domains / families / science / nvflare-autofl

nvflare-autofl

separate

Use for agent-assisted Auto-FL optimization of an existing NVFLARE job in simulation, POC, or production. Do not use for code conversion, diagnosis-only work, or deployment setup.

When to use

• Improve a measured objective for an existing NVFLARE job through isolated, reproducible candidate changes while the campaign runner preserves the live best source, comparison budget, metric semantics, state, and evidence.

When not to use

• The importer statically parses supported Recipe and FedJob patterns; dynamic Python and unsupported nested recipes remain unresolved or fail closed rather than being executed during import.
• Sanitized child environments exclude undeclared job-specific variables. Declare only required names through `environment.simulator_env_passthrough`; values remain runtime-only.
• Candidate source runs with the runner's host privileges. Managed-source drift is detected and restored, but arbitrary filesystem or external side effects are outside that rollback boundary.
• A single or noisy campaign does not establish robustness. Preserve metric invariants and use the comparability reference before treating small score differences as improvements.
• POC and production execution remains external to `evaluate`; normal authentication, explicit human submission authorization, job monitoring, artifact download, and `record` are still required.

Skill page

Source