---

name: evo-kit description: "Build, inspect, and safely use PARL EvoKit, the optional C++ evolution-strategy toolkit with Torch or PaddleLite prediction backends." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

PARL EvoKit

Use this sub-skill when a task involves PARL's optional C++ EvoKit toolkit: ESAgent sampling/update loops, SamplingInfo records, C++ build prerequisites, Torch/libtorch versus PaddleLite backend selection, or safe interpretation of EvoKit build helpers.

Do not use this sub-skill for ordinary Python PARL Model / Algorithm / Agent code, xparl cluster orchestration, or selecting a full Python RL recipe. For high-level RL recipe structure, consult the sibling algorithm-recipes skill conceptually; EvoKit-specific work here is C++ ES toolkit work.

Operating workflow

1. Confirm EvoKit is actually needed. EvoKit is a C++ evolution-strategy toolkit for parameter perturbation, evaluation, and parameter updates. If the task only asks for a Python PARL algorithm, stay with the Python PARL skills instead.

2. Choose one prediction backend. Use Torch when the model is a C++ torch::nn::Module and a local libtorch/ tree is available. Use PaddleLite when the model is an exported Paddle inference model and a local inference_lite_lib/ tree is available. Do not enable both backends in one EvoKit build.

3. Check local prerequisites before building. Run the bundled checker from a repository checkout or a copied EvoKit tree:

   python scripts/check_evokit_prereqs.py --project-root <evokit-root> --backend torch
   

   Replace torch with paddle when using PaddleLite, or use auto to accept either backend directory.

4. Model the ESAgent loop explicitly. The original agent owns the current parameters, sampling method, and optimizer. Clone sampling agents, call add_noise only on clones, evaluate each noisy clone, keep each SamplingInfo paired with its reward, then call update on the original agent. The update reconstructs perturbations from SamplingInfo keys and applies the optimizer to the original parameters.

5. Handle online/offline runs as data contracts. Online sampling should persist SamplingInfo plus reward records in the same order used for offline update. Offline update should reload the current model/solver state, parse the records, call update, and save the next model/solver state.

6. Avoid unsafe build shortcuts by default. Treat upstream shell helpers as examples only: they may remove build directories, install into the source tree, unzip bundled data, download libtorch, and run demos. Prefer an explicit local prerequisite check and a controlled manual CMake build in a disposable checkout or dedicated build directory.

7. Escalate by symptom. Use references/build-and-api.md for API/build planning and references/troubleshooting.md for protobuf, backend library, CMake, gflags/glog/OpenMP, and ESAgent misuse errors.

Verification status

This sub-skill is source-evidence backed. The bundled checker was written to inspect local prerequisites without downloads or builds. EvoKit itself was not fully compiled during skill construction because the optional C++ backend libraries are environment-specific.

Reference map

• references/build-and-api.md — distilled EvoKit C++ API concepts, config schema, Torch/PaddleLite choices, and safe build outline.
• references/troubleshooting.md — fixes and safety notes for protobuf/protoc, backend libraries, gflags/glog/OpenMP, build helper side effects, CMake cleanup, and ESAgent loop mistakes.
• scripts/check_evokit_prereqs.py — deterministic local prerequisite checker; it prints missing items and never downloads, builds, deletes, or writes project files.