菜单
Run ML training pipeline steps (specialists, core, evaluation, forecasts, monte-carlo). Use when user asks to run or re-run any part of the prediction pipeline.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when the user says "review my code", "check my changes", "run a review", "anything I should fix before merging", or wants a pre-commit or pre-PR quality check on local changes. Runs both CodeScene (structural code health) and cubic (AI semantic review) then surfaces all issues grouped by severity.
Run the full project verification gate. Use before committing or when user says "verify", "check", or "run checks".
Use when the user asks about codebase architecture, "how does X work", system design, onboarding to a project, or needs context about how parts of the codebase connect. Queries cubic's AI Wiki for the current repository to provide architectural context.
Use when the user says "check cubic comments", "cubic issues", "cubic feedback", "cubic code review", "fix issues", "address review", or is working on a pull request branch with open review comments. Fetches all PR issues from cubic, investigates each one, and reports back which are worth fixing before taking action.
Use when the user is writing code, reviewing code, asking about code quality standards, or wondering "what does this team care about in reviews". Pulls team-specific review learnings from cubic to apply the team's coding conventions and review preferences.
| name | run-pipeline |
| description | Run ML training pipeline steps (specialists, core, evaluation, forecasts, monte-carlo). Use when user asks to run or re-run any part of the prediction pipeline. |
| disable-model-invocation | true |
Run the ML pipeline step specified in $ARGUMENTS.
git branch --show-currentwhich python| Step | Command | Description |
|---|---|---|
specialists | python scripts/generate_specialist_signals.py | Generate Big-11 specialist signals |
features | python scripts/generate_specialist_features.py | Build specialist feature columns |
core | python -m fusion.core_training.run_pipeline | Train 4 AutoGluon ensembles (5d/21d/63d/126d) |
evaluate | python scripts/evaluate_oof.py | Evaluate OOF predictions vs actuals |
forecasts | python scripts/generate_production_forecasts.py | Generate production forecast outputs |
monte-carlo | python scripts/autorun_monte_carlo.py | Run L3 Monte Carlo simulation (10,000 runs) |
all | Run all steps above in order | Full pipeline end-to-end |
source .venv/bin/activateexport PYTHONPATH=src:$PYTHONPATHmake check — abort if it failssrc/fusion/core_training/config.py — it is FROZEN (2026-02-19)close.shift(-horizon)), named target_price_{h}devaluate, show the OOF summary tablecore, confirm all 4 horizons completed with WeightedEnsemble