(optional) Model-abstraction layer for vendor-independent execution

[oesteban]

Note

Exploratory / optional — may or may not be implemented. This is a forward-looking probe, not a scoped commitment; its status depends on priorities and resources. Nothing on the project's critical path depends on it.

Objective

Build a thin model-abstraction layer so the Workflow Architect (#1) is not tied to a single model vendor's API, and document how an open-weights model would slot in behind the same interface.

Rationale

Keeping the orchestrator vendor-independent avoids architectural lock-in and opens the door to running the agent on open-weights models — the precondition for a programmatic, open-weights multiverse later on. Because none of the funded deliverables depend on it, it is deliberately scoped as exploratory.

Methodological approach

1. Factor the orchestrator's model calls behind a small provider-agnostic interface.
2. Document the contract an alternative (e.g. open-weights) model must satisfy to slot in.
3. If resources allow: stand up one open-weights backend and smoke-test a single compilation through it.

Outcomes/impact

1. A provider-agnostic interface for the orchestrator's model calls (exploratory).
2. Documented requirements for an open-weights backend — enabling the optional open-weights arm of the counterfactual experiment (Counterfactual-KB test: Does the KB drive synthesis? #4) as a future stretch.

Direct upstream issues

(issues this one depends on / builds upon)

• Workflow Architect Agent 1.0 #1

Direct downstream issues

(issues that depend on this one)

None on the board yet.