Work Done
Reviewed the current mono Factory controller and dashboard changes plus the coordinated infra_k8s diff. Checked campaign-scoped Kubernetes Lease acquisition, expiry takeover, renewal, release, crash recovery, per-case duplicate prevention, one-worker dispatch behavior, API and CLI errors, Pixi-only CLI setup, backend and TypeScript case unions, removal of Factory tests and wiring, image mappings, and Lease RBAC. Reported three findings: the Lease is not renewed through the final Kubernetes list/create section, the backend union lacks an explicit OpenAPI discriminator, and obsolete Factory Dockerfile mappings remain in the Tekton policy. Confirmed the Lease RBAC targets the service account used by the Factory pod. Ran Factory typecheck, lint, and format-check; all passed. Did not run or request tests because the user explicitly prohibited Factory tests.
Lessons Learned: Pitfalls
The product workstream has no SOP index, so infrastructure was the useful context for this cross-repository review. A post-evaluation Lease ownership check narrows the critical race but does not prove ownership through later unbounded Kubernetes calls. Exact-batch workflow names prevent only identical-batch duplicates, not overlapping batches with different hashes.
Lessons Learned: Improvements
Factory operational guidance should state that a campaign Lease must remain owned through workflow creation, and that deterministic batch names are a fallback for exact retries rather than a substitute for per-case locking. The image-name policy should be kept aligned with Dockerfile moves so obsolete mappings do not accumulate.