Skip to content

2026-09-02 Infrastructure Factory inference restoration review

Adversarial review of the Factory ad-hoc inference API, CLI, and Argo submission restoration.

Work Done

Reviewed the uncommitted Factory inference restoration in the virdx/mono campaign-controller branch. Traced the FastAPI and CLI surfaces through request validation, worker selection, case batching, Kubernetes Leases, deterministic Argo names, and the worker payload entrypoints. Compared the change with the Factory inference surface that existed before the campaign-controller rewrite. Ran frozen formatting, lint, type checking, lock validation, OpenAPI inspection, and focused executable probes without editing the mono worktree.

The initial review found that inference Lease contention was not translated correctly by either new surface, and that the restored histo endpoint had lost its prior per-request memory sizing field while silently ignoring it. A focused follow-up verified both fixes: HTTP and CLI now report contention correctly, and histo memory sizes only the worker while the worker payload remains the narrower algorithm parameter model.

Lessons Learned: Pitfalls

The general infrastructure and histo indexes did not record the prior Factory ad-hoc inference request contract. Git history was required to distinguish intentional campaign-controller changes from compatibility regressions.

Lessons Learned: Improvements

A short Factory API compatibility note should list the stable ad-hoc routes, body shape, response shape, worker batch sizes, and caller-controlled resource fields. It should also state how Lease contention maps to HTTP and CLI errors.

Navigation

Type to search…

↑↓ navigate↵ selectEsc close