through your name, Look, Providers, Profiles, Assign, Persona, Safety and Install, then Set it up saves each choice through the same calls Settings uses, one card at a time, and stops at the first failure. Full Install also turns on local memory for the workspace (the embedding model, a search-index build and Inspect status until it reports healthy), then installs and checks local SearXNG search and scraping. Limited Install installs nothing; add those later from Settings. Setup runs before any project exists, so it works from the workspace’s own chat folder and registers no project.

On Windows the installer asks about local AI before Synoec first opens: a running Ollama is connected at first launch (and, if chosen, started whenever Synoec starts), and a chosen Bonsai 2 appears on Providers as soon as its setup begins. Providers does not add Ollama by itself on Windows. On a Mac, when Providers opens, Synoec asks http://127.0.0.1:11434 whether Ollama is running. If it is, and no Connection reaches it yet, Synoec adds an Ollama no-key Connection (the Ollama provider, which talks to Ollama’s native API) and activates every pulled chat model (embedding models are skipped). The native API lets Synoec choose each model’s context window per request: a Modelfile num_ctx when one is set, or else the model’s own length capped at the larger of 48K and the server’s default (OLLAMA_CONTEXT_LENGTH or the Ollama app’s context slider, read from a model already loaded; Synoec loads nothing to find out). That window is then cut to what fits on the GPU: 90% of the largest NVIDIA card’s memory (on a Mac, two thirds of unified memory), less the model’s weights and a GiB for compute, divided by the model’s KV-cache bytes per token. It never goes below 32K, which Synoec’s Agent requests need. So a 256K slider on a 16 GB card no longer pushes the model into system memory. The output limit is an eighth of the window, at most 8,192 tokens. Profiles are not changed. The Local tab’s Ollama tile then shows the connection, or says the context is too small when a Modelfile sets less than 32K tokens.

Opening Settings → Profiles & Providers checks that Ollama again and activates a model with a new window when its default or Modelfile changed (for example after raising OLLAMA_CONTEXT_LENGTH). Only the Connection setup added is updated; models you turned off stay off.