Harness and Lakebridge

Lakebridge is the deterministic analyzer/converter. Airlift invokes its pre-provisioned workspace jobs through the Harness Databricks public adapter and records pinned versions, run references, artifact references, and digests.

Harness also defines the finite repair agent used for admitted residue. The agent returns one typed candidate under strict runtime, model, size, iteration, tool, network, filesystem, and command bounds. It never receives decision authority.

The in-App Airlift Migration Assistant composes the same ownership boundary. Airlift derives the canonical blocker, owner, next action, and required evidence from its projection. A finite Harness guidance agent can add explanation and bounded diagnostics through Unity AI Gateway, Model Serving, or a secret-backed provider. An optional Databricks Genie Agent remains a read-only analytics specialist. The Harness repair agent is the bounded lane that may produce a candidate artifact, and Platform remains the only mutation path.

Airlift projection -> deterministic answer (always available)
                   -> Harness Migration Assistant explanation (optional, App principal)
                   -> Genie analytics (optional, user OBO)
                   -> highlighted governed Airlift action
                   -> Harness repair candidate when the selected lane allows it
                   -> Experiments validation -> independent review

Harness Temporal owns agent-session durability and common worker/connection plumbing. Airlift owns its application-domain cutover workflow using the Temporal TypeScript SDK; all I/O remains in activities. This ownership distinction is enforced by the family decision map and capability tests.

Conversion worker contract

For each running batch, a worker calls conversion_start with batchId, objectId, method, and the exact toolVersion pinned on the batch. Airlift rejects mismatches. The worker then calls conversion_record with the terminal outcome and provenance. For a successful attempt, it also calls artifact_register with the stored output's SHA-256 digest. For a terminal deterministic failure, it calls residue_create rather than burying the diagnostic in job output.

The repair agent may create and estimate a residue and may submit a resolution after a person assigns the work. It cannot assign work, review its own candidate, certify parity, approve a wave, waive policy, deploy, or cut over.

Modernization advisor

The bounded modernization advisor receives only governed metadata: source system, object type, candidate Databricks services, and the certified-baseline digest. It returns up to five typed, measurable proposals. It has no tools, network, filesystem, deployment, or approval capability.

The worker adapter submits each proposal through airlift.modernization_recommend with a stable replay key and immutable Harness run reference. That action is the only authority the advisor receives. A natural person must disposition the proposal, Runway must deploy the separate release, Experiments must compare it with the baseline, and a different person must promote it.