Context
Mabud noticed BigSet still has code-level model choices. Valid ask: we should let users/admins choose cheaper or better models instead of baking one expensive path into source.
Current origin/main proof:
backend/src/pipeline/schema-inference.ts returns openrouter("anthropic/claude-sonnet-4-6").backend/src/mastra/agents/populate.ts uses openrouter("moonshotai/kimi-k2-0905") for the populate orchestrator.backend/src/mastra/agents/investigate.ts uses openrouter("moonshotai/kimi-k2-0905") for investigate subagents..env.example only exposes OPENROUTER_API_KEY; there is no product/admin provider + model config surface.
OpenRouter is already wired, so the first useful version can support multiple OpenRouter model IDs before adding direct provider adapters.
Problem
AI model choice is not product/config controlled. That makes it hard to:
- switch away from expensive Claude/Sonnet paths when cost spikes;
- test cheaper models like Gemini Flash or Kimi variants;
- route schema inference, populate orchestration, and investigate subagents to different models;
- add provider-specific key/config management later without another rewrite.
Suggested scope
- Add a backend-owned config contract for provider/model selection.
- Support OpenRouter multi-model selection first.
- Keep direct Gemini/Grok/Minimax/GLM provider adapters as follow-up unless we explicitly need direct billing/key management now.
- Add a Settings page for provider/model configuration if this should be user/admin configurable, not env-only.
- Validate provider/model config in the backend; frontend should not be source of truth.
- Keep API logic in the backend. Frontend should call backend settings endpoints rather than adding Next API routes.
Acceptance criteria
- No hardcoded production model literals remain in schema inference / populate / investigate paths except typed defaults or registry entries.
- Backend has typed config for at least:
- schema inference model;
- populate orchestrator model;
- investigate subagent model.
- OpenRouter model options include current defaults plus at least one cheaper alternative.
- Missing or invalid model/provider config fails with a clear error.
- Settings UI, if included, shows current provider/model and persists updates through backend-owned API.
- Docs and
.env.example describe the default model settings without committing secrets.
- Tests cover config parsing/defaults and model selection behavior.
Context
Mabud noticed BigSet still has code-level model choices. Valid ask: we should let users/admins choose cheaper or better models instead of baking one expensive path into source.
Current
origin/mainproof:backend/src/pipeline/schema-inference.tsreturnsopenrouter("anthropic/claude-sonnet-4-6").backend/src/mastra/agents/populate.tsusesopenrouter("moonshotai/kimi-k2-0905")for the populate orchestrator.backend/src/mastra/agents/investigate.tsusesopenrouter("moonshotai/kimi-k2-0905")for investigate subagents..env.exampleonly exposesOPENROUTER_API_KEY; there is no product/admin provider + model config surface.OpenRouter is already wired, so the first useful version can support multiple OpenRouter model IDs before adding direct provider adapters.
Problem
AI model choice is not product/config controlled. That makes it hard to:
Suggested scope
Acceptance criteria
.env.exampledescribe the default model settings without committing secrets.