Repository navigation
v0.37.0
A provider and a model for every LLM layer
A project is no longer limited to one LLM provider and one model. It can hold multiple provider entries, and choose a provider and model per layer — with per-layer reasoning, prompt-cache and output-cap settings.
Every layer that calls an LLM resolves its own: exploration, analysis, recommendations, reflection, schema blurbs, questions, verification and refutation, Ask and its router and suggestions, and the rest.
Why it matters
Different layers want different models. Exploration is a long, cheap, high-volume loop; verification wants the most capable model available; blurb generation is throwaway. Forcing all of them onto one model meant paying top-tier prices for throwaway work, or accepting weak verification to keep the bill down.
A project with no models chosen is sent to settings
Rather than failing mid-run, a project that has not picked its models is gated to the settings page. Managed deployments are exempt — they have no AI configuration surface to visit.
Verified end to end
Exercised with a different model behind every layer, driven through a managed gateway: 18 of 18 layers took real traffic and each landed on its own model.
Cloud: no action required
On a managed deployment each layer derives its own gateway alias (decisionbox-<layer>-model). The gateway falls back to the analysis alias for any layer it has not been told about, so an existing tenant keeps working with no routing change and no provisioning change.
Images are cosign-signed; the tag is GPG-signed. See docs/verifying-artifacts.md.