skilloptlisted

# 🛠️ Skill: SkillOpt (Rule 71 Compliance) ## Domain - **Category:** Infrastructure / Optimization - **Origin:** Third-Party Concept Integration (Microsoft Research) - **Status:** `ACTIVE_LOCAL` ## Technical Logic This tool integrates Microsoft's SkillOpt using a separate isolated virtual environment (`.agents/venv_skillopt/`) and runtime monkeypatching. It registers the `agents_opt` benchmark adapter and routes model generation to `gemini_backend.py` without mutating PyPI package binaries. ## Provisioning (on demand) The session bootstrap only installs the lean core (`requirements-core.txt`). Before the FIRST training run in an environment, install this skill's heavy stack: ```bash .agents/venv_skillopt/bin/pip install -r .agents/requirements-skillopt.txt ``` This transitively pulls in `azure-*` packages (recorded in `docs/audits/SKILLOPT_TRANSITIVE_CLOSURE.md`, which replaced `requirements-freeze.txt` in Sprint 023 `C7` so that a scanner stops reading an optional closure as this framework's own manifest) — the vendored `skillopt` package ships an `azure_openai` backend module by default. It is never called: `train_runner.py apply_monkeypatches()` imports that module only as a patch target and overwrites its functions to route to `gemini_backend.py` (or the Claude Agent SDK, via `model_backend: claude` in the config — see `configs/agents_opt.yaml`). Do not "clean up" the azure packages; they are a real, if unused, transitive dependency of the pinned `skillopt` version. #