---

name: strategies-and-mods description: "Select, customize, and debug Flower strategies, aggregation settings, client mods, and related differential-privacy or secure-aggregation behavior." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Flower Strategies and Mods

Use this sub-skill when the task is about choosing a strategy, customizing the round lifecycle, aggregating metrics, or understanding Flower client mods and strategy wrappers.

Read this when

• the task names FedAvg, FedAdam, FedAdagrad, FedProx, FedAvgM, FedYogi, FedMedian, FedTrimmedAvg, Krum, MultiKrum, Bulyan, or QFedAvg;
• the task says "custom strategy", "override configure_train", "change the evaluation config", or "use a mod";
• the task mentions differential privacy, secure aggregation, or aggregation callbacks.

Do not route here

• Basic app wiring, ServerApp, ClientApp, or tool.flwr.app config → ../app-development/
• Local simulation, SuperLink/SuperNode, or flwr run profile routing → ../simulation-and-deployment/
• Flower Datasets partitioners and example dependency catalogs → ../datasets-and-examples/
• Contributor commands, protobufs, migrations, and repo tests → ../repository-maintenance/

What to read next

• references/api-reference.md for verified strategy and mod signatures.
• references/workflows.md for strategy selection and customization patterns.
• references/troubleshooting.md for skipped rounds, aggregation keys, and mod-order issues.
• ../../scripts/check_flower_install.py if you also want a safe import/version/CLI smoke from the current environment.

Core model

A Flower strategy decides which nodes participate in a round, what message payload to send, how to aggregate replies, and how to interpret the round's metrics. Built-in strategies give different defaults for sampling, robust aggregation, optimizer-like behavior, or privacy wrappers.

Client mods are callables that wrap a ClientApp before and after a message is handled. They are the right tool when the behavior change is orthogonal to the main training or evaluation code.

When this sub-skill is enough

If the user only needs the strategy or mod logic, stay here. If the request turns into app structure, runtime routing, or dataset preparation, hand off to that owner instead of widening this one.